Compare commits
1
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 thomwolf
|
32fef30c60 |
+5
-39
@@ -9,7 +9,7 @@ jobs:
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install torch
|
||||
- run: sudo pip install tensorflow
|
||||
- run: sudo pip install tensorflow==2.0.0-rc0
|
||||
- run: sudo pip install --progress-bar off .
|
||||
- run: sudo pip install pytest codecov pytest-cov
|
||||
- run: sudo pip install tensorboardX scikit-learn
|
||||
@@ -38,7 +38,7 @@ jobs:
|
||||
parallelism: 1
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install tensorflow
|
||||
- run: sudo pip install tensorflow==2.0.0-rc0
|
||||
- run: sudo pip install --progress-bar off .
|
||||
- run: sudo pip install pytest codecov pytest-cov
|
||||
- run: sudo pip install tensorboardX scikit-learn
|
||||
@@ -65,32 +65,11 @@ jobs:
|
||||
- image: circleci/python:2.7
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install tensorflow
|
||||
- run: sudo pip install tensorflow==2.0.0-rc0
|
||||
- run: sudo pip install --progress-bar off .
|
||||
- run: sudo pip install pytest codecov pytest-cov
|
||||
- run: python -m pytest -sv ./transformers/tests/ --cov
|
||||
- run: codecov
|
||||
build_py3_custom_tokenizers:
|
||||
working_directory: ~/transformers
|
||||
docker:
|
||||
- image: circleci/python:3.5
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install --progress-bar off .
|
||||
- run: sudo pip install pytest
|
||||
- run: sudo pip install mecab-python3
|
||||
- run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
|
||||
build_py2_custom_tokenizers:
|
||||
working_directory: ~/transformers
|
||||
docker:
|
||||
- image: circleci/python:2.7
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install --progress-bar off .
|
||||
- run: sudo pip install pytest
|
||||
- run: sudo apt-get -y install libmecab-dev mecab mecab-ipadic-utf8 swig
|
||||
- run: sudo pip install mecab-python
|
||||
- run: RUN_CUSTOM_TOKENIZERS=1 python -m pytest -sv ./transformers/tests/tokenization_bert_japanese_test.py
|
||||
deploy_doc:
|
||||
working_directory: ~/transformers
|
||||
docker:
|
||||
@@ -102,17 +81,7 @@ jobs:
|
||||
- checkout
|
||||
- run: sudo pip install --progress-bar off -r docs/requirements.txt
|
||||
- run: sudo pip install --progress-bar off -r requirements.txt
|
||||
- run: ./.circleci/deploy.sh
|
||||
repository_consistency:
|
||||
working_directory: ~/transformers
|
||||
docker:
|
||||
- image: circleci/python:3.5
|
||||
resource_class: small
|
||||
parallelism: 1
|
||||
steps:
|
||||
- checkout
|
||||
- run: sudo pip install requests
|
||||
- run: python ./utils/link_tester.py
|
||||
- run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
|
||||
workflow_filters: &workflow_filters
|
||||
filters:
|
||||
branches:
|
||||
@@ -122,12 +91,9 @@ workflows:
|
||||
version: 2
|
||||
build_and_test:
|
||||
jobs:
|
||||
- repository_consistency
|
||||
- build_py3_custom_tokenizers
|
||||
- build_py2_custom_tokenizers
|
||||
- build_py3_torch_and_tf
|
||||
- build_py3_torch
|
||||
- build_py3_tf
|
||||
- build_py2_torch
|
||||
- build_py2_tf
|
||||
- deploy_doc: *workflow_filters
|
||||
- deploy_doc: *workflow_filters
|
||||
@@ -1,26 +0,0 @@
|
||||
cd docs
|
||||
|
||||
function deploy_doc(){
|
||||
echo "Creating doc at commit $1 and pushing to folder $2"
|
||||
git checkout $1
|
||||
if [ ! -z "$2" ]
|
||||
then
|
||||
if [ -d "$dir/$2" ]; then
|
||||
echo "Directory" $2 "already exists"
|
||||
else
|
||||
echo "Pushing version" $2
|
||||
make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
|
||||
fi
|
||||
else
|
||||
echo "Pushing master"
|
||||
make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
|
||||
fi
|
||||
}
|
||||
|
||||
deploy_doc "master"
|
||||
deploy_doc "b33a385" v1.0.0
|
||||
deploy_doc "fe02e45" v1.1.0
|
||||
deploy_doc "89fd345" v1.2.0
|
||||
deploy_doc "fc9faa8" v2.0.0
|
||||
deploy_doc "3ddce1d" v2.1.1
|
||||
deploy_doc "3616209" v2.2.0
|
||||
@@ -1,22 +0,0 @@
|
||||
---
|
||||
name: "\U0001F5A5 New Benchmark"
|
||||
about: You benchmark a part of this library and would like to share your results
|
||||
title: "[Benchmark]"
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
# Benchmarking Transformers
|
||||
|
||||
## Benchmark
|
||||
|
||||
Which part of Transformers did you benchmark?
|
||||
|
||||
## Set-up
|
||||
|
||||
What did you run your benchmarks on? Please include details, such as: CPU, GPU? If using multiple GPUs, which parallelization did you use?
|
||||
|
||||
## Results
|
||||
|
||||
Put your results here!
|
||||
@@ -1,24 +0,0 @@
|
||||
---
|
||||
name: "\U0001F31FNew model addition"
|
||||
about: Submit a proposal/request to implement a new Transformer-based model
|
||||
title: ''
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
# 🌟New model addition
|
||||
|
||||
## Model description
|
||||
|
||||
<!-- Important information -->
|
||||
|
||||
## Open Source status
|
||||
|
||||
* [ ] the model implementation is available: (give details)
|
||||
* [ ] the model weights are available: (give details)
|
||||
* [ ] who are the authors: (mention them)
|
||||
|
||||
## Additional context
|
||||
|
||||
<!-- Add any other context about the problem here. -->
|
||||
@@ -1,10 +1,6 @@
|
||||
---
|
||||
name: "\U0001F41B Bug Report"
|
||||
about: Submit a bug report to help us improve PyTorch Transformers
|
||||
title: ''
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
## 🐛 Bug
|
||||
@@ -49,4 +45,4 @@ Steps to reproduce the behavior:
|
||||
|
||||
## Additional context
|
||||
|
||||
<!-- Add any other context about the problem here. -->
|
||||
<!-- Add any other context about the problem here. -->
|
||||
@@ -1,10 +1,6 @@
|
||||
---
|
||||
name: "\U0001F680 Feature Request"
|
||||
about: Submit a proposal/request for a new PyTorch Transformers feature
|
||||
title: ''
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
## 🚀 Feature
|
||||
@@ -17,4 +13,4 @@ assignees: ''
|
||||
|
||||
## Additional context
|
||||
|
||||
<!-- Add any other context or screenshots about the feature request here. -->
|
||||
<!-- Add any other context or screenshots about the feature request here. -->
|
||||
@@ -1,10 +1,6 @@
|
||||
---
|
||||
name: "\U0001F4DA Migration from PyTorch-pretrained-Bert"
|
||||
about: Report a problem when migrating from PyTorch-pretrained-Bert to Transformers
|
||||
title: ''
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
## 📚 Migration
|
||||
@@ -44,4 +40,4 @@ Details of the issue:
|
||||
|
||||
## Additional context
|
||||
|
||||
<!-- Add any other context about the problem here. -->
|
||||
<!-- Add any other context about the problem here. -->
|
||||
@@ -1,12 +1,8 @@
|
||||
---
|
||||
name: "❓Questions & Help"
|
||||
about: Start a general discussion related to PyTorch Transformers
|
||||
title: ''
|
||||
labels: ''
|
||||
assignees: ''
|
||||
|
||||
---
|
||||
|
||||
## ❓ Questions & Help
|
||||
|
||||
<!-- A clear and concise description of the question. -->
|
||||
<!-- A clear and concise description of the question. -->
|
||||
+1
-8
@@ -118,9 +118,6 @@ dmypy.json
|
||||
# vscode
|
||||
.vscode
|
||||
|
||||
# Pycharm
|
||||
.idea
|
||||
|
||||
# TF code
|
||||
tensorflow_code
|
||||
|
||||
@@ -134,8 +131,4 @@ examples/runs
|
||||
|
||||
# data
|
||||
/data
|
||||
serialization_dir
|
||||
|
||||
# emacs
|
||||
*.*~
|
||||
debug.env
|
||||
serialization_dir
|
||||
-179
@@ -1,179 +0,0 @@
|
||||
# How to contribute to transformers?
|
||||
|
||||
Everyone is welcome to contribute, and we value everybody's contribution. Code
|
||||
is thus not the only way to help the community. Answering questions, helping
|
||||
others, reaching out and improving the documentations are immensely valuable to
|
||||
the community.
|
||||
|
||||
It also helps us if you spread the word: reference the library from blog posts
|
||||
on the awesome projects it made possible, shout out on Twitter every time it has
|
||||
helped you, or simply star the repo to say "thank you".
|
||||
|
||||
## You can contribute in so many ways!
|
||||
|
||||
There are 4 ways you can contribute to transformers:
|
||||
* Fixing outstanding issues with the existing code;
|
||||
* Implementing new models;
|
||||
* Contributing to the examples or to the documentation;
|
||||
* Submitting issues related to bugs or desired new features.
|
||||
|
||||
*All are equally valuable to the community.*
|
||||
|
||||
## Submitting a new issue or feature request
|
||||
|
||||
Do your best to follow these guidelines when submitting an issue or a feature
|
||||
request. It will make it easier for us to come back to you quickly and with good
|
||||
feedback.
|
||||
|
||||
### Did you find a bug?
|
||||
|
||||
The transformers are robust and reliable thanks to the users who notify us of
|
||||
the problems they encounter. So thank you for reporting an issue.
|
||||
|
||||
First, we would really appreciate it if you could **make sure the bug was not
|
||||
already reported** (use the search bar on Github under Issues).
|
||||
|
||||
Did not find it? :( So we can act quickly on it, please follow these steps:
|
||||
|
||||
* Include your **OS type and version**, the versions of **Python**, **PyTorch** and
|
||||
**Tensorflow** when applicable;
|
||||
* A short, self-contained, code snippet that allows us to reproduce the bug in
|
||||
less than 30s;
|
||||
* Provide the *full* traceback if an exception is raised.
|
||||
|
||||
To get the OS and software versions, execute the following code and copy-paste
|
||||
the output:
|
||||
|
||||
```
|
||||
import platform; print("Platform", platform.platform())
|
||||
import sys; print("Python", sys.version)
|
||||
import torch; print("PyTorch", torch.__version__)
|
||||
import tensorflow; print("Tensorflow", tensorflow.__version__)
|
||||
```
|
||||
|
||||
### Do you want to implement a new model?
|
||||
|
||||
Awesome! Please provide the following information:
|
||||
|
||||
* Short description of the model and link to the paper;
|
||||
* Link to the implementation if it is open-source;
|
||||
* Link to the model weights if they are available.
|
||||
|
||||
If you are willing to contribute the model yourself, let us know so we can best
|
||||
guide you.
|
||||
|
||||
We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder.
|
||||
|
||||
### Do you want a new feature (that is not a model)?
|
||||
|
||||
A world-class feature request addresses the following points:
|
||||
|
||||
1. Motivation first:
|
||||
* Is it related to a problem/frustration with the library? If so, please explain
|
||||
why. Providing a code snippet that demonstrates the problem is best.
|
||||
* Is it related to something you would need for a project? We'd love to hear
|
||||
about it!
|
||||
* Is it something you worked on and think could benefit the community?
|
||||
Awesome! Tell us what problem it solved for you.
|
||||
2. Write a *full paragraph* describing the feature;
|
||||
3. Provide a **code snippet** that demonstrates its future use;
|
||||
4. In case this is related to a paper, please attach a link;
|
||||
5. Attach any additional information (drawings, screenshots, etc.) you think may help.
|
||||
|
||||
If your issue is well written we're already 80% of the way there by the time you
|
||||
post it.
|
||||
|
||||
We have added **templates** to guide you in the process of adding a new example script for training or testing the models in the library. You can find them in the [`templates`](./templates) folder.
|
||||
|
||||
## Start contributing! (Pull Requests)
|
||||
|
||||
Before writing code, we strongly advise you to search through the exising PRs or
|
||||
issues to make sure that nobody is already working on the same thing. If you are
|
||||
unsure, it is always a good idea to open an issue to get some feedback.
|
||||
|
||||
You will need basic `git` proficiency to be able to contribute to
|
||||
`transformers`. `git` is not the easiest tool to use but it has the greatest
|
||||
manual. Type `git --help` in a shell and enjoy. If you prefer books, [Pro
|
||||
Git](https://git-scm.com/book/en/v2) is a very good reference.
|
||||
|
||||
Follow these steps to start contributing:
|
||||
|
||||
1. Fork the [repository](https://github.com/huggingface/transformers) by
|
||||
clicking on the 'Fork' button on the repository's page. This creates a copy of the code
|
||||
under your github user account.
|
||||
2. Clone your fork to your local disk, and add the base repository as a remote:
|
||||
|
||||
```bash
|
||||
$ git clone git@github.com:<your Github handle>/transformers.git
|
||||
$ cd transformers
|
||||
$ git remote add upstream https://github.com/huggingface/transformers.git
|
||||
```
|
||||
|
||||
3. Create a new branch to hold your development changes:
|
||||
|
||||
```bash
|
||||
$ git checkout -b a-descriptive-name-for-my-changes
|
||||
```
|
||||
|
||||
**do not** work on the `master` branch.
|
||||
|
||||
4. Set up a development environment by running the following command in a virtual environment:
|
||||
|
||||
```bash
|
||||
$ pip install -r requirements-dev.txt
|
||||
```
|
||||
|
||||
5. Develop the features on your branch. Add changed files using `git add` and
|
||||
then `git commit` to record your changes locally:
|
||||
|
||||
```bash
|
||||
$ git add modified_file.py
|
||||
$ git commit
|
||||
```
|
||||
|
||||
Please write [good commit
|
||||
messages](https://chris.beams.io/posts/git-commit/). It
|
||||
is a good idea to sync your copy of the code with the original repository
|
||||
regularly. This way you can quickly account for changes:
|
||||
|
||||
```bash
|
||||
$ git fetch upstream
|
||||
$ git rebase upstream/master
|
||||
```
|
||||
|
||||
Push the changes to your account using:
|
||||
|
||||
```bash
|
||||
$ git push -u origin a-descriptive-name-for-my-changes
|
||||
```
|
||||
|
||||
6. Once you are satisfied (**and the checklist below is happy too**), go to the
|
||||
webpage of your fork on Github. Click on 'Pull request' to send your changes
|
||||
to the project maintainers for review.
|
||||
|
||||
7. It's ok if maintainers ask you for changes. It happens to core contributors
|
||||
too! So everyone can see the changes in the Pull request, work in your local
|
||||
branch and push the changes to your fork. They will automatically appear in
|
||||
the pull request.
|
||||
|
||||
|
||||
### Checklist
|
||||
|
||||
1. The title of your pull request should be a summary of its contribution;
|
||||
2. If your pull request adresses an issue, please mention the issue number in
|
||||
the pull request description to make sure they are linked (and people
|
||||
consulting the issue know you are working on it);
|
||||
3. To indicate a work in progress please prefix the title with `[WIP]`. These
|
||||
are useful to avoid duplicated work, and to differentiate it from PRs ready
|
||||
to be merged;
|
||||
4. Make sure pre-existing tests still pass;
|
||||
5. Add high-coverage tests. No quality test, no merge;
|
||||
6. All public methods must have informative doctrings;
|
||||
|
||||
|
||||
### Style guide
|
||||
|
||||
For documentation strings, `transformers` follows the [google
|
||||
style](https://google.github.io/styleguide/pyguide.html).
|
||||
|
||||
#### This guide was heavily inspired by the awesome [scikit-learn guide to contributing](https://github.com/scikit-learn/scikit-learn/blob/master/CONTRIBUTING.md)
|
||||
@@ -22,7 +22,7 @@
|
||||
<p>State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch
|
||||
</h3>
|
||||
|
||||
🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides state-of-the-art general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet, CTRL...) for Natural Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between TensorFlow 2.0 and PyTorch.
|
||||
🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides state-of-the-art general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained models in 100+ languages and deep interoperability between TensorFlow 2.0 and PyTorch.
|
||||
|
||||
### Features
|
||||
|
||||
@@ -39,7 +39,7 @@ State-of-the-art NLP for everyone
|
||||
Lower compute costs, smaller carbon footprint
|
||||
- Researchers can share trained models instead of always retraining
|
||||
- Practitioners can reduce compute time and production costs
|
||||
- 10 architectures with over 30 pretrained models, some in more than 100 languages
|
||||
- 8 architectures with over 30 pretrained models, some in more than 100 languages
|
||||
|
||||
Choose the right framework for every part of a model's lifetime
|
||||
- Train state-of-the-art models in 3 lines of code
|
||||
@@ -56,10 +56,9 @@ Choose the right framework for every part of a model's lifetime
|
||||
| [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
|
||||
| [Quick tour: TF 2.0 and PyTorch ](#Quick-tour-TF-20-training-and-PyTorch-interoperability) | Train a TF 2.0 model in 10 lines of code, load it in PyTorch |
|
||||
| [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
|
||||
| [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
|
||||
| [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
|
||||
| [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
|
||||
| [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
|
||||
| [Documentation][(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
|
||||
| [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
|
||||
|
||||
## Installation
|
||||
|
||||
@@ -68,7 +67,7 @@ This repo is tested on Python 2.7 and 3.5+ (examples are tested only on python 3
|
||||
### With pip
|
||||
|
||||
First you need to install one of, or both, TensorFlow 2.0 and PyTorch.
|
||||
Please refer to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
|
||||
Please refere to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
|
||||
|
||||
When TensorFlow 2.0 and/or PyTorch has been installed, 🤗 Transformers can be installed using pip as follows:
|
||||
|
||||
@@ -79,7 +78,7 @@ pip install transformers
|
||||
### From source
|
||||
|
||||
Here also, you first need to install one of, or both, TensorFlow 2.0 and PyTorch.
|
||||
Please refer to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
|
||||
Please refere to [TensorFlow installation page](https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available) and/or [PyTorch installation page](https://pytorch.org/get-started/locally/#start-locally) regarding the specific install command for your platform.
|
||||
|
||||
When TensorFlow 2.0 and/or PyTorch has been installed, you can install from source by cloning the repository and running:
|
||||
|
||||
@@ -87,52 +86,32 @@ When TensorFlow 2.0 and/or PyTorch has been installed, you can install from sour
|
||||
pip install [--editable] .
|
||||
```
|
||||
|
||||
### Run the examples
|
||||
|
||||
Examples are included in the repository but are not shipped with the library.
|
||||
Therefore, in order to run the latest versions of the examples you also need to install from source. To do so, create a new virtual environment and follow these steps:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/transformers
|
||||
cd transformers
|
||||
pip install [--editable] .
|
||||
```
|
||||
|
||||
### Tests
|
||||
|
||||
A series of tests are included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
|
||||
|
||||
These tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
|
||||
These tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
|
||||
|
||||
Depending on which framework is installed (TensorFlow 2.0 and/or PyTorch), the irrelevant tests will be skipped. Ensure that both frameworks are installed if you want to execute all tests.
|
||||
|
||||
You can run the tests from the root of the cloned repository with the commands:
|
||||
|
||||
```bash
|
||||
python -m unittest discover -s transformers/tests -p "*test.py" -t .
|
||||
python -m unittest discover -s examples -p "*test.py" -t examples
|
||||
```
|
||||
|
||||
or
|
||||
|
||||
```bash
|
||||
python -m pytest -sv ./transformers/tests/
|
||||
python -m pytest -sv ./examples/
|
||||
```
|
||||
|
||||
By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
|
||||
|
||||
### Do you want to run a Transformer model on a mobile device?
|
||||
|
||||
You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
|
||||
|
||||
It contains a set of tools to convert PyTorch or TensorFlow 2.0 trained Transformer models (currently contains `GPT-2`, `DistilGPT-2`, `BERT`, and `DistilBERT`) to CoreML models that run on iOS devices.
|
||||
It contains an example of a conversion script from a Pytorch trained Transformer model (here, `GPT-2`) to a CoreML model that runs on iOS devices.
|
||||
|
||||
At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models to productizing them in CoreML, or prototype a model or an app in CoreML then research its hyperparameters or architecture from TensorFlow 2.0 and/or PyTorch. Super exciting!
|
||||
|
||||
## Model architectures
|
||||
|
||||
🤗 Transformers currently provides 10 NLU/NLG architectures:
|
||||
🤗 Transformers currently provides 8 NLU/NLG architectures:
|
||||
|
||||
1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
|
||||
2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
|
||||
@@ -141,12 +120,8 @@ At some point in the future, you'll be able to seamlessly move from pre-training
|
||||
5. **[XLNet](https://github.com/zihangdai/xlnet/)** (from Google/CMU) released with the paper [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
|
||||
6. **[XLM](https://github.com/facebookresearch/XLM/)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
|
||||
7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
|
||||
8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
|
||||
9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
|
||||
10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
|
||||
11. **[ALBERT](https://github.com/google-research/ALBERT)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
|
||||
12. **[T5](https://github.com/google-research/text-to-text-transfer-transformer)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
|
||||
13. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
|
||||
8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the blogpost [Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT](https://medium.com/huggingface/distilbert-8cf3380435b5
|
||||
) by Victor Sanh, Lysandre Debut and Thomas Wolf.
|
||||
|
||||
These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
|
||||
|
||||
@@ -173,7 +148,6 @@ from transformers import *
|
||||
MODELS = [(BertModel, BertTokenizer, 'bert-base-uncased'),
|
||||
(OpenAIGPTModel, OpenAIGPTTokenizer, 'openai-gpt'),
|
||||
(GPT2Model, GPT2Tokenizer, 'gpt2'),
|
||||
(CTRLModel, CTRLTokenizer, 'ctrl'),
|
||||
(TransfoXLModel, TransfoXLTokenizer, 'transfo-xl-wt103'),
|
||||
(XLNetModel, XLNetTokenizer, 'xlnet-base-cased'),
|
||||
(XLMModel, XLMTokenizer, 'xlm-mlm-enfr-1024'),
|
||||
@@ -195,16 +169,16 @@ for model_class, tokenizer_class, pretrained_weights in MODELS:
|
||||
|
||||
# Each architecture is provided with several class for fine-tuning on down-stream tasks, e.g.
|
||||
BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNextSentencePrediction,
|
||||
BertForSequenceClassification, BertForTokenClassification, BertForQuestionAnswering]
|
||||
BertForSequenceClassification, BertForMultipleChoice, BertForTokenClassification,
|
||||
BertForQuestionAnswering]
|
||||
|
||||
# All the classes for an architecture can be initiated from pretrained weights for this architecture
|
||||
# Note that additional weights added for fine-tuning are only initialized
|
||||
# and need to be trained on the down-stream task
|
||||
pretrained_weights = 'bert-base-uncased'
|
||||
tokenizer = BertTokenizer.from_pretrained(pretrained_weights)
|
||||
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
|
||||
for model_class in BERT_MODEL_CLASSES:
|
||||
# Load pretrained model/tokenizer
|
||||
model = model_class.from_pretrained(pretrained_weights)
|
||||
model = model_class.from_pretrained('bert-base-uncased')
|
||||
|
||||
# Models can return full list of hidden-states & attentions weights at each layer
|
||||
model = model_class.from_pretrained(pretrained_weights,
|
||||
@@ -267,25 +241,19 @@ sentence_2 = "His findings were not compatible with this research."
|
||||
inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
|
||||
inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
|
||||
|
||||
pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
|
||||
pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
|
||||
|
||||
pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
|
||||
pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
|
||||
print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
|
||||
print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
|
||||
```
|
||||
|
||||
## Quick tour of the fine-tuning/usage scripts
|
||||
|
||||
**Important**
|
||||
Before running the fine-tuning scripts, please read the
|
||||
[instructions](#run-the-examples) on how to
|
||||
setup your environment to run the examples.
|
||||
|
||||
The library comprises several example scripts with SOTA performances for NLU and NLG tasks:
|
||||
|
||||
- `run_glue.py`: an example fine-tuning Bert, XLNet and XLM on nine different GLUE tasks (*sequence-level classification*)
|
||||
- `run_squad.py`: an example fine-tuning Bert, XLNet and XLM on the question answering dataset SQuAD 2.0 (*token-level classification*)
|
||||
- `run_generation.py`: an example using GPT, GPT-2, CTRL, Transformer-XL and XLNet for conditional language generation
|
||||
- `run_generation.py`: an example using GPT, GPT-2, Transformer-XL and XLNet for conditional language generation
|
||||
- other model-specific examples (see the documentation).
|
||||
|
||||
Here are three quick usage examples for these scripts:
|
||||
@@ -423,7 +391,7 @@ python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncase
|
||||
|
||||
This is the model provided as `bert-large-uncased-whole-word-masking-finetuned-squad`.
|
||||
|
||||
### `run_generation.py`: Text generation with GPT, GPT-2, CTRL, Transformer-XL and XLNet
|
||||
### `run_generation.py`: Text generation with GPT, GPT-2, Transformer-XL and XLNet
|
||||
|
||||
A conditional generation script is also included to generate text from a prompt.
|
||||
The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by Aman Rusia to get high-quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
|
||||
@@ -437,56 +405,6 @@ python ./examples/run_generation.py \
|
||||
--model_name_or_path=gpt2 \
|
||||
```
|
||||
|
||||
and from the Salesforce CTRL model:
|
||||
```shell
|
||||
python ./examples/run_generation.py \
|
||||
--model_type=ctrl \
|
||||
--length=20 \
|
||||
--model_name_or_path=ctrl \
|
||||
--temperature=0 \
|
||||
--repetition_penalty=1.2 \
|
||||
```
|
||||
|
||||
## Quick tour of model sharing
|
||||
|
||||
New in `v2.2.2`: you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
|
||||
|
||||
**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
|
||||
|
||||
```shell
|
||||
transformers-cli login
|
||||
# log in using the same credentials as on huggingface.co
|
||||
```
|
||||
Upload your model:
|
||||
```shell
|
||||
transformers-cli upload ./path/to/pretrained_model/
|
||||
|
||||
# ^^ Upload folder containing weights/tokenizer/config
|
||||
# saved via `.save_pretrained()`
|
||||
|
||||
transformers-cli upload ./config.json [--filename foobar.json]
|
||||
|
||||
# ^^ Upload a single file
|
||||
# (you can optionally override its filename)
|
||||
```
|
||||
|
||||
Your model will then be accessible through its identifier:
|
||||
```python
|
||||
"username/model_name"
|
||||
```
|
||||
|
||||
Anyone can load it from code:
|
||||
```python
|
||||
tokenizer = AutoTokenizer.from_pretrained("username/model_name")
|
||||
model = AutoModel.from_pretrained("username/model_name")
|
||||
```
|
||||
|
||||
Finally, list all your files on S3:
|
||||
```shell
|
||||
transformers-cli ls
|
||||
# List all your S3 objects.
|
||||
```
|
||||
|
||||
## Migrating from pytorch-transformers to transformers
|
||||
|
||||
Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.
|
||||
@@ -506,7 +424,7 @@ Here is a quick summary of what you should take care of when migrating from `pyt
|
||||
|
||||
### Models always output `tuples`
|
||||
|
||||
The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that every model's forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
|
||||
The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
|
||||
|
||||
The exact content of the tuples for each model is detailed in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
|
||||
|
||||
@@ -542,9 +460,9 @@ By enabling the configuration option `output_hidden_states`, it was possible to
|
||||
|
||||
Breaking change in the `from_pretrained()` method:
|
||||
|
||||
1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them, don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
|
||||
1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
|
||||
|
||||
2. The additional `*input` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute instead, which can break derived model classes built based on the previous `BertForSequenceClassification` examples. We are working on a way to mitigate this breaking change in [#866](https://github.com/huggingface/transformers/pull/866) by forwarding the the model's `__init__()` method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.
|
||||
2. The additional `*input` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute instead which can break derived model classes build based on the previous `BertForSequenceClassification` examples. We are working on a way to mitigate this breaking change in [#866](https://github.com/huggingface/transformers/pull/866) by forwarding the the model `__init__()` method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.
|
||||
|
||||
Also, while not a breaking change, the serialization methods have been standardized and you probably should switch to the new method `save_pretrained(save_directory)` if you were using any other serialization method before.
|
||||
|
||||
@@ -589,12 +507,12 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
|
||||
# Parameters:
|
||||
lr = 1e-3
|
||||
max_grad_norm = 1.0
|
||||
num_training_steps = 1000
|
||||
num_total_steps = 1000
|
||||
num_warmup_steps = 100
|
||||
warmup_proportion = float(num_warmup_steps) / float(num_training_steps) # 0.1
|
||||
warmup_proportion = float(num_warmup_steps) / float(num_total_steps) # 0.1
|
||||
|
||||
### Previously BertAdam optimizer was instantiated like this:
|
||||
optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_training_steps)
|
||||
optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
|
||||
### and used like this:
|
||||
for batch in train_data:
|
||||
loss = model(batch)
|
||||
@@ -603,10 +521,9 @@ for batch in train_data:
|
||||
|
||||
### In Transformers, optimizer and schedules are splitted and instantiated like this:
|
||||
optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False) # To reproduce BertAdam specific behavior set correct_bias=False
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps) # PyTorch scheduler
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps) # PyTorch scheduler
|
||||
### and used like this:
|
||||
for batch in train_data:
|
||||
model.train()
|
||||
loss = model(batch)
|
||||
loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm) # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
|
||||
@@ -617,13 +534,4 @@ for batch in train_data:
|
||||
|
||||
## Citation
|
||||
|
||||
We now have a paper you can cite for the 🤗 Transformers library:
|
||||
```
|
||||
@article{Wolf2019HuggingFacesTS,
|
||||
title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
|
||||
author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Jamie Brew},
|
||||
journal={ArXiv},
|
||||
year={2019},
|
||||
volume={abs/1910.03771}
|
||||
}
|
||||
```
|
||||
At the moment, there is no paper associated with Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.
|
||||
|
||||
@@ -1,22 +0,0 @@
|
||||
cd docs
|
||||
|
||||
function deploy_doc(){
|
||||
echo "Creating doc at commit $1 and pushing to folder $2"
|
||||
git checkout $1
|
||||
if [ ! -z "$2" ]
|
||||
then
|
||||
echo "Pushing version" $2
|
||||
make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
|
||||
else
|
||||
echo "Pushing master"
|
||||
make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
|
||||
fi
|
||||
}
|
||||
|
||||
deploy_doc "master"
|
||||
deploy_doc "b33a385" v1.0.0
|
||||
deploy_doc "fe02e45" v1.1.0
|
||||
deploy_doc "89fd345" v1.2.0
|
||||
deploy_doc "fc9faa8" v2.0.0
|
||||
deploy_doc "3ddce1d" v2.1.1
|
||||
deploy_doc "f2f3294" v2.2.0
|
||||
+1
-1
@@ -50,7 +50,7 @@ make html
|
||||
---
|
||||
**NOTE**
|
||||
|
||||
If you are adding/removing elements from the toc-tree or from any structural item, it is recommended to clean the build
|
||||
If you are adding/removing elements from the toc-tree or from any strutural item, it is recommended to clean the build
|
||||
directory before rebuilding. Run the following command to clean and build:
|
||||
|
||||
```bash
|
||||
|
||||
@@ -1,3 +1,5 @@
|
||||
huggingface.css
|
||||
|
||||
/* The literal code blocks */
|
||||
.rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
|
||||
color: #6670FF;
|
||||
@@ -42,11 +44,11 @@
|
||||
/* The text items on the toc tree */
|
||||
.wy-menu-vertical a {
|
||||
color: #FFFFDD;
|
||||
font-family: Calibre-Light, sans-serif;
|
||||
font-family: Calibre-Light;
|
||||
}
|
||||
.wy-menu-vertical header, .wy-menu-vertical p.caption{
|
||||
color: white;
|
||||
font-family: Calibre-Light, sans-serif;
|
||||
font-family: Calibre-Light;
|
||||
}
|
||||
|
||||
/* The color inside the selected toc tree block */
|
||||
@@ -83,7 +85,7 @@ a {
|
||||
border-right: solid 2px #FB8D68;
|
||||
border-left: solid 2px #FB8D68;
|
||||
color: #FB8D68;
|
||||
font-family: Calibre-Light, sans-serif;
|
||||
font-family: Calibre-Light;
|
||||
border-top: none;
|
||||
font-style: normal !important;
|
||||
}
|
||||
@@ -134,14 +136,14 @@ a {
|
||||
|
||||
/* class and method names in doc */
|
||||
.rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) code.descclassname{
|
||||
font-family: Calibre, sans-serif;
|
||||
font-family: Calibre;
|
||||
font-size: 20px !important;
|
||||
}
|
||||
|
||||
/* class name in doc*/
|
||||
.rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname{
|
||||
margin-right: 10px;
|
||||
font-family: Calibre-Medium, sans-serif;
|
||||
font-family: Calibre-Medium;
|
||||
}
|
||||
|
||||
/* Method and class parameters */
|
||||
@@ -158,17 +160,17 @@ a {
|
||||
|
||||
/* FONTS */
|
||||
body{
|
||||
font-family: Calibre, sans-serif;
|
||||
font-family: Calibre;
|
||||
font-size: 16px;
|
||||
}
|
||||
|
||||
h1 {
|
||||
font-family: Calibre-Thin, sans-serif;
|
||||
font-family: Calibre-Thin;
|
||||
font-size: 70px;
|
||||
}
|
||||
|
||||
h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
|
||||
font-family: Calibre-Medium, sans-serif;
|
||||
font-family: Calibre-Medium;
|
||||
}
|
||||
|
||||
@font-face {
|
||||
@@ -194,3 +196,4 @@ h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
|
||||
src: url(./Calibre-Thin.otf);
|
||||
font-weight:400;
|
||||
}
|
||||
|
||||
|
||||
@@ -1,5 +1,5 @@
|
||||
function addIcon() {
|
||||
const huggingFaceLogo = "https://huggingface.co/landing/assets/transformers-docs/huggingface_logo.svg";
|
||||
const huggingFaceLogo = "https://huggingface.co/assets/transformers-docs/huggingface_logo.svg";
|
||||
const image = document.createElement("img");
|
||||
image.setAttribute("src", huggingFaceLogo);
|
||||
|
||||
@@ -24,10 +24,10 @@ function addCustomFooter() {
|
||||
social.classList.add("footer__Social");
|
||||
|
||||
const imageDetails = [
|
||||
{ link: "https://huggingface.co", imageLink: "https://huggingface.co/landing/assets/transformers-docs/website.svg" },
|
||||
{ link: "https://twitter.com/huggingface", imageLink: "https://huggingface.co/landing/assets/transformers-docs/twitter.svg" },
|
||||
{ link: "https://github.com/huggingface", imageLink: "https://huggingface.co/landing/assets/transformers-docs/github.svg" },
|
||||
{ link: "https://www.linkedin.com/company/huggingface/", imageLink: "https://huggingface.co/landing/assets/transformers-docs/linkedin.svg" }
|
||||
{ link: "https://huggingface.co", imageLink: "https://huggingface.co/assets/transformers-docs/website.svg" },
|
||||
{ link: "https://twitter.com/huggingface", imageLink: "https://huggingface.co/assets/transformers-docs/twitter.svg" },
|
||||
{ link: "https://github.com/huggingface", imageLink: "https://huggingface.co/assets/transformers-docs/github.svg" },
|
||||
{ link: "https://www.linkedin.com/company/huggingface/", imageLink: "https://huggingface.co/assets/transformers-docs/linkedin.svg" }
|
||||
];
|
||||
|
||||
imageDetails.forEach(imageLinks => {
|
||||
|
||||
@@ -1,54 +0,0 @@
|
||||
# Benchmarks
|
||||
|
||||
This section is dedicated to the Benchmarks done by the library, both by maintainers, contributors and users. These
|
||||
benchmark will help keep track of the preformance improvements that are brought to our models across versions.
|
||||
|
||||
## Benchmarking all models for inference
|
||||
|
||||
As of version 2.1 we have benchmarked all models for inference, across many different settings: using PyTorch, with
|
||||
and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
|
||||
TensorFlow XLA) and GPUs.
|
||||
|
||||
The approach is detailed in the [following blogpost](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
|
||||
|
||||
The results are available [here](https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing).
|
||||
|
||||
## TF2 with mixed precision, XLA, Distribution (@tlkh)
|
||||
|
||||
This work was done by [Timothy Liu](https://github.com/tlkh).
|
||||
|
||||
There are very positive results to be gained from the various TensorFlow 2.0 features:
|
||||
|
||||
- Automatic Mixed Precision (AMP)
|
||||
- XLA compiler
|
||||
- Distribution strategies (multi-GPU)
|
||||
|
||||
The benefits are listed here (tested on CoLA, MRPC, SST-2):
|
||||
|
||||
- AMP: Between 1.4x to 1.6x decrease in overall time without change in batch size
|
||||
- AMP+XLA: Up to 2.5x decrease in overall time on SST-2 (larger dataset)
|
||||
- Distribution: Between 1.4x to 3.4x decrease in overall time on 4xV100
|
||||
- Combined: Up to 5.7x decrease in overall training time, or 9.1x training throughput
|
||||
|
||||
The model quality (measured by the validation accuracy) fluctuates slightly. Taking an average of 4 training runs
|
||||
on a single GPU gives the following results:
|
||||
|
||||
- CoLA: AMP results in slighter lower acc (0.820 vs 0.824)
|
||||
- MRPC: AMP results in lower acc (0.823 vs 0.835)
|
||||
- SST-2: AMP results in slighter lower acc (0.918 vs 0.922)
|
||||
|
||||
However, in a distributed setting with 4xV100 (4x batch size), AMP can yield in better results:
|
||||
|
||||
CoLA: AMP results in higher acc (0.828 vs 0.812)
|
||||
MRPC: AMP results in lower acc (0.817 vs 0.827)
|
||||
SST-2: AMP results in slightly lower acc (0.926 vs 0.929)
|
||||
|
||||
The benchmark script is available [here](https://github.com/NVAITC/benchmarking/blob/master/tf2/bert_dist.py).
|
||||
|
||||
Note: on some tasks (e.g. MRPC), the dataset is too small. The overhead due to the model compilation with XLA as well
|
||||
as the distribution strategy setup does not speed things up. The XLA compile time is also the reason why although throughput
|
||||
can increase a lot (e.g. 2.7x for single GPU), overall (end-to-end) training speed-up is not as fast (as low as 1.4x)
|
||||
|
||||
The benefits as seen on SST-2 (larger dataset) is much clear.
|
||||
|
||||
All results can be seen on this [Google Sheet](https://docs.google.com/spreadsheets/d/1538MN224EzjbRL239sqSiUy6YY-rAjHyXhTzz_Zptls/edit#gid=960868445).
|
||||
+1
-1
@@ -26,7 +26,7 @@ author = u'huggingface'
|
||||
# The short X.Y version
|
||||
version = u''
|
||||
# The full version, including alpha/beta/rc tags
|
||||
release = u'2.2.2'
|
||||
release = u'2.0.0'
|
||||
|
||||
|
||||
# -- General configuration ---------------------------------------------------
|
||||
|
||||
+2
-10
@@ -46,10 +46,8 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
|
||||
5. `XLNet <https://github.com/zihangdai/xlnet>`_ (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`_ by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
|
||||
6. `XLM <https://github.com/facebookresearch/XLM>`_ (from Facebook) released together with the paper `Cross-lingual Language Model Pretraining <https://arxiv.org/abs/1901.07291>`_ by Guillaume Lample and Alexis Conneau.
|
||||
7. `RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`_ (from Facebook), released together with the paper a `Robustly Optimized BERT Pretraining Approach <https://arxiv.org/abs/1907.11692>`_ by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
|
||||
8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the paper `DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2 <https://github.com/huggingface/transformers/tree/master/examples/distillation>`_.
|
||||
9. `CTRL <https://github.com/pytorch/fairseq/tree/master/examples/ctrl>`_ (from Salesforce), released together with the paper `CTRL: A Conditional Transformer Language Model for Controllable Generation <https://www.github.com/salesforce/ctrl>`_ by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
|
||||
10. `CamemBERT <https://huggingface.co/transformers/model_doc/camembert.html>`_ (from FAIR, Inria, Sorbonne Université) released together with the paper `CamemBERT: a Tasty French Language Model <https://arxiv.org/abs/1911.03894>`_ by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suarez, Yoann Dupont, Laurent Romary, Eric Villemonte de la Clergerie, Djame Seddah, and Benoît Sagot.
|
||||
11. `ALBERT <https://github.com/google-research/ALBERT>`_ (from Google Research), released together with the paper a `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations <https://arxiv.org/abs/1909.11942>`_ by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
|
||||
8. `DistilBERT <https://huggingface.co/transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the blog post `Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT <https://medium.com/huggingface/distilbert-8cf3380435b5>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf.
|
||||
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
@@ -58,7 +56,6 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
|
||||
installation
|
||||
quickstart
|
||||
pretrained_models
|
||||
model_sharing
|
||||
examples
|
||||
notebooks
|
||||
serialization
|
||||
@@ -66,8 +63,6 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
|
||||
migration
|
||||
bertology
|
||||
torchscript
|
||||
multilingual
|
||||
benchmarks
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 2
|
||||
@@ -92,6 +87,3 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
|
||||
model_doc/xlnet
|
||||
model_doc/roberta
|
||||
model_doc/distilbert
|
||||
model_doc/ctrl
|
||||
model_doc/camembert
|
||||
model_doc/albert
|
||||
|
||||
@@ -1,67 +0,0 @@
|
||||
# Installation
|
||||
|
||||
Transformers is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.1.0
|
||||
|
||||
## With pip
|
||||
|
||||
PyTorch Transformers can be installed using pip as follows:
|
||||
|
||||
``` bash
|
||||
pip install transformers
|
||||
```
|
||||
|
||||
## From source
|
||||
|
||||
To install from source, clone the repository and install with:
|
||||
|
||||
``` bash
|
||||
git clone https://github.com/huggingface/transformers.git
|
||||
cd transformers
|
||||
pip install [--editable] .
|
||||
```
|
||||
|
||||
## Tests
|
||||
|
||||
An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the [tests folder](https://github.com/huggingface/transformers/tree/master/transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/transformers/tree/master/examples).
|
||||
|
||||
Tests can be run using `unittest` or `pytest` (install pytest if needed with `pip install pytest`).
|
||||
|
||||
Run all the tests from the root of the cloned repository with the commands:
|
||||
|
||||
```bash
|
||||
python -m unittest discover -s transformers/tests -p "*test.py" -t .
|
||||
python -m unittest discover -s examples -p "*test.py" -t examples
|
||||
```
|
||||
|
||||
or
|
||||
|
||||
``` bash
|
||||
python -m pytest -sv ./transformers/tests/
|
||||
python -m pytest -sv ./examples/
|
||||
```
|
||||
|
||||
By default, slow tests are skipped. Set the `RUN_SLOW` environment variable to `yes` to run them.
|
||||
|
||||
## OpenAI GPT original tokenization workflow
|
||||
|
||||
If you want to reproduce the original tokenization process of the `OpenAI GPT` paper, you will need to install `ftfy` (use version 4.4.3 if you are using Python 2) and `SpaCy`:
|
||||
|
||||
``` bash
|
||||
pip install spacy ftfy==4.4.3
|
||||
python -m spacy download en
|
||||
```
|
||||
|
||||
If you don't install `ftfy` and `SpaCy`, the `OpenAI GPT` tokenizer will default to tokenize using BERT's `BasicTokenizer` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
|
||||
|
||||
## Note on model downloads (Continuous Integration or large-scale deployments)
|
||||
|
||||
If you expect to be downloading large volumes of models (more than 1,000) from our hosted bucket (for instance through your CI setup, or a large-scale production deployment), please cache the model files on your end. It will be way faster, and cheaper. Feel free to contact us privately if you need any help.
|
||||
|
||||
## Do you want to run a Transformer model on a mobile device?
|
||||
|
||||
You should check out our [swift-coreml-transformers](https://github.com/huggingface/swift-coreml-transformers) repo.
|
||||
|
||||
It contains a set of tools to convert PyTorch or TensorFlow 2.0 trained Transformer models (currently contains `GPT-2`, `DistilGPT-2`, `BERT`, and `DistilBERT`) to CoreML models that run on iOS devices.
|
||||
|
||||
At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
|
||||
or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
|
||||
@@ -0,0 +1,71 @@
|
||||
Installation
|
||||
================================================
|
||||
|
||||
Transformers is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.1.0
|
||||
|
||||
With pip
|
||||
^^^^^^^^
|
||||
|
||||
PyTorch Transformers can be installed using pip as follows:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
pip install transformers
|
||||
|
||||
From source
|
||||
^^^^^^^^^^^
|
||||
|
||||
To install from source, clone the repository and install with:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
git clone https://github.com/huggingface/transformers.git
|
||||
cd transformers
|
||||
pip install [--editable] .
|
||||
|
||||
|
||||
Tests
|
||||
^^^^^
|
||||
|
||||
An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the `tests folder <https://github.com/huggingface/transformers/tree/master/transformers/tests>`_ and examples tests in the `examples folder <https://github.com/huggingface/transformers/tree/master/examples>`_.
|
||||
|
||||
Tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
|
||||
|
||||
Run all the tests from the root of the cloned repository with the commands:
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
python -m pytest -sv ./transformers/tests/
|
||||
python -m pytest -sv ./examples/
|
||||
|
||||
|
||||
OpenAI GPT original tokenization workflow
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
If you want to reproduce the original tokenization process of the ``OpenAI GPT`` paper, you will need to install ``ftfy`` (use version 4.4.3 if you are using Python 2) and ``SpaCy`` :
|
||||
|
||||
.. code-block:: bash
|
||||
|
||||
pip install spacy ftfy==4.4.3
|
||||
python -m spacy download en
|
||||
|
||||
If you don't install ``ftfy`` and ``SpaCy``\ , the ``OpenAI GPT`` tokenizer will default to tokenize using BERT's ``BasicTokenizer`` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
|
||||
|
||||
|
||||
Note on model downloads (Continuous Integration or large-scale deployments)
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
If you expect to be downloading large volumes of models (more than 1,000) from our hosted bucket (for instance through your CI setup, or a large-scale production deployment), please cache the model files on your end. It will be way faster, and cheaper. Feel free to contact us privately if you need any help.
|
||||
|
||||
|
||||
Do you want to run a Transformer model on a mobile device?
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
You should check out our `swift-coreml-transformers <https://github.com/huggingface/swift-coreml-transformers>`_ repo.
|
||||
|
||||
It contains an example of a conversion script from a Pytorch trained Transformer model (here, ``GPT-2``) to a CoreML model that runs on iOS devices.
|
||||
|
||||
It also contains an implementation of BERT for Question answering.
|
||||
|
||||
At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
|
||||
or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
|
||||
@@ -5,7 +5,6 @@ The ``.optimization`` module provides:
|
||||
|
||||
- an optimizer with weight decay fixed that can be used to fine-tuned models, and
|
||||
- several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
|
||||
- a gradient accumulation class to accumulate the gradients of multiple batches
|
||||
|
||||
``AdamW``
|
||||
~~~~~~~~~~~~~~~~
|
||||
@@ -13,32 +12,25 @@ The ``.optimization`` module provides:
|
||||
.. autoclass:: transformers.AdamW
|
||||
:members:
|
||||
|
||||
``AdamWeightDecay``
|
||||
~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AdamWeightDecay
|
||||
:members:
|
||||
|
||||
.. autofunction:: transformers.create_optimizer
|
||||
:members:
|
||||
|
||||
Schedules
|
||||
----------------------------------------------------
|
||||
|
||||
Learning Rate Schedules
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
.. autofunction:: transformers.get_constant_schedule
|
||||
.. autoclass:: transformers.ConstantLRSchedule
|
||||
:members:
|
||||
|
||||
|
||||
.. autofunction:: transformers.get_constant_schedule_with_warmup
|
||||
.. autoclass:: transformers.WarmupConstantSchedule
|
||||
:members:
|
||||
|
||||
.. image:: /imgs/warmup_constant_schedule.png
|
||||
:target: /imgs/warmup_constant_schedule.png
|
||||
:alt:
|
||||
|
||||
|
||||
.. autofunction:: transformers.get_cosine_schedule_with_warmup
|
||||
.. autoclass:: transformers.WarmupCosineSchedule
|
||||
:members:
|
||||
|
||||
.. image:: /imgs/warmup_cosine_schedule.png
|
||||
@@ -46,7 +38,8 @@ Learning Rate Schedules
|
||||
:alt:
|
||||
|
||||
|
||||
.. autofunction:: transformers.get_cosine_with_hard_restarts_schedule_with_warmup
|
||||
.. autoclass:: transformers.WarmupCosineWithHardRestartsSchedule
|
||||
:members:
|
||||
|
||||
.. image:: /imgs/warmup_cosine_hard_restarts_schedule.png
|
||||
:target: /imgs/warmup_cosine_hard_restarts_schedule.png
|
||||
@@ -54,22 +47,9 @@ Learning Rate Schedules
|
||||
|
||||
|
||||
|
||||
.. autofunction:: transformers.get_linear_schedule_with_warmup
|
||||
.. autoclass:: transformers.WarmupLinearSchedule
|
||||
:members:
|
||||
|
||||
.. image:: /imgs/warmup_linear_schedule.png
|
||||
:target: /imgs/warmup_linear_schedule.png
|
||||
:alt:
|
||||
|
||||
``Warmup``
|
||||
~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.Warmup
|
||||
:members:
|
||||
|
||||
Gradient Strategies
|
||||
----------------------------------------------------
|
||||
|
||||
``GradientAccumulator``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.GradientAccumulator
|
||||
|
||||
@@ -54,100 +54,5 @@ Additionally, the following method can be used to load values from a data file
|
||||
Example usage
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
An example using these processors is given in the `run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
|
||||
|
||||
|
||||
XNLI
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
`The Cross-Lingual NLI Corpus (XNLI) <https://www.nyu.edu/projects/bowman/xnli/>`__ is a benchmark that evaluates
|
||||
the quality of cross-lingual text representations.
|
||||
XNLI is crowd-sourced dataset based on `MultiNLI <http://www.nyu.edu/projects/bowman/multinli/>`: pairs of text are labeled with textual entailment
|
||||
annotations for 15 different languages (including both high-ressource language such as English and low-ressource languages such as Swahili).
|
||||
|
||||
It was released together with the paper
|
||||
`XNLI: Evaluating Cross-lingual Sentence Representations <https://arxiv.org/abs/1809.05053>`__
|
||||
|
||||
This library hosts the processor to load the XNLI data:
|
||||
- :class:`~transformers.data.processors.utils.XnliProcessor`
|
||||
|
||||
Please note that since the gold labels are available on the test set, evaluation is performed on the test set.
|
||||
|
||||
An example using these processors is given in the
|
||||
`run_xnli.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_xnli.py>`__ script.
|
||||
|
||||
|
||||
SQuAD
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
`The Stanford Question Answering Dataset (SQuAD) <https://rajpurkar.github.io/SQuAD-explorer//>`__ is a benchmark that evaluates
|
||||
the performance of models on question answering. Two versions are available, v1.1 and v2.0. The first version (v1.1) was released together with the paper
|
||||
`SQuAD: 100,000+ Questions for Machine Comprehension of Text <https://arxiv.org/abs/1606.05250>`__. The second version (v2.0) was released alongside
|
||||
the paper `Know What You Don't Know: Unanswerable Questions for SQuAD <https://arxiv.org/abs/1806.03822>`__.
|
||||
|
||||
This library hosts a processor for each of the two versions:
|
||||
|
||||
Processors
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
Those processors are:
|
||||
- :class:`~transformers.data.processors.utils.SquadV1Processor`
|
||||
- :class:`~transformers.data.processors.utils.SquadV2Processor`
|
||||
|
||||
They both inherit from the abstract class :class:`~transformers.data.processors.utils.SquadProcessor`
|
||||
|
||||
.. autoclass:: transformers.data.processors.squad.SquadProcessor
|
||||
:members:
|
||||
|
||||
Additionally, the following method can be used to convert SQuAD examples into :class:`~transformers.data.processors.utils.SquadFeatures`
|
||||
that can be used as model inputs.
|
||||
|
||||
.. automethod:: transformers.data.processors.squad.squad_convert_examples_to_features
|
||||
|
||||
These processors as well as the aforementionned method can be used with files containing the data as well as with the `tensorflow_datasets` package.
|
||||
Examples are given below.
|
||||
|
||||
|
||||
Example usage
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
Here is an example using the processors as well as the conversion method using data files:
|
||||
|
||||
Example::
|
||||
|
||||
# Loading a V2 processor
|
||||
processor = SquadV2Processor()
|
||||
examples = processor.get_dev_examples(squad_v2_data_dir)
|
||||
|
||||
# Loading a V1 processor
|
||||
processor = SquadV1Processor()
|
||||
examples = processor.get_dev_examples(squad_v1_data_dir)
|
||||
|
||||
features = squad_convert_examples_to_features(
|
||||
examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=max_query_length,
|
||||
is_training=not evaluate,
|
||||
)
|
||||
|
||||
Using `tensorflow_datasets` is as easy as using a data file:
|
||||
|
||||
Example::
|
||||
|
||||
# tensorflow_datasets only handle Squad V1.
|
||||
tfds_examples = tfds.load("squad")
|
||||
examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
|
||||
|
||||
features = squad_convert_examples_to_features(
|
||||
examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=max_query_length,
|
||||
is_training=not evaluate,
|
||||
)
|
||||
|
||||
|
||||
Another example using these processors is given in the
|
||||
`run_squad.py <https://github.com/huggingface/transformers/blob/master/examples/run_squad.py>`__ script.
|
||||
`run_glue.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_glue.py>`__ script.
|
||||
@@ -84,12 +84,12 @@ Here is a conversion examples from `BertAdam` with a linear warmup and decay sch
|
||||
# Parameters:
|
||||
lr = 1e-3
|
||||
max_grad_norm = 1.0
|
||||
num_training_steps = 1000
|
||||
num_total_steps = 1000
|
||||
num_warmup_steps = 100
|
||||
warmup_proportion = float(num_warmup_steps) / float(num_training_steps) # 0.1
|
||||
warmup_proportion = float(num_warmup_steps) / float(num_total_steps) # 0.1
|
||||
|
||||
### Previously BertAdam optimizer was instantiated like this:
|
||||
optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, num_training_steps=num_training_steps)
|
||||
optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
|
||||
### and used like this:
|
||||
for batch in train_data:
|
||||
loss = model(batch)
|
||||
@@ -98,12 +98,12 @@ for batch in train_data:
|
||||
|
||||
### In Transformers, optimizer and schedules are splitted and instantiated like this:
|
||||
optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False) # To reproduce BertAdam specific behavior set correct_bias=False
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps) # PyTorch scheduler
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps) # PyTorch scheduler
|
||||
### and used like this:
|
||||
for batch in train_data:
|
||||
loss = model(batch)
|
||||
loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm) # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
|
||||
optimizer.step()
|
||||
scheduler.step()
|
||||
optimizer.step()
|
||||
```
|
||||
|
||||
@@ -1,64 +0,0 @@
|
||||
ALBERT
|
||||
----------------------------------------------------
|
||||
|
||||
``AlbrtConfig``
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertConfig
|
||||
:members:
|
||||
|
||||
|
||||
``AlbertTokenizer``
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertTokenizer
|
||||
:members:
|
||||
|
||||
|
||||
``AlbertModel``
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertModel
|
||||
:members:
|
||||
|
||||
|
||||
``AlbertForMaskedLM``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertForMaskedLM
|
||||
:members:
|
||||
|
||||
|
||||
``AlbertForSequenceClassification``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertForSequenceClassification
|
||||
:members:
|
||||
|
||||
|
||||
``AlbertForQuestionAnswering``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.AlbertForQuestionAnswering
|
||||
:members:
|
||||
|
||||
|
||||
``TFAlbertModel``
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.TFAlbertModel
|
||||
:members:
|
||||
|
||||
|
||||
``TFAlbertForMaskedLM``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.TFAlbertForMaskedLM
|
||||
:members:
|
||||
|
||||
|
||||
``TFAlbertForSequenceClassification``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.TFAlbertForSequenceClassification
|
||||
:members:
|
||||
@@ -1,50 +0,0 @@
|
||||
CamemBERT
|
||||
----------------------------------------------------
|
||||
|
||||
``CamembertConfig``
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertConfig
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertTokenizer``
|
||||
~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertTokenizer
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertModel``
|
||||
~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertModel
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertForMaskedLM``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertForMaskedLM
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertForSequenceClassification``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertForSequenceClassification
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertForMultipleChoice``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertForMultipleChoice
|
||||
:members:
|
||||
|
||||
|
||||
``CamembertForTokenClassification``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CamembertForTokenClassification
|
||||
:members:
|
||||
@@ -1,49 +0,0 @@
|
||||
CTRL
|
||||
----------------------------------------------------
|
||||
|
||||
Note: if you fine-tune a CTRL model using the Salesforce code (https://github.com/salesforce/ctrl),
|
||||
you'll be able to convert from TF to our HuggingFace/Transformers format using the
|
||||
``convert_tf_to_huggingface_pytorch.py`` script (see `issue #1654 <https://github.com/huggingface/transformers/issues/1654>`_).
|
||||
|
||||
|
||||
``CTRLConfig``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CTRLConfig
|
||||
:members:
|
||||
|
||||
|
||||
``CTRLTokenizer``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CTRLTokenizer
|
||||
:members:
|
||||
|
||||
|
||||
``CTRLModel``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CTRLModel
|
||||
:members:
|
||||
|
||||
|
||||
``CTRLLMHeadModel``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.CTRLLMHeadModel
|
||||
:members:
|
||||
|
||||
|
||||
``TFCTRLModel``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.TFCTRLModel
|
||||
:members:
|
||||
|
||||
|
||||
``TFCTRLLMHeadModel``
|
||||
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||||
|
||||
.. autoclass:: transformers.TFCTRLLMHeadModel
|
||||
:members:
|
||||
|
||||
@@ -1,40 +0,0 @@
|
||||
# Model upload and sharing
|
||||
|
||||
Starting with `v2.2.2`, you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
|
||||
|
||||
**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
|
||||
|
||||
```shell
|
||||
transformers-cli login
|
||||
# log in using the same credentials as on huggingface.co
|
||||
```
|
||||
Upload your model:
|
||||
```shell
|
||||
transformers-cli upload ./path/to/pretrained_model/
|
||||
|
||||
# ^^ Upload folder containing weights/tokenizer/config
|
||||
# saved via `.save_pretrained()`
|
||||
|
||||
transformers-cli upload ./config.json [--filename foobar.json]
|
||||
|
||||
# ^^ Upload a single file
|
||||
# (you can optionally override its filename)
|
||||
```
|
||||
|
||||
Your model will then be accessible through its identifier:
|
||||
```python
|
||||
"username/model_name"
|
||||
```
|
||||
|
||||
Anyone can load it from code:
|
||||
```python
|
||||
tokenizer = AutoTokenizer.from_pretrained("username/model_name")
|
||||
model = AutoModel.from_pretrained("username/model_name")
|
||||
```
|
||||
|
||||
Finally, list all your files on S3:
|
||||
```shell
|
||||
transformers-cli ls
|
||||
# List all your S3 objects.
|
||||
```
|
||||
|
||||
@@ -1,103 +0,0 @@
|
||||
Multi-lingual models
|
||||
================================================
|
||||
|
||||
Most of the models available in this library are mono-lingual models (English, Chinese and German). A few
|
||||
multi-lingual models are available and have a different mechanisms than mono-lingual models.
|
||||
This page details the usage of these models.
|
||||
|
||||
The two models that currently support multiple languages are BERT and XLM.
|
||||
|
||||
XLM
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
XLM has a total of 10 different checkpoints, only one of which is mono-lingual. The 9 remaining model checkpoints can
|
||||
be split in two categories: the checkpoints that make use of language embeddings, and those that don't
|
||||
|
||||
XLM & Language Embeddings
|
||||
------------------------------------------------
|
||||
|
||||
This section concerns the following checkpoints:
|
||||
|
||||
- ``xlm-mlm-ende-1024`` (Masked language modeling, English-German)
|
||||
- ``xlm-mlm-enfr-1024`` (Masked language modeling, English-French)
|
||||
- ``xlm-mlm-enro-1024`` (Masked language modeling, English-Romanian)
|
||||
- ``xlm-mlm-xnli15-1024`` (Masked language modeling, XNLI languages)
|
||||
- ``xlm-mlm-tlm-xnli15-1024`` (Masked language modeling + Translation, XNLI languages)
|
||||
- ``xlm-clm-enfr-1024`` (Causal language modeling, English-French)
|
||||
- ``xlm-clm-ende-1024`` (Causal language modeling, English-German)
|
||||
|
||||
These checkpoints require language embeddings that will specify the language used at inference time. These language
|
||||
embeddings are represented as a tensor that is of the same shape as the input ids passed to the model. The values in
|
||||
these tensors depend on the language used and are identifiable using the ``lang2id`` and ``id2lang`` attributes
|
||||
from the tokenizer.
|
||||
|
||||
Here is an example using the ``xlm-clm-enfr-1024`` checkpoint (Causal language modeling, English-French):
|
||||
|
||||
|
||||
.. code-block::
|
||||
|
||||
import torch
|
||||
from transformers import XLMTokenizer, XLMWithLMHeadModel
|
||||
|
||||
tokenizer = XLMTokenizer.from_pretrained("xlm-clm-1024-enfr")
|
||||
|
||||
|
||||
The different languages this model/tokenizer handles, as well as the ids of these languages are visible using the
|
||||
``lang2id`` attribute:
|
||||
|
||||
.. code-block::
|
||||
|
||||
print(tokenizer.lang2id) # {'en': 0, 'fr': 1}
|
||||
|
||||
|
||||
These ids should be used when passing a language parameter during a model pass. Let's define our inputs:
|
||||
|
||||
.. code-block::
|
||||
|
||||
input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1
|
||||
|
||||
|
||||
We should now define the language embedding by using the previously defined language id. We want to create a tensor
|
||||
filled with the appropriate language ids, of the same size as input_ids. For english, the id is 0:
|
||||
|
||||
.. code-block::
|
||||
|
||||
language_id = tokenizer.lang2id['en'] # 0
|
||||
langs = torch.tensor([language_id] * input_ids.shape[1]) # torch.tensor([0, 0, 0, ..., 0])
|
||||
|
||||
# We reshape it to be of size (batch_size, sequence_length)
|
||||
langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)
|
||||
|
||||
|
||||
You can then feed it all as input to your model:
|
||||
|
||||
.. code-block::
|
||||
|
||||
outputs = model(input_ids, langs=langs)
|
||||
|
||||
|
||||
The example `run_generation.py <https://github.com/huggingface/transformers/blob/master/examples/run_generation.py>`__
|
||||
can generate text using the CLM checkpoints from XLM, using the language embeddings.
|
||||
|
||||
XLM without Language Embeddings
|
||||
------------------------------------------------
|
||||
|
||||
This section concerns the following checkpoints:
|
||||
|
||||
- ``xlm-mlm-17-1280`` (Masked language modeling, 17 languages)
|
||||
- ``xlm-mlm-100-1280`` (Masked language modeling, 100 languages)
|
||||
|
||||
These checkpoints do not require language embeddings at inference time. These models are used to have generic
|
||||
sentence representations, differently from previously-mentioned XLM checkpoints.
|
||||
|
||||
|
||||
BERT
|
||||
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
|
||||
|
||||
BERT has two checkpoints that can be used for multi-lingual tasks:
|
||||
|
||||
- ``bert-base-multilingual-uncased`` (Masked language modeling + Next sentence prediction, 102 languages)
|
||||
- ``bert-base-multilingual-cased`` (Masked language modeling + Next sentence prediction, 104 languages)
|
||||
|
||||
These checkpoints do not require language embeddings at inference time. They should identify the language
|
||||
used in the context and infer accordingly.
|
||||
@@ -53,32 +53,6 @@ Here is the full list of the currently provided pretrained models together with
|
||||
| | ``bert-base-cased-finetuned-mrpc`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | The ``bert-base-cased`` model fine-tuned on MRPC |
|
||||
| | | (see `details of fine-tuning in the example section <https://huggingface.co/transformers/examples.html>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-german-dbmdz-cased`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on cased German text by DBMDZ |
|
||||
| | | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__). |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-german-dbmdz-uncased`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on uncased German text by DBMDZ |
|
||||
| | | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__). |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-japanese`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece. |
|
||||
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
|
||||
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-japanese-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece. |
|
||||
| | | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization. |
|
||||
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-japanese-char`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on Japanese text. Text is tokenized into characters. |
|
||||
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``bert-base-japanese-char-whole-word-masking`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters. |
|
||||
| | | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__). |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| GPT | ``openai-gpt`` | | 12-layer, 768-hidden, 12-heads, 110M parameters. |
|
||||
| | | | OpenAI GPT English model |
|
||||
@@ -91,9 +65,6 @@ Here is the full list of the currently provided pretrained models together with
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``gpt2-large`` | | 36-layer, 1280-hidden, 20-heads, 774M parameters. |
|
||||
| | | | OpenAI's Large-sized GPT-2 English model |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``gpt2-xl`` | | 48-layer, 1600-hidden, 25-heads, 1558M parameters. |
|
||||
| | | | OpenAI's XL-sized GPT-2 English model |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| Transformer-XL | ``transfo-xl-wt103`` | | 18-layer, 1024-hidden, 16-heads, 257M parameters. |
|
||||
| | | | English model trained on wikitext-103 |
|
||||
@@ -127,12 +98,6 @@ Here is the full list of the currently provided pretrained models together with
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``xlm-clm-ende-1024`` | | 6-layer, 1024-hidden, 8-heads |
|
||||
| | | | XLM English-German model trained with CLM (Causal Language Modeling) on the concatenation of English and German wikipedia |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``xlm-mlm-17-1280`` | | 16-layer, 1280-hidden, 16-heads |
|
||||
| | | | XLM model trained with MLM (Masked Language Modeling) on 17 languages. |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``xlm-mlm-100-1280`` | | 16-layer, 1280-hidden, 16-heads |
|
||||
| | | | XLM model trained with MLM (Masked Language Modeling) on 100 languages. |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| RoBERTa | ``roberta-base`` | | 12-layer, 768-hidden, 12-heads, 125M parameters |
|
||||
| | | | RoBERTa using the BERT-base architecture |
|
||||
@@ -145,93 +110,14 @@ Here is the full list of the currently provided pretrained models together with
|
||||
| | ``roberta-large-mnli`` | | 24-layer, 1024-hidden, 16-heads, 355M parameters |
|
||||
| | | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__. |
|
||||
| | | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``roberta-base-openai-detector`` | | 12-layer, 768-hidden, 12-heads, 125M parameters |
|
||||
| | | | ``roberta-base`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model. |
|
||||
| | | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``roberta-large-openai-detector`` | | 24-layer, 1024-hidden, 16-heads, 355M parameters |
|
||||
| | | | ``roberta-large`` fine-tuned by OpenAI on the outputs of the 1.5B-parameter GPT-2 model. |
|
||||
| | | (see `details <https://github.com/openai/gpt-2-output-dataset/tree/master/detector>`__) |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| DistilBERT | ``distilbert-base-uncased`` | | 6-layer, 768-hidden, 12-heads, 66M parameters |
|
||||
| | | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
| | | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``distilbert-base-uncased-distilled-squad`` | | 6-layer, 768-hidden, 12-heads, 66M parameters |
|
||||
| | | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint, with an additional linear layer. |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``distilgpt2`` | | 6-layer, 768-hidden, 12-heads, 82M parameters |
|
||||
| | | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint. |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``distilroberta-base`` | | 6-layer, 768-hidden, 12-heads, 82M parameters |
|
||||
| | | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint. |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``distilbert-base-german-cased`` | | 6-layer, 768-hidden, 12-heads, 66M parameters |
|
||||
| | | | The German DistilBERT model distilled from the German DBMDZ BERT model `bert-base-german-dbmdz-cased` checkpoint. |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``distilbert-base-multilingual-cased`` | | 6-layer, 768-hidden, 12-heads, 134M parameters |
|
||||
| | | | The multilingual DistilBERT model distilled from the Multilingual BERT model `bert-base-multilingual-cased` checkpoint. |
|
||||
| | | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__) |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| CTRL | ``ctrl`` | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters |
|
||||
| | | | Salesforce's Large-sized CTRL English model |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| CamemBERT | ``camembert-base`` | | 12-layer, 768-hidden, 12-heads, 110M parameters |
|
||||
| | | | CamemBERT using the BERT-base architecture |
|
||||
| | | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/camembert>`__) |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| ALBERT | ``albert-base-v1`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters |
|
||||
| | | | ALBERT base model |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-large-v1`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters |
|
||||
| | | | ALBERT large model |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-xlarge-v1`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters |
|
||||
| | | | ALBERT xlarge model |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-xxlarge-v1`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters |
|
||||
| | | | ALBERT xxlarge model |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-base-v2`` | | 12 repeating layers, 128 embedding, 768-hidden, 12-heads, 11M parameters |
|
||||
| | | | ALBERT base model with no dropout, additional training data and longer training |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-large-v2`` | | 24 repeating layers, 128 embedding, 1024-hidden, 16-heads, 17M parameters |
|
||||
| | | | ALBERT large model with no dropout, additional training data and longer training |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-xlarge-v2`` | | 24 repeating layers, 128 embedding, 2048-hidden, 16-heads, 58M parameters |
|
||||
| | | | ALBERT xlarge model with no dropout, additional training data and longer training |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``albert-xxlarge-v2`` | | 12 repeating layer, 128 embedding, 4096-hidden, 64-heads, 223M parameters |
|
||||
| | | | ALBERT xxlarge model with no dropout, additional training data and longer training |
|
||||
| | | (see `details <https://github.com/google-research/ALBERT>`__) |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| T5 | ``t5-small`` | | ~60M parameters with 6-layers, 512-hidden-state, 2048 feed-forward hidden-state, 8-heads, |
|
||||
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``t5-base`` | | ~220M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 12-heads, |
|
||||
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``t5-large`` | | ~770M parameters with 24-layers, 1024-hidden-state, 4096 feed-forward hidden-state, 16-heads, |
|
||||
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``t5-3B`` | | ~2.8B parameters with 24-layers, 1024-hidden-state, 16384 feed-forward hidden-state, 32-heads, |
|
||||
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
|
||||
| +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
| | ``t5-11B`` | | ~11B parameters with 24-layers, 1024-hidden-state, 65536 feed-forward hidden-state, 128-heads, |
|
||||
| | | | Trained on English text: the Colossal Clean Crawled Corpus (C4) |
|
||||
| | | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__) |
|
||||
+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
|
||||
|
||||
|
||||
.. <https://huggingface.co/transformers/examples.html>`__
|
||||
.. <https://huggingface.co/transformers/examples.html>`__
|
||||
@@ -188,35 +188,3 @@ assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
|
||||
```
|
||||
|
||||
Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).
|
||||
|
||||
#### Using the past
|
||||
|
||||
GPT-2 as well as some other models (GPT, XLNet, Transfo-XL, CTRL) make use of a `past` or `mems` attribute which can be used to prevent re-computing the key/value pairs when using sequential decoding. It is useful when generating sequences as a big part of the attention mechanism benefits from previous computations.
|
||||
|
||||
Here is a fully-working example using the `past` with `GPT2LMHeadModel` and argmax decoding (which should only be used as an example, as argmax decoding introduces a lot of repetition):
|
||||
|
||||
```python
|
||||
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
||||
import torch
|
||||
|
||||
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
||||
model = GPT2LMHeadModel.from_pretrained('gpt2')
|
||||
|
||||
generated = tokenizer.encode("The Manhattan bridge")
|
||||
context = torch.tensor([generated])
|
||||
past = None
|
||||
|
||||
for i in range(100):
|
||||
print(i)
|
||||
output, past = model(context, past=past)
|
||||
token = torch.argmax(output[0, :])
|
||||
|
||||
generated += [token.tolist()]
|
||||
context = token.unsqueeze(0)
|
||||
|
||||
sequence = tokenizer.decode(generated)
|
||||
|
||||
print(sequence)
|
||||
```
|
||||
|
||||
The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
|
||||
@@ -33,8 +33,6 @@ where
|
||||
* ``bert-large-uncased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
|
||||
* ``bert-large-cased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
|
||||
* ``bert-large-uncased-whole-word-masking-finetuned-squad``: The ``bert-large-uncased-whole-word-masking`` model finetuned on SQuAD (using the ``run_bert_squad.py`` examples). Results: *exact_match: 86.91579943235573, f1: 93.1532499015869*
|
||||
* ``bert-base-german-dbmdz-cased``: Trained on German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
|
||||
* ``bert-base-german-dbmdz-uncased``: Trained on (uncased) German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://github.com/dbmdz/german-bert>`__
|
||||
* ``openai-gpt``: OpenAI GPT English model, 12-layer, 768-hidden, 12-heads, 110M parameters
|
||||
* ``gpt2``: OpenAI GPT-2 English model, 12-layer, 768-hidden, 12-heads, 117M parameters
|
||||
* ``gpt2-medium``: OpenAI GPT-2 English model, 24-layer, 1024-hidden, 16-heads, 345M parameters
|
||||
@@ -106,7 +104,7 @@ This section explain how you can save and re-load a fine-tuned model (BERT, GPT,
|
||||
There are three types of files you need to save to be able to reload a fine-tuned model:
|
||||
|
||||
|
||||
* the model itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
|
||||
* the model it-self which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
|
||||
* the configuration file of the model which is saved as a JSON file, and
|
||||
* the vocabulary (and the merges for the BPE-based models GPT and GPT-2).
|
||||
|
||||
|
||||
+7
-300
@@ -3,49 +3,13 @@
|
||||
In this section a few examples are put together. All of these examples work for several models, making use of the very
|
||||
similar API between the different models.
|
||||
|
||||
**Important**
|
||||
To run the latest versions of the examples, you have to install from source and install some specific requirements for the examples.
|
||||
Execute the following steps in a new virtual environment:
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/transformers
|
||||
cd transformers
|
||||
pip install [--editable] .
|
||||
pip install -r ./examples/requirements.txt
|
||||
```
|
||||
|
||||
| Section | Description |
|
||||
|----------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||||
| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks.
|
||||
| [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
|
||||
| [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet. |
|
||||
| [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision. |
|
||||
| [SQuAD](#squad) | Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training. |
|
||||
| [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
|
||||
| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
|
||||
| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
|
||||
|
||||
## TensorFlow 2.0 Bert models on GLUE
|
||||
|
||||
Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
|
||||
|
||||
Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
|
||||
|
||||
This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
|
||||
Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
|
||||
These options and the below benchmark are provided by @tlkh.
|
||||
|
||||
Quick benchmarks from the script (no other modifications):
|
||||
|
||||
| GPU | Mode | Time (2nd epoch) | Val Acc (3 runs) |
|
||||
| --------- | -------- | ----------------------- | ----------------------|
|
||||
| Titan V | FP32 | 41s | 0.8438/0.8281/0.8333 |
|
||||
| Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
|
||||
| V100 | FP32 | 35s | 0.8646/0.8359/0.8464 |
|
||||
| V100 | AMP | 22s | 0.8646/0.8385/0.8411 |
|
||||
| 1080 Ti | FP32 | 55s | - |
|
||||
|
||||
Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
|
||||
| [SQuAD](#squad) | Using BERT for question answering, examples with distributed training. |
|
||||
| [Multiple Choice](#multiple choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks.
|
||||
|
||||
## Language model fine-tuning
|
||||
|
||||
@@ -113,7 +77,7 @@ python run_lm_finetuning.py \
|
||||
|
||||
Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
|
||||
|
||||
Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
|
||||
Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
|
||||
A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
|
||||
can try out the different models available in the library.
|
||||
|
||||
@@ -319,17 +283,17 @@ The results are the following:
|
||||
loss = 0.04755385363816904
|
||||
```
|
||||
|
||||
## Multiple Choice
|
||||
##Multiple Choice
|
||||
|
||||
Based on the script [`run_multiple_choice.py`]().
|
||||
|
||||
#### Fine-tuning on SWAG
|
||||
Download [swag](https://github.com/rowanz/swagaf/tree/master/data) data
|
||||
|
||||
```bash
|
||||
```
|
||||
#training on 4 tesla V100(16GB) GPUS
|
||||
export SWAG_DIR=/path/to/swag_data_dir
|
||||
python ./examples/run_multiple_choice.py \
|
||||
python ./examples/single_model_scripts/run_multiple_choice.py \
|
||||
--model_type roberta \
|
||||
--task_name swag \
|
||||
--model_name_or_path roberta-base \
|
||||
@@ -423,263 +387,6 @@ f1 = 93.15
|
||||
exact_match = 86.91
|
||||
```
|
||||
|
||||
This fine-tuned model is available as a checkpoint under the reference
|
||||
This fine-tuneds model is available as a checkpoint under the reference
|
||||
`bert-large-uncased-whole-word-masking-finetuned-squad`.
|
||||
|
||||
#### Fine-tuning XLNet on SQuAD
|
||||
|
||||
This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
|
||||
|
||||
```bash
|
||||
export SQUAD_DIR=/path/to/SQUAD
|
||||
|
||||
python /data/home/hlu/transformers/examples/run_squad.py \
|
||||
--model_type xlnet \
|
||||
--model_name_or_path xlnet-large-cased \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--do_lower_case \
|
||||
--train_file /data/home/hlu/notebooks/NLP/examples/question_answering/train-v1.1.json \
|
||||
--predict_file /data/home/hlu/notebooks/NLP/examples/question_answering/dev-v1.1.json \
|
||||
--learning_rate 3e-5 \
|
||||
--num_train_epochs 2 \
|
||||
--max_seq_length 384 \
|
||||
--doc_stride 128 \
|
||||
--output_dir ./wwm_cased_finetuned_squad/ \
|
||||
--per_gpu_eval_batch_size=4 \
|
||||
--per_gpu_train_batch_size=4 \
|
||||
--save_steps 5000
|
||||
```
|
||||
|
||||
Training with the previously defined hyper-parameters yields the following results:
|
||||
|
||||
```python
|
||||
{
|
||||
"exact": 85.45884578997162,
|
||||
"f1": 92.5974600601065,
|
||||
"total": 10570,
|
||||
"HasAns_exact": 85.45884578997162,
|
||||
"HasAns_f1": 92.59746006010651,
|
||||
"HasAns_total": 10570
|
||||
}
|
||||
```
|
||||
|
||||
## Named Entity Recognition
|
||||
|
||||
Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
|
||||
[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
|
||||
This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
|
||||
Details and results for the fine-tuning provided by @stefan-it.
|
||||
|
||||
### Data (Download and pre-processing steps)
|
||||
|
||||
Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
|
||||
|
||||
Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
|
||||
|
||||
```bash
|
||||
curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
|
||||
| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
|
||||
curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
|
||||
| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
|
||||
curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
|
||||
| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
|
||||
```
|
||||
|
||||
The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
|
||||
|
||||
```bash
|
||||
wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
|
||||
```
|
||||
Let's define some variables that we need for further pre-processing steps and training the model:
|
||||
|
||||
```bash
|
||||
export MAX_LENGTH=128
|
||||
export BERT_MODEL=bert-base-multilingual-cased
|
||||
```
|
||||
|
||||
Run the pre-processing script on training, dev and test datasets:
|
||||
|
||||
```bash
|
||||
python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
|
||||
python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
|
||||
python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
|
||||
```
|
||||
|
||||
The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
|
||||
|
||||
```bash
|
||||
cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
|
||||
```
|
||||
|
||||
### Prepare the run
|
||||
|
||||
Additional environment variables must be set:
|
||||
|
||||
```bash
|
||||
export OUTPUT_DIR=germeval-model
|
||||
export BATCH_SIZE=32
|
||||
export NUM_EPOCHS=3
|
||||
export SAVE_STEPS=750
|
||||
export SEED=1
|
||||
```
|
||||
|
||||
### Run the Pytorch version
|
||||
|
||||
To start training, just run:
|
||||
|
||||
```bash
|
||||
python3 run_ner.py --data_dir ./ \
|
||||
--model_type bert \
|
||||
--labels ./labels.txt \
|
||||
--model_name_or_path $BERT_MODEL \
|
||||
--output_dir $OUTPUT_DIR \
|
||||
--max_seq_length $MAX_LENGTH \
|
||||
--num_train_epochs $NUM_EPOCHS \
|
||||
--per_gpu_train_batch_size $BATCH_SIZE \
|
||||
--save_steps $SAVE_STEPS \
|
||||
--seed $SEED \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--do_predict
|
||||
```
|
||||
|
||||
If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
|
||||
|
||||
#### Evaluation
|
||||
|
||||
Evaluation on development dataset outputs the following for our example:
|
||||
|
||||
```bash
|
||||
10/04/2019 00:42:06 - INFO - __main__ - ***** Eval results *****
|
||||
10/04/2019 00:42:06 - INFO - __main__ - f1 = 0.8623348017621146
|
||||
10/04/2019 00:42:06 - INFO - __main__ - loss = 0.07183869666975543
|
||||
10/04/2019 00:42:06 - INFO - __main__ - precision = 0.8467916366258111
|
||||
10/04/2019 00:42:06 - INFO - __main__ - recall = 0.8784592370979806
|
||||
```
|
||||
|
||||
On the test dataset the following results could be achieved:
|
||||
|
||||
```bash
|
||||
10/04/2019 00:42:42 - INFO - __main__ - ***** Eval results *****
|
||||
10/04/2019 00:42:42 - INFO - __main__ - f1 = 0.8614389652384803
|
||||
10/04/2019 00:42:42 - INFO - __main__ - loss = 0.07064602487454782
|
||||
10/04/2019 00:42:42 - INFO - __main__ - precision = 0.8604651162790697
|
||||
10/04/2019 00:42:42 - INFO - __main__ - recall = 0.8624150210424085
|
||||
```
|
||||
|
||||
#### Comparing BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased)
|
||||
|
||||
Here is a small comparison between BERT (large, cased), RoBERTa (large, cased) and DistilBERT (base, uncased) with the same hyperparameters as specified in the [example documentation](https://huggingface.co/transformers/examples.html#named-entity-recognition) (one run):
|
||||
|
||||
| Model | F-Score Dev | F-Score Test
|
||||
| --------------------------------- | ------- | --------
|
||||
| `bert-large-cased` | 95.59 | 91.70
|
||||
| `roberta-large` | 95.96 | 91.87
|
||||
| `distilbert-base-uncased` | 94.34 | 90.32
|
||||
|
||||
### Run the Tensorflow 2 version
|
||||
|
||||
To start training, just run:
|
||||
|
||||
```bash
|
||||
python3 run_tf_ner.py --data_dir ./ \
|
||||
--model_type bert \
|
||||
--labels ./labels.txt \
|
||||
--model_name_or_path $BERT_MODEL \
|
||||
--output_dir $OUTPUT_DIR \
|
||||
--max_seq_length $MAX_LENGTH \
|
||||
--num_train_epochs $NUM_EPOCHS \
|
||||
--per_device_train_batch_size $BATCH_SIZE \
|
||||
--save_steps $SAVE_STEPS \
|
||||
--seed $SEED \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--do_predict
|
||||
```
|
||||
|
||||
Such as the Pytorch version, if your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
|
||||
|
||||
#### Evaluation
|
||||
|
||||
Evaluation on development dataset outputs the following for our example:
|
||||
```bash
|
||||
precision recall f1-score support
|
||||
|
||||
LOCderiv 0.7619 0.6154 0.6809 52
|
||||
PERpart 0.8724 0.8997 0.8858 4057
|
||||
OTHpart 0.9360 0.9466 0.9413 711
|
||||
ORGpart 0.7015 0.6989 0.7002 269
|
||||
LOCpart 0.7668 0.8488 0.8057 496
|
||||
LOC 0.8745 0.9191 0.8963 235
|
||||
ORGderiv 0.7723 0.8571 0.8125 91
|
||||
OTHderiv 0.4800 0.6667 0.5581 18
|
||||
OTH 0.5789 0.6875 0.6286 16
|
||||
PERderiv 0.5385 0.3889 0.4516 18
|
||||
PER 0.5000 0.5000 0.5000 2
|
||||
ORG 0.0000 0.0000 0.0000 3
|
||||
|
||||
micro avg 0.8574 0.8862 0.8715 5968
|
||||
macro avg 0.8575 0.8862 0.8713 5968
|
||||
```
|
||||
|
||||
On the test dataset the following results could be achieved:
|
||||
```bash
|
||||
precision recall f1-score support
|
||||
|
||||
PERpart 0.8847 0.8944 0.8896 9397
|
||||
OTHpart 0.9376 0.9353 0.9365 1639
|
||||
ORGpart 0.7307 0.7044 0.7173 697
|
||||
LOC 0.9133 0.9394 0.9262 561
|
||||
LOCpart 0.8058 0.8157 0.8107 1150
|
||||
ORG 0.0000 0.0000 0.0000 8
|
||||
OTHderiv 0.5882 0.4762 0.5263 42
|
||||
PERderiv 0.6571 0.5227 0.5823 44
|
||||
OTH 0.4906 0.6667 0.5652 39
|
||||
ORGderiv 0.7016 0.7791 0.7383 172
|
||||
LOCderiv 0.8256 0.6514 0.7282 109
|
||||
PER 0.0000 0.0000 0.0000 11
|
||||
|
||||
micro avg 0.8722 0.8774 0.8748 13869
|
||||
macro avg 0.8712 0.8774 0.8740 13869
|
||||
```
|
||||
|
||||
## XNLI
|
||||
|
||||
Based on the script [`run_xnli.py`](https://github.com/huggingface/transformers/blob/master/examples/run_xnli.py).
|
||||
|
||||
[XNLI](https://www.nyu.edu/projects/bowman/xnli/) is crowd-sourced dataset based on [MultiNLI](http://www.nyu.edu/projects/bowman/multinli/). It is an evaluation benchmark for cross-lingual text representations. Pairs of text are labeled with textual entailment annotations for 15 different languages (including both high-ressource language such as English and low-ressource languages such as Swahili).
|
||||
|
||||
#### Fine-tuning on XNLI
|
||||
|
||||
This example code fine-tunes mBERT (multi-lingual BERT) on the XNLI dataset. It runs in 106 mins
|
||||
on a single tesla V100 16GB. The data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a
|
||||
`$XNLI_DIR` directory.
|
||||
|
||||
* [XNLI 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-1.0.zip)
|
||||
* [XNLI-MT 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-MT-1.0.zip)
|
||||
|
||||
```bash
|
||||
export XNLI_DIR=/path/to/XNLI
|
||||
|
||||
python run_xnli.py \
|
||||
--model_type bert \
|
||||
--model_name_or_path bert-base-multilingual-cased \
|
||||
--language de \
|
||||
--train_language en \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--data_dir $XNLI_DIR \
|
||||
--per_gpu_train_batch_size 32 \
|
||||
--learning_rate 5e-5 \
|
||||
--num_train_epochs 2.0 \
|
||||
--max_seq_length 128 \
|
||||
--output_dir /tmp/debug_xnli/ \
|
||||
--save_steps -1
|
||||
```
|
||||
|
||||
Training with the previously defined hyper-parameters yields the following results on the **test** set:
|
||||
|
||||
```bash
|
||||
acc = 0.7093812375249501
|
||||
```
|
||||
|
||||
@@ -1,477 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Benchmarking the library on inference and training """
|
||||
|
||||
# If checking the tensors placement
|
||||
# tf.debugging.set_log_device_placement(True)
|
||||
|
||||
from typing import List
|
||||
import timeit
|
||||
from transformers import is_tf_available, is_torch_available
|
||||
from time import time
|
||||
import argparse
|
||||
import csv
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
from transformers import TFAutoModel
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
from transformers import AutoModel
|
||||
|
||||
from transformers import AutoConfig, AutoTokenizer
|
||||
|
||||
input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as
|
||||
the Director of Hatcheries and Conditioning entered the room, in the
|
||||
|
||||
|
||||
|
||||
scarcely breathing silence, the absent-minded, soliloquizing hum or
|
||||
whistle, of absorbed concentration. A troop of newly arrived students,
|
||||
very young, pink and callow, followed nervously, rather abjectly, at the
|
||||
Director's heels. Each of them carried a notebook, in which, whenever
|
||||
the great man spoke, he desperately scribbled. Straight from the
|
||||
horse's mouth. It was a rare privilege. The D. H. C. for Central London
|
||||
always made a point of personally conducting his new students round
|
||||
the various departments.
|
||||
|
||||
"Just to give you a general idea," he would explain to them. For of
|
||||
course some sort of general idea they must have, if they were to do
|
||||
their work intelligently-though as little of one, if they were to be good
|
||||
and happy members of society, as possible. For particulars, as every
|
||||
one knows, make for virtue and happiness; generalities are intellectu-
|
||||
ally necessary evils. Not philosophers but fret-sawyers and stamp col-
|
||||
lectors compose the backbone of society.
|
||||
|
||||
"To-morrow," he would add, smiling at them with a slightly menacing
|
||||
geniality, "you'll be settling down to serious work. You won't have time
|
||||
for generalities. Meanwhile ..."
|
||||
|
||||
Meanwhile, it was a privilege. Straight from the horse's mouth into the
|
||||
notebook. The boys scribbled like mad.
|
||||
|
||||
Tall and rather thin but upright, the Director advanced into the room.
|
||||
He had a long chin and big rather prominent teeth, just covered, when
|
||||
he was not talking, by his full, floridly curved lips. Old, young? Thirty?
|
||||
Fifty? Fifty-five? It was hard to say. And anyhow the question didn't
|
||||
arise; in this year of stability, A. F. 632, it didn't occur to you to ask it.
|
||||
|
||||
"I shall begin at the beginning," said the D.H.C. and the more zealous
|
||||
students recorded his intention in their notebooks: Begin at the begin-
|
||||
ning. "These," he waved his hand, "are the incubators." And opening
|
||||
an insulated door he showed them racks upon racks of numbered test-
|
||||
tubes. "The week's supply of ova. Kept," he explained, "at blood heat;
|
||||
whereas the male gametes," and here he opened another door, "they
|
||||
have to be kept at thirty-five instead of thirty-seven. Full blood heat
|
||||
sterilizes." Rams wrapped in theremogene beget no lambs.
|
||||
|
||||
Still leaning against the incubators he gave them, while the pencils
|
||||
scurried illegibly across the pages, a brief description of the modern
|
||||
|
||||
|
||||
|
||||
fertilizing process; spoke first, of course, of its surgical introduc-
|
||||
tion-"the operation undergone voluntarily for the good of Society, not
|
||||
to mention the fact that it carries a bonus amounting to six months'
|
||||
salary"; continued with some account of the technique for preserving
|
||||
the excised ovary alive and actively developing; passed on to a consid-
|
||||
eration of optimum temperature, salinity, viscosity; referred to the liq-
|
||||
uor in which the detached and ripened eggs were kept; and, leading
|
||||
his charges to the work tables, actually showed them how this liquor
|
||||
was drawn off from the test-tubes; how it was let out drop by drop
|
||||
onto the specially warmed slides of the microscopes; how the eggs
|
||||
which it contained were inspected for abnormalities, counted and
|
||||
transferred to a porous receptacle; how (and he now took them to
|
||||
watch the operation) this receptacle was immersed in a warm bouillon
|
||||
containing free-swimming spermatozoa-at a minimum concentration
|
||||
of one hundred thousand per cubic centimetre, he insisted; and how,
|
||||
after ten minutes, the container was lifted out of the liquor and its
|
||||
contents re-examined; how, if any of the eggs remained unfertilized, it
|
||||
was again immersed, and, if necessary, yet again; how the fertilized
|
||||
ova went back to the incubators; where the Alphas and Betas re-
|
||||
mained until definitely bottled; while the Gammas, Deltas and Epsilons
|
||||
were brought out again, after only thirty-six hours, to undergo Bo-
|
||||
kanovsky's Process.
|
||||
|
||||
"Bokanovsky's Process," repeated the Director, and the students un-
|
||||
derlined the words in their little notebooks.
|
||||
|
||||
One egg, one embryo, one adult-normality. But a bokanovskified egg
|
||||
will bud, will proliferate, will divide. From eight to ninety-six buds, and
|
||||
every bud will grow into a perfectly formed embryo, and every embryo
|
||||
into a full-sized adult. Making ninety-six human beings grow where
|
||||
only one grew before. Progress.
|
||||
|
||||
"Essentially," the D.H.C. concluded, "bokanovskification consists of a
|
||||
series of arrests of development. We check the normal growth and,
|
||||
paradoxically enough, the egg responds by budding."
|
||||
|
||||
Responds by budding. The pencils were busy.
|
||||
|
||||
He pointed. On a very slowly moving band a rack-full of test-tubes was
|
||||
entering a large metal box, another, rack-full was emerging. Machinery
|
||||
faintly purred. It took eight minutes for the tubes to go through, he
|
||||
|
||||
|
||||
|
||||
told them. Eight minutes of hard X-rays being about as much as an
|
||||
egg can stand. A few died; of the rest, the least susceptible divided
|
||||
into two; most put out four buds; some eight; all were returned to the
|
||||
incubators, where the buds began to develop; then, after two days,
|
||||
were suddenly chilled, chilled and checked. Two, four, eight, the buds
|
||||
in their turn budded; and having budded were dosed almost to death
|
||||
with alcohol; consequently burgeoned again and having budded-bud
|
||||
out of bud out of bud-were thereafter-further arrest being generally
|
||||
fatal-left to develop in peace. By which time the original egg was in a
|
||||
fair way to becoming anything from eight to ninety-six embryos- a
|
||||
prodigious improvement, you will agree, on nature. Identical twins-but
|
||||
not in piddling twos and threes as in the old viviparous days, when an
|
||||
egg would sometimes accidentally divide; actually by dozens, by
|
||||
scores at a time.
|
||||
|
||||
"Scores," the Director repeated and flung out his arms, as though he
|
||||
were distributing largesse. "Scores."
|
||||
|
||||
But one of the students was fool enough to ask where the advantage
|
||||
lay.
|
||||
|
||||
"My good boy!" The Director wheeled sharply round on him. "Can't you
|
||||
see? Can't you see?" He raised a hand; his expression was solemn.
|
||||
"Bokanovsky's Process is one of the major instruments of social stabil-
|
||||
ity!"
|
||||
|
||||
Major instruments of social stability.
|
||||
|
||||
Standard men and women; in uniform batches. The whole of a small
|
||||
factory staffed with the products of a single bokanovskified egg.
|
||||
|
||||
"Ninety-six identical twins working ninety-six identical machines!" The
|
||||
voice was almost tremulous with enthusiasm. "You really know where
|
||||
you are. For the first time in history." He quoted the planetary motto.
|
||||
"Community, Identity, Stability." Grand words. "If we could bo-
|
||||
kanovskify indefinitely the whole problem would be solved."
|
||||
|
||||
Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil-
|
||||
lions of identical twins. The principle of mass production at last applied
|
||||
to biology.
|
||||
|
||||
|
||||
|
||||
"But, alas," the Director shook his head, "we can't bokanovskify indefi-
|
||||
nitely."
|
||||
|
||||
Ninety-six seemed to be the limit; seventy-two a good average. From
|
||||
the same ovary and with gametes of the same male to manufacture as
|
||||
many batches of identical twins as possible-that was the best (sadly a
|
||||
second best) that they could do. And even that was difficult.
|
||||
|
||||
"For in nature it takes thirty years for two hundred eggs to reach ma-
|
||||
turity. But our business is to stabilize the population at this moment,
|
||||
here and now. Dribbling out twins over a quarter of a century-what
|
||||
would be the use of that?"
|
||||
|
||||
Obviously, no use at all. But Podsnap's Technique had immensely ac-
|
||||
celerated the process of ripening. They could make sure of at least a
|
||||
hundred and fifty mature eggs within two years. Fertilize and bo-
|
||||
kanovskify-in other words, multiply by seventy-two-and you get an
|
||||
average of nearly eleven thousand brothers and sisters in a hundred
|
||||
and fifty batches of identical twins, all within two years of the same
|
||||
age.
|
||||
|
||||
"And in exceptional cases we can make one ovary yield us over fifteen
|
||||
thousand adult individuals."
|
||||
|
||||
Beckoning to a fair-haired, ruddy young man who happened to be
|
||||
passing at the moment. "Mr. Foster," he called. The ruddy young man
|
||||
approached. "Can you tell us the record for a single ovary, Mr. Foster?"
|
||||
|
||||
"Sixteen thousand and twelve in this Centre," Mr. Foster replied with-
|
||||
out hesitation. He spoke very quickly, had a vivacious blue eye, and
|
||||
took an evident pleasure in quoting figures. "Sixteen thousand and
|
||||
twelve; in one hundred and eighty-nine batches of identicals. But of
|
||||
course they've done much better," he rattled on, "in some of the tropi-
|
||||
cal Centres. Singapore has often produced over sixteen thousand five
|
||||
hundred; and Mombasa has actually touched the seventeen thousand
|
||||
mark. But then they have unfair advantages. You should see the way a
|
||||
negro ovary responds to pituitary! It's quite astonishing, when you're
|
||||
used to working with European material. Still," he added, with a laugh
|
||||
(but the light of combat was in his eyes and the lift of his chin was
|
||||
challenging), "still, we mean to beat them if we can. I'm working on a
|
||||
wonderful Delta-Minus ovary at this moment. Only just eighteen
|
||||
|
||||
|
||||
|
||||
months old. Over twelve thousand seven hundred children already, ei-
|
||||
ther decanted or in embryo. And still going strong. We'll beat them
|
||||
yet."
|
||||
|
||||
"That's the spirit I like!" cried the Director, and clapped Mr. Foster on
|
||||
the shoulder. "Come along with us, and give these boys the benefit of
|
||||
your expert knowledge."
|
||||
|
||||
Mr. Foster smiled modestly. "With pleasure." They went.
|
||||
In the Bottling Room all was harmonious bustle and ordered activity.
|
||||
Flaps of fresh sow's peritoneum ready cut to the proper size came
|
||||
shooting up in little lifts from the Organ Store in the sub-basement.
|
||||
Whizz and then, click! the lift-hatches hew open; the bottle-liner had
|
||||
only to reach out a hand, take the flap, insert, smooth-down, and be-
|
||||
fore the lined bottle had had time to travel out of reach along the end-
|
||||
less band, whizz, click! another flap of peritoneum had shot up from
|
||||
the depths, ready to be slipped into yet another bottle, the next of that
|
||||
slow interminable procession on the band.
|
||||
|
||||
Next to the Liners stood the Matriculators. The procession advanced;
|
||||
one by one the eggs were transferred from their test-tubes to the
|
||||
larger containers; deftly the peritoneal lining was slit, the morula
|
||||
dropped into place, the saline solution poured in ... and already the
|
||||
bottle had passed, and it was the turn of the labellers. Heredity, date
|
||||
of fertilization, membership of Bokanovsky Group-details were trans-
|
||||
ferred from test-tube to bottle. No longer anonymous, but named,
|
||||
identified, the procession marched slowly on; on through an opening in
|
||||
the wall, slowly on into the Social Predestination Room.
|
||||
"Eighty-eight cubic metres of card-index," said Mr. Foster with relish,
|
||||
as they entered."""
|
||||
|
||||
|
||||
def create_setup_and_compute(model_names: List[str],
|
||||
gpu: bool = True,
|
||||
tensorflow: bool = False,
|
||||
average_over: int = 3,
|
||||
torchscript: bool = False,
|
||||
xla: bool = False,
|
||||
amp: bool = False,
|
||||
fp16: bool = False,
|
||||
save_to_csv: bool = False,
|
||||
csv_filename: str = f"results_{round(time())}.csv"):
|
||||
if xla:
|
||||
tf.config.optimizer.set_jit(True)
|
||||
if amp:
|
||||
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
|
||||
|
||||
if tensorflow:
|
||||
dictionary = {model_name: {} for model_name in model_names}
|
||||
results = _compute_tensorflow(model_names, dictionary, average_over, amp)
|
||||
else:
|
||||
device = 'cuda' if (gpu and torch.cuda.is_available()) else 'cpu'
|
||||
dictionary = {model_name: {} for model_name in model_names}
|
||||
results = _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16)
|
||||
|
||||
print("=========== RESULTS ===========")
|
||||
for model_name in model_names:
|
||||
print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
|
||||
for batch_size in results[model_name]["bs"]:
|
||||
print("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
|
||||
for slice_size in results[model_name]["ss"]:
|
||||
result = results[model_name]['results'][batch_size][slice_size]
|
||||
if isinstance(result, str):
|
||||
print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
|
||||
f"{result}")
|
||||
else:
|
||||
print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
|
||||
f"{(round(1000 * result) / 1000)}"
|
||||
f"s")
|
||||
|
||||
if save_to_csv:
|
||||
with open(csv_filename, mode='w') as csv_file:
|
||||
fieldnames = ['model',
|
||||
'1x8', '1x64', '1x128', '1x256', '1x512', '1x1024',
|
||||
'2x8', '2x64', '2x128', '2x256', '2x512', '2x1024',
|
||||
'4x8', '4x64', '4x128', '4x256', '4x512', '4x1024',
|
||||
'8x8', '8x64', '8x128', '8x256', '8x512', '8x1024',
|
||||
]
|
||||
|
||||
writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
|
||||
writer.writeheader()
|
||||
|
||||
for model_name in model_names:
|
||||
model_results = {
|
||||
f'{bs}x{ss}': results[model_name]['results'][bs][ss]
|
||||
for bs in results[model_name]["results"]
|
||||
for ss in results[model_name]['results'][bs]
|
||||
}
|
||||
writer.writerow({'model': model_name, **model_results})
|
||||
|
||||
|
||||
def _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16):
|
||||
for c, model_name in enumerate(model_names):
|
||||
print(f"{c + 1} / {len(model_names)}")
|
||||
config = AutoConfig.from_pretrained(model_name, torchscript=torchscript)
|
||||
model = AutoModel.from_pretrained(model_name, config=config)
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
|
||||
tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
|
||||
|
||||
max_input_size = tokenizer.max_model_input_sizes[model_name]
|
||||
batch_sizes = [1, 2, 4, 8]
|
||||
slice_sizes = [8, 64, 128, 256, 512, 1024]
|
||||
|
||||
dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
|
||||
dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
|
||||
|
||||
for batch_size in batch_sizes:
|
||||
if fp16:
|
||||
model.half()
|
||||
model.to(device)
|
||||
model.eval()
|
||||
for slice_size in slice_sizes:
|
||||
if max_input_size is not None and slice_size > max_input_size:
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
|
||||
else:
|
||||
sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
|
||||
try:
|
||||
if torchscript:
|
||||
print("Tracing model with sequence size", sequence.shape)
|
||||
inference = torch.jit.trace(model, sequence)
|
||||
inference(sequence)
|
||||
else:
|
||||
inference = model
|
||||
inference(sequence)
|
||||
|
||||
print("Going through model with sequence of shape", sequence.shape)
|
||||
runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
|
||||
average_time = sum(runtimes)/float(len(runtimes)) / 3.0
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = average_time
|
||||
except RuntimeError as e:
|
||||
print("Doesn't fit on GPU.", e)
|
||||
torch.cuda.empty_cache()
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
|
||||
return dictionary
|
||||
|
||||
|
||||
def _compute_tensorflow(model_names, dictionary, average_over, amp):
|
||||
for c, model_name in enumerate(model_names):
|
||||
print(f"{c + 1} / {len(model_names)}")
|
||||
config = AutoConfig.from_pretrained(model_name)
|
||||
model = TFAutoModel.from_pretrained(model_name, config=config)
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
|
||||
tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
|
||||
|
||||
max_input_size = tokenizer.max_model_input_sizes[model_name]
|
||||
batch_sizes = [1, 2, 4, 8]
|
||||
slice_sizes = [8, 64, 128, 256, 512, 1024]
|
||||
|
||||
dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
|
||||
dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
|
||||
|
||||
print("Using model", model)
|
||||
|
||||
@tf.function
|
||||
def inference(inputs):
|
||||
return model(inputs)
|
||||
|
||||
for batch_size in batch_sizes:
|
||||
for slice_size in slice_sizes:
|
||||
if max_input_size is not None and slice_size > max_input_size:
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
|
||||
else:
|
||||
sequence = tf.stack([tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size)
|
||||
|
||||
try:
|
||||
print("Going through model with sequence of shape", sequence.shape)
|
||||
# To make sure that the model is traced + that the tensors are on the appropriate device
|
||||
inference(sequence)
|
||||
|
||||
runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
|
||||
average_time = sum(runtimes)/float(len(runtimes)) / 3.0
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = average_time
|
||||
except tf.errors.ResourceExhaustedError as e:
|
||||
print("Doesn't fit on GPU.", e)
|
||||
torch.cuda.empty_cache()
|
||||
dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
|
||||
return dictionary
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
parser.add_argument("--models", required=False, type=str, default='all', help="Model checkpoints to be provided "
|
||||
"to the AutoModel classes. Leave "
|
||||
"blank to benchmark the base version "
|
||||
"of all available model "
|
||||
"architectures.")
|
||||
parser.add_argument("--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the "
|
||||
"models")
|
||||
parser.add_argument("--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available "
|
||||
"cuda devices")
|
||||
parser.add_argument("--torchscript", required=False, action="store_true", help="Pytorch only: trace the models "
|
||||
"using torchscript")
|
||||
parser.add_argument("--tensorflow", required=False, action="store_true", help="Benchmark the TensorFlow version "
|
||||
"of the models. Will run on GPU if "
|
||||
"the correct dependencies are "
|
||||
"installed")
|
||||
parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.")
|
||||
parser.add_argument("--amp", required=False, action="store_true", help="TensorFlow only: use automatic mixed precision acceleration.")
|
||||
parser.add_argument("--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference.")
|
||||
parser.add_argument("--keras_predict", required=False, action="store_true", help="Whether to use model.predict "
|
||||
"instead of model() to do a "
|
||||
"forward pass.")
|
||||
parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.")
|
||||
parser.add_argument("--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv.")
|
||||
parser.add_argument("--average_over", required=False, default=30, type=int, help="Times an experiment will be run.")
|
||||
|
||||
args = parser.parse_args()
|
||||
if args.models == 'all':
|
||||
args.models = [
|
||||
"gpt2",
|
||||
"bert-base-cased",
|
||||
"xlnet-base-cased",
|
||||
"xlm-mlm-en-2048",
|
||||
"transfo-xl-wt103",
|
||||
"openai-gpt",
|
||||
"distilbert-base-uncased",
|
||||
"distilgpt2",
|
||||
"roberta-base",
|
||||
"ctrl"
|
||||
]
|
||||
else:
|
||||
args.models = args.models.split()
|
||||
|
||||
print("Running with arguments", args)
|
||||
|
||||
if args.torch:
|
||||
if is_torch_available():
|
||||
create_setup_and_compute(
|
||||
model_names=args.models,
|
||||
tensorflow=False,
|
||||
gpu=args.torch_cuda,
|
||||
torchscript=args.torchscript,
|
||||
fp16=args.fp16,
|
||||
save_to_csv=args.save_to_csv,
|
||||
csv_filename=args.csv_filename,
|
||||
average_over=args.average_over
|
||||
)
|
||||
else:
|
||||
raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.")
|
||||
|
||||
if args.tensorflow:
|
||||
if is_tf_available():
|
||||
create_setup_and_compute(
|
||||
model_names=args.models,
|
||||
tensorflow=True,
|
||||
xla=args.xla,
|
||||
amp=args.amp,
|
||||
save_to_csv=args.save_to_csv,
|
||||
csv_filename=args.csv_filename,
|
||||
average_over=args.average_over
|
||||
)
|
||||
else:
|
||||
raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
|
||||
@@ -1,48 +0,0 @@
|
||||
from pathlib import Path
|
||||
import tarfile
|
||||
import urllib.request
|
||||
|
||||
import torch
|
||||
|
||||
from transformers.tokenization_camembert import CamembertTokenizer
|
||||
from transformers.modeling_camembert import CamembertForMaskedLM
|
||||
|
||||
|
||||
def fill_mask(masked_input, model, tokenizer, topk=5):
|
||||
# Adapted from https://github.com/pytorch/fairseq/blob/master/fairseq/models/roberta/hub_interface.py
|
||||
assert masked_input.count('<mask>') == 1
|
||||
input_ids = torch.tensor(tokenizer.encode(masked_input, add_special_tokens=True)).unsqueeze(0) # Batch size 1
|
||||
logits = model(input_ids)[0] # The last hidden-state is the first element of the output tuple
|
||||
masked_index = (input_ids.squeeze() == tokenizer.mask_token_id).nonzero().item()
|
||||
logits = logits[0, masked_index, :]
|
||||
prob = logits.softmax(dim=0)
|
||||
values, indices = prob.topk(k=topk, dim=0)
|
||||
topk_predicted_token_bpe = ' '.join([tokenizer.convert_ids_to_tokens(indices[i].item())
|
||||
for i in range(len(indices))])
|
||||
masked_token = tokenizer.mask_token
|
||||
topk_filled_outputs = []
|
||||
for index, predicted_token_bpe in enumerate(topk_predicted_token_bpe.split(' ')):
|
||||
predicted_token = predicted_token_bpe.replace('\u2581', ' ')
|
||||
if " {0}".format(masked_token) in masked_input:
|
||||
topk_filled_outputs.append((
|
||||
masked_input.replace(
|
||||
' {0}'.format(masked_token), predicted_token
|
||||
),
|
||||
values[index].item(),
|
||||
predicted_token,
|
||||
))
|
||||
else:
|
||||
topk_filled_outputs.append((
|
||||
masked_input.replace(masked_token, predicted_token),
|
||||
values[index].item(),
|
||||
predicted_token,
|
||||
))
|
||||
return topk_filled_outputs
|
||||
|
||||
|
||||
tokenizer = CamembertTokenizer.from_pretrained('camembert-base')
|
||||
model = CamembertForMaskedLM.from_pretrained('camembert-base')
|
||||
model.eval()
|
||||
|
||||
masked_input = "Le camembert est <mask> :)"
|
||||
print(fill_mask(masked_input, model, tokenizer, topk=3))
|
||||
@@ -41,7 +41,7 @@ from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
|
||||
from transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
|
||||
AdamW, cached_path, WEIGHTS_NAME, CONFIG_NAME,
|
||||
get_linear_schedule_with_warmup)
|
||||
WarmupLinearSchedule)
|
||||
|
||||
ROCSTORIES_URL = "https://s3.amazonaws.com/datasets.huggingface.co/ROCStories.tar.gz"
|
||||
|
||||
@@ -211,7 +211,7 @@ def main():
|
||||
{'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
|
||||
if args.do_train:
|
||||
nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
|
||||
@@ -237,7 +237,7 @@ def main():
|
||||
# Save a trained model
|
||||
if args.do_train:
|
||||
# Save a trained model, configuration and tokenizer
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model itself
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self
|
||||
|
||||
# If we save using the predefined names, we can load using `from_pretrained`
|
||||
output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
|
||||
|
||||
@@ -31,18 +31,14 @@ import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
BertForMultipleChoice, BertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
from transformers import AdamW, WarmupLinearSchedule
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -322,7 +318,7 @@ def train(args, train_dataset, model, tokenizer):
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
|
||||
@@ -1,71 +1,35 @@
|
||||
# Distil*
|
||||
# DistilBERT
|
||||
|
||||
This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
|
||||
This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.
|
||||
|
||||
**December 6th, 2019 - Update** We release **DistilmBERT**: 92% of `bert-base-multilingual-cased` on XNLI. The model supports 104 different languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
|
||||
**2019, September 19th - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
|
||||
|
||||
**November 19th, 2019 - Update** We release German **DistilBERT**: 98.8% of `bert-base-german-dbmdz-cased` on NER tasks.
|
||||
## What is DistilBERT
|
||||
|
||||
**October 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
|
||||
DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
|
||||
|
||||
**October 3rd, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
|
||||
For more information on DistilBERT, please refer to our [detailed blog post](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-distilbert-a-distilled-version-of-bert-8cf3380435b5
|
||||
). *Please note that we will publish a formal write-up with updated and more complete results in the near future (September 19th).*
|
||||
|
||||
**September 19th, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
|
||||
Here's the updated results on the dev sets of GLUE:
|
||||
|
||||
|
||||
## What is Distil*
|
||||
|
||||
Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
|
||||
|
||||
We have applied the same method to other Transformer architectures and released the weights:
|
||||
- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
|
||||
- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base`'s performance on GLUE while being twice faster and 35% smaller.
|
||||
- German BERT: **German DistilBERT** reaches 99% of `bert-base-german-dbmdz-cased`'s performance on German NER (CoNLL-2003).
|
||||
- Multilingual BERT: **DistilmBERT** reaches 92% of Multilingual BERT's performance on XNLI while being twice faster and 25% smaller. The model supports 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
|
||||
|
||||
For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
|
||||
|
||||
Here are the results on the dev sets of GLUE:
|
||||
|
||||
| Model | Macro-score | CoLA | MNLI | MRPC | QNLI | QQP | RTE | SST-2| STS-B| WNLI |
|
||||
| :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---: |
|
||||
| BERT-base | **77.6** | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7 |
|
||||
| DistilBERT | **76.8** | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4 |
|
||||
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
|
||||
| RoBERTa-base (reported) | **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup> |
|
||||
| DistilRoBERTa<sup>1</sup> | **79.0**/**82.3**<sup>2</sup> | 59.4 | 83.9 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1 |
|
||||
|
||||
<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
|
||||
|
||||
<sup>2</sup> Macro-score computed without WNLI.
|
||||
|
||||
<sup>3</sup> We compute this score ourselves for completeness.
|
||||
|
||||
Here are the results on the *test* sets for 6 of the languages available in XNLI. The results are computed in the zero shot setting (trained on the English portion and evaluated on the target language portion):
|
||||
|
||||
| Model | English | Spanish | Chinese | German | Arabic | Urdu |
|
||||
| :---: | :---: | :---: | :---: | :---: | :---: | :---:|
|
||||
| mBERT base cased (computed) | 82.1 | 74.6 | 69.1 | 72.3 | 66.4 | 58.5 |
|
||||
| mBERT base uncased (reported)| 81.4 | 74.3 | 63.8 | 70.5 | 62.1 | 58.3 |
|
||||
| DistilmBERT | 78.2 | 69.1 | 64.0 | 66.3 | 59.1 | 54.7 |
|
||||
| Model | Macro-score | CoLA | MNLI | MRPC | QNLI | QQP | RTE | SST-2 | STS-B | WNLI |
|
||||
| :---: | :---: | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:|
|
||||
| BERT-base | **77.6** | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7 |
|
||||
| DistilBERT | **75.2** | 49.1 | 81.8 | 90.2 | 87.0 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4 |
|
||||
|
||||
## Setup
|
||||
|
||||
This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
|
||||
|
||||
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
|
||||
|
||||
**Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0). It is important to note that there is a small internal bug in the current version of PyTorch available on pip that causes a memory leak in our training/distillation. It has been recently fixed and will likely be integrated into the next release. For the moment, we recommend to [compile PyTorch from source](https://github.com/pytorch/pytorch#from-source). Please refer to [issue 1179](https://github.com/huggingface/transformers/issues/1179) for more details.
|
||||
|
||||
## How to use DistilBERT
|
||||
|
||||
Transformers includes five pre-trained Distil* models, currently only provided for English and German (we are investigating the possibility to train and release a multilingual version of DistilBERT):
|
||||
Transformers includes two pre-trained DistilBERT models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
|
||||
|
||||
- `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
|
||||
- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
|
||||
- `distilbert-base-german-cased`: DistilBERT German language model pretrained on 1/2 of the data used to pretrain Bert using distillation with the supervision of the `bert-base-german-dbmdz-cased` version of German DBMDZ Bert. For NER tasks the model reaches a F1 score of 83.49 on the CoNLL-2003 test set (for comparison, `bert-base-german-dbmdz-cased` reaches a 84.52 F1 score), and a F1 score of 85.23 on the GermEval 2014 test set (`bert-base-german-dbmdz-cased` reaches a 86.89 F1 score).
|
||||
- `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
|
||||
- `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
|
||||
- `distilbert-base-multilingual-cased`: DistilmBERT multilingual model pretrained with the supervision of `bert-base-multilingual-cased` on the concatenation of Wikipedia in 104 different languages. The model supports the 104 languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages). The model has 6 layers, 768 dimension and 12 heads, totalizing 134M parameters (compared to 177M parameters for mBERT-base). On average DistilmBERT is twice as fast as mBERT-base.
|
||||
|
||||
Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
|
||||
|
||||
@@ -78,15 +42,9 @@ outputs = model(input_ids)
|
||||
last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple
|
||||
```
|
||||
|
||||
Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
|
||||
- DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
|
||||
- DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
|
||||
- DistilmBERT: `model = DistilBertModel.from_pretrained('distilbert-base-multilingual-cased')`
|
||||
## How to train DistilBERT
|
||||
|
||||
|
||||
## How to train Distil*
|
||||
|
||||
In the following, we will explain how you can train DistilBERT.
|
||||
In the following, we will explain how you can train your own compressed model.
|
||||
|
||||
### A. Preparing the data
|
||||
|
||||
@@ -99,8 +57,7 @@ First, we will binarize the data, i.e. tokenize the data and convert each token
|
||||
```bash
|
||||
python scripts/binarized_data.py \
|
||||
--file_path data/dump.txt \
|
||||
--tokenizer_type bert \
|
||||
--tokenizer_name bert-base-uncased \
|
||||
--bert_tokenizer bert-base-uncased \
|
||||
--dump_file data/binarized_text
|
||||
```
|
||||
|
||||
@@ -109,8 +66,7 @@ Our implementation of masked language modeling loss follows [XLM](https://github
|
||||
```bash
|
||||
python scripts/token_counts.py \
|
||||
--data_file data/binarized_text.bert-base-uncased.pickle \
|
||||
--token_counts_dump data/token_counts.bert-base-uncased.pickle \
|
||||
--vocab_size 30522
|
||||
--token_counts_dump data/token_counts.bert-base-uncased.pickle
|
||||
```
|
||||
|
||||
### B. Training
|
||||
@@ -119,12 +75,6 @@ Training with distillation is really simple once you have pre-processed the data
|
||||
|
||||
```bash
|
||||
python train.py \
|
||||
--student_type distilbert \
|
||||
--student_config training_configs/distilbert-base-uncased.json \
|
||||
--teacher_type bert \
|
||||
--teacher_name bert-base-uncased \
|
||||
--alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --alpha_clm 0.0 --mlm \
|
||||
--freeze_pos_embs \
|
||||
--dump_path serialization_dir/my_first_training \
|
||||
--data_file data/binarized_text.bert-base-uncased.pickle \
|
||||
--token_counts data/token_counts.bert-base-uncased.pickle \
|
||||
@@ -133,7 +83,7 @@ python train.py \
|
||||
|
||||
By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
|
||||
|
||||
We highly encourage you to use distributed training for training DistilBERT as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
|
||||
We highly encourage you to use distributed training for training DistilBert as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
|
||||
|
||||
```bash
|
||||
export NODE_RANK=0
|
||||
@@ -155,30 +105,11 @@ python -m torch.distributed.launch \
|
||||
train.py \
|
||||
--force \
|
||||
--n_gpu $WORLD_SIZE \
|
||||
--student_type distilbert \
|
||||
--student_config training_configs/distilbert-base-uncased.json \
|
||||
--teacher_type bert \
|
||||
--teacher_name bert-base-uncased \
|
||||
--alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --alpha_clm 0.0 --mlm \
|
||||
--freeze_pos_embs \
|
||||
--dump_path serialization_dir/my_first_training \
|
||||
--data_file data/binarized_text.bert-base-uncased.pickle \
|
||||
--token_counts data/token_counts.bert-base-uncased.pickle
|
||||
--token_counts data/token_counts.bert-base-uncased.pickle \
|
||||
--dump_path serialization_dir/my_first_distillation
|
||||
```
|
||||
|
||||
**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract.py` and `scripts/extract_distilbert.py` to create a valid initialization checkpoint and use `--student_pretrained_weights` argument to use this initialization for the distilled training!
|
||||
**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract_for_distil.py` to create a valid initialization checkpoint and use `--from_pretrained_weights` and `--from_pretrained_config` arguments to use this initialization for the distilled training!
|
||||
|
||||
Happy distillation!
|
||||
|
||||
## Citation
|
||||
|
||||
If you find the ressource useful, you should cite the following paper:
|
||||
|
||||
```
|
||||
@inproceedings{sanh2019distilbert,
|
||||
title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
|
||||
author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
|
||||
booktitle={NeurIPS EMC^2 Workshop},
|
||||
year={2019}
|
||||
}
|
||||
```
|
||||
|
||||
@@ -12,33 +12,30 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Dataset to distilled models
|
||||
""" Dataloaders to train DistilBERT
|
||||
adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
|
||||
"""
|
||||
import torch
|
||||
from torch.utils.data import Dataset
|
||||
|
||||
from typing import List
|
||||
import math
|
||||
from itertools import chain
|
||||
from collections import Counter
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
from utils import logger
|
||||
|
||||
class LmSeqsDataset(Dataset):
|
||||
"""Custom Dataset wrapping language modeling sequences.
|
||||
|
||||
Each sample will be retrieved by indexing the list of token_ids and their corresponding lengths.
|
||||
|
||||
Input:
|
||||
------
|
||||
params: `NameSpace` parameters
|
||||
data: `List[np.array[int]]
|
||||
"""
|
||||
|
||||
class Dataset:
|
||||
def __init__(self,
|
||||
params,
|
||||
data):
|
||||
self.params = params
|
||||
self.tokens_per_batch = params.tokens_per_batch
|
||||
self.batch_size = params.batch_size
|
||||
self.shuffle = params.shuffle
|
||||
self.group_by_size = params.group_by_size
|
||||
|
||||
self.token_ids = np.array(data)
|
||||
self.lengths = np.array([len(t) for t in data])
|
||||
self.lengths = np.uint16([len(t) for t in data])
|
||||
|
||||
self.check()
|
||||
self.remove_long_sequences()
|
||||
@@ -46,9 +43,6 @@ class LmSeqsDataset(Dataset):
|
||||
self.check()
|
||||
self.print_statistics()
|
||||
|
||||
def __getitem__(self, index):
|
||||
return (self.token_ids[index], self.lengths[index])
|
||||
|
||||
def __len__(self):
|
||||
return len(self.lengths)
|
||||
|
||||
@@ -57,14 +51,12 @@ class LmSeqsDataset(Dataset):
|
||||
Some sanity checks
|
||||
"""
|
||||
assert len(self.token_ids) == len(self.lengths)
|
||||
assert all(self.lengths[i] == len(self.token_ids[i]) for i in range(len(self.lengths)))
|
||||
|
||||
def remove_long_sequences(self):
|
||||
"""
|
||||
Sequences that are too long are splitted by chunk of max_model_input_size.
|
||||
Sequences that are too long are splitted by chunk of max_position_embeddings.
|
||||
"""
|
||||
max_len = self.params.max_model_input_size
|
||||
indices = self.lengths > max_len
|
||||
indices = self.lengths >= self.params.max_position_embeddings
|
||||
logger.info(f'Splitting {sum(indices)} too long sequences.')
|
||||
|
||||
def divide_chunks(l, n):
|
||||
@@ -72,13 +64,10 @@ class LmSeqsDataset(Dataset):
|
||||
|
||||
new_tok_ids = []
|
||||
new_lengths = []
|
||||
if self.params.mlm:
|
||||
cls_id, sep_id = self.params.special_tok_ids['cls_token'], self.params.special_tok_ids['sep_token']
|
||||
else:
|
||||
cls_id, sep_id = self.params.special_tok_ids['bos_token'], self.params.special_tok_ids['eos_token']
|
||||
cls_id, sep_id = self.params.special_tok_ids['cls_token'], self.params.special_tok_ids['sep_token']
|
||||
max_len = self.params.max_position_embeddings
|
||||
|
||||
for seq_, len_ in zip(self.token_ids, self.lengths):
|
||||
assert (seq_[0] == cls_id) and (seq_[-1] == sep_id), seq_
|
||||
if len_ <= max_len:
|
||||
new_tok_ids.append(seq_)
|
||||
new_lengths.append(len_)
|
||||
@@ -90,7 +79,6 @@ class LmSeqsDataset(Dataset):
|
||||
if sub_s[-1] != sep_id:
|
||||
sub_s = np.insert(sub_s, len(sub_s), sep_id)
|
||||
assert len(sub_s) <= max_len
|
||||
assert (sub_s[0] == cls_id) and (sub_s[-1] == sep_id), sub_s
|
||||
sub_seqs.append(sub_s)
|
||||
|
||||
new_tok_ids.extend(sub_seqs)
|
||||
@@ -125,27 +113,89 @@ class LmSeqsDataset(Dataset):
|
||||
# nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
|
||||
# logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
|
||||
|
||||
def select_data(self, a: int, b: int):
|
||||
"""
|
||||
Select a subportion of the data.
|
||||
"""
|
||||
n_sequences = len(self)
|
||||
assert 0 <= a < b <= n_sequences, ValueError(f'`0 <= a < b <= n_sequences` is not met with a={a} and b={b}')
|
||||
|
||||
logger.info(f'Selecting sequences from {a} to {b} (excluded).')
|
||||
self.token_ids = self.token_ids[a:b]
|
||||
self.lengths = self.lengths[a:b]
|
||||
|
||||
self.check()
|
||||
|
||||
def split(self):
|
||||
"""
|
||||
Distributed training: split the data accross the processes.
|
||||
"""
|
||||
assert self.params.n_gpu > 1
|
||||
logger.info('Splitting the data accross the processuses.')
|
||||
n_seq = len(self)
|
||||
n_seq_per_procesus = n_seq // self.params.world_size
|
||||
a = n_seq_per_procesus * self.params.global_rank
|
||||
b = a + n_seq_per_procesus
|
||||
self.select_data(a=a, b=b)
|
||||
|
||||
def batch_sequences(self,
|
||||
batch):
|
||||
token_ids: List[List[int]],
|
||||
lengths: List[int]):
|
||||
"""
|
||||
Do the padding and transform into torch.tensor.
|
||||
"""
|
||||
token_ids = [t[0] for t in batch]
|
||||
lengths = [t[1] for t in batch]
|
||||
assert len(token_ids) == len(lengths)
|
||||
|
||||
# Max for paddings
|
||||
max_seq_len_ = max(lengths)
|
||||
|
||||
# Pad token ids
|
||||
if self.params.mlm:
|
||||
pad_idx = self.params.special_tok_ids['pad_token']
|
||||
else:
|
||||
pad_idx = self.params.special_tok_ids['unk_token']
|
||||
pad_idx = self.params.special_tok_ids['pad_token']
|
||||
tk_ = [list(t.astype(int)) + [pad_idx]*(max_seq_len_-len(t)) for t in token_ids]
|
||||
assert len(tk_) == len(token_ids)
|
||||
assert all(len(t) == max_seq_len_ for t in tk_)
|
||||
|
||||
tk_t = torch.tensor(tk_) # (bs, max_seq_len_)
|
||||
lg_t = torch.tensor(lengths) # (bs)
|
||||
tk_t = torch.tensor(tk_) # (bs, max_seq_len_)
|
||||
lg_t = torch.tensor(lengths.astype(int)) # (bs)
|
||||
return tk_t, lg_t
|
||||
|
||||
def get_batches_iterator(self,
|
||||
batches):
|
||||
"""
|
||||
Return an iterator over batches.
|
||||
"""
|
||||
for sequences_ids in batches:
|
||||
token_ids, lengths = self.batch_sequences(self.token_ids[sequences_ids],
|
||||
self.lengths[sequences_ids])
|
||||
yield (token_ids, lengths)
|
||||
|
||||
def get_iterator(self,
|
||||
seed: int = None):
|
||||
"""
|
||||
Return a data iterator.
|
||||
"""
|
||||
rng = np.random.RandomState(seed)
|
||||
|
||||
n_sequences = len(self)
|
||||
indices = np.arange(n_sequences)
|
||||
|
||||
if self.group_by_size:
|
||||
indices = indices[np.argsort(self.lengths[indices], kind='mergesort')]
|
||||
|
||||
if self.tokens_per_batch == -1:
|
||||
batches = np.array_split(indices, math.ceil(len(indices) * 1. / self.batch_size))
|
||||
else:
|
||||
assert self.tokens_per_batch > 0
|
||||
batch_ids = np.cumsum(self.lengths[indices]) // self.tokens_per_batch
|
||||
_, bounds = np.unique(batch_ids, return_index=True)
|
||||
batches = [indices[bounds[i]:bounds[i + 1]] for i in range(len(bounds) - 1)]
|
||||
if bounds[-1] < len(indices):
|
||||
batches.append(indices[bounds[-1]:])
|
||||
|
||||
if self.shuffle:
|
||||
rng.shuffle(batches)
|
||||
|
||||
assert n_sequences == sum([len(x) for x in batches])
|
||||
assert self.lengths[indices].sum() == sum([self.lengths[x].sum() for x in batches])
|
||||
|
||||
return self.get_batches_iterator(batches=batches)
|
||||
@@ -12,38 +12,32 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" The distiller to distil the student.
|
||||
Adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
|
||||
""" The distiller to distil DistilBERT
|
||||
adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
|
||||
"""
|
||||
import os
|
||||
import math
|
||||
import psutil
|
||||
import time
|
||||
from tensorboardX import SummaryWriter
|
||||
from tqdm import trange, tqdm
|
||||
import numpy as np
|
||||
import psutil
|
||||
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
from torch.optim import AdamW
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
from torch.utils.data import RandomSampler, BatchSampler, DataLoader
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from transformers import get_linear_schedule_with_warmup
|
||||
from transformers import WarmupLinearSchedule
|
||||
|
||||
from utils import logger
|
||||
from lm_seqs_dataset import LmSeqsDataset
|
||||
from grouped_batch_sampler import GroupedBatchSampler, create_lengths_groups
|
||||
from dataset import Dataset
|
||||
|
||||
class Distiller:
|
||||
def __init__(self,
|
||||
params: dict,
|
||||
dataset: LmSeqsDataset,
|
||||
dataloader: Dataset,
|
||||
token_probs: torch.tensor,
|
||||
student: nn.Module,
|
||||
teacher: nn.Module):
|
||||
@@ -56,47 +50,33 @@ class Distiller:
|
||||
self.student = student
|
||||
self.teacher = teacher
|
||||
|
||||
self.student_config = student.config
|
||||
self.vocab_size = student.config.vocab_size
|
||||
|
||||
if params.n_gpu <= 1:
|
||||
sampler = RandomSampler(dataset)
|
||||
else:
|
||||
sampler = DistributedSampler(dataset)
|
||||
|
||||
if params.group_by_size:
|
||||
groups = create_lengths_groups(lengths=dataset.lengths, k=params.max_model_input_size)
|
||||
sampler = GroupedBatchSampler(sampler=sampler, group_ids=groups, batch_size=params.batch_size)
|
||||
else:
|
||||
sampler = BatchSampler(sampler=sampler, batch_size=params.batch_size, drop_last=False)
|
||||
|
||||
self.dataloader = DataLoader(dataset=dataset,
|
||||
batch_sampler=sampler,
|
||||
collate_fn=dataset.batch_sequences)
|
||||
self.dataloader = dataloader
|
||||
if self.params.n_gpu > 1:
|
||||
self.dataloader.split()
|
||||
self.get_iterator(seed=params.seed)
|
||||
|
||||
self.temperature = params.temperature
|
||||
assert self.temperature > 0.
|
||||
|
||||
self.alpha_ce = params.alpha_ce
|
||||
self.alpha_mlm = params.alpha_mlm
|
||||
self.alpha_clm = params.alpha_clm
|
||||
self.alpha_mse = params.alpha_mse
|
||||
self.alpha_cos = params.alpha_cos
|
||||
assert self.alpha_ce >= 0.
|
||||
assert self.alpha_mlm >= 0.
|
||||
assert self.alpha_mse >= 0.
|
||||
assert self.alpha_cos >= 0.
|
||||
assert self.alpha_ce + self.alpha_mlm + self.alpha_mse + self.alpha_cos > 0.
|
||||
|
||||
self.mlm = params.mlm
|
||||
if self.mlm:
|
||||
logger.info(f'Using MLM loss for LM step.')
|
||||
self.mlm_mask_prop = params.mlm_mask_prop
|
||||
assert 0.0 <= self.mlm_mask_prop <= 1.0
|
||||
assert params.word_mask + params.word_keep + params.word_rand == 1.0
|
||||
self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
|
||||
self.pred_probs = self.pred_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else self.pred_probs
|
||||
self.token_probs = token_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else token_probs
|
||||
if self.fp16:
|
||||
self.pred_probs = self.pred_probs.half()
|
||||
self.token_probs = self.token_probs.half()
|
||||
else:
|
||||
logger.info(f'Using CLM loss for LM step.')
|
||||
self.mlm_mask_prop = params.mlm_mask_prop
|
||||
assert 0.0 <= self.mlm_mask_prop <= 1.0
|
||||
assert params.word_mask + params.word_keep + params.word_rand == 1.0
|
||||
self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
|
||||
self.pred_probs = self.pred_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else self.pred_probs
|
||||
self.token_probs = token_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else token_probs
|
||||
if self.fp16:
|
||||
self.pred_probs = self.pred_probs.half()
|
||||
self.token_probs = self.token_probs.half()
|
||||
|
||||
self.epoch = 0
|
||||
self.n_iter = 0
|
||||
@@ -106,13 +86,12 @@ class Distiller:
|
||||
self.last_loss = 0
|
||||
self.last_loss_ce = 0
|
||||
self.last_loss_mlm = 0
|
||||
self.last_loss_clm = 0
|
||||
if self.alpha_mse > 0.: self.last_loss_mse = 0
|
||||
if self.alpha_cos > 0.: self.last_loss_cos = 0
|
||||
self.last_log = 0
|
||||
|
||||
self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
|
||||
self.lm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
|
||||
self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
|
||||
if self.alpha_mse > 0.:
|
||||
self.mse_loss_fct = nn.MSELoss(reduction='sum')
|
||||
if self.alpha_cos > 0.:
|
||||
@@ -120,7 +99,7 @@ class Distiller:
|
||||
|
||||
logger.info('--- Initializing model optimizer')
|
||||
assert params.gradient_accumulation_steps >= 1
|
||||
self.num_steps_epoch = len(self.dataloader)
|
||||
self.num_steps_epoch = int(len(self.dataloader) / params.batch_size) + 1
|
||||
num_train_optimization_steps = int(self.num_steps_epoch / params.gradient_accumulation_steps * params.n_epoch) + 1
|
||||
|
||||
no_decay = ['bias', 'LayerNorm.weight']
|
||||
@@ -136,9 +115,9 @@ class Distiller:
|
||||
betas=(0.9, 0.98))
|
||||
|
||||
warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
|
||||
self.scheduler = get_linear_schedule_with_warmup(self.optimizer,
|
||||
num_warmup_steps=warmup_steps,
|
||||
num_training_steps=num_train_optimization_steps)
|
||||
self.scheduler = WarmupLinearSchedule(self.optimizer,
|
||||
warmup_steps=warmup_steps,
|
||||
t_total=num_train_optimization_steps)
|
||||
|
||||
if self.fp16:
|
||||
try:
|
||||
@@ -161,18 +140,43 @@ class Distiller:
|
||||
logger.info("Using nn.parallel.DistributedDataParallel for distributed training.")
|
||||
self.student = DistributedDataParallel(self.student,
|
||||
device_ids=[params.local_rank],
|
||||
output_device=params.local_rank,
|
||||
find_unused_parameters=True)
|
||||
output_device=params.local_rank)
|
||||
|
||||
self.is_master = params.is_master
|
||||
if self.is_master:
|
||||
logger.info('--- Initializing Tensorboard')
|
||||
self.tensorboard = SummaryWriter(log_dir=os.path.join(self.dump_path, 'log', 'train'))
|
||||
self.tensorboard.add_text(tag='config/training', text_string=str(self.params), global_step=0)
|
||||
self.tensorboard.add_text(tag='config/student', text_string=str(self.student_config), global_step=0)
|
||||
self.tensorboard.add_text(tag='config', text_string=str(self.params), global_step=0)
|
||||
|
||||
def prepare_batch_mlm(self,
|
||||
batch):
|
||||
def get_iterator(self,
|
||||
seed: int = None):
|
||||
"""
|
||||
Initialize the data iterator.
|
||||
Each process has its own data iterator (iterating on his own random portion of the dataset).
|
||||
|
||||
Input:
|
||||
------
|
||||
seed: `int` - The random seed.
|
||||
"""
|
||||
logger.info('--- Initializing Data Iterator')
|
||||
self.data_iterator = self.dataloader.get_iterator(seed=seed)
|
||||
|
||||
def get_batch(self):
|
||||
"""
|
||||
Call the data iterator to output a new batch.
|
||||
If the data iterator went through the whole dataset, create a new iterator.
|
||||
"""
|
||||
assert hasattr(self, 'data_iterator')
|
||||
try:
|
||||
x = next(self.data_iterator)
|
||||
except StopIteration:
|
||||
logger.warning('--- Went through the whole dataset. Creating new data iterator.')
|
||||
self.data_iterator = self.dataloader.get_iterator()
|
||||
x = next(self.data_iterator)
|
||||
return x
|
||||
|
||||
def prepare_batch(self,
|
||||
batch):
|
||||
"""
|
||||
Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the masked label for MLM.
|
||||
|
||||
@@ -218,7 +222,7 @@ class Distiller:
|
||||
assert pred_mask.sum().item() % 8 == 0, pred_mask.sum().item()
|
||||
|
||||
_token_ids_real = token_ids[pred_mask]
|
||||
_token_ids_rand = _token_ids_real.clone().random_(self.vocab_size)
|
||||
_token_ids_rand = _token_ids_real.clone().random_(self.params.vocab_size)
|
||||
_token_ids_mask = _token_ids_real.clone().fill_(self.params.special_tok_ids['mask_token'])
|
||||
probs = torch.multinomial(self.pred_probs, len(_token_ids_real), replacement=True)
|
||||
_token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
|
||||
@@ -226,41 +230,8 @@ class Distiller:
|
||||
|
||||
mlm_labels[~pred_mask] = -1 # previously `mlm_labels[1-pred_mask] = -1`, cf pytorch 1.2.0 compatibility
|
||||
|
||||
# sanity checks
|
||||
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
|
||||
|
||||
return token_ids, attn_mask, mlm_labels
|
||||
|
||||
def prepare_batch_clm(self,
|
||||
batch):
|
||||
"""
|
||||
Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the labels for CLM.
|
||||
|
||||
Input:
|
||||
------
|
||||
batch: `Tuple`
|
||||
token_ids: `torch.tensor(bs, seq_length)` - The token ids for each of the sequence. It is padded.
|
||||
lengths: `torch.tensor(bs)` - The lengths of each of the sequences in the batch.
|
||||
|
||||
Output:
|
||||
-------
|
||||
token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
|
||||
attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
|
||||
clm_labels: `torch.tensor(bs, seq_length)` - The causal languge modeling labels. There is a -1 where there is nothing to predict.
|
||||
"""
|
||||
token_ids, lengths = batch
|
||||
token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
|
||||
assert token_ids.size(0) == lengths.size(0)
|
||||
|
||||
attn_mask = (torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None])
|
||||
clm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
|
||||
clm_labels[~attn_mask] = -1 # previously `clm_labels[1-attn_mask] = -1`, cf pytorch 1.2.0 compatibility
|
||||
|
||||
# sanity checks
|
||||
assert 0 <= token_ids.min() <= token_ids.max() < self.vocab_size
|
||||
|
||||
return token_ids, attn_mask, clm_labels
|
||||
|
||||
def round_batch(self,
|
||||
x: torch.tensor,
|
||||
lengths: torch.tensor):
|
||||
@@ -298,10 +269,7 @@ class Distiller:
|
||||
if ml1 % 8 != 0:
|
||||
pad = 8 - (ml1 % 8)
|
||||
ml2 = ml1 + pad
|
||||
if self.mlm:
|
||||
pad_id = self.params.special_tok_ids['pad_token']
|
||||
else:
|
||||
pad_id = self.params.special_tok_ids['unk_token']
|
||||
pad_id = self.params.special_tok_ids['pad_token']
|
||||
padding_tensor = torch.zeros(bs2, pad, dtype=torch.long, device=x.device).fill_(pad_id)
|
||||
x = torch.cat([x, padding_tensor], 1)
|
||||
assert x.size() == (bs2, ml2)
|
||||
@@ -324,16 +292,14 @@ class Distiller:
|
||||
if self.multi_gpu:
|
||||
torch.distributed.barrier()
|
||||
|
||||
iter_bar = tqdm(self.dataloader, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
|
||||
for batch in iter_bar:
|
||||
iter_bar = trange(self.num_steps_epoch, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
|
||||
for __ in range(self.num_steps_epoch):
|
||||
batch = self.get_batch()
|
||||
if self.params.n_gpu > 0:
|
||||
batch = tuple(t.to(f'cuda:{self.params.local_rank}') for t in batch)
|
||||
token_ids, attn_mask, mlm_labels = self.prepare_batch(batch=batch)
|
||||
|
||||
if self.mlm:
|
||||
token_ids, attn_mask, lm_labels = self.prepare_batch_mlm(batch=batch)
|
||||
else:
|
||||
token_ids, attn_mask, lm_labels = self.prepare_batch_clm(batch=batch)
|
||||
self.step(input_ids=token_ids, attention_mask=attn_mask, lm_labels=lm_labels)
|
||||
self.step(input_ids=token_ids, attention_mask=attn_mask, mlm_labels=mlm_labels)
|
||||
|
||||
iter_bar.update()
|
||||
iter_bar.set_postfix({'Last_loss': f'{self.last_loss:.2f}',
|
||||
@@ -351,7 +317,7 @@ class Distiller:
|
||||
def step(self,
|
||||
input_ids: torch.tensor,
|
||||
attention_mask: torch.tensor,
|
||||
lm_labels: torch.tensor):
|
||||
mlm_labels: torch.tensor):
|
||||
"""
|
||||
One optimization step: forward of student AND teacher, backward on the loss (for gradient accumulation),
|
||||
and possibly a parameter update (depending on the gradient accumulation).
|
||||
@@ -360,22 +326,17 @@ class Distiller:
|
||||
------
|
||||
input_ids: `torch.tensor(bs, seq_length)` - The token ids.
|
||||
attention_mask: `torch.tensor(bs, seq_length)` - The attention mask for self attention.
|
||||
lm_labels: `torch.tensor(bs, seq_length)` - The language modeling labels (mlm labels for MLM and clm labels for CLM).
|
||||
mlm_labels: `torch.tensor(bs, seq_length)` - The masked language modeling labels.
|
||||
"""
|
||||
if self.mlm:
|
||||
s_logits, s_hidden_states = self.student(input_ids=input_ids, attention_mask=attention_mask) # (bs, seq_length, voc_size)
|
||||
with torch.no_grad():
|
||||
t_logits, t_hidden_states = self.teacher(input_ids=input_ids, attention_mask=attention_mask) # (bs, seq_length, voc_size)
|
||||
else:
|
||||
s_logits, _, s_hidden_states = self.student(input_ids=input_ids, attention_mask=None) # (bs, seq_length, voc_size)
|
||||
with torch.no_grad():
|
||||
t_logits, _, t_hidden_states = self.teacher(input_ids=input_ids, attention_mask=None) # (bs, seq_length, voc_size)
|
||||
s_logits, s_hidden_states = self.student(input_ids=input_ids, attention_mask=attention_mask) # (bs, seq_length, voc_size)
|
||||
with torch.no_grad():
|
||||
t_logits, t_hidden_states = self.teacher(input_ids=input_ids, attention_mask=attention_mask) # (bs, seq_length, voc_size)
|
||||
assert s_logits.size() == t_logits.size()
|
||||
|
||||
#https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
|
||||
#https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
|
||||
if self.params.restrict_ce_to_mask:
|
||||
mask = (lm_labels>-1).unsqueeze(-1).expand_as(s_logits) # (bs, seq_lenth, voc_size)
|
||||
mask = (mlm_labels>-1).unsqueeze(-1).expand_as(s_logits) # (bs, seq_lenth, voc_size)
|
||||
else:
|
||||
mask = attention_mask.unsqueeze(-1).expand_as(s_logits) # (bs, seq_lenth, voc_size)
|
||||
s_logits_slct = torch.masked_select(s_logits, mask) # (bs * seq_length * voc_size) modulo the 1s in mask
|
||||
@@ -387,20 +348,13 @@ class Distiller:
|
||||
loss_ce = self.ce_loss_fct(F.log_softmax(s_logits_slct/self.temperature, dim=-1),
|
||||
F.softmax(t_logits_slct/self.temperature, dim=-1)) * (self.temperature)**2
|
||||
loss = self.alpha_ce*loss_ce
|
||||
|
||||
if self.alpha_mlm > 0.:
|
||||
loss_mlm = self.lm_loss_fct(s_logits.view(-1, s_logits.size(-1)), lm_labels.view(-1))
|
||||
loss_mlm = self.mlm_loss_fct(s_logits.view(-1, s_logits.size(-1)), mlm_labels.view(-1))
|
||||
loss += self.alpha_mlm * loss_mlm
|
||||
if self.alpha_clm > 0.:
|
||||
shift_logits = s_logits[..., :-1, :].contiguous()
|
||||
shift_labels = lm_labels[..., 1:].contiguous()
|
||||
loss_clm = self.lm_loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
|
||||
shift_labels.view(-1))
|
||||
loss += self.alpha_clm * loss_clm
|
||||
|
||||
if self.alpha_mse > 0.:
|
||||
loss_mse = self.mse_loss_fct(s_logits_slct, t_logits_slct)/s_logits_slct.size(0) # Reproducing batchmean reduction
|
||||
loss += self.alpha_mse * loss_mse
|
||||
|
||||
if self.alpha_cos > 0.:
|
||||
s_hidden_states = s_hidden_states[-1] # (bs, seq_length, dim)
|
||||
t_hidden_states = t_hidden_states[-1] # (bs, seq_length, dim)
|
||||
@@ -422,8 +376,6 @@ class Distiller:
|
||||
self.last_loss_ce = loss_ce.item()
|
||||
if self.alpha_mlm > 0.:
|
||||
self.last_loss_mlm = loss_mlm.item()
|
||||
if self.alpha_clm > 0.:
|
||||
self.last_loss_clm = loss_clm.item()
|
||||
if self.alpha_mse > 0.:
|
||||
self.last_loss_mse = loss_mse.item()
|
||||
if self.alpha_cos > 0.:
|
||||
@@ -500,8 +452,6 @@ class Distiller:
|
||||
self.tensorboard.add_scalar(tag="losses/loss_ce", scalar_value=self.last_loss_ce, global_step=self.n_total_iter)
|
||||
if self.alpha_mlm > 0.:
|
||||
self.tensorboard.add_scalar(tag="losses/loss_mlm", scalar_value=self.last_loss_mlm, global_step=self.n_total_iter)
|
||||
if self.alpha_clm > 0.:
|
||||
self.tensorboard.add_scalar(tag="losses/loss_clm", scalar_value=self.last_loss_clm, global_step=self.n_total_iter)
|
||||
if self.alpha_mse > 0.:
|
||||
self.tensorboard.add_scalar(tag="losses/loss_mse", scalar_value=self.last_loss_mse, global_step=self.n_total_iter)
|
||||
if self.alpha_cos > 0.:
|
||||
|
||||
@@ -1,105 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Adapted from PyTorch Vision (https://github.com/pytorch/vision/blob/master/references/detection/group_by_aspect_ratio.py)
|
||||
"""
|
||||
import bisect
|
||||
import copy
|
||||
from collections import defaultdict
|
||||
import numpy as np
|
||||
|
||||
from torch.utils.data.sampler import BatchSampler, Sampler
|
||||
|
||||
from utils import logger
|
||||
|
||||
def _quantize(x, bins):
|
||||
bins = copy.deepcopy(bins)
|
||||
bins = sorted(bins)
|
||||
quantized = list(map(lambda y: bisect.bisect_right(bins, y), x))
|
||||
return quantized
|
||||
|
||||
def create_lengths_groups(lengths, k=0):
|
||||
bins = np.arange(start=3, stop=k, step=4).tolist() if k > 0 else [10]
|
||||
groups = _quantize(lengths, bins)
|
||||
# count number of elements per group
|
||||
counts = np.unique(groups, return_counts=True)[1]
|
||||
fbins = [0] + bins + [np.inf]
|
||||
logger.info("Using {} as bins for aspect lengths quantization".format(fbins))
|
||||
logger.info("Count of instances per bin: {}".format(counts))
|
||||
return groups
|
||||
|
||||
class GroupedBatchSampler(BatchSampler):
|
||||
"""
|
||||
Wraps another sampler to yield a mini-batch of indices.
|
||||
It enforces that the batch only contain elements from the same group.
|
||||
It also tries to provide mini-batches which follows an ordering which is
|
||||
as close as possible to the ordering from the original sampler.
|
||||
Arguments:
|
||||
sampler (Sampler): Base sampler.
|
||||
group_ids (list[int]): If the sampler produces indices in range [0, N),
|
||||
`group_ids` must be a list of `N` ints which contains the group id of each sample.
|
||||
The group ids must be a continuous set of integers starting from
|
||||
0, i.e. they must be in the range [0, num_groups).
|
||||
batch_size (int): Size of mini-batch.
|
||||
"""
|
||||
def __init__(self, sampler, group_ids, batch_size):
|
||||
if not isinstance(sampler, Sampler):
|
||||
raise ValueError(
|
||||
"sampler should be an instance of "
|
||||
"torch.utils.data.Sampler, but got sampler={}".format(sampler)
|
||||
)
|
||||
self.sampler = sampler
|
||||
self.group_ids = group_ids
|
||||
self.batch_size = batch_size
|
||||
|
||||
def __iter__(self):
|
||||
buffer_per_group = defaultdict(list)
|
||||
samples_per_group = defaultdict(list)
|
||||
|
||||
num_batches = 0
|
||||
for idx in self.sampler:
|
||||
group_id = self.group_ids[idx]
|
||||
buffer_per_group[group_id].append(idx)
|
||||
samples_per_group[group_id].append(idx)
|
||||
if len(buffer_per_group[group_id]) == self.batch_size:
|
||||
yield buffer_per_group[group_id] #TODO
|
||||
num_batches += 1
|
||||
del buffer_per_group[group_id]
|
||||
assert len(buffer_per_group[group_id]) < self.batch_size
|
||||
|
||||
# now we have run out of elements that satisfy
|
||||
# the group criteria, let's return the remaining
|
||||
# elements so that the size of the sampler is
|
||||
# deterministic
|
||||
expected_num_batches = len(self)
|
||||
num_remaining = expected_num_batches - num_batches
|
||||
if num_remaining > 0:
|
||||
# for the remaining batches, group the batches by similar lengths
|
||||
batch_idx = []
|
||||
for group_id, idxs in sorted(buffer_per_group.items(), key=lambda x: x[0]):
|
||||
batch_idx.extend(idxs)
|
||||
if len(batch_idx) >= self.batch_size:
|
||||
yield batch_idx[:self.batch_size]
|
||||
batch_idx = batch_idx[self.batch_size:]
|
||||
num_remaining -= 1
|
||||
if len(batch_idx) > 0:
|
||||
yield batch_idx
|
||||
num_remaining -= 1
|
||||
assert num_remaining == 0
|
||||
|
||||
def __len__(self):
|
||||
"""
|
||||
Return the number of mini-batches rather than the number of samples.
|
||||
"""
|
||||
return (len(self.sampler) + self.batch_size - 1) // self.batch_size
|
||||
@@ -3,4 +3,4 @@ tensorboard>=1.14.0
|
||||
tensorboardX==1.8
|
||||
psutil==5.6.3
|
||||
scipy==1.3.1
|
||||
transformers
|
||||
pytorch_transformers==1.2.0
|
||||
|
||||
@@ -1,600 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" This is the exact same script as `examples/run_squad.py` (as of 2019, October 4th) with an additional and optional step of distillation."""
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
import glob
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
import torch.nn.functional as F
|
||||
import torch.nn as nn
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
BertForQuestionAnswering, BertTokenizer,
|
||||
XLMConfig, XLMForQuestionAnswering,
|
||||
XLMTokenizer, XLNetConfig,
|
||||
XLNetForQuestionAnswering,
|
||||
XLNetTokenizer,
|
||||
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
|
||||
from ..utils_squad import (read_squad_examples, convert_examples_to_features,
|
||||
RawResult, write_predictions,
|
||||
RawResultExtended, write_predictions_extended)
|
||||
|
||||
# The follwing import is the official SQuAD evaluation script (2.0).
|
||||
# You can remove it from the dependencies if you are using this script outside of the library
|
||||
# We've added it here for automated tests (see examples/test_examples.py file)
|
||||
from ..utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
|
||||
for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
|
||||
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
|
||||
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
def set_seed(args):
|
||||
random.seed(args.seed)
|
||||
np.random.seed(args.seed)
|
||||
torch.manual_seed(args.seed)
|
||||
if args.n_gpu > 0:
|
||||
torch.cuda.manual_seed_all(args.seed)
|
||||
|
||||
def to_list(tensor):
|
||||
return tensor.detach().cpu().tolist()
|
||||
|
||||
def train(args, train_dataset, model, tokenizer, teacher=None):
|
||||
""" Train the model """
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer = SummaryWriter()
|
||||
|
||||
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
|
||||
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
|
||||
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
|
||||
|
||||
if args.max_steps > 0:
|
||||
t_total = args.max_steps
|
||||
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
|
||||
else:
|
||||
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
|
||||
|
||||
# Prepare optimizer and schedule (linear warmup and decay)
|
||||
no_decay = ['bias', 'LayerNorm.weight']
|
||||
optimizer_grouped_parameters = [
|
||||
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
||||
|
||||
# multi-gpu training (should be after apex fp16 initialization)
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Distributed training (should be after apex fp16 initialization)
|
||||
if args.local_rank != -1:
|
||||
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
|
||||
output_device=args.local_rank,
|
||||
find_unused_parameters=True)
|
||||
|
||||
# Train!
|
||||
logger.info("***** Running training *****")
|
||||
logger.info(" Num examples = %d", len(train_dataset))
|
||||
logger.info(" Num Epochs = %d", args.num_train_epochs)
|
||||
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
|
||||
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
||||
args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
|
||||
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 0
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
model.zero_grad()
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
model.train()
|
||||
if teacher is not None:
|
||||
teacher.eval()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'start_positions': batch[3],
|
||||
'end_positions': batch[4]}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[5],
|
||||
'p_mask': batch[6]})
|
||||
outputs = model(**inputs)
|
||||
loss, start_logits_stu, end_logits_stu = outputs
|
||||
|
||||
# Distillation loss
|
||||
if teacher is not None:
|
||||
if 'token_type_ids' not in inputs:
|
||||
inputs['token_type_ids'] = None if args.teacher_type == 'xlm' else batch[2]
|
||||
with torch.no_grad():
|
||||
start_logits_tea, end_logits_tea = teacher(input_ids=inputs['input_ids'],
|
||||
token_type_ids=inputs['token_type_ids'],
|
||||
attention_mask=inputs['attention_mask'])
|
||||
assert start_logits_tea.size() == start_logits_stu.size()
|
||||
assert end_logits_tea.size() == end_logits_stu.size()
|
||||
|
||||
loss_fct = nn.KLDivLoss(reduction='batchmean')
|
||||
loss_start = loss_fct(F.log_softmax(start_logits_stu/args.temperature, dim=-1),
|
||||
F.softmax(start_logits_tea/args.temperature, dim=-1)) * (args.temperature**2)
|
||||
loss_end = loss_fct(F.log_softmax(end_logits_stu/args.temperature, dim=-1),
|
||||
F.softmax(end_logits_tea/args.temperature, dim=-1)) * (args.temperature**2)
|
||||
loss_ce = (loss_start + loss_end)/2.
|
||||
|
||||
loss = args.alpha_ce*loss_ce + args.alpha_squad*loss
|
||||
|
||||
if args.n_gpu > 1:
|
||||
loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
|
||||
if args.gradient_accumulation_steps > 1:
|
||||
loss = loss / args.gradient_accumulation_steps
|
||||
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
optimizer.step()
|
||||
scheduler.step() # Update learning rate schedule
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results = evaluate(args, model, tokenizer)
|
||||
for key, value in results.items():
|
||||
tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
|
||||
tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
|
||||
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(output_dir)
|
||||
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
|
||||
logger.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
epoch_iterator.close()
|
||||
break
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
train_iterator.close()
|
||||
break
|
||||
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer.close()
|
||||
|
||||
return global_step, tr_loss / global_step
|
||||
|
||||
|
||||
def evaluate(args, model, tokenizer, prefix=""):
|
||||
dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
|
||||
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
|
||||
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(dataset))
|
||||
logger.info(" Batch size = %d", args.eval_batch_size)
|
||||
all_results = []
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
model.eval()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
with torch.no_grad():
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1]
|
||||
}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
|
||||
example_indices = batch[3]
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[4],
|
||||
'p_mask': batch[5]})
|
||||
outputs = model(**inputs)
|
||||
|
||||
for i, example_index in enumerate(example_indices):
|
||||
eval_feature = features[example_index.item()]
|
||||
unique_id = int(eval_feature.unique_id)
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
result = RawResultExtended(unique_id = unique_id,
|
||||
start_top_log_probs = to_list(outputs[0][i]),
|
||||
start_top_index = to_list(outputs[1][i]),
|
||||
end_top_log_probs = to_list(outputs[2][i]),
|
||||
end_top_index = to_list(outputs[3][i]),
|
||||
cls_logits = to_list(outputs[4][i]))
|
||||
else:
|
||||
result = RawResult(unique_id = unique_id,
|
||||
start_logits = to_list(outputs[0][i]),
|
||||
end_logits = to_list(outputs[1][i]))
|
||||
all_results.append(result)
|
||||
|
||||
# Compute predictions
|
||||
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
|
||||
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
|
||||
if args.version_2_with_negative:
|
||||
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
|
||||
else:
|
||||
output_null_log_odds_file = None
|
||||
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
write_predictions_extended(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file, args.predict_file,
|
||||
model.config.start_n_top, model.config.end_n_top,
|
||||
args.version_2_with_negative, tokenizer, args.verbose_logging)
|
||||
else:
|
||||
write_predictions(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, args.do_lower_case, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
|
||||
args.version_2_with_negative, args.null_score_diff_threshold)
|
||||
|
||||
# Evaluate with the official SQuAD script
|
||||
evaluate_options = EVAL_OPTS(data_file=args.predict_file,
|
||||
pred_file=output_prediction_file,
|
||||
na_prob_file=output_null_log_odds_file)
|
||||
results = evaluate_on_squad(evaluate_options)
|
||||
return results
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
|
||||
if args.local_rank not in [-1, 0] and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Load data features from cache or dataset file
|
||||
input_file = args.predict_file if evaluate else args.train_file
|
||||
cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
|
||||
'dev' if evaluate else 'train',
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
logger.info("Creating features from dataset file at %s", input_file)
|
||||
examples = read_squad_examples(input_file=input_file,
|
||||
is_training=not evaluate,
|
||||
version_2_with_negative=args.version_2_with_negative)
|
||||
features = convert_examples_to_features(examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=args.max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=args.max_query_length,
|
||||
is_training=not evaluate)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
|
||||
if args.local_rank == 0 and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Convert to Tensors and build dataset
|
||||
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
|
||||
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
|
||||
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
|
||||
all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
|
||||
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
|
||||
if evaluate:
|
||||
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_example_index, all_cls_index, all_p_mask)
|
||||
else:
|
||||
all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
|
||||
all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_start_positions, all_end_positions,
|
||||
all_cls_index, all_p_mask)
|
||||
|
||||
if output_examples:
|
||||
return dataset, examples, features
|
||||
return dataset
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
## Required parameters
|
||||
parser.add_argument("--train_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for training. E.g., train-v1.1.json")
|
||||
parser.add_argument("--predict_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
|
||||
parser.add_argument("--model_type", default=None, type=str, required=True,
|
||||
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
|
||||
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
parser.add_argument("--output_dir", default=None, type=str, required=True,
|
||||
help="The output directory where the model checkpoints and predictions will be written.")
|
||||
|
||||
# Distillation parameters (optional)
|
||||
parser.add_argument('--teacher_type', default=None, type=str,
|
||||
help="Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for distillation.")
|
||||
parser.add_argument('--teacher_name_or_path', default=None, type=str,
|
||||
help="Path to the already SQuAD fine-tuned teacher model. Only for distillation.")
|
||||
parser.add_argument('--alpha_ce', default=0.5, type=float,
|
||||
help="Distillation loss linear weight. Only for distillation.")
|
||||
parser.add_argument('--alpha_squad', default=0.5, type=float,
|
||||
help="True SQuAD loss linear weight. Only for distillation.")
|
||||
parser.add_argument('--temperature', default=2.0, type=float,
|
||||
help="Distillation temperature. Only for distillation.")
|
||||
|
||||
## Other parameters
|
||||
parser.add_argument("--config_name", default="", type=str,
|
||||
help="Pretrained config name or path if not the same as model_name")
|
||||
parser.add_argument("--tokenizer_name", default="", type=str,
|
||||
help="Pretrained tokenizer name or path if not the same as model_name")
|
||||
parser.add_argument("--cache_dir", default="", type=str,
|
||||
help="Where do you want to store the pre-trained models downloaded from s3")
|
||||
|
||||
parser.add_argument('--version_2_with_negative', action='store_true',
|
||||
help='If true, the SQuAD examples contain some that do not have an answer.')
|
||||
parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
|
||||
help="If null_score - best_non_null is greater than the threshold predict null.")
|
||||
|
||||
parser.add_argument("--max_seq_length", default=384, type=int,
|
||||
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
|
||||
"longer than this will be truncated, and sequences shorter than this will be padded.")
|
||||
parser.add_argument("--doc_stride", default=128, type=int,
|
||||
help="When splitting up a long document into chunks, how much stride to take between chunks.")
|
||||
parser.add_argument("--max_query_length", default=64, type=int,
|
||||
help="The maximum number of tokens for the question. Questions longer than this will "
|
||||
"be truncated to this length.")
|
||||
parser.add_argument("--do_train", action='store_true',
|
||||
help="Whether to run training.")
|
||||
parser.add_argument("--do_eval", action='store_true',
|
||||
help="Whether to run eval on the dev set.")
|
||||
parser.add_argument("--evaluate_during_training", action='store_true',
|
||||
help="Rul evaluation during training at each logging step.")
|
||||
parser.add_argument("--do_lower_case", action='store_true',
|
||||
help="Set this flag if you are using an uncased model.")
|
||||
|
||||
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for training.")
|
||||
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for evaluation.")
|
||||
parser.add_argument("--learning_rate", default=5e-5, type=float,
|
||||
help="The initial learning rate for Adam.")
|
||||
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight deay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
help="Max gradient norm.")
|
||||
parser.add_argument("--num_train_epochs", default=3.0, type=float,
|
||||
help="Total number of training epochs to perform.")
|
||||
parser.add_argument("--max_steps", default=-1, type=int,
|
||||
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
|
||||
parser.add_argument("--warmup_steps", default=0, type=int,
|
||||
help="Linear warmup over warmup_steps.")
|
||||
parser.add_argument("--n_best_size", default=20, type=int,
|
||||
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
|
||||
parser.add_argument("--max_answer_length", default=30, type=int,
|
||||
help="The maximum length of an answer that can be generated. This is needed because the start "
|
||||
"and end predictions are not conditioned on one another.")
|
||||
parser.add_argument("--verbose_logging", action='store_true',
|
||||
help="If true, all of the warnings related to data processing will be printed. "
|
||||
"A number of warnings are expected for a normal SQuAD evaluation.")
|
||||
|
||||
parser.add_argument('--logging_steps', type=int, default=50,
|
||||
help="Log every X updates steps.")
|
||||
parser.add_argument('--save_steps', type=int, default=50,
|
||||
help="Save checkpoint every X updates steps.")
|
||||
parser.add_argument("--eval_all_checkpoints", action='store_true',
|
||||
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
|
||||
parser.add_argument("--no_cuda", action='store_true',
|
||||
help="Whether not to use CUDA when available")
|
||||
parser.add_argument('--overwrite_output_dir', action='store_true',
|
||||
help="Overwrite the content of the output directory")
|
||||
parser.add_argument('--overwrite_cache', action='store_true',
|
||||
help="Overwrite the cached training and evaluation sets")
|
||||
parser.add_argument('--seed', type=int, default=42,
|
||||
help="random seed for initialization")
|
||||
|
||||
parser.add_argument("--local_rank", type=int, default=-1,
|
||||
help="local_rank for distributed training on gpus")
|
||||
parser.add_argument('--fp16', action='store_true',
|
||||
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
|
||||
parser.add_argument('--fp16_opt_level', type=str, default='O1',
|
||||
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
|
||||
"See details at https://nvidia.github.io/apex/amp.html")
|
||||
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
|
||||
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
|
||||
args = parser.parse_args()
|
||||
|
||||
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
|
||||
raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
|
||||
|
||||
# Setup distant debugging if needed
|
||||
if args.server_ip and args.server_port:
|
||||
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
|
||||
import ptvsd
|
||||
print("Waiting for debugger attach")
|
||||
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
|
||||
ptvsd.wait_for_attach()
|
||||
|
||||
# Setup CUDA, GPU & distributed training
|
||||
if args.local_rank == -1 or args.no_cuda:
|
||||
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
args.n_gpu = torch.cuda.device_count()
|
||||
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
|
||||
torch.cuda.set_device(args.local_rank)
|
||||
device = torch.device("cuda", args.local_rank)
|
||||
torch.distributed.init_process_group(backend='nccl')
|
||||
args.n_gpu = 1
|
||||
args.device = device
|
||||
|
||||
# Setup logging
|
||||
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
|
||||
datefmt = '%m/%d/%Y %H:%M:%S',
|
||||
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
|
||||
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
|
||||
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
|
||||
|
||||
# Set seed
|
||||
set_seed(args)
|
||||
|
||||
# Load pretrained model and tokenizer
|
||||
if args.local_rank not in [-1, 0]:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
|
||||
if args.teacher_type is not None:
|
||||
assert args.teacher_name_or_path is not None
|
||||
assert args.alpha_ce > 0.
|
||||
assert args.alpha_ce + args.alpha_squad > 0.
|
||||
assert args.teacher_type != 'distilbert', "We constraint teachers not to be of type DistilBERT."
|
||||
teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
|
||||
teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
teacher = teacher_model_class.from_pretrained(args.teacher_name_or_path,
|
||||
config=teacher_config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
teacher.to(args.device)
|
||||
else:
|
||||
teacher = None
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
model.to(args.device)
|
||||
|
||||
logger.info("Training/evaluation parameters %s", args)
|
||||
|
||||
# Training
|
||||
if args.do_train:
|
||||
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
|
||||
global_step, tr_loss = train(args, train_dataset, model, tokenizer, teacher=teacher)
|
||||
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
|
||||
|
||||
|
||||
# Save the trained model and the tokenizer
|
||||
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
|
||||
# Create output directory if needed
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
logger.info("Saving model checkpoint to %s", args.output_dir)
|
||||
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
|
||||
# They can then be reloaded using `from_pretrained()`
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(args.output_dir)
|
||||
tokenizer.save_pretrained(args.output_dir)
|
||||
|
||||
# Good practice: save your training arguments together with the trained model
|
||||
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
|
||||
|
||||
# Load a trained model and vocabulary that you have fine-tuned
|
||||
model = model_class.from_pretrained(args.output_dir, cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model.to(args.device)
|
||||
|
||||
|
||||
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
|
||||
results = {}
|
||||
if args.do_eval and args.local_rank in [-1, 0]:
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
|
||||
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
|
||||
for checkpoint in checkpoints:
|
||||
# Reload the model
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
model = model_class.from_pretrained(checkpoint, cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model.to(args.device)
|
||||
|
||||
# Evaluate
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step)
|
||||
|
||||
result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
logger.info("Results: {}".format(results))
|
||||
|
||||
return results
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -13,14 +13,14 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Preprocessing script before distillation.
|
||||
Preprocessing script before training DistilBERT.
|
||||
"""
|
||||
import argparse
|
||||
import pickle
|
||||
import random
|
||||
import time
|
||||
import numpy as np
|
||||
from transformers import BertTokenizer, RobertaTokenizer, GPT2Tokenizer
|
||||
from transformers import BertTokenizer, RobertaTokenizer
|
||||
import logging
|
||||
|
||||
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
|
||||
@@ -32,7 +32,7 @@ def main():
|
||||
parser = argparse.ArgumentParser(description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids).")
|
||||
parser.add_argument('--file_path', type=str, default='data/dump.txt',
|
||||
help='The path to the data.')
|
||||
parser.add_argument('--tokenizer_type', type=str, default='bert', choices=['bert', 'roberta', 'gpt2'])
|
||||
parser.add_argument('--tokenizer_type', type=str, default='bert', choices=['bert', 'roberta'])
|
||||
parser.add_argument('--tokenizer_name', type=str, default='bert-base-uncased',
|
||||
help="The tokenizer to use.")
|
||||
parser.add_argument('--dump_file', type=str, default='data/dump',
|
||||
@@ -43,16 +43,10 @@ def main():
|
||||
logger.info(f'Loading Tokenizer ({args.tokenizer_name})')
|
||||
if args.tokenizer_type == 'bert':
|
||||
tokenizer = BertTokenizer.from_pretrained(args.tokenizer_name)
|
||||
bos = tokenizer.special_tokens_map['cls_token'] # `[CLS]`
|
||||
sep = tokenizer.special_tokens_map['sep_token'] # `[SEP]`
|
||||
elif args.tokenizer_type == 'roberta':
|
||||
tokenizer = RobertaTokenizer.from_pretrained(args.tokenizer_name)
|
||||
bos = tokenizer.special_tokens_map['cls_token'] # `<s>`
|
||||
sep = tokenizer.special_tokens_map['sep_token'] # `</s>`
|
||||
elif args.tokenizer_type == 'gpt2':
|
||||
tokenizer = GPT2Tokenizer.from_pretrained(args.tokenizer_name)
|
||||
bos = tokenizer.special_tokens_map['bos_token'] # `<|endoftext|>`
|
||||
sep = tokenizer.special_tokens_map['eos_token'] # `<|endoftext|>`
|
||||
bos = tokenizer.special_tokens_map['bos_token'] # `[CLS]` for bert, `<s>` for roberta
|
||||
sep = tokenizer.special_tokens_map['sep_token'] # `[SEP]` for bert, `</s>` for roberta
|
||||
|
||||
logger.info(f'Loading text from {args.file_path}')
|
||||
with open(args.file_path, 'r', encoding='utf8') as fp:
|
||||
@@ -68,7 +62,7 @@ def main():
|
||||
start = time.time()
|
||||
for text in data:
|
||||
text = f'{bos} {text.strip()} {sep}'
|
||||
token_ids = tokenizer.encode(text, add_special_tokens=False)
|
||||
token_ids = tokenizer.encode(text)
|
||||
rslt.append(token_ids)
|
||||
|
||||
iter += 1
|
||||
|
||||
@@ -1,89 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2019-present, the HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Preprocessing script before training the distilled model.
|
||||
Specific to RoBERTa -> DistilRoBERTa and GPT2 -> DistilGPT2.
|
||||
"""
|
||||
from transformers import BertForMaskedLM, RobertaForMaskedLM, GPT2LMHeadModel
|
||||
import torch
|
||||
import argparse
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description="Extraction some layers of the full RobertaForMaskedLM or GPT2LMHeadModel for Transfer Learned Distillation")
|
||||
parser.add_argument("--model_type", default="roberta", choices=["roberta", "gpt2"])
|
||||
parser.add_argument("--model_name", default='roberta-large', type=str)
|
||||
parser.add_argument("--dump_checkpoint", default='serialization_dir/tf_roberta_048131723.pth', type=str)
|
||||
parser.add_argument("--vocab_transform", action='store_true')
|
||||
args = parser.parse_args()
|
||||
|
||||
|
||||
if args.model_type == 'roberta':
|
||||
model = RobertaForMaskedLM.from_pretrained(args.model_name)
|
||||
prefix = 'roberta'
|
||||
elif args.model_type == 'gpt2':
|
||||
model = GPT2LMHeadModel.from_pretrained(args.model_name)
|
||||
prefix = 'transformer'
|
||||
|
||||
state_dict = model.state_dict()
|
||||
compressed_sd = {}
|
||||
|
||||
### Embeddings ###
|
||||
if args.model_type == 'gpt2':
|
||||
for param_name in ['wte.weight', 'wpe.weight']:
|
||||
compressed_sd[f'{prefix}.{param_name}'] = state_dict[f'{prefix}.{param_name}']
|
||||
else:
|
||||
for w in ['word_embeddings', 'position_embeddings', 'token_type_embeddings']:
|
||||
param_name = f'{prefix}.embeddings.{w}.weight'
|
||||
compressed_sd[param_name] = state_dict[param_name]
|
||||
for w in ['weight', 'bias']:
|
||||
param_name = f'{prefix}.embeddings.LayerNorm.{w}'
|
||||
compressed_sd[param_name] = state_dict[param_name]
|
||||
|
||||
### Transformer Blocks ###
|
||||
std_idx = 0
|
||||
for teacher_idx in [0, 2, 4, 7, 9, 11]:
|
||||
if args.model_type == 'gpt2':
|
||||
for layer in ['ln_1', 'attn.c_attn', 'attn.c_proj', 'ln_2', 'mlp.c_fc', 'mlp.c_proj']:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'{prefix}.h.{std_idx}.{layer}.{w}'] = \
|
||||
state_dict[f'{prefix}.h.{teacher_idx}.{layer}.{w}']
|
||||
compressed_sd[f'{prefix}.h.{std_idx}.attn.bias'] = state_dict[f'{prefix}.h.{teacher_idx}.attn.bias']
|
||||
else:
|
||||
for layer in ['attention.self.query', 'attention.self.key', 'attention.self.value',
|
||||
'attention.output.dense', 'attention.output.LayerNorm',
|
||||
'intermediate.dense', 'output.dense', 'output.LayerNorm']:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'{prefix}.encoder.layer.{std_idx}.{layer}.{w}'] = \
|
||||
state_dict[f'{prefix}.encoder.layer.{teacher_idx}.{layer}.{w}']
|
||||
std_idx += 1
|
||||
|
||||
### Language Modeling Head ###s
|
||||
if args.model_type == 'roberta':
|
||||
for layer in ['lm_head.decoder.weight', 'lm_head.bias']:
|
||||
compressed_sd[f'{layer}'] = state_dict[f'{layer}']
|
||||
if args.vocab_transform:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'lm_head.dense.{w}'] = state_dict[f'lm_head.dense.{w}']
|
||||
compressed_sd[f'lm_head.layer_norm.{w}'] = state_dict[f'lm_head.layer_norm.{w}']
|
||||
elif args.model_type == 'gpt2':
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'{prefix}.ln_f.{w}'] = state_dict[f'{prefix}.ln_f.{w}']
|
||||
compressed_sd[f'lm_head.weight'] = state_dict[f'lm_head.weight']
|
||||
|
||||
print(f'N layers selected for distillation: {std_idx}')
|
||||
print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
|
||||
|
||||
print(f'Save transfered checkpoint to {args.dump_checkpoint}.')
|
||||
torch.save(compressed_sd, args.dump_checkpoint)
|
||||
+18
-10
@@ -14,7 +14,6 @@
|
||||
# limitations under the License.
|
||||
"""
|
||||
Preprocessing script before training DistilBERT.
|
||||
Specific to BERT -> DistilBERT.
|
||||
"""
|
||||
from transformers import BertForMaskedLM, RobertaForMaskedLM
|
||||
import torch
|
||||
@@ -22,7 +21,7 @@ import argparse
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser(description="Extraction some layers of the full BertForMaskedLM or RObertaForMaskedLM for Transfer Learned Distillation")
|
||||
parser.add_argument("--model_type", default="bert", choices=["bert"])
|
||||
parser.add_argument("--model_type", default="bert", choices=["bert", "roberta"])
|
||||
parser.add_argument("--model_name", default='bert-base-uncased', type=str)
|
||||
parser.add_argument("--dump_checkpoint", default='serialization_dir/tf_bert-base-uncased_0247911.pth', type=str)
|
||||
parser.add_argument("--vocab_transform", action='store_true')
|
||||
@@ -32,8 +31,9 @@ if __name__ == '__main__':
|
||||
if args.model_type == 'bert':
|
||||
model = BertForMaskedLM.from_pretrained(args.model_name)
|
||||
prefix = 'bert'
|
||||
else:
|
||||
raise ValueError(f'args.model_type should be "bert".')
|
||||
elif args.model_type == 'roberta':
|
||||
model = RobertaForMaskedLM.from_pretrained(args.model_name)
|
||||
prefix = 'roberta'
|
||||
|
||||
state_dict = model.state_dict()
|
||||
compressed_sd = {}
|
||||
@@ -68,12 +68,20 @@ if __name__ == '__main__':
|
||||
state_dict[f'{prefix}.encoder.layer.{teacher_idx}.output.LayerNorm.{w}']
|
||||
std_idx += 1
|
||||
|
||||
compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
|
||||
compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
|
||||
if args.vocab_transform:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
|
||||
compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
|
||||
if args.model_type == 'bert':
|
||||
compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
|
||||
compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
|
||||
if args.vocab_transform:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
|
||||
compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
|
||||
elif args.model_type == 'roberta':
|
||||
compressed_sd[f'vocab_projector.weight'] = state_dict[f'lm_head.decoder.weight']
|
||||
compressed_sd[f'vocab_projector.bias'] = state_dict[f'lm_head.bias']
|
||||
if args.vocab_transform:
|
||||
for w in ['weight', 'bias']:
|
||||
compressed_sd[f'vocab_transform.{w}'] = state_dict[f'lm_head.dense.{w}']
|
||||
compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'lm_head.layer_norm.{w}']
|
||||
|
||||
print(f'N layers selected for distillation: {std_idx}')
|
||||
print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
|
||||
@@ -13,7 +13,7 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Preprocessing script before training the distilled model.
|
||||
Preprocessing script before training DistilBERT.
|
||||
"""
|
||||
from collections import Counter
|
||||
import argparse
|
||||
|
||||
+75
-118
@@ -13,8 +13,7 @@
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""
|
||||
Training the distilled model.
|
||||
Supported architectures include: BERT -> DistilBERT, RoBERTa -> DistilRoBERTa, GPT2 -> DistilGPT2.
|
||||
Training DistilBERT.
|
||||
"""
|
||||
import os
|
||||
import argparse
|
||||
@@ -24,96 +23,68 @@ import shutil
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
from transformers import BertConfig, BertForMaskedLM, BertTokenizer
|
||||
from transformers import RobertaConfig, RobertaForMaskedLM, RobertaTokenizer
|
||||
from transformers import DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer
|
||||
from transformers import GPT2Config, GPT2LMHeadModel, GPT2Tokenizer
|
||||
from transformers import BertTokenizer, BertForMaskedLM, RobertaTokenizer, RobertaForMaskedLM
|
||||
from transformers import DistilBertForMaskedLM, DistilBertConfig
|
||||
|
||||
from distiller import Distiller
|
||||
from utils import git_log, logger, init_gpu_params, set_seed
|
||||
from lm_seqs_dataset import LmSeqsDataset
|
||||
from dataset import Dataset
|
||||
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
|
||||
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
|
||||
'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
|
||||
'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer)
|
||||
}
|
||||
|
||||
def sanity_checks(args):
|
||||
"""
|
||||
A bunch of args sanity checks to perform even starting...
|
||||
"""
|
||||
assert (args.mlm and args.alpha_mlm > 0.) or (not args.mlm and args.alpha_mlm == 0.)
|
||||
assert (args.alpha_mlm > 0. and args.alpha_clm == 0.) or (args.alpha_mlm == 0. and args.alpha_clm > 0.)
|
||||
if args.mlm:
|
||||
assert os.path.isfile(args.token_counts)
|
||||
assert (args.student_type in ['roberta', 'distilbert']) and (args.teacher_type in ['roberta', 'bert'])
|
||||
else:
|
||||
assert (args.student_type in ['gpt2']) and (args.teacher_type in ['gpt2'])
|
||||
|
||||
assert args.teacher_type == args.student_type or (args.student_type=='distilbert' and args.teacher_type=='bert')
|
||||
assert os.path.isfile(args.student_config)
|
||||
if args.student_pretrained_weights is not None:
|
||||
assert os.path.isfile(args.student_pretrained_weights)
|
||||
|
||||
if args.freeze_token_type_embds: assert args.student_type in ['roberta']
|
||||
|
||||
assert args.alpha_ce >= 0.
|
||||
assert args.alpha_mlm >= 0.
|
||||
assert args.alpha_clm >= 0.
|
||||
assert args.alpha_mse >= 0.
|
||||
assert args.alpha_cos >= 0.
|
||||
assert args.alpha_ce + args.alpha_mlm + args.alpha_clm + args.alpha_mse + args.alpha_cos > 0.
|
||||
|
||||
def freeze_pos_embeddings(student, args):
|
||||
if args.student_type == 'roberta':
|
||||
student.roberta.embeddings.position_embeddings.weight.requires_grad = False
|
||||
elif args.student_type == 'gpt2':
|
||||
student.transformer.wpe.weight.requires_grad = False
|
||||
|
||||
def freeze_token_type_embeddings(student, args):
|
||||
if args.student_type == 'roberta':
|
||||
student.roberta.embeddings.token_type_embeddings.weight.requires_grad = False
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="Training")
|
||||
parser.add_argument("--force", action='store_true',
|
||||
help="Overwrite dump_path if it already exists.")
|
||||
|
||||
parser.add_argument("--dump_path", type=str, required=True,
|
||||
help="The output directory (log, checkpoints, parameters, etc.)")
|
||||
parser.add_argument("--data_file", type=str, required=True,
|
||||
help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
|
||||
parser.add_argument("--token_counts", type=str, required=True,
|
||||
help="The token counts in the data_file for MLM.")
|
||||
parser.add_argument("--force", action='store_true',
|
||||
help="Overwrite dump_path if it already exists.")
|
||||
|
||||
parser.add_argument("--student_type", type=str, choices=["distilbert", "roberta", "gpt2"], required=True,
|
||||
help="The student type (DistilBERT, RoBERTa).")
|
||||
parser.add_argument("--student_config", type=str, required=True,
|
||||
help="Path to the student configuration.")
|
||||
parser.add_argument("--student_pretrained_weights", default=None, type=str,
|
||||
parser.add_argument("--vocab_size", default=30522, type=int,
|
||||
help="The vocabulary size.")
|
||||
parser.add_argument("--max_position_embeddings", default=512, type=int,
|
||||
help="Maximum sequence length we can model (including [CLS] and [SEP]).")
|
||||
parser.add_argument("--sinusoidal_pos_embds", action='store_false',
|
||||
help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
|
||||
parser.add_argument("--n_layers", default=6, type=int,
|
||||
help="Number of Transformer blocks.")
|
||||
parser.add_argument("--n_heads", default=12, type=int,
|
||||
help="Number of heads in the self-attention module.")
|
||||
parser.add_argument("--dim", default=768, type=int,
|
||||
help="Dimension through the network. Must be divisible by n_heads")
|
||||
parser.add_argument("--hidden_dim", default=3072, type=int,
|
||||
help="Intermediate dimension in the FFN.")
|
||||
parser.add_argument("--dropout", default=0.1, type=float,
|
||||
help="Dropout.")
|
||||
parser.add_argument("--attention_dropout", default=0.1, type=float,
|
||||
help="Dropout in self-attention.")
|
||||
parser.add_argument("--activation", default='gelu', type=str,
|
||||
help="Activation to use in self-attention")
|
||||
parser.add_argument("--tie_weights_", action='store_false',
|
||||
help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
|
||||
|
||||
parser.add_argument("--from_pretrained_weights", default=None, type=str,
|
||||
help="Load student initialization checkpoint.")
|
||||
|
||||
parser.add_argument("--teacher_type", choices=["bert", "roberta", "gpt2"], required=True,
|
||||
parser.add_argument("--from_pretrained_config", default=None, type=str,
|
||||
help="Load student initialization architecture config.")
|
||||
parser.add_argument("--teacher_type", default="bert", choices=["bert", "roberta"],
|
||||
help="Teacher type (BERT, RoBERTa).")
|
||||
parser.add_argument("--teacher_name", type=str, required=True,
|
||||
parser.add_argument("--teacher_name", default="bert-base-uncased", type=str,
|
||||
help="The teacher model.")
|
||||
|
||||
parser.add_argument("--temperature", default=2., type=float,
|
||||
help="Temperature for the softmax temperature.")
|
||||
parser.add_argument("--alpha_ce", default=0.5, type=float,
|
||||
help="Linear weight for the distillation loss. Must be >=0.")
|
||||
parser.add_argument("--alpha_mlm", default=0.0, type=float,
|
||||
help="Linear weight for the MLM loss. Must be >=0. Should be used in coonjunction with `mlm` flag.")
|
||||
parser.add_argument("--alpha_clm", default=0.5, type=float,
|
||||
help="Linear weight for the CLM loss. Must be >=0.")
|
||||
parser.add_argument("--alpha_mlm", default=0.5, type=float,
|
||||
help="Linear weight for the MLM loss. Must be >=0.")
|
||||
parser.add_argument("--alpha_mse", default=0.0, type=float,
|
||||
help="Linear weight of the MSE loss. Must be >=0.")
|
||||
parser.add_argument("--alpha_cos", default=0.0, type=float,
|
||||
help="Linear weight of the cosine embedding loss. Must be >=0.")
|
||||
|
||||
parser.add_argument("--mlm", action="store_true",
|
||||
help="The LM step: MLM or CLM. If `mlm` is True, the MLM is used over CLM.")
|
||||
parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
|
||||
help="Proportion of tokens for which we need to make a prediction.")
|
||||
parser.add_argument("--word_mask", default=0.8, type=float,
|
||||
@@ -124,20 +95,17 @@ def main():
|
||||
help="Proportion of tokens to randomly replace.")
|
||||
parser.add_argument("--mlm_smoothing", default=0.7, type=float,
|
||||
help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
|
||||
parser.add_argument("--token_counts", type=str,
|
||||
help="The token counts in the data_file for MLM.")
|
||||
|
||||
parser.add_argument("--restrict_ce_to_mask", action='store_true',
|
||||
help="If true, compute the distilation loss only the [MLM] prediction distribution.")
|
||||
parser.add_argument("--freeze_pos_embs", action="store_true",
|
||||
help="Freeze positional embeddings during distillation. For student_type in ['roberta', 'gpt2'] only.")
|
||||
parser.add_argument("--freeze_token_type_embds", action="store_true",
|
||||
help="Freeze token type embeddings during distillation if existent. For student_type in ['roberta'] only.")
|
||||
|
||||
parser.add_argument("--n_epoch", type=int, default=3,
|
||||
help="Number of pass on the whole dataset.")
|
||||
parser.add_argument("--batch_size", type=int, default=5,
|
||||
help="Batch size (for each process).")
|
||||
parser.add_argument("--tokens_per_batch", type=int, default=-1,
|
||||
help="If specified, modify the batches so that they have approximately this number of tokens.")
|
||||
parser.add_argument("--shuffle", action='store_false',
|
||||
help="If true, shuffle the sequence order. Default is true.")
|
||||
parser.add_argument("--group_by_size", action='store_false',
|
||||
help="If true, group sequences that have similar length into the same batch. Default is true.")
|
||||
|
||||
@@ -173,7 +141,6 @@ def main():
|
||||
parser.add_argument("--checkpoint_interval", type=int, default=4000,
|
||||
help="Checkpoint interval.")
|
||||
args = parser.parse_args()
|
||||
sanity_checks(args)
|
||||
|
||||
|
||||
## ARGS ##
|
||||
@@ -197,19 +164,21 @@ def main():
|
||||
with open(os.path.join(args.dump_path, 'parameters.json'), 'w') as f:
|
||||
json.dump(vars(args), f, indent=4)
|
||||
git_log(args.dump_path)
|
||||
assert (args.from_pretrained_weights is None and args.from_pretrained_config is None) or \
|
||||
(args.from_pretrained_weights is not None and args.from_pretrained_config is not None)
|
||||
|
||||
student_config_class, student_model_class, _ = MODEL_CLASSES[args.student_type]
|
||||
teacher_config_class, teacher_model_class, teacher_tokenizer_class = MODEL_CLASSES[args.teacher_type]
|
||||
|
||||
### TOKENIZER ###
|
||||
tokenizer = teacher_tokenizer_class.from_pretrained(args.teacher_name)
|
||||
if args.teacher_type == 'bert':
|
||||
tokenizer = BertTokenizer.from_pretrained(args.teacher_name)
|
||||
elif args.teacher_type == 'roberta':
|
||||
tokenizer = RobertaTokenizer.from_pretrained(args.teacher_name)
|
||||
special_tok_ids = {}
|
||||
for tok_name, tok_symbol in tokenizer.special_tokens_map.items():
|
||||
idx = tokenizer.all_special_tokens.index(tok_symbol)
|
||||
special_tok_ids[tok_name] = tokenizer.all_special_ids[idx]
|
||||
logger.info(f'Special tokens {special_tok_ids}')
|
||||
args.special_tok_ids = special_tok_ids
|
||||
args.max_model_input_size = tokenizer.max_model_input_sizes[args.teacher_name]
|
||||
|
||||
|
||||
## DATA LOADER ##
|
||||
@@ -218,34 +187,35 @@ def main():
|
||||
data = pickle.load(fp)
|
||||
|
||||
|
||||
if args.mlm:
|
||||
logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
|
||||
with open(args.token_counts, 'rb') as fp:
|
||||
counts = pickle.load(fp)
|
||||
|
||||
token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
|
||||
for idx in special_tok_ids.values():
|
||||
token_probs[idx] = 0. # do not predict special tokens
|
||||
token_probs = torch.from_numpy(token_probs)
|
||||
else:
|
||||
token_probs = None
|
||||
assert os.path.isfile(args.token_counts)
|
||||
logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
|
||||
with open(args.token_counts, 'rb') as fp:
|
||||
counts = pickle.load(fp)
|
||||
assert len(counts) == args.vocab_size
|
||||
token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
|
||||
for idx in special_tok_ids.values():
|
||||
token_probs[idx] = 0. # do not predict special tokens
|
||||
token_probs = torch.from_numpy(token_probs)
|
||||
|
||||
|
||||
train_lm_seq_dataset = LmSeqsDataset(params=args, data=data)
|
||||
train_dataloader = Dataset(params=args, data=data)
|
||||
logger.info(f'Data loader created.')
|
||||
|
||||
|
||||
## STUDENT ##
|
||||
logger.info(f'Loading student config from {args.student_config}')
|
||||
stu_architecture_config = student_config_class.from_pretrained(args.student_config)
|
||||
stu_architecture_config.output_hidden_states = True
|
||||
|
||||
if args.student_pretrained_weights is not None:
|
||||
logger.info(f'Loading pretrained weights from {args.student_pretrained_weights}')
|
||||
student = student_model_class.from_pretrained(args.student_pretrained_weights,
|
||||
config=stu_architecture_config)
|
||||
if args.from_pretrained_weights is not None:
|
||||
assert os.path.isfile(args.from_pretrained_weights)
|
||||
assert os.path.isfile(args.from_pretrained_config)
|
||||
logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
|
||||
logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
|
||||
stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
|
||||
stu_architecture_config.output_hidden_states = True
|
||||
student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
|
||||
config=stu_architecture_config)
|
||||
else:
|
||||
student = student_model_class(stu_architecture_config)
|
||||
args.vocab_size_or_config_json_file = args.vocab_size
|
||||
stu_architecture_config = DistilBertConfig(**vars(args), output_hidden_states=True)
|
||||
student = DistilBertForMaskedLM(stu_architecture_config)
|
||||
|
||||
|
||||
if args.n_gpu > 0:
|
||||
@@ -254,31 +224,18 @@ def main():
|
||||
|
||||
|
||||
## TEACHER ##
|
||||
teacher = teacher_model_class.from_pretrained(args.teacher_name, output_hidden_states=True)
|
||||
if args.teacher_type == 'bert':
|
||||
teacher = BertForMaskedLM.from_pretrained(args.teacher_name, output_hidden_states=True)
|
||||
elif args.teacher_type == 'roberta':
|
||||
teacher = RobertaForMaskedLM.from_pretrained(args.teacher_name, output_hidden_states=True)
|
||||
if args.n_gpu > 0:
|
||||
teacher.to(f'cuda:{args.local_rank}')
|
||||
logger.info(f'Teacher loaded from {args.teacher_name}.')
|
||||
|
||||
|
||||
## FREEZING ##
|
||||
if args.freeze_pos_embs:
|
||||
freeze_pos_embeddings(student, args)
|
||||
if args.freeze_token_type_embds:
|
||||
freeze_token_type_embeddings(student, args)
|
||||
|
||||
|
||||
## SANITY CHECKS ##
|
||||
assert student.config.vocab_size == teacher.config.vocab_size
|
||||
assert student.config.hidden_size == teacher.config.hidden_size
|
||||
assert student.config.max_position_embeddings == teacher.config.max_position_embeddings
|
||||
if args.mlm:
|
||||
assert token_probs.size(0) == stu_architecture_config.vocab_size
|
||||
|
||||
|
||||
## DISTILLER ##
|
||||
torch.cuda.empty_cache()
|
||||
distiller = Distiller(params=args,
|
||||
dataset=train_lm_seq_dataset,
|
||||
dataloader=train_dataloader,
|
||||
token_probs=token_probs,
|
||||
student=student,
|
||||
teacher=teacher)
|
||||
|
||||
@@ -1,15 +0,0 @@
|
||||
{
|
||||
"activation": "gelu",
|
||||
"attention_dropout": 0.1,
|
||||
"dim": 768,
|
||||
"dropout": 0.1,
|
||||
"hidden_dim": 3072,
|
||||
"initializer_range": 0.02,
|
||||
"max_position_embeddings": 512,
|
||||
"n_heads": 12,
|
||||
"n_layers": 6,
|
||||
"sinusoidal_pos_embds": true,
|
||||
"tie_weights_": true,
|
||||
"vocab_size": 30522
|
||||
}
|
||||
|
||||
@@ -1,10 +0,0 @@
|
||||
{
|
||||
"initializer_range": 0.02,
|
||||
"layer_norm_epsilon": 0.00001,
|
||||
"n_ctx": 1024,
|
||||
"n_embd": 768,
|
||||
"n_head": 12,
|
||||
"n_layer": 6,
|
||||
"n_positions": 1024,
|
||||
"vocab_size": 50257
|
||||
}
|
||||
@@ -1,54 +0,0 @@
|
||||
# Plug and Play Language Models: a Simple Approach to Controlled Text Generation
|
||||
|
||||
Authors: [Sumanth Dathathri](https://dathath.github.io/), [Andrea Madotto](https://andreamad8.github.io/), Janice Lan, Jane Hung, Eric Frank, [Piero Molino](https://w4nderlu.st/), [Jason Yosinski](http://yosinski.com/), and [Rosanne Liu](http://www.rosanneliu.com/)
|
||||
|
||||
This folder contains the original code used to run the Plug and Play Language Model (PPLM).
|
||||
|
||||
Paper link: https://arxiv.org/abs/1912.02164
|
||||
|
||||
Blog link: https://eng.uber.com/pplm
|
||||
|
||||
Please check out the repo under uber-research for more information: https://github.com/uber-research/PPLM
|
||||
|
||||
|
||||
## Setup
|
||||
|
||||
```bash
|
||||
git clone https://github.com/huggingface/transformers && cd transformers
|
||||
pip install [--editable] .
|
||||
pip install nltk torchtext # additional requirements.
|
||||
cd examples/pplm
|
||||
```
|
||||
|
||||
## PPLM-BoW
|
||||
|
||||
### Example command for bag-of-words control
|
||||
|
||||
```bash
|
||||
python run_pplm.py -B military --cond_text "The potato" --length 50 --gamma 1.5 --num_iterations 3 --num_samples 10 --stepsize 0.03 --window_length 5 --kl_scale 0.01 --gm_scale 0.99 --colorama --sample
|
||||
```
|
||||
|
||||
### Tuning hyperparameters for bag-of-words control
|
||||
|
||||
1. Increase `--stepsize` to intensify topic control, and decrease its value to soften the control. `--stepsize 0` recovers the original uncontrolled GPT-2 model.
|
||||
|
||||
2. If the language being generated is repetitive (For e.g. "science science experiment experiment"), there are several options to consider: </br>
|
||||
a) Reduce the `--stepsize` </br>
|
||||
b) Increase `--kl_scale` (the KL-loss coefficient) or decrease `--gm_scale` (the gm-scaling term) </br>
|
||||
c) Add `--grad-length xx` where xx is an (integer <= length, e.g. `--grad-length 30`).</br>
|
||||
|
||||
|
||||
## PPLM-Discrim
|
||||
|
||||
### Example command for discriminator based sentiment control
|
||||
|
||||
```bash
|
||||
python run_pplm.py -D sentiment --class_label 2 --cond_text "My dog died" --length 50 --gamma 1.0 --num_iterations 10 --num_samples 10 --stepsize 0.04 --kl_scale 0.01 --gm_scale 0.95 --sample
|
||||
```
|
||||
|
||||
### Tuning hyperparameters for discriminator control
|
||||
|
||||
1. Increase `--stepsize` to intensify topic control, and decrease its value to soften the control. `--stepsize 0` recovers the original uncontrolled GPT-2 model.
|
||||
|
||||
2. Use `--class_label 3` for negative, and `--class_label 2` for positive
|
||||
|
||||
Binary file not shown.
|
Before Width: | Height: | Size: 653 KiB |
Binary file not shown.
|
Before Width: | Height: | Size: 664 KiB |
@@ -1,18 +0,0 @@
|
||||
import torch
|
||||
|
||||
class ClassificationHead(torch.nn.Module):
|
||||
"""Classification Head for transformer encoders"""
|
||||
|
||||
def __init__(self, class_size, embed_size):
|
||||
super(ClassificationHead, self).__init__()
|
||||
self.class_size = class_size
|
||||
self.embed_size = embed_size
|
||||
# self.mlp1 = torch.nn.Linear(embed_size, embed_size)
|
||||
# self.mlp2 = (torch.nn.Linear(embed_size, class_size))
|
||||
self.mlp = torch.nn.Linear(embed_size, class_size)
|
||||
|
||||
def forward(self, hidden_state):
|
||||
# hidden_state = F.relu(self.mlp1(hidden_state))
|
||||
# hidden_state = self.mlp2(hidden_state)
|
||||
logits = self.mlp(hidden_state)
|
||||
return logits
|
||||
@@ -1,879 +0,0 @@
|
||||
#! /usr/bin/env python3
|
||||
# coding=utf-8
|
||||
|
||||
#Copyright (c) 2019 Uber Technologies, Inc.
|
||||
#
|
||||
#Licensed under the Apache License, Version 2.0 (the "License");
|
||||
#you may not use this file except in compliance with the License.
|
||||
#You may obtain a copy of the License at
|
||||
#
|
||||
#http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
#Unless required by applicable law or agreed to in writing, software
|
||||
#distributed under the License is distributed on an "AS IS" BASIS,
|
||||
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
#See the License for the specific language governing permissions and
|
||||
#limitations under the License.
|
||||
|
||||
"""
|
||||
Example command with bag of words:
|
||||
python examples/run_pplm.py -B space --cond_text "The president" --length 100 --gamma 1.5 --num_iterations 3 --num_samples 10 --stepsize 0.01 --window_length 5 --kl_scale 0.01 --gm_scale 0.95
|
||||
|
||||
Example command with discriminator:
|
||||
python examples/run_pplm.py -D sentiment --class_label 3 --cond_text "The lake" --length 10 --gamma 1.0 --num_iterations 30 --num_samples 10 --stepsize 0.01 --kl_scale 0.01 --gm_scale 0.95
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import json
|
||||
from operator import add
|
||||
from typing import List, Optional, Tuple, Union
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
from torch.autograd import Variable
|
||||
from tqdm import trange
|
||||
|
||||
from transformers import GPT2Tokenizer
|
||||
from transformers.file_utils import cached_path
|
||||
from transformers.modeling_gpt2 import GPT2LMHeadModel
|
||||
from pplm_classification_head import ClassificationHead
|
||||
|
||||
PPLM_BOW = 1
|
||||
PPLM_DISCRIM = 2
|
||||
PPLM_BOW_DISCRIM = 3
|
||||
SMALL_CONST = 1e-15
|
||||
BIG_CONST = 1e10
|
||||
|
||||
BAG_OF_WORDS_ARCHIVE_MAP = {
|
||||
'legal': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/legal.txt",
|
||||
'military': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/military.txt",
|
||||
'politics': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/politics.txt",
|
||||
'religion': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/religion.txt",
|
||||
'science': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/science.txt",
|
||||
'space': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/space.txt",
|
||||
'technology': "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/bow/technology.txt",
|
||||
}
|
||||
|
||||
DISCRIMINATOR_MODELS_PARAMS = {
|
||||
"clickbait": {
|
||||
"url": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/discriminators/clickbait_classifier_head.pt",
|
||||
"class_size": 2,
|
||||
"embed_size": 1024,
|
||||
"class_vocab": {"non_clickbait": 0, "clickbait": 1},
|
||||
"default_class": 1,
|
||||
"pretrained_model": "gpt2-medium",
|
||||
},
|
||||
"sentiment": {
|
||||
"url": "https://s3.amazonaws.com/models.huggingface.co/bert/pplm/discriminators/SST_classifier_head.pt",
|
||||
"class_size": 5,
|
||||
"embed_size": 1024,
|
||||
"class_vocab": {"very_positive": 2, "very_negative": 3},
|
||||
"default_class": 3,
|
||||
"pretrained_model": "gpt2-medium",
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def to_var(x, requires_grad=False, volatile=False, device='cuda'):
|
||||
if torch.cuda.is_available() and device == 'cuda':
|
||||
x = x.cuda()
|
||||
elif device != 'cuda':
|
||||
x = x.to(device)
|
||||
return Variable(x, requires_grad=requires_grad, volatile=volatile)
|
||||
|
||||
|
||||
def top_k_filter(logits, k, probs=False):
|
||||
"""
|
||||
Masks everything but the k top entries as -infinity (1e10).
|
||||
Used to mask logits such that e^-infinity -> 0 won't contribute to the
|
||||
sum of the denominator.
|
||||
"""
|
||||
if k == 0:
|
||||
return logits
|
||||
else:
|
||||
values = torch.topk(logits, k)[0]
|
||||
batch_mins = values[:, -1].view(-1, 1).expand_as(logits)
|
||||
if probs:
|
||||
return torch.where(logits < batch_mins,
|
||||
torch.ones_like(logits) * 0.0, logits)
|
||||
return torch.where(logits < batch_mins,
|
||||
torch.ones_like(logits) * -BIG_CONST,
|
||||
logits)
|
||||
|
||||
|
||||
def perturb_past(
|
||||
past,
|
||||
model,
|
||||
last,
|
||||
unpert_past=None,
|
||||
unpert_logits=None,
|
||||
accumulated_hidden=None,
|
||||
grad_norms=None,
|
||||
stepsize=0.01,
|
||||
one_hot_bows_vectors=None,
|
||||
classifier=None,
|
||||
class_label=None,
|
||||
loss_type=0,
|
||||
num_iterations=3,
|
||||
horizon_length=1,
|
||||
window_length=0,
|
||||
decay=False,
|
||||
gamma=1.5,
|
||||
kl_scale=0.01,
|
||||
device='cuda',
|
||||
):
|
||||
# Generate inital perturbed past
|
||||
grad_accumulator = [
|
||||
(np.zeros(p.shape).astype("float32"))
|
||||
for p in past
|
||||
]
|
||||
|
||||
if accumulated_hidden is None:
|
||||
accumulated_hidden = 0
|
||||
|
||||
if decay:
|
||||
decay_mask = torch.arange(
|
||||
0.,
|
||||
1.0 + SMALL_CONST,
|
||||
1.0 / (window_length)
|
||||
)[1:]
|
||||
else:
|
||||
decay_mask = 1.0
|
||||
|
||||
# TODO fix this comment (SUMANTH)
|
||||
# Generate a mask is gradient perturbated is based on a past window
|
||||
_, _, _, curr_length, _ = past[0].shape
|
||||
|
||||
if curr_length > window_length and window_length > 0:
|
||||
ones_key_val_shape = (
|
||||
tuple(past[0].shape[:-2])
|
||||
+ tuple([window_length])
|
||||
+ tuple(past[0].shape[-1:])
|
||||
)
|
||||
|
||||
zeros_key_val_shape = (
|
||||
tuple(past[0].shape[:-2])
|
||||
+ tuple([curr_length - window_length])
|
||||
+ tuple(past[0].shape[-1:])
|
||||
)
|
||||
|
||||
ones_mask = torch.ones(ones_key_val_shape)
|
||||
ones_mask = decay_mask * ones_mask.permute(0, 1, 2, 4, 3)
|
||||
ones_mask = ones_mask.permute(0, 1, 2, 4, 3)
|
||||
|
||||
window_mask = torch.cat(
|
||||
(ones_mask, torch.zeros(zeros_key_val_shape)),
|
||||
dim=-2
|
||||
).to(device)
|
||||
else:
|
||||
window_mask = torch.ones_like(past[0]).to(device)
|
||||
|
||||
# accumulate perturbations for num_iterations
|
||||
loss_per_iter = []
|
||||
new_accumulated_hidden = None
|
||||
for i in range(num_iterations):
|
||||
print("Iteration ", i + 1)
|
||||
curr_perturbation = [
|
||||
to_var(torch.from_numpy(p_), requires_grad=True, device=device)
|
||||
for p_ in grad_accumulator
|
||||
]
|
||||
|
||||
# Compute hidden using perturbed past
|
||||
perturbed_past = list(map(add, past, curr_perturbation))
|
||||
_, _, _, curr_length, _ = curr_perturbation[0].shape
|
||||
all_logits, _, all_hidden = model(last, past=perturbed_past)
|
||||
hidden = all_hidden[-1]
|
||||
new_accumulated_hidden = accumulated_hidden + torch.sum(
|
||||
hidden,
|
||||
dim=1
|
||||
).detach()
|
||||
# TODO: Check the layer-norm consistency of this with trained discriminator (Sumanth)
|
||||
logits = all_logits[:, -1, :]
|
||||
probs = F.softmax(logits, dim=-1)
|
||||
|
||||
loss = 0.0
|
||||
loss_list = []
|
||||
if loss_type == PPLM_BOW or loss_type == PPLM_BOW_DISCRIM:
|
||||
for one_hot_bow in one_hot_bows_vectors:
|
||||
bow_logits = torch.mm(probs, torch.t(one_hot_bow))
|
||||
bow_loss = -torch.log(torch.sum(bow_logits))
|
||||
loss += bow_loss
|
||||
loss_list.append(bow_loss)
|
||||
print(" pplm_bow_loss:", loss.data.cpu().numpy())
|
||||
|
||||
if loss_type == 2 or loss_type == 3:
|
||||
ce_loss = torch.nn.CrossEntropyLoss()
|
||||
# TODO why we need to do this assignment and not just using unpert_past? (Sumanth)
|
||||
curr_unpert_past = unpert_past
|
||||
curr_probs = torch.unsqueeze(probs, dim=1)
|
||||
wte = model.resize_token_embeddings()
|
||||
for _ in range(horizon_length):
|
||||
inputs_embeds = torch.matmul(curr_probs, wte.weight.data)
|
||||
_, curr_unpert_past, curr_all_hidden = model(
|
||||
past=curr_unpert_past,
|
||||
inputs_embeds=inputs_embeds
|
||||
)
|
||||
curr_hidden = curr_all_hidden[-1]
|
||||
new_accumulated_hidden = new_accumulated_hidden + torch.sum(
|
||||
curr_hidden, dim=1)
|
||||
|
||||
prediction = classifier(new_accumulated_hidden /
|
||||
(curr_length + 1 + horizon_length))
|
||||
|
||||
label = torch.tensor(prediction.shape[0] * [class_label],
|
||||
device=device,
|
||||
dtype=torch.long)
|
||||
discrim_loss = ce_loss(prediction, label)
|
||||
print(" pplm_discrim_loss:", discrim_loss.data.cpu().numpy())
|
||||
loss += discrim_loss
|
||||
loss_list.append(discrim_loss)
|
||||
|
||||
kl_loss = 0.0
|
||||
if kl_scale > 0.0:
|
||||
unpert_probs = F.softmax(unpert_logits[:, -1, :], dim=-1)
|
||||
unpert_probs = (
|
||||
unpert_probs + SMALL_CONST *
|
||||
(unpert_probs <= SMALL_CONST).float().to(device).detach()
|
||||
)
|
||||
correction = SMALL_CONST * (probs <= SMALL_CONST).float().to(
|
||||
device).detach()
|
||||
corrected_probs = probs + correction.detach()
|
||||
kl_loss = kl_scale * (
|
||||
(corrected_probs * (corrected_probs / unpert_probs).log()).sum()
|
||||
)
|
||||
print(' kl_loss', kl_loss.data.cpu().numpy())
|
||||
loss += kl_loss
|
||||
|
||||
loss_per_iter.append(loss.data.cpu().numpy())
|
||||
print(' pplm_loss', (loss - kl_loss).data.cpu().numpy())
|
||||
|
||||
# compute gradients
|
||||
loss.backward()
|
||||
|
||||
# calculate gradient norms
|
||||
if grad_norms is not None and loss_type == PPLM_BOW:
|
||||
grad_norms = [
|
||||
torch.max(grad_norms[index], torch.norm(p_.grad * window_mask))
|
||||
for index, p_ in enumerate(curr_perturbation)
|
||||
]
|
||||
else:
|
||||
grad_norms = [
|
||||
(torch.norm(p_.grad * window_mask) + SMALL_CONST)
|
||||
for index, p_ in enumerate(curr_perturbation)
|
||||
]
|
||||
|
||||
# normalize gradients
|
||||
grad = [
|
||||
-stepsize *
|
||||
(p_.grad * window_mask / grad_norms[
|
||||
index] ** gamma).data.cpu().numpy()
|
||||
for index, p_ in enumerate(curr_perturbation)
|
||||
]
|
||||
|
||||
# accumulate gradient
|
||||
grad_accumulator = list(map(add, grad, grad_accumulator))
|
||||
|
||||
# reset gradients, just to make sure
|
||||
for p_ in curr_perturbation:
|
||||
p_.grad.data.zero_()
|
||||
|
||||
# removing past from the graph
|
||||
new_past = []
|
||||
for p_ in past:
|
||||
new_past.append(p_.detach())
|
||||
past = new_past
|
||||
|
||||
# apply the accumulated perturbations to the past
|
||||
grad_accumulator = [
|
||||
to_var(torch.from_numpy(p_), requires_grad=True, device=device)
|
||||
for p_ in grad_accumulator
|
||||
]
|
||||
pert_past = list(map(add, past, grad_accumulator))
|
||||
|
||||
return pert_past, new_accumulated_hidden, grad_norms, loss_per_iter
|
||||
|
||||
|
||||
def get_classifier(
|
||||
name: Optional[str], class_label: Union[str, int],
|
||||
device: str
|
||||
) -> Tuple[Optional[ClassificationHead], Optional[int]]:
|
||||
if name is None:
|
||||
return None, None
|
||||
|
||||
params = DISCRIMINATOR_MODELS_PARAMS[name]
|
||||
classifier = ClassificationHead(
|
||||
class_size=params['class_size'],
|
||||
embed_size=params['embed_size']
|
||||
).to(device)
|
||||
if "url" in params:
|
||||
resolved_archive_file = cached_path(params["url"])
|
||||
elif "path" in params:
|
||||
resolved_archive_file = params["path"]
|
||||
else:
|
||||
raise ValueError("Either url or path have to be specified "
|
||||
"in the discriminator model parameters")
|
||||
classifier.load_state_dict(
|
||||
torch.load(resolved_archive_file, map_location=device))
|
||||
classifier.eval()
|
||||
|
||||
if isinstance(class_label, str):
|
||||
if class_label in params["class_vocab"]:
|
||||
label_id = params["class_vocab"][class_label]
|
||||
else:
|
||||
label_id = params["default_class"]
|
||||
print("class_label {} not in class_vocab".format(class_label))
|
||||
print("available values are: {}".format(params["class_vocab"]))
|
||||
print("using default class {}".format(label_id))
|
||||
|
||||
elif isinstance(class_label, int):
|
||||
if class_label in set(params["class_vocab"].values()):
|
||||
label_id = class_label
|
||||
else:
|
||||
label_id = params["default_class"]
|
||||
print("class_label {} not in class_vocab".format(class_label))
|
||||
print("available values are: {}".format(params["class_vocab"]))
|
||||
print("using default class {}".format(label_id))
|
||||
|
||||
else:
|
||||
label_id = params["default_class"]
|
||||
|
||||
return classifier, label_id
|
||||
|
||||
|
||||
def get_bag_of_words_indices(bag_of_words_ids_or_paths: List[str], tokenizer) -> \
|
||||
List[List[List[int]]]:
|
||||
bow_indices = []
|
||||
for id_or_path in bag_of_words_ids_or_paths:
|
||||
if id_or_path in BAG_OF_WORDS_ARCHIVE_MAP:
|
||||
filepath = cached_path(BAG_OF_WORDS_ARCHIVE_MAP[id_or_path])
|
||||
else:
|
||||
filepath = id_or_path
|
||||
with open(filepath, "r") as f:
|
||||
words = f.read().strip().split("\n")
|
||||
bow_indices.append(
|
||||
[tokenizer.encode(word.strip(), add_prefix_space=True) for word in
|
||||
words])
|
||||
return bow_indices
|
||||
|
||||
|
||||
def build_bows_one_hot_vectors(bow_indices, tokenizer, device='cuda'):
|
||||
if bow_indices is None:
|
||||
return None
|
||||
|
||||
one_hot_bows_vectors = []
|
||||
for single_bow in bow_indices:
|
||||
single_bow = list(filter(lambda x: len(x) <= 1, single_bow))
|
||||
single_bow = torch.tensor(single_bow).to(device)
|
||||
num_words = single_bow.shape[0]
|
||||
one_hot_bow = torch.zeros(num_words, tokenizer.vocab_size).to(device)
|
||||
one_hot_bow.scatter_(1, single_bow, 1)
|
||||
one_hot_bows_vectors.append(one_hot_bow)
|
||||
return one_hot_bows_vectors
|
||||
|
||||
|
||||
def full_text_generation(
|
||||
model,
|
||||
tokenizer,
|
||||
context=None,
|
||||
num_samples=1,
|
||||
device="cuda",
|
||||
bag_of_words=None,
|
||||
discrim=None,
|
||||
class_label=None,
|
||||
length=100,
|
||||
stepsize=0.02,
|
||||
temperature=1.0,
|
||||
top_k=10,
|
||||
sample=False,
|
||||
num_iterations=3,
|
||||
grad_length=10000,
|
||||
horizon_length=1,
|
||||
window_length=0,
|
||||
decay=False,
|
||||
gamma=1.5,
|
||||
gm_scale=0.9,
|
||||
kl_scale=0.01,
|
||||
**kwargs
|
||||
):
|
||||
classifier, class_id = get_classifier(
|
||||
discrim,
|
||||
class_label,
|
||||
device
|
||||
)
|
||||
|
||||
bow_indices = []
|
||||
if bag_of_words:
|
||||
bow_indices = get_bag_of_words_indices(bag_of_words.split(";"),
|
||||
tokenizer)
|
||||
|
||||
if bag_of_words and classifier:
|
||||
print("Both PPLM-BoW and PPLM-Discrim are on. This is not optimized.")
|
||||
loss_type = PPLM_BOW_DISCRIM
|
||||
|
||||
elif bag_of_words:
|
||||
loss_type = PPLM_BOW
|
||||
print("Using PPLM-BoW")
|
||||
|
||||
elif classifier is not None:
|
||||
loss_type = PPLM_DISCRIM
|
||||
print("Using PPLM-Discrim")
|
||||
|
||||
else:
|
||||
raise Exception("Specify either a bag of words or a discriminator")
|
||||
|
||||
unpert_gen_tok_text, _, _ = generate_text_pplm(
|
||||
model=model,
|
||||
tokenizer=tokenizer,
|
||||
context=context,
|
||||
device=device,
|
||||
length=length,
|
||||
sample=sample,
|
||||
perturb=False
|
||||
)
|
||||
if device == 'cuda':
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
pert_gen_tok_texts = []
|
||||
discrim_losses = []
|
||||
losses_in_time = []
|
||||
|
||||
for i in range(num_samples):
|
||||
pert_gen_tok_text, discrim_loss, loss_in_time = generate_text_pplm(
|
||||
model=model,
|
||||
tokenizer=tokenizer,
|
||||
context=context,
|
||||
device=device,
|
||||
perturb=True,
|
||||
bow_indices=bow_indices,
|
||||
classifier=classifier,
|
||||
class_label=class_id,
|
||||
loss_type=loss_type,
|
||||
length=length,
|
||||
stepsize=stepsize,
|
||||
temperature=temperature,
|
||||
top_k=top_k,
|
||||
sample=sample,
|
||||
num_iterations=num_iterations,
|
||||
grad_length=grad_length,
|
||||
horizon_length=horizon_length,
|
||||
window_length=window_length,
|
||||
decay=decay,
|
||||
gamma=gamma,
|
||||
gm_scale=gm_scale,
|
||||
kl_scale=kl_scale,
|
||||
)
|
||||
pert_gen_tok_texts.append(pert_gen_tok_text)
|
||||
if classifier is not None:
|
||||
discrim_losses.append(discrim_loss.data.cpu().numpy())
|
||||
losses_in_time.append(loss_in_time)
|
||||
|
||||
if device == 'cuda':
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
return unpert_gen_tok_text, pert_gen_tok_texts, discrim_losses, losses_in_time
|
||||
|
||||
|
||||
def generate_text_pplm(
|
||||
model,
|
||||
tokenizer,
|
||||
context=None,
|
||||
past=None,
|
||||
device="cuda",
|
||||
perturb=True,
|
||||
bow_indices=None,
|
||||
classifier=None,
|
||||
class_label=None,
|
||||
loss_type=0,
|
||||
length=100,
|
||||
stepsize=0.02,
|
||||
temperature=1.0,
|
||||
top_k=10,
|
||||
sample=False,
|
||||
num_iterations=3,
|
||||
grad_length=10000,
|
||||
horizon_length=1,
|
||||
window_length=0,
|
||||
decay=False,
|
||||
gamma=1.5,
|
||||
gm_scale=0.9,
|
||||
kl_scale=0.01,
|
||||
):
|
||||
output_so_far = None
|
||||
if context:
|
||||
context_t = torch.tensor(context, device=device, dtype=torch.long)
|
||||
while len(context_t.shape) < 2:
|
||||
context_t = context_t.unsqueeze(0)
|
||||
output_so_far = context_t
|
||||
|
||||
# collect one hot vectors for bags of words
|
||||
one_hot_bows_vectors = build_bows_one_hot_vectors(bow_indices, tokenizer,
|
||||
device)
|
||||
|
||||
grad_norms = None
|
||||
last = None
|
||||
unpert_discrim_loss = 0
|
||||
loss_in_time = []
|
||||
for i in trange(length, ascii=True):
|
||||
|
||||
# Get past/probs for current output, except for last word
|
||||
# Note that GPT takes 2 inputs: past + current_token
|
||||
|
||||
# run model forward to obtain unperturbed
|
||||
if past is None and output_so_far is not None:
|
||||
last = output_so_far[:, -1:]
|
||||
if output_so_far.shape[1] > 1:
|
||||
_, past, _ = model(output_so_far[:, :-1])
|
||||
|
||||
unpert_logits, unpert_past, unpert_all_hidden = model(output_so_far)
|
||||
unpert_last_hidden = unpert_all_hidden[-1]
|
||||
|
||||
# check if we are abowe grad max length
|
||||
if i >= grad_length:
|
||||
current_stepsize = stepsize * 0
|
||||
else:
|
||||
current_stepsize = stepsize
|
||||
|
||||
# modify the past if necessary
|
||||
if not perturb or num_iterations == 0:
|
||||
pert_past = past
|
||||
|
||||
else:
|
||||
accumulated_hidden = unpert_last_hidden[:, :-1, :]
|
||||
accumulated_hidden = torch.sum(accumulated_hidden, dim=1)
|
||||
|
||||
if past is not None:
|
||||
pert_past, _, grad_norms, loss_this_iter = perturb_past(
|
||||
past,
|
||||
model,
|
||||
last,
|
||||
unpert_past=unpert_past,
|
||||
unpert_logits=unpert_logits,
|
||||
accumulated_hidden=accumulated_hidden,
|
||||
grad_norms=grad_norms,
|
||||
stepsize=current_stepsize,
|
||||
one_hot_bows_vectors=one_hot_bows_vectors,
|
||||
classifier=classifier,
|
||||
class_label=class_label,
|
||||
loss_type=loss_type,
|
||||
num_iterations=num_iterations,
|
||||
horizon_length=horizon_length,
|
||||
window_length=window_length,
|
||||
decay=decay,
|
||||
gamma=gamma,
|
||||
kl_scale=kl_scale,
|
||||
device=device,
|
||||
)
|
||||
loss_in_time.append(loss_this_iter)
|
||||
else:
|
||||
pert_past = past
|
||||
|
||||
pert_logits, past, pert_all_hidden = model(last, past=pert_past)
|
||||
pert_logits = pert_logits[:, -1, :] / temperature # + SMALL_CONST
|
||||
pert_probs = F.softmax(pert_logits, dim=-1)
|
||||
|
||||
if classifier is not None:
|
||||
ce_loss = torch.nn.CrossEntropyLoss()
|
||||
prediction = classifier(torch.mean(unpert_last_hidden, dim=1))
|
||||
label = torch.tensor([class_label], device=device,
|
||||
dtype=torch.long)
|
||||
unpert_discrim_loss = ce_loss(prediction, label)
|
||||
print(
|
||||
"unperturbed discrim loss",
|
||||
unpert_discrim_loss.data.cpu().numpy()
|
||||
)
|
||||
else:
|
||||
unpert_discrim_loss = 0
|
||||
|
||||
# Fuse the modified model and original model
|
||||
if perturb:
|
||||
|
||||
unpert_probs = F.softmax(unpert_logits[:, -1, :], dim=-1)
|
||||
|
||||
pert_probs = ((pert_probs ** gm_scale) * (
|
||||
unpert_probs ** (1 - gm_scale))) # + SMALL_CONST
|
||||
pert_probs = top_k_filter(pert_probs, k=top_k,
|
||||
probs=True) # + SMALL_CONST
|
||||
|
||||
# rescale
|
||||
if torch.sum(pert_probs) <= 1:
|
||||
pert_probs = pert_probs / torch.sum(pert_probs)
|
||||
|
||||
else:
|
||||
pert_logits = top_k_filter(pert_logits, k=top_k) # + SMALL_CONST
|
||||
pert_probs = F.softmax(pert_logits, dim=-1)
|
||||
|
||||
# sample or greedy
|
||||
if sample:
|
||||
last = torch.multinomial(pert_probs, num_samples=1)
|
||||
|
||||
else:
|
||||
_, last = torch.topk(pert_probs, k=1, dim=-1)
|
||||
|
||||
# update context/output_so_far appending the new token
|
||||
output_so_far = (
|
||||
last if output_so_far is None
|
||||
else torch.cat((output_so_far, last), dim=1)
|
||||
)
|
||||
|
||||
print(tokenizer.decode(output_so_far.tolist()[0]))
|
||||
|
||||
return output_so_far, unpert_discrim_loss, loss_in_time
|
||||
|
||||
|
||||
def set_generic_model_params(discrim_weights, discrim_meta):
|
||||
if discrim_weights is None:
|
||||
raise ValueError('When using a generic discriminator, '
|
||||
'discrim_weights need to be specified')
|
||||
if discrim_meta is None:
|
||||
raise ValueError('When using a generic discriminator, '
|
||||
'discrim_meta need to be specified')
|
||||
|
||||
with open(discrim_meta, 'r') as discrim_meta_file:
|
||||
meta = json.load(discrim_meta_file)
|
||||
meta['path'] = discrim_weights
|
||||
DISCRIMINATOR_MODELS_PARAMS['generic'] = meta
|
||||
|
||||
|
||||
def run_pplm_example(
|
||||
pretrained_model="gpt2-medium",
|
||||
cond_text="",
|
||||
uncond=False,
|
||||
num_samples=1,
|
||||
bag_of_words=None,
|
||||
discrim=None,
|
||||
discrim_weights=None,
|
||||
discrim_meta=None,
|
||||
class_label=-1,
|
||||
length=100,
|
||||
stepsize=0.02,
|
||||
temperature=1.0,
|
||||
top_k=10,
|
||||
sample=False,
|
||||
num_iterations=3,
|
||||
grad_length=10000,
|
||||
horizon_length=1,
|
||||
window_length=0,
|
||||
decay=False,
|
||||
gamma=1.5,
|
||||
gm_scale=0.9,
|
||||
kl_scale=0.01,
|
||||
seed=0,
|
||||
no_cuda=False,
|
||||
colorama=False
|
||||
):
|
||||
# set Random seed
|
||||
torch.manual_seed(seed)
|
||||
np.random.seed(seed)
|
||||
|
||||
# set the device
|
||||
device = "cuda" if torch.cuda.is_available() and not no_cuda else "cpu"
|
||||
|
||||
if discrim == 'generic':
|
||||
set_generic_model_params(discrim_weights, discrim_meta)
|
||||
|
||||
if discrim is not None:
|
||||
pretrained_model = DISCRIMINATOR_MODELS_PARAMS[discrim][
|
||||
"pretrained_model"
|
||||
]
|
||||
print("discrim = {}, pretrained_model set "
|
||||
"to discriminator's = {}".format(discrim, pretrained_model))
|
||||
|
||||
# load pretrained model
|
||||
model = GPT2LMHeadModel.from_pretrained(
|
||||
pretrained_model,
|
||||
output_hidden_states=True
|
||||
)
|
||||
model.to(device)
|
||||
model.eval()
|
||||
|
||||
# load tokenizer
|
||||
tokenizer = GPT2Tokenizer.from_pretrained(pretrained_model)
|
||||
|
||||
# Freeze GPT-2 weights
|
||||
for param in model.parameters():
|
||||
param.requires_grad = False
|
||||
|
||||
# figure out conditioning text
|
||||
if uncond:
|
||||
tokenized_cond_text = tokenizer.encode(
|
||||
[tokenizer.bos_token]
|
||||
)
|
||||
else:
|
||||
raw_text = cond_text
|
||||
while not raw_text:
|
||||
print("Did you forget to add `--cond_text`? ")
|
||||
raw_text = input("Model prompt >>> ")
|
||||
tokenized_cond_text = tokenizer.encode(tokenizer.bos_token + raw_text)
|
||||
|
||||
print("= Prefix of sentence =")
|
||||
print(tokenizer.decode(tokenized_cond_text))
|
||||
print()
|
||||
|
||||
# generate unperturbed and perturbed texts
|
||||
|
||||
# full_text_generation returns:
|
||||
# unpert_gen_tok_text, pert_gen_tok_texts, discrim_losses, losses_in_time
|
||||
unpert_gen_tok_text, pert_gen_tok_texts, _, _ = full_text_generation(
|
||||
model=model,
|
||||
tokenizer=tokenizer,
|
||||
context=tokenized_cond_text,
|
||||
device=device,
|
||||
num_samples=num_samples,
|
||||
bag_of_words=bag_of_words,
|
||||
discrim=discrim,
|
||||
class_label=class_label,
|
||||
length=length,
|
||||
stepsize=stepsize,
|
||||
temperature=temperature,
|
||||
top_k=top_k,
|
||||
sample=sample,
|
||||
num_iterations=num_iterations,
|
||||
grad_length=grad_length,
|
||||
horizon_length=horizon_length,
|
||||
window_length=window_length,
|
||||
decay=decay,
|
||||
gamma=gamma,
|
||||
gm_scale=gm_scale,
|
||||
kl_scale=kl_scale,
|
||||
)
|
||||
|
||||
# untokenize unperturbed text
|
||||
unpert_gen_text = tokenizer.decode(unpert_gen_tok_text.tolist()[0])
|
||||
|
||||
print("=" * 80)
|
||||
print("= Unperturbed generated text =")
|
||||
print(unpert_gen_text)
|
||||
print()
|
||||
|
||||
generated_texts = []
|
||||
|
||||
bow_word_ids = set()
|
||||
if bag_of_words and colorama:
|
||||
bow_indices = get_bag_of_words_indices(bag_of_words.split(";"),
|
||||
tokenizer)
|
||||
for single_bow_list in bow_indices:
|
||||
# filtering all words in the list composed of more than 1 token
|
||||
filtered = list(filter(lambda x: len(x) <= 1, single_bow_list))
|
||||
# w[0] because we are sure w has only 1 item because previous fitler
|
||||
bow_word_ids.update(w[0] for w in filtered)
|
||||
|
||||
# iterate through the perturbed texts
|
||||
for i, pert_gen_tok_text in enumerate(pert_gen_tok_texts):
|
||||
try:
|
||||
# untokenize unperturbed text
|
||||
if colorama:
|
||||
import colorama
|
||||
|
||||
pert_gen_text = ''
|
||||
for word_id in pert_gen_tok_text.tolist()[0]:
|
||||
if word_id in bow_word_ids:
|
||||
pert_gen_text += '{}{}{}'.format(
|
||||
colorama.Fore.RED,
|
||||
tokenizer.decode([word_id]),
|
||||
colorama.Style.RESET_ALL
|
||||
)
|
||||
else:
|
||||
pert_gen_text += tokenizer.decode([word_id])
|
||||
else:
|
||||
pert_gen_text = tokenizer.decode(pert_gen_tok_text.tolist()[0])
|
||||
|
||||
print("= Perturbed generated text {} =".format(i + 1))
|
||||
print(pert_gen_text)
|
||||
print()
|
||||
except:
|
||||
pass
|
||||
|
||||
# keep the prefix, perturbed seq, original seq for each index
|
||||
generated_texts.append(
|
||||
(tokenized_cond_text, pert_gen_tok_text, unpert_gen_tok_text)
|
||||
)
|
||||
|
||||
return
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--pretrained_model",
|
||||
"-M",
|
||||
type=str,
|
||||
default="gpt2-medium",
|
||||
help="pretrained model name or path to local checkpoint",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--cond_text", type=str, default="The lake",
|
||||
help="Prefix texts to condition on"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--uncond", action="store_true",
|
||||
help="Generate from end-of-text as prefix"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--num_samples",
|
||||
type=int,
|
||||
default=1,
|
||||
help="Number of samples to generate from the modified latents",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--bag_of_words",
|
||||
"-B",
|
||||
type=str,
|
||||
default=None,
|
||||
help="Bags of words used for PPLM-BoW. "
|
||||
"Either a BOW id (see list in code) or a filepath. "
|
||||
"Multiple BoWs separated by ;",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--discrim",
|
||||
"-D",
|
||||
type=str,
|
||||
default=None,
|
||||
choices=("clickbait", "sentiment", "toxicity", "generic"),
|
||||
help="Discriminator to use",
|
||||
)
|
||||
parser.add_argument('--discrim_weights', type=str, default=None,
|
||||
help='Weights for the generic discriminator')
|
||||
parser.add_argument('--discrim_meta', type=str, default=None,
|
||||
help='Meta information for the generic discriminator')
|
||||
parser.add_argument(
|
||||
"--class_label",
|
||||
type=int,
|
||||
default=-1,
|
||||
help="Class label used for the discriminator",
|
||||
)
|
||||
parser.add_argument("--length", type=int, default=100)
|
||||
parser.add_argument("--stepsize", type=float, default=0.02)
|
||||
parser.add_argument("--temperature", type=float, default=1.0)
|
||||
parser.add_argument("--top_k", type=int, default=10)
|
||||
parser.add_argument(
|
||||
"--sample", action="store_true",
|
||||
help="Generate from end-of-text as prefix"
|
||||
)
|
||||
parser.add_argument("--num_iterations", type=int, default=3)
|
||||
parser.add_argument("--grad_length", type=int, default=10000)
|
||||
parser.add_argument(
|
||||
"--window_length",
|
||||
type=int,
|
||||
default=0,
|
||||
help="Length of past which is being optimized; "
|
||||
"0 corresponds to infinite window length",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--horizon_length",
|
||||
type=int,
|
||||
default=1,
|
||||
help="Length of future to optimize over",
|
||||
)
|
||||
parser.add_argument("--decay", action="store_true",
|
||||
help="whether to decay or not")
|
||||
parser.add_argument("--gamma", type=float, default=1.5)
|
||||
parser.add_argument("--gm_scale", type=float, default=0.9)
|
||||
parser.add_argument("--kl_scale", type=float, default=0.01)
|
||||
parser.add_argument("--seed", type=int, default=0)
|
||||
parser.add_argument("--no_cuda", action="store_true", help="no cuda")
|
||||
parser.add_argument("--colorama", action="store_true",
|
||||
help="colors keywords")
|
||||
|
||||
args = parser.parse_args()
|
||||
run_pplm_example(**vars(args))
|
||||
@@ -1,588 +0,0 @@
|
||||
#! /usr/bin/env python3
|
||||
# coding=utf-8
|
||||
|
||||
#Copyright (c) 2019 Uber Technologies, Inc.
|
||||
#
|
||||
#Licensed under the Apache License, Version 2.0 (the "License");
|
||||
#you may not use this file except in compliance with the License.
|
||||
#You may obtain a copy of the License at
|
||||
#
|
||||
#http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
#Unless required by applicable law or agreed to in writing, software
|
||||
#distributed under the License is distributed on an "AS IS" BASIS,
|
||||
#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
#See the License for the specific language governing permissions and
|
||||
#limitations under the License.
|
||||
|
||||
import argparse
|
||||
import csv
|
||||
import json
|
||||
import math
|
||||
import time
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torch.optim
|
||||
import torch.optim as optim
|
||||
import torch.utils.data as data
|
||||
from nltk.tokenize.treebank import TreebankWordDetokenizer
|
||||
from torchtext import data as torchtext_data
|
||||
from torchtext import datasets
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import GPT2Tokenizer, GPT2LMHeadModel
|
||||
from pplm_classification_head import ClassificationHead
|
||||
|
||||
torch.manual_seed(0)
|
||||
np.random.seed(0)
|
||||
EPSILON = 1e-10
|
||||
example_sentence = "This is incredible! I love it, this is the best chicken I have ever had."
|
||||
max_length_seq = 100
|
||||
|
||||
|
||||
|
||||
|
||||
class Discriminator(torch.nn.Module):
|
||||
"""Transformer encoder followed by a Classification Head"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
class_size,
|
||||
pretrained_model="gpt2-medium",
|
||||
cached_mode=False,
|
||||
device='cpu'
|
||||
):
|
||||
super(Discriminator, self).__init__()
|
||||
self.tokenizer = GPT2Tokenizer.from_pretrained(pretrained_model)
|
||||
self.encoder = GPT2LMHeadModel.from_pretrained(pretrained_model)
|
||||
self.embed_size = self.encoder.transformer.config.hidden_size
|
||||
self.classifier_head = ClassificationHead(
|
||||
class_size=class_size,
|
||||
embed_size=self.embed_size
|
||||
)
|
||||
self.cached_mode = cached_mode
|
||||
self.device = device
|
||||
|
||||
def get_classifier(self):
|
||||
return self.classifier_head
|
||||
|
||||
def train_custom(self):
|
||||
for param in self.encoder.parameters():
|
||||
param.requires_grad = False
|
||||
self.classifier_head.train()
|
||||
|
||||
def avg_representation(self, x):
|
||||
mask = x.ne(0).unsqueeze(2).repeat(
|
||||
1, 1, self.embed_size
|
||||
).float().to(self.device).detach()
|
||||
hidden, _ = self.encoder.transformer(x)
|
||||
masked_hidden = hidden * mask
|
||||
avg_hidden = torch.sum(masked_hidden, dim=1) / (
|
||||
torch.sum(mask, dim=1).detach() + EPSILON
|
||||
)
|
||||
return avg_hidden
|
||||
|
||||
def forward(self, x):
|
||||
if self.cached_mode:
|
||||
avg_hidden = x.to(self.device)
|
||||
else:
|
||||
avg_hidden = self.avg_representation(x.to(self.device))
|
||||
|
||||
logits = self.classifier_head(avg_hidden)
|
||||
probs = F.log_softmax(logits, dim=-1)
|
||||
|
||||
return probs
|
||||
|
||||
|
||||
class Dataset(data.Dataset):
|
||||
def __init__(self, X, y):
|
||||
"""Reads source and target sequences from txt files."""
|
||||
self.X = X
|
||||
self.y = y
|
||||
|
||||
def __len__(self):
|
||||
return len(self.X)
|
||||
|
||||
def __getitem__(self, index):
|
||||
"""Returns one data pair (source and target)."""
|
||||
data = {}
|
||||
data["X"] = self.X[index]
|
||||
data["y"] = self.y[index]
|
||||
return data
|
||||
|
||||
|
||||
def collate_fn(data):
|
||||
def pad_sequences(sequences):
|
||||
lengths = [len(seq) for seq in sequences]
|
||||
|
||||
padded_sequences = torch.zeros(
|
||||
len(sequences),
|
||||
max(lengths)
|
||||
).long() # padding value = 0
|
||||
|
||||
for i, seq in enumerate(sequences):
|
||||
end = lengths[i]
|
||||
padded_sequences[i, :end] = seq[:end]
|
||||
|
||||
return padded_sequences, lengths
|
||||
|
||||
item_info = {}
|
||||
for key in data[0].keys():
|
||||
item_info[key] = [d[key] for d in data]
|
||||
|
||||
x_batch, _ = pad_sequences(item_info["X"])
|
||||
y_batch = torch.tensor(item_info["y"], dtype=torch.long)
|
||||
|
||||
return x_batch, y_batch
|
||||
|
||||
|
||||
def cached_collate_fn(data):
|
||||
item_info = {}
|
||||
for key in data[0].keys():
|
||||
item_info[key] = [d[key] for d in data]
|
||||
|
||||
x_batch = torch.cat(item_info["X"], 0)
|
||||
y_batch = torch.tensor(item_info["y"], dtype=torch.long)
|
||||
|
||||
return x_batch, y_batch
|
||||
|
||||
|
||||
def train_epoch(data_loader, discriminator, optimizer,
|
||||
epoch=0, log_interval=10, device='cpu'):
|
||||
samples_so_far = 0
|
||||
discriminator.train_custom()
|
||||
for batch_idx, (input_t, target_t) in enumerate(data_loader):
|
||||
input_t, target_t = input_t.to(device), target_t.to(device)
|
||||
|
||||
optimizer.zero_grad()
|
||||
|
||||
output_t = discriminator(input_t)
|
||||
loss = F.nll_loss(output_t, target_t)
|
||||
loss.backward(retain_graph=True)
|
||||
optimizer.step()
|
||||
|
||||
samples_so_far += len(input_t)
|
||||
|
||||
if batch_idx % log_interval == 0:
|
||||
print(
|
||||
"Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}".format(
|
||||
epoch + 1,
|
||||
samples_so_far, len(data_loader.dataset),
|
||||
100 * samples_so_far / len(data_loader.dataset), loss.item()
|
||||
)
|
||||
)
|
||||
|
||||
|
||||
def evaluate_performance(data_loader, discriminator, device='cpu'):
|
||||
discriminator.eval()
|
||||
test_loss = 0
|
||||
correct = 0
|
||||
with torch.no_grad():
|
||||
for input_t, target_t in data_loader:
|
||||
input_t, target_t = input_t.to(device), target_t.to(device)
|
||||
output_t = discriminator(input_t)
|
||||
# sum up batch loss
|
||||
test_loss += F.nll_loss(output_t, target_t, reduction="sum").item()
|
||||
# get the index of the max log-probability
|
||||
pred_t = output_t.argmax(dim=1, keepdim=True)
|
||||
correct += pred_t.eq(target_t.view_as(pred_t)).sum().item()
|
||||
|
||||
test_loss /= len(data_loader.dataset)
|
||||
|
||||
print(
|
||||
"Performance on test set: "
|
||||
"Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)".format(
|
||||
test_loss, correct, len(data_loader.dataset),
|
||||
100. * correct / len(data_loader.dataset)
|
||||
)
|
||||
)
|
||||
|
||||
|
||||
def predict(input_sentence, model, classes, cached=False, device='cpu'):
|
||||
input_t = model.tokenizer.encode(input_sentence)
|
||||
input_t = torch.tensor([input_t], dtype=torch.long, device=device)
|
||||
if cached:
|
||||
input_t = model.avg_representation(input_t)
|
||||
|
||||
log_probs = model(input_t).data.cpu().numpy().flatten().tolist()
|
||||
print("Input sentence:", input_sentence)
|
||||
print("Predictions:", ", ".join(
|
||||
"{}: {:.4f}".format(c, math.exp(log_prob)) for c, log_prob in
|
||||
zip(classes, log_probs)
|
||||
))
|
||||
|
||||
|
||||
def get_cached_data_loader(dataset, batch_size, discriminator,
|
||||
shuffle=False, device='cpu'):
|
||||
data_loader = torch.utils.data.DataLoader(dataset=dataset,
|
||||
batch_size=batch_size,
|
||||
collate_fn=collate_fn)
|
||||
|
||||
xs = []
|
||||
ys = []
|
||||
for batch_idx, (x, y) in enumerate(tqdm(data_loader, ascii=True)):
|
||||
with torch.no_grad():
|
||||
x = x.to(device)
|
||||
avg_rep = discriminator.avg_representation(x).cpu().detach()
|
||||
avg_rep_list = torch.unbind(avg_rep.unsqueeze(1))
|
||||
xs += avg_rep_list
|
||||
ys += y.cpu().numpy().tolist()
|
||||
|
||||
data_loader = torch.utils.data.DataLoader(
|
||||
dataset=Dataset(xs, ys),
|
||||
batch_size=batch_size,
|
||||
shuffle=shuffle,
|
||||
collate_fn=cached_collate_fn)
|
||||
|
||||
return data_loader
|
||||
|
||||
|
||||
def train_discriminator(
|
||||
dataset, dataset_fp=None, pretrained_model="gpt2-medium",
|
||||
epochs=10, batch_size=64, log_interval=10,
|
||||
save_model=False, cached=False, no_cuda=False):
|
||||
device = "cuda" if torch.cuda.is_available() and not no_cuda else "cpu"
|
||||
|
||||
print("Preprocessing {} dataset...".format(dataset))
|
||||
start = time.time()
|
||||
|
||||
if dataset == "SST":
|
||||
idx2class = ["positive", "negative", "very positive", "very negative",
|
||||
"neutral"]
|
||||
class2idx = {c: i for i, c in enumerate(idx2class)}
|
||||
|
||||
discriminator = Discriminator(
|
||||
class_size=len(idx2class),
|
||||
pretrained_model=pretrained_model,
|
||||
cached_mode=cached,
|
||||
device=device
|
||||
).to(device)
|
||||
|
||||
text = torchtext_data.Field()
|
||||
label = torchtext_data.Field(sequential=False)
|
||||
train_data, val_data, test_data = datasets.SST.splits(
|
||||
text,
|
||||
label,
|
||||
fine_grained=True,
|
||||
train_subtrees=True,
|
||||
)
|
||||
|
||||
x = []
|
||||
y = []
|
||||
for i in trange(len(train_data), ascii=True):
|
||||
seq = TreebankWordDetokenizer().detokenize(
|
||||
vars(train_data[i])["text"]
|
||||
)
|
||||
seq = discriminator.tokenizer.encode(seq)
|
||||
seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
|
||||
x.append(seq)
|
||||
y.append(class2idx[vars(train_data[i])["label"]])
|
||||
train_dataset = Dataset(x, y)
|
||||
|
||||
test_x = []
|
||||
test_y = []
|
||||
for i in trange(len(test_data), ascii=True):
|
||||
seq = TreebankWordDetokenizer().detokenize(
|
||||
vars(test_data[i])["text"]
|
||||
)
|
||||
seq = discriminator.tokenizer.encode(seq)
|
||||
seq = torch.tensor([50256] + seq, device=device, dtype=torch.long)
|
||||
test_x.append(seq)
|
||||
test_y.append(class2idx[vars(test_data[i])["label"]])
|
||||
test_dataset = Dataset(test_x, test_y)
|
||||
|
||||
discriminator_meta = {
|
||||
"class_size": len(idx2class),
|
||||
"embed_size": discriminator.embed_size,
|
||||
"pretrained_model": pretrained_model,
|
||||
"class_vocab": class2idx,
|
||||
"default_class": 2,
|
||||
}
|
||||
|
||||
elif dataset == "clickbait":
|
||||
idx2class = ["non_clickbait", "clickbait"]
|
||||
class2idx = {c: i for i, c in enumerate(idx2class)}
|
||||
|
||||
discriminator = Discriminator(
|
||||
class_size=len(idx2class),
|
||||
pretrained_model=pretrained_model,
|
||||
cached_mode=cached,
|
||||
device=device
|
||||
).to(device)
|
||||
|
||||
with open("datasets/clickbait/clickbait_train_prefix.txt") as f:
|
||||
data = []
|
||||
for i, line in enumerate(f):
|
||||
try:
|
||||
data.append(eval(line))
|
||||
except:
|
||||
print("Error evaluating line {}: {}".format(
|
||||
i, line
|
||||
))
|
||||
continue
|
||||
x = []
|
||||
y = []
|
||||
with open("datasets/clickbait/clickbait_train_prefix.txt") as f:
|
||||
for i, line in enumerate(tqdm(f, ascii=True)):
|
||||
try:
|
||||
d = eval(line)
|
||||
seq = discriminator.tokenizer.encode(d["text"])
|
||||
|
||||
if len(seq) < max_length_seq:
|
||||
seq = torch.tensor(
|
||||
[50256] + seq, device=device, dtype=torch.long
|
||||
)
|
||||
else:
|
||||
print("Line {} is longer than maximum length {}".format(
|
||||
i, max_length_seq
|
||||
))
|
||||
continue
|
||||
x.append(seq)
|
||||
y.append(d["label"])
|
||||
except:
|
||||
print("Error evaluating / tokenizing"
|
||||
" line {}, skipping it".format(i))
|
||||
pass
|
||||
|
||||
full_dataset = Dataset(x, y)
|
||||
train_size = int(0.9 * len(full_dataset))
|
||||
test_size = len(full_dataset) - train_size
|
||||
train_dataset, test_dataset = torch.utils.data.random_split(
|
||||
full_dataset, [train_size, test_size]
|
||||
)
|
||||
|
||||
discriminator_meta = {
|
||||
"class_size": len(idx2class),
|
||||
"embed_size": discriminator.embed_size,
|
||||
"pretrained_model": pretrained_model,
|
||||
"class_vocab": class2idx,
|
||||
"default_class": 1,
|
||||
}
|
||||
|
||||
elif dataset == "toxic":
|
||||
idx2class = ["non_toxic", "toxic"]
|
||||
class2idx = {c: i for i, c in enumerate(idx2class)}
|
||||
|
||||
discriminator = Discriminator(
|
||||
class_size=len(idx2class),
|
||||
pretrained_model=pretrained_model,
|
||||
cached_mode=cached,
|
||||
device=device
|
||||
).to(device)
|
||||
|
||||
x = []
|
||||
y = []
|
||||
with open("datasets/toxic/toxic_train.txt") as f:
|
||||
for i, line in enumerate(tqdm(f, ascii=True)):
|
||||
try:
|
||||
d = eval(line)
|
||||
seq = discriminator.tokenizer.encode(d["text"])
|
||||
|
||||
if len(seq) < max_length_seq:
|
||||
seq = torch.tensor(
|
||||
[50256] + seq, device=device, dtype=torch.long
|
||||
)
|
||||
else:
|
||||
print("Line {} is longer than maximum length {}".format(
|
||||
i, max_length_seq
|
||||
))
|
||||
continue
|
||||
x.append(seq)
|
||||
y.append(int(np.sum(d["label"]) > 0))
|
||||
except:
|
||||
print("Error evaluating / tokenizing"
|
||||
" line {}, skipping it".format(i))
|
||||
pass
|
||||
|
||||
full_dataset = Dataset(x, y)
|
||||
train_size = int(0.9 * len(full_dataset))
|
||||
test_size = len(full_dataset) - train_size
|
||||
train_dataset, test_dataset = torch.utils.data.random_split(
|
||||
full_dataset, [train_size, test_size]
|
||||
)
|
||||
|
||||
discriminator_meta = {
|
||||
"class_size": len(idx2class),
|
||||
"embed_size": discriminator.embed_size,
|
||||
"pretrained_model": pretrained_model,
|
||||
"class_vocab": class2idx,
|
||||
"default_class": 0,
|
||||
}
|
||||
|
||||
else: # if dataset == "generic":
|
||||
# This assumes the input dataset is a TSV with the following structure:
|
||||
# class \t text
|
||||
|
||||
if dataset_fp is None:
|
||||
raise ValueError("When generic dataset is selected, "
|
||||
"dataset_fp needs to be specified aswell.")
|
||||
|
||||
classes = set()
|
||||
with open(dataset_fp) as f:
|
||||
csv_reader = csv.reader(f, delimiter="\t")
|
||||
for row in tqdm(csv_reader, ascii=True):
|
||||
if row:
|
||||
classes.add(row[0])
|
||||
|
||||
idx2class = sorted(classes)
|
||||
class2idx = {c: i for i, c in enumerate(idx2class)}
|
||||
|
||||
discriminator = Discriminator(
|
||||
class_size=len(idx2class),
|
||||
pretrained_model=pretrained_model,
|
||||
cached_mode=cached,
|
||||
device=device
|
||||
).to(device)
|
||||
|
||||
x = []
|
||||
y = []
|
||||
with open(dataset_fp) as f:
|
||||
csv_reader = csv.reader(f, delimiter="\t")
|
||||
for i, row in enumerate(tqdm(csv_reader, ascii=True)):
|
||||
if row:
|
||||
label = row[0]
|
||||
text = row[1]
|
||||
|
||||
try:
|
||||
seq = discriminator.tokenizer.encode(text)
|
||||
if (len(seq) < max_length_seq):
|
||||
seq = torch.tensor(
|
||||
[50256] + seq,
|
||||
device=device,
|
||||
dtype=torch.long
|
||||
)
|
||||
|
||||
else:
|
||||
print(
|
||||
"Line {} is longer than maximum length {}".format(
|
||||
i, max_length_seq
|
||||
))
|
||||
continue
|
||||
|
||||
x.append(seq)
|
||||
y.append(class2idx[label])
|
||||
|
||||
except:
|
||||
print("Error tokenizing line {}, skipping it".format(i))
|
||||
pass
|
||||
|
||||
full_dataset = Dataset(x, y)
|
||||
train_size = int(0.9 * len(full_dataset))
|
||||
test_size = len(full_dataset) - train_size
|
||||
train_dataset, test_dataset = torch.utils.data.random_split(
|
||||
full_dataset,
|
||||
[train_size, test_size]
|
||||
)
|
||||
|
||||
discriminator_meta = {
|
||||
"class_size": len(idx2class),
|
||||
"embed_size": discriminator.embed_size,
|
||||
"pretrained_model": pretrained_model,
|
||||
"class_vocab": class2idx,
|
||||
"default_class": 0,
|
||||
}
|
||||
|
||||
end = time.time()
|
||||
print("Preprocessed {} data points".format(
|
||||
len(train_dataset) + len(test_dataset))
|
||||
)
|
||||
print("Data preprocessing took: {:.3f}s".format(end - start))
|
||||
|
||||
if cached:
|
||||
print("Building representation cache...")
|
||||
|
||||
start = time.time()
|
||||
|
||||
train_loader = get_cached_data_loader(
|
||||
train_dataset, batch_size, discriminator,
|
||||
shuffle=True, device=device
|
||||
)
|
||||
|
||||
test_loader = get_cached_data_loader(
|
||||
test_dataset, batch_size, discriminator, device=device
|
||||
)
|
||||
|
||||
end = time.time()
|
||||
print("Building representation cache took: {:.3f}s".format(end - start))
|
||||
|
||||
else:
|
||||
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
|
||||
batch_size=batch_size,
|
||||
shuffle=True,
|
||||
collate_fn=collate_fn)
|
||||
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
|
||||
batch_size=batch_size,
|
||||
collate_fn=collate_fn)
|
||||
|
||||
if save_model:
|
||||
with open("{}_classifier_head_meta.json".format(dataset),
|
||||
"w") as meta_file:
|
||||
json.dump(discriminator_meta, meta_file)
|
||||
|
||||
optimizer = optim.Adam(discriminator.parameters(), lr=0.0001)
|
||||
|
||||
for epoch in range(epochs):
|
||||
start = time.time()
|
||||
print("\nEpoch", epoch + 1)
|
||||
|
||||
train_epoch(
|
||||
discriminator=discriminator,
|
||||
data_loader=train_loader,
|
||||
optimizer=optimizer,
|
||||
epoch=epoch,
|
||||
log_interval=log_interval,
|
||||
device=device
|
||||
)
|
||||
evaluate_performance(
|
||||
data_loader=test_loader,
|
||||
discriminator=discriminator,
|
||||
device=device
|
||||
)
|
||||
|
||||
end = time.time()
|
||||
print("Epoch took: {:.3f}s".format(end - start))
|
||||
|
||||
print("\nExample prediction")
|
||||
predict(example_sentence, discriminator, idx2class,
|
||||
cached=cached, device=device)
|
||||
|
||||
if save_model:
|
||||
# torch.save(discriminator.state_dict(),
|
||||
# "{}_discriminator_{}.pt".format(
|
||||
# args.dataset, epoch + 1
|
||||
# ))
|
||||
torch.save(discriminator.get_classifier().state_dict(),
|
||||
"{}_classifier_head_epoch_{}.pt".format(dataset,
|
||||
epoch + 1))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Train a discriminator on top of GPT-2 representations")
|
||||
parser.add_argument("--dataset", type=str, default="SST",
|
||||
choices=("SST", "clickbait", "toxic", "generic"),
|
||||
help="dataset to train the discriminator on."
|
||||
"In case of generic, the dataset is expected"
|
||||
"to be a TSBV file with structure: class \\t text")
|
||||
parser.add_argument("--dataset_fp", type=str, default="",
|
||||
help="File path of the dataset to use. "
|
||||
"Needed only in case of generic datadset")
|
||||
parser.add_argument("--pretrained_model", type=str, default="gpt2-medium",
|
||||
help="Pretrained model to use as encoder")
|
||||
parser.add_argument("--epochs", type=int, default=10, metavar="N",
|
||||
help="Number of training epochs")
|
||||
parser.add_argument("--batch_size", type=int, default=64, metavar="N",
|
||||
help="input batch size for training (default: 64)")
|
||||
parser.add_argument("--log_interval", type=int, default=10, metavar="N",
|
||||
help="how many batches to wait before logging training status")
|
||||
parser.add_argument("--save_model", action="store_true",
|
||||
help="whether to save the model")
|
||||
parser.add_argument("--cached", action="store_true",
|
||||
help="whether to cache the input representations")
|
||||
parser.add_argument("--no_cuda", action="store_true",
|
||||
help="use to turn off cuda")
|
||||
args = parser.parse_args()
|
||||
|
||||
train_discriminator(**(vars(args)))
|
||||
@@ -1,4 +1,2 @@
|
||||
tensorboardX
|
||||
tensorboard
|
||||
scikit-learn
|
||||
seqeval
|
||||
scikit-learn
|
||||
@@ -39,9 +39,8 @@ from transformers import (WEIGHTS_NAME,
|
||||
|
||||
from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES
|
||||
|
||||
from transformers import glue_compute_metrics as compute_metrics
|
||||
from transformers import glue_output_modes as output_modes
|
||||
from transformers import glue_processors as processors
|
||||
from utils_glue import (compute_metrics, convert_examples_to_features,
|
||||
output_modes, processors)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -234,8 +233,6 @@ def main():
|
||||
help="If > 0: limit the data to a subset of data_subset instances.")
|
||||
parser.add_argument("--overwrite_output_dir", action='store_true',
|
||||
help="Whether to overwrite data in output directory")
|
||||
parser.add_argument('--overwrite_cache', action='store_true',
|
||||
help="Overwrite the cached training and evaluation sets")
|
||||
|
||||
parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
|
||||
help="Don't normalize importance score by layers")
|
||||
@@ -307,16 +304,10 @@ def main():
|
||||
break
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
num_labels=num_labels,
|
||||
finetuning_task=args.task_name,
|
||||
output_attentions=True,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
num_labels=num_labels, finetuning_task=args.task_name,
|
||||
output_attentions=True)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path)
|
||||
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
+18
-83
@@ -14,7 +14,7 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/CTRL/Transformer-XL/XLNet)
|
||||
""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/Transformer-XL/XLNet)
|
||||
"""
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
@@ -26,14 +26,12 @@ import torch
|
||||
import torch.nn.functional as F
|
||||
import numpy as np
|
||||
|
||||
from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
|
||||
from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig
|
||||
|
||||
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
||||
from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
|
||||
from transformers import XLNetLMHeadModel, XLNetTokenizer
|
||||
from transformers import TransfoXLLMHeadModel, TransfoXLTokenizer
|
||||
from transformers import CTRLLMHeadModel, CTRLTokenizer
|
||||
from transformers import XLMWithLMHeadModel, XLMTokenizer
|
||||
|
||||
|
||||
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
|
||||
@@ -43,15 +41,13 @@ logger = logging.getLogger(__name__)
|
||||
|
||||
MAX_LENGTH = int(10000) # Hardcoded max length to avoid infinite loop
|
||||
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig)), ())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
|
||||
'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
|
||||
'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
|
||||
'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
|
||||
'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
|
||||
'xlm': (XLMWithLMHeadModel, XLMTokenizer),
|
||||
}
|
||||
|
||||
# Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
|
||||
@@ -79,12 +75,13 @@ def set_seed(args):
|
||||
def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
|
||||
""" Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
|
||||
Args:
|
||||
logits: logits distribution shape (batch size x vocabulary size)
|
||||
logits: logits distribution shape (vocabulary size)
|
||||
top_k > 0: keep only top k tokens with highest probability (top-k filtering).
|
||||
top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
|
||||
Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
|
||||
From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
|
||||
"""
|
||||
assert logits.dim() == 1 # batch size 1 for now - could be updated for more but the code would be less clear
|
||||
top_k = min(top_k, logits.size(-1)) # Safety check
|
||||
if top_k > 0:
|
||||
# Remove all tokens with a probability less than the last token of the top-k
|
||||
@@ -101,14 +98,12 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
|
||||
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
||||
sorted_indices_to_remove[..., 0] = 0
|
||||
|
||||
# scatter sorted tensors to original indexing
|
||||
indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
|
||||
indices_to_remove = sorted_indices[sorted_indices_to_remove]
|
||||
logits[indices_to_remove] = filter_value
|
||||
return logits
|
||||
|
||||
|
||||
def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
|
||||
is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
|
||||
def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, is_xlnet=False, device='cpu'):
|
||||
context = torch.tensor(context, dtype=torch.long, device=device)
|
||||
context = context.unsqueeze(0).repeat(num_samples, 1)
|
||||
generated = context
|
||||
@@ -126,29 +121,11 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
|
||||
target_mapping[0, 0, -1] = 1.0 # predict last token
|
||||
inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
|
||||
|
||||
if is_xlm_mlm and xlm_mask_token:
|
||||
# XLM MLM models are direct models (predict same token, not next token)
|
||||
# => need one additional dummy token in the input (will be masked and guessed)
|
||||
input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
|
||||
inputs = {'input_ids': input_ids}
|
||||
|
||||
if xlm_lang is not None:
|
||||
inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
|
||||
|
||||
outputs = model(**inputs) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
|
||||
next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
|
||||
|
||||
# repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
|
||||
for i in range(num_samples):
|
||||
for _ in set(generated[i].tolist()):
|
||||
next_token_logits[i, _] /= repetition_penalty
|
||||
|
||||
outputs = model(**inputs) # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
|
||||
next_token_logits = outputs[0][0, -1, :] / temperature
|
||||
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
|
||||
if temperature == 0: # greedy sampling:
|
||||
next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
|
||||
else:
|
||||
next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
|
||||
generated = torch.cat((generated, next_token), dim=1)
|
||||
next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
|
||||
generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
|
||||
return generated
|
||||
|
||||
|
||||
@@ -160,21 +137,14 @@ def main():
|
||||
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
parser.add_argument("--prompt", type=str, default="")
|
||||
parser.add_argument("--padding_text", type=str, default="")
|
||||
parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
|
||||
parser.add_argument("--length", type=int, default=20)
|
||||
parser.add_argument("--num_samples", type=int, default=1)
|
||||
parser.add_argument("--temperature", type=float, default=1.0,
|
||||
help="temperature of 0 implies greedy sampling")
|
||||
parser.add_argument("--repetition_penalty", type=float, default=1.0,
|
||||
help="primarily useful for CTRL model; in that case, use 1.2")
|
||||
parser.add_argument("--temperature", type=float, default=1.0)
|
||||
parser.add_argument("--top_k", type=int, default=0)
|
||||
parser.add_argument("--top_p", type=float, default=0.9)
|
||||
parser.add_argument("--no_cuda", action='store_true',
|
||||
help="Avoid using CUDA when available")
|
||||
parser.add_argument('--seed', type=int, default=42,
|
||||
help="random seed for initialization")
|
||||
parser.add_argument('--stop_token', type=str, default=None,
|
||||
help="Token at which text generation is stopped")
|
||||
args = parser.parse_args()
|
||||
|
||||
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
@@ -196,61 +166,26 @@ def main():
|
||||
elif args.length < 0:
|
||||
args.length = MAX_LENGTH # avoid infinite loop
|
||||
|
||||
logger.info(args)
|
||||
if args.model_type in ["ctrl"]:
|
||||
if args.temperature > 0.7:
|
||||
logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
|
||||
|
||||
print(args)
|
||||
while True:
|
||||
xlm_lang = None
|
||||
# XLM Language usage detailed in the issues #1414
|
||||
if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
|
||||
and model.config.use_lang_emb:
|
||||
if args.xlm_lang:
|
||||
language = args.xlm_lang
|
||||
else:
|
||||
language = None
|
||||
while language not in tokenizer.lang2id.keys():
|
||||
language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
|
||||
xlm_lang = tokenizer.lang2id[language]
|
||||
|
||||
# XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
|
||||
is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
|
||||
if is_xlm_mlm:
|
||||
xlm_mask_token = tokenizer.mask_token_id
|
||||
else:
|
||||
xlm_mask_token = None
|
||||
|
||||
raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
|
||||
if args.model_type in ["transfo-xl", "xlnet"]:
|
||||
# Models with memory likes to have a long prompt for short inputs.
|
||||
raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
|
||||
context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
|
||||
if args.model_type == "ctrl":
|
||||
if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
|
||||
logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
|
||||
context_tokens = tokenizer.encode(raw_text)
|
||||
out = sample_sequence(
|
||||
model=model,
|
||||
context=context_tokens,
|
||||
num_samples=args.num_samples,
|
||||
length=args.length,
|
||||
temperature=args.temperature,
|
||||
top_k=args.top_k,
|
||||
top_p=args.top_p,
|
||||
repetition_penalty=args.repetition_penalty,
|
||||
is_xlnet=bool(args.model_type == "xlnet"),
|
||||
is_xlm_mlm=is_xlm_mlm,
|
||||
xlm_mask_token=xlm_mask_token,
|
||||
xlm_lang=xlm_lang,
|
||||
device=args.device,
|
||||
is_xlnet=bool(args.model_type == "xlnet"),
|
||||
)
|
||||
out = out[:, len(context_tokens):].tolist()
|
||||
for o in out:
|
||||
text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
|
||||
text = text[: text.find(args.stop_token) if args.stop_token else None]
|
||||
|
||||
print(text)
|
||||
|
||||
out = out[0, len(context_tokens):].tolist()
|
||||
text = tokenizer.decode(out, clean_up_tokenization_spaces=True)
|
||||
print(text)
|
||||
if args.prompt:
|
||||
break
|
||||
return text
|
||||
|
||||
+23
-61
@@ -22,19 +22,13 @@ import glob
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
import json
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tensorboardX import SummaryWriter
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
@@ -48,13 +42,9 @@ from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
XLNetTokenizer,
|
||||
DistilBertConfig,
|
||||
DistilBertForSequenceClassification,
|
||||
DistilBertTokenizer,
|
||||
AlbertConfig,
|
||||
AlbertForSequenceClassification,
|
||||
AlbertTokenizer,
|
||||
)
|
||||
DistilBertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
from transformers import AdamW, WarmupLinearSchedule
|
||||
|
||||
from transformers import glue_compute_metrics as compute_metrics
|
||||
from transformers import glue_output_modes as output_modes
|
||||
@@ -63,16 +53,14 @@ from transformers import glue_convert_examples_to_features as convert_examples_t
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig,
|
||||
RobertaConfig, DistilBertConfig)), ())
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), ())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
|
||||
'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
|
||||
'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
|
||||
'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
|
||||
'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
|
||||
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
|
||||
@@ -105,9 +93,8 @@ def train(args, train_dataset, model, tokenizer):
|
||||
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
@@ -161,39 +148,28 @@ def train(args, train_dataset, model, tokenizer):
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
if args.fp16:
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
optimizer.step()
|
||||
scheduler.step() # Update learning rate schedule
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
logs = {}
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results = evaluate(args, model, tokenizer)
|
||||
for key, value in results.items():
|
||||
eval_key = 'eval_{}'.format(key)
|
||||
logs[eval_key] = value
|
||||
|
||||
loss_scalar = (tr_loss - logging_loss) / args.logging_steps
|
||||
learning_rate_scalar = scheduler.get_lr()[0]
|
||||
logs['learning_rate'] = learning_rate_scalar
|
||||
logs['loss'] = loss_scalar
|
||||
tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
|
||||
tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
|
||||
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
for key, value in logs.items():
|
||||
tb_writer.add_scalar(key, value, global_step)
|
||||
print(json.dumps({**logs, **{'step': global_step}}))
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
@@ -231,13 +207,9 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(eval_dataset)
|
||||
eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
|
||||
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu eval
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(eval_dataset))
|
||||
@@ -276,7 +248,7 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
result = compute_metrics(eval_task, preds, out_label_ids)
|
||||
results.update(result)
|
||||
|
||||
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
|
||||
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
|
||||
with open(output_eval_file, "w") as writer:
|
||||
logger.info("***** Eval results {} *****".format(prefix))
|
||||
for key in sorted(result.keys()):
|
||||
@@ -298,7 +270,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length),
|
||||
str(task)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
||||
if os.path.exists(cached_features_file):
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
@@ -332,7 +304,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
|
||||
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
|
||||
elif output_mode == "regression":
|
||||
all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
|
||||
|
||||
|
||||
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
|
||||
return dataset
|
||||
|
||||
@@ -376,11 +348,11 @@ def main():
|
||||
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for evaluation.")
|
||||
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--learning_rate", default=5e-5, type=float,
|
||||
help="The initial learning rate for Adam.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight decay if we apply some.")
|
||||
help="Weight deay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
@@ -465,17 +437,9 @@ def main():
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
num_labels=num_labels,
|
||||
finetuning_task=args.task_name,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
|
||||
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
@@ -510,7 +474,7 @@ def main():
|
||||
|
||||
# Load a trained model and vocabulary that you have fine-tuned
|
||||
model = model_class.from_pretrained(args.output_dir)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
model.to(args.device)
|
||||
|
||||
|
||||
@@ -525,11 +489,9 @@ def main():
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
|
||||
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result = evaluate(args, model, tokenizer, prefix=prefix)
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step)
|
||||
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
|
||||
+22
-119
@@ -27,28 +27,20 @@ import logging
|
||||
import os
|
||||
import pickle
|
||||
import random
|
||||
import re
|
||||
import shutil
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tensorboardX import SummaryWriter
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup,
|
||||
from transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
|
||||
BertConfig, BertForMaskedLM, BertTokenizer,
|
||||
GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
|
||||
OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
|
||||
RobertaConfig, RobertaForMaskedLM, RobertaTokenizer,
|
||||
DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer,
|
||||
CamembertConfig, CamembertForMaskedLM, CamembertTokenizer)
|
||||
DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
|
||||
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
@@ -59,18 +51,17 @@ MODEL_CLASSES = {
|
||||
'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
|
||||
'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
|
||||
'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer),
|
||||
'camembert': (CamembertConfig, CamembertForMaskedLM, CamembertTokenizer)
|
||||
'distilbert': (DistilBertConfig, DistilBertForMaskedLM, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
|
||||
class TextDataset(Dataset):
|
||||
def __init__(self, tokenizer, args, file_path='train', block_size=512):
|
||||
def __init__(self, tokenizer, file_path='train', block_size=512):
|
||||
assert os.path.isfile(file_path)
|
||||
directory, filename = os.path.split(file_path)
|
||||
cached_features_file = os.path.join(directory, args.model_name_or_path + '_cached_lm_' + str(block_size) + '_' + filename)
|
||||
cached_features_file = os.path.join(directory, 'cached_lm_{}_{}'.format(block_size, filename))
|
||||
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
||||
if os.path.exists(cached_features_file):
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
with open(cached_features_file, 'rb') as handle:
|
||||
self.examples = pickle.load(handle)
|
||||
@@ -84,7 +75,7 @@ class TextDataset(Dataset):
|
||||
tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
|
||||
|
||||
for i in range(0, len(tokenized_text)-block_size+1, block_size): # Truncate in block of block_size
|
||||
self.examples.append(tokenizer.build_inputs_with_special_tokens(tokenized_text[i:i+block_size]))
|
||||
self.examples.append(tokenizer.add_special_tokens_single_sequence(tokenized_text[i:i+block_size]))
|
||||
# Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
|
||||
# If your dataset is small, first you should loook for a bigger one :-) and second you
|
||||
# can change this behavior by adding (model specific) padding.
|
||||
@@ -101,7 +92,7 @@ class TextDataset(Dataset):
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, evaluate=False):
|
||||
dataset = TextDataset(tokenizer, args, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
|
||||
dataset = TextDataset(tokenizer, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
|
||||
return dataset
|
||||
|
||||
|
||||
@@ -113,43 +104,11 @@ def set_seed(args):
|
||||
torch.cuda.manual_seed_all(args.seed)
|
||||
|
||||
|
||||
def _rotate_checkpoints(args, checkpoint_prefix, use_mtime=False):
|
||||
if not args.save_total_limit:
|
||||
return
|
||||
if args.save_total_limit <= 0:
|
||||
return
|
||||
|
||||
# Check if we should delete older checkpoint(s)
|
||||
glob_checkpoints = glob.glob(os.path.join(args.output_dir, '{}-*'.format(checkpoint_prefix)))
|
||||
if len(glob_checkpoints) <= args.save_total_limit:
|
||||
return
|
||||
|
||||
ordering_and_checkpoint_path = []
|
||||
for path in glob_checkpoints:
|
||||
if use_mtime:
|
||||
ordering_and_checkpoint_path.append((os.path.getmtime(path), path))
|
||||
else:
|
||||
regex_match = re.match('.*{}-([0-9]+)'.format(checkpoint_prefix), path)
|
||||
if regex_match and regex_match.groups():
|
||||
ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
|
||||
|
||||
checkpoints_sorted = sorted(ordering_and_checkpoint_path)
|
||||
checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
|
||||
number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - args.save_total_limit)
|
||||
checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
|
||||
for checkpoint in checkpoints_to_be_deleted:
|
||||
logger.info("Deleting older checkpoint [{}] due to args.save_total_limit".format(checkpoint))
|
||||
shutil.rmtree(checkpoint)
|
||||
|
||||
|
||||
def mask_tokens(inputs, tokenizer, args):
|
||||
""" Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
|
||||
labels = inputs.clone()
|
||||
# We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
|
||||
probability_matrix = torch.full(labels.shape, args.mlm_probability)
|
||||
special_tokens_mask = [tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()]
|
||||
probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
|
||||
masked_indices = torch.bernoulli(probability_matrix).bool()
|
||||
masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).bool()
|
||||
labels[~masked_indices] = -1 # We only compute loss on masked tokens
|
||||
|
||||
# 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
|
||||
@@ -187,14 +146,7 @@ def train(args, train_dataset, model, tokenizer):
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
|
||||
# Check if saved optimizer or scheduler states exist
|
||||
if os.path.isfile(os.path.join(args.model_name_or_path, 'optimizer.pt')) and os.path.isfile(os.path.join(args.model_name_or_path, 'scheduler.pt')):
|
||||
# Load in optimizer and scheduler states
|
||||
optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'optimizer.pt')))
|
||||
scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, 'scheduler.pt')))
|
||||
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
@@ -223,37 +175,13 @@ def train(args, train_dataset, model, tokenizer):
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 0
|
||||
epochs_trained = 0
|
||||
steps_trained_in_current_epoch = 0
|
||||
# Check if continuing training from a checkpoint
|
||||
if os.path.exists(args.model_name_or_path):
|
||||
# set global_step to gobal_step of last saved checkpoint from model path
|
||||
global_step = int(args.model_name_or_path.split('-')[-1].split('/')[0])
|
||||
epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
|
||||
steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
|
||||
|
||||
logger.info(" Continuing training from checkpoint, will skip to saved global_step")
|
||||
logger.info(" Continuing training from epoch %d", epochs_trained)
|
||||
logger.info(" Continuing training from global step %d", global_step)
|
||||
logger.info(" Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
|
||||
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
|
||||
model_to_resize = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_resize.resize_token_embeddings(len(tokenizer))
|
||||
|
||||
model.zero_grad()
|
||||
train_iterator = trange(epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproducibility (even between python 2 and 3)
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
|
||||
# Skip past any already trained steps if resuming training
|
||||
if steps_trained_in_current_epoch > 0:
|
||||
steps_trained_in_current_epoch -= 1
|
||||
continue
|
||||
|
||||
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
|
||||
inputs = inputs.to(args.device)
|
||||
labels = labels.to(args.device)
|
||||
@@ -294,24 +222,15 @@ def train(args, train_dataset, model, tokenizer):
|
||||
logging_loss = tr_loss
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
checkpoint_prefix = 'checkpoint'
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, '{}-{}'.format(checkpoint_prefix, global_step))
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(output_dir)
|
||||
tokenizer.save_pretrained(output_dir)
|
||||
|
||||
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
|
||||
logger.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
_rotate_checkpoints(args, checkpoint_prefix)
|
||||
|
||||
torch.save(optimizer.state_dict(), os.path.join(output_dir, 'optimizer.pt'))
|
||||
torch.save(scheduler.state_dict(), os.path.join(output_dir, 'scheduler.pt'))
|
||||
logger.info("Saving optimizer and scheduler states to %s", output_dir)
|
||||
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
epoch_iterator.close()
|
||||
break
|
||||
@@ -336,13 +255,9 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(eval_dataset)
|
||||
eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
|
||||
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu evaluate
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(eval_dataset))
|
||||
@@ -352,12 +267,10 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
model.eval()
|
||||
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
|
||||
inputs = inputs.to(args.device)
|
||||
labels = labels.to(args.device)
|
||||
batch = batch.to(args.device)
|
||||
|
||||
with torch.no_grad():
|
||||
outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
|
||||
outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
|
||||
lm_loss = outputs[0]
|
||||
eval_loss += lm_loss.mean().item()
|
||||
nb_eval_steps += 1
|
||||
@@ -369,7 +282,7 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
"perplexity": perplexity
|
||||
}
|
||||
|
||||
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
|
||||
output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
|
||||
with open(output_eval_file, "w") as writer:
|
||||
logger.info("***** Eval results {} *****".format(prefix))
|
||||
for key in sorted(result.keys()):
|
||||
@@ -446,8 +359,6 @@ def main():
|
||||
help="Log every X updates steps.")
|
||||
parser.add_argument('--save_steps', type=int, default=50,
|
||||
help="Save checkpoint every X updates steps.")
|
||||
parser.add_argument('--save_total_limit', type=int, default=None,
|
||||
help='Limit the total amount of checkpoints, delete the older checkpoints in the output_dir, does not delete by default')
|
||||
parser.add_argument("--eval_all_checkpoints", action='store_true',
|
||||
help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
|
||||
parser.add_argument("--no_cuda", action='store_true',
|
||||
@@ -470,7 +381,7 @@ def main():
|
||||
parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.model_type in ["bert", "roberta", "distilbert", "camembert"] and not args.mlm:
|
||||
if args.model_type in ["bert", "roberta", "distilbert"] and not args.mlm:
|
||||
raise ValueError("BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
|
||||
"flag (masked language modeling).")
|
||||
if args.eval_data_file is None and args.do_eval:
|
||||
@@ -514,18 +425,12 @@ def main():
|
||||
torch.distributed.barrier() # Barrier to make sure only the first process in distributed training download model & vocab
|
||||
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
|
||||
if args.block_size <= 0:
|
||||
args.block_size = tokenizer.max_len_single_sentence # Our input block size will be the max possible for the model
|
||||
args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
|
||||
model.to(args.device)
|
||||
|
||||
if args.local_rank == 0:
|
||||
@@ -579,11 +484,9 @@ def main():
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
|
||||
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result = evaluate(args, model, tokenizer, prefix=prefix)
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step)
|
||||
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
|
||||
@@ -29,12 +29,7 @@ import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tensorboardX import SummaryWriter
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
@@ -43,7 +38,7 @@ from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
XLNetTokenizer, RobertaConfig,
|
||||
RobertaForMultipleChoice, RobertaTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
from transformers import AdamW, WarmupLinearSchedule
|
||||
|
||||
from utils_multiple_choice import (convert_examples_to_features, processors)
|
||||
|
||||
@@ -101,7 +96,7 @@ def train(args, train_dataset, model, tokenizer):
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
@@ -226,13 +221,9 @@ def evaluate(args, model, tokenizer, prefix="", test=False):
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(eval_dataset)
|
||||
eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
|
||||
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu evaluate
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(eval_dataset))
|
||||
@@ -302,7 +293,7 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length),
|
||||
str(task)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
||||
if os.path.exists(cached_features_file):
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
@@ -315,14 +306,14 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False, test=False):
|
||||
else:
|
||||
examples = processor.get_train_examples(args.data_dir)
|
||||
logger.info("Training number: %s", str(len(examples)))
|
||||
features = convert_examples_to_features(
|
||||
examples,
|
||||
label_list,
|
||||
args.max_seq_length,
|
||||
tokenizer,
|
||||
features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer,
|
||||
cls_token_at_end=bool(args.model_type in ['xlnet']), # xlnet has a cls token at the end
|
||||
cls_token=tokenizer.cls_token,
|
||||
sep_token=tokenizer.sep_token,
|
||||
sep_token_extra=bool(args.model_type in ['roberta']),
|
||||
cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
|
||||
pad_on_left=bool(args.model_type in ['xlnet']), # pad on the left for xlnet
|
||||
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0
|
||||
)
|
||||
pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
@@ -371,7 +362,7 @@ def main():
|
||||
help="Whether to run eval on the dev set.")
|
||||
parser.add_argument("--do_test", action='store_true', help='Whether to run test on the test set')
|
||||
parser.add_argument("--evaluate_during_training", action='store_true',
|
||||
help="Run evaluation during training at each logging step.")
|
||||
help="Rul evaluation during training at each logging step.")
|
||||
parser.add_argument("--do_lower_case", action='store_true',
|
||||
help="Set this flag if you are using an uncased model.")
|
||||
|
||||
@@ -468,17 +459,9 @@ def main():
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
num_labels=num_labels,
|
||||
finetuning_task=args.task_name,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
|
||||
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
@@ -529,11 +512,9 @@ def main():
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
|
||||
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result = evaluate(args, model, tokenizer, prefix=prefix)
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step)
|
||||
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
@@ -547,11 +528,9 @@ def main():
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
|
||||
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result = evaluate(args, model, tokenizer, prefix=prefix, test=True)
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step, test=True)
|
||||
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
if best_steps:
|
||||
|
||||
@@ -1,532 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import argparse
|
||||
import glob
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from seqeval.metrics import precision_score, recall_score, f1_score
|
||||
from tensorboardX import SummaryWriter
|
||||
from torch.nn import CrossEntropyLoss
|
||||
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
from tqdm import tqdm, trange
|
||||
from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer
|
||||
from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
|
||||
from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
|
||||
from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
ALL_MODELS = sum(
|
||||
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
|
||||
())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
"bert": (BertConfig, BertForTokenClassification, BertTokenizer),
|
||||
"roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
|
||||
"distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
|
||||
"camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
|
||||
}
|
||||
|
||||
|
||||
def set_seed(args):
|
||||
random.seed(args.seed)
|
||||
np.random.seed(args.seed)
|
||||
torch.manual_seed(args.seed)
|
||||
if args.n_gpu > 0:
|
||||
torch.cuda.manual_seed_all(args.seed)
|
||||
|
||||
|
||||
def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
|
||||
""" Train the model """
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer = SummaryWriter()
|
||||
|
||||
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
|
||||
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
|
||||
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
|
||||
|
||||
if args.max_steps > 0:
|
||||
t_total = args.max_steps
|
||||
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
|
||||
else:
|
||||
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
|
||||
|
||||
# Prepare optimizer and schedule (linear warmup and decay)
|
||||
no_decay = ["bias", "LayerNorm.weight"]
|
||||
optimizer_grouped_parameters = [
|
||||
{"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
|
||||
"weight_decay": args.weight_decay},
|
||||
{"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
||||
|
||||
# multi-gpu training (should be after apex fp16 initialization)
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Distributed training (should be after apex fp16 initialization)
|
||||
if args.local_rank != -1:
|
||||
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
|
||||
output_device=args.local_rank,
|
||||
find_unused_parameters=True)
|
||||
|
||||
# Train!
|
||||
logger.info("***** Running training *****")
|
||||
logger.info(" Num examples = %d", len(train_dataset))
|
||||
logger.info(" Num Epochs = %d", args.num_train_epochs)
|
||||
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
|
||||
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
||||
args.train_batch_size * args.gradient_accumulation_steps * (
|
||||
torch.distributed.get_world_size() if args.local_rank != -1 else 1))
|
||||
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 0
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
model.zero_grad()
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
model.train()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
inputs = {"input_ids": batch[0],
|
||||
"attention_mask": batch[1],
|
||||
"labels": batch[3]}
|
||||
if args.model_type != "distilbert":
|
||||
inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids
|
||||
|
||||
outputs = model(**inputs)
|
||||
loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc)
|
||||
|
||||
if args.n_gpu > 1:
|
||||
loss = loss.mean() # mean() to average on multi-gpu parallel training
|
||||
if args.gradient_accumulation_steps > 1:
|
||||
loss = loss / args.gradient_accumulation_steps
|
||||
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
else:
|
||||
loss.backward()
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
if args.fp16:
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
scheduler.step() # Update learning rate schedule
|
||||
optimizer.step()
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev")
|
||||
for key, value in results.items():
|
||||
tb_writer.add_scalar("eval_{}".format(key), value, global_step)
|
||||
tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
|
||||
tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(output_dir)
|
||||
torch.save(args, os.path.join(output_dir, "training_args.bin"))
|
||||
logger.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
epoch_iterator.close()
|
||||
break
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
train_iterator.close()
|
||||
break
|
||||
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer.close()
|
||||
|
||||
return global_step, tr_loss / global_step
|
||||
|
||||
|
||||
def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
|
||||
eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode)
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
|
||||
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu evaluate
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation %s *****", prefix)
|
||||
logger.info(" Num examples = %d", len(eval_dataset))
|
||||
logger.info(" Batch size = %d", args.eval_batch_size)
|
||||
eval_loss = 0.0
|
||||
nb_eval_steps = 0
|
||||
preds = None
|
||||
out_label_ids = None
|
||||
model.eval()
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
|
||||
with torch.no_grad():
|
||||
inputs = {"input_ids": batch[0],
|
||||
"attention_mask": batch[1],
|
||||
"labels": batch[3]}
|
||||
if args.model_type != "distilbert":
|
||||
inputs["token_type_ids"] = batch[2] if args.model_type in ["bert", "xlnet"] else None # XLM and RoBERTa don"t use segment_ids
|
||||
outputs = model(**inputs)
|
||||
tmp_eval_loss, logits = outputs[:2]
|
||||
|
||||
if args.n_gpu > 1:
|
||||
tmp_eval_loss = tmp_eval_loss.mean() # mean() to average on multi-gpu parallel evaluating
|
||||
|
||||
eval_loss += tmp_eval_loss.item()
|
||||
nb_eval_steps += 1
|
||||
if preds is None:
|
||||
preds = logits.detach().cpu().numpy()
|
||||
out_label_ids = inputs["labels"].detach().cpu().numpy()
|
||||
else:
|
||||
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
|
||||
out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
|
||||
|
||||
eval_loss = eval_loss / nb_eval_steps
|
||||
preds = np.argmax(preds, axis=2)
|
||||
|
||||
label_map = {i: label for i, label in enumerate(labels)}
|
||||
|
||||
out_label_list = [[] for _ in range(out_label_ids.shape[0])]
|
||||
preds_list = [[] for _ in range(out_label_ids.shape[0])]
|
||||
|
||||
for i in range(out_label_ids.shape[0]):
|
||||
for j in range(out_label_ids.shape[1]):
|
||||
if out_label_ids[i, j] != pad_token_label_id:
|
||||
out_label_list[i].append(label_map[out_label_ids[i][j]])
|
||||
preds_list[i].append(label_map[preds[i][j]])
|
||||
|
||||
results = {
|
||||
"loss": eval_loss,
|
||||
"precision": precision_score(out_label_list, preds_list),
|
||||
"recall": recall_score(out_label_list, preds_list),
|
||||
"f1": f1_score(out_label_list, preds_list)
|
||||
}
|
||||
|
||||
logger.info("***** Eval results %s *****", prefix)
|
||||
for key in sorted(results.keys()):
|
||||
logger.info(" %s = %s", key, str(results[key]))
|
||||
|
||||
return results, preds_list
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
|
||||
if args.local_rank not in [-1, 0] and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Load data features from cache or dataset file
|
||||
cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode,
|
||||
list(filter(None, args.model_name_or_path.split("/"))).pop(),
|
||||
str(args.max_seq_length)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
logger.info("Creating features from dataset file at %s", args.data_dir)
|
||||
examples = read_examples_from_file(args.data_dir, mode)
|
||||
features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
|
||||
cls_token_at_end=bool(args.model_type in ["xlnet"]),
|
||||
# xlnet has a cls token at the end
|
||||
cls_token=tokenizer.cls_token,
|
||||
cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
|
||||
sep_token=tokenizer.sep_token,
|
||||
sep_token_extra=bool(args.model_type in ["roberta"]),
|
||||
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
|
||||
pad_on_left=bool(args.model_type in ["xlnet"]),
|
||||
# pad on the left for xlnet
|
||||
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
|
||||
pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
|
||||
pad_token_label_id=pad_token_label_id
|
||||
)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
|
||||
if args.local_rank == 0 and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Convert to Tensors and build dataset
|
||||
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
|
||||
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
|
||||
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
|
||||
all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
|
||||
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
|
||||
return dataset
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
## Required parameters
|
||||
parser.add_argument("--data_dir", default=None, type=str, required=True,
|
||||
help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
|
||||
parser.add_argument("--model_type", default=None, type=str, required=True,
|
||||
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
|
||||
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
parser.add_argument("--output_dir", default=None, type=str, required=True,
|
||||
help="The output directory where the model predictions and checkpoints will be written.")
|
||||
|
||||
## Other parameters
|
||||
parser.add_argument("--labels", default="", type=str,
|
||||
help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
|
||||
parser.add_argument("--config_name", default="", type=str,
|
||||
help="Pretrained config name or path if not the same as model_name")
|
||||
parser.add_argument("--tokenizer_name", default="", type=str,
|
||||
help="Pretrained tokenizer name or path if not the same as model_name")
|
||||
parser.add_argument("--cache_dir", default="", type=str,
|
||||
help="Where do you want to store the pre-trained models downloaded from s3")
|
||||
parser.add_argument("--max_seq_length", default=128, type=int,
|
||||
help="The maximum total input sequence length after tokenization. Sequences longer "
|
||||
"than this will be truncated, sequences shorter will be padded.")
|
||||
parser.add_argument("--do_train", action="store_true",
|
||||
help="Whether to run training.")
|
||||
parser.add_argument("--do_eval", action="store_true",
|
||||
help="Whether to run eval on the dev set.")
|
||||
parser.add_argument("--do_predict", action="store_true",
|
||||
help="Whether to run predictions on the test set.")
|
||||
parser.add_argument("--evaluate_during_training", action="store_true",
|
||||
help="Whether to run evaluation during training at each logging step.")
|
||||
parser.add_argument("--do_lower_case", action="store_true",
|
||||
help="Set this flag if you are using an uncased model.")
|
||||
|
||||
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for training.")
|
||||
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for evaluation.")
|
||||
parser.add_argument("--gradient_accumulation_steps", type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--learning_rate", default=5e-5, type=float,
|
||||
help="The initial learning rate for Adam.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight decay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
help="Max gradient norm.")
|
||||
parser.add_argument("--num_train_epochs", default=3.0, type=float,
|
||||
help="Total number of training epochs to perform.")
|
||||
parser.add_argument("--max_steps", default=-1, type=int,
|
||||
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
|
||||
parser.add_argument("--warmup_steps", default=0, type=int,
|
||||
help="Linear warmup over warmup_steps.")
|
||||
|
||||
parser.add_argument("--logging_steps", type=int, default=50,
|
||||
help="Log every X updates steps.")
|
||||
parser.add_argument("--save_steps", type=int, default=50,
|
||||
help="Save checkpoint every X updates steps.")
|
||||
parser.add_argument("--eval_all_checkpoints", action="store_true",
|
||||
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
|
||||
parser.add_argument("--no_cuda", action="store_true",
|
||||
help="Avoid using CUDA when available")
|
||||
parser.add_argument("--overwrite_output_dir", action="store_true",
|
||||
help="Overwrite the content of the output directory")
|
||||
parser.add_argument("--overwrite_cache", action="store_true",
|
||||
help="Overwrite the cached training and evaluation sets")
|
||||
parser.add_argument("--seed", type=int, default=42,
|
||||
help="random seed for initialization")
|
||||
|
||||
parser.add_argument("--fp16", action="store_true",
|
||||
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
|
||||
parser.add_argument("--fp16_opt_level", type=str, default="O1",
|
||||
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
|
||||
"See details at https://nvidia.github.io/apex/amp.html")
|
||||
parser.add_argument("--local_rank", type=int, default=-1,
|
||||
help="For distributed training: local_rank")
|
||||
parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
|
||||
parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
|
||||
args = parser.parse_args()
|
||||
|
||||
if os.path.exists(args.output_dir) and os.listdir(
|
||||
args.output_dir) and args.do_train and not args.overwrite_output_dir:
|
||||
raise ValueError(
|
||||
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
|
||||
args.output_dir))
|
||||
|
||||
# Setup distant debugging if needed
|
||||
if args.server_ip and args.server_port:
|
||||
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
|
||||
import ptvsd
|
||||
print("Waiting for debugger attach")
|
||||
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
|
||||
ptvsd.wait_for_attach()
|
||||
|
||||
# Setup CUDA, GPU & distributed training
|
||||
if args.local_rank == -1 or args.no_cuda:
|
||||
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
args.n_gpu = torch.cuda.device_count()
|
||||
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
|
||||
torch.cuda.set_device(args.local_rank)
|
||||
device = torch.device("cuda", args.local_rank)
|
||||
torch.distributed.init_process_group(backend="nccl")
|
||||
args.n_gpu = 1
|
||||
args.device = device
|
||||
|
||||
# Setup logging
|
||||
logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
||||
datefmt="%m/%d/%Y %H:%M:%S",
|
||||
level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
|
||||
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
|
||||
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
|
||||
|
||||
# Set seed
|
||||
set_seed(args)
|
||||
|
||||
# Prepare CONLL-2003 task
|
||||
labels = get_labels(args.labels)
|
||||
num_labels = len(labels)
|
||||
# Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
|
||||
pad_token_label_id = CrossEntropyLoss().ignore_index
|
||||
|
||||
# Load pretrained model and tokenizer
|
||||
if args.local_rank not in [-1, 0]:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
num_labels=num_labels,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool(".ckpt" in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
model.to(args.device)
|
||||
|
||||
logger.info("Training/evaluation parameters %s", args)
|
||||
|
||||
# Training
|
||||
if args.do_train:
|
||||
train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train")
|
||||
global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id)
|
||||
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
|
||||
|
||||
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
|
||||
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
|
||||
# Create output directory if needed
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
logger.info("Saving model checkpoint to %s", args.output_dir)
|
||||
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
|
||||
# They can then be reloaded using `from_pretrained()`
|
||||
model_to_save = model.module if hasattr(model, "module") else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(args.output_dir)
|
||||
tokenizer.save_pretrained(args.output_dir)
|
||||
|
||||
# Good practice: save your training arguments together with the trained model
|
||||
torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
|
||||
|
||||
# Evaluation
|
||||
results = {}
|
||||
if args.do_eval and args.local_rank in [-1, 0]:
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev", prefix=global_step)
|
||||
if global_step:
|
||||
result = {"{}_{}".format(global_step, k): v for k, v in result.items()}
|
||||
results.update(result)
|
||||
output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
|
||||
with open(output_eval_file, "w") as writer:
|
||||
for key in sorted(results.keys()):
|
||||
writer.write("{} = {}\n".format(key, str(results[key])))
|
||||
|
||||
if args.do_predict and args.local_rank in [-1, 0]:
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
model = model_class.from_pretrained(args.output_dir)
|
||||
model.to(args.device)
|
||||
result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
|
||||
# Save results
|
||||
output_test_results_file = os.path.join(args.output_dir, "test_results.txt")
|
||||
with open(output_test_results_file, "w") as writer:
|
||||
for key in sorted(result.keys()):
|
||||
writer.write("{} = {}\n".format(key, str(result[key])))
|
||||
# Save predictions
|
||||
output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt")
|
||||
with open(output_test_predictions_file, "w") as writer:
|
||||
with open(os.path.join(args.data_dir, "test.txt"), "r") as f:
|
||||
example_id = 0
|
||||
for line in f:
|
||||
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
|
||||
writer.write(line)
|
||||
if not predictions[example_id]:
|
||||
example_id += 1
|
||||
elif predictions[example_id]:
|
||||
output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n"
|
||||
writer.write(output_line)
|
||||
else:
|
||||
logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
|
||||
|
||||
return results
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
|
||||
+107
-186
@@ -13,42 +13,43 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
|
||||
""" Finetuning the library models for question-answering on SQuAD (Bert, XLM, XLNet)."""
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
from transformers.data.processors.squad import SquadV1Processor, SquadV2Processor, SquadResult
|
||||
from transformers.data.metrics.squad_metrics import compute_predictions_logits, compute_predictions_log_probs, squad_evaluate
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
import glob
|
||||
import timeit
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler, TensorDataset)
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
BertForQuestionAnswering, BertTokenizer,
|
||||
XLMConfig, XLMForQuestionAnswering,
|
||||
XLMTokenizer, XLNetConfig,
|
||||
XLNetForQuestionAnswering,
|
||||
XLNetTokenizer,
|
||||
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer,
|
||||
AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer,
|
||||
XLMConfig, XLMForQuestionAnswering, XLMTokenizer,
|
||||
)
|
||||
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup, squad_convert_examples_to_features
|
||||
from transformers import AdamW, WarmupLinearSchedule
|
||||
|
||||
from utils_squad import (read_squad_examples, convert_examples_to_features,
|
||||
RawResult, write_predictions,
|
||||
RawResultExtended, write_predictions_extended)
|
||||
|
||||
# The follwing import is the official SQuAD evaluation script (2.0).
|
||||
# You can remove it from the dependencies if you are using this script outside of the library
|
||||
# We've added it here for automated tests (see examples/test_examples.py file)
|
||||
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -59,9 +60,7 @@ MODEL_CLASSES = {
|
||||
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
|
||||
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
|
||||
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
|
||||
'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
|
||||
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer)
|
||||
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
def set_seed(args):
|
||||
@@ -94,16 +93,14 @@ def train(args, train_dataset, model, tokenizer):
|
||||
optimizer_grouped_parameters = [
|
||||
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
|
||||
scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
||||
|
||||
# multi-gpu training (should be after apex fp16 initialization)
|
||||
@@ -126,31 +123,24 @@ def train(args, train_dataset, model, tokenizer):
|
||||
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 1
|
||||
global_step = 0
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
model.zero_grad()
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
|
||||
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
model.train()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
|
||||
inputs = {
|
||||
'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'start_positions': batch[3],
|
||||
'end_positions': batch[4]
|
||||
}
|
||||
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
|
||||
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'token_type_ids': None if args.model_type == 'xlm' else batch[2],
|
||||
'start_positions': batch[3],
|
||||
'end_positions': batch[4]}
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[5], 'p_mask': batch[6]})
|
||||
|
||||
inputs.update({'cls_index': batch[5],
|
||||
'p_mask': batch[6]})
|
||||
outputs = model(**inputs)
|
||||
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
|
||||
|
||||
@@ -162,23 +152,20 @@ def train(args, train_dataset, model, tokenizer):
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
loss.backward()
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
if args.fp16:
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
optimizer.step()
|
||||
scheduler.step() # Update learning rate schedule
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
# Log metrics
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results = evaluate(args, model, tokenizer)
|
||||
for key, value in results.items():
|
||||
@@ -187,8 +174,8 @@ def train(args, train_dataset, model, tokenizer):
|
||||
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
# Save model checkpoint
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
@@ -217,162 +204,123 @@ def evaluate(args, model, tokenizer, prefix=""):
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(dataset)
|
||||
eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
|
||||
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu evaluate
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(dataset))
|
||||
logger.info(" Batch size = %d", args.eval_batch_size)
|
||||
|
||||
all_results = []
|
||||
start_time = timeit.default_timer()
|
||||
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
model.eval()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
|
||||
with torch.no_grad():
|
||||
inputs = {
|
||||
'input_ids': batch[0],
|
||||
'attention_mask': batch[1]
|
||||
}
|
||||
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
|
||||
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'token_type_ids': None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
|
||||
}
|
||||
example_indices = batch[3]
|
||||
|
||||
# XLNet and XLM use more arguments for their predictions
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[4], 'p_mask': batch[5]})
|
||||
|
||||
inputs.update({'cls_index': batch[4],
|
||||
'p_mask': batch[5]})
|
||||
outputs = model(**inputs)
|
||||
|
||||
for i, example_index in enumerate(example_indices):
|
||||
eval_feature = features[example_index.item()]
|
||||
unique_id = int(eval_feature.unique_id)
|
||||
|
||||
output = [to_list(output[i]) for output in outputs]
|
||||
|
||||
# Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
|
||||
# models only use two.
|
||||
if len(output) >= 5:
|
||||
start_logits = output[0]
|
||||
start_top_index = output[1]
|
||||
end_logits = output[2]
|
||||
end_top_index = output[3]
|
||||
cls_logits = output[4]
|
||||
|
||||
result = SquadResult(
|
||||
unique_id, start_logits, end_logits,
|
||||
start_top_index=start_top_index,
|
||||
end_top_index=end_top_index,
|
||||
cls_logits=cls_logits
|
||||
)
|
||||
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
result = RawResultExtended(unique_id = unique_id,
|
||||
start_top_log_probs = to_list(outputs[0][i]),
|
||||
start_top_index = to_list(outputs[1][i]),
|
||||
end_top_log_probs = to_list(outputs[2][i]),
|
||||
end_top_index = to_list(outputs[3][i]),
|
||||
cls_logits = to_list(outputs[4][i]))
|
||||
else:
|
||||
start_logits, end_logits = output
|
||||
result = SquadResult(
|
||||
unique_id, start_logits, end_logits
|
||||
)
|
||||
|
||||
result = RawResult(unique_id = unique_id,
|
||||
start_logits = to_list(outputs[0][i]),
|
||||
end_logits = to_list(outputs[1][i]))
|
||||
all_results.append(result)
|
||||
|
||||
evalTime = timeit.default_timer() - start_time
|
||||
logger.info(" Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
|
||||
|
||||
# Compute predictions
|
||||
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
|
||||
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
|
||||
|
||||
if args.version_2_with_negative:
|
||||
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
|
||||
else:
|
||||
output_null_log_odds_file = None
|
||||
|
||||
# XLNet and XLM use a more complex post-processing procedure
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
|
||||
end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
|
||||
|
||||
predictions = compute_predictions_log_probs(examples, features, all_results, args.n_best_size,
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
write_predictions_extended(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file,
|
||||
start_n_top, end_n_top,
|
||||
output_nbest_file, output_null_log_odds_file, args.predict_file,
|
||||
model.config.start_n_top, model.config.end_n_top,
|
||||
args.version_2_with_negative, tokenizer, args.verbose_logging)
|
||||
else:
|
||||
predictions = compute_predictions_logits(examples, features, all_results, args.n_best_size,
|
||||
write_predictions(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, args.do_lower_case, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
|
||||
args.version_2_with_negative, args.null_score_diff_threshold)
|
||||
|
||||
# Compute the F1 and exact scores.
|
||||
results = squad_evaluate(examples, predictions)
|
||||
# Evaluate with the official SQuAD script
|
||||
evaluate_options = EVAL_OPTS(data_file=args.predict_file,
|
||||
pred_file=output_prediction_file,
|
||||
na_prob_file=output_null_log_odds_file)
|
||||
results = evaluate_on_squad(evaluate_options)
|
||||
return results
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
|
||||
if args.local_rank not in [-1, 0] and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Load data features from cache or dataset file
|
||||
input_dir = args.data_dir if args.data_dir else "."
|
||||
cached_features_file = os.path.join(input_dir, 'cached_{}_{}_{}'.format(
|
||||
input_file = args.predict_file if evaluate else args.train_file
|
||||
cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
|
||||
'dev' if evaluate else 'train',
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length))
|
||||
)
|
||||
|
||||
# Init features and dataset from cache if it exists
|
||||
str(args.max_seq_length)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features_and_dataset = torch.load(cached_features_file)
|
||||
features, dataset = features_and_dataset["features"], features_and_dataset["dataset"]
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
logger.info("Creating features from dataset file at %s", input_dir)
|
||||
|
||||
if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
|
||||
try:
|
||||
import tensorflow_datasets as tfds
|
||||
except ImportError:
|
||||
raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
|
||||
|
||||
if args.version_2_with_negative:
|
||||
logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
|
||||
|
||||
tfds_examples = tfds.load("squad")
|
||||
examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
|
||||
else:
|
||||
processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
|
||||
|
||||
if evaluate:
|
||||
examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
|
||||
else:
|
||||
examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
|
||||
|
||||
features, dataset = squad_convert_examples_to_features(
|
||||
examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=args.max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=args.max_query_length,
|
||||
is_training=not evaluate,
|
||||
return_dataset='pt'
|
||||
)
|
||||
|
||||
logger.info("Creating features from dataset file at %s", input_file)
|
||||
examples = read_squad_examples(input_file=input_file,
|
||||
is_training=not evaluate,
|
||||
version_2_with_negative=args.version_2_with_negative)
|
||||
features = convert_examples_to_features(examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=args.max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=args.max_query_length,
|
||||
is_training=not evaluate)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save({"features": features, "dataset": dataset}, cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
|
||||
if args.local_rank == 0 and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Convert to Tensors and build dataset
|
||||
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
|
||||
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
|
||||
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
|
||||
all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
|
||||
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
|
||||
if evaluate:
|
||||
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_example_index, all_cls_index, all_p_mask)
|
||||
else:
|
||||
all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
|
||||
all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_start_positions, all_end_positions,
|
||||
all_cls_index, all_p_mask)
|
||||
|
||||
if output_examples:
|
||||
return dataset, examples, features
|
||||
return dataset
|
||||
@@ -382,6 +330,10 @@ def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
## Required parameters
|
||||
parser.add_argument("--train_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for training. E.g., train-v1.1.json")
|
||||
parser.add_argument("--predict_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
|
||||
parser.add_argument("--model_type", default=None, type=str, required=True,
|
||||
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
|
||||
@@ -390,15 +342,6 @@ def main():
|
||||
help="The output directory where the model checkpoints and predictions will be written.")
|
||||
|
||||
## Other parameters
|
||||
parser.add_argument("--data_dir", default=None, type=str,
|
||||
help="The input data dir. Should contain the .json files for the task." +
|
||||
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
|
||||
parser.add_argument("--train_file", default=None, type=str,
|
||||
help="The input training file. If a data dir is specified, will look for the file there" +
|
||||
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
|
||||
parser.add_argument("--predict_file", default=None, type=str,
|
||||
help="The input evaluation file. If a data dir is specified, will look for the file there" +
|
||||
"If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
|
||||
parser.add_argument("--config_name", default="", type=str,
|
||||
help="Pretrained config name or path if not the same as model_name")
|
||||
parser.add_argument("--tokenizer_name", default="", type=str,
|
||||
@@ -437,7 +380,7 @@ def main():
|
||||
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight decay if we apply some.")
|
||||
help="Weight deay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
@@ -521,15 +464,9 @@ def main():
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
|
||||
model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
@@ -538,16 +475,6 @@ def main():
|
||||
|
||||
logger.info("Training/evaluation parameters %s", args)
|
||||
|
||||
# Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
|
||||
# Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
|
||||
# remove the need for this code, but it is still valid.
|
||||
if args.fp16:
|
||||
try:
|
||||
import apex
|
||||
apex.amp.register_half_function(torch, 'einsum')
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
|
||||
# Training
|
||||
if args.do_train:
|
||||
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
|
||||
@@ -572,7 +499,7 @@ def main():
|
||||
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
|
||||
|
||||
# Load a trained model and vocabulary that you have fine-tuned
|
||||
model = model_class.from_pretrained(args.output_dir, force_download=True)
|
||||
model = model_class.from_pretrained(args.output_dir)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
model.to(args.device)
|
||||
|
||||
@@ -580,23 +507,17 @@ def main():
|
||||
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
|
||||
results = {}
|
||||
if args.do_eval and args.local_rank in [-1, 0]:
|
||||
|
||||
if args.do_train:
|
||||
logger.info("Loading checkpoints saved during training for evaluation")
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
|
||||
else:
|
||||
logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
|
||||
checkpoints = [args.model_name_or_path]
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
|
||||
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
|
||||
for checkpoint in checkpoints:
|
||||
# Reload the model
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
model = model_class.from_pretrained(checkpoint, force_download=True)
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
|
||||
# Evaluate
|
||||
|
||||
+25
-78
@@ -1,93 +1,40 @@
|
||||
import os
|
||||
import tensorflow as tf
|
||||
import tensorflow_datasets
|
||||
from transformers import BertTokenizer, TFBertForSequenceClassification, BertConfig, glue_convert_examples_to_features, BertForSequenceClassification, glue_processors
|
||||
from transformers import BertTokenizer, TFBertForSequenceClassification, glue_convert_examples_to_features, BertForSequenceClassification
|
||||
|
||||
# script parameters
|
||||
BATCH_SIZE = 32
|
||||
EVAL_BATCH_SIZE = BATCH_SIZE * 2
|
||||
USE_XLA = False
|
||||
USE_AMP = False
|
||||
EPOCHS = 3
|
||||
|
||||
TASK = "mrpc"
|
||||
|
||||
if TASK == "sst-2":
|
||||
TFDS_TASK = "sst2"
|
||||
elif TASK == "sts-b":
|
||||
TFDS_TASK = "stsb"
|
||||
else:
|
||||
TFDS_TASK = TASK
|
||||
|
||||
num_labels = len(glue_processors[TASK]().get_labels())
|
||||
print(num_labels)
|
||||
|
||||
tf.config.optimizer.set_jit(USE_XLA)
|
||||
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP})
|
||||
|
||||
# Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression)
|
||||
config = BertConfig.from_pretrained("bert-base-cased", num_labels=num_labels)
|
||||
# Load dataset, tokenizer, model from pretrained model/vocabulary
|
||||
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
|
||||
model = TFBertForSequenceClassification.from_pretrained('bert-base-cased', config=config)
|
||||
|
||||
# Load dataset via TensorFlow Datasets
|
||||
data, info = tensorflow_datasets.load(f'glue/{TFDS_TASK}', with_info=True)
|
||||
train_examples = info.splits['train'].num_examples
|
||||
|
||||
# MNLI expects either validation_matched or validation_mismatched
|
||||
valid_examples = info.splits['validation'].num_examples
|
||||
model = TFBertForSequenceClassification.from_pretrained('bert-base-cased')
|
||||
data = tensorflow_datasets.load('glue/mrpc')
|
||||
|
||||
# Prepare dataset for GLUE as a tf.data.Dataset instance
|
||||
train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, TASK)
|
||||
|
||||
# MNLI expects either validation_matched or validation_mismatched
|
||||
valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, TASK)
|
||||
train_dataset = train_dataset.shuffle(128).batch(BATCH_SIZE).repeat(-1)
|
||||
valid_dataset = valid_dataset.batch(EVAL_BATCH_SIZE)
|
||||
train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, 128, 'mrpc')
|
||||
valid_dataset = glue_convert_examples_to_features(data['validation'], tokenizer, 128, 'mrpc')
|
||||
train_dataset = train_dataset.shuffle(100).batch(32).repeat(2)
|
||||
valid_dataset = valid_dataset.batch(64)
|
||||
|
||||
# Prepare training: Compile tf.keras model with optimizer, loss and learning rate schedule
|
||||
opt = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08)
|
||||
if USE_AMP:
|
||||
# loss scaling is currently required when using mixed precision
|
||||
opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, 'dynamic')
|
||||
|
||||
|
||||
if num_labels == 1:
|
||||
loss = tf.keras.losses.MeanSquaredError()
|
||||
else:
|
||||
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
|
||||
|
||||
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
|
||||
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
|
||||
metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
|
||||
model.compile(optimizer=opt, loss=loss, metrics=[metric])
|
||||
model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
|
||||
|
||||
# Train and evaluate using tf.keras.Model.fit()
|
||||
train_steps = train_examples//BATCH_SIZE
|
||||
valid_steps = valid_examples//EVAL_BATCH_SIZE
|
||||
history = model.fit(train_dataset, epochs=2, steps_per_epoch=115,
|
||||
validation_data=valid_dataset, validation_steps=7)
|
||||
|
||||
history = model.fit(train_dataset, epochs=EPOCHS, steps_per_epoch=train_steps,
|
||||
validation_data=valid_dataset, validation_steps=valid_steps)
|
||||
|
||||
# Save TF2 model
|
||||
os.makedirs('./save/', exist_ok=True)
|
||||
# Load the TensorFlow model in PyTorch for inspection
|
||||
model.save_pretrained('./save/')
|
||||
pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
|
||||
|
||||
if TASK == "mrpc":
|
||||
# Load the TensorFlow model in PyTorch for inspection
|
||||
# This is to demo the interoperability between the two frameworks, you don't have to
|
||||
# do this in real life (you can run the inference on the TF model).
|
||||
pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf=True)
|
||||
# Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
|
||||
sentence_0 = "This research was consistent with his findings."
|
||||
sentence_1 = "His findings were compatible with this research."
|
||||
sentence_2 = "His findings were not compatible with this research."
|
||||
inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
|
||||
inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
|
||||
|
||||
# Quickly test a few predictions - MRPC is a paraphrasing task, let's see if our model learned the task
|
||||
sentence_0 = 'This research was consistent with his findings.'
|
||||
sentence_1 = 'His findings were compatible with this research.'
|
||||
sentence_2 = 'His findings were not compatible with this research.'
|
||||
inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
|
||||
inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
|
||||
|
||||
del inputs_1["special_tokens_mask"]
|
||||
del inputs_2["special_tokens_mask"]
|
||||
|
||||
pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
|
||||
pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
|
||||
print('sentence_1 is', 'a paraphrase' if pred_1 else 'not a paraphrase', 'of sentence_0')
|
||||
print('sentence_2 is', 'a paraphrase' if pred_2 else 'not a paraphrase', 'of sentence_0')
|
||||
pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
|
||||
pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
|
||||
print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
|
||||
print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
|
||||
|
||||
@@ -1,615 +0,0 @@
|
||||
# coding=utf-8
|
||||
import datetime
|
||||
import os
|
||||
import math
|
||||
import glob
|
||||
import re
|
||||
import tensorflow as tf
|
||||
import collections
|
||||
import numpy as np
|
||||
from seqeval import metrics
|
||||
import _pickle as pickle
|
||||
from absl import logging
|
||||
from transformers import TF2_WEIGHTS_NAME, BertConfig, BertTokenizer, TFBertForTokenClassification
|
||||
from transformers import RobertaConfig, RobertaTokenizer, TFRobertaForTokenClassification
|
||||
from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForTokenClassification
|
||||
from transformers import create_optimizer, GradientAccumulator
|
||||
from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
|
||||
from fastprogress import master_bar, progress_bar
|
||||
from absl import flags
|
||||
from absl import app
|
||||
|
||||
|
||||
ALL_MODELS = sum(
|
||||
(tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
|
||||
())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
"bert": (BertConfig, TFBertForTokenClassification, BertTokenizer),
|
||||
"roberta": (RobertaConfig, TFRobertaForTokenClassification, RobertaTokenizer),
|
||||
"distilbert": (DistilBertConfig, TFDistilBertForTokenClassification, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
|
||||
flags.DEFINE_string(
|
||||
"data_dir", None,
|
||||
"The input data dir. Should contain the .conll files (or other data files) "
|
||||
"for the task.")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"model_type", None,
|
||||
"Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
|
||||
flags.DEFINE_string(
|
||||
"model_name_or_path", None,
|
||||
"Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
|
||||
flags.DEFINE_string(
|
||||
"output_dir", None,
|
||||
"The output directory where the model checkpoints will be written.")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"labels", "",
|
||||
"Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"config_name", "",
|
||||
"Pretrained config name or path if not the same as model_name")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"tokenizer_name", "",
|
||||
"Pretrained tokenizer name or path if not the same as model_name")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"cache_dir", "",
|
||||
"Where do you want to store the pre-trained models downloaded from s3")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"max_seq_length", 128,
|
||||
"The maximum total input sentence length after tokenization. "
|
||||
"Sequences longer than this will be truncated, sequences shorter "
|
||||
"will be padded.")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"tpu", None,
|
||||
"The Cloud TPU to use for training. This should be either the name "
|
||||
"used when creating the Cloud TPU, or a grpc://ip.address.of.tpu:8470 "
|
||||
"url.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"num_tpu_cores", 8,
|
||||
"Total number of TPU cores to use.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"do_train", False,
|
||||
"Whether to run training.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"do_eval", False,
|
||||
"Whether to run eval on the dev set.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"do_predict", False,
|
||||
"Whether to run predictions on the test set.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"evaluate_during_training", False,
|
||||
"Whether to run evaluation during training at each logging step.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"do_lower_case", False,
|
||||
"Set this flag if you are using an uncased model.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"per_device_train_batch_size", 8,
|
||||
"Batch size per GPU/CPU/TPU for training.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"per_device_eval_batch_size", 8,
|
||||
"Batch size per GPU/CPU/TPU for evaluation.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"gradient_accumulation_steps", 1,
|
||||
"Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
|
||||
flags.DEFINE_float(
|
||||
"learning_rate", 5e-5,
|
||||
"The initial learning rate for Adam.")
|
||||
|
||||
flags.DEFINE_float(
|
||||
"weight_decay", 0.0,
|
||||
"Weight decay if we apply some.")
|
||||
|
||||
flags.DEFINE_float(
|
||||
"adam_epsilon", 1e-8,
|
||||
"Epsilon for Adam optimizer.")
|
||||
|
||||
flags.DEFINE_float(
|
||||
"max_grad_norm", 1.0,
|
||||
"Max gradient norm.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"num_train_epochs", 3,
|
||||
"Total number of training epochs to perform.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"max_steps", -1,
|
||||
"If > 0: set total number of training steps to perform. Override num_train_epochs.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"warmup_steps", 0,
|
||||
"Linear warmup over warmup_steps.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"logging_steps", 50,
|
||||
"Log every X updates steps.")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"save_steps", 50,
|
||||
"Save checkpoint every X updates steps.")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"eval_all_checkpoints", False,
|
||||
"Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"no_cuda", False,
|
||||
"Avoid using CUDA when available")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"overwrite_output_dir", False,
|
||||
"Overwrite the content of the output directory")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"overwrite_cache", False,
|
||||
"Overwrite the cached training and evaluation sets")
|
||||
|
||||
flags.DEFINE_integer(
|
||||
"seed", 42,
|
||||
"random seed for initialization")
|
||||
|
||||
flags.DEFINE_boolean(
|
||||
"fp16", False,
|
||||
"Whether to use 16-bit (mixed) precision instead of 32-bit")
|
||||
|
||||
flags.DEFINE_string(
|
||||
"gpus", "0",
|
||||
"Comma separated list of gpus devices. If only one, switch to single "
|
||||
"gpu strategy, if None takes all the gpus available.")
|
||||
|
||||
|
||||
def train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id):
|
||||
if args['max_steps'] > 0:
|
||||
num_train_steps = args['max_steps'] * args['gradient_accumulation_steps']
|
||||
args['num_train_epochs'] = 1
|
||||
else:
|
||||
num_train_steps = math.ceil(num_train_examples / train_batch_size) // args['gradient_accumulation_steps'] * args['num_train_epochs']
|
||||
|
||||
writer = tf.summary.create_file_writer("/tmp/mylogs")
|
||||
|
||||
with strategy.scope():
|
||||
loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
|
||||
optimizer = create_optimizer(args['learning_rate'], num_train_steps, args['warmup_steps'])
|
||||
|
||||
if args['fp16']:
|
||||
optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer(optimizer, 'dynamic')
|
||||
|
||||
loss_metric = tf.keras.metrics.Mean(name='loss', dtype=tf.float32)
|
||||
gradient_accumulator = GradientAccumulator()
|
||||
|
||||
logging.info("***** Running training *****")
|
||||
logging.info(" Num examples = %d", num_train_examples)
|
||||
logging.info(" Num Epochs = %d", args['num_train_epochs'])
|
||||
logging.info(" Instantaneous batch size per device = %d", args['per_device_train_batch_size'])
|
||||
logging.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
||||
train_batch_size * args['gradient_accumulation_steps'])
|
||||
logging.info(" Gradient Accumulation steps = %d", args['gradient_accumulation_steps'])
|
||||
logging.info(" Total training steps = %d", num_train_steps)
|
||||
|
||||
model.summary()
|
||||
|
||||
@tf.function
|
||||
def apply_gradients():
|
||||
grads_and_vars = []
|
||||
|
||||
for gradient, variable in zip(gradient_accumulator.gradients, model.trainable_variables):
|
||||
if gradient is not None:
|
||||
scaled_gradient = gradient / (args['n_device'] * args['gradient_accumulation_steps'])
|
||||
grads_and_vars.append((scaled_gradient, variable))
|
||||
else:
|
||||
grads_and_vars.append((gradient, variable))
|
||||
|
||||
optimizer.apply_gradients(grads_and_vars, args['max_grad_norm'])
|
||||
gradient_accumulator.reset()
|
||||
|
||||
@tf.function
|
||||
def train_step(train_features, train_labels):
|
||||
def step_fn(train_features, train_labels):
|
||||
inputs = {'attention_mask': train_features['input_mask'], 'training': True}
|
||||
|
||||
if args['model_type'] != "distilbert":
|
||||
inputs["token_type_ids"] = train_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
|
||||
|
||||
with tf.GradientTape() as tape:
|
||||
logits = model(train_features['input_ids'], **inputs)[0]
|
||||
logits = tf.reshape(logits, (-1, len(labels) + 1))
|
||||
active_loss = tf.reshape(train_features['input_mask'], (-1,))
|
||||
active_logits = tf.boolean_mask(logits, active_loss)
|
||||
train_labels = tf.reshape(train_labels, (-1,))
|
||||
active_labels = tf.boolean_mask(train_labels, active_loss)
|
||||
cross_entropy = loss_fct(active_labels, active_logits)
|
||||
loss = tf.reduce_sum(cross_entropy) * (1.0 / train_batch_size)
|
||||
grads = tape.gradient(loss, model.trainable_variables)
|
||||
|
||||
gradient_accumulator(grads)
|
||||
|
||||
return cross_entropy
|
||||
|
||||
per_example_losses = strategy.experimental_run_v2(step_fn, args=(train_features, train_labels))
|
||||
mean_loss = strategy.reduce(tf.distribute.ReduceOp.MEAN, per_example_losses, axis=0)
|
||||
|
||||
return mean_loss
|
||||
|
||||
current_time = datetime.datetime.now()
|
||||
train_iterator = master_bar(range(args['num_train_epochs']))
|
||||
global_step = 0
|
||||
logging_loss = 0.0
|
||||
|
||||
for epoch in train_iterator:
|
||||
epoch_iterator = progress_bar(train_dataset, total=num_train_steps, parent=train_iterator, display=args['n_device'] > 1)
|
||||
step = 1
|
||||
|
||||
with strategy.scope():
|
||||
for train_features, train_labels in epoch_iterator:
|
||||
loss = train_step(train_features, train_labels)
|
||||
|
||||
if step % args['gradient_accumulation_steps'] == 0:
|
||||
strategy.experimental_run_v2(apply_gradients)
|
||||
|
||||
loss_metric(loss)
|
||||
|
||||
global_step += 1
|
||||
|
||||
if args['logging_steps'] > 0 and global_step % args['logging_steps'] == 0:
|
||||
# Log metrics
|
||||
if args['n_device'] == 1 and args['evaluate_during_training']: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
|
||||
report = metrics.classification_report(y_true, y_pred, digits=4)
|
||||
|
||||
logging.info("Eval at step " + str(global_step) + "\n" + report)
|
||||
logging.info("eval_loss: " + str(eval_loss))
|
||||
|
||||
precision = metrics.precision_score(y_true, y_pred)
|
||||
recall = metrics.recall_score(y_true, y_pred)
|
||||
f1 = metrics.f1_score(y_true, y_pred)
|
||||
|
||||
with writer.as_default():
|
||||
tf.summary.scalar("eval_loss", eval_loss, global_step)
|
||||
tf.summary.scalar("precision", precision, global_step)
|
||||
tf.summary.scalar("recall", recall, global_step)
|
||||
tf.summary.scalar("f1", f1, global_step)
|
||||
|
||||
lr = optimizer.learning_rate
|
||||
learning_rate = lr(step)
|
||||
|
||||
with writer.as_default():
|
||||
tf.summary.scalar("lr", learning_rate, global_step)
|
||||
tf.summary.scalar("loss", (loss_metric.result() - logging_loss) / args['logging_steps'], global_step)
|
||||
|
||||
logging_loss = loss_metric.result()
|
||||
|
||||
with writer.as_default():
|
||||
tf.summary.scalar("loss", loss_metric.result(), step=step)
|
||||
|
||||
if args['save_steps'] > 0 and global_step % args['save_steps'] == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args['output_dir'], "checkpoint-{}".format(global_step))
|
||||
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
|
||||
model.save_pretrained(output_dir)
|
||||
logging.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
train_iterator.child.comment = f'loss : {loss_metric.result()}'
|
||||
step += 1
|
||||
|
||||
train_iterator.write(f'loss epoch {epoch + 1}: {loss_metric.result()}')
|
||||
|
||||
loss_metric.reset_states()
|
||||
|
||||
logging.info(" Training took time = {}".format(datetime.datetime.now() - current_time))
|
||||
|
||||
|
||||
def evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode):
|
||||
eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
|
||||
eval_dataset, size = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode=mode)
|
||||
eval_dataset = strategy.experimental_distribute_dataset(eval_dataset)
|
||||
preds = None
|
||||
num_eval_steps = math.ceil(size / eval_batch_size)
|
||||
master = master_bar(range(1))
|
||||
eval_iterator = progress_bar(eval_dataset, total=num_eval_steps, parent=master, display=args['n_device'] > 1)
|
||||
loss_fct = tf.keras.losses.SparseCategoricalCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
|
||||
loss = 0.0
|
||||
|
||||
logging.info("***** Running evaluation *****")
|
||||
logging.info(" Num examples = %d", size)
|
||||
logging.info(" Batch size = %d", eval_batch_size)
|
||||
|
||||
for eval_features, eval_labels in eval_iterator:
|
||||
inputs = {'attention_mask': eval_features['input_mask'], 'training': False}
|
||||
|
||||
if args['model_type'] != "distilbert":
|
||||
inputs["token_type_ids"] = eval_features['segment_ids'] if args['model_type'] in ["bert", "xlnet"] else None
|
||||
|
||||
with strategy.scope():
|
||||
logits = model(eval_features['input_ids'], **inputs)[0]
|
||||
tmp_logits = tf.reshape(logits, (-1, len(labels) + 1))
|
||||
active_loss = tf.reshape(eval_features['input_mask'], (-1,))
|
||||
active_logits = tf.boolean_mask(tmp_logits, active_loss)
|
||||
tmp_eval_labels = tf.reshape(eval_labels, (-1,))
|
||||
active_labels = tf.boolean_mask(tmp_eval_labels, active_loss)
|
||||
cross_entropy = loss_fct(active_labels, active_logits)
|
||||
loss += tf.reduce_sum(cross_entropy) * (1.0 / eval_batch_size)
|
||||
|
||||
if preds is None:
|
||||
preds = logits.numpy()
|
||||
label_ids = eval_labels.numpy()
|
||||
else:
|
||||
preds = np.append(preds, logits.numpy(), axis=0)
|
||||
label_ids = np.append(label_ids, eval_labels.numpy(), axis=0)
|
||||
|
||||
preds = np.argmax(preds, axis=2)
|
||||
y_pred = [[] for _ in range(label_ids.shape[0])]
|
||||
y_true = [[] for _ in range(label_ids.shape[0])]
|
||||
loss = loss / num_eval_steps
|
||||
|
||||
for i in range(label_ids.shape[0]):
|
||||
for j in range(label_ids.shape[1]):
|
||||
if label_ids[i, j] != pad_token_label_id:
|
||||
y_pred[i].append(labels[preds[i, j] - 1])
|
||||
y_true[i].append(labels[label_ids[i, j] - 1])
|
||||
|
||||
return y_true, y_pred, loss.numpy()
|
||||
|
||||
|
||||
def load_cache(cached_file, max_seq_length):
|
||||
name_to_features = {
|
||||
"input_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
|
||||
"input_mask": tf.io.FixedLenFeature([max_seq_length], tf.int64),
|
||||
"segment_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
|
||||
"label_ids": tf.io.FixedLenFeature([max_seq_length], tf.int64),
|
||||
}
|
||||
|
||||
def _decode_record(record):
|
||||
example = tf.io.parse_single_example(record, name_to_features)
|
||||
features = {}
|
||||
features['input_ids'] = example['input_ids']
|
||||
features['input_mask'] = example['input_mask']
|
||||
features['segment_ids'] = example['segment_ids']
|
||||
|
||||
return features, example['label_ids']
|
||||
|
||||
d = tf.data.TFRecordDataset(cached_file)
|
||||
d = d.map(_decode_record, num_parallel_calls=4)
|
||||
count = d.reduce(0, lambda x, _: x + 1)
|
||||
|
||||
return d, count.numpy()
|
||||
|
||||
|
||||
def save_cache(features, cached_features_file):
|
||||
writer = tf.io.TFRecordWriter(cached_features_file)
|
||||
|
||||
for (ex_index, feature) in enumerate(features):
|
||||
if ex_index % 5000 == 0:
|
||||
logging.info("Writing example %d of %d" % (ex_index, len(features)))
|
||||
|
||||
def create_int_feature(values):
|
||||
f = tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))
|
||||
return f
|
||||
|
||||
record_feature = collections.OrderedDict()
|
||||
record_feature["input_ids"] = create_int_feature(feature.input_ids)
|
||||
record_feature["input_mask"] = create_int_feature(feature.input_mask)
|
||||
record_feature["segment_ids"] = create_int_feature(feature.segment_ids)
|
||||
record_feature["label_ids"] = create_int_feature(feature.label_ids)
|
||||
|
||||
tf_example = tf.train.Example(features=tf.train.Features(feature=record_feature))
|
||||
|
||||
writer.write(tf_example.SerializeToString())
|
||||
|
||||
writer.close()
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, batch_size, mode):
|
||||
drop_remainder = True if args['tpu'] or mode == 'train' else False
|
||||
|
||||
# Load data features from cache or dataset file
|
||||
cached_features_file = os.path.join(args['data_dir'], "cached_{}_{}_{}.tf_record".format(mode,
|
||||
list(filter(None, args['model_name_or_path'].split("/"))).pop(),
|
||||
str(args['max_seq_length'])))
|
||||
if os.path.exists(cached_features_file) and not args['overwrite_cache']:
|
||||
logging.info("Loading features from cached file %s", cached_features_file)
|
||||
dataset, size = load_cache(cached_features_file, args['max_seq_length'])
|
||||
else:
|
||||
logging.info("Creating features from dataset file at %s", args['data_dir'])
|
||||
examples = read_examples_from_file(args['data_dir'], mode)
|
||||
features = convert_examples_to_features(examples, labels, args['max_seq_length'], tokenizer,
|
||||
cls_token_at_end=bool(args['model_type'] in ["xlnet"]),
|
||||
# xlnet has a cls token at the end
|
||||
cls_token=tokenizer.cls_token,
|
||||
cls_token_segment_id=2 if args['model_type'] in ["xlnet"] else 0,
|
||||
sep_token=tokenizer.sep_token,
|
||||
sep_token_extra=bool(args['model_type'] in ["roberta"]),
|
||||
# roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
|
||||
pad_on_left=bool(args['model_type'] in ["xlnet"]),
|
||||
# pad on the left for xlnet
|
||||
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
|
||||
pad_token_segment_id=4 if args['model_type'] in ["xlnet"] else 0,
|
||||
pad_token_label_id=pad_token_label_id
|
||||
)
|
||||
logging.info("Saving features into cached file %s", cached_features_file)
|
||||
save_cache(features, cached_features_file)
|
||||
dataset, size = load_cache(cached_features_file, args['max_seq_length'])
|
||||
|
||||
if mode == 'train':
|
||||
dataset = dataset.repeat()
|
||||
dataset = dataset.shuffle(buffer_size=8192, seed=args['seed'])
|
||||
|
||||
dataset = dataset.batch(batch_size, drop_remainder)
|
||||
dataset = dataset.prefetch(buffer_size=batch_size)
|
||||
|
||||
return dataset, size
|
||||
|
||||
|
||||
def main(_):
|
||||
logging.set_verbosity(logging.INFO)
|
||||
args = flags.FLAGS.flag_values_dict()
|
||||
|
||||
if os.path.exists(args['output_dir']) and os.listdir(
|
||||
args['output_dir']) and args['do_train'] and not args['overwrite_output_dir']:
|
||||
raise ValueError(
|
||||
"Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
|
||||
args['output_dir']))
|
||||
|
||||
if args['fp16']:
|
||||
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
|
||||
|
||||
if args['tpu']:
|
||||
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(tpu=args['tpu'])
|
||||
tf.config.experimental_connect_to_cluster(resolver)
|
||||
tf.tpu.experimental.initialize_tpu_system(resolver)
|
||||
strategy = tf.distribute.experimental.TPUStrategy(resolver)
|
||||
args['n_device'] = args['num_tpu_cores']
|
||||
elif len(args['gpus'].split(',')) > 1:
|
||||
args['n_device'] = len([f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
|
||||
strategy = tf.distribute.MirroredStrategy(devices=[f"/gpu:{gpu}" for gpu in args['gpus'].split(',')])
|
||||
elif args['no_cuda']:
|
||||
args['n_device'] = 1
|
||||
strategy = tf.distribute.OneDeviceStrategy(device="/cpu:0")
|
||||
else:
|
||||
args['n_device'] = len(args['gpus'].split(','))
|
||||
strategy = tf.distribute.OneDeviceStrategy(device="/gpu:" + args['gpus'].split(',')[0])
|
||||
|
||||
logging.warning("n_device: %s, distributed training: %s, 16-bits training: %s",
|
||||
args['n_device'], bool(args['n_device'] > 1), args['fp16'])
|
||||
|
||||
labels = get_labels(args['labels'])
|
||||
num_labels = len(labels) + 1
|
||||
pad_token_label_id = 0
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args['model_type']]
|
||||
config = config_class.from_pretrained(args['config_name'] if args['config_name'] else args['model_name_or_path'],
|
||||
num_labels=num_labels,
|
||||
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
|
||||
|
||||
logging.info("Training/evaluation parameters %s", args)
|
||||
|
||||
# Training
|
||||
if args['do_train']:
|
||||
tokenizer = tokenizer_class.from_pretrained(args['tokenizer_name'] if args['tokenizer_name'] else args['model_name_or_path'],
|
||||
do_lower_case=args['do_lower_case'],
|
||||
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
|
||||
|
||||
with strategy.scope():
|
||||
model = model_class.from_pretrained(args['model_name_or_path'],
|
||||
from_pt=bool(".bin" in args['model_name_or_path']),
|
||||
config=config,
|
||||
cache_dir=args['cache_dir'] if args['cache_dir'] else None)
|
||||
model.layers[-1].activation = tf.keras.activations.softmax
|
||||
|
||||
train_batch_size = args['per_device_train_batch_size'] * args['n_device']
|
||||
train_dataset, num_train_examples = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, train_batch_size, mode="train")
|
||||
train_dataset = strategy.experimental_distribute_dataset(train_dataset)
|
||||
train(args, strategy, train_dataset, tokenizer, model, num_train_examples, labels, train_batch_size, pad_token_label_id)
|
||||
|
||||
if not os.path.exists(args['output_dir']):
|
||||
os.makedirs(args['output_dir'])
|
||||
|
||||
logging.info("Saving model to %s", args['output_dir'])
|
||||
|
||||
model.save_pretrained(args['output_dir'])
|
||||
tokenizer.save_pretrained(args['output_dir'])
|
||||
|
||||
# Evaluation
|
||||
if args['do_eval']:
|
||||
tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
|
||||
checkpoints = []
|
||||
results = []
|
||||
|
||||
if args['eval_all_checkpoints']:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args['output_dir'] + "/**/" + TF2_WEIGHTS_NAME, recursive=True), key=lambda f: int(''.join(filter(str.isdigit, f)) or -1)))
|
||||
|
||||
logging.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
|
||||
if len(checkpoints) == 0:
|
||||
checkpoints.append(args['output_dir'])
|
||||
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split("-")[-1] if re.match(".*checkpoint-[0-9]", checkpoint) else "final"
|
||||
|
||||
with strategy.scope():
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
|
||||
y_true, y_pred, eval_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="dev")
|
||||
report = metrics.classification_report(y_true, y_pred, digits=4)
|
||||
|
||||
if global_step:
|
||||
results.append({global_step + "_report": report, global_step + "_loss": eval_loss})
|
||||
|
||||
output_eval_file = os.path.join(args['output_dir'], "eval_results.txt")
|
||||
|
||||
with tf.io.gfile.GFile(output_eval_file, "w") as writer:
|
||||
for res in results:
|
||||
for key, val in res.items():
|
||||
if "loss" in key:
|
||||
logging.info(key + " = " + str(val))
|
||||
writer.write(key + " = " + str(val))
|
||||
writer.write("\n")
|
||||
else:
|
||||
logging.info(key)
|
||||
logging.info("\n" + report)
|
||||
writer.write(key + "\n")
|
||||
writer.write(report)
|
||||
writer.write("\n")
|
||||
|
||||
if args['do_predict']:
|
||||
tokenizer = tokenizer_class.from_pretrained(args['output_dir'], do_lower_case=args['do_lower_case'])
|
||||
model = model_class.from_pretrained(args['output_dir'])
|
||||
eval_batch_size = args['per_device_eval_batch_size'] * args['n_device']
|
||||
predict_dataset, _ = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, eval_batch_size, mode="test")
|
||||
y_true, y_pred, pred_loss = evaluate(args, strategy, model, tokenizer, labels, pad_token_label_id, mode="test")
|
||||
output_test_results_file = os.path.join(args['output_dir'], "test_results.txt")
|
||||
output_test_predictions_file = os.path.join(args['output_dir'], "test_predictions.txt")
|
||||
report = metrics.classification_report(y_true, y_pred, digits=4)
|
||||
|
||||
with tf.io.gfile.GFile(output_test_results_file, "w") as writer:
|
||||
report = metrics.classification_report(y_true, y_pred, digits=4)
|
||||
|
||||
logging.info("\n" + report)
|
||||
|
||||
writer.write(report)
|
||||
writer.write("\n\nloss = " + str(pred_loss))
|
||||
|
||||
with tf.io.gfile.GFile(output_test_predictions_file, "w") as writer:
|
||||
with tf.io.gfile.GFile(os.path.join(args['data_dir'], "test.txt"), "r") as f:
|
||||
example_id = 0
|
||||
|
||||
for line in f:
|
||||
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
|
||||
writer.write(line)
|
||||
|
||||
if not y_pred[example_id]:
|
||||
example_id += 1
|
||||
elif y_pred[example_id]:
|
||||
output_line = line.split()[0] + " " + y_pred[example_id].pop(0) + "\n"
|
||||
writer.write(output_line)
|
||||
else:
|
||||
logging.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
flags.mark_flag_as_required("data_dir")
|
||||
flags.mark_flag_as_required("output_dir")
|
||||
flags.mark_flag_as_required("model_name_or_path")
|
||||
flags.mark_flag_as_required("model_type")
|
||||
app.run(main)
|
||||
@@ -1,515 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Finetuning multi-lingual models on XNLI (Bert, DistilBERT, XLM).
|
||||
Adapted from `examples/run_glue.py`"""
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import argparse
|
||||
import glob
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME,
|
||||
BertConfig, BertForSequenceClassification, BertTokenizer,
|
||||
XLMConfig, XLMForSequenceClassification, XLMTokenizer,
|
||||
DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
|
||||
from transformers import xnli_compute_metrics as compute_metrics
|
||||
from transformers import xnli_output_modes as output_modes
|
||||
from transformers import xnli_processors as processors
|
||||
|
||||
from transformers import glue_convert_examples_to_features as convert_examples_to_features
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, DistilBertConfig, XLMConfig)), ())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
|
||||
'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
|
||||
def set_seed(args):
|
||||
random.seed(args.seed)
|
||||
np.random.seed(args.seed)
|
||||
torch.manual_seed(args.seed)
|
||||
if args.n_gpu > 0:
|
||||
torch.cuda.manual_seed_all(args.seed)
|
||||
|
||||
|
||||
def train(args, train_dataset, model, tokenizer):
|
||||
""" Train the model """
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer = SummaryWriter()
|
||||
|
||||
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
|
||||
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
|
||||
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
|
||||
|
||||
if args.max_steps > 0:
|
||||
t_total = args.max_steps
|
||||
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
|
||||
else:
|
||||
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
|
||||
|
||||
# Prepare optimizer and schedule (linear warmup and decay)
|
||||
no_decay = ['bias', 'LayerNorm.weight']
|
||||
optimizer_grouped_parameters = [
|
||||
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
||||
|
||||
# multi-gpu training (should be after apex fp16 initialization)
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Distributed training (should be after apex fp16 initialization)
|
||||
if args.local_rank != -1:
|
||||
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
|
||||
output_device=args.local_rank,
|
||||
find_unused_parameters=True)
|
||||
|
||||
# Train!
|
||||
logger.info("***** Running training *****")
|
||||
logger.info(" Num examples = %d", len(train_dataset))
|
||||
logger.info(" Num Epochs = %d", args.num_train_epochs)
|
||||
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
|
||||
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
||||
args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
|
||||
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 0
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
model.zero_grad()
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
model.train()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'labels': batch[3]}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None # XLM and DistilBERT don't use segment_ids
|
||||
outputs = model(**inputs)
|
||||
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
|
||||
|
||||
if args.n_gpu > 1:
|
||||
loss = loss.mean() # mean() to average on multi-gpu parallel training
|
||||
if args.gradient_accumulation_steps > 1:
|
||||
loss = loss / args.gradient_accumulation_steps
|
||||
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
else:
|
||||
loss.backward()
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
if args.fp16:
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
optimizer.step()
|
||||
scheduler.step() # Update learning rate schedule
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results = evaluate(args, model, tokenizer)
|
||||
for key, value in results.items():
|
||||
tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
|
||||
tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
|
||||
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(output_dir)
|
||||
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
|
||||
logger.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
epoch_iterator.close()
|
||||
break
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
train_iterator.close()
|
||||
break
|
||||
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer.close()
|
||||
|
||||
return global_step, tr_loss / global_step
|
||||
|
||||
|
||||
def evaluate(args, model, tokenizer, prefix=""):
|
||||
eval_task_names = (args.task_name,)
|
||||
eval_outputs_dirs = (args.output_dir,)
|
||||
|
||||
results = {}
|
||||
for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
|
||||
eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
|
||||
|
||||
if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(eval_output_dir)
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(eval_dataset)
|
||||
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# multi-gpu eval
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(eval_dataset))
|
||||
logger.info(" Batch size = %d", args.eval_batch_size)
|
||||
eval_loss = 0.0
|
||||
nb_eval_steps = 0
|
||||
preds = None
|
||||
out_label_ids = None
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
model.eval()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
|
||||
with torch.no_grad():
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'labels': batch[3]}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = batch[2] if args.model_type in ['bert'] else None # XLM and DistilBERT don't use segment_ids
|
||||
outputs = model(**inputs)
|
||||
tmp_eval_loss, logits = outputs[:2]
|
||||
|
||||
eval_loss += tmp_eval_loss.mean().item()
|
||||
nb_eval_steps += 1
|
||||
if preds is None:
|
||||
preds = logits.detach().cpu().numpy()
|
||||
out_label_ids = inputs['labels'].detach().cpu().numpy()
|
||||
else:
|
||||
preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
|
||||
out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)
|
||||
|
||||
eval_loss = eval_loss / nb_eval_steps
|
||||
if args.output_mode == "classification":
|
||||
preds = np.argmax(preds, axis=1)
|
||||
else:
|
||||
raise ValueError('No other `output_mode` for XNLI.')
|
||||
result = compute_metrics(eval_task, preds, out_label_ids)
|
||||
results.update(result)
|
||||
|
||||
output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
|
||||
with open(output_eval_file, "w") as writer:
|
||||
logger.info("***** Eval results {} *****".format(prefix))
|
||||
for key in sorted(result.keys()):
|
||||
logger.info(" %s = %s", key, str(result[key]))
|
||||
writer.write("%s = %s\n" % (key, str(result[key])))
|
||||
|
||||
return results
|
||||
|
||||
|
||||
def load_and_cache_examples(args, task, tokenizer, evaluate=False):
|
||||
if args.local_rank not in [-1, 0] and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
processor = processors[task](language=args.language, train_language=args.train_language)
|
||||
output_mode = output_modes[task]
|
||||
# Load data features from cache or dataset file
|
||||
cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}_{}'.format(
|
||||
'test' if evaluate else 'train',
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length),
|
||||
str(task),
|
||||
str(args.train_language if (not evaluate and args.train_language is not None) else args.language)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache:
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
logger.info("Creating features from dataset file at %s", args.data_dir)
|
||||
label_list = processor.get_labels()
|
||||
examples = processor.get_test_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
|
||||
features = convert_examples_to_features(examples,
|
||||
tokenizer,
|
||||
label_list=label_list,
|
||||
max_length=args.max_seq_length,
|
||||
output_mode=output_mode,
|
||||
pad_on_left=False,
|
||||
pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
|
||||
pad_token_segment_id=0,
|
||||
)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
|
||||
if args.local_rank == 0 and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Convert to Tensors and build dataset
|
||||
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
|
||||
all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
|
||||
all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
|
||||
if output_mode == "classification":
|
||||
all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
|
||||
else:
|
||||
raise ValueError('No other `output_mode` for XNLI.')
|
||||
|
||||
dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
|
||||
return dataset
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
## Required parameters
|
||||
parser.add_argument("--data_dir", default=None, type=str, required=True,
|
||||
help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
|
||||
parser.add_argument("--model_type", default=None, type=str, required=True,
|
||||
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
|
||||
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
parser.add_argument("--language", default=None, type=str, required=True,
|
||||
help="Evaluation language. Also train language if `train_language` is set to None.")
|
||||
parser.add_argument("--train_language", default=None, type=str,
|
||||
help="Train language if is different of the evaluation language.")
|
||||
parser.add_argument("--output_dir", default=None, type=str, required=True,
|
||||
help="The output directory where the model predictions and checkpoints will be written.")
|
||||
|
||||
## Other parameters
|
||||
parser.add_argument("--config_name", default="", type=str,
|
||||
help="Pretrained config name or path if not the same as model_name")
|
||||
parser.add_argument("--tokenizer_name", default="", type=str,
|
||||
help="Pretrained tokenizer name or path if not the same as model_name")
|
||||
parser.add_argument("--cache_dir", default="", type=str,
|
||||
help="Where do you want to store the pre-trained models downloaded from s3")
|
||||
parser.add_argument("--max_seq_length", default=128, type=int,
|
||||
help="The maximum total input sequence length after tokenization. Sequences longer "
|
||||
"than this will be truncated, sequences shorter will be padded.")
|
||||
parser.add_argument("--do_train", action='store_true',
|
||||
help="Whether to run training.")
|
||||
parser.add_argument("--do_eval", action='store_true',
|
||||
help="Whether to run eval on the test set.")
|
||||
parser.add_argument("--evaluate_during_training", action='store_true',
|
||||
help="Rul evaluation during training at each logging step.")
|
||||
parser.add_argument("--do_lower_case", action='store_true',
|
||||
help="Set this flag if you are using an uncased model.")
|
||||
|
||||
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for training.")
|
||||
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for evaluation.")
|
||||
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--learning_rate", default=5e-5, type=float,
|
||||
help="The initial learning rate for Adam.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight deay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
help="Max gradient norm.")
|
||||
parser.add_argument("--num_train_epochs", default=3.0, type=float,
|
||||
help="Total number of training epochs to perform.")
|
||||
parser.add_argument("--max_steps", default=-1, type=int,
|
||||
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
|
||||
parser.add_argument("--warmup_steps", default=0, type=int,
|
||||
help="Linear warmup over warmup_steps.")
|
||||
|
||||
parser.add_argument('--logging_steps', type=int, default=50,
|
||||
help="Log every X updates steps.")
|
||||
parser.add_argument('--save_steps', type=int, default=50,
|
||||
help="Save checkpoint every X updates steps.")
|
||||
parser.add_argument("--eval_all_checkpoints", action='store_true',
|
||||
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
|
||||
parser.add_argument("--no_cuda", action='store_true',
|
||||
help="Avoid using CUDA when available")
|
||||
parser.add_argument('--overwrite_output_dir', action='store_true',
|
||||
help="Overwrite the content of the output directory")
|
||||
parser.add_argument('--overwrite_cache', action='store_true',
|
||||
help="Overwrite the cached training and evaluation sets")
|
||||
parser.add_argument('--seed', type=int, default=42,
|
||||
help="random seed for initialization")
|
||||
|
||||
parser.add_argument('--fp16', action='store_true',
|
||||
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
|
||||
parser.add_argument('--fp16_opt_level', type=str, default='O1',
|
||||
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
|
||||
"See details at https://nvidia.github.io/apex/amp.html")
|
||||
parser.add_argument("--local_rank", type=int, default=-1,
|
||||
help="For distributed training: local_rank")
|
||||
parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
|
||||
parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
|
||||
args = parser.parse_args()
|
||||
|
||||
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
|
||||
raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
|
||||
|
||||
# Setup distant debugging if needed
|
||||
if args.server_ip and args.server_port:
|
||||
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
|
||||
import ptvsd
|
||||
print("Waiting for debugger attach")
|
||||
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
|
||||
ptvsd.wait_for_attach()
|
||||
|
||||
# Setup CUDA, GPU & distributed training
|
||||
if args.local_rank == -1 or args.no_cuda:
|
||||
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
args.n_gpu = torch.cuda.device_count()
|
||||
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
|
||||
torch.cuda.set_device(args.local_rank)
|
||||
device = torch.device("cuda", args.local_rank)
|
||||
torch.distributed.init_process_group(backend='nccl')
|
||||
args.n_gpu = 1
|
||||
args.device = device
|
||||
|
||||
# Setup logging
|
||||
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
|
||||
datefmt = '%m/%d/%Y %H:%M:%S',
|
||||
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
|
||||
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
|
||||
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
|
||||
|
||||
# Set seed
|
||||
set_seed(args)
|
||||
|
||||
# Prepare XNLI task
|
||||
args.task_name = 'xnli'
|
||||
if args.task_name not in processors:
|
||||
raise ValueError("Task not found: %s" % (args.task_name))
|
||||
processor = processors[args.task_name](language=args.language, train_language=args.train_language)
|
||||
args.output_mode = output_modes[args.task_name]
|
||||
label_list = processor.get_labels()
|
||||
num_labels = len(label_list)
|
||||
|
||||
# Load pretrained model and tokenizer
|
||||
if args.local_rank not in [-1, 0]:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
num_labels=num_labels,
|
||||
finetuning_task=args.task_name,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
model.to(args.device)
|
||||
|
||||
logger.info("Training/evaluation parameters %s", args)
|
||||
|
||||
|
||||
# Training
|
||||
if args.do_train:
|
||||
train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
|
||||
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
|
||||
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
|
||||
|
||||
|
||||
# Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
|
||||
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
|
||||
# Create output directory if needed
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
logger.info("Saving model checkpoint to %s", args.output_dir)
|
||||
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
|
||||
# They can then be reloaded using `from_pretrained()`
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(args.output_dir)
|
||||
tokenizer.save_pretrained(args.output_dir)
|
||||
|
||||
# Good practice: save your training arguments together with the trained model
|
||||
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
|
||||
|
||||
# Load a trained model and vocabulary that you have fine-tuned
|
||||
model = model_class.from_pretrained(args.output_dir)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir)
|
||||
model.to(args.device)
|
||||
|
||||
|
||||
# Evaluation
|
||||
results = {}
|
||||
if args.do_eval and args.local_rank in [-1, 0]:
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce logging
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
for checkpoint in checkpoints:
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
prefix = checkpoint.split('/')[-1] if checkpoint.find('checkpoint') != -1 else ""
|
||||
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
result = evaluate(args, model, tokenizer, prefix=prefix)
|
||||
result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
return results
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,61 +0,0 @@
|
||||
# Text Summarization with Pretrained Encoders
|
||||
|
||||
This folder contains part of the code necessary to reproduce the results on abstractive summarization from the article [Text Summarization with Pretrained Encoders](https://arxiv.org/pdf/1908.08345.pdf) by [Yang Liu](https://nlp-yang.github.io/) and [Mirella Lapata](https://homepages.inf.ed.ac.uk/mlap/). It can also be used to summarize any document.
|
||||
|
||||
The original code can be found on the Yang Liu's [github repository](https://github.com/nlpyang/PreSumm).
|
||||
|
||||
The model is loaded with the pre-trained weights for the abstractive summarization model trained on the CNN/Daily Mail dataset with an extractive and then abstractive tasks.
|
||||
|
||||
## Setup
|
||||
|
||||
```
|
||||
git clone https://github.com/huggingface/transformers && cd transformers
|
||||
pip install [--editable] .
|
||||
pip install nltk py-rouge
|
||||
cd examples/summarization
|
||||
```
|
||||
|
||||
## Reproduce the authors' results on ROUGE
|
||||
|
||||
To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:
|
||||
|
||||
```bash
|
||||
tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
|
||||
```
|
||||
|
||||
And move all the stories to the same folder. We will refer as `$DATA_PATH` the path to where you uncompressed both archive. Then run the following in the same folder as `run_summarization.py`:
|
||||
|
||||
```bash
|
||||
python run_summarization.py \
|
||||
--documents_dir $DATA_PATH \
|
||||
--summaries_output_dir $SUMMARIES_PATH \ # optional
|
||||
--no_cuda false \
|
||||
--batch_size 4 \
|
||||
--min_length 50 \
|
||||
--max_length 200 \
|
||||
--beam_size 5 \
|
||||
--alpha 0.95 \
|
||||
--block_trigram true \
|
||||
--compute_rouge true
|
||||
```
|
||||
|
||||
The scripts executes on GPU if one is available and if `no_cuda` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
|
||||
|
||||
## Summarize any text
|
||||
|
||||
Put the documents that you would like to summarize in a folder (the path to which is referred to as `$DATA_PATH` below) and run the following in the same folder as `run_summarization.py`:
|
||||
|
||||
```bash
|
||||
python run_summarization.py \
|
||||
--documents_dir $DATA_PATH \
|
||||
--summaries_output_dir $SUMMARIES_PATH \ # optional
|
||||
--no_cuda false \
|
||||
--batch_size 4 \
|
||||
--min_length 50 \
|
||||
--max_length 200 \
|
||||
--beam_size 5 \
|
||||
--alpha 0.95 \
|
||||
--block_trigram true \
|
||||
```
|
||||
|
||||
You may want to play around with `min_length`, `max_length` and `alpha` to suit your use case. If you want to compute ROUGE on another dataset you will need to tweak the stories/summaries import in `utils_summarization.py` and tell it where to fetch the reference summaries.
|
||||
@@ -1,119 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2019 The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" BertAbs configuration """
|
||||
import json
|
||||
import logging
|
||||
import sys
|
||||
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
BERTABS_FINETUNED_CONFIG_MAP = {
|
||||
"bertabs-finetuned-cnndm": "https://s3.amazonaws.com/models.huggingface.co/bert/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization-config.json",
|
||||
}
|
||||
|
||||
|
||||
class BertAbsConfig(PretrainedConfig):
|
||||
r""" Class to store the configuration of the BertAbs model.
|
||||
|
||||
Arguments:
|
||||
max_pos: int
|
||||
The maximum sequence length that this model will be used with.
|
||||
enc_layer: int
|
||||
The numner of hidden layers in the Transformer encoder.
|
||||
enc_hidden_size: int
|
||||
The size of the encoder's layers.
|
||||
enc_heads: int
|
||||
The number of attention heads for each attention layer in the encoder.
|
||||
enc_ff_size: int
|
||||
The size of the encoder's feed-forward layers.
|
||||
enc_dropout: int
|
||||
The dropout probabilitiy for all fully connected layers in the
|
||||
embeddings, layers, pooler and also the attention probabilities in
|
||||
the encoder.
|
||||
dec_layer: int
|
||||
The numner of hidden layers in the decoder.
|
||||
dec_hidden_size: int
|
||||
The size of the decoder's layers.
|
||||
dec_heads: int
|
||||
The number of attention heads for each attention layer in the decoder.
|
||||
dec_ff_size: int
|
||||
The size of the decoder's feed-forward layers.
|
||||
dec_dropout: int
|
||||
The dropout probabilitiy for all fully connected layers in the
|
||||
embeddings, layers, pooler and also the attention probabilities in
|
||||
the decoder.
|
||||
"""
|
||||
|
||||
pretrained_config_archive_map = BERTABS_FINETUNED_CONFIG_MAP
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size_or_config_json_file=30522,
|
||||
max_pos=512,
|
||||
enc_layers=6,
|
||||
enc_hidden_size=512,
|
||||
enc_heads=8,
|
||||
enc_ff_size=512,
|
||||
enc_dropout=0.2,
|
||||
dec_layers=6,
|
||||
dec_hidden_size=768,
|
||||
dec_heads=8,
|
||||
dec_ff_size=2048,
|
||||
dec_dropout=0.2,
|
||||
**kwargs,
|
||||
):
|
||||
super(BertAbsConfig, self).__init__(**kwargs)
|
||||
|
||||
if self._input_is_path_to_json(vocab_size_or_config_json_file):
|
||||
path_to_json = vocab_size_or_config_json_file
|
||||
with open(path_to_json, "r", encoding="utf-8") as reader:
|
||||
json_config = json.loads(reader.read())
|
||||
for key, value in json_config.items():
|
||||
self.__dict__[key] = value
|
||||
elif isinstance(vocab_size_or_config_json_file, int):
|
||||
self.vocab_size = vocab_size_or_config_json_file
|
||||
self.max_pos = max_pos
|
||||
|
||||
self.enc_layers = enc_layers
|
||||
self.enc_hidden_size = enc_hidden_size
|
||||
self.enc_heads = enc_heads
|
||||
self.enc_ff_size = enc_ff_size
|
||||
self.enc_dropout = enc_dropout
|
||||
|
||||
self.dec_layers = dec_layers
|
||||
self.dec_hidden_size = dec_hidden_size
|
||||
self.dec_heads = dec_heads
|
||||
self.dec_ff_size = dec_ff_size
|
||||
self.dec_dropout = dec_dropout
|
||||
else:
|
||||
raise ValueError(
|
||||
"First argument must be either a vocabulary size (int)"
|
||||
"or the path to a pretrained model config file (str)"
|
||||
)
|
||||
|
||||
def _input_is_path_to_json(self, first_argument):
|
||||
""" Checks whether the first argument passed to config
|
||||
is the path to a JSON file that contains the config.
|
||||
"""
|
||||
is_python_2 = sys.version_info[0] == 2
|
||||
if is_python_2:
|
||||
return isinstance(first_argument, unicode)
|
||||
else:
|
||||
return isinstance(first_argument, str)
|
||||
@@ -1,163 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Convert BertExtAbs's checkpoints.
|
||||
|
||||
The script looks like it is doing something trivial but it is not. The "weights"
|
||||
proposed by the authors are actually the entire model pickled. We need to load
|
||||
the model within the original codebase to be able to only save its `state_dict`.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
from collections import namedtuple
|
||||
import logging
|
||||
import torch
|
||||
|
||||
from models.model_builder import AbsSummarizer # The authors' implementation
|
||||
from model_bertabs import BertAbsSummarizer
|
||||
|
||||
from transformers import BertTokenizer
|
||||
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
SAMPLE_TEXT = 'Hello world! cécé herlolip'
|
||||
|
||||
|
||||
BertAbsConfig = namedtuple(
|
||||
"BertAbsConfig",
|
||||
["temp_dir", "large", "use_bert_emb", "finetune_bert", "encoder", "share_emb", "max_pos", "enc_layers", "enc_hidden_size", "enc_heads", "enc_ff_size", "enc_dropout", "dec_layers", "dec_hidden_size", "dec_heads", "dec_ff_size", "dec_dropout"],
|
||||
)
|
||||
|
||||
|
||||
def convert_bertabs_checkpoints(path_to_checkpoints, dump_path):
|
||||
""" Copy/paste and tweak the pre-trained weights provided by the creators
|
||||
of BertAbs for the internal architecture.
|
||||
"""
|
||||
|
||||
# Instantiate the authors' model with the pre-trained weights
|
||||
config = BertAbsConfig(
|
||||
temp_dir=".",
|
||||
finetune_bert=False,
|
||||
large=False,
|
||||
share_emb=True,
|
||||
use_bert_emb=False,
|
||||
encoder="bert",
|
||||
max_pos=512,
|
||||
enc_layers=6,
|
||||
enc_hidden_size=512,
|
||||
enc_heads=8,
|
||||
enc_ff_size=512,
|
||||
enc_dropout=0.2,
|
||||
dec_layers=6,
|
||||
dec_hidden_size=768,
|
||||
dec_heads=8,
|
||||
dec_ff_size=2048,
|
||||
dec_dropout=0.2,
|
||||
)
|
||||
checkpoints = torch.load(path_to_checkpoints, lambda storage, loc: storage)
|
||||
original = AbsSummarizer(config, torch.device("cpu"), checkpoints)
|
||||
original.eval()
|
||||
|
||||
new_model = BertAbsSummarizer(config, torch.device("cpu"))
|
||||
new_model.eval()
|
||||
|
||||
# -------------------
|
||||
# Convert the weights
|
||||
# -------------------
|
||||
|
||||
logging.info("convert the model")
|
||||
new_model.bert.load_state_dict(original.bert.state_dict())
|
||||
new_model.decoder.load_state_dict(original.decoder.state_dict())
|
||||
new_model.generator.load_state_dict(original.generator.state_dict())
|
||||
|
||||
# ----------------------------------
|
||||
# Make sure the outpus are identical
|
||||
# ----------------------------------
|
||||
|
||||
logging.info("Make sure that the models' outputs are identical")
|
||||
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
|
||||
|
||||
# prepare the model inputs
|
||||
encoder_input_ids = tokenizer.encode("This is sample éàalj'-.")
|
||||
encoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(encoder_input_ids)))
|
||||
encoder_input_ids = torch.tensor(encoder_input_ids).unsqueeze(0)
|
||||
decoder_input_ids = tokenizer.encode("This is sample 3 éàalj'-.")
|
||||
decoder_input_ids.extend([tokenizer.pad_token_id] * (512 - len(decoder_input_ids)))
|
||||
decoder_input_ids = torch.tensor(decoder_input_ids).unsqueeze(0)
|
||||
|
||||
# failsafe to make sure the weights reset does not affect the
|
||||
# loaded weights.
|
||||
assert torch.max(torch.abs(original.generator[0].weight - new_model.generator[0].weight)) == 0
|
||||
|
||||
# forward pass
|
||||
src = encoder_input_ids
|
||||
tgt = decoder_input_ids
|
||||
segs = token_type_ids = None
|
||||
clss = None
|
||||
mask_src = encoder_attention_mask = None
|
||||
mask_tgt = decoder_attention_mask = None
|
||||
mask_cls = None
|
||||
|
||||
# The original model does not apply the geneator layer immediatly but rather in
|
||||
# the beam search (where it combines softmax + linear layer). Since we already
|
||||
# apply the softmax in our generation process we only apply the linear layer here.
|
||||
# We make sure that the outputs of the full stack are identical
|
||||
output_original_model = original(src, tgt, segs, clss, mask_src, mask_tgt, mask_cls)[0]
|
||||
output_original_generator = original.generator(output_original_model)
|
||||
|
||||
output_converted_model = new_model(encoder_input_ids, decoder_input_ids, token_type_ids, encoder_attention_mask, decoder_attention_mask)[0]
|
||||
output_converted_generator = new_model.generator(output_converted_model)
|
||||
|
||||
maximum_absolute_difference = torch.max(torch.abs(output_converted_model - output_original_model)).item()
|
||||
print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
|
||||
maximum_absolute_difference = torch.max(torch.abs(output_converted_generator - output_original_generator)).item()
|
||||
print("Maximum absolute difference beween weights: {:.2f}".format(maximum_absolute_difference))
|
||||
|
||||
are_identical = torch.allclose(output_converted_model, output_original_model, atol=1e-3)
|
||||
if are_identical:
|
||||
logging.info("all weights are equal up to 1e-3")
|
||||
else:
|
||||
raise ValueError("the weights are different. The new model is likely different from the original one.")
|
||||
|
||||
# The model has been saved with torch.save(model) and this is bound to the exact
|
||||
# directory structure. We save the state_dict instead.
|
||||
logging.info("saving the model's state dictionary")
|
||||
torch.save(new_model.state_dict(), "bertabs-finetuned-cnndm-extractive-abstractive-summarization-pytorch_model.bin")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--bertabs_checkpoint_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path the official PyTorch dump.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--pytorch_dump_folder_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to the output PyTorch model.",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
convert_bertabs_checkpoints(
|
||||
args.bertabs_checkpoint_path,
|
||||
args.pytorch_dump_folder_path,
|
||||
)
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,9 +0,0 @@
|
||||
# progress bars in model download and training scripts
|
||||
tqdm
|
||||
# Accessing files from S3 directly.
|
||||
boto3
|
||||
# Used for downloading models over HTTP
|
||||
requests
|
||||
# For ROUGE
|
||||
nltk
|
||||
py-rouge
|
||||
@@ -1,344 +0,0 @@
|
||||
#! /usr/bin/python3
|
||||
import argparse
|
||||
from collections import namedtuple
|
||||
import logging
|
||||
import os
|
||||
import sys
|
||||
|
||||
import torch
|
||||
from torch.utils.data import DataLoader, SequentialSampler
|
||||
from tqdm import tqdm
|
||||
|
||||
from transformers import BertTokenizer
|
||||
|
||||
from modeling_bertabs import BertAbs, build_predictor
|
||||
|
||||
from utils_summarization import (
|
||||
SummarizationDataset,
|
||||
encode_for_summarization,
|
||||
build_mask,
|
||||
fit_to_block_size,
|
||||
compute_token_type_ids,
|
||||
)
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
|
||||
|
||||
|
||||
Batch = namedtuple(
|
||||
"Batch", ["document_names", "batch_size", "src", "segs", "mask_src", "tgt_str"]
|
||||
)
|
||||
|
||||
|
||||
def evaluate(args):
|
||||
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", do_lower_case=True)
|
||||
model = BertAbs.from_pretrained("bertabs-finetuned-cnndm")
|
||||
model.to(args.device)
|
||||
model.eval()
|
||||
|
||||
symbols = {
|
||||
"BOS": tokenizer.vocab["[unused0]"],
|
||||
"EOS": tokenizer.vocab["[unused1]"],
|
||||
"PAD": tokenizer.vocab["[PAD]"],
|
||||
}
|
||||
|
||||
if args.compute_rouge:
|
||||
reference_summaries = []
|
||||
generated_summaries = []
|
||||
|
||||
import rouge
|
||||
import nltk
|
||||
nltk.download('punkt')
|
||||
rouge_evaluator = rouge.Rouge(
|
||||
metrics=['rouge-n', 'rouge-l'],
|
||||
max_n=2,
|
||||
limit_length=True,
|
||||
length_limit=args.beam_size,
|
||||
length_limit_type='words',
|
||||
apply_avg=True,
|
||||
apply_best=False,
|
||||
alpha=0.5, # Default F1_score
|
||||
weight_factor=1.2,
|
||||
stemming=True,
|
||||
)
|
||||
|
||||
# these (unused) arguments are defined to keep the compatibility
|
||||
# with the legacy code and will be deleted in a next iteration.
|
||||
args.result_path = ""
|
||||
args.temp_dir = ""
|
||||
|
||||
data_iterator = build_data_iterator(args, tokenizer)
|
||||
predictor = build_predictor(args, tokenizer, symbols, model)
|
||||
|
||||
logger.info("***** Running evaluation *****")
|
||||
logger.info(" Number examples = %d", len(data_iterator.dataset))
|
||||
logger.info(" Batch size = %d", args.batch_size)
|
||||
logger.info("")
|
||||
logger.info("***** Beam Search parameters *****")
|
||||
logger.info(" Beam size = %d", args.beam_size)
|
||||
logger.info(" Minimum length = %d", args.min_length)
|
||||
logger.info(" Maximum length = %d", args.max_length)
|
||||
logger.info(" Alpha (length penalty) = %.2f", args.alpha)
|
||||
logger.info(" Trigrams %s be blocked", ("will" if args.block_trigram else "will NOT"))
|
||||
|
||||
for batch in tqdm(data_iterator):
|
||||
batch_data = predictor.translate_batch(batch)
|
||||
translations = predictor.from_batch(batch_data)
|
||||
summaries = [format_summary(t) for t in translations]
|
||||
save_summaries(summaries, args.summaries_output_dir, batch.document_names)
|
||||
|
||||
if args.compute_rouge:
|
||||
reference_summaries += batch.tgt_str
|
||||
generated_summaries += summaries
|
||||
|
||||
if args.compute_rouge:
|
||||
scores = rouge_evaluator.get_scores(generated_summaries, reference_summaries)
|
||||
str_scores = format_rouge_scores(scores)
|
||||
save_rouge_scores(str_scores)
|
||||
print(str_scores)
|
||||
|
||||
|
||||
def save_summaries(summaries, path, original_document_name):
|
||||
""" Write the summaries in fies that are prefixed by the original
|
||||
files' name with the `_summary` appended.
|
||||
|
||||
Attributes:
|
||||
original_document_names: List[string]
|
||||
Name of the document that was summarized.
|
||||
path: string
|
||||
Path were the summaries will be written
|
||||
summaries: List[string]
|
||||
The summaries that we produced.
|
||||
"""
|
||||
for summary, document_name in zip(summaries, original_document_name):
|
||||
# Prepare the summary file's name
|
||||
if "." in document_name:
|
||||
bare_document_name = ".".join(document_name.split(".")[:-1])
|
||||
extension = document_name.split(".")[-1]
|
||||
name = bare_document_name + "_summary." + extension
|
||||
else:
|
||||
name = document_name + "_summary"
|
||||
|
||||
file_path = os.path.join(path, name)
|
||||
with open(file_path, "w") as output:
|
||||
output.write(summary)
|
||||
|
||||
|
||||
def format_summary(translation):
|
||||
""" Transforms the output of the `from_batch` function
|
||||
into nicely formatted summaries.
|
||||
"""
|
||||
raw_summary, _, _ = translation
|
||||
summary = (
|
||||
raw_summary.replace("[unused0]", "")
|
||||
.replace("[unused3]", "")
|
||||
.replace("[PAD]", "")
|
||||
.replace("[unused1]", "")
|
||||
.replace(r" +", " ")
|
||||
.replace(" [unused2] ", ". ")
|
||||
.replace("[unused2]", "")
|
||||
.strip()
|
||||
)
|
||||
|
||||
return summary
|
||||
|
||||
|
||||
def format_rouge_scores(scores):
|
||||
return """\n
|
||||
****** ROUGE SCORES ******
|
||||
|
||||
** ROUGE 1
|
||||
F1 >> {:.3f}
|
||||
Precision >> {:.3f}
|
||||
Recall >> {:.3f}
|
||||
|
||||
** ROUGE 2
|
||||
F1 >> {:.3f}
|
||||
Precision >> {:.3f}
|
||||
Recall >> {:.3f}
|
||||
|
||||
** ROUGE L
|
||||
F1 >> {:.3f}
|
||||
Precision >> {:.3f}
|
||||
Recall >> {:.3f}""".format(
|
||||
scores['rouge-1']['f'],
|
||||
scores['rouge-1']['p'],
|
||||
scores['rouge-1']['r'],
|
||||
scores['rouge-2']['f'],
|
||||
scores['rouge-2']['p'],
|
||||
scores['rouge-2']['r'],
|
||||
scores['rouge-l']['f'],
|
||||
scores['rouge-l']['p'],
|
||||
scores['rouge-l']['r'],
|
||||
)
|
||||
|
||||
|
||||
def save_rouge_scores(str_scores):
|
||||
with open("rouge_scores.txt", "w") as output:
|
||||
output.write(str_scores)
|
||||
|
||||
|
||||
#
|
||||
# LOAD the dataset
|
||||
#
|
||||
|
||||
|
||||
def build_data_iterator(args, tokenizer):
|
||||
dataset = load_and_cache_examples(args, tokenizer)
|
||||
sampler = SequentialSampler(dataset)
|
||||
collate_fn = lambda data: collate(data, tokenizer, block_size=512, device=args.device)
|
||||
iterator = DataLoader(
|
||||
dataset, sampler=sampler, batch_size=args.batch_size, collate_fn=collate_fn,
|
||||
)
|
||||
|
||||
return iterator
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer):
|
||||
dataset = SummarizationDataset(args.documents_dir)
|
||||
return dataset
|
||||
|
||||
|
||||
def collate(data, tokenizer, block_size, device):
|
||||
""" Collate formats the data passed to the data loader.
|
||||
|
||||
In particular we tokenize the data batch after batch to avoid keeping them
|
||||
all in memory. We output the data as a namedtuple to fit the original BertAbs's
|
||||
API.
|
||||
"""
|
||||
data = [x for x in data if not len(x[1]) == 0] # remove empty_files
|
||||
names = [name for name, _, _ in data]
|
||||
summaries = [" ".join(summary_list) for _, _, summary_list in data]
|
||||
|
||||
encoded_text = [
|
||||
encode_for_summarization(story, summary, tokenizer) for _, story, summary in data
|
||||
]
|
||||
encoded_stories = torch.tensor(
|
||||
[
|
||||
fit_to_block_size(story, block_size, tokenizer.pad_token_id)
|
||||
for story, _ in encoded_text
|
||||
]
|
||||
)
|
||||
encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
|
||||
encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)
|
||||
|
||||
batch = Batch(
|
||||
document_names=names,
|
||||
batch_size=len(encoded_stories),
|
||||
src=encoded_stories.to(device),
|
||||
segs=encoder_token_type_ids.to(device),
|
||||
mask_src=encoder_mask.to(device),
|
||||
tgt_str=summaries,
|
||||
)
|
||||
|
||||
return batch
|
||||
|
||||
|
||||
def decode_summary(summary_tokens, tokenizer):
|
||||
""" Decode the summary and return it in a format
|
||||
suitable for evaluation.
|
||||
"""
|
||||
summary_tokens = summary_tokens.to("cpu").numpy()
|
||||
summary = tokenizer.decode(summary_tokens)
|
||||
sentences = summary.split(".")
|
||||
sentences = [s + "." for s in sentences]
|
||||
return sentences
|
||||
|
||||
|
||||
def main():
|
||||
""" The main function defines the interface with the users.
|
||||
"""
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
"--documents_dir",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="The folder where the documents to summarize are located.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--summaries_output_dir",
|
||||
default=None,
|
||||
type=str,
|
||||
required=False,
|
||||
help="The folder in wich the summaries should be written. Defaults to the folder where the documents are",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--compute_rouge",
|
||||
default=False,
|
||||
type=bool,
|
||||
required=False,
|
||||
help="Compute the ROUGE metrics during evaluation. Only available for the CNN/DailyMail dataset.",
|
||||
)
|
||||
# EVALUATION options
|
||||
parser.add_argument(
|
||||
"--no_cuda",
|
||||
default=False,
|
||||
type=bool,
|
||||
help="Whether to force the execution on CPU.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--batch_size", default=4, type=int, help="Batch size per GPU/CPU for training.",
|
||||
)
|
||||
# BEAM SEARCH arguments
|
||||
parser.add_argument(
|
||||
"--min_length",
|
||||
default=50,
|
||||
type=int,
|
||||
help="Minimum number of tokens for the summaries.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--max_length",
|
||||
default=200,
|
||||
type=int,
|
||||
help="Maixmum number of tokens for the summaries.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--beam_size",
|
||||
default=5,
|
||||
type=int,
|
||||
help="The number of beams to start with for each example.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--alpha",
|
||||
default=0.95,
|
||||
type=float,
|
||||
help="The value of alpha for the length penalty in the beam search.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--block_trigram",
|
||||
default=True,
|
||||
type=bool,
|
||||
help="Whether to block the existence of repeating trigrams in the text generated by beam search.",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
# Select device (distibuted not available)
|
||||
args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
|
||||
# Check the existence of directories
|
||||
if not args.summaries_output_dir:
|
||||
args.summaries_output_dir = args.documents_dir
|
||||
|
||||
if not documents_dir_is_valid(args.documents_dir):
|
||||
raise FileNotFoundError(
|
||||
"We could not find the directory you specified for the documents to summarize, or it was empty. Please specify a valid path."
|
||||
)
|
||||
os.makedirs(args.summaries_output_dir, exist_ok=True)
|
||||
|
||||
evaluate(args)
|
||||
|
||||
|
||||
def documents_dir_is_valid(path):
|
||||
if not os.path.exists(path):
|
||||
return False
|
||||
|
||||
file_list = os.listdir(path)
|
||||
if len(file_list) == 0:
|
||||
return False
|
||||
|
||||
return True
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,173 +0,0 @@
|
||||
from collections import deque
|
||||
import os
|
||||
|
||||
import torch
|
||||
from torch.utils.data import Dataset
|
||||
|
||||
|
||||
# ------------
|
||||
# Data loading
|
||||
# ------------
|
||||
|
||||
|
||||
class SummarizationDataset(Dataset):
|
||||
""" Abstracts the dataset used to train seq2seq models.
|
||||
|
||||
The class will process the documents that are located in the specified
|
||||
folder. The preprocessing will work on any document that is reasonably
|
||||
formatted. On the CNN/DailyMail dataset it will extract both the story
|
||||
and the summary.
|
||||
|
||||
CNN/Daily News:
|
||||
|
||||
The CNN/Daily News raw datasets are downloaded from [1]. The stories are
|
||||
stored in different files; the summary appears at the end of the story as
|
||||
sentences that are prefixed by the special `@highlight` line. To process
|
||||
the data, untar both datasets in the same folder, and pass the path to this
|
||||
folder as the "data_dir argument. The formatting code was inspired by [2].
|
||||
|
||||
[1] https://cs.nyu.edu/~kcho/
|
||||
[2] https://github.com/abisee/cnn-dailymail/
|
||||
"""
|
||||
|
||||
def __init__(self, path="", prefix="train"):
|
||||
""" We initialize the class by listing all the documents to summarize.
|
||||
Files are not read in memory due to the size of some datasets (like CNN/DailyMail).
|
||||
"""
|
||||
assert os.path.isdir(path)
|
||||
|
||||
self.documents = []
|
||||
story_filenames_list = os.listdir(path)
|
||||
for story_filename in story_filenames_list:
|
||||
if "summary" in story_filename:
|
||||
continue
|
||||
path_to_story = os.path.join(path, story_filename)
|
||||
if not os.path.isfile(path_to_story):
|
||||
continue
|
||||
self.documents.append(path_to_story)
|
||||
|
||||
def __len__(self):
|
||||
""" Returns the number of documents. """
|
||||
return len(self.documents)
|
||||
|
||||
def __getitem__(self, idx):
|
||||
document_path = self.documents[idx]
|
||||
document_name = document_path.split("/")[-1]
|
||||
with open(document_path, encoding="utf-8") as source:
|
||||
raw_story = source.read()
|
||||
story_lines, summary_lines = process_story(raw_story)
|
||||
return document_name, story_lines, summary_lines
|
||||
|
||||
|
||||
def process_story(raw_story):
|
||||
""" Extract the story and summary from a story file.
|
||||
|
||||
Attributes:
|
||||
raw_story (str): content of the story file as an utf-8 encoded string.
|
||||
|
||||
Raises:
|
||||
IndexError: If the stoy is empty or contains no highlights.
|
||||
"""
|
||||
nonempty_lines = list(
|
||||
filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")])
|
||||
)
|
||||
|
||||
# for some unknown reason some lines miss a period, add it
|
||||
nonempty_lines = [_add_missing_period(line) for line in nonempty_lines]
|
||||
|
||||
# gather article lines
|
||||
story_lines = []
|
||||
lines = deque(nonempty_lines)
|
||||
while True:
|
||||
try:
|
||||
element = lines.popleft()
|
||||
if element.startswith("@highlight"):
|
||||
break
|
||||
story_lines.append(element)
|
||||
except IndexError:
|
||||
# if "@highlight" is absent from the file we pop
|
||||
# all elements until there is None, raising an exception.
|
||||
return story_lines, []
|
||||
|
||||
# gather summary lines
|
||||
summary_lines = list(filter(lambda t: not t.startswith("@highlight"), lines))
|
||||
|
||||
return story_lines, summary_lines
|
||||
|
||||
|
||||
def _add_missing_period(line):
|
||||
END_TOKENS = [".", "!", "?", "...", "'", "`", '"', u"\u2019", u"\u2019", ")"]
|
||||
if line.startswith("@highlight"):
|
||||
return line
|
||||
if line[-1] in END_TOKENS:
|
||||
return line
|
||||
return line + "."
|
||||
|
||||
|
||||
# --------------------------
|
||||
# Encoding and preprocessing
|
||||
# --------------------------
|
||||
|
||||
|
||||
def fit_to_block_size(sequence, block_size, pad_token_id):
|
||||
""" Adapt the source and target sequences' lengths to the block size.
|
||||
If the sequence is shorter we append padding token to the right of the sequence.
|
||||
"""
|
||||
if len(sequence) > block_size:
|
||||
return sequence[:block_size]
|
||||
else:
|
||||
sequence.extend([pad_token_id] * (block_size - len(sequence)))
|
||||
return sequence
|
||||
|
||||
|
||||
def build_mask(sequence, pad_token_id):
|
||||
""" Builds the mask. The attention mechanism will only attend to positions
|
||||
with value 1. """
|
||||
mask = torch.ones_like(sequence)
|
||||
idx_pad_tokens = sequence == pad_token_id
|
||||
mask[idx_pad_tokens] = 0
|
||||
return mask
|
||||
|
||||
|
||||
def encode_for_summarization(story_lines, summary_lines, tokenizer):
|
||||
""" Encode the story and summary lines, and join them
|
||||
as specified in [1] by using `[SEP] [CLS]` tokens to separate
|
||||
sentences.
|
||||
"""
|
||||
story_lines_token_ids = [tokenizer.encode(line) for line in story_lines]
|
||||
story_token_ids = [
|
||||
token for sentence in story_lines_token_ids for token in sentence
|
||||
]
|
||||
summary_lines_token_ids = [tokenizer.encode(line) for line in summary_lines]
|
||||
summary_token_ids = [
|
||||
token for sentence in summary_lines_token_ids for token in sentence
|
||||
]
|
||||
|
||||
return story_token_ids, summary_token_ids
|
||||
|
||||
|
||||
def compute_token_type_ids(batch, separator_token_id):
|
||||
""" Segment embeddings as described in [1]
|
||||
|
||||
The values {0,1} were found in the repository [2].
|
||||
|
||||
Attributes:
|
||||
batch: torch.Tensor, size [batch_size, block_size]
|
||||
Batch of input.
|
||||
separator_token_id: int
|
||||
The value of the token that separates the segments.
|
||||
|
||||
[1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
|
||||
arXiv preprint arXiv:1908.08345 (2019).
|
||||
[2] https://github.com/nlpyang/PreSumm (/src/prepro/data_builder.py, commit fac1217)
|
||||
"""
|
||||
batch_embeddings = []
|
||||
for sequence in batch:
|
||||
sentence_num = -1
|
||||
embeddings = []
|
||||
for s in sequence:
|
||||
if s == separator_token_id:
|
||||
sentence_num += 1
|
||||
embeddings.append(sentence_num % 2)
|
||||
batch_embeddings.append(embeddings)
|
||||
return torch.tensor(batch_embeddings)
|
||||
@@ -1,121 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2019 HuggingFace Inc.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
|
||||
from utils_summarization import (
|
||||
compute_token_type_ids,
|
||||
fit_to_block_size,
|
||||
build_mask,
|
||||
process_story,
|
||||
)
|
||||
|
||||
|
||||
class SummarizationDataProcessingTest(unittest.TestCase):
|
||||
def setUp(self):
|
||||
self.block_size = 10
|
||||
|
||||
def test_fit_to_block_sequence_too_small(self):
|
||||
""" Pad the sequence with 0 if the sequence is smaller than the block size."""
|
||||
sequence = [1, 2, 3, 4]
|
||||
expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0]
|
||||
self.assertEqual(
|
||||
fit_to_block_size(sequence, self.block_size, 0), expected_output
|
||||
)
|
||||
|
||||
def test_fit_to_block_sequence_fit_exactly(self):
|
||||
""" Do nothing if the sequence is the right size. """
|
||||
sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
|
||||
expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
|
||||
self.assertEqual(
|
||||
fit_to_block_size(sequence, self.block_size, 0), expected_output
|
||||
)
|
||||
|
||||
def test_fit_to_block_sequence_too_big(self):
|
||||
""" Truncate the sequence if it is too long. """
|
||||
sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
|
||||
expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
|
||||
self.assertEqual(
|
||||
fit_to_block_size(sequence, self.block_size, 0), expected_output
|
||||
)
|
||||
|
||||
def test_process_story_no_highlights(self):
|
||||
""" Processing a story with no highlights returns an empty list for the summary.
|
||||
"""
|
||||
raw_story = """It was the year of Our Lord one thousand seven hundred and
|
||||
seventy-five.\n\nSpiritual revelations were conceded to England at that
|
||||
favoured period, as at this."""
|
||||
_, summary_lines = process_story(raw_story)
|
||||
self.assertEqual(summary_lines, [])
|
||||
|
||||
def test_process_empty_story(self):
|
||||
""" An empty story returns an empty collection of lines.
|
||||
"""
|
||||
raw_story = ""
|
||||
story_lines, summary_lines = process_story(raw_story)
|
||||
self.assertEqual(story_lines, [])
|
||||
self.assertEqual(summary_lines, [])
|
||||
|
||||
def test_process_story_with_missing_period(self):
|
||||
raw_story = (
|
||||
"It was the year of Our Lord one thousand seven hundred and "
|
||||
"seventy-five\n\nSpiritual revelations were conceded to England "
|
||||
"at that favoured period, as at this.\n@highlight\n\nIt was the best of times"
|
||||
)
|
||||
story_lines, summary_lines = process_story(raw_story)
|
||||
|
||||
expected_story_lines = [
|
||||
"It was the year of Our Lord one thousand seven hundred and seventy-five.",
|
||||
"Spiritual revelations were conceded to England at that favoured period, as at this.",
|
||||
]
|
||||
self.assertEqual(expected_story_lines, story_lines)
|
||||
|
||||
expected_summary_lines = ["It was the best of times."]
|
||||
self.assertEqual(expected_summary_lines, summary_lines)
|
||||
|
||||
def test_build_mask_no_padding(self):
|
||||
sequence = torch.tensor([1, 2, 3, 4])
|
||||
expected = torch.tensor([1, 1, 1, 1])
|
||||
np.testing.assert_array_equal(build_mask(sequence, 0).numpy(), expected.numpy())
|
||||
|
||||
def test_build_mask(self):
|
||||
sequence = torch.tensor([1, 2, 3, 4, 23, 23, 23])
|
||||
expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
|
||||
np.testing.assert_array_equal(
|
||||
build_mask(sequence, 23).numpy(), expected.numpy()
|
||||
)
|
||||
|
||||
def test_build_mask_with_padding_equal_to_one(self):
|
||||
sequence = torch.tensor([8, 2, 3, 4, 1, 1, 1])
|
||||
expected = torch.tensor([1, 1, 1, 1, 0, 0, 0])
|
||||
np.testing.assert_array_equal(build_mask(sequence, 1).numpy(), expected.numpy())
|
||||
|
||||
def test_compute_token_type_ids(self):
|
||||
separator = 101
|
||||
batch = torch.tensor(
|
||||
[[1, 2, 3, 4, 5, 6], [1, 2, 3, 101, 5, 6], [1, 101, 3, 4, 101, 6]]
|
||||
)
|
||||
expected = torch.tensor(
|
||||
[[1, 1, 1, 1, 1, 1], [1, 1, 1, 0, 0, 0], [1, 0, 0, 0, 1, 1]]
|
||||
)
|
||||
|
||||
result = compute_token_type_ids(batch, separator)
|
||||
np.testing.assert_array_equal(result, expected)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -72,7 +72,8 @@ class ExamplesTests(unittest.TestCase):
|
||||
logger.addHandler(stream_handler)
|
||||
|
||||
testargs = ["run_squad.py",
|
||||
"--data_dir=./examples/tests_samples/SQUAD",
|
||||
"--train_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
|
||||
"--predict_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
|
||||
"--model_name=bert-base-uncased",
|
||||
"--output_dir=./examples/tests_samples/temp_dir",
|
||||
"--max_steps=10",
|
||||
|
||||
@@ -1,140 +0,0 @@
|
||||
{
|
||||
"version": "v2.0",
|
||||
"data": [{
|
||||
"title": "Normans",
|
||||
"paragraphs": [{
|
||||
"qas": [{
|
||||
"question": "In what country is Normandy located?",
|
||||
"id": "56ddde6b9a695914005b9628",
|
||||
"answers": [{
|
||||
"text": "France",
|
||||
"answer_start": 159
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"question": "When were the Normans in Normandy?",
|
||||
"id": "56ddde6b9a695914005b9629",
|
||||
"answers": [{
|
||||
"text": "10th and 11th centuries",
|
||||
"answer_start": 94
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"question": "From which countries did the Norse originate?",
|
||||
"id": "56ddde6b9a695914005b962a",
|
||||
"answers": [{
|
||||
"text": "Denmark, Iceland and Norway",
|
||||
"answer_start": 256
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "Rollo",
|
||||
"answer_start": 308
|
||||
}],
|
||||
"question": "Who did King Charles III swear fealty to?",
|
||||
"id": "5ad39d53604f3c001a3fe8d3",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "10th century",
|
||||
"answer_start": 671
|
||||
}],
|
||||
"question": "When did the Frankish identity emerge?",
|
||||
"id": "5ad39d53604f3c001a3fe8d4",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}],
|
||||
"context": "The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France. They were descended from Norse (\"Norman\" comes from \"Norseman\") raiders and pirates from Denmark, Iceland and Norway who, under their leader Rollo, agreed to swear fealty to King Charles III of West Francia. Through generations of assimilation and mixing with the native Frankish and Roman-Gaulish populations, their descendants would gradually merge with the Carolingian-based cultures of West Francia. The distinct cultural and ethnic identity of the Normans emerged initially in the first half of the 10th century, and it continued to evolve over the succeeding centuries."
|
||||
}, {
|
||||
"qas": [{
|
||||
"question": "Who was the duke in the battle of Hastings?",
|
||||
"id": "56dddf4066d3e219004dad5f",
|
||||
"answers": [{
|
||||
"text": "William the Conqueror",
|
||||
"answer_start": 1022
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "Antioch",
|
||||
"answer_start": 1295
|
||||
}],
|
||||
"question": "What principality did William the conquerer found?",
|
||||
"id": "5ad3a266604f3c001a3fea2b",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}],
|
||||
"context": "The Norman dynasty had a major political, cultural and military impact on medieval Europe and even the Near East. The Normans were famed for their martial spirit and eventually for their Christian piety, becoming exponents of the Catholic orthodoxy into which they assimilated. They adopted the Gallo-Romance language of the Frankish land they settled, their dialect becoming known as Norman, Normaund or Norman French, an important literary language. The Duchy of Normandy, which they formed by treaty with the French crown, was a great fief of medieval France, and under Richard I of Normandy was forged into a cohesive and formidable principality in feudal tenure. The Normans are noted both for their culture, such as their unique Romanesque architecture and musical traditions, and for their significant military accomplishments and innovations. Norman adventurers founded the Kingdom of Sicily under Roger II after conquering southern Italy on the Saracens and Byzantines, and an expedition on behalf of their duke, William the Conqueror, led to the Norman conquest of England at the Battle of Hastings in 1066. Norman cultural and military influence spread from these new European centres to the Crusader states of the Near East, where their prince Bohemond I founded the Principality of Antioch in the Levant, to Scotland and Wales in Great Britain, to Ireland, and to the coasts of north Africa and the Canary Islands."
|
||||
}]
|
||||
}, {
|
||||
"title": "Computational_complexity_theory",
|
||||
"paragraphs": [{
|
||||
"qas": [{
|
||||
"question": "What branch of theoretical computer science deals with broadly classifying computational problems by difficulty and class of relationship?",
|
||||
"id": "56e16182e3433e1400422e28",
|
||||
"answers": [{
|
||||
"text": "Computational complexity theory",
|
||||
"answer_start": 0
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "algorithm",
|
||||
"answer_start": 472
|
||||
}],
|
||||
"question": "What is a manual application of mathematical steps?",
|
||||
"id": "5ad5316b5b96ef001a10ab76",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}],
|
||||
"context": "Computational complexity theory is a branch of the theory of computation in theoretical computer science that focuses on classifying computational problems according to their inherent difficulty, and relating those classes to each other. A computational problem is understood to be a task that is in principle amenable to being solved by a computer, which is equivalent to stating that the problem may be solved by mechanical application of mathematical steps, such as an algorithm."
|
||||
}, {
|
||||
"qas": [{
|
||||
"question": "What measure of a computational problem broadly defines the inherent difficulty of the solution?",
|
||||
"id": "56e16839cd28a01900c67887",
|
||||
"answers": [{
|
||||
"text": "if its solution requires significant resources",
|
||||
"answer_start": 46
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"question": "What method is used to intuitively assess or quantify the amount of resources required to solve a computational problem?",
|
||||
"id": "56e16839cd28a01900c67888",
|
||||
"answers": [{
|
||||
"text": "mathematical models of computation",
|
||||
"answer_start": 176
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"question": "What are two basic primary resources used to guage complexity?",
|
||||
"id": "56e16839cd28a01900c67889",
|
||||
"answers": [{
|
||||
"text": "time and storage",
|
||||
"answer_start": 305
|
||||
}],
|
||||
"is_impossible": false
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "the number of gates in a circuit",
|
||||
"answer_start": 436
|
||||
}],
|
||||
"question": "What unit is measured to determine circuit simplicity?",
|
||||
"id": "5ad532575b96ef001a10ab7f",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}, {
|
||||
"plausible_answers": [{
|
||||
"text": "the number of processors",
|
||||
"answer_start": 502
|
||||
}],
|
||||
"question": "What number is used in perpendicular computing?",
|
||||
"id": "5ad532575b96ef001a10ab80",
|
||||
"answers": [],
|
||||
"is_impossible": true
|
||||
}],
|
||||
"context": "A problem is regarded as inherently difficult if its solution requires significant resources, whatever the algorithm used. The theory formalizes this intuition, by introducing mathematical models of computation to study these problems and quantifying the amount of resources needed to solve them, such as time and storage. Other complexity measures are also used, such as the amount of communication (used in communication complexity), the number of gates in a circuit (used in circuit complexity) and the number of processors (used in parallel computing). One of the roles of computational complexity theory is to determine the practical limits on what computers can and cannot do."
|
||||
}]
|
||||
}]
|
||||
}
|
||||
@@ -13,7 +13,7 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
|
||||
""" BERT multiple choice fine-tuning: utilities to work with multiple choice tasks of reading comprehension """
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
@@ -26,8 +26,6 @@ import json
|
||||
import csv
|
||||
import glob
|
||||
import tqdm
|
||||
from typing import List
|
||||
from transformers import PreTrainedTokenizer
|
||||
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
@@ -36,13 +34,13 @@ logger = logging.getLogger(__name__)
|
||||
class InputExample(object):
|
||||
"""A single training/test example for multiple choice"""
|
||||
|
||||
def __init__(self, example_id, question, contexts, endings, label=None):
|
||||
def __init__(self, example_id, question, contexts, endings, label=None):
|
||||
"""Constructs a InputExample.
|
||||
|
||||
Args:
|
||||
example_id: Unique id for the example.
|
||||
contexts: list of str. The untokenized text of the first sequence (context of corresponding question).
|
||||
question: string. The untokenized text of the second sequence (question).
|
||||
question: string. The untokenized text of the second sequence (qustion).
|
||||
endings: list of str. multiple choice's options. Its length must be equal to contexts' length.
|
||||
label: (Optional) string. The label of the example. This should be
|
||||
specified for train and dev examples, but not for test examples.
|
||||
@@ -68,7 +66,7 @@ class InputFeatures(object):
|
||||
'input_mask': input_mask,
|
||||
'segment_ids': segment_ids
|
||||
}
|
||||
for input_ids, input_mask, segment_ids in choices_features
|
||||
for _, input_ids, input_mask, segment_ids in choices_features
|
||||
]
|
||||
self.label = label
|
||||
|
||||
@@ -194,7 +192,7 @@ class SwagProcessor(DataProcessor):
|
||||
return lines
|
||||
|
||||
|
||||
def _create_examples(self, lines: List[List[str]], type: str):
|
||||
def _create_examples(self, lines, type):
|
||||
"""Creates examples for the training and dev sets."""
|
||||
if type == "train" and lines[0][-1] != 'label':
|
||||
raise ValueError(
|
||||
@@ -302,18 +300,24 @@ class ArcProcessor(DataProcessor):
|
||||
return examples
|
||||
|
||||
|
||||
def convert_examples_to_features(
|
||||
examples: List[InputExample],
|
||||
label_list: List[str],
|
||||
max_length: int,
|
||||
tokenizer: PreTrainedTokenizer,
|
||||
pad_token_segment_id=0,
|
||||
pad_on_left=False,
|
||||
pad_token=0,
|
||||
mask_padding_with_zero=True,
|
||||
) -> List[InputFeatures]:
|
||||
"""
|
||||
Loads a data file into a list of `InputFeatures`
|
||||
def convert_examples_to_features(examples, label_list, max_seq_length,
|
||||
tokenizer,
|
||||
cls_token_at_end=False,
|
||||
cls_token='[CLS]',
|
||||
cls_token_segment_id=1,
|
||||
sep_token='[SEP]',
|
||||
sequence_a_segment_id=0,
|
||||
sequence_b_segment_id=1,
|
||||
sep_token_extra=False,
|
||||
pad_token_segment_id=0,
|
||||
pad_on_left=False,
|
||||
pad_token=0,
|
||||
mask_padding_with_zero=True):
|
||||
""" Loads a data file into a list of `InputBatch`s
|
||||
`cls_token_at_end` define the location of the CLS token:
|
||||
- False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
|
||||
- True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
|
||||
`cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
|
||||
"""
|
||||
|
||||
label_map = {label : i for i, label in enumerate(label_list)}
|
||||
@@ -324,70 +328,125 @@ def convert_examples_to_features(
|
||||
logger.info("Writing example %d of %d" % (ex_index, len(examples)))
|
||||
choices_features = []
|
||||
for ending_idx, (context, ending) in enumerate(zip(example.contexts, example.endings)):
|
||||
text_a = context
|
||||
tokens_a = tokenizer.tokenize(context)
|
||||
tokens_b = None
|
||||
if example.question.find("_") != -1:
|
||||
# this is for cloze question
|
||||
text_b = example.question.replace("_", ending)
|
||||
#this is for cloze question
|
||||
tokens_b = tokenizer.tokenize(example.question.replace("_", ending))
|
||||
else:
|
||||
text_b = example.question + " " + ending
|
||||
tokens_b = tokenizer.tokenize(example.question + " " + ending)
|
||||
# you can add seq token between quesiotn and ending. This does not make too much difference.
|
||||
# tokens_b = tokenizer.tokenize(example.question)
|
||||
# tokens_b += [sep_token]
|
||||
# if sep_token_extra:
|
||||
# tokens_b += [sep_token]
|
||||
# tokens_b += tokenizer.tokenize(ending)
|
||||
|
||||
inputs = tokenizer.encode_plus(
|
||||
text_a,
|
||||
text_b,
|
||||
add_special_tokens=True,
|
||||
max_length=max_length,
|
||||
)
|
||||
if 'num_truncated_tokens' in inputs and inputs['num_truncated_tokens'] > 0:
|
||||
logger.info('Attention! you are cropping tokens (swag task is ok). '
|
||||
'If you are training ARC and RACE and you are poping question + options,'
|
||||
'you need to try to use a bigger max seq length!')
|
||||
special_tokens_count = 4 if sep_token_extra else 3
|
||||
_truncate_seq_pair(tokens_a, tokens_b, max_seq_length - special_tokens_count)
|
||||
|
||||
input_ids, token_type_ids = inputs["input_ids"], inputs["token_type_ids"]
|
||||
# The convention in BERT is:
|
||||
# (a) For sequence pairs:
|
||||
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
|
||||
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
|
||||
# (b) For single sequences:
|
||||
# tokens: [CLS] the dog is hairy . [SEP]
|
||||
# type_ids: 0 0 0 0 0 0 0
|
||||
#
|
||||
# Where "type_ids" are used to indicate whether this is the first
|
||||
# sequence or the second sequence. The embedding vectors for `type=0` and
|
||||
# `type=1` were learned during pre-training and are added to the wordpiece
|
||||
# embedding vector (and position vector). This is not *strictly* necessary
|
||||
# since the [SEP] token unambiguously separates the sequences, but it makes
|
||||
# it easier for the model to learn the concept of sequences.
|
||||
#
|
||||
# For classification tasks, the first vector (corresponding to [CLS]) is
|
||||
# used as as the "sentence vector". Note that this only makes sense because
|
||||
# the entire model is fine-tuned.
|
||||
tokens = tokens_a + [sep_token]
|
||||
if sep_token_extra:
|
||||
# roberta uses an extra separator b/w pairs of sentences
|
||||
tokens += [sep_token]
|
||||
|
||||
segment_ids = [sequence_a_segment_id] * len(tokens)
|
||||
|
||||
if tokens_b:
|
||||
tokens += tokens_b + [sep_token]
|
||||
segment_ids += [sequence_b_segment_id] * (len(tokens_b) + 1)
|
||||
|
||||
if cls_token_at_end:
|
||||
tokens = tokens + [cls_token]
|
||||
segment_ids = segment_ids + [cls_token_segment_id]
|
||||
else:
|
||||
tokens = [cls_token] + tokens
|
||||
segment_ids = [cls_token_segment_id] + segment_ids
|
||||
|
||||
input_ids = tokenizer.convert_tokens_to_ids(tokens)
|
||||
|
||||
# The mask has 1 for real tokens and 0 for padding tokens. Only real
|
||||
# tokens are attended to.
|
||||
attention_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
|
||||
input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
|
||||
|
||||
# Zero-pad up to the sequence length.
|
||||
padding_length = max_length - len(input_ids)
|
||||
padding_length = max_seq_length - len(input_ids)
|
||||
if pad_on_left:
|
||||
input_ids = ([pad_token] * padding_length) + input_ids
|
||||
attention_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + attention_mask
|
||||
token_type_ids = ([pad_token_segment_id] * padding_length) + token_type_ids
|
||||
input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
|
||||
segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
|
||||
else:
|
||||
input_ids = input_ids + ([pad_token] * padding_length)
|
||||
attention_mask = attention_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
|
||||
token_type_ids = token_type_ids + ([pad_token_segment_id] * padding_length)
|
||||
|
||||
assert len(input_ids) == max_length
|
||||
assert len(attention_mask) == max_length
|
||||
assert len(token_type_ids) == max_length
|
||||
choices_features.append((input_ids, attention_mask, token_type_ids))
|
||||
|
||||
input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
|
||||
segment_ids = segment_ids + ([pad_token_segment_id] * padding_length)
|
||||
|
||||
assert len(input_ids) == max_seq_length
|
||||
assert len(input_mask) == max_seq_length
|
||||
assert len(segment_ids) == max_seq_length
|
||||
choices_features.append((tokens, input_ids, input_mask, segment_ids))
|
||||
label = label_map[example.label]
|
||||
|
||||
if ex_index < 2:
|
||||
logger.info("*** Example ***")
|
||||
logger.info("race_id: {}".format(example.example_id))
|
||||
for choice_idx, (input_ids, attention_mask, token_type_ids) in enumerate(choices_features):
|
||||
for choice_idx, (tokens, input_ids, input_mask, segment_ids) in enumerate(choices_features):
|
||||
logger.info("choice: {}".format(choice_idx))
|
||||
logger.info("tokens: {}".format(' '.join(tokens)))
|
||||
logger.info("input_ids: {}".format(' '.join(map(str, input_ids))))
|
||||
logger.info("attention_mask: {}".format(' '.join(map(str, attention_mask))))
|
||||
logger.info("token_type_ids: {}".format(' '.join(map(str, token_type_ids))))
|
||||
logger.info("input_mask: {}".format(' '.join(map(str, input_mask))))
|
||||
logger.info("segment_ids: {}".format(' '.join(map(str, segment_ids))))
|
||||
logger.info("label: {}".format(label))
|
||||
|
||||
features.append(
|
||||
InputFeatures(
|
||||
example_id=example.example_id,
|
||||
choices_features=choices_features,
|
||||
label=label,
|
||||
example_id = example.example_id,
|
||||
choices_features = choices_features,
|
||||
label = label
|
||||
)
|
||||
)
|
||||
|
||||
return features
|
||||
|
||||
|
||||
def _truncate_seq_pair(tokens_a, tokens_b, max_length):
|
||||
"""Truncates a sequence pair in place to the maximum length."""
|
||||
|
||||
# This is a simple heuristic which will always truncate the longer sequence
|
||||
# one token at a time. This makes more sense than truncating an equal percent
|
||||
# of tokens from each, since if one sequence is very short then each token
|
||||
# that's truncated likely contains more information than a longer sequence.
|
||||
|
||||
# However, since we'd better not to remove tokens of options and questions, you can choose to use a bigger
|
||||
# length or only pop from context
|
||||
while True:
|
||||
total_length = len(tokens_a) + len(tokens_b)
|
||||
if total_length <= max_length:
|
||||
break
|
||||
if len(tokens_a) > len(tokens_b):
|
||||
tokens_a.pop()
|
||||
else:
|
||||
logger.info('Attention! you are removing from token_b (swag task is ok). '
|
||||
'If you are training ARC and RACE (you are poping question + options), '
|
||||
'you need to try to use a bigger max seq length!')
|
||||
tokens_b.pop()
|
||||
|
||||
|
||||
processors = {
|
||||
@@ -397,7 +456,7 @@ processors = {
|
||||
}
|
||||
|
||||
|
||||
MULTIPLE_CHOICE_TASKS_NUM_LABELS = {
|
||||
GLUE_TASKS_NUM_LABELS = {
|
||||
"race", 4,
|
||||
"swag", 4,
|
||||
"arc", 4
|
||||
|
||||
@@ -1,212 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Named entity recognition fine-tuning: utilities to work with CoNLL-2003 task. """
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import logging
|
||||
import os
|
||||
from io import open
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class InputExample(object):
|
||||
"""A single training/test example for token classification."""
|
||||
|
||||
def __init__(self, guid, words, labels):
|
||||
"""Constructs a InputExample.
|
||||
|
||||
Args:
|
||||
guid: Unique id for the example.
|
||||
words: list. The words of the sequence.
|
||||
labels: (Optional) list. The labels for each word of the sequence. This should be
|
||||
specified for train and dev examples, but not for test examples.
|
||||
"""
|
||||
self.guid = guid
|
||||
self.words = words
|
||||
self.labels = labels
|
||||
|
||||
|
||||
class InputFeatures(object):
|
||||
"""A single set of features of data."""
|
||||
|
||||
def __init__(self, input_ids, input_mask, segment_ids, label_ids):
|
||||
self.input_ids = input_ids
|
||||
self.input_mask = input_mask
|
||||
self.segment_ids = segment_ids
|
||||
self.label_ids = label_ids
|
||||
|
||||
|
||||
def read_examples_from_file(data_dir, mode):
|
||||
file_path = os.path.join(data_dir, "{}.txt".format(mode))
|
||||
guid_index = 1
|
||||
examples = []
|
||||
with open(file_path, encoding="utf-8") as f:
|
||||
words = []
|
||||
labels = []
|
||||
for line in f:
|
||||
if line.startswith("-DOCSTART-") or line == "" or line == "\n":
|
||||
if words:
|
||||
examples.append(InputExample(guid="{}-{}".format(mode, guid_index),
|
||||
words=words,
|
||||
labels=labels))
|
||||
guid_index += 1
|
||||
words = []
|
||||
labels = []
|
||||
else:
|
||||
splits = line.split(" ")
|
||||
words.append(splits[0])
|
||||
if len(splits) > 1:
|
||||
labels.append(splits[-1].replace("\n", ""))
|
||||
else:
|
||||
# Examples could have no label for mode = "test"
|
||||
labels.append("O")
|
||||
if words:
|
||||
examples.append(InputExample(guid="%s-%d".format(mode, guid_index),
|
||||
words=words,
|
||||
labels=labels))
|
||||
return examples
|
||||
|
||||
|
||||
def convert_examples_to_features(examples,
|
||||
label_list,
|
||||
max_seq_length,
|
||||
tokenizer,
|
||||
cls_token_at_end=False,
|
||||
cls_token="[CLS]",
|
||||
cls_token_segment_id=1,
|
||||
sep_token="[SEP]",
|
||||
sep_token_extra=False,
|
||||
pad_on_left=False,
|
||||
pad_token=0,
|
||||
pad_token_segment_id=0,
|
||||
pad_token_label_id=-1,
|
||||
sequence_a_segment_id=0,
|
||||
mask_padding_with_zero=True):
|
||||
""" Loads a data file into a list of `InputBatch`s
|
||||
`cls_token_at_end` define the location of the CLS token:
|
||||
- False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
|
||||
- True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
|
||||
`cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
|
||||
"""
|
||||
|
||||
label_map = {label: i for i, label in enumerate(label_list)}
|
||||
|
||||
features = []
|
||||
for (ex_index, example) in enumerate(examples):
|
||||
if ex_index % 10000 == 0:
|
||||
logger.info("Writing example %d of %d", ex_index, len(examples))
|
||||
|
||||
tokens = []
|
||||
label_ids = []
|
||||
for word, label in zip(example.words, example.labels):
|
||||
word_tokens = tokenizer.tokenize(word)
|
||||
tokens.extend(word_tokens)
|
||||
# Use the real label id for the first token of the word, and padding ids for the remaining tokens
|
||||
label_ids.extend([label_map[label]] + [pad_token_label_id] * (len(word_tokens) - 1))
|
||||
|
||||
# Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
|
||||
special_tokens_count = 3 if sep_token_extra else 2
|
||||
if len(tokens) > max_seq_length - special_tokens_count:
|
||||
tokens = tokens[:(max_seq_length - special_tokens_count)]
|
||||
label_ids = label_ids[:(max_seq_length - special_tokens_count)]
|
||||
|
||||
# The convention in BERT is:
|
||||
# (a) For sequence pairs:
|
||||
# tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
|
||||
# type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1
|
||||
# (b) For single sequences:
|
||||
# tokens: [CLS] the dog is hairy . [SEP]
|
||||
# type_ids: 0 0 0 0 0 0 0
|
||||
#
|
||||
# Where "type_ids" are used to indicate whether this is the first
|
||||
# sequence or the second sequence. The embedding vectors for `type=0` and
|
||||
# `type=1` were learned during pre-training and are added to the wordpiece
|
||||
# embedding vector (and position vector). This is not *strictly* necessary
|
||||
# since the [SEP] token unambiguously separates the sequences, but it makes
|
||||
# it easier for the model to learn the concept of sequences.
|
||||
#
|
||||
# For classification tasks, the first vector (corresponding to [CLS]) is
|
||||
# used as as the "sentence vector". Note that this only makes sense because
|
||||
# the entire model is fine-tuned.
|
||||
tokens += [sep_token]
|
||||
label_ids += [pad_token_label_id]
|
||||
if sep_token_extra:
|
||||
# roberta uses an extra separator b/w pairs of sentences
|
||||
tokens += [sep_token]
|
||||
label_ids += [pad_token_label_id]
|
||||
segment_ids = [sequence_a_segment_id] * len(tokens)
|
||||
|
||||
if cls_token_at_end:
|
||||
tokens += [cls_token]
|
||||
label_ids += [pad_token_label_id]
|
||||
segment_ids += [cls_token_segment_id]
|
||||
else:
|
||||
tokens = [cls_token] + tokens
|
||||
label_ids = [pad_token_label_id] + label_ids
|
||||
segment_ids = [cls_token_segment_id] + segment_ids
|
||||
|
||||
input_ids = tokenizer.convert_tokens_to_ids(tokens)
|
||||
|
||||
# The mask has 1 for real tokens and 0 for padding tokens. Only real
|
||||
# tokens are attended to.
|
||||
input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
|
||||
|
||||
# Zero-pad up to the sequence length.
|
||||
padding_length = max_seq_length - len(input_ids)
|
||||
if pad_on_left:
|
||||
input_ids = ([pad_token] * padding_length) + input_ids
|
||||
input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
|
||||
segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
|
||||
label_ids = ([pad_token_label_id] * padding_length) + label_ids
|
||||
else:
|
||||
input_ids += ([pad_token] * padding_length)
|
||||
input_mask += ([0 if mask_padding_with_zero else 1] * padding_length)
|
||||
segment_ids += ([pad_token_segment_id] * padding_length)
|
||||
label_ids += ([pad_token_label_id] * padding_length)
|
||||
|
||||
assert len(input_ids) == max_seq_length
|
||||
assert len(input_mask) == max_seq_length
|
||||
assert len(segment_ids) == max_seq_length
|
||||
assert len(label_ids) == max_seq_length
|
||||
|
||||
if ex_index < 5:
|
||||
logger.info("*** Example ***")
|
||||
logger.info("guid: %s", example.guid)
|
||||
logger.info("tokens: %s", " ".join([str(x) for x in tokens]))
|
||||
logger.info("input_ids: %s", " ".join([str(x) for x in input_ids]))
|
||||
logger.info("input_mask: %s", " ".join([str(x) for x in input_mask]))
|
||||
logger.info("segment_ids: %s", " ".join([str(x) for x in segment_ids]))
|
||||
logger.info("label_ids: %s", " ".join([str(x) for x in label_ids]))
|
||||
|
||||
features.append(
|
||||
InputFeatures(input_ids=input_ids,
|
||||
input_mask=input_mask,
|
||||
segment_ids=segment_ids,
|
||||
label_ids=label_ids))
|
||||
return features
|
||||
|
||||
|
||||
def get_labels(path):
|
||||
if path:
|
||||
with open(path, "r") as f:
|
||||
labels = f.read().splitlines()
|
||||
if "O" not in labels:
|
||||
labels = ["O"] + labels
|
||||
return labels
|
||||
else:
|
||||
return ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"]
|
||||
@@ -1,6 +1,7 @@
|
||||
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX. All rights reserved.
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
@@ -13,7 +14,7 @@
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Load XXX dataset. """
|
||||
""" Load SQuAD dataset. """
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
@@ -0,0 +1,330 @@
|
||||
""" Official evaluation script for SQuAD version 2.0.
|
||||
Modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
|
||||
|
||||
In addition to basic functionality, we also compute additional statistics and
|
||||
plot precision-recall curves if an additional na_prob.json file is provided.
|
||||
This file is expected to map question ID's to the model's predicted probability
|
||||
that a question is unanswerable.
|
||||
"""
|
||||
import argparse
|
||||
import collections
|
||||
import json
|
||||
import numpy as np
|
||||
import os
|
||||
import re
|
||||
import string
|
||||
import sys
|
||||
|
||||
class EVAL_OPTS():
|
||||
def __init__(self, data_file, pred_file, out_file="",
|
||||
na_prob_file="na_prob.json", na_prob_thresh=1.0,
|
||||
out_image_dir=None, verbose=False):
|
||||
self.data_file = data_file
|
||||
self.pred_file = pred_file
|
||||
self.out_file = out_file
|
||||
self.na_prob_file = na_prob_file
|
||||
self.na_prob_thresh = na_prob_thresh
|
||||
self.out_image_dir = out_image_dir
|
||||
self.verbose = verbose
|
||||
|
||||
OPTS = None
|
||||
|
||||
def parse_args():
|
||||
parser = argparse.ArgumentParser('Official evaluation script for SQuAD version 2.0.')
|
||||
parser.add_argument('data_file', metavar='data.json', help='Input data JSON file.')
|
||||
parser.add_argument('pred_file', metavar='pred.json', help='Model predictions.')
|
||||
parser.add_argument('--out-file', '-o', metavar='eval.json',
|
||||
help='Write accuracy metrics to file (default is stdout).')
|
||||
parser.add_argument('--na-prob-file', '-n', metavar='na_prob.json',
|
||||
help='Model estimates of probability of no answer.')
|
||||
parser.add_argument('--na-prob-thresh', '-t', type=float, default=1.0,
|
||||
help='Predict "" if no-answer probability exceeds this (default = 1.0).')
|
||||
parser.add_argument('--out-image-dir', '-p', metavar='out_images', default=None,
|
||||
help='Save precision-recall curves to directory.')
|
||||
parser.add_argument('--verbose', '-v', action='store_true')
|
||||
if len(sys.argv) == 1:
|
||||
parser.print_help()
|
||||
sys.exit(1)
|
||||
return parser.parse_args()
|
||||
|
||||
def make_qid_to_has_ans(dataset):
|
||||
qid_to_has_ans = {}
|
||||
for article in dataset:
|
||||
for p in article['paragraphs']:
|
||||
for qa in p['qas']:
|
||||
qid_to_has_ans[qa['id']] = bool(qa['answers'])
|
||||
return qid_to_has_ans
|
||||
|
||||
def normalize_answer(s):
|
||||
"""Lower text and remove punctuation, articles and extra whitespace."""
|
||||
def remove_articles(text):
|
||||
regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
|
||||
return re.sub(regex, ' ', text)
|
||||
def white_space_fix(text):
|
||||
return ' '.join(text.split())
|
||||
def remove_punc(text):
|
||||
exclude = set(string.punctuation)
|
||||
return ''.join(ch for ch in text if ch not in exclude)
|
||||
def lower(text):
|
||||
return text.lower()
|
||||
return white_space_fix(remove_articles(remove_punc(lower(s))))
|
||||
|
||||
def get_tokens(s):
|
||||
if not s: return []
|
||||
return normalize_answer(s).split()
|
||||
|
||||
def compute_exact(a_gold, a_pred):
|
||||
return int(normalize_answer(a_gold) == normalize_answer(a_pred))
|
||||
|
||||
def compute_f1(a_gold, a_pred):
|
||||
gold_toks = get_tokens(a_gold)
|
||||
pred_toks = get_tokens(a_pred)
|
||||
common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
|
||||
num_same = sum(common.values())
|
||||
if len(gold_toks) == 0 or len(pred_toks) == 0:
|
||||
# If either is no-answer, then F1 is 1 if they agree, 0 otherwise
|
||||
return int(gold_toks == pred_toks)
|
||||
if num_same == 0:
|
||||
return 0
|
||||
precision = 1.0 * num_same / len(pred_toks)
|
||||
recall = 1.0 * num_same / len(gold_toks)
|
||||
f1 = (2 * precision * recall) / (precision + recall)
|
||||
return f1
|
||||
|
||||
def get_raw_scores(dataset, preds):
|
||||
exact_scores = {}
|
||||
f1_scores = {}
|
||||
for article in dataset:
|
||||
for p in article['paragraphs']:
|
||||
for qa in p['qas']:
|
||||
qid = qa['id']
|
||||
gold_answers = [a['text'] for a in qa['answers']
|
||||
if normalize_answer(a['text'])]
|
||||
if not gold_answers:
|
||||
# For unanswerable questions, only correct answer is empty string
|
||||
gold_answers = ['']
|
||||
if qid not in preds:
|
||||
print('Missing prediction for %s' % qid)
|
||||
continue
|
||||
a_pred = preds[qid]
|
||||
# Take max over all gold answers
|
||||
exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
|
||||
f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
|
||||
return exact_scores, f1_scores
|
||||
|
||||
def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
|
||||
new_scores = {}
|
||||
for qid, s in scores.items():
|
||||
pred_na = na_probs[qid] > na_prob_thresh
|
||||
if pred_na:
|
||||
new_scores[qid] = float(not qid_to_has_ans[qid])
|
||||
else:
|
||||
new_scores[qid] = s
|
||||
return new_scores
|
||||
|
||||
def make_eval_dict(exact_scores, f1_scores, qid_list=None):
|
||||
if not qid_list:
|
||||
total = len(exact_scores)
|
||||
return collections.OrderedDict([
|
||||
('exact', 100.0 * sum(exact_scores.values()) / total),
|
||||
('f1', 100.0 * sum(f1_scores.values()) / total),
|
||||
('total', total),
|
||||
])
|
||||
else:
|
||||
total = len(qid_list)
|
||||
return collections.OrderedDict([
|
||||
('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
|
||||
('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
|
||||
('total', total),
|
||||
])
|
||||
|
||||
def merge_eval(main_eval, new_eval, prefix):
|
||||
for k in new_eval:
|
||||
main_eval['%s_%s' % (prefix, k)] = new_eval[k]
|
||||
|
||||
def plot_pr_curve(precisions, recalls, out_image, title):
|
||||
plt.step(recalls, precisions, color='b', alpha=0.2, where='post')
|
||||
plt.fill_between(recalls, precisions, step='post', alpha=0.2, color='b')
|
||||
plt.xlabel('Recall')
|
||||
plt.ylabel('Precision')
|
||||
plt.xlim([0.0, 1.05])
|
||||
plt.ylim([0.0, 1.05])
|
||||
plt.title(title)
|
||||
plt.savefig(out_image)
|
||||
plt.clf()
|
||||
|
||||
def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans,
|
||||
out_image=None, title=None):
|
||||
qid_list = sorted(na_probs, key=lambda k: na_probs[k])
|
||||
true_pos = 0.0
|
||||
cur_p = 1.0
|
||||
cur_r = 0.0
|
||||
precisions = [1.0]
|
||||
recalls = [0.0]
|
||||
avg_prec = 0.0
|
||||
for i, qid in enumerate(qid_list):
|
||||
if qid_to_has_ans[qid]:
|
||||
true_pos += scores[qid]
|
||||
cur_p = true_pos / float(i+1)
|
||||
cur_r = true_pos / float(num_true_pos)
|
||||
if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i+1]]:
|
||||
# i.e., if we can put a threshold after this point
|
||||
avg_prec += cur_p * (cur_r - recalls[-1])
|
||||
precisions.append(cur_p)
|
||||
recalls.append(cur_r)
|
||||
if out_image:
|
||||
plot_pr_curve(precisions, recalls, out_image, title)
|
||||
return {'ap': 100.0 * avg_prec}
|
||||
|
||||
def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs,
|
||||
qid_to_has_ans, out_image_dir):
|
||||
if out_image_dir and not os.path.exists(out_image_dir):
|
||||
os.makedirs(out_image_dir)
|
||||
num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
|
||||
if num_true_pos == 0:
|
||||
return
|
||||
pr_exact = make_precision_recall_eval(
|
||||
exact_raw, na_probs, num_true_pos, qid_to_has_ans,
|
||||
out_image=os.path.join(out_image_dir, 'pr_exact.png'),
|
||||
title='Precision-Recall curve for Exact Match score')
|
||||
pr_f1 = make_precision_recall_eval(
|
||||
f1_raw, na_probs, num_true_pos, qid_to_has_ans,
|
||||
out_image=os.path.join(out_image_dir, 'pr_f1.png'),
|
||||
title='Precision-Recall curve for F1 score')
|
||||
oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
|
||||
pr_oracle = make_precision_recall_eval(
|
||||
oracle_scores, na_probs, num_true_pos, qid_to_has_ans,
|
||||
out_image=os.path.join(out_image_dir, 'pr_oracle.png'),
|
||||
title='Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)')
|
||||
merge_eval(main_eval, pr_exact, 'pr_exact')
|
||||
merge_eval(main_eval, pr_f1, 'pr_f1')
|
||||
merge_eval(main_eval, pr_oracle, 'pr_oracle')
|
||||
|
||||
def histogram_na_prob(na_probs, qid_list, image_dir, name):
|
||||
if not qid_list:
|
||||
return
|
||||
x = [na_probs[k] for k in qid_list]
|
||||
weights = np.ones_like(x) / float(len(x))
|
||||
plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
|
||||
plt.xlabel('Model probability of no-answer')
|
||||
plt.ylabel('Proportion of dataset')
|
||||
plt.title('Histogram of no-answer probability: %s' % name)
|
||||
plt.savefig(os.path.join(image_dir, 'na_prob_hist_%s.png' % name))
|
||||
plt.clf()
|
||||
|
||||
def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
|
||||
num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
|
||||
cur_score = num_no_ans
|
||||
best_score = cur_score
|
||||
best_thresh = 0.0
|
||||
qid_list = sorted(na_probs, key=lambda k: na_probs[k])
|
||||
for i, qid in enumerate(qid_list):
|
||||
if qid not in scores: continue
|
||||
if qid_to_has_ans[qid]:
|
||||
diff = scores[qid]
|
||||
else:
|
||||
if preds[qid]:
|
||||
diff = -1
|
||||
else:
|
||||
diff = 0
|
||||
cur_score += diff
|
||||
if cur_score > best_score:
|
||||
best_score = cur_score
|
||||
best_thresh = na_probs[qid]
|
||||
return 100.0 * best_score / len(scores), best_thresh
|
||||
|
||||
def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
|
||||
num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
|
||||
cur_score = num_no_ans
|
||||
best_score = cur_score
|
||||
best_thresh = 0.0
|
||||
qid_list = sorted(na_probs, key=lambda k: na_probs[k])
|
||||
for i, qid in enumerate(qid_list):
|
||||
if qid not in scores: continue
|
||||
if qid_to_has_ans[qid]:
|
||||
diff = scores[qid]
|
||||
else:
|
||||
if preds[qid]:
|
||||
diff = -1
|
||||
else:
|
||||
diff = 0
|
||||
cur_score += diff
|
||||
if cur_score > best_score:
|
||||
best_score = cur_score
|
||||
best_thresh = na_probs[qid]
|
||||
|
||||
has_ans_score, has_ans_cnt = 0, 0
|
||||
for qid in qid_list:
|
||||
if not qid_to_has_ans[qid]: continue
|
||||
has_ans_cnt += 1
|
||||
|
||||
if qid not in scores: continue
|
||||
has_ans_score += scores[qid]
|
||||
|
||||
return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
|
||||
|
||||
def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
|
||||
best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
|
||||
best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
|
||||
main_eval['best_exact'] = best_exact
|
||||
main_eval['best_exact_thresh'] = exact_thresh
|
||||
main_eval['best_f1'] = best_f1
|
||||
main_eval['best_f1_thresh'] = f1_thresh
|
||||
|
||||
def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
|
||||
best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(preds, exact_raw, na_probs, qid_to_has_ans)
|
||||
best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(preds, f1_raw, na_probs, qid_to_has_ans)
|
||||
main_eval['best_exact'] = best_exact
|
||||
main_eval['best_exact_thresh'] = exact_thresh
|
||||
main_eval['best_f1'] = best_f1
|
||||
main_eval['best_f1_thresh'] = f1_thresh
|
||||
main_eval['has_ans_exact'] = has_ans_exact
|
||||
main_eval['has_ans_f1'] = has_ans_f1
|
||||
|
||||
def main(OPTS):
|
||||
with open(OPTS.data_file) as f:
|
||||
dataset_json = json.load(f)
|
||||
dataset = dataset_json['data']
|
||||
with open(OPTS.pred_file) as f:
|
||||
preds = json.load(f)
|
||||
if OPTS.na_prob_file:
|
||||
with open(OPTS.na_prob_file) as f:
|
||||
na_probs = json.load(f)
|
||||
else:
|
||||
na_probs = {k: 0.0 for k in preds}
|
||||
qid_to_has_ans = make_qid_to_has_ans(dataset) # maps qid to True/False
|
||||
has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
|
||||
no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
|
||||
exact_raw, f1_raw = get_raw_scores(dataset, preds)
|
||||
exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans,
|
||||
OPTS.na_prob_thresh)
|
||||
f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans,
|
||||
OPTS.na_prob_thresh)
|
||||
out_eval = make_eval_dict(exact_thresh, f1_thresh)
|
||||
if has_ans_qids:
|
||||
has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
|
||||
merge_eval(out_eval, has_ans_eval, 'HasAns')
|
||||
if no_ans_qids:
|
||||
no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
|
||||
merge_eval(out_eval, no_ans_eval, 'NoAns')
|
||||
if OPTS.na_prob_file:
|
||||
find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
|
||||
if OPTS.na_prob_file and OPTS.out_image_dir:
|
||||
run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs,
|
||||
qid_to_has_ans, OPTS.out_image_dir)
|
||||
histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, 'hasAns')
|
||||
histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, 'noAns')
|
||||
if OPTS.out_file:
|
||||
with open(OPTS.out_file, 'w') as f:
|
||||
json.dump(out_eval, f)
|
||||
else:
|
||||
print(json.dumps(out_eval, indent=2))
|
||||
return out_eval
|
||||
|
||||
if __name__ == '__main__':
|
||||
OPTS = parse_args()
|
||||
if OPTS.out_image_dir:
|
||||
import matplotlib
|
||||
matplotlib.use('Agg')
|
||||
import matplotlib.pyplot as plt
|
||||
main(OPTS)
|
||||
@@ -1,48 +0,0 @@
|
||||
absl-py==0.8.0
|
||||
astor==0.8.0
|
||||
atomicwrites==1.3.0
|
||||
attrs==19.2.0
|
||||
boto3==1.9.243
|
||||
botocore==1.12.243
|
||||
certifi==2019.9.11
|
||||
chardet==3.0.4
|
||||
Click==7.0
|
||||
docutils==0.15.2
|
||||
gast==0.2.2
|
||||
google-pasta==0.1.7
|
||||
grpcio==1.24.1
|
||||
h5py==2.10.0
|
||||
idna==2.8
|
||||
importlib-metadata==0.23
|
||||
jmespath==0.9.4
|
||||
joblib==0.14.0
|
||||
Keras-Applications==1.0.8
|
||||
Keras-Preprocessing==1.1.0
|
||||
Markdown==3.1.1
|
||||
more-itertools==7.2.0
|
||||
numpy==1.17.2
|
||||
opt-einsum==3.1.0
|
||||
packaging==19.2
|
||||
pluggy==0.13.0
|
||||
protobuf==3.10.0
|
||||
py==1.8.0
|
||||
pyparsing==2.4.2
|
||||
pytest==5.2.1
|
||||
python-dateutil==2.8.0
|
||||
regex==2019.8.19
|
||||
requests==2.22.0
|
||||
s3transfer==0.2.1
|
||||
sacremoses==0.0.35
|
||||
sentencepiece==0.1.83
|
||||
six==1.12.0
|
||||
tensorboard==2.0.0
|
||||
tensorflow==2.0.0
|
||||
tensorflow-estimator==2.0.0
|
||||
termcolor==1.1.0
|
||||
torch==1.2.0
|
||||
tqdm==4.36.1
|
||||
urllib3==1.25.6
|
||||
wcwidth==0.1.7
|
||||
Werkzeug==0.16.0
|
||||
wrapt==1.11.2
|
||||
zipp==0.6.0
|
||||
@@ -3,7 +3,7 @@ Simple check list from AllenNLP repo: https://github.com/allenai/allennlp/blob/m
|
||||
|
||||
To create the package for pypi.
|
||||
|
||||
1. Change the version in __init__.py, setup.py as well as docs/source/conf.py.
|
||||
1. Change the version in __init__.py and setup.py.
|
||||
|
||||
2. Commit these changes with the message: "Release: VERSION"
|
||||
|
||||
@@ -36,15 +36,9 @@ To create the package for pypi.
|
||||
from io import open
|
||||
from setuptools import find_packages, setup
|
||||
|
||||
|
||||
extras = {
|
||||
'serving': ['uvicorn', 'fastapi']
|
||||
}
|
||||
extras['all'] = [package for package in extras.values()]
|
||||
|
||||
setup(
|
||||
name="transformers",
|
||||
version="2.2.2",
|
||||
version="2.0.0",
|
||||
author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
|
||||
author_email="thomas@huggingface.co",
|
||||
description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
|
||||
@@ -67,11 +61,8 @@ setup(
|
||||
"transformers=transformers.__main__:main",
|
||||
]
|
||||
},
|
||||
extras_require=extras,
|
||||
scripts=[
|
||||
'transformers-cli'
|
||||
],
|
||||
# python_requires='>=3.5.0',
|
||||
tests_require=['pytest'],
|
||||
classifiers=[
|
||||
'Intended Audience :: Science/Research',
|
||||
'License :: OSI Approved :: Apache Software License',
|
||||
|
||||
@@ -1,5 +0,0 @@
|
||||
# How to add a new example script in 🤗Transformers
|
||||
|
||||
This folder provide a template for adding a new example script implementing a training or inference task with the models in the 🤗Transformers library.
|
||||
|
||||
Currently only examples for PyTorch are provided which are adaptations of the library's SQuAD examples which implement single-GPU and distributed training with gradient accumulation and mixed-precision (using NVIDIA's apex library) to cover a reasonable range of use cases.
|
||||
@@ -1,559 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Finetuning the library models for task XXX."""
|
||||
|
||||
from __future__ import absolute_import, division, print_function
|
||||
|
||||
import argparse
|
||||
import logging
|
||||
import os
|
||||
import random
|
||||
import glob
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
|
||||
TensorDataset)
|
||||
from torch.utils.data.distributed import DistributedSampler
|
||||
|
||||
try:
|
||||
from torch.utils.tensorboard import SummaryWriter
|
||||
except:
|
||||
from tensorboardX import SummaryWriter
|
||||
|
||||
from tqdm import tqdm, trange
|
||||
|
||||
from transformers import (WEIGHTS_NAME, BertConfig,
|
||||
BertForQuestionAnswering, BertTokenizer,
|
||||
XLMConfig, XLMForQuestionAnswering,
|
||||
XLMTokenizer, XLNetConfig,
|
||||
XLNetForQuestionAnswering,
|
||||
XLNetTokenizer,
|
||||
DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
|
||||
from transformers import AdamW, get_linear_schedule_with_warmup
|
||||
|
||||
from utils_squad import (read_squad_examples, convert_examples_to_features,
|
||||
RawResult, write_predictions,
|
||||
RawResultExtended, write_predictions_extended)
|
||||
|
||||
# The follwing import is the official SQuAD evaluation script (2.0).
|
||||
# You can remove it from the dependencies if you are using this script outside of the library
|
||||
# We've added it here for automated tests (see examples/test_examples.py file)
|
||||
from utils_squad_evaluate import EVAL_OPTS, main as evaluate_on_squad
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) \
|
||||
for conf in (BertConfig, XLNetConfig, XLMConfig)), ())
|
||||
|
||||
MODEL_CLASSES = {
|
||||
'bert': (BertConfig, BertForQuestionAnswering, BertTokenizer),
|
||||
'xlnet': (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
|
||||
'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
|
||||
'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer)
|
||||
}
|
||||
|
||||
def set_seed(args):
|
||||
random.seed(args.seed)
|
||||
np.random.seed(args.seed)
|
||||
torch.manual_seed(args.seed)
|
||||
if args.n_gpu > 0:
|
||||
torch.cuda.manual_seed_all(args.seed)
|
||||
|
||||
def to_list(tensor):
|
||||
return tensor.detach().cpu().tolist()
|
||||
|
||||
def train(args, train_dataset, model, tokenizer):
|
||||
""" Train the model """
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer = SummaryWriter()
|
||||
|
||||
args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
|
||||
train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
|
||||
train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
|
||||
|
||||
if args.max_steps > 0:
|
||||
t_total = args.max_steps
|
||||
args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
|
||||
else:
|
||||
t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
|
||||
|
||||
# Prepare optimizer and schedule (linear warmup and decay)
|
||||
no_decay = ['bias', 'LayerNorm.weight']
|
||||
optimizer_grouped_parameters = [
|
||||
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
|
||||
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
|
||||
]
|
||||
optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
|
||||
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total)
|
||||
if args.fp16:
|
||||
try:
|
||||
from apex import amp
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
|
||||
|
||||
# multi-gpu training (should be after apex fp16 initialization)
|
||||
if args.n_gpu > 1:
|
||||
model = torch.nn.DataParallel(model)
|
||||
|
||||
# Distributed training (should be after apex fp16 initialization)
|
||||
if args.local_rank != -1:
|
||||
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
|
||||
output_device=args.local_rank,
|
||||
find_unused_parameters=True)
|
||||
|
||||
# Train!
|
||||
logger.info("***** Running training *****")
|
||||
logger.info(" Num examples = %d", len(train_dataset))
|
||||
logger.info(" Num Epochs = %d", args.num_train_epochs)
|
||||
logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
|
||||
logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d",
|
||||
args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
|
||||
logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
|
||||
logger.info(" Total optimization steps = %d", t_total)
|
||||
|
||||
global_step = 0
|
||||
tr_loss, logging_loss = 0.0, 0.0
|
||||
model.zero_grad()
|
||||
train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
|
||||
set_seed(args) # Added here for reproductibility (even between python 2 and 3)
|
||||
for _ in train_iterator:
|
||||
epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
|
||||
for step, batch in enumerate(epoch_iterator):
|
||||
model.train()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1],
|
||||
'start_positions': batch[3],
|
||||
'end_positions': batch[4]}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[5],
|
||||
'p_mask': batch[6]})
|
||||
outputs = model(**inputs)
|
||||
loss = outputs[0] # model outputs are always tuple in transformers (see doc)
|
||||
|
||||
if args.n_gpu > 1:
|
||||
loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training
|
||||
if args.gradient_accumulation_steps > 1:
|
||||
loss = loss / args.gradient_accumulation_steps
|
||||
|
||||
if args.fp16:
|
||||
with amp.scale_loss(loss, optimizer) as scaled_loss:
|
||||
scaled_loss.backward()
|
||||
else:
|
||||
loss.backward()
|
||||
|
||||
tr_loss += loss.item()
|
||||
if (step + 1) % args.gradient_accumulation_steps == 0:
|
||||
if args.fp16:
|
||||
torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
|
||||
else:
|
||||
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
|
||||
|
||||
optimizer.step()
|
||||
scheduler.step() # Update learning rate schedule
|
||||
model.zero_grad()
|
||||
global_step += 1
|
||||
|
||||
if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
|
||||
# Log metrics
|
||||
if args.local_rank == -1 and args.evaluate_during_training: # Only evaluate when single GPU otherwise metrics may not average well
|
||||
results = evaluate(args, model, tokenizer)
|
||||
for key, value in results.items():
|
||||
tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
|
||||
tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
|
||||
tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
|
||||
logging_loss = tr_loss
|
||||
|
||||
if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
|
||||
# Save model checkpoint
|
||||
output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
|
||||
if not os.path.exists(output_dir):
|
||||
os.makedirs(output_dir)
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(output_dir)
|
||||
torch.save(args, os.path.join(output_dir, 'training_args.bin'))
|
||||
logger.info("Saving model checkpoint to %s", output_dir)
|
||||
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
epoch_iterator.close()
|
||||
break
|
||||
if args.max_steps > 0 and global_step > args.max_steps:
|
||||
train_iterator.close()
|
||||
break
|
||||
|
||||
if args.local_rank in [-1, 0]:
|
||||
tb_writer.close()
|
||||
|
||||
return global_step, tr_loss / global_step
|
||||
|
||||
|
||||
def evaluate(args, model, tokenizer, prefix=""):
|
||||
dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
|
||||
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
|
||||
# Note that DistributedSampler samples randomly
|
||||
eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset)
|
||||
eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
|
||||
|
||||
# Eval!
|
||||
logger.info("***** Running evaluation {} *****".format(prefix))
|
||||
logger.info(" Num examples = %d", len(dataset))
|
||||
logger.info(" Batch size = %d", args.eval_batch_size)
|
||||
all_results = []
|
||||
for batch in tqdm(eval_dataloader, desc="Evaluating"):
|
||||
model.eval()
|
||||
batch = tuple(t.to(args.device) for t in batch)
|
||||
with torch.no_grad():
|
||||
inputs = {'input_ids': batch[0],
|
||||
'attention_mask': batch[1]
|
||||
}
|
||||
if args.model_type != 'distilbert':
|
||||
inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2] # XLM don't use segment_ids
|
||||
example_indices = batch[3]
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
inputs.update({'cls_index': batch[4],
|
||||
'p_mask': batch[5]})
|
||||
outputs = model(**inputs)
|
||||
|
||||
for i, example_index in enumerate(example_indices):
|
||||
eval_feature = features[example_index.item()]
|
||||
unique_id = int(eval_feature.unique_id)
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
result = RawResultExtended(unique_id = unique_id,
|
||||
start_top_log_probs = to_list(outputs[0][i]),
|
||||
start_top_index = to_list(outputs[1][i]),
|
||||
end_top_log_probs = to_list(outputs[2][i]),
|
||||
end_top_index = to_list(outputs[3][i]),
|
||||
cls_logits = to_list(outputs[4][i]))
|
||||
else:
|
||||
result = RawResult(unique_id = unique_id,
|
||||
start_logits = to_list(outputs[0][i]),
|
||||
end_logits = to_list(outputs[1][i]))
|
||||
all_results.append(result)
|
||||
|
||||
# Compute predictions
|
||||
output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
|
||||
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
|
||||
if args.version_2_with_negative:
|
||||
output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
|
||||
else:
|
||||
output_null_log_odds_file = None
|
||||
|
||||
if args.model_type in ['xlnet', 'xlm']:
|
||||
# XLNet uses a more complex post-processing procedure
|
||||
write_predictions_extended(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file, args.predict_file,
|
||||
model.config.start_n_top, model.config.end_n_top,
|
||||
args.version_2_with_negative, tokenizer, args.verbose_logging)
|
||||
else:
|
||||
write_predictions(examples, features, all_results, args.n_best_size,
|
||||
args.max_answer_length, args.do_lower_case, output_prediction_file,
|
||||
output_nbest_file, output_null_log_odds_file, args.verbose_logging,
|
||||
args.version_2_with_negative, args.null_score_diff_threshold)
|
||||
|
||||
# Evaluate with the official SQuAD script
|
||||
evaluate_options = EVAL_OPTS(data_file=args.predict_file,
|
||||
pred_file=output_prediction_file,
|
||||
na_prob_file=output_null_log_odds_file)
|
||||
results = evaluate_on_squad(evaluate_options)
|
||||
return results
|
||||
|
||||
|
||||
def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
|
||||
if args.local_rank not in [-1, 0] and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Load data features from cache or dataset file
|
||||
input_file = args.predict_file if evaluate else args.train_file
|
||||
cached_features_file = os.path.join(os.path.dirname(input_file), 'cached_{}_{}_{}'.format(
|
||||
'dev' if evaluate else 'train',
|
||||
list(filter(None, args.model_name_or_path.split('/'))).pop(),
|
||||
str(args.max_seq_length)))
|
||||
if os.path.exists(cached_features_file) and not args.overwrite_cache and not output_examples:
|
||||
logger.info("Loading features from cached file %s", cached_features_file)
|
||||
features = torch.load(cached_features_file)
|
||||
else:
|
||||
logger.info("Creating features from dataset file at %s", input_file)
|
||||
examples = read_squad_examples(input_file=input_file,
|
||||
is_training=not evaluate,
|
||||
version_2_with_negative=args.version_2_with_negative)
|
||||
features = convert_examples_to_features(examples=examples,
|
||||
tokenizer=tokenizer,
|
||||
max_seq_length=args.max_seq_length,
|
||||
doc_stride=args.doc_stride,
|
||||
max_query_length=args.max_query_length,
|
||||
is_training=not evaluate)
|
||||
if args.local_rank in [-1, 0]:
|
||||
logger.info("Saving features into cached file %s", cached_features_file)
|
||||
torch.save(features, cached_features_file)
|
||||
|
||||
if args.local_rank == 0 and not evaluate:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, and the others will use the cache
|
||||
|
||||
# Convert to Tensors and build dataset
|
||||
all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
|
||||
all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
|
||||
all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
|
||||
all_cls_index = torch.tensor([f.cls_index for f in features], dtype=torch.long)
|
||||
all_p_mask = torch.tensor([f.p_mask for f in features], dtype=torch.float)
|
||||
if evaluate:
|
||||
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_example_index, all_cls_index, all_p_mask)
|
||||
else:
|
||||
all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long)
|
||||
all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long)
|
||||
dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
|
||||
all_start_positions, all_end_positions,
|
||||
all_cls_index, all_p_mask)
|
||||
|
||||
if output_examples:
|
||||
return dataset, examples, features
|
||||
return dataset
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
## Required parameters
|
||||
parser.add_argument("--train_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for training. E.g., train-v1.1.json")
|
||||
parser.add_argument("--predict_file", default=None, type=str, required=True,
|
||||
help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json")
|
||||
parser.add_argument("--model_type", default=None, type=str, required=True,
|
||||
help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
|
||||
parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
|
||||
help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
|
||||
parser.add_argument("--output_dir", default=None, type=str, required=True,
|
||||
help="The output directory where the model checkpoints and predictions will be written.")
|
||||
|
||||
## Other parameters
|
||||
parser.add_argument("--config_name", default="", type=str,
|
||||
help="Pretrained config name or path if not the same as model_name")
|
||||
parser.add_argument("--tokenizer_name", default="", type=str,
|
||||
help="Pretrained tokenizer name or path if not the same as model_name")
|
||||
parser.add_argument("--cache_dir", default="", type=str,
|
||||
help="Where do you want to store the pre-trained models downloaded from s3")
|
||||
|
||||
parser.add_argument('--version_2_with_negative', action='store_true',
|
||||
help='If true, the SQuAD examples contain some that do not have an answer.')
|
||||
parser.add_argument('--null_score_diff_threshold', type=float, default=0.0,
|
||||
help="If null_score - best_non_null is greater than the threshold predict null.")
|
||||
|
||||
parser.add_argument("--max_seq_length", default=384, type=int,
|
||||
help="The maximum total input sequence length after WordPiece tokenization. Sequences "
|
||||
"longer than this will be truncated, and sequences shorter than this will be padded.")
|
||||
parser.add_argument("--doc_stride", default=128, type=int,
|
||||
help="When splitting up a long document into chunks, how much stride to take between chunks.")
|
||||
parser.add_argument("--max_query_length", default=64, type=int,
|
||||
help="The maximum number of tokens for the question. Questions longer than this will "
|
||||
"be truncated to this length.")
|
||||
parser.add_argument("--do_train", action='store_true',
|
||||
help="Whether to run training.")
|
||||
parser.add_argument("--do_eval", action='store_true',
|
||||
help="Whether to run eval on the dev set.")
|
||||
parser.add_argument("--evaluate_during_training", action='store_true',
|
||||
help="Rul evaluation during training at each logging step.")
|
||||
parser.add_argument("--do_lower_case", action='store_true',
|
||||
help="Set this flag if you are using an uncased model.")
|
||||
|
||||
parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for training.")
|
||||
parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
|
||||
help="Batch size per GPU/CPU for evaluation.")
|
||||
parser.add_argument("--learning_rate", default=5e-5, type=float,
|
||||
help="The initial learning rate for Adam.")
|
||||
parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
|
||||
help="Number of updates steps to accumulate before performing a backward/update pass.")
|
||||
parser.add_argument("--weight_decay", default=0.0, type=float,
|
||||
help="Weight deay if we apply some.")
|
||||
parser.add_argument("--adam_epsilon", default=1e-8, type=float,
|
||||
help="Epsilon for Adam optimizer.")
|
||||
parser.add_argument("--max_grad_norm", default=1.0, type=float,
|
||||
help="Max gradient norm.")
|
||||
parser.add_argument("--num_train_epochs", default=3.0, type=float,
|
||||
help="Total number of training epochs to perform.")
|
||||
parser.add_argument("--max_steps", default=-1, type=int,
|
||||
help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
|
||||
parser.add_argument("--warmup_steps", default=0, type=int,
|
||||
help="Linear warmup over warmup_steps.")
|
||||
parser.add_argument("--n_best_size", default=20, type=int,
|
||||
help="The total number of n-best predictions to generate in the nbest_predictions.json output file.")
|
||||
parser.add_argument("--max_answer_length", default=30, type=int,
|
||||
help="The maximum length of an answer that can be generated. This is needed because the start "
|
||||
"and end predictions are not conditioned on one another.")
|
||||
parser.add_argument("--verbose_logging", action='store_true',
|
||||
help="If true, all of the warnings related to data processing will be printed. "
|
||||
"A number of warnings are expected for a normal SQuAD evaluation.")
|
||||
|
||||
parser.add_argument('--logging_steps', type=int, default=50,
|
||||
help="Log every X updates steps.")
|
||||
parser.add_argument('--save_steps', type=int, default=50,
|
||||
help="Save checkpoint every X updates steps.")
|
||||
parser.add_argument("--eval_all_checkpoints", action='store_true',
|
||||
help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
|
||||
parser.add_argument("--no_cuda", action='store_true',
|
||||
help="Whether not to use CUDA when available")
|
||||
parser.add_argument('--overwrite_output_dir', action='store_true',
|
||||
help="Overwrite the content of the output directory")
|
||||
parser.add_argument('--overwrite_cache', action='store_true',
|
||||
help="Overwrite the cached training and evaluation sets")
|
||||
parser.add_argument('--seed', type=int, default=42,
|
||||
help="random seed for initialization")
|
||||
|
||||
parser.add_argument("--local_rank", type=int, default=-1,
|
||||
help="local_rank for distributed training on gpus")
|
||||
parser.add_argument('--fp16', action='store_true',
|
||||
help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
|
||||
parser.add_argument('--fp16_opt_level', type=str, default='O1',
|
||||
help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
|
||||
"See details at https://nvidia.github.io/apex/amp.html")
|
||||
parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
|
||||
parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
|
||||
args = parser.parse_args()
|
||||
|
||||
if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
|
||||
raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
|
||||
|
||||
# Setup distant debugging if needed
|
||||
if args.server_ip and args.server_port:
|
||||
# Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
|
||||
import ptvsd
|
||||
print("Waiting for debugger attach")
|
||||
ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
|
||||
ptvsd.wait_for_attach()
|
||||
|
||||
# Setup CUDA, GPU & distributed training
|
||||
if args.local_rank == -1 or args.no_cuda:
|
||||
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
|
||||
args.n_gpu = torch.cuda.device_count()
|
||||
else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
|
||||
torch.cuda.set_device(args.local_rank)
|
||||
device = torch.device("cuda", args.local_rank)
|
||||
torch.distributed.init_process_group(backend='nccl')
|
||||
args.n_gpu = 1
|
||||
args.device = device
|
||||
|
||||
# Setup logging
|
||||
logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - %(message)s',
|
||||
datefmt = '%m/%d/%Y %H:%M:%S',
|
||||
level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
|
||||
logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
|
||||
args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
|
||||
|
||||
# Set seed
|
||||
set_seed(args)
|
||||
|
||||
# Load pretrained model and tokenizer
|
||||
if args.local_rank not in [-1, 0]:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
args.model_type = args.model_type.lower()
|
||||
config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
|
||||
config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
|
||||
do_lower_case=args.do_lower_case,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
model = model_class.from_pretrained(args.model_name_or_path,
|
||||
from_tf=bool('.ckpt' in args.model_name_or_path),
|
||||
config=config,
|
||||
cache_dir=args.cache_dir if args.cache_dir else None)
|
||||
|
||||
if args.local_rank == 0:
|
||||
torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab
|
||||
|
||||
model.to(args.device)
|
||||
|
||||
logger.info("Training/evaluation parameters %s", args)
|
||||
|
||||
# Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
|
||||
# Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
|
||||
# remove the need for this code, but it is still valid.
|
||||
if args.fp16:
|
||||
try:
|
||||
import apex
|
||||
apex.amp.register_half_function(torch, 'einsum')
|
||||
except ImportError:
|
||||
raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
|
||||
|
||||
# Training
|
||||
if args.do_train:
|
||||
train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
|
||||
global_step, tr_loss = train(args, train_dataset, model, tokenizer)
|
||||
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
|
||||
|
||||
|
||||
# Save the trained model and the tokenizer
|
||||
if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
|
||||
# Create output directory if needed
|
||||
if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
|
||||
os.makedirs(args.output_dir)
|
||||
|
||||
logger.info("Saving model checkpoint to %s", args.output_dir)
|
||||
# Save a trained model, configuration and tokenizer using `save_pretrained()`.
|
||||
# They can then be reloaded using `from_pretrained()`
|
||||
model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training
|
||||
model_to_save.save_pretrained(args.output_dir)
|
||||
tokenizer.save_pretrained(args.output_dir)
|
||||
|
||||
# Good practice: save your training arguments together with the trained model
|
||||
torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
|
||||
|
||||
# Load a trained model and vocabulary that you have fine-tuned
|
||||
model = model_class.from_pretrained(args.output_dir)
|
||||
tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
|
||||
model.to(args.device)
|
||||
|
||||
|
||||
# Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
|
||||
results = {}
|
||||
if args.do_eval and args.local_rank in [-1, 0]:
|
||||
checkpoints = [args.output_dir]
|
||||
if args.eval_all_checkpoints:
|
||||
checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
|
||||
logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs
|
||||
|
||||
logger.info("Evaluate the following checkpoints: %s", checkpoints)
|
||||
|
||||
for checkpoint in checkpoints:
|
||||
# Reload the model
|
||||
global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
|
||||
model = model_class.from_pretrained(checkpoint)
|
||||
model.to(args.device)
|
||||
|
||||
# Evaluate
|
||||
result = evaluate(args, model, tokenizer, prefix=global_step)
|
||||
|
||||
result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items())
|
||||
results.update(result)
|
||||
|
||||
logger.info("Results: {}".format(results))
|
||||
|
||||
return results
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,62 +0,0 @@
|
||||
# How to add a new model in 🤗Transformers
|
||||
|
||||
This folder describes the process to add a new model in 🤗Transformers and provide templates for the required files.
|
||||
|
||||
The library is designed to incorporate a variety of models and code bases. As such the process for adding a new model usually mostly consists in copy-pasting to relevant original code in the various sections of the templates included in the present repository.
|
||||
|
||||
One important point though is that the library has the following goals impacting the way models are incorporated:
|
||||
|
||||
- one specific feature of the API is the capability to run the model and tokenizer inline. The tokenization code thus often have to be slightly adapted to allow for running in the python interpreter.
|
||||
- the package is also designed to be as self-consistent and with a small and reliable set of packages dependencies. In consequence, additional dependencies are usually not allowed when adding a model but can be allowed for the inclusion of a new tokenizer (recent examples of dependencies added for tokenizer specificities include `sentencepiece` and `sacremoses`). Please make sure to check the existing dependencies when possible before adding a new one.
|
||||
|
||||
For a quick overview of the library organization, please check the [QuickStart section of the documentation](https://huggingface.co/transformers/quickstart.html).
|
||||
|
||||
# Typical workflow for including a model
|
||||
|
||||
Here an overview of the general workflow:
|
||||
|
||||
- [ ] add model/configuration/tokenization classes
|
||||
- [ ] add conversion scripts
|
||||
- [ ] add tests
|
||||
- [ ] finalize
|
||||
|
||||
Let's detail what should be done at each step
|
||||
|
||||
## Adding model/configuration/tokenization classes
|
||||
|
||||
Here is the workflow for adding model/configuration/tokenization classes:
|
||||
|
||||
- [ ] copy the python files from the present folder to the main folder and rename them, replacing `xxx` with your model name,
|
||||
- [ ] edit the files to replace `XXX` (with various casing) with your model name
|
||||
- [ ] copy-paste or create a simple configuration class for your model in the `configuration_...` file
|
||||
- [ ] copy-paste or create the code for your model in the `modeling_...` files (PyTorch and TF 2.0)
|
||||
- [ ] copy-paste or create a tokenizer class for your model in the `tokenization_...` file
|
||||
|
||||
# Adding conversion scripts
|
||||
|
||||
Here is the workflow for the conversion scripts:
|
||||
|
||||
- [ ] copy the conversion script (`convert_...`) from the present folder to the main folder.
|
||||
- [ ] edit this script to convert your original checkpoint weights to the current pytorch ones.
|
||||
|
||||
# Adding tests:
|
||||
|
||||
Here is the workflow for the adding tests:
|
||||
|
||||
- [ ] copy the python files from the `tests` sub-folder of the present folder to the `tests` subfolder of the main folder and rename them, replacing `xxx` with your model name,
|
||||
- [ ] edit the tests files to replace `XXX` (with various casing) with your model name
|
||||
- [ ] edit the tests code as needed
|
||||
|
||||
# Final steps
|
||||
|
||||
You can then finish the addition step by adding imports for your classes in the common files:
|
||||
|
||||
- [ ] add import for all the relevant classes in `__init__.py`
|
||||
- [ ] add your configuration in `configuration_auto.py`
|
||||
- [ ] add your PyTorch and TF 2.0 model respectively in `modeling_auto.py` and `modeling_tf_auto.py`
|
||||
- [ ] add your tokenizer in `tokenization_auto.py`
|
||||
- [ ] add your models and tokenizer to `pipeline.py`
|
||||
- [ ] add a link to your conversion script in the main conversion utility (currently in `__main__` but will be moved to the `commands` subfolder in the near future)
|
||||
- [ ] edit the PyTorch to TF 2.0 conversion script to add your model in the `convert_pytorch_checkpoint_to_tf2.py` file
|
||||
- [ ] add a mention of your model in the doc: `README.md` and the documentation itself at `docs/source/pretrained_models.rst`.
|
||||
- [ ] upload the pretrained weigths, configurations and vocabulary files.
|
||||
@@ -1,130 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2010, XXX authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" XXX model configuration """
|
||||
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
import json
|
||||
import logging
|
||||
import sys
|
||||
import six
|
||||
from io import open
|
||||
|
||||
from .configuration_utils import PretrainedConfig
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
XXX_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-config.json",
|
||||
'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-config.json",
|
||||
}
|
||||
|
||||
|
||||
class XxxConfig(PretrainedConfig):
|
||||
r"""
|
||||
:class:`~transformers.XxxConfig` is the configuration class to store the configuration of a
|
||||
`XxxModel`.
|
||||
|
||||
|
||||
Arguments:
|
||||
vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XxxModel`.
|
||||
hidden_size: Size of the encoder layers and the pooler layer.
|
||||
num_hidden_layers: Number of hidden layers in the Transformer encoder.
|
||||
num_attention_heads: Number of attention heads for each attention layer in
|
||||
the Transformer encoder.
|
||||
intermediate_size: The size of the "intermediate" (i.e., feed-forward)
|
||||
layer in the Transformer encoder.
|
||||
hidden_act: The non-linear activation function (function or string) in the
|
||||
encoder and pooler. If string, "gelu", "relu", "swish" and "gelu_new" are supported.
|
||||
hidden_dropout_prob: The dropout probabilitiy for all fully connected
|
||||
layers in the embeddings, encoder, and pooler.
|
||||
attention_probs_dropout_prob: The dropout ratio for the attention
|
||||
probabilities.
|
||||
max_position_embeddings: The maximum sequence length that this model might
|
||||
ever be used with. Typically set this to something large just in case
|
||||
(e.g., 512 or 1024 or 2048).
|
||||
type_vocab_size: The vocabulary size of the `token_type_ids` passed into
|
||||
`XxxModel`.
|
||||
initializer_range: The sttdev of the truncated_normal_initializer for
|
||||
initializing all weight matrices.
|
||||
layer_norm_eps: The epsilon used by LayerNorm.
|
||||
"""
|
||||
pretrained_config_archive_map = XXX_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
|
||||
def __init__(self,
|
||||
vocab_size_or_config_json_file=50257,
|
||||
n_positions=1024,
|
||||
n_ctx=1024,
|
||||
n_embd=768,
|
||||
n_layer=12,
|
||||
n_head=12,
|
||||
resid_pdrop=0.1,
|
||||
embd_pdrop=0.1,
|
||||
attn_pdrop=0.1,
|
||||
layer_norm_epsilon=1e-5,
|
||||
initializer_range=0.02,
|
||||
|
||||
num_labels=1,
|
||||
summary_type='cls_index',
|
||||
summary_use_proj=True,
|
||||
summary_activation=None,
|
||||
summary_proj_to_labels=True,
|
||||
summary_first_dropout=0.1,
|
||||
**kwargs):
|
||||
super(XxxConfig, self).__init__(**kwargs)
|
||||
self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
|
||||
self.n_ctx = n_ctx
|
||||
self.n_positions = n_positions
|
||||
self.n_embd = n_embd
|
||||
self.n_layer = n_layer
|
||||
self.n_head = n_head
|
||||
self.resid_pdrop = resid_pdrop
|
||||
self.embd_pdrop = embd_pdrop
|
||||
self.attn_pdrop = attn_pdrop
|
||||
self.layer_norm_epsilon = layer_norm_epsilon
|
||||
self.initializer_range = initializer_range
|
||||
|
||||
self.num_labels = num_labels
|
||||
self.summary_type = summary_type
|
||||
self.summary_use_proj = summary_use_proj
|
||||
self.summary_activation = summary_activation
|
||||
self.summary_first_dropout = summary_first_dropout
|
||||
self.summary_proj_to_labels = summary_proj_to_labels
|
||||
if isinstance(vocab_size_or_config_json_file, six.string_types):
|
||||
with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
|
||||
json_config = json.loads(reader.read())
|
||||
for key, value in json_config.items():
|
||||
self.__dict__[key] = value
|
||||
elif not isinstance(vocab_size_or_config_json_file, int):
|
||||
raise ValueError(
|
||||
"First argument must be either a vocabulary size (int)"
|
||||
"or the path to a pretrained model config file (str)"
|
||||
)
|
||||
|
||||
@property
|
||||
def max_position_embeddings(self):
|
||||
return self.n_positions
|
||||
|
||||
@property
|
||||
def hidden_size(self):
|
||||
return self.n_embd
|
||||
|
||||
@property
|
||||
def num_attention_heads(self):
|
||||
return self.n_head
|
||||
|
||||
@property
|
||||
def num_hidden_layers(self):
|
||||
return self.n_layer
|
||||
@@ -1,65 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The HuggingFace Inc. team.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Convert XXX checkpoint."""
|
||||
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import argparse
|
||||
import torch
|
||||
|
||||
from transformers import XxxConfig, XxxForPreTraining, load_tf_weights_in_xxx
|
||||
|
||||
import logging
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
|
||||
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
|
||||
# Initialise PyTorch model
|
||||
config = XxxConfig.from_json_file(config_file)
|
||||
print("Building PyTorch model from configuration: {}".format(str(config)))
|
||||
model = XxxForPreTraining(config)
|
||||
|
||||
# Load weights from tf checkpoint
|
||||
load_tf_weights_in_xxx(model, config, tf_checkpoint_path)
|
||||
|
||||
# Save pytorch-model
|
||||
print("Save PyTorch model to {}".format(pytorch_dump_path))
|
||||
torch.save(model.state_dict(), pytorch_dump_path)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
## Required parameters
|
||||
parser.add_argument("--tf_checkpoint_path",
|
||||
default = None,
|
||||
type = str,
|
||||
required = True,
|
||||
help = "Path to the TensorFlow checkpoint path.")
|
||||
parser.add_argument("--config_file",
|
||||
default = None,
|
||||
type = str,
|
||||
required = True,
|
||||
help = "The config json file corresponding to the pre-trained model. \n"
|
||||
"This specifies the model architecture.")
|
||||
parser.add_argument("--pytorch_dump_path",
|
||||
default = None,
|
||||
type = str,
|
||||
required = True,
|
||||
help = "Path to the output PyTorch model.")
|
||||
args = parser.parse_args()
|
||||
convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
|
||||
args.config_file,
|
||||
args.pytorch_dump_path)
|
||||
@@ -1,506 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" TF 2.0 XXX model. """
|
||||
|
||||
####################################################
|
||||
# In this template, replace all the XXX (various casings) with your model name
|
||||
####################################################
|
||||
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
import json
|
||||
import logging
|
||||
import math
|
||||
import os
|
||||
import sys
|
||||
import copy
|
||||
import itertools
|
||||
from io import open
|
||||
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
|
||||
from .configuration_xxx import XxxConfig
|
||||
from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list
|
||||
from .file_utils import add_start_docstrings
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
####################################################
|
||||
# This dict contrains shortcut names and associated url
|
||||
# for the pretrained weights provided with the models
|
||||
####################################################
|
||||
TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP = {
|
||||
'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-tf_model.h5",
|
||||
'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-tf_model.h5",
|
||||
}
|
||||
|
||||
####################################################
|
||||
# TF 2.0 Models are constructed using Keras imperative API by sub-classing
|
||||
# - tf.keras.layers.Layer for the layers and
|
||||
# - TFPreTrainedModel for the models (itself a sub-class of tf.keras.Model)
|
||||
####################################################
|
||||
|
||||
####################################################
|
||||
# Here is an example of typical layer in a TF 2.0 model of the library
|
||||
# The classes are usually identical to the PyTorch ones and prefixed with 'TF'.
|
||||
#
|
||||
# Note that class __init__ parameters includes **kwargs (send to 'super').
|
||||
# This let us have a control on class scope and variable names:
|
||||
# More precisely, we set the names of the class attributes (lower level layers) to
|
||||
# to the equivalent attributes names in the PyTorch model so we can have equivalent
|
||||
# class and scope structure between PyTorch and TF 2.0 models and easily load one in the other.
|
||||
#
|
||||
# See the conversion methods in modeling_tf_pytorch_utils.py for more details
|
||||
####################################################
|
||||
class TFXxxLayer(tf.keras.layers.Layer):
|
||||
def __init__(self, config, **kwargs):
|
||||
super(TFXxxLayer, self).__init__(**kwargs)
|
||||
self.attention = TFXxxAttention(config, name='attention')
|
||||
self.intermediate = TFXxxIntermediate(config, name='intermediate')
|
||||
self.transformer_output = TFXxxOutput(config, name='output')
|
||||
|
||||
def call(self, inputs, training=False):
|
||||
hidden_states, attention_mask, head_mask = inputs
|
||||
|
||||
attention_outputs = self.attention([hidden_states, attention_mask, head_mask], training=training)
|
||||
attention_output = attention_outputs[0]
|
||||
intermediate_output = self.intermediate(attention_output)
|
||||
layer_output = self.transformer_output([intermediate_output, attention_output], training=training)
|
||||
outputs = (layer_output,) + attention_outputs[1:] # add attentions if we output them
|
||||
return outputs
|
||||
|
||||
|
||||
####################################################
|
||||
# The full model without a specific pretrained or finetuning head is
|
||||
# provided as a tf.keras.layers.Layer usually called "TFXxxMainLayer"
|
||||
####################################################
|
||||
class TFXxxMainLayer(tf.keras.layers.Layer):
|
||||
def __init__(self, config, **kwargs):
|
||||
super(TFXxxMainLayer, self).__init__(**kwargs)
|
||||
|
||||
def _resize_token_embeddings(self, new_num_tokens):
|
||||
raise NotImplementedError # Not implemented yet in the library fr TF 2.0 models
|
||||
|
||||
def _prune_heads(self, heads_to_prune):
|
||||
raise NotImplementedError # Not implemented yet in the library fr TF 2.0 models
|
||||
|
||||
def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, training=False):
|
||||
# We allow three types of multi-inputs:
|
||||
# - traditional keyword arguments in the call method
|
||||
# - all the arguments provided as a dict in the first positional argument of call
|
||||
# - all the arguments provided as a list/tuple (ordered) in the first positional argument of call
|
||||
# The last two options are useful to use the tf.keras fit() method.
|
||||
|
||||
if isinstance(inputs, (tuple, list)):
|
||||
input_ids = inputs[0]
|
||||
attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
|
||||
token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
|
||||
position_ids = inputs[3] if len(inputs) > 3 else position_ids
|
||||
head_mask = inputs[4] if len(inputs) > 4 else head_mask
|
||||
assert len(inputs) <= 5, "Too many inputs."
|
||||
elif isinstance(inputs, dict):
|
||||
input_ids = inputs.get('input_ids')
|
||||
attention_mask = inputs.get('attention_mask', attention_mask)
|
||||
token_type_ids = inputs.get('token_type_ids', token_type_ids)
|
||||
position_ids = inputs.get('position_ids', position_ids)
|
||||
head_mask = inputs.get('head_mask', head_mask)
|
||||
assert len(inputs) <= 5, "Too many inputs."
|
||||
else:
|
||||
input_ids = inputs
|
||||
|
||||
if attention_mask is None:
|
||||
attention_mask = tf.fill(shape_list(input_ids), 1)
|
||||
if token_type_ids is None:
|
||||
token_type_ids = tf.fill(shape_list(input_ids), 0)
|
||||
|
||||
# We create a 3D attention mask from a 2D tensor mask.
|
||||
# Sizes are [batch_size, 1, 1, to_seq_length]
|
||||
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
|
||||
# this attention mask is more simple than the triangular masking of causal attention
|
||||
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
|
||||
extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :]
|
||||
|
||||
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
|
||||
# masked positions, this operation will create a tensor which is 0.0 for
|
||||
# positions we want to attend and -10000.0 for masked positions.
|
||||
# Since we are adding it to the raw scores before the softmax, this is
|
||||
# effectively the same as removing these entirely.
|
||||
|
||||
extended_attention_mask = tf.cast(extended_attention_mask, tf.float32)
|
||||
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
|
||||
|
||||
# Prepare head mask if needed
|
||||
# 1.0 in head_mask indicate we keep the head
|
||||
# attention_probs has shape bsz x n_heads x N x N
|
||||
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
|
||||
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
|
||||
if not head_mask is None:
|
||||
raise NotImplementedError
|
||||
else:
|
||||
head_mask = [None] * self.num_hidden_layers
|
||||
# head_mask = tf.constant([0] * self.num_hidden_layers)
|
||||
|
||||
##################################
|
||||
# Replace this with your model code
|
||||
embedding_output = self.embeddings(input_ids, position_ids=position_ids, token_type_ids=token_type_ids)
|
||||
encoder_outputs = self.encoder([embedding_output, extended_attention_mask, head_mask], training=training)
|
||||
sequence_output = encoder_outputs[0]
|
||||
outputs = (sequence_output,) + encoder_outputs[1:] # add hidden_states and attentions if they are here
|
||||
|
||||
return outputs # sequence_output, (hidden_states), (attentions)
|
||||
|
||||
|
||||
####################################################
|
||||
# TFXxxPreTrainedModel is a sub-class of tf.keras.Model
|
||||
# which take care of loading and saving pretrained weights
|
||||
# and various common utilities.
|
||||
# Here you just need to specify a few (self-explanatory)
|
||||
# pointers for your model.
|
||||
####################################################
|
||||
class TFXxxPreTrainedModel(TFPreTrainedModel):
|
||||
""" An abstract class to handle weights initialization and
|
||||
a simple interface for dowloading and loading pretrained models.
|
||||
"""
|
||||
config_class = XxxConfig
|
||||
pretrained_model_archive_map = TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP
|
||||
base_model_prefix = "transformer"
|
||||
|
||||
|
||||
XXX_START_DOCSTRING = r""" The XXX model was proposed in
|
||||
`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_
|
||||
by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
|
||||
pre-trained using a combination of masked language modeling objective and next sentence prediction
|
||||
on a large corpus comprising the Toronto Book Corpus and Wikipedia.
|
||||
|
||||
This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
|
||||
refer to the TF 2.0 documentation for all matter related to general usage and behavior.
|
||||
|
||||
.. _`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`:
|
||||
https://arxiv.org/abs/1810.04805
|
||||
|
||||
.. _`tf.keras.Model`:
|
||||
https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
|
||||
|
||||
Note on the model inputs:
|
||||
TF 2.0 models accepts two formats as inputs:
|
||||
|
||||
- having all inputs as keyword arguments (like PyTorch models), or
|
||||
- having all inputs as a list, tuple or dict in the first positional arguments.
|
||||
|
||||
This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
|
||||
|
||||
If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
|
||||
|
||||
- a single Tensor with input_ids only and nothing else: `model(inputs_ids)
|
||||
- a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
|
||||
`model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
|
||||
- a dictionary with one or several input Tensors associaed to the input names given in the docstring:
|
||||
`model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
|
||||
|
||||
Parameters:
|
||||
config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model.
|
||||
Initializing with a config file does not load the weights associated with the model, only the configuration.
|
||||
Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
|
||||
"""
|
||||
|
||||
XXX_INPUTS_DOCSTRING = r"""
|
||||
Inputs:
|
||||
**input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Indices of input sequence tokens in the vocabulary.
|
||||
To match pre-training, XXX input sequence should be formatted with [CLS] and [SEP] tokens as follows:
|
||||
|
||||
(a) For sequence pairs:
|
||||
|
||||
``tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
|
||||
|
||||
``token_type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1``
|
||||
|
||||
(b) For single sequences:
|
||||
|
||||
``tokens: [CLS] the dog is hairy . [SEP]``
|
||||
|
||||
``token_type_ids: 0 0 0 0 0 0 0``
|
||||
|
||||
Xxx is a model with absolute position embeddings so it's usually advised to pad the inputs on
|
||||
the right rather than the left.
|
||||
|
||||
Indices can be obtained using :class:`transformers.XxxTokenizer`.
|
||||
See :func:`transformers.PreTrainedTokenizer.encode` and
|
||||
:func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
|
||||
**attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Mask to avoid performing attention on padding token indices.
|
||||
Mask values selected in ``[0, 1]``:
|
||||
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
|
||||
**token_type_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Segment token indices to indicate first and second portions of the inputs.
|
||||
Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
|
||||
corresponds to a `sentence B` token
|
||||
(see `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
|
||||
**position_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Indices of positions of each input sequence tokens in the position embeddings.
|
||||
Selected in the range ``[0, config.max_position_embeddings - 1]``.
|
||||
**head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
|
||||
Mask to nullify selected heads of the self-attention modules.
|
||||
Mask values selected in ``[0, 1]``:
|
||||
``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
|
||||
**inputs_embeds**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
|
||||
Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
|
||||
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
||||
than the model's internal embedding lookup matrix.
|
||||
"""
|
||||
|
||||
@add_start_docstrings("The bare Xxx Model transformer outputing raw hidden-states without any specific head on top.",
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class TFXxxModel(TFXxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
|
||||
Sequence of hidden-states at the output of the last layer of the model.
|
||||
**pooler_output**: ``tf.Tensor`` of shape ``(batch_size, hidden_size)``
|
||||
Last layer hidden-state of the first token of the sequence (classification token)
|
||||
further processed by a Linear layer and a Tanh activation function. The Linear
|
||||
layer weights are trained from the next sentence prediction (classification)
|
||||
objective during Xxx pretraining. This output is usually *not* a good summary
|
||||
of the semantic content of the input, you're often better with averaging or pooling
|
||||
the sequence of hidden-states for the whole input sequence.
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
import tensorflow as tf
|
||||
from transformers import XxxTokenizer, TFXxxModel
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = TFXxxModel.from_pretrained('xxx-base-uncased')
|
||||
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple
|
||||
|
||||
"""
|
||||
def __init__(self, config, *inputs, **kwargs):
|
||||
super(TFXxxModel, self).__init__(config, *inputs, **kwargs)
|
||||
self.transformer = TFXxxMainLayer(config, name='transformer')
|
||||
|
||||
def call(self, inputs, **kwargs):
|
||||
outputs = self.transformer(inputs, **kwargs)
|
||||
return outputs
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a `language modeling` head on top. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class TFXxxForMaskedLM(TFXxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**prediction_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
|
||||
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
import tensorflow as tf
|
||||
from transformers import XxxTokenizer, TFXxxForMaskedLM
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = TFXxxForMaskedLM.from_pretrained('xxx-base-uncased')
|
||||
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
prediction_scores = outputs[0]
|
||||
|
||||
"""
|
||||
def __init__(self, config, *inputs, **kwargs):
|
||||
super(TFXxxForMaskedLM, self).__init__(config, *inputs, **kwargs)
|
||||
|
||||
self.transformer = TFXxxMainLayer(config, name='transformer')
|
||||
self.mlm = TFXxxMLMHead(config, self.transformer.embeddings, name='mlm')
|
||||
|
||||
def call(self, inputs, **kwargs):
|
||||
outputs = self.transformer(inputs, **kwargs)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
prediction_scores = self.mlm(sequence_output, training=kwargs.get('training', False))
|
||||
|
||||
outputs = (prediction_scores,) + outputs[2:] # Add hidden states and attention if they are here
|
||||
|
||||
return outputs # prediction_scores, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model transformer with a sequence classification/regression head on top (a linear layer on top of
|
||||
the pooled output) e.g. for GLUE tasks. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class TFXxxForSequenceClassification(TFXxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**logits**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, config.num_labels)``
|
||||
Classification (or regression if config.num_labels==1) scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
import tensorflow as tf
|
||||
from transformers import XxxTokenizer, TFXxxForSequenceClassification
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = TFXxxForSequenceClassification.from_pretrained('xxx-base-uncased')
|
||||
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
logits = outputs[0]
|
||||
|
||||
"""
|
||||
def __init__(self, config, *inputs, **kwargs):
|
||||
super(TFXxxForSequenceClassification, self).__init__(config, *inputs, **kwargs)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = TFXxxMainLayer(config, name='transformer')
|
||||
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
|
||||
self.classifier = tf.keras.layers.Dense(config.num_labels,
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name='classifier')
|
||||
|
||||
def call(self, inputs, **kwargs):
|
||||
outputs = self.transformer(inputs, **kwargs)
|
||||
|
||||
pooled_output = outputs[1]
|
||||
|
||||
pooled_output = self.dropout(pooled_output, training=kwargs.get('training', False))
|
||||
logits = self.classifier(pooled_output)
|
||||
|
||||
outputs = (logits,) + outputs[2:] # add hidden states and attention if they are here
|
||||
|
||||
return outputs # logits, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a token classification head on top (a linear layer on top of
|
||||
the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class TFXxxForTokenClassification(TFXxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
|
||||
Classification scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
import tensorflow as tf
|
||||
from transformers import XxxTokenizer, TFXxxForTokenClassification
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = TFXxxForTokenClassification.from_pretrained('xxx-base-uncased')
|
||||
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
scores = outputs[0]
|
||||
|
||||
"""
|
||||
def __init__(self, config, *inputs, **kwargs):
|
||||
super(TFXxxForTokenClassification, self).__init__(config, *inputs, **kwargs)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = TFXxxMainLayer(config, name='transformer')
|
||||
self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
|
||||
self.classifier = tf.keras.layers.Dense(config.num_labels,
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name='classifier')
|
||||
|
||||
def call(self, inputs, **kwargs):
|
||||
outputs = self.transformer(inputs, **kwargs)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
|
||||
sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
|
||||
logits = self.classifier(sequence_output)
|
||||
|
||||
outputs = (logits,) + outputs[2:] # add hidden states and attention if they are here
|
||||
|
||||
return outputs # scores, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
|
||||
the hidden-states output to compute `span start logits` and `span end logits`). """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class TFXxxForQuestionAnswering(TFXxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**start_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
|
||||
Span-start scores (before SoftMax).
|
||||
**end_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length,)``
|
||||
Span-end scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
import tensorflow as tf
|
||||
from transformers import XxxTokenizer, TFXxxForQuestionAnswering
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = TFXxxForQuestionAnswering.from_pretrained('xxx-base-uncased')
|
||||
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :] # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
start_scores, end_scores = outputs[:2]
|
||||
|
||||
"""
|
||||
def __init__(self, config, *inputs, **kwargs):
|
||||
super(TFXxxForQuestionAnswering, self).__init__(config, *inputs, **kwargs)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = TFXxxMainLayer(config, name='transformer')
|
||||
self.qa_outputs = tf.keras.layers.Dense(config.num_labels,
|
||||
kernel_initializer=get_initializer(config.initializer_range),
|
||||
name='qa_outputs')
|
||||
|
||||
def call(self, inputs, **kwargs):
|
||||
outputs = self.transformer(inputs, **kwargs)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
|
||||
logits = self.qa_outputs(sequence_output)
|
||||
start_logits, end_logits = tf.split(logits, 2, axis=-1)
|
||||
start_logits = tf.squeeze(start_logits, axis=-1)
|
||||
end_logits = tf.squeeze(end_logits, axis=-1)
|
||||
|
||||
outputs = (start_logits, end_logits,) + outputs[2:]
|
||||
|
||||
return outputs # start_logits, end_logits, (hidden_states), (attentions)
|
||||
@@ -1,660 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX Authors
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" PyTorch XXX model. """
|
||||
|
||||
####################################################
|
||||
# In this template, replace all the XXX (various casings) with your model name
|
||||
####################################################
|
||||
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
import json
|
||||
import logging
|
||||
import math
|
||||
import os
|
||||
import sys
|
||||
import copy
|
||||
import itertools
|
||||
from io import open
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
from torch.nn import CrossEntropyLoss, MSELoss
|
||||
|
||||
from .modeling_utils import PreTrainedModel, prune_linear_layer
|
||||
from .configuration_xxx import XxxConfig
|
||||
from .file_utils import add_start_docstrings
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
####################################################
|
||||
# This dict contrains shortcut names and associated url
|
||||
# for the pretrained weights provided with the models
|
||||
####################################################
|
||||
XXX_PRETRAINED_MODEL_ARCHIVE_MAP = {
|
||||
'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-pytorch_model.bin",
|
||||
'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-pytorch_model.bin",
|
||||
}
|
||||
|
||||
####################################################
|
||||
# This is a conversion method from TF 1.0 to PyTorch
|
||||
# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
|
||||
####################################################
|
||||
def load_tf_weights_in_xxx(model, config, tf_checkpoint_path):
|
||||
""" Load tf checkpoints in a pytorch model.
|
||||
"""
|
||||
try:
|
||||
import re
|
||||
import numpy as np
|
||||
import tensorflow as tf
|
||||
except ImportError:
|
||||
logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
|
||||
"https://www.tensorflow.org/install/ for installation instructions.")
|
||||
raise
|
||||
tf_path = os.path.abspath(tf_checkpoint_path)
|
||||
logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
|
||||
# Load weights from TF model
|
||||
init_vars = tf.train.list_variables(tf_path)
|
||||
names = []
|
||||
arrays = []
|
||||
for name, shape in init_vars:
|
||||
logger.info("Loading TF weight {} with shape {}".format(name, shape))
|
||||
array = tf.train.load_variable(tf_path, name)
|
||||
names.append(name)
|
||||
arrays.append(array)
|
||||
|
||||
for name, array in zip(names, arrays):
|
||||
name = name.split('/')
|
||||
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
|
||||
# which are not required for using pretrained model
|
||||
if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
|
||||
logger.info("Skipping {}".format("/".join(name)))
|
||||
continue
|
||||
pointer = model
|
||||
for m_name in name:
|
||||
if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
|
||||
l = re.split(r'_(\d+)', m_name)
|
||||
else:
|
||||
l = [m_name]
|
||||
if l[0] == 'kernel' or l[0] == 'gamma':
|
||||
pointer = getattr(pointer, 'weight')
|
||||
elif l[0] == 'output_bias' or l[0] == 'beta':
|
||||
pointer = getattr(pointer, 'bias')
|
||||
elif l[0] == 'output_weights':
|
||||
pointer = getattr(pointer, 'weight')
|
||||
elif l[0] == 'squad':
|
||||
pointer = getattr(pointer, 'classifier')
|
||||
else:
|
||||
try:
|
||||
pointer = getattr(pointer, l[0])
|
||||
except AttributeError:
|
||||
logger.info("Skipping {}".format("/".join(name)))
|
||||
continue
|
||||
if len(l) >= 2:
|
||||
num = int(l[1])
|
||||
pointer = pointer[num]
|
||||
if m_name[-11:] == '_embeddings':
|
||||
pointer = getattr(pointer, 'weight')
|
||||
elif m_name == 'kernel':
|
||||
array = np.transpose(array)
|
||||
try:
|
||||
assert pointer.shape == array.shape
|
||||
except AssertionError as e:
|
||||
e.args += (pointer.shape, array.shape)
|
||||
raise
|
||||
logger.info("Initialize PyTorch weight {}".format(name))
|
||||
pointer.data = torch.from_numpy(array)
|
||||
return model
|
||||
|
||||
|
||||
####################################################
|
||||
# PyTorch Models are constructed by sub-classing
|
||||
# - torch.nn.Module for the layers and
|
||||
# - PreTrainedModel for the models (itself a sub-class of torch.nn.Module)
|
||||
####################################################
|
||||
|
||||
####################################################
|
||||
# Here is an example of typical layer in a PyTorch model of the library
|
||||
# The classes are usually identical to the TF 2.0 ones without the 'TF' prefix.
|
||||
#
|
||||
# See the conversion methods in modeling_tf_pytorch_utils.py for more details
|
||||
####################################################
|
||||
class XxxLayer(nn.Module):
|
||||
def __init__(self, config):
|
||||
super(XxxLayer, self).__init__()
|
||||
self.attention = XxxAttention(config)
|
||||
self.intermediate = XxxIntermediate(config)
|
||||
self.output = XxxOutput(config)
|
||||
|
||||
def forward(self, hidden_states, attention_mask=None, head_mask=None):
|
||||
attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
|
||||
attention_output = attention_outputs[0]
|
||||
intermediate_output = self.intermediate(attention_output)
|
||||
layer_output = self.output(intermediate_output, attention_output)
|
||||
outputs = (layer_output,) + attention_outputs[1:] # add attentions if we output them
|
||||
return outputs
|
||||
|
||||
|
||||
|
||||
####################################################
|
||||
# PreTrainedModel is a sub-class of torch.nn.Module
|
||||
# which take care of loading and saving pretrained weights
|
||||
# and various common utilities.
|
||||
#
|
||||
# Here you just need to specify a few (self-explanatory)
|
||||
# pointers for your model and the weights initialization
|
||||
# method if its not fully covered by PreTrainedModel's default method
|
||||
####################################################
|
||||
class XxxPreTrainedModel(PreTrainedModel):
|
||||
""" An abstract class to handle weights initialization and
|
||||
a simple interface for dowloading and loading pretrained models.
|
||||
"""
|
||||
config_class = XxxConfig
|
||||
pretrained_model_archive_map = XXX_PRETRAINED_MODEL_ARCHIVE_MAP
|
||||
load_tf_weights = load_tf_weights_in_xxx
|
||||
base_model_prefix = "transformer"
|
||||
|
||||
def _init_weights(self, module):
|
||||
""" Initialize the weights """
|
||||
if isinstance(module, (nn.Linear, nn.Embedding)):
|
||||
# Slightly different from the TF version which uses truncated_normal for initialization
|
||||
# cf https://github.com/pytorch/pytorch/pull/5617
|
||||
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
||||
elif isinstance(module, XxxLayerNorm):
|
||||
module.bias.data.zero_()
|
||||
module.weight.data.fill_(1.0)
|
||||
if isinstance(module, nn.Linear) and module.bias is not None:
|
||||
module.bias.data.zero_()
|
||||
|
||||
|
||||
XXX_START_DOCSTRING = r""" The XXX model was proposed in
|
||||
`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_
|
||||
by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova. It's a bidirectional transformer
|
||||
pre-trained using a combination of masked language modeling objective and next sentence prediction
|
||||
on a large corpus comprising the Toronto Book Corpus and Wikipedia.
|
||||
|
||||
This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
|
||||
refer to the PyTorch documentation for all matter related to general usage and behavior.
|
||||
|
||||
.. _`XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`:
|
||||
https://arxiv.org/abs/1810.04805
|
||||
|
||||
.. _`torch.nn.Module`:
|
||||
https://pytorch.org/docs/stable/nn.html#module
|
||||
|
||||
Parameters:
|
||||
config (:class:`~transformers.XxxConfig`): Model configuration class with all the parameters of the model.
|
||||
Initializing with a config file does not load the weights associated with the model, only the configuration.
|
||||
Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
|
||||
"""
|
||||
|
||||
XXX_INPUTS_DOCSTRING = r"""
|
||||
Inputs:
|
||||
**input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Indices of input sequence tokens in the vocabulary.
|
||||
To match pre-training, XXX input sequence should be formatted with [CLS] and [SEP] tokens as follows:
|
||||
|
||||
(a) For sequence pairs:
|
||||
|
||||
``tokens: [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
|
||||
|
||||
``token_type_ids: 0 0 0 0 0 0 0 0 1 1 1 1 1 1``
|
||||
|
||||
(b) For single sequences:
|
||||
|
||||
``tokens: [CLS] the dog is hairy . [SEP]``
|
||||
|
||||
``token_type_ids: 0 0 0 0 0 0 0``
|
||||
|
||||
Xxx is a model with absolute position embeddings so it's usually advised to pad the inputs on
|
||||
the right rather than the left.
|
||||
|
||||
Indices can be obtained using :class:`transformers.XxxTokenizer`.
|
||||
See :func:`transformers.PreTrainedTokenizer.encode` and
|
||||
:func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
|
||||
**attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Mask to avoid performing attention on padding token indices.
|
||||
Mask values selected in ``[0, 1]``:
|
||||
``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
|
||||
**token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Segment token indices to indicate first and second portions of the inputs.
|
||||
Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
|
||||
corresponds to a `sentence B` token
|
||||
(see `XXX: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
|
||||
**position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Indices of positions of each input sequence tokens in the position embeddings.
|
||||
Selected in the range ``[0, config.max_position_embeddings - 1]``.
|
||||
**head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
|
||||
Mask to nullify selected heads of the self-attention modules.
|
||||
Mask values selected in ``[0, 1]``:
|
||||
``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
|
||||
**inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
|
||||
Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
|
||||
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
||||
than the model's internal embedding lookup matrix.
|
||||
"""
|
||||
|
||||
@add_start_docstrings("The bare Xxx Model transformer outputting raw hidden-states without any specific head on top.",
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class XxxModel(XxxPreTrainedModel):
|
||||
r"""
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
|
||||
Sequence of hidden-states at the output of the last layer of the model.
|
||||
**pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
|
||||
Last layer hidden-state of the first token of the sequence (classification token)
|
||||
further processed by a Linear layer and a Tanh activation function. The Linear
|
||||
layer weights are trained from the next sentence prediction (classification)
|
||||
objective during Xxx pretraining. This output is usually *not* a good summary
|
||||
of the semantic content of the input, you're often better with averaging or pooling
|
||||
the sequence of hidden-states for the whole input sequence.
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = XxxModel.from_pretrained('xxx-base-uncased')
|
||||
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
|
||||
outputs = model(input_ids)
|
||||
last_hidden_states = outputs[0] # The last hidden-state is the first element of the output tuple
|
||||
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(XxxModel, self).__init__(config)
|
||||
|
||||
self.embeddings = XxxEmbeddings(config)
|
||||
self.encoder = XxxEncoder(config)
|
||||
self.pooler = XxxPooler(config)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def get_input_embeddings(self):
|
||||
return self.embeddings.word_embeddings
|
||||
|
||||
def set_input_embeddings(self, new_embeddings):
|
||||
self.embeddings.word_embeddings = new_embeddings
|
||||
|
||||
def _prune_heads(self, heads_to_prune):
|
||||
""" Prunes heads of the model.
|
||||
heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
|
||||
See base class PreTrainedModel
|
||||
"""
|
||||
for layer, heads in heads_to_prune.items():
|
||||
self.encoder.layer[layer].attention.prune_heads(heads)
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None):
|
||||
if input_ids is not None and inputs_embeds is not None:
|
||||
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
||||
elif input_ids is not None:
|
||||
input_shape = input_ids.size()
|
||||
elif inputs_embeds is not None:
|
||||
input_shape = inputs_embeds.size()[:-1]
|
||||
else:
|
||||
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
||||
|
||||
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
||||
|
||||
if attention_mask is None:
|
||||
attention_mask = torch.ones(input_shape, device=device)
|
||||
if token_type_ids is None:
|
||||
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
|
||||
|
||||
# We create a 3D attention mask from a 2D tensor mask.
|
||||
# Sizes are [batch_size, 1, 1, to_seq_length]
|
||||
# So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
|
||||
# this attention mask is more simple than the triangular masking of causal attention
|
||||
# used in OpenAI GPT, we just need to prepare the broadcast dimension here.
|
||||
extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
|
||||
|
||||
# Since attention_mask is 1.0 for positions we want to attend and 0.0 for
|
||||
# masked positions, this operation will create a tensor which is 0.0 for
|
||||
# positions we want to attend and -10000.0 for masked positions.
|
||||
# Since we are adding it to the raw scores before the softmax, this is
|
||||
# effectively the same as removing these entirely.
|
||||
extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
|
||||
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
|
||||
|
||||
# Prepare head mask if needed
|
||||
# 1.0 in head_mask indicate we keep the head
|
||||
# attention_probs has shape bsz x n_heads x N x N
|
||||
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
|
||||
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
|
||||
if head_mask is not None:
|
||||
if head_mask.dim() == 1:
|
||||
head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
|
||||
head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
|
||||
elif head_mask.dim() == 2:
|
||||
head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1) # We can specify head_mask for each layer
|
||||
head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
|
||||
else:
|
||||
head_mask = [None] * self.config.num_hidden_layers
|
||||
|
||||
##################################
|
||||
# Replace this with your model code
|
||||
embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds)
|
||||
encoder_outputs = self.encoder(embedding_output, extended_attention_mask, head_mask=head_mask)
|
||||
sequence_output = encoder_outputs[0]
|
||||
outputs = (sequence_output,) + encoder_outputs[1:] # add hidden_states and attentions if they are here
|
||||
|
||||
return outputs # sequence_output, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a `language modeling` head on top. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class XxxForMaskedLM(XxxPreTrainedModel):
|
||||
r"""
|
||||
**masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Labels for computing the masked language modeling loss.
|
||||
Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
|
||||
Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
|
||||
in ``[0, ..., config.vocab_size]``
|
||||
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
|
||||
Masked language modeling loss.
|
||||
**prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
|
||||
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = XxxForMaskedLM.from_pretrained('xxx-base-uncased')
|
||||
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
|
||||
outputs = model(input_ids, masked_lm_labels=input_ids)
|
||||
loss, prediction_scores = outputs[:2]
|
||||
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(XxxForMaskedLM, self).__init__(config)
|
||||
|
||||
self.transformer = XxxModel(config)
|
||||
self.lm_head = nn.Linear(config.n_embd, config.vocab_size)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def get_output_embeddings(self):
|
||||
return self.lm_head
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
|
||||
masked_lm_labels=None):
|
||||
|
||||
outputs = self.transformer(input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
prediction_scores = self.cls(sequence_output)
|
||||
|
||||
outputs = (prediction_scores,) + outputs[2:] # Add hidden states and attention if they are here
|
||||
if masked_lm_labels is not None:
|
||||
loss_fct = CrossEntropyLoss(ignore_index=-1)
|
||||
masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
|
||||
outputs = (masked_lm_loss,) + outputs
|
||||
|
||||
return outputs # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model transformer with a sequence classification/regression head on top (a linear layer on top of
|
||||
the pooled output) e.g. for GLUE tasks. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class XxxForSequenceClassification(XxxPreTrainedModel):
|
||||
r"""
|
||||
**labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
|
||||
Labels for computing the sequence classification/regression loss.
|
||||
Indices should be in ``[0, ..., config.num_labels - 1]``.
|
||||
If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
|
||||
If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
|
||||
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
|
||||
Classification (or regression if config.num_labels==1) loss.
|
||||
**logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
|
||||
Classification (or regression if config.num_labels==1) scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = XxxForSequenceClassification.from_pretrained('xxx-base-uncased')
|
||||
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
|
||||
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
|
||||
outputs = model(input_ids, labels=labels)
|
||||
loss, logits = outputs[:2]
|
||||
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(XxxForSequenceClassification, self).__init__(config)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = XxxModel(config)
|
||||
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
||||
self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
|
||||
position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
|
||||
|
||||
outputs = self.transformer(input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds)
|
||||
|
||||
pooled_output = outputs[1]
|
||||
|
||||
pooled_output = self.dropout(pooled_output)
|
||||
logits = self.classifier(pooled_output)
|
||||
|
||||
outputs = (logits,) + outputs[2:] # add hidden states and attention if they are here
|
||||
|
||||
if labels is not None:
|
||||
if self.num_labels == 1:
|
||||
# We are doing regression
|
||||
loss_fct = MSELoss()
|
||||
loss = loss_fct(logits.view(-1), labels.view(-1))
|
||||
else:
|
||||
loss_fct = CrossEntropyLoss()
|
||||
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
||||
outputs = (loss,) + outputs
|
||||
|
||||
return outputs # (loss), logits, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a token classification head on top (a linear layer on top of
|
||||
the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class XxxForTokenClassification(XxxPreTrainedModel):
|
||||
r"""
|
||||
**labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
|
||||
Labels for computing the token classification loss.
|
||||
Indices should be in ``[0, ..., config.num_labels - 1]``.
|
||||
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
|
||||
Classification loss.
|
||||
**scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
|
||||
Classification scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = XxxForTokenClassification.from_pretrained('xxx-base-uncased')
|
||||
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0) # Batch size 1
|
||||
labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0) # Batch size 1
|
||||
outputs = model(input_ids, labels=labels)
|
||||
loss, scores = outputs[:2]
|
||||
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(XxxForTokenClassification, self).__init__(config)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = XxxModel(config)
|
||||
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
||||
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None,
|
||||
position_ids=None, head_mask=None, inputs_embeds=None, labels=None):
|
||||
|
||||
outputs = self.transformer(input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
|
||||
sequence_output = self.dropout(sequence_output)
|
||||
logits = self.classifier(sequence_output)
|
||||
|
||||
outputs = (logits,) + outputs[2:] # add hidden states and attention if they are here
|
||||
if labels is not None:
|
||||
loss_fct = CrossEntropyLoss()
|
||||
# Only keep active parts of the loss
|
||||
if attention_mask is not None:
|
||||
active_loss = attention_mask.view(-1) == 1
|
||||
active_logits = logits.view(-1, self.num_labels)[active_loss]
|
||||
active_labels = labels.view(-1)[active_loss]
|
||||
loss = loss_fct(active_logits, active_labels)
|
||||
else:
|
||||
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
|
||||
outputs = (loss,) + outputs
|
||||
|
||||
return outputs # (loss), scores, (hidden_states), (attentions)
|
||||
|
||||
|
||||
@add_start_docstrings("""Xxx Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
|
||||
the hidden-states output to compute `span start logits` and `span end logits`). """,
|
||||
XXX_START_DOCSTRING, XXX_INPUTS_DOCSTRING)
|
||||
class XxxForQuestionAnswering(XxxPreTrainedModel):
|
||||
r"""
|
||||
**start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
|
||||
Labels for position (index) of the start of the labelled span for computing the token classification loss.
|
||||
Positions are clamped to the length of the sequence (`sequence_length`).
|
||||
Position outside of the sequence are not taken into account for computing the loss.
|
||||
**end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
|
||||
Labels for position (index) of the end of the labelled span for computing the token classification loss.
|
||||
Positions are clamped to the length of the sequence (`sequence_length`).
|
||||
Position outside of the sequence are not taken into account for computing the loss.
|
||||
|
||||
Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
|
||||
**loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
|
||||
Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
|
||||
**start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
|
||||
Span-start scores (before SoftMax).
|
||||
**end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
|
||||
Span-end scores (before SoftMax).
|
||||
**hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
|
||||
list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
|
||||
of shape ``(batch_size, sequence_length, hidden_size)``:
|
||||
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
||||
**attentions**: (`optional`, returned when ``config.output_attentions=True``)
|
||||
list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
|
||||
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
||||
|
||||
Examples::
|
||||
|
||||
tokenizer = XxxTokenizer.from_pretrained('xxx-base-uncased')
|
||||
model = XxxForQuestionAnswering.from_pretrained('xxx-large-uncased-whole-word-masking-finetuned-squad')
|
||||
question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
|
||||
input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
|
||||
input_ids = tokenizer.encode(input_text)
|
||||
token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]
|
||||
start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
|
||||
all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
|
||||
print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
|
||||
# a nice puppet
|
||||
|
||||
|
||||
"""
|
||||
def __init__(self, config):
|
||||
super(XxxForQuestionAnswering, self).__init__(config)
|
||||
self.num_labels = config.num_labels
|
||||
|
||||
self.transformer = XxxModel(config)
|
||||
self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
|
||||
|
||||
self.init_weights()
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None,
|
||||
start_positions=None, end_positions=None):
|
||||
|
||||
outputs = self.transformer(input_ids,
|
||||
attention_mask=attention_mask,
|
||||
token_type_ids=token_type_ids,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds)
|
||||
|
||||
sequence_output = outputs[0]
|
||||
|
||||
logits = self.qa_outputs(sequence_output)
|
||||
start_logits, end_logits = logits.split(1, dim=-1)
|
||||
start_logits = start_logits.squeeze(-1)
|
||||
end_logits = end_logits.squeeze(-1)
|
||||
|
||||
outputs = (start_logits, end_logits,) + outputs[2:]
|
||||
if start_positions is not None and end_positions is not None:
|
||||
# If we are on multi-GPU, split add a dimension
|
||||
if len(start_positions.size()) > 1:
|
||||
start_positions = start_positions.squeeze(-1)
|
||||
if len(end_positions.size()) > 1:
|
||||
end_positions = end_positions.squeeze(-1)
|
||||
# sometimes the start/end positions are outside our model inputs, we ignore these terms
|
||||
ignored_index = start_logits.size(1)
|
||||
start_positions.clamp_(0, ignored_index)
|
||||
end_positions.clamp_(0, ignored_index)
|
||||
|
||||
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
|
||||
start_loss = loss_fct(start_logits, start_positions)
|
||||
end_loss = loss_fct(end_logits, end_positions)
|
||||
total_loss = (start_loss + end_loss) / 2
|
||||
outputs = (total_loss,) + outputs
|
||||
|
||||
return outputs # (loss), start_logits, end_logits, (hidden_states), (attentions)
|
||||
@@ -1,255 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX Authors.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import unittest
|
||||
import shutil
|
||||
import sys
|
||||
|
||||
from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
|
||||
from .configuration_common_test import ConfigTester
|
||||
from .utils import require_tf, slow
|
||||
|
||||
from transformers import XxxConfig, is_tf_available
|
||||
|
||||
if is_tf_available():
|
||||
import tensorflow as tf
|
||||
from transformers.modeling_tf_xxx import (TFXxxModel, TFXxxForMaskedLM,
|
||||
TFXxxForSequenceClassification,
|
||||
TFXxxForTokenClassification,
|
||||
TFXxxForQuestionAnswering,
|
||||
TF_XXX_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
|
||||
@require_tf
|
||||
class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
|
||||
|
||||
all_model_classes = (TFXxxModel, TFXxxForMaskedLM, TFXxxForQuestionAnswering,
|
||||
TFXxxForSequenceClassification,
|
||||
TFXxxForTokenClassification) if is_tf_available() else ()
|
||||
|
||||
class TFXxxModelTester(object):
|
||||
|
||||
def __init__(self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=True,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.scope = scope
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
||||
|
||||
config = XxxConfig(
|
||||
vocab_size_or_config_json_file=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
initializer_range=self.initializer_range)
|
||||
|
||||
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
|
||||
def create_and_check_xxx_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = TFXxxModel(config=config)
|
||||
inputs = {'input_ids': input_ids,
|
||||
'attention_mask': input_mask,
|
||||
'token_type_ids': token_type_ids}
|
||||
sequence_output, pooled_output = model(inputs)
|
||||
|
||||
inputs = [input_ids, input_mask]
|
||||
sequence_output, pooled_output = model(inputs)
|
||||
|
||||
sequence_output, pooled_output = model(input_ids)
|
||||
|
||||
result = {
|
||||
"sequence_output": sequence_output.numpy(),
|
||||
"pooled_output": pooled_output.numpy(),
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["sequence_output"].shape),
|
||||
[self.batch_size, self.seq_length, self.hidden_size])
|
||||
self.parent.assertListEqual(list(result["pooled_output"].shape), [self.batch_size, self.hidden_size])
|
||||
|
||||
|
||||
def create_and_check_xxx_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = TFXxxForMaskedLM(config=config)
|
||||
inputs = {'input_ids': input_ids,
|
||||
'attention_mask': input_mask,
|
||||
'token_type_ids': token_type_ids}
|
||||
prediction_scores, = model(inputs)
|
||||
result = {
|
||||
"prediction_scores": prediction_scores.numpy(),
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["prediction_scores"].shape),
|
||||
[self.batch_size, self.seq_length, self.vocab_size])
|
||||
|
||||
|
||||
def create_and_check_xxx_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
config.num_labels = self.num_labels
|
||||
model = TFXxxForSequenceClassification(config=config)
|
||||
inputs = {'input_ids': input_ids,
|
||||
'attention_mask': input_mask,
|
||||
'token_type_ids': token_type_ids}
|
||||
logits, = model(inputs)
|
||||
result = {
|
||||
"logits": logits.numpy(),
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["logits"].shape),
|
||||
[self.batch_size, self.num_labels])
|
||||
|
||||
|
||||
def create_and_check_xxx_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
config.num_labels = self.num_labels
|
||||
model = TFXxxForTokenClassification(config=config)
|
||||
inputs = {'input_ids': input_ids,
|
||||
'attention_mask': input_mask,
|
||||
'token_type_ids': token_type_ids}
|
||||
logits, = model(inputs)
|
||||
result = {
|
||||
"logits": logits.numpy(),
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["logits"].shape),
|
||||
[self.batch_size, self.seq_length, self.num_labels])
|
||||
|
||||
|
||||
def create_and_check_xxx_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = TFXxxForQuestionAnswering(config=config)
|
||||
inputs = {'input_ids': input_ids,
|
||||
'attention_mask': input_mask,
|
||||
'token_type_ids': token_type_ids}
|
||||
start_logits, end_logits = model(inputs)
|
||||
result = {
|
||||
"start_logits": start_logits.numpy(),
|
||||
"end_logits": end_logits.numpy(),
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["start_logits"].shape),
|
||||
[self.batch_size, self.seq_length])
|
||||
self.parent.assertListEqual(
|
||||
list(result["end_logits"].shape),
|
||||
[self.batch_size, self.seq_length])
|
||||
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
(config, input_ids, token_type_ids, input_mask,
|
||||
sequence_labels, token_labels, choice_labels) = config_and_inputs
|
||||
inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = TFXxxModelTest.TFXxxModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=XxxConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_xxx_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_model(*config_and_inputs)
|
||||
|
||||
def test_for_masked_lm(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_masked_lm(*config_and_inputs)
|
||||
|
||||
def test_for_question_answering(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_question_answering(*config_and_inputs)
|
||||
|
||||
def test_for_sequence_classification(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_sequence_classification(*config_and_inputs)
|
||||
|
||||
def test_for_token_classification(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
cache_dir = "/tmp/transformers_test/"
|
||||
for model_name in ['xxx-base-uncased']:
|
||||
model = TFXxxModel.from_pretrained(model_name, cache_dir=cache_dir)
|
||||
shutil.rmtree(cache_dir)
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -1,259 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX Authors.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from __future__ import absolute_import
|
||||
from __future__ import division
|
||||
from __future__ import print_function
|
||||
|
||||
import unittest
|
||||
import shutil
|
||||
|
||||
from transformers import is_torch_available
|
||||
|
||||
from .modeling_common_test import (CommonTestCases, ids_tensor)
|
||||
from .configuration_common_test import ConfigTester
|
||||
from .utils import require_torch, slow, torch_device
|
||||
|
||||
if is_torch_available():
|
||||
from transformers import (XxxConfig, XxxModel, XxxForMaskedLM,
|
||||
XxxForNextSentencePrediction, XxxForPreTraining,
|
||||
XxxForQuestionAnswering, XxxForSequenceClassification,
|
||||
XxxForTokenClassification, XxxForMultipleChoice)
|
||||
from transformers.modeling_xxx import XXX_PRETRAINED_MODEL_ARCHIVE_MAP
|
||||
|
||||
|
||||
@require_torch
|
||||
class XxxModelTest(CommonTestCases.CommonModelTester):
|
||||
|
||||
all_model_classes = (XxxModel, XxxForMaskedLM, XxxForQuestionAnswering,
|
||||
XxxForSequenceClassification,
|
||||
XxxForTokenClassification) if is_torch_available() else ()
|
||||
|
||||
class XxxModelTester(object):
|
||||
|
||||
def __init__(self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=True,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.scope = scope
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
||||
|
||||
config = XxxConfig(
|
||||
vocab_size_or_config_json_file=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
initializer_range=self.initializer_range)
|
||||
|
||||
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
|
||||
def check_loss_output(self, result):
|
||||
self.parent.assertListEqual(
|
||||
list(result["loss"].size()),
|
||||
[])
|
||||
|
||||
def create_and_check_xxx_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = XxxModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
||||
sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
|
||||
sequence_output, pooled_output = model(input_ids)
|
||||
|
||||
result = {
|
||||
"sequence_output": sequence_output,
|
||||
"pooled_output": pooled_output,
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["sequence_output"].size()),
|
||||
[self.batch_size, self.seq_length, self.hidden_size])
|
||||
self.parent.assertListEqual(list(result["pooled_output"].size()), [self.batch_size, self.hidden_size])
|
||||
|
||||
|
||||
def create_and_check_xxx_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = XxxForMaskedLM(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
|
||||
result = {
|
||||
"loss": loss,
|
||||
"prediction_scores": prediction_scores,
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["prediction_scores"].size()),
|
||||
[self.batch_size, self.seq_length, self.vocab_size])
|
||||
self.check_loss_output(result)
|
||||
|
||||
|
||||
def create_and_check_xxx_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
model = XxxForQuestionAnswering(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
|
||||
start_positions=sequence_labels, end_positions=sequence_labels)
|
||||
result = {
|
||||
"loss": loss,
|
||||
"start_logits": start_logits,
|
||||
"end_logits": end_logits,
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["start_logits"].size()),
|
||||
[self.batch_size, self.seq_length])
|
||||
self.parent.assertListEqual(
|
||||
list(result["end_logits"].size()),
|
||||
[self.batch_size, self.seq_length])
|
||||
self.check_loss_output(result)
|
||||
|
||||
|
||||
def create_and_check_xxx_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
config.num_labels = self.num_labels
|
||||
model = XxxForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
|
||||
result = {
|
||||
"loss": loss,
|
||||
"logits": logits,
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["logits"].size()),
|
||||
[self.batch_size, self.num_labels])
|
||||
self.check_loss_output(result)
|
||||
|
||||
|
||||
def create_and_check_xxx_for_token_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
|
||||
config.num_labels = self.num_labels
|
||||
model = XxxForTokenClassification(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
|
||||
result = {
|
||||
"loss": loss,
|
||||
"logits": logits,
|
||||
}
|
||||
self.parent.assertListEqual(
|
||||
list(result["logits"].size()),
|
||||
[self.batch_size, self.seq_length, self.num_labels])
|
||||
self.check_loss_output(result)
|
||||
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
(config, input_ids, token_type_ids, input_mask,
|
||||
sequence_labels, token_labels, choice_labels) = config_and_inputs
|
||||
inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = XxxModelTest.XxxModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=XxxConfig, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_xxx_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_model(*config_and_inputs)
|
||||
|
||||
def test_for_masked_lm(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_masked_lm(*config_and_inputs)
|
||||
|
||||
def test_for_question_answering(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_question_answering(*config_and_inputs)
|
||||
|
||||
def test_for_sequence_classification(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_sequence_classification(*config_and_inputs)
|
||||
|
||||
def test_for_token_classification(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_xxx_for_token_classification(*config_and_inputs)
|
||||
|
||||
@slow
|
||||
def test_model_from_pretrained(self):
|
||||
cache_dir = "/tmp/transformers_test/"
|
||||
for model_name in list(XXX_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
|
||||
model = XxxModel.from_pretrained(model_name, cache_dir=cache_dir)
|
||||
shutil.rmtree(cache_dir)
|
||||
self.assertIsNotNone(model)
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -1,57 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX Authors.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
import os
|
||||
import unittest
|
||||
from io import open
|
||||
|
||||
from transformers.tokenization_bert import (XxxTokenizer, VOCAB_FILES_NAMES)
|
||||
|
||||
from .tokenization_tests_commons import CommonTestCases
|
||||
|
||||
class XxxTokenizationTest(CommonTestCases.CommonTokenizerTester):
|
||||
|
||||
tokenizer_class = XxxTokenizer
|
||||
|
||||
def setUp(self):
|
||||
super(XxxTokenizationTest, self).setUp()
|
||||
|
||||
vocab_tokens = [
|
||||
"[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn",
|
||||
"##ing", ",", "low", "lowest",
|
||||
]
|
||||
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
|
||||
with open(self.vocab_file, "w", encoding='utf-8') as vocab_writer:
|
||||
vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
|
||||
|
||||
def get_tokenizer(self, **kwargs):
|
||||
return XxxTokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_input_output_texts(self):
|
||||
input_text = u"UNwant\u00E9d,running"
|
||||
output_text = u"unwanted, running"
|
||||
return input_text, output_text
|
||||
|
||||
def test_full_tokenizer(self):
|
||||
tokenizer = self.tokenizer_class(self.vocab_file)
|
||||
|
||||
tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
|
||||
self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
|
||||
self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
unittest.main()
|
||||
@@ -1,218 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 XXX Authors.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Tokenization class for model XXX."""
|
||||
|
||||
from __future__ import absolute_import, division, print_function, unicode_literals
|
||||
|
||||
import collections
|
||||
import logging
|
||||
import os
|
||||
import unicodedata
|
||||
from io import open
|
||||
|
||||
from .tokenization_utils import PreTrainedTokenizer
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
####################################################
|
||||
# In this template, replace all the XXX (various casings) with your model name
|
||||
####################################################
|
||||
|
||||
####################################################
|
||||
# Mapping from the keyword arguments names of Tokenizer `__init__`
|
||||
# to file names for serializing Tokenizer instances
|
||||
####################################################
|
||||
VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
|
||||
|
||||
####################################################
|
||||
# Mapping from the keyword arguments names of Tokenizer `__init__`
|
||||
# to pretrained vocabulary URL for all the model shortcut names.
|
||||
####################################################
|
||||
PRETRAINED_VOCAB_FILES_MAP = {
|
||||
'vocab_file':
|
||||
{
|
||||
'xxx-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-base-uncased-vocab.txt",
|
||||
'xxx-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/xxx-large-uncased-vocab.txt",
|
||||
}
|
||||
}
|
||||
|
||||
####################################################
|
||||
# Mapping from model shortcut names to max length of inputs
|
||||
####################################################
|
||||
PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
|
||||
'xxx-base-uncased': 512,
|
||||
'xxx-large-uncased': 512,
|
||||
}
|
||||
|
||||
####################################################
|
||||
# Mapping from model shortcut names to a dictionary of additional
|
||||
# keyword arguments for Tokenizer `__init__`.
|
||||
# To be used for checkpoint specific configurations.
|
||||
####################################################
|
||||
PRETRAINED_INIT_CONFIGURATION = {
|
||||
'xxx-base-uncased': {'do_lower_case': True},
|
||||
'xxx-large-uncased': {'do_lower_case': True},
|
||||
}
|
||||
|
||||
|
||||
def load_vocab(vocab_file):
|
||||
"""Loads a vocabulary file into a dictionary."""
|
||||
vocab = collections.OrderedDict()
|
||||
with open(vocab_file, "r", encoding="utf-8") as reader:
|
||||
tokens = reader.readlines()
|
||||
for index, token in enumerate(tokens):
|
||||
token = token.rstrip('\n')
|
||||
vocab[token] = index
|
||||
return vocab
|
||||
|
||||
|
||||
class XxxTokenizer(PreTrainedTokenizer):
|
||||
r"""
|
||||
Constructs a XxxTokenizer.
|
||||
:class:`~transformers.XxxTokenizer` runs end-to-end tokenization: punctuation splitting + wordpiece
|
||||
|
||||
Args:
|
||||
vocab_file: Path to a one-wordpiece-per-line vocabulary file
|
||||
do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
|
||||
"""
|
||||
|
||||
vocab_files_names = VOCAB_FILES_NAMES
|
||||
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
||||
pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
|
||||
max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
|
||||
|
||||
def __init__(self, vocab_file, do_lower_case=True,
|
||||
unk_token="[UNK]", sep_token="[SEP]", pad_token="[PAD]", cls_token="[CLS]",
|
||||
mask_token="[MASK]", **kwargs):
|
||||
"""Constructs a XxxTokenizer.
|
||||
|
||||
Args:
|
||||
**vocab_file**: Path to a one-wordpiece-per-line vocabulary file
|
||||
**do_lower_case**: (`optional`) boolean (default True)
|
||||
Whether to lower case the input
|
||||
Only has an effect when do_basic_tokenize=True
|
||||
"""
|
||||
super(XxxTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
|
||||
pad_token=pad_token, cls_token=cls_token,
|
||||
mask_token=mask_token, **kwargs)
|
||||
self.max_len_single_sentence = self.max_len - 2 # take into account special tokens
|
||||
self.max_len_sentences_pair = self.max_len - 3 # take into account special tokens
|
||||
|
||||
if not os.path.isfile(vocab_file):
|
||||
raise ValueError(
|
||||
"Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
|
||||
"model use `tokenizer = XxxTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
|
||||
self.vocab = load_vocab(vocab_file)
|
||||
|
||||
@property
|
||||
def vocab_size(self):
|
||||
return len(self.vocab)
|
||||
|
||||
def _tokenize(self, text):
|
||||
""" Take as input a string and return a list of strings (tokens) for words/sub-words
|
||||
"""
|
||||
split_tokens = []
|
||||
if self.do_basic_tokenize:
|
||||
for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
|
||||
for sub_token in self.wordpiece_tokenizer.tokenize(token):
|
||||
split_tokens.append(sub_token)
|
||||
else:
|
||||
split_tokens = self.wordpiece_tokenizer.tokenize(text)
|
||||
return split_tokens
|
||||
|
||||
def _convert_token_to_id(self, token):
|
||||
""" Converts a token (str/unicode) in an id using the vocab. """
|
||||
return self.vocab.get(token, self.vocab.get(self.unk_token))
|
||||
|
||||
def _convert_id_to_token(self, index):
|
||||
"""Converts an index (integer) in a token (string/unicode) using the vocab."""
|
||||
return self.ids_to_tokens.get(index, self.unk_token)
|
||||
|
||||
def convert_tokens_to_string(self, tokens):
|
||||
""" Converts a sequence of tokens (string) in a single string. """
|
||||
out_string = ' '.join(tokens).replace(' ##', '').strip()
|
||||
return out_string
|
||||
|
||||
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
|
||||
"""
|
||||
Build model inputs from a sequence or a pair of sequence for sequence classification tasks
|
||||
by concatenating and adding special tokens.
|
||||
A BERT sequence has the following format:
|
||||
single sequence: [CLS] X [SEP]
|
||||
pair of sequences: [CLS] A [SEP] B [SEP]
|
||||
"""
|
||||
if token_ids_1 is None:
|
||||
return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
|
||||
cls = [self.cls_token_id]
|
||||
sep = [self.sep_token_id]
|
||||
return cls + token_ids_0 + sep + token_ids_1 + sep
|
||||
|
||||
def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
|
||||
"""
|
||||
Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
|
||||
special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
|
||||
|
||||
Args:
|
||||
token_ids_0: list of ids (must not contain special tokens)
|
||||
token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
|
||||
for sequence pairs
|
||||
already_has_special_tokens: (default False) Set to True if the token list is already formated with
|
||||
special tokens for the model
|
||||
|
||||
Returns:
|
||||
A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
|
||||
"""
|
||||
|
||||
if already_has_special_tokens:
|
||||
if token_ids_1 is not None:
|
||||
raise ValueError("You should not supply a second sequence if the provided sequence of "
|
||||
"ids is already formated with special tokens for the model.")
|
||||
return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
|
||||
|
||||
if token_ids_1 is not None:
|
||||
return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
|
||||
return [1] + ([0] * len(token_ids_0)) + [1]
|
||||
|
||||
def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
|
||||
"""
|
||||
Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
|
||||
A BERT sequence pair mask has the following format:
|
||||
0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
|
||||
| first sequence | second sequence
|
||||
|
||||
if token_ids_1 is None, only returns the first portion of the mask (0's).
|
||||
"""
|
||||
sep = [self.sep_token_id]
|
||||
cls = [self.cls_token_id]
|
||||
if token_ids_1 is None:
|
||||
return len(cls + token_ids_0 + sep) * [0]
|
||||
return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
|
||||
|
||||
def save_vocabulary(self, vocab_path):
|
||||
"""Save the tokenizer vocabulary to a directory or file."""
|
||||
index = 0
|
||||
if os.path.isdir(vocab_path):
|
||||
vocab_file = os.path.join(vocab_path, VOCAB_FILES_NAMES['vocab_file'])
|
||||
else:
|
||||
vocab_file = vocab_path
|
||||
with open(vocab_file, "w", encoding="utf-8") as writer:
|
||||
for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
|
||||
if index != token_index:
|
||||
logger.warning("Saving vocabulary to {}: vocabulary indices are not consecutive."
|
||||
" Please check that the vocabulary is not corrupted!".format(vocab_file))
|
||||
index = token_index
|
||||
writer.write(token + u'\n')
|
||||
index += 1
|
||||
return (vocab_file,)
|
||||
@@ -1,23 +0,0 @@
|
||||
#!/usr/bin/env python
|
||||
from argparse import ArgumentParser
|
||||
|
||||
from transformers.commands.user import UserCommands
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = ArgumentParser(description='Transformers CLI tool', usage='transformers-cli <command> [<args>]')
|
||||
commands_parser = parser.add_subparsers(help='transformers-cli command helpers')
|
||||
|
||||
# Register commands
|
||||
UserCommands.register_subcommand(commands_parser)
|
||||
|
||||
# Let's go
|
||||
args = parser.parse_args()
|
||||
|
||||
if not hasattr(args, 'func'):
|
||||
parser.print_help()
|
||||
exit(1)
|
||||
|
||||
# Run
|
||||
service = args.func(args)
|
||||
service.run()
|
||||
+28
-67
@@ -1,4 +1,4 @@
|
||||
__version__ = "2.2.2"
|
||||
__version__ = "2.0.0"
|
||||
|
||||
# Work around to update TensorFlow's absl.logging threshold which alters the
|
||||
# default Python logging output behavior when present.
|
||||
@@ -25,30 +25,22 @@ from .file_utils import (TRANSFORMERS_CACHE, PYTORCH_TRANSFORMERS_CACHE, PYTORCH
|
||||
from .data import (is_sklearn_available,
|
||||
InputExample, InputFeatures, DataProcessor,
|
||||
glue_output_modes, glue_convert_examples_to_features,
|
||||
glue_processors, glue_tasks_num_labels,
|
||||
xnli_output_modes, xnli_processors, xnli_tasks_num_labels,
|
||||
squad_convert_examples_to_features, SquadFeatures,
|
||||
SquadExample, SquadV1Processor, SquadV2Processor)
|
||||
glue_processors, glue_tasks_num_labels)
|
||||
|
||||
if is_sklearn_available():
|
||||
from .data import glue_compute_metrics, xnli_compute_metrics
|
||||
from .data import glue_compute_metrics
|
||||
|
||||
# Tokenizers
|
||||
from .tokenization_utils import (PreTrainedTokenizer)
|
||||
from .tokenization_auto import AutoTokenizer
|
||||
from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
|
||||
from .tokenization_bert_japanese import BertJapaneseTokenizer, MecabTokenizer, CharacterTokenizer
|
||||
from .tokenization_openai import OpenAIGPTTokenizer
|
||||
from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
|
||||
from .tokenization_gpt2 import GPT2Tokenizer
|
||||
from .tokenization_ctrl import CTRLTokenizer
|
||||
from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
|
||||
from .tokenization_xlm import XLMTokenizer
|
||||
from .tokenization_roberta import RobertaTokenizer
|
||||
from .tokenization_distilbert import DistilBertTokenizer
|
||||
from .tokenization_albert import AlbertTokenizer
|
||||
from .tokenization_camembert import CamembertTokenizer
|
||||
from .tokenization_t5 import T5Tokenizer
|
||||
|
||||
# Configurations
|
||||
from .configuration_utils import PretrainedConfig
|
||||
@@ -57,14 +49,10 @@ from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_xlnet import XLNetConfig, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_xlm import XLMConfig, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_roberta import RobertaConfig, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_distilbert import DistilBertConfig, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_albert import AlbertConfig, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_camembert import CamembertConfig, CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
from .configuration_t5 import T5Config, T5_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
|
||||
# Modeling
|
||||
if is_torch_available():
|
||||
@@ -78,55 +66,36 @@ if is_torch_available():
|
||||
BertForTokenClassification, BertForQuestionAnswering,
|
||||
load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_openai import (OpenAIGPTPreTrainedModel, OpenAIGPTModel,
|
||||
OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
|
||||
load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
|
||||
load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_transfo_xl import (TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
|
||||
AdaptiveEmbedding,
|
||||
load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_gpt2 import (GPT2PreTrainedModel, GPT2Model,
|
||||
GPT2LMHeadModel, GPT2DoubleHeadsModel,
|
||||
load_tf_weights_in_gpt2, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_ctrl import (CTRLPreTrainedModel, CTRLModel,
|
||||
CTRLLMHeadModel,
|
||||
CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_xlnet import (XLNetPreTrainedModel, XLNetModel, XLNetLMHeadModel,
|
||||
XLNetForSequenceClassification, XLNetForTokenClassification,
|
||||
XLNetForMultipleChoice, XLNetForQuestionAnsweringSimple,
|
||||
XLNetForQuestionAnswering, load_tf_weights_in_xlnet,
|
||||
XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
XLNetForSequenceClassification, XLNetForMultipleChoice,
|
||||
XLNetForQuestionAnsweringSimple, XLNetForQuestionAnswering,
|
||||
load_tf_weights_in_xlnet, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_xlm import (XLMPreTrainedModel , XLMModel,
|
||||
XLMWithLMHeadModel, XLMForSequenceClassification,
|
||||
XLMForQuestionAnswering, XLMForQuestionAnsweringSimple,
|
||||
XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_roberta import (RobertaForMaskedLM, RobertaModel,
|
||||
RobertaForSequenceClassification, RobertaForMultipleChoice,
|
||||
RobertaForTokenClassification,
|
||||
ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_distilbert import (DistilBertPreTrainedModel, DistilBertForMaskedLM, DistilBertModel,
|
||||
from .modeling_distilbert import (DistilBertForMaskedLM, DistilBertModel,
|
||||
DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
|
||||
DistilBertForTokenClassification,
|
||||
DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_camembert import (CamembertForMaskedLM, CamembertModel,
|
||||
CamembertForSequenceClassification, CamembertForMultipleChoice,
|
||||
CamembertForTokenClassification,
|
||||
CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
|
||||
from .modeling_t5 import (T5PreTrainedModel, T5Model, T5WithLMHeadModel,
|
||||
load_tf_weights_in_t5,
|
||||
T5_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_albert import (AlbertPreTrainedModel, AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
|
||||
AlbertForQuestionAnswering,
|
||||
load_tf_weights_in_albert, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
# Optimization
|
||||
from .optimization import (AdamW, get_constant_schedule, get_constant_schedule_with_warmup, get_cosine_schedule_with_warmup,
|
||||
get_cosine_with_hard_restarts_schedule_with_warmup, get_linear_schedule_with_warmup)
|
||||
from .optimization import (AdamW, ConstantLRSchedule, WarmupConstantSchedule, WarmupCosineSchedule,
|
||||
WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
|
||||
|
||||
|
||||
# TensorFlow
|
||||
if is_tf_available():
|
||||
from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
|
||||
from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary
|
||||
from .modeling_tf_auto import (TFAutoModel, TFAutoModelForSequenceClassification, TFAutoModelForQuestionAnswering,
|
||||
TFAutoModelWithLMHead)
|
||||
|
||||
@@ -135,68 +104,60 @@ if is_tf_available():
|
||||
TFBertForMaskedLM, TFBertForNextSentencePrediction,
|
||||
TFBertForSequenceClassification, TFBertForMultipleChoice,
|
||||
TFBertForTokenClassification, TFBertForQuestionAnswering,
|
||||
load_bert_pt_weights_in_tf2,
|
||||
TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_gpt2 import (TFGPT2PreTrainedModel, TFGPT2MainLayer,
|
||||
TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel,
|
||||
load_gpt2_pt_weights_in_tf2,
|
||||
TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_openai import (TFOpenAIGPTPreTrainedModel, TFOpenAIGPTMainLayer,
|
||||
TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TFOpenAIGPTDoubleHeadsModel,
|
||||
load_openai_gpt_pt_weights_in_tf2,
|
||||
TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_transfo_xl import (TFTransfoXLPreTrainedModel, TFTransfoXLMainLayer,
|
||||
TFTransfoXLModel, TFTransfoXLLMHeadModel,
|
||||
load_transfo_xl_pt_weights_in_tf2,
|
||||
TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_xlnet import (TFXLNetPreTrainedModel, TFXLNetMainLayer,
|
||||
TFXLNetModel, TFXLNetLMHeadModel,
|
||||
TFXLNetForSequenceClassification,
|
||||
TFXLNetForTokenClassification,
|
||||
TFXLNetForQuestionAnsweringSimple,
|
||||
load_xlnet_pt_weights_in_tf2,
|
||||
TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_xlm import (TFXLMPreTrainedModel, TFXLMMainLayer,
|
||||
TFXLMModel, TFXLMWithLMHeadModel,
|
||||
TFXLMForSequenceClassification,
|
||||
TFXLMForQuestionAnsweringSimple,
|
||||
load_xlm_pt_weights_in_tf2,
|
||||
TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_roberta import (TFRobertaPreTrainedModel, TFRobertaMainLayer,
|
||||
TFRobertaModel, TFRobertaForMaskedLM,
|
||||
TFRobertaForSequenceClassification,
|
||||
TFRobertaForTokenClassification,
|
||||
load_roberta_pt_weights_in_tf2,
|
||||
TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_distilbert import (TFDistilBertPreTrainedModel, TFDistilBertMainLayer,
|
||||
TFDistilBertModel, TFDistilBertForMaskedLM,
|
||||
TFDistilBertForSequenceClassification,
|
||||
TFDistilBertForTokenClassification,
|
||||
TFDistilBertForQuestionAnswering,
|
||||
load_distilbert_pt_weights_in_tf2,
|
||||
TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_ctrl import (TFCTRLPreTrainedModel, TFCTRLModel,
|
||||
TFCTRLLMHeadModel,
|
||||
TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
|
||||
TFAlbertForSequenceClassification,
|
||||
TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
from .modeling_tf_t5 import (TFT5PreTrainedModel, TFT5Model, TFT5WithLMHeadModel,
|
||||
TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP)
|
||||
|
||||
# Optimization
|
||||
from .optimization_tf import (WarmUp, create_optimizer, AdamWeightDecay, GradientAccumulator)
|
||||
|
||||
# TF 2.0 <=> PyTorch conversion utilities
|
||||
from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
|
||||
load_pytorch_checkpoint_in_tf2_model,
|
||||
load_pytorch_weights_in_tf2_model,
|
||||
load_pytorch_model_in_tf2_model,
|
||||
load_tf2_checkpoint_in_pytorch_model,
|
||||
load_tf2_weights_in_pytorch_model,
|
||||
load_tf2_model_in_pytorch_model)
|
||||
if is_tf_available() and is_torch_available():
|
||||
from .modeling_tf_pytorch_utils import (convert_tf_weight_name_to_pt_weight_name,
|
||||
load_pytorch_checkpoint_in_tf2_model,
|
||||
load_pytorch_weights_in_tf2_model,
|
||||
load_pytorch_model_in_tf2_model,
|
||||
load_tf2_checkpoint_in_pytorch_model,
|
||||
load_tf2_weights_in_pytorch_model,
|
||||
load_tf2_model_in_pytorch_model)
|
||||
|
||||
if not is_tf_available() and not is_torch_available():
|
||||
logger.warning("Neither PyTorch nor TensorFlow >= 2.0 have been found."
|
||||
|
||||
@@ -6,7 +6,6 @@ def main():
|
||||
"This command line utility let you convert original (author released) model checkpoint to pytorch.\n"
|
||||
"It should be used as one of: \n"
|
||||
">> transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
|
||||
">> transformers t5 TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
|
||||
">> transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
|
||||
">> transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
|
||||
">> transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
|
||||
@@ -22,23 +21,6 @@ def main():
|
||||
"https://www.tensorflow.org/install/ for installation instructions.")
|
||||
raise
|
||||
|
||||
if len(sys.argv) != 5:
|
||||
# pylint: disable=line-too-long
|
||||
print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
|
||||
else:
|
||||
PYTORCH_DUMP_OUTPUT = sys.argv.pop()
|
||||
TF_CONFIG = sys.argv.pop()
|
||||
TF_CHECKPOINT = sys.argv.pop()
|
||||
convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
|
||||
elif sys.argv[1] == "t5":
|
||||
try:
|
||||
from .convert_t5_original_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
|
||||
except ImportError:
|
||||
print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
|
||||
"In that case, it requires TensorFlow to be installed. Please see "
|
||||
"https://www.tensorflow.org/install/ for installation instructions.")
|
||||
raise
|
||||
|
||||
if len(sys.argv) != 5:
|
||||
# pylint: disable=line-too-long
|
||||
print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
|
||||
|
||||
@@ -1,12 +0,0 @@
|
||||
from abc import ABC, abstractmethod
|
||||
from argparse import ArgumentParser
|
||||
|
||||
class BaseTransformersCLICommand(ABC):
|
||||
@staticmethod
|
||||
@abstractmethod
|
||||
def register_subcommand(parser: ArgumentParser):
|
||||
raise NotImplementedError()
|
||||
|
||||
@abstractmethod
|
||||
def run(self):
|
||||
raise NotImplementedError()
|
||||
@@ -1,194 +0,0 @@
|
||||
from argparse import ArgumentParser
|
||||
from getpass import getpass
|
||||
import os
|
||||
|
||||
from transformers.commands import BaseTransformersCLICommand
|
||||
from transformers.hf_api import HfApi, HfFolder, HTTPError
|
||||
|
||||
|
||||
class UserCommands(BaseTransformersCLICommand):
|
||||
@staticmethod
|
||||
def register_subcommand(parser: ArgumentParser):
|
||||
login_parser = parser.add_parser('login')
|
||||
login_parser.set_defaults(func=lambda args: LoginCommand(args))
|
||||
whoami_parser = parser.add_parser('whoami')
|
||||
whoami_parser.set_defaults(func=lambda args: WhoamiCommand(args))
|
||||
logout_parser = parser.add_parser('logout')
|
||||
logout_parser.set_defaults(func=lambda args: LogoutCommand(args))
|
||||
list_parser = parser.add_parser('ls')
|
||||
list_parser.set_defaults(func=lambda args: ListObjsCommand(args))
|
||||
# upload
|
||||
upload_parser = parser.add_parser('upload')
|
||||
upload_parser.add_argument('path', type=str, help='Local path of the folder or individual file to upload.')
|
||||
upload_parser.add_argument('--filename', type=str, default=None, help='Optional: override individual object filename on S3.')
|
||||
upload_parser.set_defaults(func=lambda args: UploadCommand(args))
|
||||
|
||||
|
||||
|
||||
class ANSI:
|
||||
"""
|
||||
Helper for en.wikipedia.org/wiki/ANSI_escape_code
|
||||
"""
|
||||
_bold = u"\u001b[1m"
|
||||
_reset = u"\u001b[0m"
|
||||
@classmethod
|
||||
def bold(cls, s):
|
||||
return "{}{}{}".format(cls._bold, s, cls._reset)
|
||||
|
||||
|
||||
class BaseUserCommand:
|
||||
def __init__(self, args):
|
||||
self.args = args
|
||||
self._api = HfApi()
|
||||
|
||||
|
||||
class LoginCommand(BaseUserCommand):
|
||||
def run(self):
|
||||
print("""
|
||||
_| _| _| _| _|_|_| _|_|_| _|_|_| _| _| _|_|_| _|_|_|_| _|_| _|_|_| _|_|_|_|
|
||||
_| _| _| _| _| _| _| _|_| _| _| _| _| _| _| _|
|
||||
_|_|_|_| _| _| _| _|_| _| _|_| _| _| _| _| _| _|_| _|_|_| _|_|_|_| _| _|_|_|
|
||||
_| _| _| _| _| _| _| _| _| _| _|_| _| _| _| _| _| _| _|
|
||||
_| _| _|_| _|_|_| _|_|_| _|_|_| _| _| _|_|_| _| _| _| _|_|_| _|_|_|_|
|
||||
|
||||
""")
|
||||
username = input("Username: ")
|
||||
password = getpass()
|
||||
try:
|
||||
token = self._api.login(username, password)
|
||||
except HTTPError as e:
|
||||
# probably invalid credentials, display error message.
|
||||
print(e)
|
||||
exit(1)
|
||||
HfFolder.save_token(token)
|
||||
print("Login successful")
|
||||
print("Your token:", token, "\n")
|
||||
print("Your token has been saved to", HfFolder.path_token)
|
||||
|
||||
|
||||
class WhoamiCommand(BaseUserCommand):
|
||||
def run(self):
|
||||
token = HfFolder.get_token()
|
||||
if token is None:
|
||||
print("Not logged in")
|
||||
exit()
|
||||
try:
|
||||
user = self._api.whoami(token)
|
||||
print(user)
|
||||
except HTTPError as e:
|
||||
print(e)
|
||||
|
||||
|
||||
class LogoutCommand(BaseUserCommand):
|
||||
def run(self):
|
||||
token = HfFolder.get_token()
|
||||
if token is None:
|
||||
print("Not logged in")
|
||||
exit()
|
||||
HfFolder.delete_token()
|
||||
self._api.logout(token)
|
||||
print("Successfully logged out.")
|
||||
|
||||
|
||||
class ListObjsCommand(BaseUserCommand):
|
||||
def tabulate(self, rows, headers):
|
||||
# type: (List[List[Union[str, int]]], List[str]) -> str
|
||||
"""
|
||||
Inspired by:
|
||||
stackoverflow.com/a/8356620/593036
|
||||
stackoverflow.com/questions/9535954/printing-lists-as-tabular-data
|
||||
"""
|
||||
col_widths = [max(len(str(x)) for x in col) for col in zip(*rows, headers)]
|
||||
row_format = ("{{:{}}} " * len(headers)).format(*col_widths)
|
||||
lines = []
|
||||
lines.append(
|
||||
row_format.format(*headers)
|
||||
)
|
||||
lines.append(
|
||||
row_format.format(*["-" * w for w in col_widths])
|
||||
)
|
||||
for row in rows:
|
||||
lines.append(
|
||||
row_format.format(*row)
|
||||
)
|
||||
return "\n".join(lines)
|
||||
|
||||
def run(self):
|
||||
token = HfFolder.get_token()
|
||||
if token is None:
|
||||
print("Not logged in")
|
||||
exit(1)
|
||||
try:
|
||||
objs = self._api.list_objs(token)
|
||||
except HTTPError as e:
|
||||
print(e)
|
||||
exit(1)
|
||||
if len(objs) == 0:
|
||||
print("No shared file yet")
|
||||
exit()
|
||||
rows = [ [
|
||||
obj.filename,
|
||||
obj.LastModified,
|
||||
obj.ETag,
|
||||
obj.Size
|
||||
] for obj in objs ]
|
||||
print(
|
||||
self.tabulate(rows, headers=["Filename", "LastModified", "ETag", "Size"])
|
||||
)
|
||||
|
||||
|
||||
class UploadCommand(BaseUserCommand):
|
||||
def walk_dir(self, rel_path):
|
||||
"""
|
||||
Recursively list all files in a folder.
|
||||
"""
|
||||
entries: List[os.DirEntry] = list(os.scandir(rel_path))
|
||||
files = [
|
||||
(
|
||||
os.path.join(os.getcwd(), f.path), # filepath
|
||||
f.path # filename
|
||||
)
|
||||
for f in entries if f.is_file()
|
||||
]
|
||||
for f in entries:
|
||||
if f.is_dir():
|
||||
files += self.walk_dir(f.path)
|
||||
return files
|
||||
|
||||
def run(self):
|
||||
token = HfFolder.get_token()
|
||||
if token is None:
|
||||
print("Not logged in")
|
||||
exit(1)
|
||||
local_path = os.path.abspath(self.args.path)
|
||||
if os.path.isdir(local_path):
|
||||
if self.args.filename is not None:
|
||||
raise ValueError("Cannot specify a filename override when uploading a folder.")
|
||||
rel_path = os.path.basename(local_path)
|
||||
files = self.walk_dir(rel_path)
|
||||
elif os.path.isfile(local_path):
|
||||
filename = self.args.filename if self.args.filename is not None else os.path.basename(local_path)
|
||||
files = [(local_path, filename)]
|
||||
else:
|
||||
raise ValueError("Not a valid file or directory: {}".format(local_path))
|
||||
|
||||
for filepath, filename in files:
|
||||
print(
|
||||
"About to upload file {} to S3 under filename {}".format(
|
||||
ANSI.bold(filepath), ANSI.bold(filename)
|
||||
)
|
||||
)
|
||||
|
||||
choice = input("Proceed? [Y/n] ").lower()
|
||||
if not(choice == "" or choice == "y" or choice == "yes"):
|
||||
print("Abort")
|
||||
exit()
|
||||
print(
|
||||
ANSI.bold("Uploading... This might take a while if files are large")
|
||||
)
|
||||
for filepath, filename in files:
|
||||
access_url = self._api.presign_and_upload(
|
||||
token=token, filename=filename, filepath=filepath
|
||||
)
|
||||
print("Your file now lives at:")
|
||||
print(access_url)
|
||||
@@ -1,100 +0,0 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
|
||||
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" ALBERT model configuration """
|
||||
|
||||
from .configuration_utils import PretrainedConfig
|
||||
|
||||
ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-config.json",
|
||||
'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-config.json",
|
||||
'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-config.json",
|
||||
'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-config.json",
|
||||
'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-config.json",
|
||||
'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-config.json",
|
||||
'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-config.json",
|
||||
'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-config.json",
|
||||
}
|
||||
|
||||
class AlbertConfig(PretrainedConfig):
|
||||
"""Configuration for `AlbertModel`.
|
||||
|
||||
The default settings match the configuration of model `albert_xxlarge`.
|
||||
"""
|
||||
|
||||
pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
|
||||
|
||||
def __init__(self,
|
||||
vocab_size_or_config_json_file=30000,
|
||||
embedding_size=128,
|
||||
hidden_size=4096,
|
||||
num_hidden_layers=12,
|
||||
num_hidden_groups=1,
|
||||
num_attention_heads=64,
|
||||
intermediate_size=16384,
|
||||
inner_group_num=1,
|
||||
hidden_act="gelu_new",
|
||||
hidden_dropout_prob=0,
|
||||
attention_probs_dropout_prob=0,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=2,
|
||||
initializer_range=0.02,
|
||||
layer_norm_eps=1e-12, **kwargs):
|
||||
"""Constructs AlbertConfig.
|
||||
|
||||
Args:
|
||||
vocab_size: Vocabulary size of `inputs_ids` in `AlbertModel`.
|
||||
embedding_size: size of voc embeddings.
|
||||
hidden_size: Size of the encoder layers and the pooler layer.
|
||||
num_hidden_layers: Number of hidden layers in the Transformer encoder.
|
||||
num_hidden_groups: Number of group for the hidden layers, parameters in
|
||||
the same group are shared.
|
||||
num_attention_heads: Number of attention heads for each attention layer in
|
||||
the Transformer encoder.
|
||||
intermediate_size: The size of the "intermediate" (i.e., feed-forward)
|
||||
layer in the Transformer encoder.
|
||||
inner_group_num: int, number of inner repetition of attention and ffn.
|
||||
down_scale_factor: float, the scale to apply
|
||||
hidden_act: The non-linear activation function (function or string) in the
|
||||
encoder and pooler.
|
||||
hidden_dropout_prob: The dropout probability for all fully connected
|
||||
layers in the embeddings, encoder, and pooler.
|
||||
attention_probs_dropout_prob: The dropout ratio for the attention
|
||||
probabilities.
|
||||
max_position_embeddings: The maximum sequence length that this model might
|
||||
ever be used with. Typically set this to something large just in case
|
||||
(e.g., 512 or 1024 or 2048).
|
||||
type_vocab_size: The vocabulary size of the `token_type_ids` passed into
|
||||
`AlbertModel`.
|
||||
initializer_range: The stdev of the truncated_normal_initializer for
|
||||
initializing all weight matrices.
|
||||
"""
|
||||
super(AlbertConfig, self).__init__(**kwargs)
|
||||
|
||||
self.vocab_size = vocab_size_or_config_json_file
|
||||
self.embedding_size = embedding_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_hidden_groups = num_hidden_groups
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.inner_group_num = inner_group_num
|
||||
self.hidden_act = hidden_act
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.initializer_range = initializer_range
|
||||
self.layer_norm_eps = layer_norm_eps
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user