save_copy

prophetnet modified

modify codes as suggested v1

add prophetnet test files

.circleci/deploy.sh

@@ -49,5 +49,4 @@ deploy_doc "10d7239" v2.10.0
 deploy_doc "b42586e" v2.11.0
 deploy_doc "7fb8bdf" v3.0.2
 deploy_doc "4b3ee9c" v3.1.0
-deploy_doc "3ebb1b3" v3.2.0
-deploy_doc "0613f05" # v3.3.0 Latest stable release
+deploy_doc "3ebb1b3" # v3.2.0 Latest stable release

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,61 +1,2 @@
-# What does this PR do?
-
-<!--
-Congratulations! You've made it this far! You're not quite done yet though.
-
-Once merged, your PR is going to appear in the release notes with the title you set, so make sure it's a great title that fully reflects the extent of your awesome contribution.
-
-Then, please replace this with a description of the change and which issue is fixed (if applicable). Please also include relevant motivation and context. List any dependencies (if any) that are required for this change.
-
-Once you're done, someone will review your PR shortly (see the section "Who can review?" below to tag some potential reviewers). They may suggest changes to make the code even better. If no one reviewed your PR after a week has passed, don't hesitate to post a new comment @-mentioning the same persons---sometimes notifications get lost.
--->
-
-<!-- Remove if not applicable -->
-
-Fixes # (issue)
-
-
-## Before submitting
-- [ ] This PR fixes a typo or improves the docs (you can dimiss the other checks if that's the case).
-- [ ] Did you read the [contributor guideline](https://github.com/huggingface/transformers/blob/master/CONTRIBUTING.md#start-contributing-pull-requests), 
-      Pull Request section?
-- [ ] Was this discussed/approved via a Github issue or the [forum](https://discuss.huggingface.co/)? Please add a link
-      to the it if that's the case.
-- [ ] Did you make sure to update the documentation with your changes? Here are the
-      [documentation guidelines](https://github.com/huggingface/transformers/tree/master/docs), and
-      [here are tips on formatting docstrings](https://github.com/huggingface/transformers/tree/master/docs#writing-source-documentation).
-- [ ] Did you write any new necessary tests? 
-
-
-## Who can review?
-
-Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
-members/contributors which may be interested in your PR.
-
-<!-- Your PR will be replied to more quickly if you can figure out the right person to tag with @
-
- If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
- Please tag fewer than 3 people.
-
- albert, bert, GPT2, XLM: @LysandreJik 
- tokenizers: @mfuntowicz
- Trainer: @sgugger
- Speed and Memory Benchmarks: @patrickvonplaten
- Model Cards: @julien-c
- Translation: @sshleifer
- Summarization: @sshleifer
- TextGeneration: @TevenLeScao 
- examples/distillation: @VictorSanh
- nlp datasets: [different repo](https://github.com/huggingface/nlp)
- rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
- Text Generation: @TevenLeScao
- Blenderbot, Bart, Marian, Pegasus: @sshleifer
- T5: @patrickvonplaten
- Longformer/Reformer: @patrickvonplaten
- TransfoXL/XLNet: @TevenLeScao 
- examples/seq2seq: @sshleifer
- examples/bert-loses-patience: @JetRunner
- tensorflow: @jplu
- examples/token-classification: @stefan-it
- documentation: @sgugger
- -->
+<!-- This line specifies which issue to close after the pull request is merged. -->
+Fixes #{issue number}

CODE_OF_CONDUCT.md

@@ -1,129 +0,0 @@
-
-# Contributor Covenant Code of Conduct
-
-## Our Pledge
-
-We as members, contributors, and leaders pledge to make participation in our
-community a harassment-free experience for everyone, regardless of age, body
-size, visible or invisible disability, ethnicity, sex characteristics, gender
-identity and expression, level of experience, education, socio-economic status,
-nationality, personal appearance, race, religion, or sexual identity
-and orientation.
-
-We pledge to act and interact in ways that contribute to an open, welcoming,
-diverse, inclusive, and healthy community.
-
-## Our Standards
-
-Examples of behavior that contributes to a positive environment for our
-community include:
-
-* Demonstrating empathy and kindness toward other people
-* Being respectful of differing opinions, viewpoints, and experiences
-* Giving and gracefully accepting constructive feedback
-* Accepting responsibility and apologizing to those affected by our mistakes,
-  and learning from the experience
-* Focusing on what is best not just for us as individuals, but for the
-  overall community
-
-Examples of unacceptable behavior include:
-
-* The use of sexualized language or imagery, and sexual attention or
-  advances of any kind
-* Trolling, insulting or derogatory comments, and personal or political attacks
-* Public or private harassment
-* Publishing others' private information, such as a physical or email
-  address, without their explicit permission
-* Other conduct which could reasonably be considered inappropriate in a
-  professional setting
-
-## Enforcement Responsibilities
-
-Community leaders are responsible for clarifying and enforcing our standards of
-acceptable behavior and will take appropriate and fair corrective action in
-response to any behavior that they deem inappropriate, threatening, offensive,
-or harmful.
-
-Community leaders have the right and responsibility to remove, edit, or reject
-comments, commits, code, wiki edits, issues, and other contributions that are
-not aligned to this Code of Conduct, and will communicate reasons for moderation
-decisions when appropriate.
-
-## Scope
-
-This Code of Conduct applies within all community spaces, and also applies when
-an individual is officially representing the community in public spaces.
-Examples of representing our community include using an official e-mail address,
-posting via an official social media account, or acting as an appointed
-representative at an online or offline event.
-
-## Enforcement
-
-Instances of abusive, harassing, or otherwise unacceptable behavior may be
-reported to the community leaders responsible for enforcement at
-feedback@huggingface.co.
-All complaints will be reviewed and investigated promptly and fairly.
-
-All community leaders are obligated to respect the privacy and security of the
-reporter of any incident.
-
-## Enforcement Guidelines
-
-Community leaders will follow these Community Impact Guidelines in determining
-the consequences for any action they deem in violation of this Code of Conduct:
-
-### 1. Correction
-
-**Community Impact**: Use of inappropriate language or other behavior deemed
-unprofessional or unwelcome in the community.
-
-**Consequence**: A private, written warning from community leaders, providing
-clarity around the nature of the violation and an explanation of why the
-behavior was inappropriate. A public apology may be requested.
-
-### 2. Warning
-
-**Community Impact**: A violation through a single incident or series
-of actions.
-
-**Consequence**: A warning with consequences for continued behavior. No
-interaction with the people involved, including unsolicited interaction with
-those enforcing the Code of Conduct, for a specified period of time. This
-includes avoiding interactions in community spaces as well as external channels
-like social media. Violating these terms may lead to a temporary or
-permanent ban.
-
-### 3. Temporary Ban
-
-**Community Impact**: A serious violation of community standards, including
-sustained inappropriate behavior.
-
-**Consequence**: A temporary ban from any sort of interaction or public
-communication with the community for a specified period of time. No public or
-private interaction with the people involved, including unsolicited interaction
-with those enforcing the Code of Conduct, is allowed during this period.
-Violating these terms may lead to a permanent ban.
-
-### 4. Permanent Ban
-
-**Community Impact**: Demonstrating a pattern of violation of community
-standards, including sustained inappropriate behavior,  harassment of an
-individual, or aggression toward or disparagement of classes of individuals.
-
-**Consequence**: A permanent ban from any sort of public interaction within
-the community.
-
-## Attribution
-
-This Code of Conduct is adapted from the [Contributor Covenant][homepage],
-version 2.0, available at
-https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
-
-Community Impact Guidelines were inspired by [Mozilla's code of conduct
-enforcement ladder](https://github.com/mozilla/diversity).
-
-[homepage]: https://www.contributor-covenant.org
-
-For answers to common questions about this code of conduct, see the FAQ at
-https://www.contributor-covenant.org/faq. Translations are available at
-https://www.contributor-covenant.org/translations.

CONTRIBUTING.md

@@ -9,9 +9,6 @@ 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".
 
-Whichever way you choose to contribute, please be mindful to respect our
-[code of conduct](https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md).
-
 ## You can contribute in so many ways!
 
 There are 4 ways you can contribute to transformers:
@@ -179,14 +176,13 @@ Follow these steps to start contributing:
    ```bash
    $ make quality
    ```
+
    You can do the automatic style corrections and code verifications that can't be automated in one go:
 
    ```bash
    $ make fixup
    ```
 
-   This target is also optimized to only work with files modified by the PR you're working on.
-
    If you're modifying documents under `docs/source`, make sure to validate that
    they can still be built. This check also runs in CI. To run a local check
    make sure you have installed the documentation builder requirements, by

Makefile

@@ -1,46 +1,24 @@
-.PHONY: modified_only_fixup extra_quality_checks quality style fixup fix-copies test test-examples docs
-
-
-check_dirs := examples templates tests src utils
-
-# get modified files since the branch was made
-fork_point_sha := $(shell git merge-base --fork-point master)
-joined_dirs    := $(shell echo $(check_dirs) | tr " " "|")
-modified_files := $(shell git diff --name-only $(fork_point_sha) | egrep '^($(joined_dirs))')
-#$(info modified files are: $(modified_files))
-
-modified_only_fixup:
-	@if [ -n "$(modified_files)" ]; then \
-		echo "Checking/fixing $(modified_files)"; \
-		black $(modified_files); \
-		isort $(modified_files); \
-		flake8 $(modified_files); \
-	else \
-		echo "No relevant files were modified"; \
-	fi
+.PHONY: quality_checks quality style fixup test test-examples docs
 
 # Check that source code meets quality standards
 
-extra_quality_checks:
+quality_checks:
+	flake8 examples templates tests src utils
 	python utils/check_copies.py
 	python utils/check_repo.py
 
-# this target runs checks on all files
 quality:
-	black --check $(check_dirs)
-	isort --check-only $(check_dirs)
-	flake8 $(check_dirs)
-	${MAKE} extra_quality_checks
+	black --check examples templates tests src utils
+	isort --check-only examples templates tests src utils
+	${MAKE} quality_checks
 
 # Format source code automatically and check is there are any problems left that need manual fixing
 
 style:
-	black $(check_dirs)
-	isort $(check_dirs)
+	black examples templates tests src utils
+	isort examples templates tests src utils
 
-# Super fast fix and check target that only works on relevant modified files since the branch was made
-
-fixup: modified_only_fixup extra_quality_checks
+fixup: style quality_checks
 
 # Make marked copies of snippets of codes conform to the original
 

README.md

@@ -16,9 +16,6 @@
     <a href="https://github.com/huggingface/transformers/releases">
         <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/transformers.svg">
     </a>
-    <a href="https://github.com/huggingface/transformers/blob/master/CODE_OF_CONDUCT.md">
-        <img alt="Contributor Covenant" src="https://img.shields.io/badge/Contributor%20Covenant-v2.0%20adopted-ff69b4.svg">
-    </a>
 </p>
 
 <h3 align="center">
@@ -186,9 +183,19 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 24. **[MBart](https://github.com/pytorch/fairseq/tree/master/examples/mbart)** (from Facebook) released with the paper  [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.  
 25. **[LXMERT](https://github.com/airsplay/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
 26. **[Funnel Transformer](https://github.com/laiguokun/Funnel-Transformer)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
+<<<<<<< HEAD
+27. **[ProphetNet](https://github.com/microsoft/ProphetNet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
+28. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
+29. 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.
+
+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 Pearson 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).
+
+## Online demo
+=======
 27. **[LayoutLM](https://github.com/microsoft/unilm/tree/master/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 28. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
 29. 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.
+>>>>>>> main/master
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations. You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 

docs/source/_static/css/huggingface.css

@@ -125,12 +125,6 @@ a.copybtn {
     background-color: #6670FF;
 }
 
-/* The section headers in the toc tree */
-.wy-menu-vertical p.caption{
-    background-color: #4d59ff;
-    line-height: 40px;
-}
-
 /* The selected items in the toc tree */
 .wy-menu-vertical li.current{
     background-color: #A6B0FF;

docs/source/_static/js/custom.js

@@ -1,11 +1,10 @@
 // These two things need to be updated at each release for the version selector.
 // Last stable version
-const stableVersion = "v3.3.0"
+const stableVersion = "v3.2.0"
 // Dictionary doc folder to label
 const versionMapping = {
     "master": "master",
-    "": "v3.3.0",
-    "v3.2.0": "v3.2.0",
+    "": "v3.2.0",
     "v3.1.0": "v3.1.0 (stable)",
     "v3.0.2": "v3.0.0/v3.0.1/v3.0.2",
     "v2.11.0": "v2.11.0",

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'3.3.1'
+release = u'3.2.0'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/index.rst

@@ -46,10 +46,7 @@ The documentation is organized in five parts:
 - **ADVANCED GUIDES** contains more advanced guides that are more specific to a given script or part of the library.
 - **RESEARCH** focuses on tutorials that have less to do with how to use the library but more about general resarch in
   transformers model
-- The three last section contain the documentation of each public class and function, grouped in:
-    - **MAIN CLASSES** for the main classes exposing the important APIs of the library.
-    - **MODELS** for the classes and functions related to each model implemented in the library.
-    - **INTERNAL HELPERS** for the classes and functions we use internally.
+- **PACKAGE REFERENCE** contains the documentation of each public class and function.
 
 The library currently contains PyTorch and Tensorflow implementations, pre-trained model weights, usage scripts and
 conversion utilities for the following models:
@@ -191,60 +188,49 @@ conversion utilities for the following models:
 
 .. toctree::
     :maxdepth: 2
-    :caption: Main Classes
+    :caption: Package Reference
 
     main_classes/configuration
-    main_classes/logging
-    main_classes/model
-    main_classes/optimizer_schedules
     main_classes/output
-    main_classes/pipelines
-    main_classes/processors
+    main_classes/model
     main_classes/tokenizer
+    main_classes/pipelines
     main_classes/trainer
-
-.. toctree::
-    :maxdepth: 2
-    :caption: Models
-
-    model_doc/albert
+    main_classes/optimizer_schedules
+    main_classes/processors
+    main_classes/logging
     model_doc/auto
-    model_doc/bart
-    model_doc/bert
-    model_doc/bertgeneration
-    model_doc/camembert
-    model_doc/ctrl
-    model_doc/dialogpt
-    model_doc/distilbert
-    model_doc/dpr
-    model_doc/electra
     model_doc/encoderdecoder
+    model_doc/bert
+    model_doc/gpt
+    model_doc/transformerxl
+    model_doc/gpt2
+    model_doc/xlm
+    model_doc/xlnet
+    model_doc/roberta
+    model_doc/distilbert
+    model_doc/ctrl
+    model_doc/camembert
+    model_doc/albert
+    model_doc/xlmroberta
     model_doc/flaubert
+    model_doc/bart
+    model_doc/t5
+    model_doc/electra
+    model_doc/dialogpt
+    model_doc/reformer
+    model_doc/marian
+    model_doc/longformer
+    model_doc/retribert
+    model_doc/mobilebert
+    model_doc/dpr
+    model_doc/pegasus
+    model_doc/mbart
     model_doc/fsmt
     model_doc/funnel
-    model_doc/layoutlm
-    model_doc/longformer
     model_doc/lxmert
-    model_doc/marian
-    model_doc/mbart
-    model_doc/mobilebert
-    model_doc/gpt
-    model_doc/gpt2
-    model_doc/pegasus
-    model_doc/rag
-    model_doc/reformer
-    model_doc/retribert
-    model_doc/roberta
-    model_doc/t5
-    model_doc/transformerxl
-    model_doc/xlm
-    model_doc/xlmroberta
-    model_doc/xlnet
-
-.. toctree::
-    :maxdepth: 2
-    :caption: Internal Helpers
-
+    model_doc/bertgeneration
+    model_doc/layoutlm
     internal/modeling_utils
-    internal/pipelines_utils
     internal/tokenization_utils
+    internal/pipelines_utils

docs/source/model_doc/rag.rst

@@ -1,91 +0,0 @@
-RAG
-----------------------------------------------------
-
-Overview
-~~~~~~~~~~~~~~~~~~~~~
-
-Retrieval-augmented generation ("RAG") models combine the powers of pretrained dense retrieval (DPR) and
-sequence-to-sequence models. RAG models retrieve documents, pass them to a seq2seq model, then marginalize to generate
-outputs. The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing
-both retrieval and generation to adapt to downstream tasks.
-
-It is based on the paper `Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
-<https://arxiv.org/abs/2005.11401>`__ by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir
-Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
-
-The abstract from the paper is the following:
-
-*Large pre-trained language models have been shown to store factual knowledge
-in their parameters, and achieve state-of-the-art results when fine-tuned on
-downstream NLP tasks. However, their ability to access and precisely manipulate
-knowledge is still limited, and hence on knowledge-intensive tasks, their
-performance lags behind task-specific architectures. Additionally, providing
-provenance for their decisions and updating their world knowledge remain open
-research problems. Pre-trained models with a differentiable access mechanism to
-explicit nonparametric memory can overcome this issue, but have so far been only
-investigated for extractive downstream tasks. We explore a general-purpose
-fine-tuning recipe for retrieval-augmented generation (RAG) — models which combine
-pre-trained parametric and non-parametric memory for language generation. We
-introduce RAG models where the parametric memory is a pre-trained seq2seq model and
-the non-parametric memory is a dense vector index of Wikipedia, accessed with
-a pre-trained neural retriever. We compare two RAG formulations, one which
-conditions on the same retrieved passages across the whole generated sequence, the
-other can use different passages per token. We fine-tune and evaluate our models
-on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art
-on three open domain QA tasks, outperforming parametric seq2seq models and
-task-specific retrieve-and-extract architectures. For language generation tasks, we
-find that RAG models generate more specific, diverse and factual language than a
-state-of-the-art parametric-only seq2seq baseline.*
-
-
-
-RagConfig
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagConfig
-    :members:
-
-
-RagTokenizer
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagTokenizer
-    :members: prepare_seq2seq_batch
-
-
-Rag specific outputs
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_rag.RetrievAugLMMarginOutput
-    :members:
-
-.. autoclass:: transformers.modeling_rag.RetrievAugLMOutput
-    :members:
-
-
-RAGRetriever
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagRetriever
-    :members:
-
-
-RagModel
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagModel
-    :members: forward
-
-
-RagSequenceForGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagSequenceForGeneration
-    :members: forward, generate
-
-
-RagTokenForGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RagTokenForGeneration
-    :members: forward, generate

docs/source/model_sharing.rst

@@ -112,7 +112,8 @@ Make sure there are no garbage files in the directory you'll upload. It should o
 - a `tf_model.h5` file, which is the TensorFlow checkpoint (unless you can't have it for some reason) ;
 - a `special_tokens_map.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
 - a `tokenizer_config.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
-- files named `vocab.json`, `vocab.txt`, `merges.txt`, or similar, which contain the vocabulary of your tokenizer, part of your :doc:`tokenizer <main_classes/tokenizer>` save;
+- a `vocab.txt`, which is the vocabulary of your tokenizer, part of your :doc:`tokenizer <main_classes/tokenizer>`
+  save;
 - maybe a `added_tokens.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save.
 
 Other files can safely be deleted.
@@ -220,3 +221,4 @@ You can also delete unneeded files with
 .. code-block::
 
     transformers-cli s3 rm awesome-name-you-picked/filename
+

docs/source/model_summary.rst

@@ -672,27 +672,6 @@ DPR consists in three models:
 
 DPR's pipeline (not implemented yet) uses a retrieval step to find the top k contexts given a certain question, and then it calls the reader with the question and the retrieved documents to get the answer.
 
-RAG
------------------------------------------------------------------------------------------------------------------------
-
-.. raw:: html
-
-   <a href="https://huggingface.co/models?filter=rag">
-       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-rag-blueviolet">
-   </a>
-   <a href="model_doc/rag.html">
-       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-rag-blueviolet">
-   </a>
-
-`Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks <https://arxiv.org/abs/2005.11401>`_,
-Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela
-
-Retrieval-augmented generation ("RAG") models combine the powers of pretrained dense retrieval (DPR) and Seq2Seq models.
-RAG models retrieve docs, pass them to a seq2seq model, then marginalize to generate outputs.
-The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing both retrieval and generation to adapt to downstream tasks.
-
-The two models RAG-Token and RAG-Sequence are available for generation.
-
 More technical aspects
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 

examples/seq2seq/rouge_cli.py

@@ -9,7 +9,7 @@ def calculate_rouge_path(pred_path, tgt_path, save_path=None, **kwargs):
     tgt_lns = [x.strip() for x in open(tgt_path).readlines()][: len(pred_lns)]
     metrics = calculate_rouge(pred_lns, tgt_lns, **kwargs)
     if save_path is not None:
-        save_json(metrics, save_path, indent=None)
+        save_json(metrics, save_path)
     return metrics  # these print nicely
 
 

examples/seq2seq/run_eval.py

@@ -152,7 +152,8 @@ def run_generate(verbose=True):
         print(scores)
 
     if args.score_path is not None:
-        json.dump(scores, open(args.score_path, "w"))
+        path = args.score_path
+        json.dump(scores, open(path, "w"))
 
     return scores
 

model_cards/TypicaAI/magbert-ner/README.md

@@ -1,55 +0,0 @@
----
-language: fr
-widget:
-- text: "Je m'appelle Hicham et je vis a Fès"
----
-
-# MagBERT-NER: a state-of-the-art NER model for Moroccan French language (Maghreb)
-
-## Introduction
-
-[MagBERT-NER] is a state-of-the-art NER model for Moroccan French language (Maghreb). The MagBERT-NER model was fine-tuned for NER Task based the language model for French Camembert (based on the RoBERTa architecture).
-
-For further information or requests, please go to [Typica.AI Website](https://typicasoft.io/)
-
-## How to use MagBERT-NER with HuggingFace
-
-##### Load MagBERT-NER and its sub-word tokenizer :
-```python
-from transformers import AutoTokenizer, AutoModelForTokenClassification
-
-tokenizer = AutoTokenizer.from_pretrained("TypicaAI/magbert-ner")
-model = AutoModelForTokenClassification.from_pretrained("TypicaAI/magbert-ner")
-
-
-##### Process text sample (from wikipedia about the current Prime Minister of Morocco) Using NER pipeline  
-
-from transformers import pipeline
-
-nlp = pipeline('ner', model=model, tokenizer=tokenizer, grouped_entities=True)
-nlp("Saad Dine El Otmani, né le 16 janvier 1956 à Inezgane, est un homme d'État marocain, chef du gouvernement du Maroc depuis le 5 avril 2017")
-
-
-#[{'entity_group': 'I-PERSON',
-#  'score': 0.8941445276141167,
-#  'word': 'Saad Dine El Otmani'},
-# {'entity_group': 'B-DATE',
-#  'score': 0.5967703461647034,
-#  'word': '16 janvier 1956'},
-# {'entity_group': 'B-GPE', 'score': 0.7160899192094803, 'word': 'Inezgane'},
-# {'entity_group': 'B-NORP', 'score': 0.7971733212471008, 'word': 'marocain'},
-# {'entity_group': 'B-GPE', 'score': 0.8921478390693665, 'word': 'Maroc'},
-# {'entity_group': 'B-DATE',
-#  'score': 0.5760444005330404,
-#  'word': '5 avril 2017'}]
-
-```
-
-```
-
-
-## Authors 
-
-MagBert-NER was trained and evaluated by Hicham Assoudi, Ph.D.
-
-

model_cards/microsoft/prophetnet-large-uncased-cnndm/README.md

@@ -0,0 +1,30 @@
+## prophetnet-large-uncased-cnndm
+Fine-tuned weights(converted from [original fairseq version repo](https://github.com/microsoft/ProphetNet)) for [ProphetNet](https://arxiv.org/abs/2001.04063) on summarization task CNN/DailyMail.  
+ProphetNet is a new pre-trained language model for sequence-to-sequence learning with a novel self-supervised objective called future n-gram prediction.  
+ProphetNet is able to predict more future tokens with a n-stream decoder. The original implementation is Fairseq version at [github repo](https://github.com/microsoft/ProphetNet).   
+
+### Usage
+A quick usage is like: 
+```
+from transformers import ProphetNetTokenizer, ProphetNetForConditionalGeneration, ProphetNetConfig
+
+model = ProphetNetForConditionalGeneration.from_pretrained('microsoft/prophetnet-large-uncased-cnndm')
+tokenizer = ProphetNetTokenizer.from_pretrained('microsoft/prophetnet-large-uncased-cnndm')
+
+ARTICLE_TO_SUMMARIZE = "USTC was founded in Beijing by the Chinese Academy of Sciences (CAS) in September 1958. The Director of CAS, Mr. Guo Moruo was appointed the first president of USTC. USTC's founding mission was to develop a high-level science and technology workforce, as deemed critical for development of China's economy, defense, and science and technology education. The establishment was hailed as \"A Major Event in the History of Chinese Education and Science.\" CAS has supported USTC by combining most of its institutes with the departments of the university. USTC is listed in the top 16 national key universities, becoming the youngest national key university.".lower()
+inputs = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=100, return_tensors='pt')
+
+# Generate Summary
+summary_ids = model.generate(inputs['input_ids'], num_beams=4, max_length=512, early_stopping=True)
+print([tokenizer.decode(g) for g in summary_ids])
+```
+Here, [X_SEP] is used as a special token to seperate sentences.
+### Citation
+```bibtex
+@article{yan2020prophetnet,
+  title={Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training},
+  author={Yan, Yu and Qi, Weizhen and Gong, Yeyun and Liu, Dayiheng and Duan, Nan and Chen, Jiusheng and Zhang, Ruofei and Zhou, Ming},
+  journal={arXiv preprint arXiv:2001.04063},
+  year={2020}
+}
+```

model_cards/microsoft/prophetnet-large-uncased/README.md

@@ -0,0 +1,17 @@
+## prophetnet-large-uncased
+Pretrained weights for [ProphetNet](https://arxiv.org/abs/2001.04063).  
+ProphetNet is a new pre-trained language model for sequence-to-sequence learning with a novel self-supervised objective called future n-gram prediction.  
+ProphetNet is able to predict more future tokens with a n-stream decoder. The original implementation is Fairseq version at [github repo](https://github.com/microsoft/ProphetNet).   
+
+### Usage
+Please see [the official repository](https://github.com/microsoft/ProphetNet) for details.
+
+### Citation
+```bibtex
+@article{yan2020prophetnet,
+  title={Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training},
+  author={Yan, Yu and Qi, Weizhen and Gong, Yeyun and Liu, Dayiheng and Duan, Nan and Chen, Jiusheng and Zhang, Ruofei and Zhou, Ming},
+  journal={arXiv preprint arXiv:2001.04063},
+  year={2020}
+}
+```

model_cards/microsoft/xprophetnet-large-wiki100-cased-xglue-ntg/README.md

@@ -0,0 +1,32 @@
+## xprophetnet-large-wiki100-cased-xglue-ntg
+Cross-lingual version [ProphetNet](https://arxiv.org/abs/2001.04063), pretrained on [wiki100 xGLUE dataset](https://arxiv.org/abs/2004.01401) and finetuned on xGLUE cross-lingual News Titles Generation task.  
+ProphetNet is a new pre-trained language model for sequence-to-sequence learning with a novel self-supervised objective called future n-gram prediction.  
+ProphetNet is able to predict more future tokens with a n-stream decoder. The original implementation is Fairseq version at [github repo](https://github.com/microsoft/ProphetNet).   
+
+xProphetNet is also served as the baseline model for xGLUE cross-lingual natural language generation tasks.  
+For xGLUE corss-lingual NLG tasks, xProphetNet is finetuned with English data, but inference with both English and other zero-shot language data.  
+### Usage
+A quick usage is like: 
+```
+from transformers import ProphetNetTokenizer, ProphetNetForConditionalGeneration, ProphetNetConfig
+
+model = ProphetNetForConditionalGeneration.from_pretrained('microsoft/xprophetnet-large-wiki100-cased-xglue-ntg')
+tokenizer = ProphetNetTokenizer.from_pretrained('microsoft/xprophetnet-large-wiki100-cased-xglue-ntg')
+
+EN_SENTENCE = "Microsoft Corporation intends to officially end free support for the Windows 7 operating system after January 14, 2020, according to the official portal of the organization. From that day, users of this system will not be able to receive security updates, which could make their computers vulnerable to cyber attacks."
+RU_SENTENCE = "орпорация Microsoft намерена официально прекратить бесплатную поддержку операционной системы Windows 7 после 14 января 2020 года, сообщается на официальном портале организации . С указанного дня пользователи этой системы не смогут получать обновления безопасности, из-за чего их компьютеры могут стать уязвимыми к кибератакам."
+ZH_SENTENCE = "根据该组织的官方门户网站，微软公司打算在2020年1月14日之后正式终止对Windows 7操作系统的免费支持。从那时起，该系统的用户将无法接收安全更新，这可能会使他们的计算机容易受到网络攻击。"
+inputs = tokenizer([EN_SENTENCE, RU_SENTENCE, ZH_SENTENCE], padding=True, max_length=256, return_tensors='pt')
+
+summary_ids = model.generate(inputs['input_ids'], num_beams=4, max_length=100, early_stopping=True)
+print([tokenizer.decode(g) for g in summary_ids])  
+```
+### Citation
+```bibtex
+@article{yan2020prophetnet,
+  title={Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training},
+  author={Yan, Yu and Qi, Weizhen and Gong, Yeyun and Liu, Dayiheng and Duan, Nan and Chen, Jiusheng and Zhang, Ruofei and Zhou, Ming},
+  journal={arXiv preprint arXiv:2001.04063},
+  year={2020}
+}
+```

model_cards/microsoft/xprophetnet-large-wiki100-cased-xglue-qg/README.md

@@ -0,0 +1,31 @@
+## xprophetnet-large-wiki100-cased-xglue-ntg
+Cross-lingual version [ProphetNet](https://arxiv.org/abs/2001.04063), pretrained on [wiki100 xGLUE dataset](https://arxiv.org/abs/2004.01401) and finetuned on xGLUE cross-lingual Question Generation task.  
+ProphetNet is a new pre-trained language model for sequence-to-sequence learning with a novel self-supervised objective called future n-gram prediction.  
+ProphetNet is able to predict more future tokens with a n-stream decoder. The original implementation is Fairseq version at [github repo](https://github.com/microsoft/ProphetNet).   
+
+xProphetNet is also served as the baseline model for xGLUE cross-lingual natural language generation tasks.  
+For xGLUE corss-lingual NLG tasks, xProphetNet is finetuned with English data, but inference with both English and other zero-shot language data.  
+### Usage
+A quick usage is like: 
+```
+from transformers import ProphetNetTokenizer, ProphetNetForConditionalGeneration, ProphetNetConfig
+
+model = ProphetNetForConditionalGeneration.from_pretrained('microsoft/xprophetnet-large-wiki100-cased-xglue-qg')
+tokenizer = ProphetNetTokenizer.from_pretrained('microsoft/xprophetnet-large-wiki100-cased-xglue-qg')
+
+EN_SENTENCE = "Google left China in 2010"
+ZH_SENTENCE = "Google在2010年离开中国"
+inputs = tokenizer([EN_SENTENCE, ZH_SENTENCE], padding=True, max_length=256, return_tensors='pt')
+
+summary_ids = model.generate(inputs['input_ids'], num_beams=4, max_length=100, early_stopping=True)
+print([tokenizer.decode(g) for g in summary_ids])  
+```
+### Citation
+```bibtex
+@article{yan2020prophetnet,
+  title={Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training},
+  author={Yan, Yu and Qi, Weizhen and Gong, Yeyun and Liu, Dayiheng and Duan, Nan and Chen, Jiusheng and Zhang, Ruofei and Zhou, Ming},
+  journal={arXiv preprint arXiv:2001.04063},
+  year={2020}
+}
+```

model_cards/microsoft/xprophetnet-large-wiki100-cased/README.md

@@ -0,0 +1,19 @@
+## xprophetnet-large-wiki100-cased
+Cross-lingual version [ProphetNet](https://arxiv.org/abs/2001.04063), pretrained on [wiki100 xGLUE dataset](https://arxiv.org/abs/2004.01401).  
+ProphetNet is a new pre-trained language model for sequence-to-sequence learning with a novel self-supervised objective called future n-gram prediction.  
+ProphetNet is able to predict more future tokens with a n-stream decoder. The original implementation is Fairseq version at [github repo](https://github.com/microsoft/ProphetNet).   
+
+xProphetNet is also served as the baseline model for xGLUE cross-lingual natural language generation tasks.  
+For xGLUE corss-lingual NLG tasks, xProphetNet is finetuned with English data, but inference with both English and other zero-shot language data.  
+### Usage
+Please see [the official repository](https://github.com/microsoft/ProphetNet/tree/master/xProphetNet) for details.
+
+### Citation
+```bibtex
+@article{yan2020prophetnet,
+  title={Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training},
+  author={Yan, Yu and Qi, Weizhen and Gong, Yeyun and Liu, Dayiheng and Duan, Nan and Chen, Jiusheng and Zhang, Ruofei and Zhou, Ming},
+  journal={arXiv preprint arXiv:2001.04063},
+  year={2020}
+}
+```

model_cards/mrm8488/electricidad-base-discriminator/README.md

@@ -59,21 +59,9 @@ predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)
 el rapido  zorro  marro    ##n   amar  sobre     el  perro   pere ##zoso    0.0    0.0    0.0    0.0    0.0    0.0    1.0    1.0    0.0    0.0    0.0    0.0    0.0[None, None, None, None, None, None, None, None, None, None, None, None, None
 '''
 ```
+
 As you can see there are **1s** in the places where the model detected a fake token. So, it works! 🎉
 
-
-### Some models fine-tuned on a downstream task 🛠️
-
-[Question Answering](https://huggingface.co/mrm8488/electricidad-base-finetuned-squadv1-es)
-
-[POS](https://huggingface.co/mrm8488/electricidad-base-finetuned-pos)
-
-[NER](https://huggingface.co/mrm8488/electricidad-base-finetuned-ner)
-
-[Paraphrase Identification](https://huggingface.co/mrm8488/RuPERTa-base-finetuned-pawsx-es)
-
-
-
 ## Acknowledgments
 
 I thank [🤗/transformers team](https://github.com/huggingface/transformers) for allowing me to train the model (specially to [Julien Chaumond](https://twitter.com/julien_c)).

model_cards/unideeplearning/polibert_sa/README.md

@@ -12,22 +12,15 @@ widget:
   
 ## Model description  
   
-This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets. You can try it out at https://www.unideeplearning.com/twitter_sa/ (in italian!)
+This model performs sentiment analysis on Italian political twitter sentences. It was trained starting from an instance of "bert-base-italian-uncased-xxl" and fine-tuned on an Italian dataset of tweets.
   
 #### Hands-on  
   
 ```python
 import torch
 from torch import nn 
-from transformers import AutoTokenizer, AutoModelForSequenceClassification
 
-tokenizer = AutoTokenizer.from_pretrained("unideeplearning/polibert_sa")
-model = AutoModelForSequenceClassification.from_pretrained("unideeplearning/polibert_sa")
-			
-
-
-
-text = "Giuseppe Rossi è un pessimo politico"
+text = "Giueseppe Rossi è un pessimo politico"
 input_ids = tokenizer.encode(text, add_special_tokens=True, return_tensors= 'pt')
 
 logits, = model(input_ids)
@@ -48,6 +41,4 @@ print(prob.argmax().tolist())
 ## Acknowledgments
 
 Thanks to the support from: 
-the [Hugging Face](https://huggingface.co/), https://www.unioneprofessionisti.com
-
-https://www.unideeplearning.com/
+the [Hugging Face](https://huggingface.co/), Unione Professionisti (https://www.unioneprofessionisti.com/)

setup.py

@@ -5,7 +5,7 @@ To create the package for pypi.
 
 1. Change the version in __init__.py, setup.py as well as docs/source/conf.py.
 
-2. Unpin specific versions from setup.py that use a git install.
+2. Unpin specific versions from setup.py (like isort).
 
 2. Commit these changes with the message: "Release: VERSION"
 
@@ -98,7 +98,7 @@ extras["dev"] = extras["testing"] + extras["quality"] + extras["ja"] + ["scikit-
 
 setup(
     name="transformers",
-    version="3.3.1",
+    version="3.2.0",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Sylvain Gugger, 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",

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "3.3.1"
+__version__ = "3.2.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -52,6 +52,7 @@ from .configuration_mmbt import MMBTConfig
 from .configuration_mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
 from .configuration_pegasus import PegasusConfig
+from .configuration_prophetnet import PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ProphetNetConfig
 from .configuration_rag import RagConfig
 from .configuration_reformer import REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
@@ -60,6 +61,7 @@ from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
+from .configuration_xlm_prophetnet import XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMProphetNetConfig
 from .configuration_xlm_roberta import XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMRobertaConfig
 from .configuration_xlnet import XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLNetConfig
 from .data import (
@@ -178,6 +180,7 @@ from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFas
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_pegasus import PegasusTokenizer
 from .tokenization_phobert import PhobertTokenizer
+from .tokenization_prophetnet import ProphetNetTokenizer
 from .tokenization_rag import RagTokenizer
 from .tokenization_reformer import ReformerTokenizer
 from .tokenization_retribert import RetriBertTokenizer, RetriBertTokenizerFast
@@ -195,6 +198,7 @@ from .tokenization_utils_base import (
 )
 from .tokenization_utils_fast import PreTrainedTokenizerFast
 from .tokenization_xlm import XLMTokenizer
+from .tokenization_xlm_prophetnet import XLMProphetNetTokenizer
 from .tokenization_xlm_roberta import XLMRobertaTokenizer
 from .tokenization_xlnet import SPIECE_UNDERLINE, XLNetTokenizer
 
@@ -423,6 +427,12 @@ if is_torch_available():
         load_tf_weights_in_openai_gpt,
     )
     from .modeling_pegasus import PegasusForConditionalGeneration
+    from .modeling_prophetnet import (
+        PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST,
+        ProphetNetForConditionalGeneration,
+        ProphetNetModel,
+        ProphetNetPreTrainedModel,
+    )
     from .modeling_rag import RagModel, RagSequenceForGeneration, RagTokenForGeneration
     from .modeling_reformer import (
         REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -472,6 +482,11 @@ if is_torch_available():
         XLMPreTrainedModel,
         XLMWithLMHeadModel,
     )
+    from .modeling_xlm_prophetnet import (
+        XLM_PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST,
+        XLMProphetNetForConditionalGeneration,
+        XLMProphetNetModel,
+    )
     from .modeling_xlm_roberta import (
         XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
         XLMRobertaForCausalLM,

src/transformers/configuration_auto.py

@@ -39,6 +39,7 @@ from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfi
 from .configuration_mobilebert import MobileBertConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
 from .configuration_pegasus import PegasusConfig
+from .configuration_prophetnet import PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, ProphetNetConfig
 from .configuration_rag import RagConfig
 from .configuration_reformer import ReformerConfig
 from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, RetriBertConfig
@@ -47,6 +48,7 @@ from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
+from .configuration_xlm_prophetnet import XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMProphetNetConfig
 from .configuration_xlm_roberta import XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMRobertaConfig
 from .configuration_xlnet import XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLNetConfig
 
@@ -78,6 +80,8 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
         DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -116,6 +120,8 @@ CONFIG_MAPPING = OrderedDict(
         ("dpr", DPRConfig),
         ("layoutlm", LayoutLMConfig),
         ("rag", RagConfig),
+        ("prophetnet", ProphetNetConfig),
+        ("xlm-prophetnet", XLMProphetNetConfig),
     ]
 )
 
@@ -152,6 +158,8 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("layoutlm", "LayoutLM"),
         ("dpr", "DPR"),
         ("rag", "RAG"),
+        ("prophetnet", "ProphetNet"),
+        ("xlm-prophetnet", "XLMProphetNet"),
     ]
 )
 

src/transformers/configuration_gpt2.py

@@ -103,8 +103,6 @@ class GPT2Config(PretrainedConfig):
             :class:`~transformers.GPT2DoubleHeadsModel` and :class:`~transformers.TFGPT2DoubleHeadsModel`.
 
             The dropout ratio to be used after the projection and activation.
-        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.
 
     Example::
 
@@ -144,7 +142,6 @@ class GPT2Config(PretrainedConfig):
         summary_first_dropout=0.1,
         bos_token_id=50256,
         eos_token_id=50256,
-        gradient_checkpointing=False,
         **kwargs
     ):
         super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
@@ -167,7 +164,6 @@ class GPT2Config(PretrainedConfig):
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        self.gradient_checkpointing = gradient_checkpointing
 
         self.bos_token_id = bos_token_id
         self.eos_token_id = eos_token_id

src/transformers/configuration_prophetnet.py

@@ -0,0 +1,104 @@
+# coding=utf-8
+# Copyright 2020 The Microsoft Authors and 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.
+""" ProphetNet model configuration """
+
+
+import logging
+
+from .configuration_utils import PretrainedConfig
+
+
+logger = logging.getLogger(__name__)
+
+PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/prophetnet-large-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/microsoft/prophetnet-large-uncased/config.json",
+}
+
+
+class ProphetNetConfig(PretrainedConfig):
+    r"""
+    Configuration class for ProphetNet. Parameters are renamed from the fairseq implementation
+    """
+    model_type = "prophetnet"
+
+    def __init__(
+        self,
+        activation_dropout=0.1,
+        activation_function="gelu",
+        vocab_size=30522,
+        hidden_size=1024,
+        encoder_ffn_dim=4096,
+        num_encoder_layers=12,
+        num_encoder_attention_heads=16,
+        decoder_ffn_dim=4096,
+        num_decoder_layers=12,
+        num_decoder_attention_heads=16,
+        encoder_layerdrop=0.1,
+        decoder_layerdrop=0.1,
+        attention_dropout=0.1,
+        dropout=0.1,
+        max_position_embeddings=512,
+        init_std=0.02,
+        is_encoder_decoder=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        ngram=2,
+        num_buckets=32,
+        relative_max_distance=128,
+        disable_ngram_loss=False,
+        eps=0.0,
+        **common_kwargs
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            **common_kwargs,
+        )
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.num_encoder_layers = num_encoder_layers
+        self.num_encoder_attention_heads = num_encoder_attention_heads
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.num_decoder_layers = num_decoder_layers
+        self.num_decoder_attention_heads = num_decoder_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.init_std = init_std  # Normal(0, this parameter)
+        self.activation_function = activation_function
+
+        # parameters for prophetnet
+        self.ngram = ngram
+        self.num_buckets = num_buckets
+        self.relative_max_distance = relative_max_distance
+        self.disable_ngram_loss = disable_ngram_loss
+        self.eps = eps
+
+        # 3 Types of Dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.dropout = dropout
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.num_encoder_attention_heads
+
+    @property
+    def num_hidden_layers(self) -> int:
+        return self.num_encoder_layers + self.num_decoder_layers

src/transformers/configuration_xlm_prophetnet.py

@@ -0,0 +1,22 @@
+""" XLM-ProphetNet model configuration """
+
+
+import logging
+
+from .configuration_prophetnet import ProphetNetConfig
+
+
+logger = logging.getLogger(__name__)
+
+XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/xprophetnet-large-wiki100-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/microsoft/xprophetnet-large-wiki100-cased/config.json",
+}
+
+
+class XLMProphetNetConfig(ProphetNetConfig):
+    """
+    This class overrides :class:`~transformers.RobertaConfig`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
+    """
+
+    model_type = "xlm-prophetnet"

src/transformers/file_utils.py

@@ -68,12 +68,8 @@ except (ImportError, AssertionError):
 try:
     import datasets  # noqa: F401
 
-    # Check we're not importing a "datasets" directory somewhere
-    _datasets_available = hasattr(datasets, "__version__") and hasattr(datasets, "load_dataset")
-    if _datasets_available:
-        logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
-    else:
-        logger.debug("Imported a datasets object but this doesn't seem to be the 🤗 datasets library.")
+    _datasets_available = True
+    logger.debug(f"Succesfully imported datasets version {datasets.__version__}")
 
 except ImportError:
     _datasets_available = False
@@ -167,7 +163,8 @@ def is_torch_available():
 
 
 def is_tf_available():
-    return _tf_available
+    return False
+#    return _tf_available
 
 
 def is_torch_tpu_available():

src/transformers/modeling_auto.py

@@ -41,12 +41,14 @@ from .configuration_auto import (
     MobileBertConfig,
     OpenAIGPTConfig,
     PegasusConfig,
+    ProphetNetConfig,
     ReformerConfig,
     RetriBertConfig,
     RobertaConfig,
     T5Config,
     TransfoXLConfig,
     XLMConfig,
+    XLMProphetNetConfig,
     XLMRobertaConfig,
     XLNetConfig,
     replace_list_option_in_docstrings,
@@ -150,6 +152,7 @@ from .modeling_mobilebert import (
 )
 from .modeling_openai import OpenAIGPTLMHeadModel, OpenAIGPTModel
 from .modeling_pegasus import PegasusForConditionalGeneration
+from .modeling_prophetnet import ProphetNetForConditionalGeneration, ProphetNetModel
 from .modeling_rag import (  # noqa: F401 - need to import all RagModels to be in globals() function
     RagModel,
     RagSequenceForGeneration,
@@ -181,6 +184,7 @@ from .modeling_xlm import (
     XLMModel,
     XLMWithLMHeadModel,
 )
+from .modeling_xlm_prophetnet import XLMProphetNetForConditionalGeneration, XLMProphetNetModel
 from .modeling_xlm_roberta import (
     XLMRobertaForCausalLM,
     XLMRobertaForMaskedLM,
@@ -232,6 +236,8 @@ MODEL_MAPPING = OrderedDict(
         (LxmertConfig, LxmertModel),
         (BertGenerationConfig, BertGenerationEncoder),
         (DPRConfig, DPRQuestionEncoder),
+        (ProphetNetConfig, ProphetNetModel),
+        (XLMProphetNetConfig, XLMProphetNetModel),
     ]
 )
 
@@ -340,6 +346,8 @@ MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
         (BartConfig, BartForConditionalGeneration),
         (FSMTConfig, FSMTForConditionalGeneration),
         (EncoderDecoderConfig, EncoderDecoderModel),
+        (ProphetNetConfig, ProphetNetForConditionalGeneration),
+        (XLMProphetNetConfig, XLMProphetNetForConditionalGeneration),
     ]
 )
 

src/transformers/modeling_gpt2.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 """PyTorch OpenAI GPT-2 model."""
 
+
 import os
 import warnings
 from dataclasses import dataclass
@@ -623,35 +624,16 @@ class GPT2Model(GPT2PreTrainedModel):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
 
-            if getattr(self.config, "gradient_checkpointing", False):
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        # checkpointing only works with tuple returns, not with lists
-                        return tuple(output for output in module(*inputs, use_cache, output_attentions))
-
-                    return custom_forward
-
-                outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
-                    hidden_states,
-                    layer_past,
-                    attention_mask,
-                    head_mask[i],
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                )
-            else:
-                outputs = block(
-                    hidden_states,
-                    layer_past=layer_past,
-                    attention_mask=attention_mask,
-                    head_mask=head_mask[i],
-                    encoder_hidden_states=encoder_hidden_states,
-                    encoder_attention_mask=encoder_attention_mask,
-                    use_cache=use_cache,
-                    output_attentions=output_attentions,
-                )
+            outputs = block(
+                hidden_states,
+                layer_past=layer_past,
+                attention_mask=attention_mask,
+                head_mask=head_mask[i],
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+            )
 
             hidden_states, present = outputs[:2]
             if use_cache is True:

src/transformers/modeling_xlm_prophetnet.py

@@ -0,0 +1,31 @@
+from .configuration_xlm_prophetnet import XLMProphetNetConfig
+from .modeling_prophetnet import ProphetNetForConditionalGeneration, ProphetNetModel
+from .utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+_TOKENIZER_FOR_DOC = "XLMProphetNetTokenizer"
+
+XLM_PROPHETNET_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/xprophetnet-large-wiki100-cased",
+    # See all ProphetNet models at https://huggingface.co/models?filter=xprophetnet
+]
+
+
+class XLMProphetNetModel(ProphetNetModel):
+    """
+    This class overrides :class:`~transformers.ProphetNetModel`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
+    """
+
+    config_class = XLMProphetNetConfig
+
+
+class XLMProphetNetForConditionalGeneration(ProphetNetForConditionalGeneration):
+    """
+    This class overrides :class:`~transformers.ProphetNetForConditionalGeneration`. Please check the
+    superclass for the appropriate documentation alongside usage examples.
+    """
+
+    config_class = XLMProphetNetConfig

src/transformers/tokenization_auto.py

@@ -41,6 +41,7 @@ from .configuration_auto import (
     MobileBertConfig,
     OpenAIGPTConfig,
     PegasusConfig,
+    ProphetNetConfig,
     RagConfig,
     ReformerConfig,
     RetriBertConfig,
@@ -48,6 +49,7 @@ from .configuration_auto import (
     T5Config,
     TransfoXLConfig,
     XLMConfig,
+    XLMProphetNetConfig,
     XLMRobertaConfig,
     XLNetConfig,
     replace_list_option_in_docstrings,
@@ -77,6 +79,7 @@ from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFas
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_pegasus import PegasusTokenizer
 from .tokenization_phobert import PhobertTokenizer
+from .tokenization_prophetnet import ProphetNetTokenizer
 from .tokenization_rag import RagTokenizer
 from .tokenization_reformer import ReformerTokenizer
 from .tokenization_retribert import RetriBertTokenizer, RetriBertTokenizerFast
@@ -84,6 +87,7 @@ from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
 from .tokenization_t5 import T5Tokenizer
 from .tokenization_transfo_xl import TransfoXLTokenizer, TransfoXLTokenizerFast
 from .tokenization_xlm import XLMTokenizer
+from .tokenization_xlm_prophetnet import XLMProphetNetTokenizer
 from .tokenization_xlm_roberta import XLMRobertaTokenizer
 from .tokenization_xlnet import XLNetTokenizer
 from .utils import logging
@@ -127,6 +131,8 @@ TOKENIZER_MAPPING = OrderedDict(
         (BertGenerationConfig, (BertGenerationTokenizer, None)),
         (LayoutLMConfig, (LayoutLMTokenizer, None)),
         (RagConfig, (RagTokenizer, None)),
+        (ProphetNetConfig, (ProphetNetTokenizer, None)),
+        (XLMProphetNetConfig, (XLMProphetNetTokenizer, None)),
     ]
 )
 

src/transformers/tokenization_prophetnet.py

@@ -0,0 +1,290 @@
+# coding=utf-8
+# Copyright 2020 The Microsoft Authors and 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.
+
+import collections
+import logging
+import os
+from typing import List, Optional
+
+from .tokenization_bert import BasicTokenizer, WordpieceTokenizer
+from .tokenization_utils import PreTrainedTokenizer
+
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "prophetnet.tokenizer"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "microsoft/prophetnet-large-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/microsoft/prophetnet-large-uncased/prophetnet.tokenizer",
+    }
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "microsoft/prophetnet-large-uncased": {"do_lower_case": True, "xprophetnet_tokenizer": False},
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "microsoft/prophetnet-large-uncased": 512,
+}
+
+
+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 ProphetNetTokenizer(PreTrainedTokenizer):
+    r"""
+    Construct a ProphetNetTokenizer. Based on WordPiece.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            File containing the vocabulary.
+        do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to lowercase the input when tokenizing.
+        do_basic_tokenize (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to do basic tokenization before WordPiece.
+        never_split (:obj:`Iterable`, `optional`):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            :obj:`do_basic_tokenize=True`
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"[UNK]"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences
+            for sequence classification or for a text and a question for question answering.
+            It is also used as the last token of a sequence built with special tokens.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"[PAD]"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole
+            sequence instead of per-token classification). It is the first token of the sequence when built with
+            special tokens.
+        mask_token (:obj:`str`, `optional`, defaults to :obj:`"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        tokenize_chinese_chars (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to tokenize Chinese characters.
+
+            This should likely be deactivated for Japanese (see this `issue
+            <https://github.com/huggingface/transformers/issues/328>`__).
+        strip_accents: (:obj:`bool`, `optional`):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for :obj:`lowercase` (as in the original BERT).
+    """
+
+    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,
+        do_basic_tokenize=True,
+        never_split=None,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        x_sep_token="[X_SEP]",
+        pad_token="[PAD]",
+        mask_token="[MASK]",
+        tokenize_chinese_chars=True,
+        strip_accents=None,
+        **kwargs
+    ):
+        super().__init__(
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            x_sep_token=x_sep_token,
+            **kwargs,
+        )
+        self.unique_no_split_tokens.append(x_sep_token)
+
+        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 = ProphetNetTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file)
+            )
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])
+        self.do_basic_tokenize = do_basic_tokenize
+        if do_basic_tokenize:
+            self.basic_tokenizer = BasicTokenizer(
+                do_lower_case=do_lower_case,
+                never_split=never_split,
+                tokenize_chinese_chars=tokenize_chinese_chars,
+                strip_accents=strip_accents,
+            )
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=self.unk_token)
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def get_vocab(self):
+        return dict(self.vocab, **self.added_tokens_encoder)
+
+    def _tokenize(self, text):
+        split_tokens = []
+        if self.do_basic_tokenize:
+            for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
+
+                # If the token is part of the never_split set
+                if token in self.basic_tokenizer.never_split:
+                    split_tokens.append(token)
+                else:
+                    split_tokens += self.wordpiece_tokenizer.tokenize(token)
+        else:
+            split_tokens = self.wordpiece_tokenizer.tokenize(text)
+        return split_tokens
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str) 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 (str) 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 get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve 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`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`List[int]`: 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: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A ProphetNet sequence pair mask has the following format:
+
+        ::
+
+            0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+            | first sequence    | second sequence |
+
+        If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
+            sequence(s).
+        """
+        sep = [self.sep_token_id]
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, vocab_path):
+        """
+        Save the vocabulary (copy original file) and special tokens file to a directory.
+
+        Args:
+            vocab_path (:obj:`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            :obj:`Tuple(str)`: Paths to the files saved.
+        """
+        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 + "\n")
+                index += 1
+        return (vocab_file,)
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        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]``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return token_ids_0 + [self.sep_token_id]
+        sep = [self.sep_token_id]
+        return token_ids_0 + sep + token_ids_1 + sep

src/transformers/tokenization_xlm_prophetnet.py

@@ -0,0 +1,314 @@
+# coding=utf-8
+# Copyright 2020 The Microsoft Authors and 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.
+
+import collections
+import logging
+import os
+from shutil import copyfile
+from typing import List, Optional
+
+from .tokenization_utils import PreTrainedTokenizer
+
+
+logger = logging.getLogger(__name__)
+
+SPIECE_UNDERLINE = "▁"
+
+VOCAB_FILES_NAMES = {"vocab_file": "prophetnet.tokenizer"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "microsoft/xprophetnet-large-wiki100-cased": "https://cdn.huggingface.co/microsoft/xprophetnet-large-wiki100-cased/prophetnet.tokenizer",
+    }
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "microsoft/xprophetnet-large-wiki100-cased": {"do_lower_case": False, "xprophetnet_tokenizer": True},
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "microsoft/xprophetnet-large-wiki100-cased": 512,
+}
+
+
+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 XLMProphetNetTokenizer(PreTrainedTokenizer):
+    """
+    Adapted from :class:`~transfomers.RobertaTokenizer` and class:`~transfomers.XLNetTokenizer`. Based on
+    `SentencePiece <https://github.com/google/sentencepiece>`__.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the main methods.
+    Users should refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (:obj:`str`):
+            Path to the vocabulary file.
+        bos_token (:obj:`str`, `optional`, defaults to :obj:`"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            .. note::
+
+                When building a sequence using special tokens, this is not the token that is used for the beginning
+                of sequence. The token used is the :obj:`cls_token`.
+        eos_token (:obj:`str`, `optional`, defaults to :obj:`"</s>"`):
+            The end of sequence token.
+
+            .. note::
+
+                When building a sequence using special tokens, this is not the token that is used for the end
+                of sequence. The token used is the :obj:`sep_token`.
+        sep_token (:obj:`str`, `optional`, defaults to :obj:`"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences
+            for sequence classification or for a text and a question for question answering.
+            It is also used as the last token of a sequence built with special tokens.
+        cls_token (:obj:`str`, `optional`, defaults to :obj:`"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole
+            sequence instead of per-token classification). It is the first token of the sequence when built with
+            special tokens.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        mask_token (:obj:`str`, `optional`, defaults to :obj:`"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
+            Additional special tokens used by the tokenizer.
+
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        bos_token="[SEP]",
+        eos_token="[SEP]",
+        sep_token="[SEP]",
+        unk_token="[UNK]",
+        pad_token="[PAD]",
+        mask_token="[MASK]",
+        **kwargs
+    ):
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            mask_token=mask_token,
+            **kwargs,
+        )
+
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning(
+                "You need to install SentencePiece to use XLMRobertaTokenizer: https://github.com/google/sentencepiece"
+                "pip install sentencepiece"
+            )
+            raise
+
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(str(vocab_file))
+        self.vocab_file = vocab_file
+
+        # Original fairseq vocab and spm vocab must be "aligned":
+        # Vocab    |    0    |    1    |   2    |    3    |  4  |  5  |  6  |   7   |   8   |  9
+        # -------- | ------- | ------- | ------ | ------- | --- | --- | --- | ----- | ----- | ----
+        # fairseq  | '<s>'   | '<pad>' | '</s>' | '<unk>' | ',' | '.' | '▁' | 's'   | '▁de' | '-'
+        # spm      | '<unk>' | '<s>'   | '</s>' | ','     | '.' | '▁' | 's' | '▁de' | '-'   | '▁a'
+
+        # put special tokens and [unused] tokens into the vocab
+        self.fairseq_tokens_to_ids = {"[PAD]": 0, "[CLS]": 1, "[SEP]": 2, "[UNK]": 3, "[MASK]": 4}
+
+        for i in range(10):
+            tok = "[unused{}]".format(i)
+            self.fairseq_tokens_to_ids[tok] = 5 + i
+
+        # The first "real" token "," has position 15 in the embedding vocab and position 3 in the spm vocab
+        self.fairseq_offset = 12
+        self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
+        for k in self.fairseq_tokens_to_ids.keys():
+            self.unique_no_split_tokens.append(k)
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning(
+                "You need to install SentencePiece to use XLMRobertaTokenizer: https://github.com/google/sentencepiece"
+                "pip install sentencepiece"
+            )
+            raise
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(self.vocab_file)
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve 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`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`List[int]`: 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 None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task.
+        XLM-RoBERTa does not make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of zeros.
+
+        """
+
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + sep) * [0]
+        return len(token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
+
+    @property
+    def vocab_size(self):
+        return len(self.sp_model) + self.fairseq_offset + 1  # Add the <mask> token
+
+    def get_vocab(self):
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        return self.sp_model.EncodeAsPieces(text)
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str) in an id using the vocab. """
+        if token in self.fairseq_tokens_to_ids:
+            return self.fairseq_tokens_to_ids[token]
+        spm_id = self.sp_model.PieceToId(token)
+
+        # Need to return unknown token if the SP model returned 0
+        return spm_id + self.fairseq_offset if spm_id else self.unk_token_id
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        if index in self.fairseq_ids_to_tokens:
+            return self.fairseq_ids_to_tokens[index]
+        return self.sp_model.IdToPiece(index - self.fairseq_offset)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = "".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory):
+        """
+        Save the sentencepiece vocabulary (copy original file) and special tokens file to a directory.
+
+        Args:
+            save_directory (:obj:`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            :obj:`Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES["vocab_file"])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
+        by concatenating and adding special tokens.
+        A ProphetNet sequence has the following format:
+
+        - single sequence: ``X [SEP]``
+        - pair of sequences: ``A [SEP] B [SEP]``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: list of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+
+        if token_ids_1 is None:
+            return token_ids_0 + [self.sep_token_id]
+        sep = [self.sep_token_id]
+        return token_ids_0 + sep + token_ids_1 + sep

src/transformers/tokenization_xlm_roberta.py

@@ -20,12 +20,13 @@ from shutil import copyfile
 from typing import List, Optional
 
 from .tokenization_utils import PreTrainedTokenizer
-from .tokenization_xlnet import SPIECE_UNDERLINE
 from .utils import logging
 
 
 logger = logging.get_logger(__name__)
 
+SPIECE_UNDERLINE = "▁"
+
 VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model"}
 
 PRETRAINED_VOCAB_FILES_MAP = {

src/transformers/trainer.py

@@ -20,7 +20,7 @@ from torch.utils.data.sampler import RandomSampler, Sampler, SequentialSampler
 from tqdm.auto import tqdm, trange
 
 from .data.data_collator import DataCollator, DataCollatorWithPadding, default_data_collator
-from .file_utils import WEIGHTS_NAME, is_datasets_available, is_torch_tpu_available
+from .file_utils import is_datasets_available, is_torch_tpu_available
 from .integrations import (
     default_hp_search_backend,
     is_comet_available,
@@ -42,7 +42,6 @@ from .trainer_utils import (
     EvaluationStrategy,
     HPSearchBackend,
     PredictionOutput,
-    TrainerState,
     TrainOutput,
     default_compute_objective,
     default_hp_space,
@@ -60,8 +59,6 @@ from .utils import logging
 _use_native_amp = False
 _use_apex = False
 
-PT_LR_SCHEDULER_WARNING = "Please also save or load the state of the optimzer when saving or loading the scheduler."
-
 # Check if Pytorch version >= 1.6 to switch between Native AMP and Apex
 if version.parse(torch.__version__) < version.parse("1.6"):
     from .file_utils import is_apex_available
@@ -102,14 +99,6 @@ if is_ray_available():
 logger = logging.get_logger(__name__)
 
 
-def reissue_pt_warnings(caught_warnings):
-    # Reissue warnings that are not the PT_LR_SCHEDULER_WARNING
-    if len(caught_warnings) > 1:
-        for w in caught_warnings:
-            if w.category != UserWarning or w.message != PT_LR_SCHEDULER_WARNING:
-                warnings.warn(w.message, w.category)
-
-
 @contextmanager
 def torch_distributed_zero_first(local_rank: int):
     """
@@ -643,7 +632,6 @@ class Trainer:
             self.args.max_steps = t_total
 
         self.create_optimizer_and_scheduler(num_training_steps=t_total)
-        self.state = TrainerState()
 
         # Check if saved optimizer or scheduler states exist
         if (
@@ -655,13 +643,7 @@ class Trainer:
             self.optimizer.load_state_dict(
                 torch.load(os.path.join(model_path, "optimizer.pt"), map_location=self.args.device)
             )
-            with warnings.catch_warnings(record=True) as caught_warnings:
-                self.lr_scheduler.load_state_dict(torch.load(os.path.join(model_path, "scheduler.pt")))
-            reissue_pt_warnings(caught_warnings)
-
-        # Check if a saved Trainer state exist
-        if model_path is not None and os.path.isfile(os.path.join(model_path, "trainer_state.json")):
-            self.state = TrainerState.load_from_json(os.path.join(model_path, "trainer_state.json"))
+            self.lr_scheduler.load_state_dict(torch.load(os.path.join(model_path, "scheduler.pt")))
 
         model = self.model
         if self.args.fp16 and _use_apex:
@@ -679,10 +661,8 @@ class Trainer:
                 model,
                 device_ids=[self.args.local_rank],
                 output_device=self.args.local_rank,
-                find_unused_parameters=not getattr(model.config, "gradient_checkpointing", False),
+                find_unused_parameters=True,
             )
-        # find_unused_parameters breaks checkpointing as per
-        # https://github.com/huggingface/transformers/pull/4659#issuecomment-643356021
 
         if self.tb_writer is not None:
             self.tb_writer.add_text("args", self.args.to_json_string())
@@ -811,15 +791,40 @@ class Trainer:
                     ):
                         metrics = self.evaluate()
                         self._report_to_hp_search(trial, epoch, metrics)
-                        if self.args.load_best_model_at_end:
-                            self._save_training(model, trial, metrics=metrics)
 
-                    if (
-                        not self.args.load_best_model_at_end
-                        and self.args.save_steps > 0
-                        and self.global_step % self.args.save_steps == 0
-                    ):
-                        self._save_training(model, trial)
+                    if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
+                        # In all cases (even distributed/parallel), self.model is always a reference
+                        # to the model we want to save.
+                        if hasattr(model, "module"):
+                            assert (
+                                model.module is self.model
+                            ), f"Module {model.module} should be a reference to self.model"
+                        else:
+                            assert model is self.model, f"Model {model} should be a reference to self.model"
+                        # Save model checkpoint
+                        checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}"
+                        if self.hp_search_backend is not None and trial is not None:
+                            run_id = (
+                                trial.number
+                                if self.hp_search_backend == HPSearchBackend.OPTUNA
+                                else tune.get_trial_id()
+                            )
+                            checkpoint_folder += f"-run-{run_id}"
+                        output_dir = os.path.join(self.args.output_dir, checkpoint_folder)
+
+                        self.store_flos()
+                        self.save_model(output_dir)
+
+                        if self.is_world_process_zero():
+                            self._rotate_checkpoints(use_mtime=True)
+
+                        if is_torch_tpu_available():
+                            xm.rendezvous("saving_optimizer_states")
+                            xm.save(self.optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                            xm.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                        elif self.is_world_process_zero():
+                            torch.save(self.optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                            torch.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
 
                 epoch_pbar.update(1)
                 if self.args.max_steps > 0 and self.global_step >= self.args.max_steps:
@@ -830,8 +835,6 @@ class Trainer:
             if self.args.evaluation_strategy == EvaluationStrategy.EPOCH:
                 metrics = self.evaluate()
                 self._report_to_hp_search(trial, epoch, metrics)
-                if self.args.load_best_model_at_end:
-                    self._save_training(model, trial, metrics=metrics)
 
             if self.args.tpu_metrics_debug or self.args.debug:
                 if is_torch_tpu_available():
@@ -853,73 +856,8 @@ class Trainer:
             delattr(self, "_past")
 
         logger.info("\n\nTraining completed. Do not forget to share your model on huggingface.co/models =)\n\n")
-        if self.args.load_best_model_at_end and self.state.best_model_checkpoint is not None:
-            logger.info(
-                f"Loading best model from {self.state.best_model_checkpoint} (score: {self.state.best_metric})."
-            )
-            if isinstance(model, PreTrainedModel):
-                self.model = model.from_pretrained(self.state.best_model_checkpoint)
-                self.model = self.model.to(self.args.device)
-            else:
-                state_dict = torch.load(os.path.join(self.state.best_model_checkpoint, WEIGHTS_NAME))
-                self.model.load_state_dict(state_dict)
-
         return TrainOutput(self.global_step, tr_loss.item() / self.global_step)
 
-    def _save_training(self, model, trial, metrics=None):
-        # In all cases (even distributed/parallel), self.model is always a reference
-        # to the model we want to save.
-        if hasattr(model, "module"):
-            assert model.module is self.model, f"Module {model.module} should be a reference to self.model"
-        else:
-            assert model is self.model, f"Model {model} should be a reference to self.model"
-        # Save model checkpoint
-        checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}"
-        if self.hp_search_backend is not None and trial is not None:
-            run_id = trial.number if self.hp_search_backend == HPSearchBackend.OPTUNA else tune.get_trial_id()
-            checkpoint_folder += f"-run-{run_id}"
-        output_dir = os.path.join(self.args.output_dir, checkpoint_folder)
-
-        self.store_flos()
-        self.save_model(output_dir)
-
-        # Save optimizer and scheduler
-        if is_torch_tpu_available():
-            xm.rendezvous("saving_optimizer_states")
-            xm.save(self.optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
-            with warnings.catch_warnings(record=True) as caught_warnings:
-                xm.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
-                reissue_pt_warnings(caught_warnings)
-        elif self.is_world_process_zero():
-            torch.save(self.optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
-            with warnings.catch_warnings(record=True) as caught_warnings:
-                torch.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
-            reissue_pt_warnings(caught_warnings)
-
-        # Determine the new best metric / best model checkpoint
-        if metrics is not None:
-            metric_to_check = self.args.metric_for_best_model
-            if not metric_to_check.startswith("eval_"):
-                metric_to_check = f"eval_{metric_to_check}"
-            metric_value = metrics[metric_to_check]
-
-            operator = np.greater if self.args.greater_is_better else np.less
-            if (
-                self.state.best_metric is None
-                or self.state.best_model_checkpoint is None
-                or operator(metric_value, self.state.best_metric)
-            ):
-                self.state.best_metric = metric_value
-                self.state.best_model_checkpoint = output_dir
-
-        # Save the Trainer state
-        if self.is_world_process_zero():
-            self.state.save_to_json(os.path.join(output_dir, "trainer_state.json"))
-
-        # Maybe delete some older checkpoints.
-        if self.is_world_process_zero():
-            self._rotate_checkpoints(use_mtime=True)
-
     def hyperparameter_search(
         self,
         hp_space: Optional[Callable[["optuna.Trial"], Dict[str, float]]] = None,
@@ -1210,13 +1148,11 @@ class Trainer:
 
         # Save a trained model and configuration using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`
-        xm.rendezvous("saving_checkpoint")
         if not isinstance(self.model, PreTrainedModel):
-            logger.info("Trainer.model is not a `PreTrainedModel`, only saving its state dict.")
-            state_dict = self.model.state_dict()
-            xm.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
-        else:
-            self.model.save_pretrained(output_dir)
+            raise ValueError("Trainer.model appears to not be a PreTrainedModel")
+
+        xm.rendezvous("saving_checkpoint")
+        self.model.save_pretrained(output_dir)
         if self.tokenizer is not None:
             self.tokenizer.save_pretrained(output_dir)
 
@@ -1227,11 +1163,8 @@ class Trainer:
         # Save a trained model and configuration using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`
         if not isinstance(self.model, PreTrainedModel):
-            logger.info("Trainer.model is not a `PreTrainedModel`, only saving its state dict.")
-            state_dict = self.model.state_dict()
-            torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
-        else:
-            self.model.save_pretrained(output_dir)
+            raise ValueError("Trainer.model appears to not be a PreTrainedModel")
+        self.model.save_pretrained(output_dir)
         if self.tokenizer is not None:
             self.tokenizer.save_pretrained(output_dir)
 
@@ -1266,13 +1199,6 @@ class Trainer:
 
         checkpoints_sorted = sorted(ordering_and_checkpoint_path)
         checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
-        # Make sure we don't delete the best model.
-        if self.state.best_model_checkpoint is not None:
-            best_model_index = checkpoints_sorted.index(self.state.best_model_checkpoint)
-            checkpoints_sorted[best_model_index], checkpoints_sorted[best_model_index][-1] = (
-                checkpoints_sorted[-1],
-                checkpoints_sorted[best_model_index],
-            )
         return checkpoints_sorted
 
     def _rotate_checkpoints(self, use_mtime=False) -> None:

src/transformers/trainer_utils.py

@@ -1,7 +1,4 @@
-import dataclasses
-import json
 import random
-from dataclasses import dataclass
 from typing import Any, Dict, List, NamedTuple, Optional, Tuple, Union
 
 import numpy as np
@@ -216,26 +213,3 @@ def distributed_broadcast_scalars(
             raise AssertionError("Not currently using distributed training")
     else:
         raise ImportError("Torch must be installed to use `distributed_broadcast_scalars`")
-
-
-@dataclass
-class TrainerState:
-    """
-    A class containing the `Trainer` fields that will be saved along the model and optimizer.
-    """
-
-    best_metric: Optional[float] = None
-    best_model_checkpoint: Optional[str] = None
-
-    def save_to_json(self, json_path: str):
-        """ Save the content of this instance in JSON format inside :obj:`json_path`."""
-        json_string = json.dumps(dataclasses.asdict(self), indent=2, sort_keys=True) + "\n"
-        with open(json_path, "w", encoding="utf-8") as f:
-            f.write(json_string)
-
-    @classmethod
-    def load_from_json(cls, json_path: str):
-        """ Create an instance from the content of :obj:`json_path`."""
-        with open(json_path, "r", encoding="utf-8") as f:
-            text = f.read()
-        return cls(**json.loads(text))

src/transformers/training_args.py

@@ -145,28 +145,6 @@ class TrainingArguments:
             Will eventually default to :obj:`["labels"]` except if the model used is one of the
             :obj:`XxxForQuestionAnswering` in which case it will default to
             :obj:`["start_positions", "end_positions"]`.
-        load_best_model_at_end (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether or not to load the best model found during training at the end of training.
-
-            .. note::
-
-                When set to :obj:`True`, the parameters :obj:`save_steps` will be ignored and the model will be saved
-                after each evaluation.
-        metric_for_best_model (:obj:`str`, `optional`)
-            Use in conjunction with :obj:`load_best_model_at_end` to specify the metric to use to compare two different
-            models. Must be the name of a metric returned by the evaluation with or without the prefix :obj:`"eval_"`.
-            Will default to :obj:`"loss"` if unspecified and :obj:`load_best_model_at_end=True` (to use the evaluation
-            loss).
-
-            If you set this value, :obj:`greater_is_better` will defaut to :obj:`True`. Don't forget to set it to
-            :obj:`False` if your metric is better when lower.
-        greater_is_better (:obj:`bool`, `optional`)
-            Use in conjunction with :obj:`load_best_model_at_end` and :obj:`metric_for_best_model` to specify if better
-            models should have a greater metric or not. Will default to:
-
-            - :obj:`True` if :obj:`metric_for_best_model` is set to a value that isn't :obj:`"loss"` or
-              :obj:`"eval_loss"`.
-            - :obj:`False` if :obj:`metric_for_best_model` is not set, or set to :obj:`"loss"` or :obj:`"eval_loss"`.
     """
 
     output_dir: str = field(
@@ -309,17 +287,6 @@ class TrainingArguments:
         default=None, metadata={"help": "The list of keys in your dictionary of inputs that correspond to the labels."}
     )
 
-    load_best_model_at_end: Optional[bool] = field(
-        default=False,
-        metadata={"help": "Whether or not to load the best model found during training at the end of training."},
-    )
-    metric_for_best_model: Optional[str] = field(
-        default=None, metadata={"help": "The metric to use to compare two different models."}
-    )
-    greater_is_better: Optional[bool] = field(
-        default=None, metadata={"help": "Whether the `metric_for_best_model` should be maximized or not."}
-    )
-
     def __post_init__(self):
         if self.disable_tqdm is None:
             self.disable_tqdm = logger.getEffectiveLevel() > logging.WARN
@@ -337,11 +304,6 @@ class TrainingArguments:
         if self.eval_steps is None:
             self.eval_steps = self.logging_steps
 
-        if self.load_best_model_at_end and self.metric_for_best_model is None:
-            self.metric_for_best_model = "loss"
-        if self.greater_is_better is None and self.metric_for_best_model is not None:
-            self.greater_is_better = self.metric_for_best_model not in ["loss", "eval_loss"]
-
     @property
     def train_batch_size(self) -> int:
         """

tests/test_modeling_common.py

@@ -111,6 +111,7 @@ class ModelTesterMixin:
             model.eval()
             with torch.no_grad():
                 outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
             out_2 = outputs[0].cpu().numpy()
             out_2[np.isnan(out_2)] = 0
 
@@ -151,6 +152,7 @@ class ModelTesterMixin:
             with torch.no_grad():
                 first = model(**self._prepare_for_class(inputs_dict, model_class))[0]
                 second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
+
             out_1 = first.cpu().numpy()
             out_2 = second.cpu().numpy()
             out_1 = out_1[~np.isnan(out_1)]
@@ -163,7 +165,7 @@ class ModelTesterMixin:
         seq_len = getattr(self.model_tester, "seq_length", None)
         decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
         encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
-        decoder_key_length = getattr(self.model_tester, "key_length", decoder_seq_length)
+        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
         encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
         chunk_length = getattr(self.model_tester, "chunk_length", None)
         if chunk_length is not None and hasattr(self.model_tester, "num_hashes"):
@@ -187,7 +189,7 @@ class ModelTesterMixin:
             model.to(torch_device)
             model.eval()
             with torch.no_grad():
-                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class), return_dict=True)
             attentions = outputs[-1]
             self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
 
@@ -215,6 +217,7 @@ class ModelTesterMixin:
                 if model_class in MODEL_FOR_QUESTION_ANSWERING_MAPPING.values():
                     correct_outlen += 1  # start_logits and end_logits instead of only 1 output
                     decoder_attention_idx += 1
+
                 self.assertEqual(out_len, correct_outlen)
 
                 decoder_attentions = outputs[decoder_attention_idx]
@@ -781,6 +784,10 @@ class ModelTesterMixin:
             model.eval()
 
             inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
+
+            with torch.no_grad():
+                hidden_states = model(**inputs)[0]
+
             if not self.is_encoder_decoder:
                 input_ids = inputs["input_ids"]
                 del inputs["input_ids"]
@@ -798,7 +805,9 @@ class ModelTesterMixin:
                 inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
 
             with torch.no_grad():
-                model(**inputs)
+                hidden_states_from_inputs_embeds = model(**inputs)[0]
+
+            self.assertTrue(torch.allclose(hidden_states, hidden_states_from_inputs_embeds, atol=1e-2))
 
     def test_lm_head_model_random_no_beam_search_generate(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

tests/test_modeling_gpt2.py

@@ -88,7 +88,7 @@ class GPT2ModelTester:
         self.bos_token_id = vocab_size - 1
         self.eos_token_id = vocab_size - 1
 
-    def prepare_config_and_inputs(self, gradient_checkpointing=False):
+    def prepare_config_and_inputs(self):
         input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
         input_mask = None
@@ -127,7 +127,6 @@ class GPT2ModelTester:
             bos_token_id=self.bos_token_id,
             eos_token_id=self.eos_token_id,
             return_dict=True,
-            gradient_checkpointing=gradient_checkpointing,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -270,15 +269,6 @@ class GPT2ModelTester:
         self.parent.assertEqual(result.loss.shape, ())
         self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
 
-    def create_and_check_forward_and_backwards(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
-        model = GPT2LMHeadModel(config)
-        model.to(torch_device)
-
-        result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
-        self.parent.assertEqual(result.loss.shape, ())
-        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
-        result.loss.backward()
-
     def create_and_check_double_lm_head_model(
         self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
     ):
@@ -365,10 +355,6 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_double_lm_head_model(*config_and_inputs)
 
-    def test_gpt2_gradient_checkpointing(self):
-        config_and_inputs = self.model_tester.prepare_config_and_inputs(gradient_checkpointing=True)
-        self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
-
     @slow
     def test_model_from_pretrained(self):
         for model_name in GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
@@ -380,34 +366,33 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
 class GPT2ModelLanguageGenerationTest(unittest.TestCase):
     @slow
     def test_lm_generate_gpt2(self):
-        for checkpointing in [True, False]:
-            model = GPT2LMHeadModel.from_pretrained("gpt2", gradient_checkpointing=checkpointing)
-            model.to(torch_device)
-            input_ids = torch.tensor([[464, 3290]], dtype=torch.long, device=torch_device)  # The dog
-            expected_output_ids = [
-                464,
-                3290,
-                373,
-                1043,
-                287,
-                257,
-                2214,
-                1474,
-                262,
-                16246,
-                286,
-                2688,
-                290,
-                2688,
-                27262,
-                13,
-                198,
-                198,
-                464,
-                3290,
-            ]  # The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
-            output_ids = model.generate(input_ids, do_sample=False)
-            self.assertListEqual(output_ids[0].tolist(), expected_output_ids)
+        model = GPT2LMHeadModel.from_pretrained("gpt2")
+        model.to(torch_device)
+        input_ids = torch.tensor([[464, 3290]], dtype=torch.long, device=torch_device)  # The dog
+        expected_output_ids = [
+            464,
+            3290,
+            373,
+            1043,
+            287,
+            257,
+            2214,
+            1474,
+            262,
+            16246,
+            286,
+            2688,
+            290,
+            2688,
+            27262,
+            13,
+            198,
+            198,
+            464,
+            3290,
+        ]  # The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
+        output_ids = model.generate(input_ids, do_sample=False)
+        self.assertListEqual(output_ids[0].tolist(), expected_output_ids)
 
     @slow
     def test_lm_generate_distilgpt2(self):

tests/test_modeling_prophetnet.py

@@ -0,0 +1,587 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team 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.
+
+
+import copy
+import tempfile
+import unittest
+
+from transformers import is_torch_available
+from transformers.file_utils import cached_property
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_common import ModelTesterMixin, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import ProphetNetConfig, ProphetNetForConditionalGeneration, ProphetNetModel, ProphetNetTokenizer
+
+
+class ProphetNetModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        hidden_size=16,
+        encoder_seq_length=7,
+        decoder_seq_length=9,
+        # For common tests
+        is_training=True,
+        use_attention_mask=True,
+        use_labels=True,
+        decoder_start_token_id=0,
+        encoder_ffn_dim=32,
+        num_encoder_layers=4,
+        num_encoder_attention_heads=4,
+        decoder_ffn_dim=32,
+        num_decoder_layers=4,
+        num_decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        ngram=1,
+        num_buckets=32,
+        relative_max_distance=128,
+        disable_ngram_loss=False,
+        scope=None,
+    ):
+
+        self.parent = parent
+        self.batch_size = batch_size
+        self.encoder_seq_length = encoder_seq_length
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_decoder_layers
+        self.num_encoder_layers = num_encoder_layers
+        self.num_decoder_layers = num_decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.num_attention_heads = num_decoder_attention_heads
+        self.num_encoder_attention_heads = num_encoder_attention_heads
+        self.num_decoder_attention_heads = num_decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.ngram = ngram
+        self.num_buckets = num_buckets
+        self.relative_max_distance = relative_max_distance
+        self.disable_ngram_loss = disable_ngram_loss
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = 2 * decoder_seq_length
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
+        decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        decoder_attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
+            decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = ProphetNetConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_encoder_layers=self.num_encoder_layers,
+            num_decoder_layers=self.num_decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_ffn_dim=self.encoder_ffn_dim,
+            num_encoder_attention_heads=self.num_encoder_attention_heads,
+            num_decoder_attention_heads=self.num_decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            ngram=self.ngram,
+            num_buckets=self.num_buckets,
+            relative_max_distance=self.relative_max_distance,
+            disable_ngram_loss=self.disable_ngram_loss,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        )
+
+    def check_prepare_lm_labels_via_shift_left(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetModel(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        # make sure that lm_labels are correctly padded from the right
+        lm_labels.masked_fill_((lm_labels == self.decoder_start_token_id), self.eos_token_id)
+
+        # add casaul pad token mask
+        triangular_mask = torch.tril(lm_labels.new_ones(lm_labels.shape)).logical_not()
+        lm_labels.masked_fill_(triangular_mask, self.pad_token_id)
+        decoder_input_ids = model._shift_right(lm_labels)
+
+        for i, (decoder_input_ids_slice, lm_labels_slice) in enumerate(zip(decoder_input_ids, lm_labels)):
+            # first item
+            self.parent.assertEqual(decoder_input_ids_slice[0].item(), self.decoder_start_token_id)
+            if i < decoder_input_ids_slice.shape[-1]:
+                if i < decoder_input_ids.shape[-1] - 1:
+                    # items before diagonal
+                    self.parent.assertListEqual(
+                        decoder_input_ids_slice[1 : i + 1].tolist(), lm_labels_slice[:i].tolist()
+                    )
+                # pad items after diagonal
+                if i < decoder_input_ids.shape[-1] - 2:
+                    self.parent.assertListEqual(
+                        decoder_input_ids_slice[i + 2 :].tolist(), lm_labels_slice[i + 1 : -1].tolist()
+                    )
+            else:
+                # all items after square
+                self.parent.assertListEqual(decoder_input_ids_slice[1:].tolist(), lm_labels_slice[:-1].tolist())
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
+        )
+        result = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+        decoder_output = result.last_hidden_state
+        decoder_past = result.past_key_values
+        encoder_output = result.encoder_last_hidden_state
+
+        self.parent.assertEqual(encoder_output.size(), (self.batch_size, self.encoder_seq_length, self.hidden_size))
+        self.parent.assertEqual(decoder_output.size(), (self.batch_size, self.decoder_seq_length, self.hidden_size))
+        # There should be `num_layers` key value embeddings stored in decoder_past
+        self.parent.assertEqual(len(decoder_past), config.num_layers)
+        # There should be a self attn key, a self attn value, a cross attn key and a cross attn value stored in each decoder_past tuple
+        self.parent.assertEqual(len(decoder_past[0]), 4)
+
+    def create_and_check_with_lm_head(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetForConditionalGeneration(config=config).to(torch_device).eval()
+        outputs = model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            labels=lm_labels,
+        )
+        self.parent.assertEqual(len(outputs), 4)
+        self.parent.assertEqual(outputs["logits"].size(), (self.batch_size, self.decoder_seq_length, self.vocab_size))
+        self.parent.assertEqual(outputs["loss"].size(), ())
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetModel(config=config).get_decoder().to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        output, past_key_value_states = outputs.to_tuple()
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_value_states=past_key_value_states)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetModel(config=config).get_decoder()
+        model.to(torch_device)
+        model.eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        output, past_key_value_states = model(input_ids, attention_mask=attn_mask, use_cache=True).to_tuple()
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # change a random masked slice from input_ids
+        random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
+        random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
+        input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
+
+        # append to next input_ids and attn_mask
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        attn_mask = torch.cat(
+            [attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_value_states=past_key_value_states, attention_mask=attn_mask)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    def create_and_check_generate_with_past_key_value_states(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetForConditionalGeneration(config=config).to(torch_device).eval()
+        torch.manual_seed(0)
+        output_without_past_cache = model.generate(
+            input_ids[:1], num_beams=2, max_length=5, do_sample=True, use_cache=False
+        )
+        torch.manual_seed(0)
+        output_with_past_cache = model.generate(input_ids[:1], num_beams=2, max_length=5, do_sample=True)
+        self.parent.assertTrue(torch.all(output_with_past_cache == output_without_past_cache))
+
+    def create_and_check_model_fp16_forward(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = ProphetNetModel(config=config).to(torch_device).half().eval()
+        output = model(input_ids, decoder_input_ids=input_ids, attention_mask=attention_mask)["last_hidden_state"]
+        self.parent.assertFalse(torch.isnan(output).any().item())
+
+    def create_and_check_encoder_decoder_shared_weights(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        for model_class in [ProphetNetModel, ProphetNetForConditionalGeneration]:
+            torch.manual_seed(0)
+            model = model_class(config=config).to(torch_device).eval()
+            # load state dict copies weights but does not tie them
+            model.encoder.load_state_dict(model.decoder.state_dict(), strict=False)
+
+            torch.manual_seed(0)
+            tied_config = copy.deepcopy(config)
+            tied_config.tie_encoder_decoder = True
+            tied_model = model_class(config=tied_config).to(torch_device).eval()
+
+            model_result = model(
+                input_ids=input_ids,
+                decoder_input_ids=decoder_input_ids,
+                attention_mask=attention_mask,
+                decoder_attention_mask=decoder_attention_mask,
+            )
+
+            tied_model_result = tied_model(
+                input_ids=input_ids,
+                decoder_input_ids=decoder_input_ids,
+                attention_mask=attention_mask,
+                decoder_attention_mask=decoder_attention_mask,
+            )
+
+            # check that models has less parameters
+            self.parent.assertLess(
+                sum(p.numel() for p in tied_model.parameters()), sum(p.numel() for p in model.parameters())
+            )
+            random_slice_idx = ids_tensor((1,), model_result[0].shape[-1]).item()
+
+            # check that outputs are equal
+            self.parent.assertTrue(
+                torch.allclose(
+                    model_result[0][0, :, random_slice_idx], tied_model_result[0][0, :, random_slice_idx], atol=1e-4
+                )
+            )
+
+            # check that outputs after saving and loading are equal
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                tied_model.save_pretrained(tmpdirname)
+                tied_model = model_class.from_pretrained(tmpdirname)
+                tied_model.to(torch_device)
+                tied_model.eval()
+
+                # check that models has less parameters
+                self.parent.assertLess(
+                    sum(p.numel() for p in tied_model.parameters()), sum(p.numel() for p in model.parameters())
+                )
+                random_slice_idx = ids_tensor((1,), model_result[0].shape[-1]).item()
+
+                tied_model_result = tied_model(
+                    input_ids=input_ids,
+                    decoder_input_ids=decoder_input_ids,
+                    attention_mask=attention_mask,
+                    decoder_attention_mask=decoder_attention_mask,
+                )
+
+                # check that outputs are equal
+                self.parent.assertTrue(
+                    torch.allclose(
+                        model_result[0][0, :, random_slice_idx],
+                        tied_model_result[0][0, :, random_slice_idx],
+                        atol=1e-4,
+                    )
+                )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "use_cache": False,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class ProphetNetModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (ProphetNetModel, ProphetNetForConditionalGeneration) if is_torch_available() else ()
+    all_generative_model_classes = (ProphetNetForConditionalGeneration,) if is_torch_available() else ()
+    test_pruning = False
+    test_torchscript = False
+    test_resize_embeddings = False
+    test_headmasking = False
+    is_encoder_decoder = True
+
+    def setUp(self):
+        self.model_tester = ProphetNetModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ProphetNetConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+
+class ProphetNetModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_pretrained_checkpoint_hidden_states(self):
+        model = ProphetNetForConditionalGeneration.from_pretrained("microsoft/prophetnet-large-uncased")
+        model.to(torch_device)
+
+        # encoder-decoder outputs
+        encoder_ids = torch.tensor(
+            [
+                [
+                    2871,
+                    102,
+                    2048,
+                    3176,
+                    2780,
+                    1997,
+                    2871,
+                    26727,
+                    2169,
+                    2097,
+                    12673,
+                    1996,
+                    8457,
+                    2006,
+                    2049,
+                    8240,
+                    2859,
+                    2799,
+                    1012,
+                    2023,
+                    6512,
+                    2038,
+                    2174,
+                    13977,
+                    2195,
+                    25962,
+                    1012,
+                    102,
+                ]
+            ]
+        ).to(torch_device)
+
+        decoder_prev_ids = torch.tensor([[102, 2129, 2116, 2372, 2024, 2006, 2169, 1997, 2122, 2048, 2780, 1029]]).to(
+            torch_device
+        )
+        output = model(
+            input_ids=encoder_ids,
+            attention_mask=None,
+            encoder_outputs=None,
+            decoder_input_ids=decoder_prev_ids,
+        )
+        output_predited_logis = output[0]
+        expected_shape = torch.Size((1, 12, 30522))
+        self.assertEqual(output_predited_logis.shape, expected_shape)
+        expected_slice = torch.tensor(
+            [[[-7.6213, -7.9008, -7.9979], [-7.6834, -7.8467, -8.2187], [-7.5326, -7.4762, -8.1914]]]
+        ).to(torch_device)
+        self.assertTrue(torch.allclose(output_predited_logis[:, :3, :3], expected_slice, atol=1e-4))
+
+        # encoder outputs
+        encoder_outputs = model.model.encoder(encoder_ids)[0]
+        expected_encoder_outputs_slice = torch.tensor(
+            [[[-0.2526, -0.1951, -0.2185], [-0.8923, 0.2992, -0.4623], [-0.4585, 0.0165, -0.6652]]]
+        ).to(torch_device)
+        expected_shape_encoder = torch.Size((1, 28, 1024))
+        self.assertEqual(encoder_outputs.shape, expected_shape_encoder)
+        self.assertTrue(torch.allclose(encoder_outputs[:, :3, :3], expected_encoder_outputs_slice, atol=1e-4))
+
+        # decoder outputs
+        decoder_outputs = model.model.decoder(
+            decoder_prev_ids, encoder_hidden_states=encoder_outputs, encoder_padding_mask=None
+        )
+        predicting_streams = decoder_outputs[0][:, 1:]
+        predicting_streams_logits = model.lm_head(predicting_streams)
+        next_first_stream_logits = predicting_streams_logits[:, 0]
+        self.assertTrue(torch.allclose(next_first_stream_logits[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_cnndm_inference(self):
+        model = ProphetNetForConditionalGeneration.from_pretrained("microsoft/prophetnet-large-uncased-cnndm")
+        model.to(torch_device)
+
+        tokenizer = ProphetNetTokenizer.from_pretrained("microsoft/prophetnet-large-uncased-cnndm")
+
+        ARTICLE_TO_SUMMARIZE = "USTC was founded in Beijing by the Chinese Academy of Sciences (CAS) in September 1958. The Director of CAS, Mr. Guo Moruo was appointed the first president of USTC. USTC's founding mission was to develop a high-level science and technology workforce, as deemed critical for development of China's economy, defense, and science and technology education. The establishment was hailed as \"A Major Event in the History of Chinese Education and Science.\" CAS has supported USTC by combining most of its institutes with the departments of the university. USTC is listed in the top 16 national key universities, becoming the youngest national key university.".lower()
+        input_ids = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=511, return_tensors="pt").input_ids
+
+        input_ids = input_ids.to(torch_device)
+
+        summary_ids = model.generate(
+            input_ids, num_beams=4, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        EXPECTED_SUMMARIZE_512 = "us ##tc was founded by the chinese academy of sciences ( cas ) in 1958 . [X_SEP] us ##tc is listed in the top 16 national key universities ."
+        generated_titles = [
+            " ".join(tokenizer.convert_ids_to_tokens(g, skip_special_tokens=True)) for g in summary_ids
+        ]
+        self.assertListEqual(
+            [EXPECTED_SUMMARIZE_512],
+            generated_titles,
+        )
+        input_ids = tokenizer([ARTICLE_TO_SUMMARIZE], max_length=99, return_tensors="pt").input_ids
+        input_ids = input_ids.to(torch_device)
+        # actually 98 tokens are used. max_length=100 contains bos and eos.
+        # print(' '.join(tokenizer.tokenize(ARTICLE_TO_SUMMARIZE)[:98]))
+        summary_ids = model.generate(
+            input_ids, num_beams=4, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        EXPECTED_SUMMARIZE_100 = (
+            r"us ##tc was founded in beijing by the chinese academy of sciences ( cas ) in 1958 . [X_SEP] us ##tc "
+            "'"
+            ' s founding mission was to develop a high - level science and technology workforce . [X_SEP] establishment hailed as " a major event in the history of chinese education and science "'
+        )
+        generated_titles = [
+            " ".join(tokenizer.convert_ids_to_tokens(g, skip_special_tokens=True)) for g in summary_ids
+        ]
+        self.assertListEqual(
+            [EXPECTED_SUMMARIZE_100],
+            generated_titles,
+        )

tests/test_modeling_t5.py

@@ -44,7 +44,6 @@ class T5ModelTester:
         encoder_seq_length=7,
         decoder_seq_length=9,
         # For common tests
-        seq_length=7,
         is_training=True,
         use_attention_mask=True,
         use_labels=True,

tests/test_modeling_xlm_prophetnet.py

@@ -0,0 +1,139 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team 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.
+
+
+import unittest
+
+from transformers import is_torch_available
+from transformers.testing_utils import slow, torch_device
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import XLMProphetNetForConditionalGeneration, XLMProphetNetTokenizer
+
+
+class XLMProphetNetModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_pretrained_checkpoint_hidden_states(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained("microsoft/xprophetnet-large-wiki100-cased")
+        model.to(torch_device)
+
+        # encoder-decoder outputs
+        encoder_ids = torch.tensor([[17, 96208, 103471, 2]]).to(torch_device)
+        decoder_prev_ids = torch.tensor(
+            [[2, 250, 9953, 34, 69489, 1620, 32, 118424, 624, 210, 105, 2913, 1032, 351]]
+        ).to(torch_device)
+        output = model(
+            input_ids=encoder_ids, attention_mask=None, encoder_outputs=None, decoder_input_ids=decoder_prev_ids
+        )
+        output_predited_logis = output[0]
+        expected_shape = torch.Size((1, 14, 250012))
+        self.assertEqual(output_predited_logis.shape, expected_shape)
+        expected_slice = torch.tensor(
+            [[[-6.6042, -8.3838, 12.4717], [-6.4426, -8.1994, 12.4542], [-6.0851, -7.8209, 12.9493]]]
+        ).to(torch_device)
+        self.assertTrue(torch.allclose(output_predited_logis[:, :3, :3], expected_slice, atol=1e-4))
+
+        # encoder outputs
+        encoder_outputs = model.model.encoder(encoder_ids)[0]
+        expected_encoder_outputs_slice = torch.tensor(
+            [[[-1.4260, -0.7628, 0.8453], [-1.4719, -0.1391, 0.7807], [-1.7678, 0.0114, 0.4646]]]
+        ).to(torch_device)
+        expected_shape_encoder = torch.Size((1, 4, 1024))
+        self.assertEqual(encoder_outputs.shape, expected_shape_encoder)
+        self.assertTrue(torch.allclose(encoder_outputs[:, :3, :3], expected_encoder_outputs_slice, atol=1e-4))
+
+        # decoder outputs
+        decoder_outputs = model.model.decoder(
+            decoder_prev_ids, encoder_hidden_states=encoder_outputs, encoder_padding_mask=None
+        )
+        predicting_streams = decoder_outputs[0][:, 1:]
+        predicting_streams_logits = model.lm_head(predicting_streams)
+        next_first_stream_logits = predicting_streams_logits[:, 0]
+        self.assertTrue(torch.allclose(next_first_stream_logits[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_ntg_hidden_states(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained(
+            "microsoft/xprophetnet-large-wiki100-cased-xglue-ntg"
+        )
+        model.to(torch_device)
+
+        encoder_ids = torch.tensor([[17, 96208, 103471, 2]]).to(torch_device)
+        decoder_prev_ids = torch.tensor(
+            [[2, 250, 9953, 34, 69489, 1620, 32, 118424, 624, 210, 105, 2913, 1032, 351]]
+        ).to(torch_device)
+        output = model(
+            input_ids=encoder_ids, attention_mask=None, encoder_outputs=None, decoder_input_ids=decoder_prev_ids
+        )
+        output_predited_logis = output[0]
+        expected_shape = torch.Size((1, 14, 250012))
+        self.assertEqual(output_predited_logis.shape, expected_shape)
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [[[-8.8815, -9.2996, -4.4506], [-6.7202, -7.8944, -0.9402], [-8.6890, -7.4528, -1.9437]]]
+        ).to(torch_device)
+
+        self.assertTrue(torch.allclose(output_predited_logis[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_xprophetnet_ntg_inference(self):
+        model = XLMProphetNetForConditionalGeneration.from_pretrained(
+            "microsoft/xprophetnet-large-wiki100-cased-xglue-ntg"
+        )
+        model.to(torch_device)
+
+        tokenizer = XLMProphetNetTokenizer.from_pretrained("microsoft/xprophetnet-large-wiki100-cased-xglue-ntg")
+
+        EN_SENTENCE = "Microsoft Corporation intends to officially end free support for the Windows 7 operating system after January 14, 2020, according to the official portal of the organization. From that day, users of this system will not be able to receive security updates, which could make their computers vulnerable to cyber attacks."
+        RU_SENTENCE = "орпорация Microsoft намерена официально прекратить бесплатную поддержку операционной системы Windows 7 после 14 января 2020 года, сообщается на официальном портале организации . С указанного дня пользователи этой системы не смогут получать обновления безопасности, из-за чего их компьютеры могут стать уязвимыми к кибератакам."
+        ZH_SENTENCE = (
+            "根据该组织的官方门户网站，微软公司打算在2020年1月14日之后正式终止对Windows 7操作系统的免费支持。从那时起，该系统的用户将无法接收安全更新，这可能会使他们的计算机容易受到网络攻击。"
+        )
+
+        input_ids = tokenizer(
+            [EN_SENTENCE, RU_SENTENCE, ZH_SENTENCE], padding=True, max_length=255, return_tensors="pt"
+        ).input_ids
+        input_ids = input_ids.to(torch_device)
+
+        summary_ids = model.generate(
+            input_ids, num_beams=10, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        generated_titles = [tokenizer.decode(g, skip_special_tokens=True) for g in summary_ids]
+        EXPECTED_TITLE_EN = "Microsoft to end Windows 7 free support after January 14, 2020"
+        EXPECTED_TITLE_RU = "Microsoft намерена прекратить бесплатную поддержку Windows 7 после 14 января 2020 года"
+        EXPECTED_TITLE_ZH = "微软打算终止对Windows 7操作系统的免费支持"
+        self.assertListEqual(
+            [EXPECTED_TITLE_EN, EXPECTED_TITLE_RU, EXPECTED_TITLE_ZH],
+            generated_titles,
+        )
+
+        summary_ids_beam1 = model.generate(
+            input_ids, num_beams=1, length_penalty=1.0, no_repeat_ngram_size=3, early_stopping=True
+        )
+        generated_titles_beam1_tok = [
+            tokenizer.convert_ids_to_tokens(g, skip_special_tokens=True) for g in summary_ids_beam1
+        ]
+        EXPECTED_TITLE_EN_BEAM1_TOK = "▁Microsoft ▁to ▁end ▁free ▁support ▁for ▁Windows ▁7".split(" ")
+        EXPECTED_TITLE_RU_BEAM1_TOK = "▁Microsoft ▁намерен а ▁прекрати ть ▁бес плат ную ▁поддержку ▁Windows ▁7 ▁после ▁14 ▁января ▁2020 ▁года".split(
+            " "
+        )
+        EXPECTED_TITLE_ZH_BEAM1_TOK = "微软 公司 打算 终止 对 Windows ▁7 操作 系统的 免费 支持".split(" ")
+        self.assertListEqual(
+            [EXPECTED_TITLE_EN_BEAM1_TOK, EXPECTED_TITLE_RU_BEAM1_TOK, EXPECTED_TITLE_ZH_BEAM1_TOK],
+            generated_titles_beam1_tok,
+        )

tests/test_trainer.py

@@ -1,13 +1,9 @@
-import json
-import os
-import tempfile
 import unittest
 
 import datasets
 import numpy as np
 
-from transformers import AutoTokenizer, PretrainedConfig, TrainingArguments, is_torch_available
-from transformers.file_utils import WEIGHTS_NAME
+from transformers import AutoTokenizer, TrainingArguments, is_torch_available
 from transformers.testing_utils import get_tests_dir, require_torch, slow
 
 
@@ -20,7 +16,6 @@ if is_torch_available():
         GlueDataset,
         GlueDataTrainingArguments,
         LineByLineTextDataset,
-        PreTrainedModel,
         Trainer,
     )
 
@@ -56,14 +51,6 @@ class AlmostAccuracy:
         return {"accuracy": true.astype(np.float32).mean().item()}
 
 
-class RegressionModelConfig(PretrainedConfig):
-    def __init__(self, a=0, b=0, double_output=False, **kwargs):
-        super().__init__(**kwargs)
-        self.a = a
-        self.b = b
-        self.double_output = double_output
-
-
 if is_torch_available():
 
     class SampleIterableDataset(IterableDataset):
@@ -92,37 +79,15 @@ if is_torch_available():
             loss = torch.nn.functional.mse_loss(y, labels)
             return (loss, y, y) if self.double_output else (loss, y)
 
-    class RegressionPreTrainedModel(PreTrainedModel):
-        config_class = RegressionModelConfig
-        base_model_prefix = "regression"
-
-        def __init__(self, config):
-            super().__init__(config)
-            self.a = torch.nn.Parameter(torch.tensor(config.a).float())
-            self.b = torch.nn.Parameter(torch.tensor(config.b).float())
-            self.double_output = config.double_output
-
-        def forward(self, input_x=None, labels=None, **kwargs):
-            y = input_x * self.a + self.b
-            if labels is None:
-                return (y, y) if self.double_output else (y,)
-            loss = torch.nn.functional.mse_loss(y, labels)
-            return (loss, y, y) if self.double_output else (loss, y)
-
-    def get_regression_trainer(a=0, b=0, double_output=False, train_len=64, eval_len=64, pretrained=True, **kwargs):
+    def get_regression_trainer(a=0, b=0, double_output=False, train_len=64, eval_len=64, **kwargs):
         label_names = kwargs.get("label_names", None)
         train_dataset = RegressionDataset(length=train_len, label_names=label_names)
         eval_dataset = RegressionDataset(length=eval_len, label_names=label_names)
-        if pretrained:
-            config = RegressionModelConfig(a=a, b=b, double_output=double_output)
-            model = RegressionPreTrainedModel(config)
-        else:
-            model = RegressionModel(a=a, b=b, double_output=double_output)
+        model = RegressionModel(a, b, double_output)
         compute_metrics = kwargs.pop("compute_metrics", None)
         data_collator = kwargs.pop("data_collator", None)
         optimizers = kwargs.pop("optimizers", (None, None))
-        output_dir = kwargs.pop("output_dir", "./regression")
-        args = TrainingArguments(output_dir, **kwargs)
+        args = TrainingArguments("./regression", **kwargs)
         return Trainer(
             model,
             args,
@@ -154,40 +119,6 @@ class TrainerIntegrationTest(unittest.TestCase):
         self.assertTrue(torch.allclose(model.a, a))
         self.assertTrue(torch.allclose(model.b, b))
 
-    def check_saved_checkpoints(self, output_dir, freq, total, is_pretrained=True):
-        file_list = [WEIGHTS_NAME, "training_args.bin", "log_history.json", "optimizer.pt", "scheduler.pt"]
-        if is_pretrained:
-            file_list.append("config.json")
-        for step in range(freq, total, freq):
-            checkpoint = os.path.join(output_dir, f"checkpoint-{step}")
-            self.assertTrue(os.path.isdir(checkpoint))
-            for filename in file_list:
-                self.assertTrue(os.path.isfile(os.path.join(checkpoint, filename)))
-
-    def check_best_model_has_been_loaded(
-        self, output_dir, freq, total, trainer, metric, greater_is_better=False, is_pretrained=True
-    ):
-        checkpoint = os.path.join(output_dir, f"checkpoint-{(total // freq) * freq}")
-        log_history = json.load(open(os.path.join(checkpoint, "log_history.json")))
-
-        values = [d[metric] for d in log_history]
-        best_value = max(values) if greater_is_better else min(values)
-        best_checkpoint = (values.index(best_value) + 1) * freq
-        checkpoint = os.path.join(output_dir, f"checkpoint-{best_checkpoint}")
-        if is_pretrained:
-            best_model = RegressionPreTrainedModel.from_pretrained(checkpoint)
-            best_model.to(trainer.args.device)
-        else:
-            best_model = RegressionModel()
-            state_dict = torch.load(os.path.join(checkpoint, WEIGHTS_NAME))
-            best_model.load_state_dict(state_dict)
-            best_model.to(trainer.args.device)
-        self.assertTrue(torch.allclose(best_model.a, trainer.model.a))
-        self.assertTrue(torch.allclose(best_model.b, trainer.model.b))
-
-        metrics = trainer.evaluate()
-        self.assertEqual(metrics[metric], best_value)
-
     def test_reproducible_training(self):
         # Checks that training worked, model trained and seed made a reproducible training.
         trainer = get_regression_trainer(learning_rate=0.1)
@@ -356,86 +287,6 @@ class TrainerIntegrationTest(unittest.TestCase):
         trainer.train()
         self.check_trained_model(trainer.model, alternate_seed=True)
 
-    def test_save_checkpoints(self):
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 5, int(self.n_epochs * 64 / self.batch_size))
-
-        # With a regular model that is not a PreTrainedModel
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(output_dir=tmpdir, save_steps=5, pretrained=False)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 5, int(self.n_epochs * 64 / self.batch_size), False)
-
-    def test_load_best_model_at_end(self):
-        total = int(self.n_epochs * 64 / self.batch_size)
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(
-                a=1.5,
-                b=2.5,
-                output_dir=tmpdir,
-                learning_rate=0.1,
-                eval_steps=5,
-                evaluation_strategy="steps",
-                load_best_model_at_end=True,
-            )
-            self.assertFalse(trainer.args.greater_is_better)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 5, total)
-            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_loss")
-
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(
-                a=1.5,
-                b=2.5,
-                output_dir=tmpdir,
-                learning_rate=0.1,
-                eval_steps=5,
-                evaluation_strategy="steps",
-                load_best_model_at_end=True,
-                metric_for_best_model="accuracy",
-                compute_metrics=AlmostAccuracy(),
-            )
-            self.assertTrue(trainer.args.greater_is_better)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 5, total)
-            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_accuracy", greater_is_better=True)
-
-        # Save is done every eval regardless of the strategy
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(
-                a=1.5,
-                b=2.5,
-                output_dir=tmpdir,
-                learning_rate=0.1,
-                evaluation_strategy="epoch",
-                load_best_model_at_end=True,
-                metric_for_best_model="accuracy",
-                compute_metrics=AlmostAccuracy(),
-            )
-            self.assertTrue(trainer.args.greater_is_better)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 64 // self.batch_size, total)
-            self.check_best_model_has_been_loaded(
-                tmpdir, 64 // self.batch_size, total, trainer, "eval_accuracy", greater_is_better=True
-            )
-
-        # Test this works with a non PreTrainedModel
-        with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = get_regression_trainer(
-                output_dir=tmpdir,
-                learning_rate=0.1,
-                eval_steps=5,
-                evaluation_strategy="steps",
-                load_best_model_at_end=True,
-                pretrained=False,
-            )
-            self.assertFalse(trainer.args.greater_is_better)
-            trainer.train()
-            self.check_saved_checkpoints(tmpdir, 5, total, is_pretrained=False)
-            self.check_best_model_has_been_loaded(tmpdir, 5, total, trainer, "eval_loss", is_pretrained=False)
-
     @slow
     def test_trainer_eval_mrpc(self):
         MODEL_ID = "bert-base-cased-finetuned-mrpc"

utils/check_repo.py

@@ -314,8 +314,9 @@ def check_repo_quality():
     print("Checking all models are properly tested.")
     check_all_decorator_order()
     check_all_models_are_tested()
-    print("Checking all models are properly documented.")
-    check_all_models_are_documented()
+    # Uncomment me when RAG is back
+    # print("Checking all models are properly documented.")
+    # check_all_models_are_documented()
 
 
 if __name__ == "__main__":