fix ==

* fix beam_search behavior when sampling

* delete print

* make correct style

.circleci/config.yml

@@ -14,23 +14,6 @@ jobs:
             - run: sudo pip install codecov pytest-cov
             - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
             - run: codecov
-    run_all_tests_torch_and_tf:
-        working_directory: ~/transformers
-        docker:
-            - image: circleci/python:3.5
-        environment:
-            OMP_NUM_THREADS: 1
-            RUN_SLOW: yes
-            RUN_CUSTOM_TOKENIZERS: yes
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - run: sudo pip install .[mecab,sklearn,tf-cpu,torch,testing]
-            - run:
-                command: python -m pytest -n 8 --dist=loadfile -s -v ./tests/
-                no_output_timeout: 4h
-
     run_tests_torch:
         working_directory: ~/transformers
         docker:
@@ -45,6 +28,21 @@ jobs:
             - run: sudo pip install codecov pytest-cov
             - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
             - run: codecov
+    run_tests_legacy_torch:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install torch==1.0.0
+            - run: sudo pip install .[sklearn,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
+            - run: codecov
     run_tests_tf:
         working_directory: ~/transformers
         docker:
@@ -131,16 +129,6 @@ workflows:
             - run_examples_torch
             - run_tests_custom_tokenizers
             - run_tests_torch_and_tf
-            - run_tests_torch
+            - run_tests_legacy_torch
             - run_tests_tf
             - deploy_doc: *workflow_filters
-    run_slow_tests:
-        triggers:
-            - schedule:
-                cron: "0 4 * * *"
-                filters:
-                    branches:
-                        only:
-                            - master
-        jobs:
-            - run_all_tests_torch_and_tf

.github/workflows/github-push.yml

@@ -0,0 +1,19 @@
+name: GitHub-hosted runner
+
+on: push
+
+jobs:
+  check_code_quality:
+    runs-on: ubuntu-18.04
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v1
+      with:
+        python-version: 3.7
+    - name: Install dependencies
+      run: |
+        pip install .[tf,torch,quality]
+
+
+

.github/workflows/self-push.yml

@@ -0,0 +1,49 @@
+name: Self-hosted runner (push)
+
+on: 
+  push:
+    branches:
+      - master
+  pull_request:
+
+
+jobs:
+  run_tests_torch_and_tf_gpu:
+    runs-on: self-hosted
+    steps:
+    - uses: actions/checkout@v2
+    - name: Python version
+      run: |
+        which python
+        python --version
+        pip --version
+    - name: Current dir
+      run: pwd
+    - run: nvidia-smi
+    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+      run: |
+        python -m venv .env
+        source .env/bin/activate
+        which python
+        python --version
+        pip --version
+    - name: Install dependencies
+      run: |
+        source .env/bin/activate
+        pip install .[sklearn,tf,torch,testing]
+
+    - name: Are GPUs recognized by our DL frameworks
+      run: |
+        source .env/bin/activate
+        python -c "import torch; print(torch.cuda.is_available())"
+        python -c "import tensorflow as tf; print(tf.test.is_built_with_cuda(), tf.config.list_physical_devices('GPU'))"
+
+    - name: Run all non-slow tests on GPU
+      env:
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
+        # TF_GPU_MEMORY_LIMIT: 4096
+        OMP_NUM_THREADS: 1
+        USE_CUDA: yes
+      run: |
+        source .env/bin/activate
+        python -m pytest -n 2 --dist=loadfile -s -v ./tests/

.github/workflows/self-scheduled.yml

@@ -0,0 +1,51 @@
+name: Self-hosted runner (scheduled)
+
+on:
+  push:
+    branches:
+      - ci_*
+  repository_dispatch:
+  schedule:
+    - cron: "0 0 * * *"
+
+jobs:
+  run_all_tests_torch_and_tf_gpu:
+    runs-on: self-hosted
+    steps:
+    - uses: actions/checkout@v2
+    - name: Python version
+      run: |
+        which python
+        python --version
+        pip --version
+    - name: Current dir
+      run: pwd
+    - run: nvidia-smi
+    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+      run: |
+        python -m venv .env
+        source .env/bin/activate
+        which python
+        python --version
+        pip --version
+    - name: Install dependencies
+      run: |
+        source .env/bin/activate
+        pip install .[sklearn,tf,torch,testing]
+
+    - name: Are GPUs recognized by our DL frameworks
+      run: |
+        source .env/bin/activate
+        python -c "import torch; print(torch.cuda.is_available())"
+        python -c "import tensorflow as tf; print(tf.test.is_built_with_cuda(), tf.config.list_physical_devices('GPU'))"
+
+    - name: Run all tests on GPU
+      env:
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
+        OMP_NUM_THREADS: 1
+        RUN_SLOW: yes
+        USE_CUDA: yes
+      run: |
+        source .env/bin/activate
+        python -m pytest -n 1 --dist=loadfile -s -v ./tests/
+        

docs/source/index.rst

@@ -80,6 +80,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     main_classes/configuration
     main_classes/model
     main_classes/tokenizer
+    main_classes/pipelines
     main_classes/optimizer_schedules
     main_classes/processors
 

docs/source/main_classes/pipelines.rst

@@ -0,0 +1,63 @@
+Pipelines
+----------------------------------------------------
+
+The pipelines are a great and easy way to use models for inference. These pipelines are objects that abstract most
+of the complex code from the library, offering a simple API dedicated to several tasks, including Named Entity
+Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering.
+
+There are two categories of pipeline abstractions to be aware about:
+
+- The :class:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines
+- The other task-specific pipelines, such as :class:`~transformers.NerPipeline`
+  or :class:`~transformers.QuestionAnsweringPipeline`
+
+The pipeline abstraction
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The `pipeline` abstraction is a wrapper around all the other available pipelines. It is instantiated as any
+other pipeline but requires an additional argument which is the `task`.
+
+.. autoclass:: transformers.pipeline
+    :members:
+
+
+The task specific pipelines
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Parent class: Pipeline
+=========================================
+
+.. autoclass:: transformers.Pipeline
+    :members: predict, transform, save_pretrained
+
+NerPipeline
+==========================================
+
+.. autoclass:: transformers.NerPipeline
+
+TokenClassificationPipeline
+==========================================
+
+This class is an alias of the :class:`~transformers.NerPipeline` defined above. Please refer to that pipeline for
+documentation and usage examples.
+
+FillMaskPipeline
+==========================================
+
+.. autoclass:: transformers.FillMaskPipeline
+
+FeatureExtractionPipeline
+==========================================
+
+.. autoclass:: transformers.FeatureExtractionPipeline
+
+TextClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.TextClassificationPipeline
+
+QuestionAnsweringPipeline
+==========================================
+
+.. autoclass:: transformers.QuestionAnsweringPipeline
+

docs/source/model_doc/bart.rst

@@ -4,20 +4,27 @@ Bart
 file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
 @sshleifer
 
-The Bart model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, Luke Zettlemoyer on 29 Oct, 2019.
-It is a sequence to sequence model where both encoder and decoder are transformers. The paper also introduces a novel pretraining objective, and demonstrates excellent summarization results.
-The authors released their code `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_
+Paper
+~~~~~
+The Bart model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer on 29 Oct, 2019.
+According to the abstract:
 
-**Abstract:**
+- Bart uses a standard seq2seq/machine translation architecture with a bidirectional encoder (like BERT) and a left-to-right decoder (like GPT).
+- The pretraining task involves randomly shuffling the order of the original sentences and a novel in-filling scheme, where spans of text are replaced with a single mask token.
+- BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE.
 
-*We present BART, a denoising autoencoder for pretraining sequence-to-sequence models. BART is trained by (1) corrupting text with an arbitrary noising function, and (2) learning a model to reconstruct the original text. It uses a standard Tranformer-based neural machine translation architecture which, despite its simplicity, can be seen as generalizing BERT (due to the bidirectional encoder), GPT (with the left-to-right decoder), and many other more recent pretraining schemes. We evaluate a number of noising approaches, finding the best performance by both randomly shuffling the order of the original sentences and using a novel in-filling scheme, where spans of text are replaced with a single mask token. BART is particularly effective when fine tuned for text generation but also works well for comprehension tasks. It matches the performance of RoBERTa with comparable training resources on GLUE and SQuAD, achieves new state-of-the-art results on a range of abstractive dialogue, question answering, and summarization tasks, with gains of up to 6 ROUGE. BART also provides a 1.1 BLEU increase over a back-translation system for machine translation, with only target language pretraining. We also report ablation experiments that replicate other pretraining schemes within the BART framework, to better measure which factors most influence end-task performance.*
-`BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension`
+The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_
 
 
-Notes:
+Implementation Notes
+~~~~~~~~~~~~~~~~~~~~
 - Bart doesn't use :obj:`token_type_ids`, for sequence classification just use BartTokenizer.encode to get the proper splitting.
 - Inputs to the decoder are created by BartModel.forward if they are not passed. This is different than some other model APIs.
 - Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to fairseq.encode starts with a space.
+- Decoder inputs are created automatically by the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``
+BartModel
+- ``MaskedLM.generate`` should be used for summarization, see the example in that docstrings
+
 
 BartModel
 ~~~~~~~~~~~~~~~~~~~~
@@ -30,7 +37,7 @@ BartForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.BartForMaskedLM
-    :members: forward
+    :members: forward, generate
 
 
 BartForSequenceClassification

docs/source/pretrained_models.rst

@@ -280,7 +280,10 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_)                                                       |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``bart-large-mnli``                                        | | Adds a 2 layer classification head with 1 million parameters                                                                        |
-|                   |                                                            | | bart-large base architecture with a classification head                                                                             |
+|                   |                                                            | | bart-large base architecture with a classification head, finetuned on MNLI                                                          |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bart-large-cnn``                                         | | 12-layer, 1024-hidden, 16-heads, 406M parameters       (same as base)                                                               |
+|                   |                                                            | | bart-large base architecture finetuned on cnn summarization task                                                                    |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 
 

docs/source/quickstart.md

@@ -220,96 +220,3 @@ print(sequence)
 ```
 
 The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
-
-### Model2Model example
-
-Encoder-decoder architectures require two tokenized inputs: one for the encoder and the other one for the decoder. Let's assume that we want to use `Model2Model` for generative question answering, and start by tokenizing the question and answer that will be fed to the model.
-
-```python
-import torch
-from transformers import BertTokenizer, Model2Model
-
-# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
-import logging
-logging.basicConfig(level=logging.INFO)
-
-# Load pre-trained model tokenizer (vocabulary)
-tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-
-# Encode the input to the encoder (the question)
-question = "Who was Jim Henson?"
-encoded_question = tokenizer.encode(question)
-
-# Encode the input to the decoder (the answer)
-answer = "Jim Henson was a puppeteer"
-encoded_answer = tokenizer.encode(answer)
-
-# Convert inputs to PyTorch tensors
-question_tensor = torch.tensor([encoded_question])
-answer_tensor = torch.tensor([encoded_answer])
-```
-
-Let's see how we can use `Model2Model` to get the value of the loss associated with this (question, answer) pair:
-
-```python
-# In order to compute the loss we need to provide language model
-# labels (the token ids that the model should have produced) to
-# the decoder.
-lm_labels =  encoded_answer
-labels_tensor = torch.tensor([lm_labels])
-
-# Load pre-trained model (weights)
-model = Model2Model.from_pretrained('bert-base-uncased')
-
-# Set the model in evaluation mode to deactivate the DropOut modules
-# This is IMPORTANT to have reproducible results during evaluation!
-model.eval()
-
-# If you have a GPU, put everything on cuda
-question_tensor = question_tensor.to('cuda')
-answer_tensor = answer_tensor.to('cuda')
-labels_tensor = labels_tensor.to('cuda')
-model.to('cuda')
-
-# Predict hidden states features for each layer
-with torch.no_grad():
-    # See the models docstrings for the detail of the inputs
-    outputs = model(question_tensor, answer_tensor, decoder_lm_labels=labels_tensor)
-    # Transformers models always output tuples.
-    # See the models docstrings for the detail of all the outputs
-    # In our case, the first element is the value of the LM loss 
-    lm_loss = outputs[0]
-```
-
-This loss can be used to fine-tune `Model2Model` on the question answering task. Assuming that we fine-tuned the model, let us now see how to generate an answer:
-
-```python
-# Let's re-use the previous question
-question = "Who was Jim Henson?"
-encoded_question = tokenizer.encode(question)
-question_tensor = torch.tensor([encoded_question])
-
-# This time we try to generate the answer, so we start with an empty sequence
-answer = "[CLS]"
-encoded_answer = tokenizer.encode(answer, add_special_tokens=False)
-answer_tensor = torch.tensor([encoded_answer])
-
-# Load pre-trained model (weights)
-model = Model2Model.from_pretrained('fine-tuned-weights')
-model.eval()
-
-# If you have a GPU, put everything on cuda
-question_tensor = question_tensor.to('cuda')
-answer_tensor = answer_tensor.to('cuda')
-model.to('cuda')
-
-# Predict all tokens
-with torch.no_grad():
-    outputs = model(question_tensor, answer_tensor)
-    predictions = outputs[0]
-
-# confirm we were able to predict 'jim'
-predicted_index = torch.argmax(predictions[0, -1]).item()
-predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
-assert predicted_token == 'jim'
-```

examples/contrib/run_swag.py

@@ -622,7 +622,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/distillation/README.md

@@ -10,14 +10,14 @@ This folder contains the original code used to train Distil* as well as examples
 
 **October 23, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
 
-**October 3, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
+**October 3, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper supersedes our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
 
 **September 19, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 99% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 
 
 ## What is Distil*
 
-Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 99% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
+Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
 
 We have applied the same method to other Transformer architectures and released the weights:
 - GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 16.3 compared to 21.1 for **DistilGPT2** (after fine-tuning on the train set).
@@ -31,15 +31,15 @@ Here are the results on the dev sets of GLUE:
 
 | Model                     | Macro-score                    | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI              |
 | :---:                     |    :---:                       | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:             |
-| BERT-base-uncased         |  **74.9**                      | 49.2 | 80.8 | 87.4 | 87.5 | 86.4 | 61.7 | 92.0 | 83.8 | 45.1              |
-| DistilBERT-base-uncased   |  **74.3**                      | 43.6 | 79.0 | 87.5 | 85.3 | 84.9 | 59.9 | 90.7 | 81.2 | 56.3              |
+| BERT-base-uncased         |  **79.5**                      | 56.3 | 84.7 | 88.6 | 91.8 | 89.6 | 69.3 | 92.7 | 89.0 | 53.5              |
+| DistilBERT-base-uncased   |  **77.0**                      | 51.3 | 82.1 | 87.5 | 89.2 | 88.5 | 59.9 | 91.3 | 86.9 | 56.3              |
 | BERT-base-cased           |  **78.2**                      | 58.2 | 83.9 | 87.8 | 91.0 | 89.2 | 66.1 | 91.7 | 89.2 | 46.5              |
 | DistilBERT-base-cased     |  **75.9**                      | 47.2 | 81.5 | 85.6 | 88.2 | 87.8 | 60.6 | 90.4 | 85.5 | 56.3              |
 | ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
 | RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup>  |
 | DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.3 | 84.0 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
 
-<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
+<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directly perform transfer learning on the pre-trained DistilRoBERTa.
 
 <sup>2</sup> Macro-score computed without WNLI.
 
@@ -65,9 +65,9 @@ This part of the library has only be tested with Python3.6+. There are few speci
 Transformers includes five pre-trained Distil* models, currently only provided for English and German (we are investigating the possibility to train and release a multilingual version of DistilBERT):
 
 - `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
-- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 79.8 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 82.3 F1 score).
+- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knowledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
 - `distilbert-base-cased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-cased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 65M parameters.
-- `distilbert-base-cased-distilled-squad`: A finetuned version of `distilbert-base-cased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 87.1 on the dev set (for comparison, Bert `bert-base-cased` version reaches a 88.7 F1 score).
+- `distilbert-base-cased-distilled-squad`: A finetuned version of `distilbert-base-cased` finetuned using (a second step of) knowledge distillation on SQuAD 1.0. This model reaches a F1 score of 87.1 on the dev set (for comparison, Bert `bert-base-cased` version reaches a 88.7 F1 score).
 - `distilbert-base-german-cased`: DistilBERT German language model pretrained on 1/2 of the data used to pretrain Bert using distillation with the supervision of the `bert-base-german-dbmdz-cased` version of German DBMDZ Bert. For NER tasks the model reaches a F1 score of 83.49 on the CoNLL-2003 test set (for comparison, `bert-base-german-dbmdz-cased` reaches a 84.52 F1 score), and a F1 score of 85.23 on the GermEval 2014 test set (`bert-base-german-dbmdz-cased` reaches a 86.89 F1 score).
 - `distilgpt2`: DistilGPT2 English language model pretrained with the supervision of `gpt2` (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
 - `distilroberta-base`: DistilRoBERTa English language model pretrained with the supervision of `roberta-base` solely on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa). The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base). On average DistilRoBERTa is twice as fast as Roberta-base.
@@ -111,7 +111,7 @@ python scripts/binarized_data.py \
     --dump_file data/binarized_text
 ```
 
-Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurences of each tokens in the data:
+Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurrences of each tokens in the data:
 
 ```bash
 python scripts/token_counts.py \

examples/distillation/run_squad_w_distillation.py

@@ -39,6 +39,9 @@ from transformers import (
     DistilBertConfig,
     DistilBertForQuestionAnswering,
     DistilBertTokenizer,
+    RobertaConfig,
+    RobertaForQuestionAnswering,
+    RobertaTokenizer,
     XLMConfig,
     XLMForQuestionAnswering,
     XLMTokenizer,
@@ -73,6 +76,7 @@ MODEL_CLASSES = {
     "xlnet": (XLNetConfig, XLNetForQuestionAnswering, XLNetTokenizer),
     "xlm": (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
     "distilbert": (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
+    "roberta": (RobertaConfig, RobertaForQuestionAnswering, RobertaTokenizer),
 }
 
 
@@ -716,7 +720,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/hans/test_hans.py

@@ -520,7 +520,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/mm-imdb/run_mmimdb.py

@@ -492,7 +492,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/ner/run_ner.py

@@ -33,6 +33,9 @@ from tqdm import tqdm, trange
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
+    AlbertConfig,
+    AlbertForTokenClassification,
+    AlbertTokenizer,
     BertConfig,
     BertForTokenClassification,
     BertTokenizer,
@@ -70,6 +73,7 @@ ALL_MODELS = sum(
 )
 
 MODEL_CLASSES = {
+    "albert": (AlbertConfig, AlbertForTokenClassification, AlbertTokenizer),
     "bert": (BertConfig, BertForTokenClassification, BertTokenizer),
     "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
     "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
@@ -77,6 +81,8 @@ MODEL_CLASSES = {
     "xlmroberta": (XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer),
 }
 
+TOKENIZER_ARGS = ["do_lower_case", "strip_accents", "keep_accents", "use_fast"]
+
 
 def set_seed(args):
     random.seed(args.seed)
@@ -462,7 +468,13 @@ def main():
     parser.add_argument(
         "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
     )
-
+    parser.add_argument(
+        "--keep_accents", action="store_const", const=True, help="Set this flag if model is trained with accents."
+    )
+    parser.add_argument(
+        "--strip_accents", action="store_const", const=True, help="Set this flag if model is trained without accents."
+    )
+    parser.add_argument("--use_fast", action="store_const", const=True, help="Set this flag to use fast tokenization.")
     parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
     parser.add_argument(
         "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
@@ -545,7 +557,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)
@@ -590,10 +602,12 @@ def main():
         label2id={label: i for i, label in enumerate(labels)},
         cache_dir=args.cache_dir if args.cache_dir else None,
     )
+    tokenizer_args = {k: v for k, v in vars(args).items() if v is not None and k in TOKENIZER_ARGS}
+    logger.info("Tokenizer arguments: %s", tokenizer_args)
     tokenizer = tokenizer_class.from_pretrained(
         args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-        do_lower_case=args.do_lower_case,
         cache_dir=args.cache_dir if args.cache_dir else None,
+        **tokenizer_args,
     )
     model = model_class.from_pretrained(
         args.model_name_or_path,
@@ -636,7 +650,7 @@ def main():
     # Evaluation
     results = {}
     if args.do_eval and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, **tokenizer_args)
         checkpoints = [args.output_dir]
         if args.eval_all_checkpoints:
             checkpoints = list(
@@ -658,7 +672,7 @@ def main():
                 writer.write("{} = {}\n".format(key, str(results[key])))
 
     if args.do_predict and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, **tokenizer_args)
         model = model_class.from_pretrained(args.output_dir)
         model.to(args.device)
         result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")

examples/ner/run_pl.sh

@@ -1,6 +1,20 @@
-# Require pytorch-lightning=0.6
+# Install newest ptl.
+pip install -U git+http://github.com/PyTorchLightning/pytorch-lightning/
+
+
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+ wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
 export MAX_LENGTH=128
 export BERT_MODEL=bert-base-multilingual-cased
+python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
 export OUTPUT_DIR=germeval-model
 export BATCH_SIZE=32
 export NUM_EPOCHS=3

examples/ner/run_pl_ner.py

@@ -7,8 +7,7 @@ import numpy as np
 import torch
 from seqeval.metrics import f1_score, precision_score, recall_score
 from torch.nn import CrossEntropyLoss
-from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
-from torch.utils.data.distributed import DistributedSampler
+from torch.utils.data import DataLoader, TensorDataset
 
 from transformer_base import BaseTransformer, add_generic_args, generic_train
 from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
@@ -25,13 +24,14 @@ class NERTransformer(BaseTransformer):
     def __init__(self, hparams):
         self.labels = get_labels(hparams.labels)
         num_labels = len(self.labels)
+        self.pad_token_label_id = CrossEntropyLoss().ignore_index
         super(NERTransformer, self).__init__(hparams, num_labels)
 
     def forward(self, **inputs):
         return self.model(**inputs)
 
     def training_step(self, batch, batch_num):
-        "Compute loss"
+        "Compute loss and log."
         inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
         if self.hparams.model_type != "distilbert":
             inputs["token_type_ids"] = (
@@ -40,25 +40,61 @@ class NERTransformer(BaseTransformer):
 
         outputs = self.forward(**inputs)
         loss = outputs[0]
-
         tensorboard_logs = {"loss": loss, "rate": self.lr_scheduler.get_last_lr()[-1]}
         return {"loss": loss, "log": tensorboard_logs}
 
+    def _feature_file(self, mode):
+        return os.path.join(
+            self.hparams.data_dir,
+            "cached_{}_{}_{}".format(
+                mode,
+                list(filter(None, self.hparams.model_name_or_path.split("/"))).pop(),
+                str(self.hparams.max_seq_length),
+            ),
+        )
+
+    def prepare_data(self):
+        "Called to initialize data. Use the call to construct features"
+        args = self.hparams
+        for mode in ["train", "dev", "test"]:
+            cached_features_file = self._feature_file(mode)
+            if not os.path.exists(cached_features_file):
+                logger.info("Creating features from dataset file at %s", args.data_dir)
+                examples = read_examples_from_file(args.data_dir, mode)
+                features = convert_examples_to_features(
+                    examples,
+                    self.labels,
+                    args.max_seq_length,
+                    self.tokenizer,
+                    cls_token_at_end=bool(args.model_type in ["xlnet"]),
+                    cls_token=self.tokenizer.cls_token,
+                    cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
+                    sep_token=self.tokenizer.sep_token,
+                    sep_token_extra=bool(args.model_type in ["roberta"]),
+                    pad_on_left=bool(args.model_type in ["xlnet"]),
+                    pad_token=self.tokenizer.convert_tokens_to_ids([self.tokenizer.pad_token])[0],
+                    pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+                    pad_token_label_id=self.pad_token_label_id,
+                )
+                logger.info("Saving features into cached file %s", cached_features_file)
+                torch.save(features, cached_features_file)
+
     def load_dataset(self, mode, batch_size):
-        labels = get_labels(self.hparams.labels)
-        self.pad_token_label_id = CrossEntropyLoss().ignore_index
-        dataset = self.load_and_cache_examples(labels, self.pad_token_label_id, mode)
-        if mode == "train":
-            if self.hparams.n_gpu > 1:
-                sampler = DistributedSampler(dataset)
-            else:
-                sampler = RandomSampler(dataset)
-        else:
-            sampler = SequentialSampler(dataset)
-        dataloader = DataLoader(dataset, sampler=sampler, batch_size=batch_size)
-        return dataloader
+        "Load datasets. Called after prepare data."
+        cached_features_file = self._feature_file(mode)
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+        all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+        all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
+        all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
+        all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
+        return DataLoader(
+            TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids), batch_size=batch_size
+        )
 
     def validation_step(self, batch, batch_nb):
+        "Compute validation"
+
         inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
         if self.hparams.model_type != "distilbert":
             inputs["token_type_ids"] = (
@@ -68,11 +104,10 @@ class NERTransformer(BaseTransformer):
         tmp_eval_loss, logits = outputs[:2]
         preds = logits.detach().cpu().numpy()
         out_label_ids = inputs["labels"].detach().cpu().numpy()
-
-        return {"val_loss": tmp_eval_loss, "pred": preds, "target": out_label_ids}
+        return {"val_loss": tmp_eval_loss.detach().cpu(), "pred": preds, "target": out_label_ids}
 
     def _eval_end(self, outputs):
-        "Task specific validation"
+        "Evaluation called for both Val and Test"
         val_loss_mean = torch.stack([x["val_loss"] for x in outputs]).mean()
         preds = np.concatenate([x["pred"] for x in outputs], axis=0)
         preds = np.argmax(preds, axis=2)
@@ -96,7 +131,6 @@ class NERTransformer(BaseTransformer):
         }
 
         if self.is_logger():
-            logger.info(self.proc_rank)
             logger.info("***** Eval results *****")
             for key in sorted(results.keys()):
                 logger.info("  %s = %s", key, str(results[key]))
@@ -140,56 +174,6 @@ class NERTransformer(BaseTransformer):
                             )
         return ret
 
-    def load_and_cache_examples(self, labels, pad_token_label_id, mode):
-        args = self.hparams
-        tokenizer = self.tokenizer
-        if self.proc_rank not in [-1, 0] and mode == "train":
-            torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-        # Load data features from cache or dataset file
-        cached_features_file = os.path.join(
-            args.data_dir,
-            "cached_{}_{}_{}".format(
-                mode, list(filter(None, args.model_name_or_path.split("/"))).pop(), str(args.max_seq_length)
-            ),
-        )
-        if os.path.exists(cached_features_file) and not args.overwrite_cache:
-            logger.info("Loading features from cached file %s", cached_features_file)
-            features = torch.load(cached_features_file)
-        else:
-            logger.info("Creating features from dataset file at %s", args.data_dir)
-            examples = read_examples_from_file(args.data_dir, mode)
-            features = convert_examples_to_features(
-                examples,
-                labels,
-                args.max_seq_length,
-                tokenizer,
-                cls_token_at_end=bool(args.model_type in ["xlnet"]),
-                cls_token=tokenizer.cls_token,
-                cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
-                sep_token=tokenizer.sep_token,
-                sep_token_extra=bool(args.model_type in ["roberta"]),
-                pad_on_left=bool(args.model_type in ["xlnet"]),
-                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
-                pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
-                pad_token_label_id=pad_token_label_id,
-            )
-            if self.proc_rank in [-1, 0]:
-                logger.info("Saving features into cached file %s", cached_features_file)
-                torch.save(features, cached_features_file)
-
-        if self.proc_rank == 0 and mode == "train":
-            torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-        # Convert to Tensors and build dataset
-        all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-        all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-        all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
-        all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
-
-        dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
-        return dataset
-
     @staticmethod
     def add_model_specific_args(parser, root_dir):
         # Add NER specific options

examples/ner/transformer_base.py

@@ -1,3 +1,4 @@
+import logging
 import os
 import random
 
@@ -26,6 +27,9 @@ from transformers import (
 )
 
 
+logger = logging.getLogger(__name__)
+
+
 ALL_MODELS = sum(
     (
         tuple(conf.pretrained_config_archive_map.keys())
@@ -77,20 +81,14 @@ class BaseTransformer(pl.LightningModule):
             cache_dir=self.hparams.cache_dir if self.hparams.cache_dir else None,
         )
         self.config, self.tokenizer, self.model = config, tokenizer, model
-        self.proc_rank = -1
 
     def is_logger(self):
-        return self.proc_rank <= 0
+        return self.trainer.proc_rank <= 0
 
     def configure_optimizers(self):
         "Prepare optimizer and schedule (linear warmup and decay)"
-        model = self.model
 
-        t_total = (
-            len(self.train_dataloader())
-            // self.hparams.gradient_accumulation_steps
-            * float(self.hparams.num_train_epochs)
-        )
+        model = self.model
         no_decay = ["bias", "LayerNorm.weight"]
         optimizer_grouped_parameters = [
             {
@@ -103,18 +101,16 @@ class BaseTransformer(pl.LightningModule):
             },
         ]
         optimizer = AdamW(optimizer_grouped_parameters, lr=self.hparams.learning_rate, eps=self.hparams.adam_epsilon)
-        scheduler = get_linear_schedule_with_warmup(
-            optimizer, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
-        )
-        self.lr_scheduler = scheduler
+        self.opt = optimizer
         return [optimizer]
 
     def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, second_order_closure=None):
-
-        # Step each time.
-        optimizer.step()
-        self.lr_scheduler.step()
+        if self.trainer.use_tpu:
+            xm.optimizer_step(optimizer)
+        else:
+            optimizer.step()
         optimizer.zero_grad()
+        self.lr_scheduler.step()
 
     def get_tqdm_dict(self):
         tqdm_dict = {"loss": "{:.3f}".format(self.trainer.avg_loss), "lr": self.lr_scheduler.get_last_lr()[-1]}
@@ -127,22 +123,27 @@ class BaseTransformer(pl.LightningModule):
     def test_end(self, outputs):
         return self.validation_end(outputs)
 
-    @pl.data_loader
     def train_dataloader(self):
-        return self.load_dataset("train", self.hparams.train_batch_size)
+        train_batch_size = self.hparams.train_batch_size
+        dataloader = self.load_dataset("train", train_batch_size)
+
+        t_total = (
+            (len(dataloader.dataset) // (train_batch_size * max(1, self.hparams.n_gpu)))
+            // self.hparams.gradient_accumulation_steps
+            * float(self.hparams.num_train_epochs)
+        )
+        scheduler = get_linear_schedule_with_warmup(
+            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
+        )
+        self.lr_scheduler = scheduler
+        return dataloader
 
-    @pl.data_loader
     def val_dataloader(self):
         return self.load_dataset("dev", self.hparams.eval_batch_size)
 
-    @pl.data_loader
     def test_dataloader(self):
         return self.load_dataset("test", self.hparams.eval_batch_size)
 
-    def init_ddp_connection(self, proc_rank, world_size):
-        self.proc_rank = proc_rank
-        super(BaseTransformer, self).init_ddp_connection(proc_rank, world_size)
-
     @staticmethod
     def add_model_specific_args(parser, root_dir):
         parser.add_argument(
@@ -213,6 +214,7 @@ def add_generic_args(parser, root_dir):
     )
 
     parser.add_argument("--n_gpu", type=int, default=1)
+    parser.add_argument("--n_tpu_cores", type=int, default=0)
     parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
     parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
     parser.add_argument("--do_predict", action="store_true", help="Whether to run predictions on the test set.")
@@ -252,13 +254,22 @@ def generic_train(model, args):
         accumulate_grad_batches=args.gradient_accumulation_steps,
         gpus=args.n_gpu,
         max_epochs=args.num_train_epochs,
+        early_stop_callback=False,
         gradient_clip_val=args.max_grad_norm,
         checkpoint_callback=checkpoint_callback,
     )
+
     if args.fp16:
         train_params["use_amp"] = args.fp16
         train_params["amp_level"] = args.fp16_opt_level
 
+    if args.n_tpu_cores > 0:
+        global xm
+        import torch_xla.core.xla_model as xm
+
+        train_params["num_tpu_cores"] = args.n_tpu_cores
+        train_params["gpus"] = 0
+
     if args.n_gpu > 1:
         train_params["distributed_backend"] = "ddp"
 

examples/run_bertology.py

@@ -338,7 +338,7 @@ def main():
     # Setup devices and distributed training
     if args.local_rank == -1 or args.no_cuda:
         args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:
         torch.cuda.set_device(args.local_rank)
         args.device = torch.device("cuda", args.local_rank)

examples/run_generation.py

@@ -189,7 +189,7 @@ def main():
     args = parser.parse_args()
 
     args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-    args.n_gpu = torch.cuda.device_count()
+    args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
 
     set_seed(args)
 

examples/run_glue.py

@@ -183,8 +183,11 @@ def train(args, train_dataset, model, tokenizer):
     steps_trained_in_current_epoch = 0
     # Check if continuing training from a checkpoint
     if os.path.exists(args.model_name_or_path):
-        # set global_step to gobal_step of last saved checkpoint from model path
-        global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
+        # set global_step to global_step of last saved checkpoint from model path
+        try:
+            global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
+        except ValueError:
+            global_step = 0
         epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
         steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
 
@@ -575,7 +578,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/run_language_modeling.py

@@ -663,7 +663,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/run_multiple_choice.py

@@ -535,7 +535,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/run_squad.py

@@ -725,7 +725,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/run_xnli.py

@@ -530,7 +530,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

examples/summarization/bart/README.md

@@ -0,0 +1,45 @@
+### Get the CNN/Daily Mail Data
+To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:
+
+```bash
+tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+```
+this should make a directory called cnn_dm/ with files like `test.source`. 
+To use your own data, copy that files format. Each article to be summarized is on its own line.
+
+### Usage
+To create summaries for each article in dataset, run:
+```bash
+python evaluate_cnn.py <path_to_test.source> cnn_test_summaries.txt
+```
+the default batch size, 8, fits in 16GB GPU memory, but may need to be adjusted to fit your system.
+
+### Where is the code?
+The core model is in `src/transformers/modeling_bart.py`. This directory only contains examples.
+
+### (WIP) Rouge Scores
+
+### Stanford CoreNLP Setup
+```
+ptb_tokenize () {
+    cat $1 | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > $2
+}
+
+sudo apt install openjdk-8-jre-headless
+sudo apt-get install ant
+wget http://nlp.stanford.edu/software/stanford-corenlp-full-2018-10-05.zip
+unzip stanford-corenlp-full-2018-10-05.zip
+cd stanford-corenlp-full-2018-10-05
+export CLASSPATH=stanford-corenlp-3.9.2.jar:stanford-corenlp-3.9.2-models.jar
+```
+### Rouge Setup
+Install `files2rouge` following the instructions at [here](https://github.com/pltrdy/files2rouge).
+I also needed to run `sudo apt-get install libxml-parser-perl`
+
+```python
+from files2rouge import files2rouge
+from files2rouge import settings
+files2rouge.run(<path_to_tokenized_hypo>,
+                <path_to_tokenized_target>,
+               saveto='rouge_output.txt')
+```

examples/summarization/bart/evaluate_cnn.py

@@ -0,0 +1,60 @@
+import argparse
+from pathlib import Path
+
+import torch
+from tqdm import tqdm
+
+from transformers import BartForMaskedLM, BartTokenizer
+
+
+DEFAULT_DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def chunks(lst, n):
+    """Yield successive n-sized chunks from lst."""
+    for i in range(0, len(lst), n):
+        yield lst[i : i + n]
+
+
+def generate_summaries(lns, out_file, batch_size=8, device=DEFAULT_DEVICE):
+    fout = Path(out_file).open("w")
+    model = BartForMaskedLM.from_pretrained("bart-large-cnn", output_past=True,)
+    tokenizer = BartTokenizer.from_pretrained("bart-large")
+    for batch in tqdm(list(chunks(lns, batch_size))):
+        dct = tokenizer.batch_encode_plus(batch, max_length=1024, return_tensors="pt", pad_to_max_length=True)
+        summaries = model.generate(
+            input_ids=dct["input_ids"].to(device),
+            attention_mask=dct["attention_mask"].to(device),
+            num_beams=4,
+            length_penalty=2.0,
+            max_length=140,
+            min_len=55,
+            no_repeat_ngram_size=3,
+        )
+        dec = [tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in summaries]
+        for hypothesis in dec:
+            fout.write(hypothesis + "\n")
+            fout.flush()
+
+
+def _run_generate():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "source_path", type=str, help="like cnn_dm/test.source",
+    )
+    parser.add_argument(
+        "output_path", type=str, help="where to save summaries",
+    )
+    parser.add_argument(
+        "--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.",
+    )
+    parser.add_argument(
+        "--bs", type=int, default=8, required=False, help="batch size: how many to summarize at a time",
+    )
+    args = parser.parse_args()
+    lns = [" " + x.rstrip() for x in open(args.source_path).readlines()]
+    generate_summaries(lns, args.output_path, batch_size=args.bs, device=args.device)
+
+
+if __name__ == "__main__":
+    _run_generate()

examples/summarization/bart/test_bart_examples.py

@@ -0,0 +1,28 @@
+import logging
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+from .evaluate_cnn import _run_generate
+
+
+articles = [" New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+
+
+class TestBartExamples(unittest.TestCase):
+    def test_bart_cnn_cli(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+        tmp = Path(tempfile.gettempdir()) / "utest_generations.hypo"
+        with tmp.open("w") as f:
+            f.write("\n".join(articles))
+        testargs = ["evaluate_cnn.py", str(tmp), "output.txt"]
+        with patch.object(sys, "argv", testargs):
+            _run_generate()
+            self.assertTrue(Path("output.txt").exists())

examples/summarization/bertabs/README.md

@@ -15,7 +15,7 @@ pip install nltk py-rouge
 cd examples/summarization
 ```
 
-## Reproduce the authors' results on ROUGE
+## Reproduce the authors'  ROUGE score
 
 To be able to reproduce the authors' results on the CNN/Daily Mail dataset you first need to download both CNN and Daily Mail datasets [from Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/) (the links next to "Stories") in the same folder. Then uncompress the archives by running:
 

examples/summarization/bertabs/run_summarization.py

@@ -11,12 +11,13 @@ from tqdm import tqdm
 
 from modeling_bertabs import BertAbs, build_predictor
 from transformers import BertTokenizer
-from utils_summarization import (
-    SummarizationDataset,
+
+from .utils_summarization import (
+    CNNDMDataset,
     build_mask,
     compute_token_type_ids,
     encode_for_summarization,
-    fit_to_block_size,
+    truncate_or_pad,
 )
 
 
@@ -194,7 +195,7 @@ def build_data_iterator(args, tokenizer):
 
 
 def load_and_cache_examples(args, tokenizer):
-    dataset = SummarizationDataset(args.documents_dir)
+    dataset = CNNDMDataset(args.documents_dir)
     return dataset
 
 
@@ -211,7 +212,7 @@ def collate(data, tokenizer, block_size, device):
 
     encoded_text = [encode_for_summarization(story, summary, tokenizer) for _, story, summary in data]
     encoded_stories = torch.tensor(
-        [fit_to_block_size(story, block_size, tokenizer.pad_token_id) for story, _ in encoded_text]
+        [truncate_or_pad(story, block_size, tokenizer.pad_token_id) for story, _ in encoded_text]
     )
     encoder_token_type_ids = compute_token_type_ids(encoded_stories, tokenizer.cls_token_id)
     encoder_mask = build_mask(encoded_stories, tokenizer.pad_token_id)

examples/summarization/bertabs/test_utils_summarization.py

@@ -17,7 +17,7 @@ import unittest
 import numpy as np
 import torch
 
-from utils_summarization import build_mask, compute_token_type_ids, fit_to_block_size, process_story
+from .utils_summarization import build_mask, compute_token_type_ids, process_story, truncate_or_pad
 
 
 class SummarizationDataProcessingTest(unittest.TestCase):
@@ -28,19 +28,19 @@ class SummarizationDataProcessingTest(unittest.TestCase):
         """ Pad the sequence with 0 if the sequence is smaller than the block size."""
         sequence = [1, 2, 3, 4]
         expected_output = [1, 2, 3, 4, 0, 0, 0, 0, 0, 0]
-        self.assertEqual(fit_to_block_size(sequence, self.block_size, 0), expected_output)
+        self.assertEqual(truncate_or_pad(sequence, self.block_size, 0), expected_output)
 
     def test_fit_to_block_sequence_fit_exactly(self):
         """ Do nothing if the sequence is the right size. """
         sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
         expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
-        self.assertEqual(fit_to_block_size(sequence, self.block_size, 0), expected_output)
+        self.assertEqual(truncate_or_pad(sequence, self.block_size, 0), expected_output)
 
     def test_fit_to_block_sequence_too_big(self):
         """ Truncate the sequence if it is too long. """
         sequence = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
         expected_output = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
-        self.assertEqual(fit_to_block_size(sequence, self.block_size, 0), expected_output)
+        self.assertEqual(truncate_or_pad(sequence, self.block_size, 0), expected_output)
 
     def test_process_story_no_highlights(self):
         """ Processing a story with no highlights returns an empty list for the summary.

examples/summarization/bertabs/utils_summarization.py

@@ -10,7 +10,7 @@ from torch.utils.data import Dataset
 # ------------
 
 
-class SummarizationDataset(Dataset):
+class CNNDMDataset(Dataset):
     """ Abstracts the dataset used to train seq2seq models.
 
     The class will process the documents that are located in the specified
@@ -62,11 +62,11 @@ class SummarizationDataset(Dataset):
 def process_story(raw_story):
     """ Extract the story and summary from a story file.
 
-    Attributes:
+    Arguments:
         raw_story (str): content of the story file as an utf-8 encoded string.
 
     Raises:
-        IndexError: If the stoy is empty or contains no highlights.
+        IndexError: If the story is empty or contains no highlights.
     """
     nonempty_lines = list(filter(lambda x: len(x) != 0, [line.strip() for line in raw_story.split("\n")]))
 
@@ -107,7 +107,7 @@ def _add_missing_period(line):
 # --------------------------
 
 
-def fit_to_block_size(sequence, block_size, pad_token_id):
+def truncate_or_pad(sequence, block_size, pad_token_id):
     """ Adapt the source and target sequences' lengths to the block size.
     If the sequence is shorter we append padding token to the right of the sequence.
     """

model_cards/asafaya/bert-base-arabic/README.md

@@ -0,0 +1,44 @@
+# Arabic BERT Model
+
+Pretrained BERT base language model for Arabic
+
+## Pretraining Corpus
+
+`arabic-bert-base` model was pretrained on ~8.2 Billion words:
+
+- Arabic version of [OSCAR](https://traces1.inria.fr/oscar/) - filtered from [Common Crawl](http://commoncrawl.org/)
+- Recent dump of Arabic [Wikipedia](https://dumps.wikimedia.org/backup-index.html)
+
+and other Arabic resources which sum up to ~95GB of text.
+
+__Notes on training data:__
+
+- Our final version of corpus contains some non-Arabic words inlines, which we did not remove from sentences since that would affect some tasks like NER.
+- Although non-Arabic characters were lowered as a preprocessing step, since Arabic characters does not have upper or lower case, there is no cased and uncased version of the model.
+- The corpus and vocabulary set are not restricted to Modern Standard Arabic, they contain some dialectical Arabic too.
+
+## Pretraining details
+
+- This model was trained using Google BERT's github [repository](https://github.com/google-research/bert) on a single TPU v3-8 provided for free from [TFRC](https://www.tensorflow.org/tfrc).
+- Our pretraining procedure follows training settings of bert with some changes: trained for 3M training steps with batchsize of 128, instead of 1M with batchsize of 256.
+
+## Load Pretrained Model
+
+You can use this model by installing `torch` or `tensorflow` and Huggingface library `transformers`. And you can use it directly by initializing it like this:  
+
+```python
+from transformers import AutoTokenizer, AutoModel
+
+tokenizer = AutoTokenizer.from_pretrained("asafaya/bert-base-arabic")
+model = AutoModel.from_pretrained("asafaya/bert-base-arabic")
+```
+
+## Results
+
+For further details on the models performance or any other queries, please refer to [Arabic-BERT](https://github.com/alisafaya/Arabic-BERT)
+
+## Acknowledgement
+
+Thanks to Google for providing free TPU for the training process and for Huggingface for hosting this model on their servers 😊
+
+

model_cards/fmikaelian/camembert-base-fquad/README.md

@@ -0,0 +1,49 @@
+---
+language: french
+---
+
+# camembert-base-fquad
+
+## Description
+
+A baseline model for question-answering in french ([CamemBERT](https://camembert-model.fr/) model fine-tuned on [FQuAD](https://fquad.illuin.tech/))
+
+## Training hyperparameters
+
+```shell
+python3 ./examples/run_squad.py \
+--model_type camembert \
+--model_name_or_path camembert-base \
+--do_train \
+--do_eval \
+--do_lower_case \
+--train_file train.json \
+--predict_file valid.json \
+--learning_rate 3e-5 \
+--num_train_epochs 2 \
+--max_seq_length 384 \
+--doc_stride 128 \
+--output_dir output \
+--per_gpu_eval_batch_size=3 \
+--per_gpu_train_batch_size=3 \
+--save_steps 10000
+``` 
+
+## Evaluation results
+
+```shell
+{"f1": 77.24515316052342, "exact_match": 52.82308657465496}
+```
+
+## Usage
+
+```python
+from transformers import pipeline
+
+nlp = pipeline('question-answering', model='fmikaelian/camembert-base-fquad', tokenizer='fmikaelian/camembert-base-fquad')
+
+nlp({
+    'question': "Qui est Claude Monet?",
+    'context': "Claude Monet, né le 14 novembre 1840 à Paris et mort le 5 décembre 1926 à Giverny, est un peintre français et l’un des fondateurs de l'impressionnisme."
+})
+```

model_cards/fmikaelian/camembert-base-squad/README.md

@@ -0,0 +1,49 @@
+---
+language: french
+---
+
+# camembert-base-squad
+
+## Description
+
+A baseline model for question-answering in french ([CamemBERT](https://camembert-model.fr/) model fine-tuned on [french-translated SQuAD 1.1 dataset](https://github.com/Alikabbadj/French-SQuAD))
+
+## Training hyperparameters
+
+```shell
+python3 ./examples/run_squad.py \
+--model_type camembert \
+--model_name_or_path camembert-base \
+--do_train \
+--do_eval \
+--do_lower_case \
+--train_file SQuAD-v1.1-train_fr_ss999_awstart2_net.json \
+--predict_file SQuAD-v1.1-dev_fr_ss999_awstart2_net.json \
+--learning_rate 3e-5 \
+--num_train_epochs 2 \
+--max_seq_length 384 \
+--doc_stride 128 \
+--output_dir output3 \
+--per_gpu_eval_batch_size=3 \
+--per_gpu_train_batch_size=3 \
+--save_steps 10000
+``` 
+
+## Evaluation results
+
+```shell
+{"f1": 79.8570684959745, "exact_match": 59.21327108373895}
+```
+
+## Usage
+
+```python
+from transformers import pipeline
+
+nlp = pipeline('question-answering', model='fmikaelian/camembert-base-squad', tokenizer='fmikaelian/camembert-base-squad')
+
+nlp({
+    'question': "Qui est Claude Monet?",
+    'context': "Claude Monet, né le 14 novembre 1840 à Paris et mort le 5 décembre 1926 à Giverny, est un peintre français et l’un des fondateurs de l'impressionnisme."
+})
+```

model_cards/mrm8488/bert-spanish-cased-finetuned-ner/README.md

@@ -19,22 +19,29 @@ I preprocessed the dataset and splitted it as train / dev (80/20)
 | Dev                    | 2.2 K |
 
 
-- [Fine-tune on NER script](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py)
+- [Fine-tune on NER script provided by Huggingface](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py)
 
-```bash
-!export NER_DIR='/content/ner_dataset'
-!python /content/transformers/examples/run_ner.py \
-  --model_type bert \
-  --model_name_or_path dccuchile/bert-base-spanish-wwm-cased \
-  --do_train \
-  --do_eval \
-  --data_dir '/content/ner_dataset' \
-  --num_train_epochs 15.0 \
-  --max_seq_length 384 \
-  --output_dir /content/model_output \
-  --save_steps 5000 \
+- Labels covered:
 
 ```
+B-LOC
+B-MISC
+B-ORG
+B-PER
+I-LOC
+I-MISC
+I-ORG
+I-PER
+O
+```
+
+## Metrics on evaluation set:
+
+|                                                      Metric                                                       |  # score  |
+| :------------------------------------------------------------------------------------: | :-------: |
+| F1                                       | **90.17**  
+| Precision                                | **89.86** | 
+| Recall                                   | **90.47** |    
 
 ## Comparison:
 
@@ -44,13 +51,24 @@ I preprocessed the dataset and splitted it as train / dev (80/20)
 | [bert-spanish-cased-finetuned-ner (this one)](https://huggingface.co/mrm8488/bert-spanish-cased-finetuned-ner) | **89.65** |
 |                                              Best Multilingual BERT                                              |   87.38   |
 
-```
- ***** All metrics on Eval results  *****
+## Model in action
 
-f1 = 0.8965040489828165
-loss = 0.11504213575173258
-precision = 0.893679858239811
-recall = 0.8993461462254805
+Fast usage with **pipelines**:
+
+```python
+from transformers import pipeline
+
+nlp_ner = pipeline(
+    "ner",
+    model="mrm8488/bert-spanish-cased-finetuned-ner",
+    tokenizer=(
+        'mrm8488/bert-spanish-cased-finetuned-ner',  
+        {"use_fast": False}
+))
+
+nlp_ner(text)
+
+#Output: [{'entity': 'B-LOC', 'score': 0.9998720288276672, 'word': 'Londres'}]
 ```
 
 > Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)

setup.py

@@ -64,7 +64,7 @@ if stale_egg_info.exists():
 extras = {}
 
 extras["mecab"] = ["mecab-python3"]
-extras["sklearn"] = ["scikit-learn"]
+extras["sklearn"] = ["scikit-learn==0.22.1"]
 extras["tf"] = ["tensorflow"]
 extras["tf-cpu"] = ["tensorflow-cpu"]
 extras["torch"] = ["torch"]

src/transformers/__init__.py

@@ -136,7 +136,7 @@ if is_sklearn_available():
 
 # Modeling
 if is_torch_available():
-    from .modeling_utils import PreTrainedModel, prune_layer, Conv1D
+    from .modeling_utils import PreTrainedModel, prune_layer, Conv1D, top_k_top_p_filtering
     from .modeling_auto import (
         AutoModel,
         AutoModelForPreTraining,
@@ -241,7 +241,7 @@ if is_torch_available():
         CamembertForTokenClassification,
         CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
-    from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
+    from .modeling_encoder_decoder import PreTrainedEncoderDecoder
     from .modeling_t5 import (
         T5PreTrainedModel,
         T5Model,
@@ -255,6 +255,7 @@ if is_torch_available():
         AlbertForMaskedLM,
         AlbertForSequenceClassification,
         AlbertForQuestionAnswering,
+        AlbertForTokenClassification,
         load_tf_weights_in_albert,
         ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
@@ -290,7 +291,13 @@ if is_torch_available():
 
 # TensorFlow
 if is_tf_available():
-    from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
+    from .modeling_tf_utils import (
+        TFPreTrainedModel,
+        TFSharedEmbeddings,
+        TFSequenceSummary,
+        shape_list,
+        tf_top_k_top_p_filtering,
+    )
     from .modeling_tf_auto import (
         TFAutoModel,
         TFAutoModelForPreTraining,

src/transformers/configuration_bart.py

@@ -22,11 +22,10 @@ from .configuration_utils import PretrainedConfig
 
 logger = logging.getLogger(__name__)
 
-_bart_large_url = "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/config.json"
 BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "bart-large": _bart_large_url,
-    "bart-large-mnli": _bart_large_url,  # fine as same
-    "bart-cnn": None,  # not done
+    "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/config.json",
+    "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/config.json",
+    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/config.json",
 }
 
 
@@ -59,6 +58,7 @@ class BartConfig(PretrainedConfig):
         classifier_dropout=0.0,
         output_past=False,
         num_labels=3,
+        bos_token_id=0,
         **common_kwargs
     ):
         r"""
@@ -67,12 +67,16 @@ class BartConfig(PretrainedConfig):
                 config = BartConfig.from_pretrained('bart-large')
                 model = BartModel(config)
         """
-        super().__init__(num_labels=num_labels, output_past=output_past, pad_token_id=pad_token_id, **common_kwargs)
-
+        super().__init__(
+            num_labels=num_labels,
+            output_past=output_past,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            **common_kwargs,
+        )
         self.vocab_size = vocab_size
         self.d_model = d_model  # encoder_embed_dim and decoder_embed_dim
         self.eos_token_id = eos_token_id
-
         self.encoder_ffn_dim = encoder_ffn_dim
         self.encoder_layers = self.num_hidden_layers = encoder_layers
         self.encoder_attention_heads = encoder_attention_heads

src/transformers/convert_bart_original_pytorch_checkpoint_to_pytorch.py

@@ -23,9 +23,11 @@ import fairseq
 import torch
 from packaging import version
 
-from transformers import BartConfig, BartForSequenceClassification, BartModel, BartTokenizer
+from transformers import BartConfig, BartForMaskedLM, BartForSequenceClassification, BartModel, BartTokenizer
 
 
+FAIRSEQ_MODELS = ["bart.large", "bart.large.mnli", "bart.large.cnn"]
+
 if version.parse(fairseq.__version__) < version.parse("0.9.0"):
     raise Exception("requires fairseq >= 0.9.0")
 
@@ -33,7 +35,7 @@ if version.parse(fairseq.__version__) < version.parse("0.9.0"):
 logging.basicConfig(level=logging.INFO)
 logger = logging.getLogger(__name__)
 
-SAMPLE_TEXT = "Hello world! cécé herlolip"
+SAMPLE_TEXT = " Hello world! cécé herlolip"
 
 rename_keys = [
     ("model.classification_heads.mnli.dense.weight", "classification_head.dense.weight"),
@@ -41,7 +43,7 @@ rename_keys = [
     ("model.classification_heads.mnli.out_proj.weight", "classification_head.out_proj.weight"),
     ("model.classification_heads.mnli.out_proj.bias", "classification_head.out_proj.bias"),
 ]
-IGNORE_KEYS = ["encoder.version", "decoder.version", "model.encoder.version", "model.decoder.version"]
+IGNORE_KEYS = ["encoder.version", "decoder.version", "model.encoder.version", "model.decoder.version", "_float_tensor"]
 
 
 def rename_key(dct, old, new):
@@ -53,36 +55,45 @@ def convert_bart_checkpoint(checkpoint_path, pytorch_dump_folder_path):
     """
     Copy/paste/tweak model's weights to our BERT structure.
     """
-    b2 = torch.hub.load("pytorch/fairseq", checkpoint_path)
-    b2.eval()  # disable dropout
-    b2.model.upgrade_state_dict(b2.model.state_dict())
-    config = BartConfig()
-    tokens = b2.encode(SAMPLE_TEXT).unsqueeze(0)
-    tokens2 = BartTokenizer.from_pretrained("bart-large").encode(SAMPLE_TEXT).unsqueeze(0)
+    bart = torch.hub.load("pytorch/fairseq", checkpoint_path)
+    bart.eval()  # disable dropout
+    bart.model.upgrade_state_dict(bart.model.state_dict())
+    hf_model_name = checkpoint_path.replace(".", "-")
+    config = BartConfig.from_pretrained(hf_model_name)
+    tokens = bart.encode(SAMPLE_TEXT).unsqueeze(0)
+    tokens2 = BartTokenizer.from_pretrained(hf_model_name).encode(SAMPLE_TEXT, return_tensors="pt").unsqueeze(0)
     assert torch.eq(tokens, tokens2).all()
 
-    # assert their_output.size() == (1, 11, 1024)
-
-    if checkpoint_path == "bart.large":
-        state_dict = b2.model.state_dict()
+    if checkpoint_path in ["bart.large", "bart.large.cnn"]:
+        state_dict = bart.model.state_dict()
+        for k in IGNORE_KEYS:
+            state_dict.pop(k, None)
         state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"]
         model = BartModel(config)
-        their_output = b2.extract_features(tokens)
-
+        their_output = bart.extract_features(tokens)
     else:  # MNLI Case
-        state_dict = b2.state_dict()
+        state_dict = bart.state_dict()
+        for k in IGNORE_KEYS:
+            state_dict.pop(k, None)
         state_dict["model.shared.weight"] = state_dict["model.decoder.embed_tokens.weight"]
         for src, dest in rename_keys:
             rename_key(state_dict, src, dest)
-        state_dict.pop("_float_tensor", None)
         model = BartForSequenceClassification(config)
-        their_output = b2.predict("mnli", tokens, return_logits=True)
-    for k in IGNORE_KEYS:
-        state_dict.pop(k, None)
+        their_output = bart.predict("mnli", tokens, return_logits=True)
+
+    # Load state dict
     model.load_state_dict(state_dict)
     model.eval()
-    our_outputs = model.forward(tokens)[0]
+    # Check results
 
+    if checkpoint_path == "bart.large.cnn":  # generate doesnt work yet
+        model = BartForMaskedLM(config, base_model=model)
+        assert "lm_head.weight" in model.state_dict()
+        assert model.lm_head.out_features == config.max_position_embeddings
+        model.eval()
+        our_outputs = model.model.forward(tokens)[0]
+    else:
+        our_outputs = model.forward(tokens)[0]
     assert their_output.shape == our_outputs.shape
     assert (their_output == our_outputs).all().item()
     Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
@@ -92,7 +103,8 @@ def convert_bart_checkpoint(checkpoint_path, pytorch_dump_folder_path):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     # Required parameters
-    parser.add_argument("fairseq_path", choices=["bart.large", "bart.large.mnli"], type=str, help="")
+    parser.add_argument("fairseq_path", choices=FAIRSEQ_MODELS, type=str, help="")
+
     parser.add_argument("pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
     args = parser.parse_args()
     convert_bart_checkpoint(

src/transformers/convert_roberta_original_pytorch_checkpoint_to_pytorch.py

@@ -46,7 +46,9 @@ logger = logging.getLogger(__name__)
 SAMPLE_TEXT = "Hello world! cécé herlolip"
 
 
-def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_folder_path, classification_head):
+def convert_roberta_checkpoint_to_pytorch(
+    roberta_checkpoint_path: str, pytorch_dump_folder_path: str, classification_head: bool
+):
     """
     Copy/paste/tweak roberta's weights to our BERT structure.
     """

src/transformers/modeling_albert.py

@@ -788,6 +788,103 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
         return outputs  # (loss), logits, (hidden_states), (attentions)
 
 
+@add_start_docstrings(
+    """Albert Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForTokenClassification(AlbertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        from transformers import AlbertTokenizer, AlbertForTokenClassification
+        import torch
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = AlbertForTokenClassification.from_pretrained('albert-base-v2')
+
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+
+        loss, scores = outputs[:2]
+
+        """
+
+        outputs = self.albert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)[active_loss]
+                active_labels = labels.view(-1)[active_loss]
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
 @add_start_docstrings(
     """Albert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
     the hidden-states output to compute `span start logits` and `span end logits`). """,

src/transformers/modeling_auto.py

@@ -42,6 +42,7 @@ from .modeling_albert import (
     AlbertForMaskedLM,
     AlbertForQuestionAnswering,
     AlbertForSequenceClassification,
+    AlbertForTokenClassification,
     AlbertModel,
 )
 from .modeling_bart import BART_PRETRAINED_MODEL_ARCHIVE_MAP, BartForMaskedLM, BartForSequenceClassification, BartModel
@@ -233,6 +234,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (RobertaConfig, RobertaForTokenClassification),
         (BertConfig, BertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
+        (AlbertConfig, AlbertForTokenClassification),
     ]
 )
 

src/transformers/modeling_bart.py

@@ -13,8 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """PyTorch BART model, ported from the fairseq repo."""
-
 import logging
+import math
 import random
 from typing import Dict, List, Optional, Tuple
 
@@ -24,7 +24,7 @@ from torch import Tensor, nn
 
 from .configuration_bart import BartConfig
 from .file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from .modeling_utils import PreTrainedModel, create_position_ids_from_input_ids
+from .modeling_utils import BeamHypotheses, PreTrainedModel, create_position_ids_from_input_ids
 
 
 logger = logging.getLogger(__name__)
@@ -33,6 +33,7 @@ logger = logging.getLogger(__name__)
 BART_PRETRAINED_MODEL_ARCHIVE_MAP = {
     "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/pytorch_model.bin",
     "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/pytorch_model.bin",
+    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/pytorch_model.bin",
 }
 
 BART_START_DOCSTRING = r"""
@@ -86,7 +87,7 @@ def _prepare_bart_decoder_inputs(
             causal_lm_mask = None
         new_shape = (bsz, tgt_len, tgt_len)
         # make it broadcastable so can just be added to the attention coefficients
-        decoder_attn_mask = _combine_masks(decoder_padding_mask, causal_lm_mask, new_shape)
+        decoder_attn_mask = _combine_masks(decoder_padding_mask, causal_lm_mask, new_shape).to(device=input_ids.device)
     assert decoder_attn_mask is None or decoder_attn_mask.shape == (bsz, 1, tgt_len, tgt_len)
     return decoder_input_ids, decoder_attn_mask
 
@@ -207,7 +208,7 @@ class EncoderLayer(nn.Module):
             encoded output of shape `(seq_len, batch, embed_dim)`
         """
         residual = x
-        x, attn_weights = self.self_attn.forward(
+        x, attn_weights = self.self_attn(
             query=x, key=x, value=x, key_padding_mask=encoder_padding_mask, need_weights=self.output_attentions,
         )
         x = F.dropout(x, p=self.dropout, training=self.training)
@@ -291,7 +292,7 @@ class BartEncoder(nn.Module):
             if self.training and (dropout_probability < self.layerdrop):  # skip the layer
                 attn = None
             else:
-                x, attn = encoder_layer.forward(x, attention_mask)
+                x, attn = encoder_layer(x, attention_mask)
 
             if self.output_attentions:
                 all_attentions.append(attn)
@@ -332,7 +333,7 @@ class DecoderLayer(nn.Module):
         x,
         encoder_hidden_states,
         encoder_attn_mask=None,
-        decoder_cached_states=None,
+        layer_state=None,
         attention_mask=None,
         need_attn_weights=False,
     ):
@@ -348,43 +349,28 @@ class DecoderLayer(nn.Module):
         Returns:
             encoded output of shape `(seq_len, batch, embed_dim)`
         """
-        if decoder_cached_states is None:
-            prev_self_attn_state, prev_attn_state = (None, None)
-        else:
-            assert len(decoder_cached_states) == 3
-            prev_self_attn_state, prev_attn_state = (
-                decoder_cached_states["self"],
-                decoder_cached_states["encoder_decoder"],
-            )
 
         residual = x
-        if prev_self_attn_state is not None:
-            saved_state = prev_self_attn_state
-            decoder_cached_states["self"] = saved_state
         y = x  # TODO(SS): figure out why fairseq did this, then hopefully delete it
 
-        x, self_attn_weights = self.self_attn.forward(
-            query=x,
-            key=y,
-            value=y,
-            decoder_cached_states=decoder_cached_states,
-            need_weights=need_attn_weights,
-            attn_mask=attention_mask,
+        if layer_state is None:
+            layer_state = {}
+        # next line mutates layer state
+        x, self_attn_weights = self.self_attn(
+            query=x, key=y, value=y, layer_state=layer_state, need_weights=need_attn_weights, attn_mask=attention_mask,
         )
         x = F.dropout(x, p=self.dropout, training=self.training)
         x = residual + x
         x = self.self_attn_layer_norm(x)
         residual = x
         assert self.encoder_attn.cache_key != self.self_attn.cache_key
-        if prev_attn_state is not None:
-            saved_state = prev_attn_state
-            decoder_cached_states["encoder_decoder"] = saved_state
-        x, encoder_attn_weights = self.encoder_attn.forward(
+
+        x, encoder_attn_weights = self.encoder_attn(
             query=x,
             key=encoder_hidden_states,  # could be None
             value=encoder_hidden_states,
             key_padding_mask=encoder_attn_mask,
-            decoder_cached_states=decoder_cached_states,
+            layer_state=layer_state,  # mutates layer state
             static_kv=True,
             need_weights=False,  # not returning it so why compute it
         )
@@ -403,15 +389,8 @@ class DecoderLayer(nn.Module):
         return (
             x,
             self_attn_weights,
-            decoder_cached_states,
-        )  # just self_attn weights for now, following t5, decoder_cached_states = cache for decoding
-
-    def _past_to_dict(self, prev_attn_state):
-        prev_key, prev_value = prev_attn_state[:2]
-        saved_state = {"prev_key": prev_key, "prev_value": prev_value}
-        if len(prev_attn_state) >= 3:
-            saved_state["prev_key_padding_mask"] = prev_attn_state[2]
-        return saved_state
+            layer_state,
+        )  # just self_attn weights for now, following t5, layer_state = cache for decoding
 
 
 class BartDecoder(nn.Module):
@@ -440,6 +419,7 @@ class BartDecoder(nn.Module):
             [DecoderLayer(config) for _ in range(config.decoder_layers)]
         )  # type: List[DecoderLayer]
         self.layernorm_embedding = LayerNorm(config.d_model)
+        self.generation_mode = False
 
     def forward(
         self,
@@ -469,11 +449,15 @@ class BartDecoder(nn.Module):
                 - attentions
         """
         # embed positions
-        positions = self.embed_positions(input_ids)
-        x = self.embed_tokens(input_ids)
+        positions = self.embed_positions(input_ids, generation_mode=self.generation_mode)
 
-        if positions is not None:
-            x += positions
+        if self.generation_mode:
+            input_ids = input_ids[:, -1:]
+            positions = positions[:, -1:]  # happens after we embed them
+            assert input_ids.ne(self.padding_idx).any()
+
+        x = self.embed_tokens(input_ids)
+        x += positions
 
         x = self.layernorm_embedding(x)
         x = F.dropout(x, p=self.dropout, training=self.training)
@@ -489,17 +473,19 @@ class BartDecoder(nn.Module):
             dropout_probability = random.uniform(0, 1)
             if self.training and (dropout_probability < self.layerdrop):
                 continue
+
             layer_state = decoder_cached_states[i] if decoder_cached_states is not None else None
-            x, layer_self_attn, layer_past = decoder_layer.forward(
+            x, layer_self_attn, layer_past = decoder_layer(
                 x,
                 encoder_hidden_states,
                 encoder_padding_mask,
-                decoder_cached_states=layer_state,
+                layer_state=layer_state,
                 attention_mask=combined_mask,
                 need_attn_weights=self.output_attentions,
             )
+
             if self.output_past:
-                next_decoder_cache.append(layer_past)
+                next_decoder_cache.append(layer_past.copy())
             if self.output_hidden_states:
                 all_hidden_states += (x,)
             if self.output_attentions:
@@ -509,7 +495,22 @@ class BartDecoder(nn.Module):
         all_hidden_states = [hidden_state.transpose(0, 1) for hidden_state in all_hidden_states]
         x = x.transpose(0, 1)
 
-        return x, next_decoder_cache, all_hidden_states, list(all_self_attns)
+        if self.output_past:
+            next_cache = ((encoder_hidden_states, encoder_padding_mask), next_decoder_cache)
+        else:
+            next_cache = None
+        return x, next_cache, all_hidden_states, list(all_self_attns)
+
+
+def reorder_attn_buffer(input_buffer, new_order):
+    """Reorder buffered internal state (for incremental generation)."""
+    # input_buffer = self._get_input_buffer(incremental_state)
+    for k in input_buffer.keys():
+        input_buffer_k = input_buffer[k]
+        if input_buffer_k is not None:
+            input_buffer[k] = input_buffer_k.index_select(0, new_order)
+        # incremental_state = self._set_input_buffer(incremental_state, input_buffer)
+    return input_buffer
 
 
 class SelfAttention(nn.Module):
@@ -557,7 +558,7 @@ class SelfAttention(nn.Module):
         key: Optional[Tensor],
         value: Optional[Tensor],
         key_padding_mask: Optional[Tensor] = None,
-        decoder_cached_states: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
+        layer_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,
         need_weights: bool = False,
         static_kv: bool = False,
         attn_mask: Optional[Tensor] = None,
@@ -579,8 +580,8 @@ class SelfAttention(nn.Module):
         assert embed_dim == self.embed_dim
         assert list(query.size()) == [tgt_len, bsz, embed_dim]
         # get here for encoder decoder cause of static_kv
-        if decoder_cached_states is not None:  # get the last k,v and mask for reuse
-            saved_state = decoder_cached_states.get(self.cache_key, {})
+        if layer_state is not None:  # get the last k,v and mask for reuse
+            saved_state = layer_state.get(self.cache_key, {})
             if "prev_key" in saved_state:
                 # previous time steps are cached - no need to recompute key and value if they are static
                 if static_kv:
@@ -588,6 +589,7 @@ class SelfAttention(nn.Module):
                     key = value = None
         else:
             saved_state = None
+            layer_state = {}
 
         q = self.q_proj(query) * self.scaling
         if self.encoder_decoder_attention:
@@ -608,17 +610,16 @@ class SelfAttention(nn.Module):
             v = self._shape(v, -1, bsz)
 
         if saved_state is not None:
-            k, v, key_padding_mask, new_state = self._use_and_update_saved_state(
-                k, v, saved_state, key_padding_mask, static_kv, bsz
-            )
-            saved_state.update(
-                {
-                    "prev_key": k.view(bsz, self.num_heads, -1, self.head_dim),
-                    "prev_value": v.view(bsz, self.num_heads, -1, self.head_dim),
-                    "prev_key_padding_mask": key_padding_mask,
-                }
-            )
-            decoder_cached_states[self.cache_key] = saved_state  # Update cache
+            k, v, key_padding_mask = self._use_saved_state(k, v, saved_state, key_padding_mask, static_kv, bsz)
+        # assert self.cache_key != 'encoder_decoder' or key_padding_mask is None
+
+        # Update cache
+        layer_state[self.cache_key] = {
+            "prev_key": k.view(bsz, self.num_heads, -1, self.head_dim),
+            "prev_value": v.view(bsz, self.num_heads, -1, self.head_dim),
+            "prev_key_padding_mask": key_padding_mask if not static_kv else None,
+        }
+
         assert k is not None
         src_len = k.size(1)
         attn_weights = torch.bmm(q, k.transpose(1, 2))
@@ -632,7 +633,7 @@ class SelfAttention(nn.Module):
         # This is part of a workaround to get around fork/join parallelism not supporting Optional types.
         if key_padding_mask is not None and key_padding_mask.dim() == 0:
             key_padding_mask = None
-        assert key_padding_mask is None or key_padding_mask.size()[:2] == (bsz, src_len)
+        assert key_padding_mask is None or key_padding_mask.size()[:2] == (bsz, src_len,)
 
         if key_padding_mask is not None:  # don't attend to padding symbols
             attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
@@ -650,7 +651,7 @@ class SelfAttention(nn.Module):
         attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
         return attn_output, attn_weights
 
-    def _use_and_update_saved_state(self, k, v, saved_state, key_padding_mask, static_kv, bsz):
+    def _use_saved_state(self, k, v, saved_state, key_padding_mask, static_kv, bsz):
         # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)
         if "prev_key" in saved_state:
             _prev_key = saved_state["prev_key"]
@@ -675,7 +676,7 @@ class SelfAttention(nn.Module):
         key_padding_mask = self._cat_prev_key_padding_mask(
             key_padding_mask, prev_key_padding_mask, bsz, k.size(1), static_kv
         )
-        return k, v, key_padding_mask, saved_state
+        return k, v, key_padding_mask
 
     @staticmethod
     def _cat_prev_key_padding_mask(
@@ -693,7 +694,6 @@ class SelfAttention(nn.Module):
         # During incremental decoding, as the padding token enters and
         # leaves the frame, there will be a time when prev or current is None
         elif prev_key_padding_mask is not None:
-
             filler = torch.zeros(batch_size, src_len - prev_key_padding_mask.size(1))
             if prev_key_padding_mask.is_cuda:
                 filler = filler.cuda()
@@ -747,9 +747,13 @@ class LearnedPositionalEmbedding(nn.Embedding):
         num_embeddings += padding_idx + 1  # WHY?
         super().__init__(num_embeddings, embedding_dim, padding_idx=padding_idx)
 
-    def forward(self, input):
+    def forward(self, input, generation_mode=False):
         """Input is expected to be of size [bsz x seqlen]."""
-        positions = create_position_ids_from_input_ids(input, self.padding_idx)
+        if generation_mode:  # the position is our current step in the decoded sequence
+            pos = int(self.padding_idx + input.size(1))
+            positions = input.data.new(1, 1).fill_(pos)
+        else:
+            positions = create_position_ids_from_input_ids(input, self.padding_idx)
         return super().forward(positions)
 
 
@@ -826,21 +830,20 @@ class BartModel(PretrainedBartModel):
             assert attention_mask.max() <= 0
 
         # make masks if user doesn't supply
-        decoder_input_ids, decoder_attn_mask = _prepare_bart_decoder_inputs(
-            self.config, input_ids, decoder_input_ids=decoder_input_ids, decoder_attn_mask=decoder_attention_mask,
-        )
-
+        if not self.decoder.generation_mode:
+            decoder_input_ids, decoder_attention_mask = _prepare_bart_decoder_inputs(
+                self.config, input_ids, decoder_input_ids=decoder_input_ids, decoder_attn_mask=decoder_attention_mask,
+            )
         assert decoder_input_ids is not None
         if encoder_outputs is None:
-            # TODO(SS): make this caching more usable when overwrite generate
-            encoder_outputs = self.encoder.forward(input_ids=input_ids, attention_mask=attention_mask)
+            encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
         assert isinstance(encoder_outputs, tuple)
-        # dec_features, decoder_cached_states, dec_hidden, dec_attn
-        decoder_outputs = self.decoder.forward(
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
             decoder_input_ids,
             encoder_outputs[0],
             attention_mask,
-            decoder_attn_mask,
+            decoder_attention_mask,
             decoder_cached_states=decoder_cached_states,
         )
         # Attention and hidden_states will be [] or None if they aren't needed
@@ -856,20 +859,26 @@ class BartModel(PretrainedBartModel):
         self.shared = value
 
     def get_output_embeddings(self):
-        return _make_linear_from_emb(self.shared)
+        return _make_linear_from_emb(self.shared)  # make it on the fly
 
 
 @add_start_docstrings(
-    "The bare BART Model with a language modeling head", BART_START_DOCSTRING,
+    "The bare BART Model with a language modeling head. This is the model used for summarization.",
+    BART_START_DOCSTRING,
 )
 class BartForMaskedLM(PretrainedBartModel):
     base_model_prefix = "model"
 
     def __init__(self, config: BartConfig):
         super().__init__(config)
-        self.model = BartModel(config)
+        # if base_model is None:
+        base_model = BartModel(config)
+        self.model = base_model
         self.lm_head = _make_linear_from_emb(self.model.shared)
 
+    def tie_weights(self):
+        pass  # hack to prevent changing lm_head.out_features. The input and output embeddings are still the same.
+
     @add_start_docstrings_to_callable(BART_INPUTS_DOCSTRING)
     def forward(
         self,
@@ -916,7 +925,7 @@ class BartForMaskedLM(PretrainedBartModel):
             outputs = model(input_ids=input_ids, lm_labels=input_ids)
             loss, prediction_scores = outputs[:2]
         """
-        outputs = self.model.forward(
+        outputs = self.model(
             input_ids,
             attention_mask=attention_mask,
             decoder_input_ids=decoder_input_ids,
@@ -924,7 +933,7 @@ class BartForMaskedLM(PretrainedBartModel):
             decoder_attention_mask=decoder_attention_mask,
             decoder_cached_states=decoder_cached_states,
         )
-        lm_logits = self.lm_head.forward(outputs[0])
+        lm_logits = self.lm_head(outputs[0])
         outputs = (lm_logits,) + outputs[1:]  # Add hidden states and attention if they are here
         if lm_labels is not None:
             loss_fct = nn.CrossEntropyLoss()
@@ -935,12 +944,309 @@ class BartForMaskedLM(PretrainedBartModel):
         return outputs
 
     @staticmethod
-    def prepare_inputs_for_generation(input_ids, past, **kwargs):
-        return {"input_ids": input_ids, "decoder_cached_states": past, "decoder_input_ids": input_ids[:, -1:]}
+    def prepare_inputs_for_generation(input_ids, past, decoder_input_ids, attention_mask):
+        if past is None:  # first step
+            encoder_outputs, decoder_cached_states = None, None
+        else:
+            encoder_outputs, decoder_cached_states = past
+        return {
+            "input_ids": input_ids,  # ignored after first pass
+            "decoder_cached_states": decoder_cached_states,
+            "decoder_input_ids": decoder_input_ids,
+            "encoder_outputs": encoder_outputs,
+            "attention_mask": attention_mask,
+            # "decoder_attention_mask": decoder_attention_mask,
+        }
+
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        ((enc_out, enc_mask), decoder_cached_states) = past
+        reordered_past = []
+        for layer_past in decoder_cached_states:
+            # get the correct batch idx from decoder layer's batch dim for cross and self-attn
+            layer_past_new = {
+                attn_key: reorder_attn_buffer(attn_cache, beam_idx) for attn_key, attn_cache in layer_past.items()
+            }
+            # reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
+            # reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
+            reordered_past.append(layer_past_new)
+        new_enc_out = enc_out if enc_out is None else enc_out.index_select(1, beam_idx)
+        new_enc_mask = enc_mask if enc_mask is None else enc_mask.index_select(0, beam_idx)
+
+        past = ((new_enc_out, new_enc_mask), reordered_past)
+        return past
 
     def get_output_embeddings(self):
         return self.lm_head
 
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids,
+        attention_mask=None,
+        max_length=20,
+        num_beams=1,
+        repetition_penalty=1.0,
+        length_penalty=1.0,
+        num_return_sequences=1,
+        min_len=0,
+        no_repeat_ngram_size=0,
+    ):
+        r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
+        and beam-search.
+
+        Adapted in part from Facebook's `XLM beam search code`_ and `Fairseq beam search code`_.
+
+        .. _`XLM beam search code`:
+           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529
+        .. _`Fairseq beam search code`:
+           https://github.com/pytorch/fairseq/blob/master/fairseq/sequence_generator.py
+
+
+        Parameters:
+
+            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
+                The sequence used as a prompt for the generation. If `None` the method initializes
+                it as an empty `torch.LongTensor` of shape `(1,)`.
+
+            max_length: (`optional`) int
+                The max length of the sequence to be generated. Does not include tokens in input_ids.
+
+            num_beams: (`optional`) int
+                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.
+
+            repetition_penalty: (`optional`) float
+                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.
+
+            length_penalty: (`optional`) float
+                Exponential penalty to the length. Default to 1.
+
+            num_return_sequences: (`optional`) int
+                The number of independently computed returned sequences for each element in the batch. Default to 1.
+
+            min_len: (`optional`) int
+
+        Returns:
+            `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
+                sequence_length is <= max_length (examples can finish early)
+
+        Examples::
+
+            config = BartConfig(vocab_size=50264, output_past=True)
+            model = AutoModelWithLMHead.from_pretrained('bart-large-cnn', config=config)
+            tokenizer = AutoTokenizer.from_pretrained('bart-large-cnn')
+            ARTICLE_TO_SUMMARIZE = "My friends are cool but they eat too many carbs."
+            inputs = tokenizer.batch_encode_plus([ARTICLE_TO_SUMMARIZE], max_length=1024, return_tensors='pt')
+            # Generate Summary
+            generated_ids = model.generate(inputs['input_ids'], attention_mask=inputs['attention_mask'], num_beams=4, max_length=5)
+            print([tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in generated_ids])
+
+        """
+        bos_token_id = self.config.bos_token_id
+        pad_token_id = self.config.pad_token_id
+        eos_token_id = self.config.eos_token_id
+        batch_size, cur_len = input_ids.shape
+        assert input_ids is not None
+        assert self.config.output_past, "Generating with bart requires instantiating a config with output_past=True"
+        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
+        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
+        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
+        assert isinstance(pad_token_id, int)
+        assert bos_token_id == 0, "configurable bos_token_id not yet supported"
+        assert length_penalty > 0, "`length_penalty` should be strictly positive."
+        assert (
+            isinstance(num_return_sequences, int) and num_return_sequences > 0
+        ), "`num_return_sequences` should be a positive integer."
+
+        # current position and vocab size
+        cur_len = input_ids.shape[1]
+        vocab_size = self.config.vocab_size
+
+        if num_return_sequences != 1:
+            # Expand input to num return sequences
+            input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
+            input_ids = input_ids.contiguous().view(
+                batch_size * num_return_sequences, cur_len
+            )  # shape: (batch_size * num_return_sequences, cur_len)
+            batch_size *= num_return_sequences
+
+        # Below here somewhat similar to PretrainedModel._generate_beam_search
+        # Expand input to num beams
+        input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)
+
+        input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len)  # (batch_size * num_beams, cur_len)
+        if attention_mask is not None:
+            attention_mask = (
+                attention_mask.unsqueeze(1)
+                .expand(batch_size, num_beams, cur_len)
+                .contiguous()
+                .view(batch_size * num_beams, cur_len)
+            )  # RESHAPE
+
+        # generated hypotheses
+        finalized_hyps = [  # they end in EOS and we wont work on them more!
+            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=True) for _ in range(batch_size)
+        ]
+
+        # scores for each sentence in the beam
+        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
+        beam_scores[:, 1:] = -1e9  # avoid ties in first step
+        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)
+
+        # decoder tokens
+        prev_output_tokens = input_ids.new(batch_size * num_beams, 1).long().fill_(-1)
+        prev_output_tokens[:, 0] = 2  # HARDCODED EOS, which will be removed at the end.
+        decoder_cache = None
+        done = [False for _ in range(batch_size)]  # done sentences
+
+        self.model.decoder.generation_mode = True  # tells decoder not to use causal mask
+        for step in range(max_length + 1):
+            decoder_input_ids = prev_output_tokens.clone()
+            model_inputs = self.prepare_inputs_for_generation(
+                input_ids, decoder_cache, decoder_input_ids, attention_mask,
+            )
+            outputs = self(**model_inputs)
+            lprobs = F.log_softmax(outputs[0][:, -1, :], dim=-1)
+
+            lprobs[lprobs != lprobs] = -math.inf  # block nans
+            lprobs[:, pad_token_id] = -math.inf
+            # TODO(SS): fairseq also takes out <unk> every step, and has unk at slot 3
+
+            if step == 0:  # Force BOS to be chosen
+                lprobs[:, bos_token_id + 1 :] = -math.inf
+            elif step < min_len:  # Prevent EOS from being chosen
+                lprobs[:, eos_token_id] = -math.inf
+            elif step == max_length:  # FORCE EOS to be chosen
+                lprobs[:, :eos_token_id] = -math.inf
+                lprobs[:, eos_token_id + 1 :] = -math.inf
+            assert self._do_output_past(outputs)
+            decoder_cache = outputs[1]
+            if repetition_penalty != 1.0:
+                self.enforce_repetition_penalty_(lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty)
+            num_hypos = batch_size * num_beams
+            if no_repeat_ngram_size > 0:  # copied from fairseq
+                # for each sentence, calculate a list of banned tokens to prevent repetitively generating the same ngrams
+                banned_tokens = self.calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step)
+                # then set their probabilities tof -inf
+                for idx in range(num_hypos):
+                    lprobs[idx, banned_tokens[idx]] = -math.inf
+            assert lprobs.size() == (batch_size * num_beams, vocab_size)
+            _scores = lprobs + beam_scores[:, None].expand_as(lprobs)  # (batch_size * num_beams, vocab_size)
+
+            # re-organize to group the beam together (we are keeping top hypothesis across beams)
+            _scores = _scores.view(batch_size, num_beams * vocab_size)  # (batch_size, num_beams * vocab_size)
+            # Take the best 2 x beam_size predictions for each example, we'll choose the first beam_size of these which don't predict eos to continue with.
+            next_scores, next_words = torch.topk(_scores, 2 * num_beams)
+            assert next_scores.size() == next_words.size() == (batch_size, 2 * num_beams)
+
+            # list of (batch_size * num_beams)
+            next_batch_beam = []  # Tuple(next score, next word, current position in the batch)
+            for batch_idx in range(batch_size):
+                # if we are done with this sentence (because we can't improve)
+                if done[batch_idx]:  # then pad all associated hypotheses
+                    assert (
+                        len(finalized_hyps[batch_idx]) >= num_beams
+                    ), "Example can only be done if at least {} beams have been generated".format(num_beams)
+                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
+                    continue
+
+                # Otherwise generate some next word choices
+                next_sent_beam = []
+                # add next words for this sentence
+                for i, (idx, score) in enumerate(zip(next_words[batch_idx], next_scores[batch_idx])):
+                    beam_id = idx // vocab_size
+                    word_id = idx % vocab_size
+                    assert prev_output_tokens.shape[1] == (step + 1)
+                    if word_id.item() == eos_token_id:
+                        if i >= num_beams:
+                            continue
+                        finalized_hyps[batch_idx].add(
+                            prev_output_tokens[batch_idx * num_beams + beam_id].clone(), score.item(),
+                        )
+                    else:
+                        next_sent_beam.append((score, word_id, batch_idx * num_beams + beam_id))
+
+                    if len(next_sent_beam) == num_beams:  # TODO(SS): can we delete this?
+                        break
+                # Check if were done so that we can save a pad step if all(done)
+                done[batch_idx] = done[batch_idx] or finalized_hyps[batch_idx].is_done(
+                    next_scores[batch_idx].max().item(), cur_len=step + 1,
+                )
+                assert len(next_sent_beam) == num_beams, "Beam should always be full"
+                next_batch_beam.extend(next_sent_beam)
+                assert len(next_batch_beam) == num_beams * (batch_idx + 1)
+
+            if all(done):
+                break
+
+            # sanity check / prepare next batch
+            assert len(next_batch_beam) == batch_size * num_beams
+            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
+            beam_words = input_ids.new([x[1] for x in next_batch_beam])
+            beam_idx = input_ids.new([x[2] for x in next_batch_beam])
+            # re-order decoder inputs to [beam_idx]
+            prev_output_tokens = prev_output_tokens[beam_idx]
+            prev_output_tokens = torch.cat([prev_output_tokens, beam_words.unsqueeze(1)], dim=-1)
+
+            # re-order internal states
+            decoder_cache = self._reorder_cache(decoder_cache, beam_idx)
+
+        for batch_idx in range(batch_size):
+            # Add all open beam hypothesis to generated_hyps
+            if done[batch_idx]:
+                continue
+            offset = batch_idx * num_beams
+            for i in range(num_beams):
+                score = beam_scores[offset + i]
+                final_tokens = prev_output_tokens[offset + i]
+                finalized_hyps[batch_idx].add(final_tokens, score.item())
+
+        # select the best hypotheses
+        sent_lengths = input_ids.new(batch_size)
+        best = []
+        for i, hypotheses in enumerate(finalized_hyps):
+            best_hyp = max(hypotheses.beams, key=lambda x: x[0])[1]
+            sent_lengths[i] = len(best_hyp)
+            best.append(best_hyp)
+
+        # shorter batches are filled with pad_token
+        if sent_lengths.min().item() != sent_lengths.max().item():
+            # TODO(SS): decoded = torch.rnn.utils.pad_sequence(best, batch_first=True, padding_value=pad_token_id)
+            sent_max_len = min(sent_lengths.max().item() + 1, max_length + 1)  # TODO(SS): same as step?
+            decoded = input_ids.new(batch_size, sent_max_len).fill_(pad_token_id)
+            # fill with hypothesis and eos_token_id if necessary
+            for i, hypo in enumerate(best):
+                decoded[i, : sent_lengths[i]] = hypo
+                if sent_lengths[i] < max_length:
+                    decoded[i, sent_lengths[i]] = eos_token_id
+        else:
+            assert (len(hypo) == max_length for hypo in best)
+            decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
+        return decoded[:, 1:]  # get rid of starting EOS
+
+    @staticmethod
+    def calc_banned_tokens(prev_output_tokens, num_hypos, no_repeat_ngram_size, step):
+        """Copied from fairseq for no_repeat_ngram in beam_search"""
+        # TODO(SS): this can go on parent if there is demand
+        if step + 2 < no_repeat_ngram_size:
+            return [
+                [] for _ in range(num_hypos)
+            ]  # no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+        gen_ngrams = [{} for _ in range(num_hypos)]
+        for idx in range(num_hypos):
+            gen_tokens = prev_output_tokens[idx].tolist()
+            for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
+                k = tuple(ngram[:-1])
+                gen_ngrams[idx][k] = gen_ngrams[idx].get(k, []) + [ngram[-1]]
+
+        def _get_generated_ngrams(hypo_idx):
+            """Before decoding the next token, prevent decoding of ngrams that have already appeared"""
+            ngram_index = tuple(prev_output_tokens[hypo_idx, step + 2 - no_repeat_ngram_size : step + 1].tolist())
+            return gen_ngrams[hypo_idx].get(ngram_index, [])
+
+        banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
+        return banned_tokens
+
 
 @add_start_docstrings(
     """Bart model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. """,
@@ -1002,7 +1308,7 @@ class BartForSequenceClassification(PretrainedBartModel):
         loss, logits = outputs[:2]
 
         """
-        outputs = self.model.forward(
+        outputs = self.model(
             input_ids,
             attention_mask=attention_mask,
             decoder_input_ids=decoder_input_ids,
@@ -1018,7 +1324,7 @@ class BartForSequenceClassification(PretrainedBartModel):
         # Prepend logits
         outputs = (logits,) + outputs[1:]  # Add hidden states and attention if they are here
         if labels is not None:  # prepend loss to output,
-            loss = F.cross_entropy(logits.view(-1, self.num_labels), labels.view(-1))
+            loss = F.cross_entropy(logits.view(-1, self.config.num_labels), labels.view(-1))
             outputs = (loss,) + outputs
 
         return outputs

src/transformers/modeling_ctrl.py

@@ -454,14 +454,12 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head
 
-    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
         # only last token for inputs_ids if past is defined in kwargs
-        if "past" in kwargs and kwargs["past"]:
+        if past:
             input_ids = input_ids[:, -1].unsqueeze(-1)
 
-        inputs = {"input_ids": input_ids}
-        inputs.update(kwargs)
-        return inputs
+        return {"input_ids": input_ids, "past": past}
 
     @add_start_docstrings_to_callable(CTRL_INPUTS_DOCSTRING)
     def forward(

src/transformers/modeling_encoder_decoder.py

@@ -234,62 +234,3 @@ class PreTrainedEncoderDecoder(nn.Module):
         decoder_outputs = self.decoder(decoder_input_ids, **kwargs_decoder)
 
         return decoder_outputs + encoder_outputs
-
-
-class Model2Model(PreTrainedEncoderDecoder):
-    r"""
-        :class:`~transformers.Model2Model` instantiates a Seq2Seq2 model
-        where both of the encoder and decoder are of the same family. If the
-        name of or that path to a pretrained model is specified the encoder and
-        the decoder will be initialized with the pretrained weight (the
-        cross-attention will be intialized randomly if its weights are not
-        present).
-
-        It is possible to override this behavior and initialize, say, the decoder randomly
-        by creating it beforehand as follows
-
-            config = BertConfig.from_pretrained()
-            decoder = BertForMaskedLM(config)
-            model = Model2Model.from_pretrained('bert-base-uncased', decoder_model=decoder)
-    """
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.tie_weights()
-
-    def tie_weights(self):
-        """ Tying the encoder and decoders' embeddings together.
-
-       We need for each to get down to the embedding weights. However the
-        different model classes are inconsistent to that respect:
-        - BertModel: embeddings.word_embeddings
-        - RoBERTa: embeddings.word_embeddings
-        - XLMModel: embeddings
-        - GPT2: wte
-        - BertForMaskedLM: bert.embeddings.word_embeddings
-        - RobertaForMaskedLM: roberta.embeddings.word_embeddings
-
-        argument of the XEmbedding layer for each model, but it is "blocked"
-        by a model-specific keyword (bert, )...
-        """
-        # self._tie_or_clone_weights(self.encoder, self.decoder)
-        pass
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
-
-        if (
-            "bert" not in pretrained_model_name_or_path
-            or "roberta" in pretrained_model_name_or_path
-            or "distilbert" in pretrained_model_name_or_path
-        ):
-            raise ValueError("Only the Bert model is currently supported.")
-
-        model = super().from_pretrained(
-            encoder_pretrained_model_name_or_path=pretrained_model_name_or_path,
-            decoder_pretrained_model_name_or_path=pretrained_model_name_or_path,
-            *args,
-            **kwargs,
-        )
-
-        return model

src/transformers/modeling_gpt2.py

@@ -276,14 +276,17 @@ GPT2_START_DOCSTRING = r"""
 
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            `input_ids_length` = `sequence_length if `past` is None else 1
             Indices of input sequence tokens in the vocabulary.
+            If using `past` as an input make sure that `input_ids` are those of the last position.
 
             Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
+
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model
@@ -294,10 +297,12 @@ GPT2_INPUTS_DOCSTRING = r"""
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`, defaults to :obj:`None`):
+            `input_ids_length` = `sequence_length if `past` is None else 1
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
+            If using `past` as an input make sure that `token_type_ids` correspond to the `input_ids` of the last position.
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -419,7 +424,8 @@ class GPT2Model(GPT2PreTrainedModel):
 
         # Attention mask.
         if attention_mask is not None:
-            attention_mask = attention_mask.view(-1, input_shape[-1])
+            batch_size = input_ids.shape[0]
+            attention_mask = attention_mask.view(batch_size, -1)
             # We create a 3D attention mask from a 2D tensor mask.
             # Sizes are [batch_size, 1, 1, to_seq_length]
             # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
@@ -519,14 +525,12 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head
 
-    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
         # only last token for inputs_ids if past is defined in kwargs
-        if "past" in kwargs and kwargs["past"]:
+        if past:
             input_ids = input_ids[:, -1].unsqueeze(-1)
 
-        inputs = {"input_ids": input_ids}
-        inputs.update(kwargs)
-        return inputs
+        return {"input_ids": input_ids, "past": past}
 
     @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
     def forward(

src/transformers/modeling_tf_ctrl.py

@@ -105,8 +105,8 @@ class TFMultiHeadAttention(tf.keras.layers.Layer):
         v = self.split_into_heads(v, batch_size)
         if layer_past is not None:
             past_key, past_value = tf.unstack(layer_past, axis=1)
-            k = tf.concat((past_key, k), dim=-2)
-            v = tf.concat((past_value, v), dim=-2)
+            k = tf.concat((past_key, k), axis=-2)
+            v = tf.concat((past_value, v), axis=-2)
         present = tf.stack((k, v), axis=1)
 
         output = scaled_dot_product_attention(q, k, v, mask, attention_mask, head_mask)
@@ -505,6 +505,13 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head.input_embeddings
 
+    def prepare_inputs_for_generation(self, inputs, past, **kwargs):
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            inputs = tf.expand_dims(inputs[:, -1], -1)
+
+        return {"inputs": inputs, "past": past}
+
     @add_start_docstrings_to_callable(CTRL_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""

src/transformers/modeling_tf_gpt2.py

@@ -500,6 +500,13 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
     def get_output_embeddings(self):
         return self.transformer.wte
 
+    def prepare_inputs_for_generation(self, inputs, past, **kwargs):
+        # only last token for inputs_ids if past is defined in kwargs
+        if past:
+            inputs = tf.expand_dims(inputs[:, -1], -1)
+
+        return {"inputs": inputs, "past": past}
+
     @add_start_docstrings_to_callable(GPT2_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""

src/transformers/modeling_tf_transfo_xl.py

@@ -826,3 +826,12 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
             outputs = [softmax_output] + outputs
 
         return outputs  # logits, new_mems, (all hidden states), (all attentions)
+
+    def prepare_inputs_for_generation(self, inputs, past, **model_kwargs):
+        inputs = {"inputs": inputs}
+
+        # if past is defined in model kwargs then use it for faster decoding
+        if past:
+            inputs["mems"] = past
+
+        return inputs

src/transformers/modeling_tf_utils.py

@@ -384,6 +384,432 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin):
 
         return model
 
+    def prepare_inputs_for_generation(self, inputs, **kwargs):
+        return {"inputs": inputs}
+
+    def _do_output_past(self, outputs):
+        has_output_past = hasattr(self.config, "output_past") and self.config.output_past
+        has_mem_len = hasattr(self.config, "mem_len") and self.config.mem_len
+
+        if has_output_past and not has_mem_len and len(outputs) > 1:
+            return True
+        elif has_mem_len and self.config.mem_len > 0 and len(outputs) > 1:
+            return True
+
+        return False
+
+    def generate(
+        self,
+        input_ids=None,
+        max_length=None,
+        do_sample=True,
+        num_beams=None,
+        temperature=None,
+        top_k=None,
+        top_p=None,
+        repetition_penalty=None,
+        bos_token_id=None,
+        pad_token_id=None,
+        eos_token_ids=None,
+        length_penalty=None,
+        num_return_sequences=None,
+    ):
+        r""" Generates sequences for models with a LM head. The method currently supports greedy or penalized greedy decoding, sampling with top-k or nucleus sampling
+        and beam-search.
+
+        Adapted in part from `Facebook's XLM beam search code`_.
+
+        .. _`Facebook's XLM beam search code`:
+           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529
+
+
+        Parameters:
+
+            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
+                The sequence used as a prompt for the generation. If `None` the method initializes
+                it as an empty `torch.LongTensor` of shape `(1,)`.
+
+            max_length: (`optional`) int
+                The max length of the sequence to be generated.  Between 1 and infinity. Default to 20.
+
+            do_sample: (`optional`) bool
+                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `True`.
+
+            num_beams: (`optional`) int
+                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.
+
+            temperature: (`optional`) float
+                The value used to module the next token probabilities. Must be strictely positive. Default to 1.0.
+
+            top_k: (`optional`) int
+                The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
+
+            top_p: (`optional`) float
+                The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
+
+            repetition_penalty: (`optional`) float
+                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.
+
+            bos_token_id: (`optional`) int
+                Beginning of sentence token if no prompt is provided. Default to 0.
+
+            eos_token_ids: (`optional`) int or list of int
+                End of sequence token or list of tokens to stop the generation. Default to 0.
+            length_penalty: (`optional`) float
+                Exponential penalty to the length. Default to 1.
+
+            num_return_sequences: (`optional`) int
+                The number of independently computed returned sequences for each element in the batch. Default to 1.
+
+        Return:
+
+            output: `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
+                sequence_length is either equal to max_length or shorter if all batches finished early due to the `eos_token_id`
+
+        Examples::
+
+            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
+            outputs = model.generate(max_length=40, bos_token_id=tokenizer.bos_token_id, eos_token_ids=tokenizer.eos_token_id, do_sample=False)  # do greedy decoding
+            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('openai-gpt')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('openai-gpt')    # Download model and configuration from S3 and cache.
+            input_context = 'The dog'
+            input_ids = torch.tensor(tokenizer.encode(input_context)).unsqueeze(0)  # encode input context
+            outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5)  # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
+            for i in range(3): #  3 output sequences were generated
+                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
+            input_context = 'The dog'
+            input_ids = torch.tensor(tokenizer.encode(input_context)).unsqueeze(0)  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, bos_token_id=tokenizer.bos_token_id, pad_token_id=tokenizer.pad_token_id, eos_token_ids=tokenizer.eos_token_id, num_return_sequences=3)  # 3 generate sequences using by sampling
+            for i in range(3): #  3 output sequences were generated
+                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('ctrl')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('ctrl')    # Download model and configuration from S3 and cache.
+            input_context = 'Legal My neighbor is'  # "Legal" is one of the control codes for ctrl
+            input_ids = torch.tensor(tokenizer.encode(input_context)).unsqueeze(0)  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
+            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+
+        """
+
+        # We cannot generate if the model does not have a LM head
+        if self.get_output_embeddings() is None:
+            raise AttributeError(
+                "You tried to generate sequences with a model that does not have a LM Head."
+                "Please use another model class (e.g. `OpenAIGPTLMHeadModel`, `XLNetLMHeadModel`, `GPT2LMHeadModel`, `CTRLLMHeadModel`, `T5WithLMHeadModel`, `TransfoXLLMHeadModel`)"
+            )
+
+        max_length = max_length if max_length is not None else self.config.max_length
+        do_sample = do_sample if do_sample is not None else self.config.do_sample
+        num_beams = num_beams if num_beams is not None else self.config.num_beams
+        temperature = temperature if temperature is not None else self.config.temperature
+        top_k = top_k if top_k is not None else self.config.top_k
+        top_p = top_p if top_p is not None else self.config.top_p
+        repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
+        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
+        pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
+        eos_token_ids = eos_token_ids if eos_token_ids is not None else self.config.eos_token_ids
+        length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
+        num_return_sequences = (
+            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
+        )
+
+        if input_ids is not None:
+            batch_size = shape_list(input_ids)[0]  # overriden by the input batch_size
+        else:
+            batch_size = 1
+        if isinstance(eos_token_ids, int):
+            eos_token_ids = [eos_token_ids]
+
+        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictely positive integer."
+        assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
+        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictely positive integer."
+        assert temperature > 0, "`temperature` should be strictely positive."
+        assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
+        assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
+        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
+        assert input_ids is not None or (
+            isinstance(bos_token_id, int) and bos_token_id >= 0
+        ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
+        assert pad_token_id is None or (
+            isinstance(pad_token_id, int) and (pad_token_id >= 0)
+        ), "`pad_token_id` should be a positive integer."
+        assert (eos_token_ids is None) or (
+            isinstance(eos_token_ids, (list, tuple)) and ((isinstance(e, int) and e >= 0) for e in eos_token_ids)
+        ), "`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
+        assert length_penalty > 0, "`length_penalty` should be strictely positive."
+        assert (
+            isinstance(num_return_sequences, int) and num_return_sequences > 0
+        ), "`num_return_sequences` should be a strictely positive integer."
+
+        if input_ids is None:
+            assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
+                "you should either supply a context to complete as `input_ids` input "
+                "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
+            )
+            input_ids = tf.fill((batch_size, 1), bos_token_id)
+        else:
+            assert len(shape_list(input_ids)) == 2, "Input prompt should be of shape (batch_size, sequence length)."
+
+        if pad_token_id is None and eos_token_ids is not None:
+            logger.warning(
+                "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_ids[0])
+            )
+            pad_token_id = eos_token_ids[0]
+
+        # current position and vocab size
+        cur_len = shape_list(input_ids)[1]
+        vocab_size = self.config.vocab_size
+
+        if num_return_sequences != 1:
+            # Expand input to num return sequences
+            input_ids = tf.broadcast_to(tf.expand_dims(input_ids, 1), (batch_size, num_return_sequences, cur_len))
+            effective_batch_size = batch_size * num_return_sequences
+            input_ids = tf.reshape(input_ids, (effective_batch_size, cur_len))
+        else:
+            effective_batch_size = batch_size
+
+        if num_beams > 1:
+            output = self._generate_beam_search(
+                input_ids,
+                cur_len,
+                max_length,
+                do_sample,
+                temperature,
+                top_k,
+                top_p,
+                repetition_penalty,
+                pad_token_id,
+                eos_token_ids,
+                effective_batch_size,
+                length_penalty,
+                num_beams,
+                vocab_size,
+            )
+        else:
+            output = self._generate_no_beam_search(
+                input_ids,
+                cur_len,
+                max_length,
+                do_sample,
+                temperature,
+                top_k,
+                top_p,
+                repetition_penalty,
+                pad_token_id,
+                eos_token_ids,
+                effective_batch_size,
+            )
+
+        return output
+
+    def _generate_no_beam_search(
+        self,
+        input_ids,
+        cur_len,
+        max_length,
+        do_sample,
+        temperature,
+        top_k,
+        top_p,
+        repetition_penalty,
+        pad_token_id,
+        eos_token_ids,
+        batch_size,
+    ):
+        """ Generate sequences for each example without beam search (num_beams == 1).
+            All returned sequence are generated independantly.
+        """
+
+        def _create_next_token_logits_penalties(input_ids, logits):
+            # create logit penalties for already seen input_ids
+            token_penalties = np.ones(shape_list(logits))
+            prev_input_ids = [np.unique(input_id) for input_id in input_ids.numpy()]
+            for i, prev_input_id in enumerate(prev_input_ids):
+                logit_penalized = logits[i].numpy()[prev_input_id]
+                # if previous logit score is < 0 then multiply repetition penalty else divide
+                logit_penalized[logit_penalized < 0] = repetition_penalty
+                logit_penalized[logit_penalized > 0] = 1 / repetition_penalty
+                np.put(token_penalties[i], prev_input_id, logit_penalized)
+            return tf.convert_to_tensor(token_penalties, dtype=tf.float32)
+
+        # current position / max lengths / length of generated sentences / unfinished sentences
+        unfinished_sents = tf.ones_like(input_ids[:, 0])
+        sent_lengths = tf.ones_like(input_ids[:, 0]) * max_length
+
+        past = None
+
+        while cur_len < max_length:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, past=past)
+            outputs = self(**model_inputs)
+            next_token_logits = outputs[0][:, -1, :]
+
+            # if model has past, then set the past variable to speed up decoding
+            if self._do_output_past(outputs):
+                past = outputs[1]
+
+            # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
+            if repetition_penalty != 1.0:
+                next_token_logits_penalties = _create_next_token_logits_penalties(input_ids, next_token_logits)
+                next_token_logits = tf.math.multiply(next_token_logits, next_token_logits_penalties)
+
+            if do_sample:
+                # Temperature (higher temperature => more likely to sample low probability tokens)
+                if temperature != 1.0:
+                    next_token_logits = next_token_logits / temperature
+                # Top-p/top-k filtering
+                next_token_logits = tf_top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
+                # Sample
+                next_token = tf.squeeze(
+                    tf.random.categorical(next_token_logits, dtype=tf.int32, num_samples=1), axis=1
+                )
+            else:
+                # Greedy decoding
+                next_token = tf.math.argmax(next_token_logits, axis=-1, output_type=tf.int32)
+
+            # update generations and finished sentences
+            if eos_token_ids is not None:
+                # pad finished sentences if eos_token_ids exist
+                tokens_to_add = next_token * unfinished_sents + (pad_token_id) * (1 - unfinished_sents)
+            else:
+                tokens_to_add = next_token
+
+            input_ids = tf.concat([input_ids, tf.expand_dims(tokens_to_add, -1)], 1)
+
+            if eos_token_ids is not None:
+                for eos_token_id in eos_token_ids:
+                    eos_in_sents = tokens_to_add == eos_token_id
+                    # if sentence is unfinished and the token to add is eos, sent_lengths is filled with current length
+                    is_sents_unfinished_and_token_to_add_is_eos = tf.math.multiply(
+                        unfinished_sents, tf.cast(eos_in_sents, tf.int32)
+                    )
+                    sent_lengths = (
+                        sent_lengths * (1 - is_sents_unfinished_and_token_to_add_is_eos)
+                        + cur_len * is_sents_unfinished_and_token_to_add_is_eos
+                    )
+
+                    # unfinished_sents is set to zero if eos in sentence
+                    unfinished_sents -= is_sents_unfinished_and_token_to_add_is_eos
+
+            cur_len = cur_len + 1
+
+            # stop when there is a </s> in each sentence, or if we exceed the maximul length
+            if tf.math.reduce_max(unfinished_sents) == 0:
+                break
+
+        # if there are different sentences lengths in the batch, some batches have to be padded
+        min_sent_length = tf.math.reduce_min(sent_lengths)
+        max_sent_length = tf.math.reduce_max(sent_lengths)
+        if min_sent_length != max_sent_length:
+            assert pad_token_id is not None, "`Pad_token_id` has to be defined if batches have different lengths"
+            # finished sents are filled with pad_token
+            padding = tf.ones([batch_size, max_sent_length.numpy()], dtype=tf.int32) * pad_token_id
+
+            # create length masks for tf.where operation
+            broad_casted_sent_lengths = tf.broadcast_to(
+                tf.expand_dims(sent_lengths, -1), [batch_size, max_sent_length]
+            )
+            broad_casted_range = tf.transpose(
+                tf.broadcast_to(tf.expand_dims(tf.range(max_length), -1), [max_length, batch_size])
+            )
+
+            decoded = tf.where(broad_casted_range < broad_casted_sent_lengths, input_ids, padding)
+        else:
+            decoded = input_ids
+
+        return decoded
+
+    def _generate_beam_search(
+        self,
+        input_ids,
+        cur_len,
+        max_length,
+        do_sample,
+        temperature,
+        top_k,
+        top_p,
+        repetition_penalty,
+        pad_token_id,
+        eos_token_ids,
+        batch_size,
+        length_penalty,
+        num_beams,
+        vocab_size,
+    ):
+        pass
+
+
+def tf_top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
+    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+        Args:
+            logits: logits distribution shape (batch size, vocabulary size)
+            if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+            if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+            Make sure we keep at least min_tokens_to_keep per batch example in the output
+        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    logits_shape = shape_list(logits)
+
+    if top_k > 0:
+        top_k = min(max(top_k, min_tokens_to_keep), logits_shape[-1])  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < tf.math.top_k(logits, k=top_k)[0][..., -1, None]
+        logits = set_tensor_by_indices_to_value(logits, indices_to_remove, filter_value)
+
+    if top_p < 1.0:
+        sorted_indices = tf.argsort(logits, direction="DESCENDING")
+        sorted_logits = tf.gather(
+            logits, sorted_indices, axis=-1, batch_dims=1
+        )  # expects logits to be of dim (batch_size, vocab_size)
+
+        cumulative_probs = tf.math.cumsum(tf.nn.softmax(sorted_logits, axis=-1), axis=-1)
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove = tf.concat(
+                [
+                    tf.zeros_like(sorted_indices_to_remove[:, :min_tokens_to_keep]),
+                    sorted_indices_to_remove[:, min_tokens_to_keep:],
+                ],
+                -1,
+            )
+
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove = tf.roll(sorted_indices_to_remove, 1, axis=-1)
+        sorted_indices_to_remove = tf.concat(
+            [tf.zeros_like(sorted_indices_to_remove[:, :1]), sorted_indices_to_remove[:, 1:]], -1,
+        )
+        # scatter sorted tensors to original indexing
+        indices_to_remove = scatter_values_on_batch_indices(sorted_indices_to_remove, sorted_indices)
+        logits = set_tensor_by_indices_to_value(logits, indices_to_remove, filter_value)
+    return logits
+
+
+def scatter_values_on_batch_indices(values, batch_indices):
+    shape = shape_list(batch_indices)
+    # broadcast batch dim to shape
+    broad_casted_batch_dims = tf.reshape(tf.broadcast_to(tf.expand_dims(tf.range(shape[0]), axis=-1), shape), [1, -1])
+    # transform batch_indices to pair_indices
+    pair_indices = tf.transpose(tf.concat([broad_casted_batch_dims, tf.reshape(batch_indices, [1, -1])], 0))
+    # scatter values to pair indices
+    return tf.scatter_nd(pair_indices, tf.reshape(values, [-1]), shape)
+
+
+def set_tensor_by_indices_to_value(tensor, indices, value):
+    # create value_tensor since tensor value assignment is not possible in TF
+    value_tensor = tf.zeros_like(tensor) + value
+    return tf.where(indices, value_tensor, tensor)
+
 
 class TFConv1D(tf.keras.layers.Layer):
     def __init__(self, nf, nx, initializer_range=0.02, **kwargs):

src/transformers/modeling_tf_xlm.py

@@ -657,6 +657,20 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
     def get_output_embeddings(self):
         return self.pred_layer.input_embeddings
 
+    def prepare_inputs_for_generation(self, inputs, **kwargs):
+        mask_token_id = self.config.mask_token_id
+        lang_id = self.config.lang_id
+
+        effective_batch_size = inputs.shape[0]
+        mask_token = tf.ones((effective_batch_size, 1), dtype=tf.int32) * mask_token_id
+        inputs = tf.concat([inputs, mask_token], axis=1)
+
+        if lang_id is not None:
+            langs = tf.ones_like(inputs) * lang_id
+        else:
+            langs = None
+        return {"inputs": inputs, "langs": langs}
+
     @add_start_docstrings_to_callable(XLM_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""

src/transformers/modeling_tf_xlnet.py

@@ -837,6 +837,32 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_loss.input_embeddings
 
+    def prepare_inputs_for_generation(self, inputs, past, **model_kwargs):
+        # Add dummy token at the end (no attention on this one)
+
+        effective_batch_size = inputs.shape[0]
+        dummy_token = tf.zeros((effective_batch_size, 1), dtype=tf.int32)
+        inputs = tf.concat([inputs, dummy_token], axis=1)
+
+        # Build permutation mask so that previous tokens don't see last token
+        sequence_length = inputs.shape[1]
+        perm_mask = tf.zeros((effective_batch_size, sequence_length, sequence_length - 1), dtype=tf.float32)
+        perm_mask_seq_end = tf.ones((effective_batch_size, sequence_length, 1), dtype=tf.float32)
+        perm_mask = tf.concat([perm_mask, perm_mask_seq_end], axis=-1)
+
+        # We'll only predict the last token
+        target_mapping = tf.zeros((effective_batch_size, 1, sequence_length - 1), dtype=tf.float32)
+        target_mapping_seq_end = tf.ones((effective_batch_size, 1, 1), dtype=tf.float32)
+        target_mapping = tf.concat([target_mapping, target_mapping_seq_end], axis=-1)
+
+        inputs = {"inputs": inputs, "perm_mask": perm_mask, "target_mapping": target_mapping}
+
+        # if past is defined in model kwargs then use it for faster decoding
+        if past:
+            inputs["mems"] = past
+
+        return inputs
+
     @add_start_docstrings_to_callable(XLNET_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""

src/transformers/modeling_transfo_xl.py

@@ -935,11 +935,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         else:
             return self.crit.out_layers[-1]
 
-    def prepare_inputs_for_generation(self, input_ids, **model_kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past, **model_kwargs):
         inputs = {"input_ids": input_ids}
 
         # if past is defined in model kwargs then use it for faster decoding
-        if "past" in model_kwargs and model_kwargs["past"]:
-            inputs["mems"] = model_kwargs["past"]
+        if past:
+            inputs["mems"] = past
 
         return inputs

src/transformers/modeling_utils.py

@@ -171,7 +171,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         else:
             output_embeddings.weight = input_embeddings.weight
 
-        if hasattr(output_embeddings, "bias") and output_embeddings.bias is not None:
+        if getattr(output_embeddings, "bias", None) is not None:
             output_embeddings.bias.data = torch.nn.functional.pad(
                 output_embeddings.bias.data,
                 (0, output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0]),
@@ -558,14 +558,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                         model.__class__.__name__, "\n\t".join(error_msgs)
                     )
                 )
-
         model.tie_weights()  # make sure word embedding weights are still tied if needed
 
         # Set model in evaluation mode to desactivate DropOut modules by default
         model.eval()
 
         if output_loading_info:
-            loading_info = {"missing_keys": missing_keys, "unexpected_keys": unexpected_keys, "error_msgs": error_msgs}
+            loading_info = {
+                "missing_keys": missing_keys,
+                "unexpected_keys": unexpected_keys,
+                "error_msgs": error_msgs,
+            }
             return model, loading_info
 
         return model
@@ -574,16 +577,25 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         return {"input_ids": input_ids}
 
     def _do_output_past(self, outputs):
-        has_output_past = hasattr(self.config, "output_past") and self.config.output_past
-        has_mem_len = hasattr(self.config, "mem_len") and self.config.mem_len
-
-        if has_output_past and not has_mem_len and len(outputs) > 1:
+        """During generation, decide whether to pass the `past` variable to the next forward pass."""
+        has_output_past = getattr(self.config, "output_past", False)
+        mem_len = getattr(self.config, "mem_len", 0)
+        if len(outputs) <= 1:
+            return False
+        if mem_len > 0 or has_output_past:
             return True
-        elif has_mem_len and self.config.mem_len > 0 and len(outputs) > 1:
-            return True
-
         return False
 
+    def enforce_repetition_penalty_(self, lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty):
+        """repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858). """
+        for i in range(batch_size * num_beams):
+            for previous_token in set(prev_output_tokens[i].tolist()):
+                # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if lprobs[i, previous_token] < 0:
+                    lprobs[i, previous_token] *= repetition_penalty
+                else:
+                    lprobs[i, previous_token] /= repetition_penalty
+
     @torch.no_grad()
     def generate(
         self,
@@ -626,7 +638,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                 Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.
 
             temperature: (`optional`) float
-                The value used to module the next token probabilities. Must be strictely positive. Default to 1.0.
+                The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.
 
             top_k: (`optional`) int
                 The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
@@ -714,10 +726,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         if isinstance(eos_token_ids, int):
             eos_token_ids = [eos_token_ids]
 
-        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictely positive integer."
+        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
         assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
-        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictely positive integer."
-        assert temperature > 0, "`temperature` should be strictely positive."
+        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
+        assert temperature > 0, "`temperature` should be strictly positive."
         assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
         assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
         assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
@@ -730,10 +742,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         assert (eos_token_ids is None) or (
             isinstance(eos_token_ids, (list, tuple)) and ((isinstance(e, int) and e >= 0) for e in eos_token_ids)
         ), "`eos_token_ids` should be a positive integer or a list/tuple of positive integers."
-        assert length_penalty > 0, "`length_penalty` should be strictely positive."
+        assert length_penalty > 0, "`length_penalty` should be strictly positive."
         assert (
             isinstance(num_return_sequences, int) and num_return_sequences > 0
-        ), "`num_return_sequences` should be a strictely positive integer."
+        ), "`num_return_sequences` should be a strictly positive integer."
 
         if input_ids is None:
             assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
@@ -746,6 +758,19 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         else:
             assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
 
+        if do_sample is False:
+            if num_beams == 1:
+                # no_beam_search greedy generation conditions
+                assert (
+                    num_return_sequences == 1
+                ), "Greedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1"
+
+            else:
+                # beam_search greedy generation conditions
+                assert (
+                    num_beams >= num_return_sequences
+                ), "Greedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequences"
+
         if pad_token_id is None and eos_token_ids is not None:
             logger.warning(
                 "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_ids[0])
@@ -756,12 +781,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         cur_len = input_ids.shape[1]
         vocab_size = self.config.vocab_size
 
-        if num_return_sequences != 1:
+        if num_return_sequences != 1 and do_sample:
             # Expand input to num return sequences
             input_ids = input_ids.unsqueeze(1).expand(batch_size, num_return_sequences, cur_len)
             input_ids = input_ids.contiguous().view(
                 batch_size * num_return_sequences, cur_len
-            )  # (batch_size * num_return_sequences, cur_len)
+            )  # shape: (batch_size * num_return_sequences, cur_len)
             effective_batch_size = batch_size * num_return_sequences
         else:
             effective_batch_size = batch_size
@@ -779,6 +804,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                 pad_token_id,
                 eos_token_ids,
                 effective_batch_size,
+                num_return_sequences,
                 length_penalty,
                 num_beams,
                 vocab_size,
@@ -818,13 +844,14 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
             All returned sequence are generated independantly.
         """
         # current position / max lengths / length of generated sentences / unfinished sentences
+
         unfinished_sents = input_ids.new(batch_size).fill_(1)
         sent_lengths = input_ids.new(batch_size).fill_(max_length)
 
         past = None
-
         while cur_len < max_length:
             model_inputs = self.prepare_inputs_for_generation(input_ids, past=past)
+
             outputs = self(**model_inputs)
             next_token_logits = outputs[0][:, -1, :]
 
@@ -834,13 +861,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
             # repetition penalty from CTRL paper (https://arxiv.org/abs/1909.05858)
             if repetition_penalty != 1.0:
-                for i in range(batch_size):
-                    for previous_token in set(input_ids[i].tolist()):
-                        # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
-                        if next_token_logits[i, previous_token] < 0:
-                            next_token_logits[i, previous_token] *= repetition_penalty
-                        else:
-                            next_token_logits[i, previous_token] /= repetition_penalty
+                self.enforce_repetition_penalty_(next_token_logits, batch_size, 1, input_ids, repetition_penalty)
 
             if do_sample:
                 # Temperature (higher temperature => more likely to sample low probability tokens)
@@ -904,13 +925,16 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         pad_token_id,
         eos_token_ids,
         batch_size,
+        num_return_sequences,
         length_penalty,
         num_beams,
         vocab_size,
     ):
         """ Generate sequences for each example with beam search.
         """
+
         # Expand input to num beams
+        # assert input_ids.shape == (batch_size * num_beams, cur_len)
         input_ids = input_ids.unsqueeze(1).expand(batch_size, num_beams, cur_len)
         input_ids = input_ids.contiguous().view(batch_size * num_beams, cur_len)  # (batch_size * num_beams, cur_len)
 
@@ -921,7 +945,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
         # scores for each sentence in the beam
         beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
-        beam_scores[:, 1:] = -1e9
+
+        # Greedy decoding it is made sure that only words of the first beam are considered to avoid sampling the exact same words three times
+        if do_sample is False:
+            beam_scores[:, 1:] = -1e9
         beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)
 
         # cache compute states
@@ -941,31 +968,34 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
             # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
             if repetition_penalty != 1.0:
-                for i in range(batch_size * num_beams):
-                    for previous_token in set(input_ids[i].tolist()):
-                        # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
-                        if scores[i, previous_token] < 0:
-                            scores[i, previous_token] *= repetition_penalty
-                        else:
-                            scores[i, previous_token] /= repetition_penalty
+                self.enforce_repetition_penalty_(scores, batch_size, num_beams, input_ids, repetition_penalty)
 
             if do_sample:
                 # Temperature (higher temperature => more likely to sample low probability tokens)
                 if temperature != 1.0:
                     scores = scores / temperature
+
+                scores = F.log_softmax(scores, dim=-1)  # (batch_size * num_beams, vocab_size)
+                _scores = scores + beam_scores[:, None].expand_as(scores)  # (batch_size * num_beams, vocab_size)
+
                 # Top-p/top-k filtering
-                scores = top_k_top_p_filtering(
-                    scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
+                _scores = top_k_top_p_filtering(
+                    _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
                 )  # (batch_size * num_beams, vocab_size)
+
+                # re-organize to group the beam together to sample from all beam_idxs
+                _scores = _scores.contiguous().view(
+                    batch_size, num_beams * vocab_size
+                )  # (batch_size, num_beams * vocab_size)
+
                 # Sample 2 next words for each beam (so we have some spare tokens and match output of greedy beam search)
-                next_words = torch.multinomial(F.softmax(scores, dim=-1), num_samples=2)  # (batch_size * num_beams, 2)
+                next_words = torch.multinomial(
+                    F.softmax(_scores, dim=-1), num_samples=2 * num_beams
+                )  # (batch_size, num_beams * 2)
+
                 # Compute next scores
-                _scores = F.log_softmax(scores, dim=-1)  # (batch_size * num_beams, vocab_size)
-                _scores = torch.gather(_scores, -1, next_words)  # (batch_size * num_beams, 2)
-                next_scores = _scores + beam_scores[:, None].expand_as(_scores)  # (batch_size * num_beams, 2)
-                # Match shape of greedy beam search
-                next_words = next_words.view(batch_size, 2 * num_beams)  # (batch_size, 2 * num_beams)
-                next_scores = next_scores.view(batch_size, 2 * num_beams)  # (batch_size, 2 * num_beams)
+                next_scores = torch.gather(_scores, -1, next_words)  # (batch_size, num_beams * 2)
+
             else:
                 # do greedy beam search
                 scores = F.log_softmax(scores, dim=-1)  # (batch_size * num_beams, vocab_size)
@@ -1012,7 +1042,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                     # add to generated hypotheses if end of sentence or last iteration
                     if eos_token_ids is not None and word_id.item() in eos_token_ids:
                         generated_hyps[batch_idx].add(
-                            input_ids[batch_idx * num_beams + beam_id, :cur_len].clone(), score.item()
+                            input_ids[batch_idx * num_beams + beam_id, :cur_len].clone(), score.item(),
                         )
                     else:
                         # add next predicted word if it is not eos_token
@@ -1039,16 +1069,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
             # re-order internal states
             if past:
-                reordered_past = []
-                for layer_past in past:
-                    # get the correct batch idx from layer past batch dim
-                    # batch dim of `past` and `mems` is at 2nd position
-                    reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
-                    reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
-                    # check that shape matches
-                    assert reordered_layer_past.shape == layer_past.shape
-                    reordered_past.append(reordered_layer_past)
-                past = tuple(reordered_past)
+                past = self._reorder_cache(past, beam_idx)
 
             # update current length
             cur_len = cur_len + 1
@@ -1069,20 +1090,28 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
                         input_ids[batch_idx * num_beams + beam_id, :cur_len].clone(), score.item()
                     )
 
+        # depending on whether greedy generation is wanted or not define different output_batch_size and output_num_return_sequences_per_batch
+        output_batch_size = batch_size if do_sample else batch_size * num_return_sequences
+        output_num_return_sequences_per_batch = 1 if do_sample else num_return_sequences
+
         # select the best hypotheses
-        sent_lengths = input_ids.new(batch_size)
+        sent_lengths = input_ids.new(output_batch_size)
         best = []
 
+        # retrieve best hypotheses
         for i, hypotheses in enumerate(generated_hyps):
-            best_hyp = max(hypotheses.beams, key=lambda x: x[0])[1]
-            sent_lengths[i] = len(best_hyp)
-            best.append(best_hyp)
+            sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0])
+            for j in range(output_num_return_sequences_per_batch):
+                effective_batch_idx = output_num_return_sequences_per_batch * i + j
+                best_hyp = sorted_hyps.pop()[1]
+                sent_lengths[effective_batch_idx] = len(best_hyp)
+                best.append(best_hyp)
 
         # shorter batches are filled with pad_token
         if sent_lengths.min().item() != sent_lengths.max().item():
             assert pad_token_id is not None, "`Pad_token_id` has to be defined"
             sent_max_len = min(sent_lengths.max().item() + 1, max_length)
-            decoded = input_ids.new(batch_size, sent_max_len).fill_(pad_token_id)
+            decoded = input_ids.new(output_batch_size, sent_max_len).fill_(pad_token_id)
 
             # fill with hypothesis and eos_token_id if necessary
             for i, hypo in enumerate(best):
@@ -1096,6 +1125,20 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
 
         return decoded
 
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = []
+        for layer_past in past:
+            # get the correct batch idx from layer past batch dim
+            # batch dim of `past` and `mems` is at 2nd position
+            reordered_layer_past = [layer_past[:, i].unsqueeze(1).clone().detach() for i in beam_idx]
+            reordered_layer_past = torch.cat(reordered_layer_past, dim=1)
+            # check that shape matches
+            assert reordered_layer_past.shape == layer_past.shape
+            reordered_past.append(reordered_layer_past)
+        past = tuple(reordered_past)
+        return past
+
 
 def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
     """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
@@ -1164,17 +1207,22 @@ class BeamHypotheses(object):
             else:
                 self.worst_score = min(score, self.worst_score)
 
-    def is_done(self, best_sum_logprobs):
+    def is_done(self, best_sum_logprobs, cur_len=None):
         """
         If there are enough hypotheses and that none of the hypotheses being generated
         can become better than the worst one in the heap, then we are done with this sentence.
         """
+
         if len(self) < self.num_beams:
             return False
         elif self.early_stopping:
             return True
         else:
-            return self.worst_score >= best_sum_logprobs / self.max_length ** self.length_penalty
+            if cur_len is None:
+                cur_len = self.max_length
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret
 
 
 class Conv1D(nn.Module):

src/transformers/modeling_xlnet.py

@@ -935,7 +935,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_loss
 
-    def prepare_inputs_for_generation(self, input_ids, **model_kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past, **model_kwargs):
         # Add dummy token at the end (no attention on this one)
 
         effective_batch_size = input_ids.shape[0]
@@ -958,8 +958,8 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         inputs = {"input_ids": input_ids, "perm_mask": perm_mask, "target_mapping": target_mapping}
 
         # if past is defined in model kwargs then use it for faster decoding
-        if "past" in model_kwargs and model_kwargs["past"]:
-            inputs["mems"] = model_kwargs["past"]
+        if past:
+            inputs["mems"] = past
 
         return inputs
 

src/transformers/pipelines.py

@@ -28,6 +28,7 @@ from typing import Dict, List, Optional, Tuple, Union
 import numpy as np
 
 from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, AutoConfig
+from .configuration_bart import BartConfig
 from .configuration_distilbert import DistilBertConfig
 from .configuration_roberta import RobertaConfig
 from .configuration_utils import PretrainedConfig
@@ -279,6 +280,9 @@ class _ScikitCompat(ABC):
 
 class Pipeline(_ScikitCompat):
     """
+    The Pipeline class is the class from which all pipelines inherit. Refer to this class for methods shared across
+    different pipelines.
+
     Base class implementing pipelined operations.
     Pipeline workflow is defined as a sequence of the following operations:
         Input -> Tokenization -> Model Inference -> Post-Processing (Task dependent) -> Output
@@ -292,39 +296,49 @@ class Pipeline(_ScikitCompat):
     pickle format.
 
     Arguments:
-        **model**: ``(str, PretrainedModel, TFPretrainedModel)``:
-            Reference to the model to use through this pipeline.
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
 
-        **tokenizer**: ``(str, PreTrainedTokenizer)``:
-            Reference to the tokenizer to use through this pipeline.
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
 
-        **args_parser**: ``ArgumentHandler``:
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
             Reference to the object in charge of parsing supplied pipeline parameters.
-
-        **device**: ``int``:
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
             Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
             on the associated CUDA device id.
-
-        **binary_output** ``bool`` (default: False):
+        binary_output (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Flag indicating if the output the pipeline should happen in a binary format (i.e. pickle) or as raw text.
 
     Return:
+        :obj:`List` or :obj:`Dict`:
         Pipeline returns list or dictionary depending on:
-         - Does the user provided multiple sample
-         - The pipeline expose multiple fields in the output object
 
-    Examples:
-        nlp = pipeline('ner')
-        nlp = pipeline('ner', model='...', config='...', tokenizer='...')
-        nlp = NerPipeline(model='...', config='...', tokenizer='...')
-        nlp = QuestionAnsweringPipeline(model=AutoModel.from_pretrained('...'), tokenizer='...')
+         - Whether the user supplied multiple samples
+         - Whether the pipeline exposes multiple fields in the output object
     """
 
     default_input_names = None
+    task = None
 
     def __init__(
         self,
-        model,
+        model: Optional = None,
         tokenizer: PreTrainedTokenizer = None,
         modelcard: Optional[ModelCard] = None,
         framework: Optional[str] = None,
@@ -336,6 +350,8 @@ class Pipeline(_ScikitCompat):
         if framework is None:
             framework = get_framework()
 
+        model, tokenizer = self.get_defaults(model, tokenizer, framework)
+
         self.model = model
         self.tokenizer = tokenizer
         self.modelcard = modelcard
@@ -412,7 +428,7 @@ class Pipeline(_ScikitCompat):
         """
         args = ["input_ids", "attention_mask"]
 
-        if not isinstance(self.model.config, (DistilBertConfig, XLMConfig, RobertaConfig)):
+        if not isinstance(self.model.config, (DistilBertConfig, XLMConfig, RobertaConfig, BartConfig)):
             args += ["token_type_ids"]
 
         # PR #1548 (CLI) There is an issue with attention_mask
@@ -467,15 +483,74 @@ class Pipeline(_ScikitCompat):
         else:
             return predictions.numpy()
 
+    def get_defaults(self, model, tokenizer, framework):
+        task_defaults = SUPPORTED_TASKS[self.task]
+        if model is None:
+            if framework == "tf":
+                model = task_defaults["tf"].from_pretrained(task_defaults["default"]["model"]["tf"])
+            elif framework == "pt":
+                model = task_defaults["pt"].from_pretrained(task_defaults["default"]["model"]["pt"])
+            else:
+                raise ValueError("Provided framework should be either 'tf' for TensorFlow or 'pt' for PyTorch.")
+
+        if tokenizer is None:
+            default_tokenizer = task_defaults["default"]["tokenizer"]
+            if isinstance(default_tokenizer, tuple):
+                # For tuple we have (tokenizer name, {kwargs})
+                tokenizer = AutoTokenizer.from_pretrained(default_tokenizer[0], **default_tokenizer[1])
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(default_tokenizer)
+
+        return model, tokenizer
+
 
 class FeatureExtractionPipeline(Pipeline):
     """
-    Feature extraction pipeline using Model head.
+    Feature extraction pipeline using Model head. This pipeline extracts the hidden states from the base transformer,
+    which can be used as features in a downstream tasks.
+
+    This feature extraction pipeline can currently be loaded from the :func:`~transformers.pipeline` method using
+    the following task identifier(s):
+
+    - "feature-extraction", for extracting features of a sequence.
+
+    All models may be used for this pipeline. See a list of all models, including community-contributed models on
+    `huggingface.co/models <https://huggingface.co/models>`__.
+
+    Arguments:
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
+            Reference to the object in charge of parsing supplied pipeline parameters.
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
+            Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
+            on the associated CUDA device id.
     """
 
+    task = "feature-extraction"
+
     def __init__(
         self,
-        model,
+        model: Optional = None,
         tokenizer: PreTrainedTokenizer = None,
         modelcard: Optional[ModelCard] = None,
         framework: Optional[str] = None,
@@ -498,9 +573,49 @@ class FeatureExtractionPipeline(Pipeline):
 
 class TextClassificationPipeline(Pipeline):
     """
-    Text classification pipeline using ModelForTextClassification head.
+    Text classification pipeline using ModelForSequenceClassification head. See the
+    `sequence classification usage <../usage.html#sequence-classification>`__ examples for more information.
+
+    This text classification pipeline can currently be loaded from the :func:`~transformers.pipeline` method using
+    the following task identifier(s):
+
+    - "sentiment-analysis", for classifying sequences according to positive or negative sentiments.
+
+    The models that this pipeline can use are models that have been fine-tuned on a sequence classification task.
+    See the list of available community models fine-tuned on such a task on
+    `huggingface.co/models <https://huggingface.co/models?search=&filter=text-classification>`__.
+
+    Arguments:
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
+            Reference to the object in charge of parsing supplied pipeline parameters.
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
+            Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
+            on the associated CUDA device id.
     """
 
+    task = "sentiment-analysis"
+
     def __call__(self, *args, **kwargs):
         outputs = super().__call__(*args, **kwargs)
         scores = np.exp(outputs) / np.exp(outputs).sum(-1)
@@ -509,12 +624,53 @@ class TextClassificationPipeline(Pipeline):
 
 class FillMaskPipeline(Pipeline):
     """
-    Masked language modeling prediction pipeline using ModelWithLMHead head.
+    Masked language modeling prediction pipeline using ModelWithLMHead head. See the
+    `masked language modeling usage <../usage.html#masked-language-modeling>`__ examples for more information.
+
+    This mask filling pipeline can currently be loaded from the :func:`~transformers.pipeline` method using
+    the following task identifier(s):
+
+    - "fill-mask", for predicting masked tokens in a sequence.
+
+    The models that this pipeline can use are models that have been trained with a masked language modeling objective,
+    which includes the bi-directional models in the library.
+    See the list of available community models on
+    `huggingface.co/models <https://huggingface.co/models?search=&filter=lm-head>`__.
+
+    Arguments:
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
+            Reference to the object in charge of parsing supplied pipeline parameters.
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
+            Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
+            on the associated CUDA device id.
     """
 
+    task = "fill-mask"
+
     def __init__(
         self,
-        model,
+        model: Optional = None,
         tokenizer: PreTrainedTokenizer = None,
         modelcard: Optional[ModelCard] = None,
         framework: Optional[str] = None,
@@ -574,14 +730,57 @@ class FillMaskPipeline(Pipeline):
 
 class NerPipeline(Pipeline):
     """
-    Named Entity Recognition pipeline using ModelForTokenClassification head.
+    Named Entity Recognition pipeline using ModelForTokenClassification head. See the
+    `named entity recognition usage <../usage.html#named-entity-recognition>`__ examples for more information.
+
+    This token recognition pipeline can currently be loaded from the :func:`~transformers.pipeline` method using
+    the following task identifier(s):
+
+    - "ner", for predicting the classes of tokens in a sequence: person, organisation, location or miscellaneous.
+
+    The models that this pipeline can use are models that have been fine-tuned on a token classification task.
+    See the list of available community models fine-tuned on such a task on
+    `huggingface.co/models <https://huggingface.co/models?search=&filter=token-classification>`__.
+
+    Arguments:
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
+            Reference to the object in charge of parsing supplied pipeline parameters.
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
+            Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
+            on the associated CUDA device id.
+
+    Example::
+
+        from transformers import pi
     """
 
     default_input_names = "sequences"
+    task = "ner"
 
     def __init__(
         self,
-        model,
+        model: Optional = None,
         tokenizer: PreTrainedTokenizer = None,
         modelcard: Optional[ModelCard] = None,
         framework: Optional[str] = None,
@@ -636,7 +835,7 @@ class NerPipeline(Pipeline):
                 if self.model.config.id2label[label_idx] not in self.ignore_labels:
                     answer += [
                         {
-                            "word": self.tokenizer.decode([int(input_ids[idx])]),
+                            "word": self.tokenizer.convert_ids_to_tokens(int(input_ids[idx])),
                             "score": score[idx][label_idx].item(),
                             "entity": self.model.config.id2label[label_idx],
                         }
@@ -716,15 +915,54 @@ class QuestionAnsweringArgumentHandler(ArgumentHandler):
 
 class QuestionAnsweringPipeline(Pipeline):
     """
-    Question Answering pipeline using ModelForQuestionAnswering head.
+    Question Answering pipeline using ModelForQuestionAnswering head. See the
+    `question answering usage <../usage.html#question-answering>`__ examples for more information.
+
+    This question answering can currently be loaded from the :func:`~transformers.pipeline` method using
+    the following task identifier(s):
+
+    - "question-answering", for answering questions given a context.
+
+    The models that this pipeline can use are models that have been fine-tuned on a question answering task.
+    See the list of available community models fine-tuned on such a task on
+    `huggingface.co/models <https://huggingface.co/models?search=&filter=question-answering>`__.
+
+    Arguments:
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        modelcard (:obj:`str` or :class:`~transformers.ModelCard`, `optional`, defaults to :obj:`None`):
+            Model card attributed to the model for this pipeline.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+        args_parser (:class:`~transformers.pipelines.ArgumentHandler`, `optional`, defaults to :obj:`None`):
+            Reference to the object in charge of parsing supplied pipeline parameters.
+        device (:obj:`int`, `optional`, defaults to :obj:`-1`):
+            Device ordinal for CPU/GPU supports. Setting this to -1 will leverage CPU, >=0 will run the model
+            on the associated CUDA device id.
     """
 
     default_input_names = "question,context"
+    task = "question-answering"
 
     def __init__(
         self,
-        model,
-        tokenizer: Optional[PreTrainedTokenizer],
+        model: Optional = None,
+        tokenizer: Optional[PreTrainedTokenizer] = None,
         modelcard: Optional[ModelCard] = None,
         framework: Optional[str] = None,
         device: int = -1,
@@ -1003,23 +1241,77 @@ def pipeline(
     model: Optional = None,
     config: Optional[Union[str, PretrainedConfig]] = None,
     tokenizer: Optional[Union[str, PreTrainedTokenizer]] = None,
-    modelcard: Optional[Union[str, ModelCard]] = None,
     framework: Optional[str] = None,
     **kwargs
 ) -> Pipeline:
     """
     Utility factory method to build a pipeline.
-    Pipeline are made of:
-        A Tokenizer instance in charge of mapping raw textual input to token
-        A Model instance
-        Some (optional) post processing for enhancing model's output
 
-    Examples:
+    Pipeline are made of:
+
+        - A Tokenizer instance in charge of mapping raw textual input to token
+        - A Model instance
+        - Some (optional) post processing for enhancing model's output
+
+
+    Args:
+        task (:obj:`str`):
+            The task defining which pipeline will be returned. Currently accepted tasks are:
+
+            - "feature-extraction": will return a :class:`~transformers.FeatureExtractionPipeline`
+            - "sentiment-analysis": will return a :class:`~transformers.TextClassificationPipeline`
+            - "ner": will return a :class:`~transformers.NerPipeline`
+            - "question-answering": will return a :class:`~transformers.QuestionAnsweringPipeline`
+            - "fill-mask": will return a :class:`~transformers.FillMaskPipeline`
+        model (:obj:`str` or :obj:`~transformers.PreTrainedModel` or :obj:`~transformers.TFPreTrainedModel`, `optional`, defaults to :obj:`None`):
+            The model that will be used by the pipeline to make predictions. This can be :obj:`None`, a string
+            checkpoint identifier or an actual pre-trained model inheriting from
+            :class:`~transformers.PreTrainedModel` for PyTorch and :class:`~transformers.TFPreTrainedModel` for
+            TensorFlow.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        config (:obj:`str` or :obj:`~transformers.PretrainedConfig`, `optional`, defaults to :obj:`None`):
+            The configuration that will be used by the pipeline to instantiate the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained model configuration inheriting from
+            :class:`~transformers.PretrainedConfig`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        tokenizer (:obj:`str` or :obj:`~transformers.PreTrainedTokenizer`, `optional`, defaults to :obj:`None`):
+            The tokenizer that will be used by the pipeline to encode data for the model. This can be :obj:`None`,
+            a string checkpoint identifier or an actual pre-trained tokenizer inheriting from
+            :class:`~transformers.PreTrainedTokenizer`.
+
+            If :obj:`None`, the default of the pipeline will be loaded.
+        framework (:obj:`str`, `optional`, defaults to :obj:`None`):
+            The framework to use, either "pt" for PyTorch or "tf" for TensorFlow. The specified framework must be
+            installed.
+
+            If no framework is specified, will default to the one currently installed. If no framework is specified
+            and both frameworks are installed, will default to PyTorch.
+
+    Returns:
+        :class:`~transformers.Pipeline`: Class inheriting from :class:`~transformers.Pipeline`, according to
+        the task.
+
+    Examples::
+
+        from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer
+
+        # Sentiment analysis pipeline
         pipeline('sentiment-analysis')
+
+        # Question answering pipeline, specifying the checkpoint identifier
         pipeline('question-answering', model='distilbert-base-cased-distilled-squad', tokenizer='bert-base-cased')
-        pipeline('ner', model=AutoModel.from_pretrained(...), tokenizer=AutoTokenizer.from_pretrained(...)
-        pipeline('ner', model='dbmdz/bert-large-cased-finetuned-conll03-english', tokenizer='bert-base-cased')
-        pipeline('ner', model='https://...pytorch-model.bin', config='https://...config.json', tokenizer='bert-base-cased')
+
+        # Named entity recognition pipeline, passing in a specific model and tokenizer
+        model = AutoModelForTokenClassification.from_pretrained("dbmdz/bert-large-cased-finetuned-conll03-english")
+        tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+        pipeline('ner', model=model, tokenizer=tokenizer)
+
+        # Named entity recognition pipeline, passing a model and configuration with a HTTPS URL.
+        model_url = "https://s3.amazonaws.com/models.huggingface.co/bert/dbmdz/bert-large-cased-finetuned-conll03-english/pytorch_model.bin"
+        config_url = "https://s3.amazonaws.com/models.huggingface.co/bert/dbmdz/bert-large-cased-finetuned-conll03-english/config.json"
+        pipeline('ner', model=model_url, config=config_url, tokenizer='bert-base-cased')
     """
     # Retrieve the task
     if task not in SUPPORTED_TASKS:
@@ -1048,13 +1340,12 @@ def pipeline(
                 "Please provided a PretrainedTokenizer class or a path/url/shortcut name to a pretrained tokenizer."
             )
 
+    modelcard = None
     # Try to infer modelcard from model or config name (if provided as str)
-    if modelcard is None:
-        # Try to fallback on one of the provided string for model or config (will replace the suffix)
-        if isinstance(model, str):
-            modelcard = model
-        elif isinstance(config, str):
-            modelcard = config
+    if isinstance(model, str):
+        modelcard = model
+    elif isinstance(config, str):
+        modelcard = config
 
     # Instantiate tokenizer if needed
     if isinstance(tokenizer, (str, tuple)):

src/transformers/tokenization_bart.py

@@ -19,11 +19,7 @@ from .tokenization_roberta import RobertaTokenizer
 # vocab and merges same as roberta
 vocab_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-vocab.json"
 merges_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-merges.txt"
-_all_bart_models = [
-    "bart-large",
-    "bart-large-mnli",
-    # "bart-large-cnn"
-]
+_all_bart_models = ["bart-large", "bart-large-mnli", "bart-large-cnn"]
 
 
 class BartTokenizer(RobertaTokenizer):

src/transformers/tokenization_utils.py

@@ -1012,6 +1012,12 @@ class PreTrainedTokenizer(object):
                 "https://github.com/huggingface/transformers/pull/2674"
             )
 
+        # Throw an error if we can pad because there is no padding token
+        if pad_to_max_length and self.pad_token_id is None:
+            raise ValueError(
+                "Unable to set proper padding strategy as the tokenizer does not have a padding token. In this case please set the `pad_token` `(tokenizer.pad_token = tokenizer.eos_token e.g.)` or add a new pad token via the function add_special_tokens if you want to use a padding strategy"
+            )
+
         first_ids = get_input_ids(text)
         second_ids = get_input_ids(text_pair) if text_pair is not None else None
 
@@ -1115,6 +1121,12 @@ class PreTrainedTokenizer(object):
                     "Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
                 )
 
+        # Throw an error if we can pad because there is no padding token
+        if pad_to_max_length and self.pad_token_id is None:
+            raise ValueError(
+                "Unable to set proper padding strategy as the tokenizer does not have a padding token. In this case please set the `pad_token` `(tokenizer.pad_token = tokenizer.eos_token e.g.)` or add a new pad token via the function add_special_tokens if you want to use a padding strategy"
+            )
+
         if return_offsets_mapping:
             raise NotImplementedError(
                 "return_offset_mapping is not available when using Python tokenizers."
@@ -1126,8 +1138,7 @@ class PreTrainedTokenizer(object):
 
         input_ids = []
         for ids_or_pair_ids in batch_text_or_text_pairs:
-            if isinstance(ids_or_pair_ids, (list, tuple)):
-                assert len(ids_or_pair_ids) == 2
+            if isinstance(ids_or_pair_ids, (list, tuple)) and len(ids_or_pair_ids) == 2:
                 ids, pair_ids = ids_or_pair_ids
             else:
                 ids, pair_ids = ids_or_pair_ids, None
@@ -1789,7 +1800,7 @@ class PreTrainedTokenizerFast(PreTrainedTokenizer):
 
         # Throw an error if we can pad because there is no padding token
         if pad_to_max_length and self.pad_token_id is None:
-            raise ValueError("Unable to set proper padding strategy as the tokenizer does have padding token")
+            raise ValueError("Unable to set proper padding strategy as the tokenizer does not have a padding token")
 
         # Set the truncation and padding strategy and restore the initial configuration
         with truncate_and_pad(

templates/adding_a_new_example_script/run_xxx.py

@@ -594,7 +594,7 @@ def main():
     # Setup CUDA, GPU & distributed training
     if args.local_rank == -1 or args.no_cuda:
         device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
     else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
         torch.cuda.set_device(args.local_rank)
         device = torch.device("cuda", args.local_rank)

tests/test_modeling_albert.py

@@ -29,6 +29,7 @@ if is_torch_available():
         AlbertModel,
         AlbertForMaskedLM,
         AlbertForSequenceClassification,
+        AlbertForTokenClassification,
         AlbertForQuestionAnswering,
     )
     from transformers.modeling_albert import ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
@@ -207,6 +208,25 @@ class AlbertModelTest(ModelTesterMixin, unittest.TestCase):
             self.parent.assertListEqual(list(result["logits"].size()), [self.batch_size, self.num_labels])
             self.check_loss_output(result)
 
+        def create_and_check_albert_for_token_classification(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            config.num_labels = self.num_labels
+            model = AlbertForTokenClassification(config=config)
+            model.to(torch_device)
+            model.eval()
+            loss, logits = model(
+                input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels
+            )
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()), [self.batch_size, self.seq_length, self.num_labels]
+            )
+            self.check_loss_output(result)
+
         def prepare_config_and_inputs_for_common(self):
             config_and_inputs = self.prepare_config_and_inputs()
             (

tests/test_modeling_bart.py

@@ -171,8 +171,8 @@ class BartHeadTests(unittest.TestCase):
 
     vocab_size = 99
 
-    def test_lm_forward(self):
-        input_ids = torch.Tensor(
+    def _get_config_and_data(self, output_past=False):
+        input_ids = torch.tensor(
             [
                 [71, 82, 18, 33, 46, 91, 2],
                 [68, 34, 26, 58, 30, 82, 2],
@@ -187,11 +187,12 @@ class BartHeadTests(unittest.TestCase):
                 [21, 5, 62, 28, 14, 76, 2],
                 [45, 98, 37, 86, 59, 48, 2],
                 [70, 70, 50, 9, 28, 0, 2],
-            ]
-        ).long()
-        batch_size = input_ids.shape[0]
-        decoder_lm_labels = ids_tensor([batch_size, input_ids.shape[1]], self.vocab_size)
+            ],
+            dtype=torch.long,
+            device=torch_device,
+        )
 
+        batch_size = input_ids.shape[0]
         config = BartConfig(
             vocab_size=self.vocab_size,
             d_model=24,
@@ -202,14 +203,27 @@ class BartHeadTests(unittest.TestCase):
             encoder_ffn_dim=32,
             decoder_ffn_dim=32,
             max_position_embeddings=48,
+            output_past=output_past,
         )
+        return config, input_ids, batch_size
+
+    def test_sequence_classification_forward(self):
+        config, input_ids, batch_size = self._get_config_and_data()
+        labels = _long_tensor([2] * batch_size).to(torch_device)
         model = BartForSequenceClassification(config)
-        outputs = model.forward(input_ids=input_ids, decoder_input_ids=input_ids)
-        logits = outputs[0]
+        model.to(torch_device)
+        outputs = model(input_ids=input_ids, decoder_input_ids=input_ids, labels=labels)
+        logits = outputs[1]
         expected_shape = torch.Size((batch_size, config.num_labels))
         self.assertEqual(logits.shape, expected_shape)
+        loss = outputs[0]
+        self.assertIsInstance(loss.item(), float)
 
+    def test_lm_forward(self):
+        config, input_ids, batch_size = self._get_config_and_data(output_past=False)
+        decoder_lm_labels = ids_tensor([batch_size, input_ids.shape[1]], self.vocab_size)
         lm_model = BartForMaskedLM(config)
+        lm_model.to(torch_device)
         loss, logits, enc_features = lm_model.forward(
             input_ids=input_ids, lm_labels=decoder_lm_labels, decoder_input_ids=input_ids
         )
@@ -236,7 +250,7 @@ class BartHeadTests(unittest.TestCase):
         expected_shape = (*summary.shape, config.vocab_size)
         self.assertEqual(logits.shape, expected_shape)
 
-    def test_generate(self):
+    def test_generate_beam_search(self):
         input_ids = torch.Tensor([[71, 82, 2], [68, 34, 2]]).long()
         config = BartConfig(
             vocab_size=self.vocab_size,
@@ -252,8 +266,12 @@ class BartHeadTests(unittest.TestCase):
         )
         lm_model = BartForMaskedLM(config)
         lm_model.eval()
-        new_input_ids = lm_model.generate(input_ids)
-        self.assertEqual(new_input_ids.shape, (input_ids.shape[0], 20))
+
+        new_input_ids = lm_model.generate(
+            input_ids.clone(), num_return_sequences=1, num_beams=2, no_repeat_ngram_size=3, max_length=5
+        )
+        self.assertEqual(new_input_ids.shape, (input_ids.shape[0], 5))
+        # TODO(SS): uneven length batches, empty inputs
 
     def test_shift_tokens_right(self):
         input_ids = torch.Tensor([[71, 82, 18, 33, 2, 1, 1], [68, 34, 26, 58, 30, 82, 2]]).long()
@@ -292,6 +310,10 @@ def _assert_tensors_equal(a, b, atol=1e-12, prefix=""):
         raise AssertionError(msg)
 
 
+def _long_tensor(tok_lst):
+    return torch.tensor(tok_lst, dtype=torch.long, device=torch_device,)
+
+
 TOLERANCE = 1e-4
 
 
@@ -299,15 +321,15 @@ TOLERANCE = 1e-4
 class BartModelIntegrationTest(unittest.TestCase):
     @slow
     def test_inference_no_head(self):
-        model = BartModel.from_pretrained("bart-large")
-        input_ids = torch.Tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]).long()
+        model = BartModel.from_pretrained("bart-large").to(torch_device)
+        input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
         inputs_dict = prepare_bart_inputs_dict(model.config, input_ids)
         with torch.no_grad():
             output = model.forward(**inputs_dict)[0]
         expected_shape = torch.Size((1, 11, 1024))
         self.assertEqual(output.shape, expected_shape)
-        expected_slice = torch.Tensor(
-            [[0.7144, 0.8143, -1.2813], [0.7144, 0.8143, -1.2813], [-0.0467, 2.5911, -2.1845]]
+        expected_slice = torch.tensor(
+            [[0.7144, 0.8143, -1.2813], [0.7144, 0.8143, -1.2813], [-0.0467, 2.5911, -2.1845]], device=torch_device
         )
         self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=TOLERANCE))
 
@@ -315,20 +337,22 @@ class BartModelIntegrationTest(unittest.TestCase):
     def test_mnli_inference(self):
 
         example_b = [0, 31414, 232, 328, 740, 1140, 69, 46078, 1588, 2, 1]
-        input_ids = torch.Tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2], example_b]).long()
+        input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2], example_b])
 
-        model = AutoModelForSequenceClassification.from_pretrained("bart-large-mnli")  # eval called in from_pre
+        model = AutoModelForSequenceClassification.from_pretrained("bart-large-mnli").to(
+            torch_device
+        )  # eval called in from_pre
         inputs_dict = prepare_bart_inputs_dict(model.config, input_ids)
         # Test that model hasn't changed
         with torch.no_grad():
             batched_logits, features = model.forward(**inputs_dict)
         expected_shape = torch.Size((2, 3))
         self.assertEqual(batched_logits.shape, expected_shape)
-        expected_slice = torch.Tensor([[0.1907, 1.4342, -1.0289]])
+        expected_slice = torch.Tensor([[0.1907, 1.4342, -1.0289]]).to(torch_device)
         logits_arr = batched_logits[0].detach()
 
         # Test that padding does not change results
-        input_ids_no_pad = torch.Tensor([example_b[:-1]]).long()
+        input_ids_no_pad = _long_tensor([example_b[:-1]])
 
         inputs_dict = prepare_bart_inputs_dict(model.config, input_ids=input_ids_no_pad)
         with torch.no_grad():
@@ -342,3 +366,55 @@ class BartModelIntegrationTest(unittest.TestCase):
         for model_name in list(BART_PRETRAINED_MODEL_ARCHIVE_MAP.keys()):
             model = BartModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
+
+    @slow
+    def test_cnn_summarization_same_as_fairseq(self):
+        hf = BartForMaskedLM.from_pretrained("bart-large-cnn", output_past=True,).to(torch_device)
+        tok = BartTokenizer.from_pretrained("bart-large")
+        text = " (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian"
+        tokens = tok.encode(text, return_tensors="pt").to(torch_device)
+        extra_len = 20
+        gen_tokens = hf.generate(tokens, num_beams=4, max_length=extra_len,)  # repetition_penalty=10.,
+        expected_result = "<s>The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday."
+        generated = [tok.decode(g,) for g in gen_tokens]
+        self.assertEqual(expected_result, generated[0])
+
+        # Harder cases with batching
+        FRANCE_ARTICLE = ' Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object.  Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the Line truncated
+        EXPECTED_SUMMARY_FRANCE = 'French prosecutor says he\'s not aware of any video footage from on board the plane. German daily Bild and French Paris Match claim to have found a cell phone video of the crash. A French Gendarmerie spokesman calls the reports "completely wrong" and "unwarranted" German airline Lufthansa confirms co-pilot Andreas Lubitz had battled depression.'
+
+        SHORTER_ARTICLE = ' (CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian territories. The formal accession was marked with a ceremony at The Hague, in the Netherlands, where the court is based. The Palestinians signed the ICC\'s founding Rome Statute in January, when they also accepted its jurisdiction over alleged crimes committed "in the occupied Palestinian territory, including East Jerusalem, since June 13, 2014." Later that month, the ICC opened a preliminary examination into the situation in Palestinian territories, paving the way for possible war crimes investigations against Israelis. As members of the court, Palestinians may be subject to counter-charges as well. Israel and the United States, neither of which is an ICC member, opposed the Palestinians\' efforts to join the body. But Palestinian Foreign Minister Riad al-Malki, speaking at Wednesday\'s ceremony, said it was a move toward greater justice. "As Palestine formally becomes a State Party to the Rome Statute today, the world is also a step closer to ending a long era of impunity and injustice," he said, according to an ICC news release. "Indeed, today brings us closer to our shared goals of justice and peace." Judge Kuniko Ozaki, a vice president of the ICC, said acceding to the treaty was just the first step for the Palestinians. "As the Rome Statute today enters into force for the State of Palestine, Palestine acquires all the rights as well as responsibilities that come with being a State Party to the Statute. These are substantive commitments, which cannot be taken lightly," she said. Rights group Human Rights Watch welcomed the development. "Governments seeking to penalize Palestine for joining the ICC should immediately end their pressure, and countries that support universal acceptance of the court\'s treaty should speak out to welcome its membership," said Balkees Jarrah, international justice counsel for the group. "What\'s objectionable is the attempts to undermine international justice, not Palestine\'s decision to join a treaty to which over 100 countries around the world are members." In January, when the preliminary ICC examination was opened, Israeli Prime Minister Benjamin Netanyahu described it as an outrage, saying the court was overstepping its boundaries. The United States also said it "strongly" disagreed with the court\'s decision. "As we have said repeatedly, we do not believe that Palestine is a state and therefore we do not believe that it is eligible to join the ICC," the State Department said in a statement. It urged the warring sides to resolve their differences through direct negotiations. "We will continue to oppose actions against Israel at the ICC as counterproductive to the cause of peace," it said. But the ICC begs to differ with the definition of a state for its purposes and refers to the territories as "Palestine." While a preliminary examination is not a formal investigation, it allows the court to review evidence and determine whether to investigate suspects on both sides. Prosecutor Fatou Bensouda said her office would "conduct its analysis in full independence and impartiality." The war between Israel and Hamas militants in Gaza last summer left more than 2,000 people dead. The inquiry will include alleged war crimes committed since June. The International Criminal Court was set up in 2002 to prosecute genocide, crimes against humanity and war crimes. CNN\'s Vasco Cotovio, Kareem Khadder and Faith Karimi contributed to this report.'
+        EXPECTED_SUMMARY_SHORTER = "The Palestinian Authority becomes the 123rd member of the International Criminal Court. The move gives the court jurisdiction over alleged crimes in Palestinian territories. Israel and the United States opposed the Palestinians' efforts to join the body. But Palestinian Foreign Minister Riad al-Malki said it was a move toward greater justice."
+
+        # The below article tests that we don't add any hypotheses outside of the top n_beams
+        IRAN_ARTICLE = " (CNN)The United States and its negotiating partners reached a very strong framework agreement with Iran in Lausanne, Switzerland, on Thursday that limits Iran's nuclear program in such a way as to effectively block it from building a nuclear weapon. Expect pushback anyway, if the recent past is any harbinger. Just last month, in an attempt to head off such an agreement, House Speaker John Boehner invited Israeli Prime Minister Benjamin Netanyahu to preemptively blast it before Congress, and 47 senators sent a letter to the Iranian leadership warning them away from a deal. The debate that has already begun since the announcement of the new framework will likely result in more heat than light. It will not be helped by the gathering swirl of dubious assumptions and doubtful assertions. Let us address some of these: . The most misleading assertion, despite universal rejection by experts, is that the negotiations' objective at the outset was the total elimination of any nuclear program in Iran. That is the position of Netanyahu and his acolytes in the U.S. Congress. But that is not and never was the objective. If it had been, there would have been no Iranian team at the negotiating table. Rather, the objective has always been to structure an agreement or series of agreements so that Iran could not covertly develop a nuclear arsenal before the United States and its allies could respond. The new framework has exceeded expectations in achieving that goal. It would reduce Iran's low-enriched uranium stockpile, cut by two-thirds its number of installed centrifuges and implement a rigorous inspection regime. Another dubious assumption of opponents is that the Iranian nuclear program is a covert weapons program. Despite sharp accusations by some in the United States and its allies, Iran denies having such a program, and U.S. intelligence contends that Iran has not yet made the decision to build a nuclear weapon. Iran's continued cooperation with International Atomic Energy Agency inspections is further evidence on this point, and we'll know even more about Iran's program in the coming months and years because of the deal. In fact, the inspections provisions that are part of this agreement are designed to protect against any covert action by the Iranians. What's more, the rhetoric of some members of Congress has implied that the negotiations have been between only the United States and Iran (i.e., the 47 senators' letter warning that a deal might be killed by Congress or a future president). This of course is not the case. The talks were between Iran and the five permanent members of the U.N. Security Council (United States, United Kingdom, France, China and Russia) plus Germany, dubbed the P5+1. While the United States has played a leading role in the effort, it negotiated the terms alongside its partners. If the agreement reached by the P5+1 is rejected by Congress, it could result in an unraveling of the sanctions on Iran and threaten NATO cohesion in other areas. Another questionable assertion is that this agreement contains a sunset clause, after which Iran will be free to do as it pleases. Again, this is not the case. Some of the restrictions on Iran's nuclear activities, such as uranium enrichment, will be eased or eliminated over time, as long as 15 years. But most importantly, the framework agreement includes Iran's ratification of the Additional Protocol, which allows IAEA inspectors expanded access to nuclear sites both declared and nondeclared. This provision will be permanent. It does not sunset. Thus, going forward, if Iran decides to enrich uranium to weapons-grade levels, monitors will be able to detect such a move in a matter of days and alert the U.N. Security Council. Many in Congress have said that the agreement should be a formal treaty requiring the Senate to \"advise and consent.\" But the issue is not suited for a treaty. Treaties impose equivalent obligations on all signatories. For example, the New START treaty limits Russia and the United States to 1,550 deployed strategic warheads. But any agreement with Iran will not be so balanced.  The restrictions and obligations in the final framework agreement will be imposed almost exclusively on Iran. The P5+1 are obligated only to ease and eventually remove most but not all economic sanctions, which were imposed as leverage to gain this final deal. Finally some insist that any agreement must address Iranian missile programs, human rights violations or support for Hamas or Hezbollah.  As important as these issues are, and they must indeed be addressed, they are unrelated to the most important aim of a nuclear deal: preventing a nuclear Iran.  To include them in the negotiations would be a poison pill. This agreement should be judged on its merits and on how it affects the security of our negotiating partners and allies, including Israel. Those judgments should be fact-based, not based on questionable assertions or dubious assLine truncated
+        EXPECTED_SUMMARY_IRAN = "The U.S. and its negotiating partners reached a very strong framework agreement with Iran. Peter Bergen: The debate that has already begun will likely result in more heat than light. He says the agreement limits Iran's nuclear program in such a way as to effectively block it from building a nuclear weapon. Bergen says the most important aim of a nuclear deal is preventing a nuclear Iran."
+
+        ARTICLE_SUBWAY = ' New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband.  Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 2010 marriage license application, according to court documents. Prosecutors said the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002.  All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages.  Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. If convicted, Barrientos faces up to four years in prison.  Her next court appearance is scheduled for May 18.'
+        EXPECTED_SUMMARY_SUBWAY = "Liana Barrientos has been married 10 times, sometimes within two weeks of each other. Prosecutors say the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx. She was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the subway."
+
+        dct = tok.batch_encode_plus(
+            [FRANCE_ARTICLE, SHORTER_ARTICLE, IRAN_ARTICLE, ARTICLE_SUBWAY],
+            max_length=1024,
+            pad_to_max_length=True,
+            return_tensors="pt",
+        )
+        self.assertEqual(1024, dct["input_ids"].shape[1])
+        hypotheses_batch = hf.generate(
+            input_ids=dct["input_ids"].to(torch_device),
+            attention_mask=dct["attention_mask"].to(torch_device),
+            num_beams=4,
+            length_penalty=2.0,
+            max_length=140,
+            min_len=55,
+            no_repeat_ngram_size=3,
+        )
+        decoded = [
+            tok.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in hypotheses_batch
+        ]
+        self.assertListEqual(
+            [EXPECTED_SUMMARY_FRANCE, EXPECTED_SUMMARY_SHORTER, EXPECTED_SUMMARY_IRAN, EXPECTED_SUMMARY_SUBWAY],
+            decoded,
+        )
+        # TODO(SS): run fairseq again with num_beams=2, min_len=20.
+        # TODO(SS): add test case that hits max_length

tests/test_modeling_common.py

@@ -36,6 +36,7 @@ if is_torch_available():
         BertModel,
         BertConfig,
         BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        top_k_top_p_filtering,
     )
 
 
@@ -68,7 +69,7 @@ class ModelTesterMixin:
             model.eval()
             with torch.no_grad():
                 outputs = model(**inputs_dict)
-            out_2 = outputs[0].numpy()
+            out_2 = outputs[0].cpu().numpy()
             out_2[np.isnan(out_2)] = 0
 
             with tempfile.TemporaryDirectory() as tmpdirname:
@@ -263,7 +264,7 @@ class ModelTesterMixin:
             # Prepare head_mask
             # Set require_grad after having prepared the tensor to avoid error (leaf variable has been moved into the graph interior)
             head_mask = torch.ones(
-                self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads, device=torch_device
+                self.model_tester.num_hidden_layers, self.model_tester.num_attention_heads, device=torch_device,
             )
             head_mask[0, 0] = 0
             head_mask[-1, :-1] = 0
@@ -303,7 +304,7 @@ class ModelTesterMixin:
             return
 
         for model_class in self.all_model_classes:
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            (config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
 
             if "head_mask" in inputs_dict:
                 del inputs_dict["head_mask"]
@@ -313,7 +314,10 @@ class ModelTesterMixin:
             model = model_class(config=config)
             model.to(torch_device)
             model.eval()
-            heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
+            heads_to_prune = {
+                0: list(range(1, self.model_tester.num_attention_heads)),
+                -1: [0],
+            }
             model.prune_heads(heads_to_prune)
             with torch.no_grad():
                 outputs = model(**inputs_dict)
@@ -329,7 +333,7 @@ class ModelTesterMixin:
             return
 
         for model_class in self.all_model_classes:
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            (config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
 
             if "head_mask" in inputs_dict:
                 del inputs_dict["head_mask"]
@@ -339,7 +343,10 @@ class ModelTesterMixin:
             model = model_class(config=config)
             model.to(torch_device)
             model.eval()
-            heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
+            heads_to_prune = {
+                0: list(range(1, self.model_tester.num_attention_heads)),
+                -1: [0],
+            }
             model.prune_heads(heads_to_prune)
 
             with tempfile.TemporaryDirectory() as temp_dir_name:
@@ -359,7 +366,7 @@ class ModelTesterMixin:
             return
 
         for model_class in self.all_model_classes:
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            (config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
 
             if "head_mask" in inputs_dict:
                 del inputs_dict["head_mask"]
@@ -367,7 +374,10 @@ class ModelTesterMixin:
             config.output_attentions = True
             config.output_hidden_states = False
 
-            heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)), -1: [0]}
+            heads_to_prune = {
+                0: list(range(1, self.model_tester.num_attention_heads)),
+                -1: [0],
+            }
             config.pruned_heads = heads_to_prune
 
             model = model_class(config=config)
@@ -387,7 +397,7 @@ class ModelTesterMixin:
             return
 
         for model_class in self.all_model_classes:
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            (config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
 
             if "head_mask" in inputs_dict:
                 del inputs_dict["head_mask"]
@@ -465,13 +475,14 @@ class ModelTesterMixin:
             )
 
     def test_resize_tokens_embeddings(self):
-        original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        (original_config, inputs_dict,) = self.model_tester.prepare_config_and_inputs_for_common()
         if not self.test_resize_embeddings:
             return
 
         for model_class in self.all_model_classes:
             config = copy.deepcopy(original_config)
             model = model_class(config)
+            model.to(torch_device)
 
             model_vocab_size = config.vocab_size
             # Retrieve the embeddings and clone theme
@@ -620,10 +631,20 @@ class ModelTesterMixin:
                 # batch_size = 1, num_beams > 1
                 self._check_generated_tokens(model.generate(max_length=5, num_beams=3))
 
+            with self.assertRaises(AssertionError):
+                # generating multiple sequences when greedy no beam generation
+                # is not allowed as it would always generate the same sequences
+                model.generate(input_ids, do_sample=False, num_return_sequences=2)
+
+            with self.assertRaises(AssertionError):
+                # generating more sequences than having beams leads is not possible
+                model.generate(input_ids, do_sample=False, num_return_sequences=3, num_beams=2)
+
             # batch_size > 1, sample
             self._check_generated_tokens(model.generate(input_ids, num_return_sequences=3))
             # batch_size > 1, greedy
-            self._check_generated_tokens(model.generate(input_ids, do_sample=False, num_return_sequences=3))
+            self._check_generated_tokens(model.generate(input_ids, do_sample=False))
+
             # batch_size > 1, num_beams > 1, sample
             self._check_generated_tokens(model.generate(input_ids, num_beams=3, num_return_sequences=3,))
             # batch_size > 1, num_beams > 1, greedy
@@ -694,3 +715,110 @@ class ModelUtilsTest(unittest.TestCase):
             self.assertEqual(model.config.output_attentions, True)
             self.assertEqual(model.config.output_hidden_states, True)
             self.assertEqual(model.config, config)
+
+
+@require_torch
+class UtilsFunctionsTest(unittest.TestCase):
+
+    # tests whether the top_k_top_p function behaves as expected
+    def test_top_k_top_p_filtering(self):
+        logits = torch.tensor(
+            [
+                [
+                    8.2220991,  # 3rd highest value; idx. 0
+                    -0.5620044,
+                    5.23229752,
+                    4.0386393,
+                    -6.8798378,
+                    -0.54785802,
+                    -3.2012153,
+                    2.92777176,
+                    1.88171953,
+                    7.35341276,  # 5th highest value; idx. 9
+                    8.43207833,  # 2nd highest value; idx. 10
+                    -9.85711836,
+                    -5.96209236,
+                    -1.13039161,
+                    -7.1115294,
+                    -0.8369633,
+                    -5.3186408,
+                    7.06427407,
+                    0.81369344,
+                    -0.82023817,
+                    -5.9179796,
+                    0.58813443,
+                    -6.99778438,
+                    4.71551189,
+                    -0.18771637,
+                    7.44020759,  # 4th highest value; idx. 25
+                    9.38450987,  # 1st highest value; idx. 26
+                    2.12662941,
+                    -9.32562038,
+                    2.35652522,
+                ],  # cummulative prob of 5 highest values <= 0.6
+                [
+                    0.58425518,
+                    4.53139238,
+                    -5.57510464,
+                    -6.28030699,
+                    -7.19529503,
+                    -4.02122551,
+                    1.39337037,
+                    -6.06707057,
+                    1.59480517,
+                    -9.643119,
+                    0.03907799,
+                    0.67231762,
+                    -8.88206726,
+                    6.27115922,  # 4th highest value; idx. 13
+                    2.28520723,
+                    4.82767506,
+                    4.30421368,
+                    8.8275313,  # 2nd highest value; idx. 17
+                    5.44029958,  # 5th highest value; idx. 18
+                    -4.4735794,
+                    7.38579536,  # 3rd highest value; idx. 20
+                    -2.91051663,
+                    2.61946077,
+                    -2.5674762,
+                    -9.48959302,
+                    -4.02922645,
+                    -1.35416918,
+                    9.67702323,  # 1st highest value; idx. 27
+                    -5.89478553,
+                    1.85370467,
+                ],  # cummulative prob of 5 highest values <= 0.6
+            ],
+            dtype=torch.float,
+            device=torch_device,
+        )
+
+        non_inf_expected_idx = torch.tensor(
+            [[0, 0], [0, 9], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 18], [1, 20], [1, 27]],
+            dtype=torch.long,
+            device=torch_device,
+        )  # expected non filtered idx as noted above
+
+        non_inf_expected_output = torch.tensor(
+            [
+                8.2221,
+                7.3534,
+                8.4321,
+                7.4402,
+                9.3845,
+                6.2712,
+                8.8275,
+                5.4403,
+                7.3858,
+                9.6770,
+            ],  # expected non filtered values as noted above
+            dtype=torch.float,
+            device=torch_device,
+        )
+
+        output = top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
+        non_inf_output = output[output != -float("inf")].to(device=torch_device)
+        non_inf_idx = (output != -float("inf")).nonzero().to(device=torch_device)
+
+        self.assertTrue(torch.allclose(non_inf_expected_output, non_inf_output, atol=1e-12))
+        self.assertTrue(torch.all(torch.eq(non_inf_expected_idx, non_inf_idx)))

tests/test_modeling_encoder_decoder.py

@@ -1,50 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Hugging Face 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 logging
-import unittest
-
-from transformers import is_torch_available
-
-from .utils import require_torch, slow
-
-
-if is_torch_available():
-    from transformers import BertModel, BertForMaskedLM, Model2Model
-    from transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
-
-
-@require_torch
-class EncoderDecoderModelTest(unittest.TestCase):
-    @slow
-    def test_model2model_from_pretrained(self):
-        logging.basicConfig(level=logging.INFO)
-        for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = Model2Model.from_pretrained(model_name)
-            self.assertIsInstance(model.encoder, BertModel)
-            self.assertIsInstance(model.decoder, BertForMaskedLM)
-            self.assertEqual(model.decoder.config.is_decoder, True)
-            self.assertEqual(model.encoder.config.is_decoder, False)
-
-    def test_model2model_from_pretrained_not_bert(self):
-        logging.basicConfig(level=logging.INFO)
-        with self.assertRaises(ValueError):
-            _ = Model2Model.from_pretrained("roberta")
-
-        with self.assertRaises(ValueError):
-            _ = Model2Model.from_pretrained("distilbert")
-
-        with self.assertRaises(ValueError):
-            _ = Model2Model.from_pretrained("does-not-exist")

tests/test_modeling_gpt2.py

@@ -170,6 +170,74 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
             )
             self.parent.assertEqual(len(result["presents"]), config.n_layer)
 
+        def create_and_check_gpt2_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = GPT2Model(config=config)
+            model.to(torch_device)
+            model.eval()
+
+            # first forward pass
+            output, past = model(input_ids, token_type_ids=token_type_ids)
+
+            # create hypothetical next token and extent to next_input_ids
+            next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+            next_token_types = ids_tensor([self.batch_size, 1], self.type_vocab_size)
+
+            # append to next input_ids and token_type_ids
+            next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+            next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
+
+            output_from_no_past, _ = model(next_input_ids, token_type_ids=next_token_type_ids)
+            output_from_past, _ = model(next_tokens, token_type_ids=next_token_types, past=past)
+
+            # 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_gpt2_model_attention_mask_past(
+            self, config, input_ids, input_mask, head_mask, token_type_ids, *args
+        ):
+            model = GPT2Model(config=config)
+            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 = self.seq_length // 2
+            attn_mask[:, half_seq_length:] = 0
+
+            # first forward pass
+            output, past = model(input_ids, attention_mask=attn_mask)
+
+            # 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)
+            output_from_past, _ = model(next_tokens, past=past, attention_mask=attn_mask)
+
+            # 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_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = GPT2LMHeadModel(config)
             model.to(torch_device)
@@ -248,6 +316,14 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_gpt2_model(*config_and_inputs)
 
+    def test_gpt2_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_gpt2_model_past(*config_and_inputs)
+
+    def test_gpt2_model_att_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_gpt2_model_attention_mask_past(*config_and_inputs)
+
     def test_gpt2_lm_head_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
@@ -310,33 +386,33 @@ class GPT2ModelLanguageGenerationTest(unittest.TestCase):
     @slow
     def test_lm_generate_distilgpt2(self):
         model = GPT2LMHeadModel.from_pretrained("distilgpt2")
-        input_ids = torch.Tensor([[464, 3290, 318, 13779]]).long()  # The dog is cute
+        input_ids = torch.Tensor([[464, 1893]]).long()  # The president
         expected_output_ids = [
             464,
-            3290,
-            318,
-            13779,
-            996,
-            339,
-            460,
-            3360,
-            655,
-            2513,
+            1893,
+            286,
+            262,
+            1578,
+            1829,
+            11,
+            290,
+            262,
+            1893,
+            286,
+            262,
+            1578,
+            7526,
+            11,
+            423,
+            587,
             287,
             262,
-            3952,
-            13,
-            632,
-            318,
-            407,
-            845,
-            3621,
-            284,
-        ]  # The dog is cute though he can sometimes just walk in the park. It is not very nice to
-        torch.manual_seed(0)
+            2635,
+        ]  # The president of the United States, and the president of the United Kingdom, have been in the White
 
         output_ids = model.generate(
             input_ids,
+            do_sample=False,
             bos_token_id=self.special_tokens["bos_token_id"],
             eos_token_ids=self.special_tokens["eos_token_id"],
         )

tests/test_modeling_roberta.py

@@ -329,10 +329,15 @@ class RobertaModelIntegrationTest(unittest.TestCase):
         expected_shape = torch.Size((1, 11, 50265))
         self.assertEqual(output.shape, expected_shape)
         # compare the actual values for a slice.
-        expected_slice = torch.Tensor(
-            [[[33.8843, -4.3107, 22.7779], [4.6533, -2.8099, 13.6252], [1.8222, -3.6898, 8.8600]]]
+        expected_slice = torch.tensor(
+            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
         )
-        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
+        # roberta.eval()
+        # expected_slice = roberta.model.forward(input_ids)[0][:, :3, :3].detach()
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
 
     @slow
     def test_inference_no_head(self):
@@ -341,10 +346,15 @@ class RobertaModelIntegrationTest(unittest.TestCase):
         input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
         output = model(input_ids)[0]
         # compare the actual values for a slice.
-        expected_slice = torch.Tensor(
-            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0539, -0.0174], [0.0548, 0.0799, 0.1687]]]
+        expected_slice = torch.tensor(
+            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
         )
-        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.base')
+        # roberta.eval()
+        # expected_slice = roberta.extract_features(input_ids)[:, :3, :3].detach()
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
 
     @slow
     def test_inference_classification_head(self):
@@ -354,5 +364,10 @@ class RobertaModelIntegrationTest(unittest.TestCase):
         output = model(input_ids)[0]
         expected_shape = torch.Size((1, 3))
         self.assertEqual(output.shape, expected_shape)
-        expected_tensor = torch.Tensor([[-0.9469, 0.3913, 0.5118]])
-        self.assertTrue(torch.allclose(output, expected_tensor, atol=1e-3))
+        expected_tensor = torch.tensor([[-0.9469, 0.3913, 0.5118]])
+
+        # roberta = torch.hub.load('pytorch/fairseq', 'roberta.large.mnli')
+        # roberta.eval()
+        # expected_tensor = roberta.predict("mnli", input_ids, return_logits=True).detach()
+
+        self.assertTrue(torch.allclose(output, expected_tensor, atol=1e-4))

tests/test_modeling_t5.py

@@ -20,7 +20,7 @@ from transformers import is_torch_available
 
 from .test_configuration_common import ConfigTester
 from .test_modeling_common import ModelTesterMixin, ids_tensor
-from .utils import CACHE_DIR, require_torch, slow
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
 if is_torch_available():
@@ -125,6 +125,7 @@ class T5ModelTest(ModelTesterMixin, unittest.TestCase):
             decoder_lm_labels,
         ):
             model = T5Model(config=config)
+            model.to(torch_device)
             model.eval()
             decoder_output, encoder_output = model(
                 encoder_input_ids=encoder_input_ids,
@@ -157,6 +158,7 @@ class T5ModelTest(ModelTesterMixin, unittest.TestCase):
             decoder_lm_labels,
         ):
             model = T5WithLMHeadModel(config=config)
+            model.to(torch_device)
             model.eval()
             outputs = model(
                 encoder_input_ids=encoder_input_ids,

tests/test_modeling_tf_common.py

@@ -18,17 +18,32 @@ import copy
 import os
 import random
 import tempfile
+import unittest
 
 from transformers import is_tf_available, is_torch_available
 
-from .utils import require_tf
+from .utils import _tf_gpu_memory_limit, require_tf
 
 
 if is_tf_available():
     import tensorflow as tf
     import numpy as np
 
-    # from transformers.modeling_bert import BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    from transformers import tf_top_k_top_p_filtering
+
+    if _tf_gpu_memory_limit is not None:
+        gpus = tf.config.list_physical_devices("GPU")
+        for gpu in gpus:
+            # Restrict TensorFlow to only allocate x GB of memory on the GPUs
+            try:
+                tf.config.experimental.set_virtual_device_configuration(
+                    gpu, [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
+                )
+                logical_gpus = tf.config.experimental.list_logical_devices("GPU")
+                print("Logical GPUs", logical_gpus)
+            except RuntimeError as e:
+                # Virtual devices must be set before GPUs have been initialized
+                print(e)
 
 
 def _config_zero_init(config):
@@ -44,6 +59,7 @@ class TFModelTesterMixin:
 
     model_tester = None
     all_model_classes = ()
+    all_generative_model_classes = ()
     test_torchscript = True
     test_pruning = True
     test_resize_embeddings = True
@@ -204,7 +220,7 @@ class TFModelTesterMixin:
             outputs_dict = model(inputs_dict)
 
             inputs_keywords = copy.deepcopy(inputs_dict)
-            input_ids = inputs_keywords.pop("input_ids" if not self.is_encoder_decoder else "decoder_input_ids", None)
+            input_ids = inputs_keywords.pop("input_ids" if not self.is_encoder_decoder else "decoder_input_ids", None,)
             outputs_keywords = model(input_ids, **inputs_keywords)
 
             output_dict = outputs_dict[0].numpy()
@@ -287,7 +303,7 @@ class TFModelTesterMixin:
             self.assertEqual(model.config.output_hidden_states, True)
             self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
             self.assertListEqual(
-                list(hidden_states[0].shape[-2:]), [self.model_tester.seq_length, self.model_tester.hidden_size]
+                list(hidden_states[0].shape[-2:]), [self.model_tester.seq_length, self.model_tester.hidden_size],
             )
 
     def test_model_common_attributes(self):
@@ -304,7 +320,10 @@ class TFModelTesterMixin:
 
         for model_class in self.all_model_classes:
             model = model_class(config)
-            first, second = model(inputs_dict, training=False)[0], model(inputs_dict, training=False)[0]
+            first, second = (
+                model(inputs_dict, training=False)[0],
+                model(inputs_dict, training=False)[0],
+            )
             out_1 = first.numpy()
             out_2 = second.numpy()
             out_1 = out_1[~np.isnan(out_1)]
@@ -326,9 +345,9 @@ class TFModelTesterMixin:
                     x = wte([input_ids, None, None, None], mode="embedding")
                 except Exception:
                     if hasattr(self.model_tester, "embedding_size"):
-                        x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32)
+                        x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32,)
                     else:
-                        x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
+                        x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32,)
         return x
 
     def test_inputs_embeds(self):
@@ -354,6 +373,37 @@ class TFModelTesterMixin:
 
             model(inputs_dict)
 
+    def test_lm_head_model_random_generate(self):
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        input_ids = inputs_dict.get(
+            "input_ids", None
+        )  # TODO (PVP): ugly workaround to make code work for t5 for the moment - has to changed when t5 is fixed.
+
+        for model_class in self.all_generative_model_classes:
+            # TODO (PVP): add beam search tests when beam search is implemented
+            model = model_class(config)
+
+            if config.bos_token_id is None:
+                with self.assertRaises(AssertionError):
+                    model.generate(max_length=5)
+                # batch_size = 1
+                self._check_generated_tokens(model.generate(input_ids))
+            else:
+                # batch_size = 1
+                self._check_generated_tokens(model.generate(max_length=5))
+                # batch_size = 1, num_beams > 1
+
+            # batch_size > 1, sample
+            self._check_generated_tokens(model.generate(input_ids, num_return_sequences=3))
+            # batch_size > 1, greedy
+            self._check_generated_tokens(model.generate(input_ids, do_sample=False, num_return_sequences=3))
+
+    def _check_generated_tokens(self, output_ids):
+        for token_id in output_ids[0].numpy().tolist():
+            self.assertGreaterEqual(token_id, 0)
+            self.assertLess(token_id, self.model_tester.vocab_size)
+
 
 def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
     """Creates a random int32 tensor of the shape within the vocab size."""
@@ -371,3 +421,98 @@ def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
     output = tf.constant(values, shape=shape, dtype=dtype if dtype is not None else tf.int32)
 
     return output
+
+
+@require_tf
+class UtilsFunctionsTest(unittest.TestCase):
+
+    # tests whether the top_k_top_p_filtering function behaves as expected
+    def test_top_k_top_p_filtering(self):
+        logits = tf.convert_to_tensor(
+            [
+                [
+                    8.2220991,  # 3rd highest value; idx. 0
+                    -0.5620044,
+                    5.23229752,
+                    4.0386393,
+                    -6.8798378,
+                    -0.54785802,
+                    -3.2012153,
+                    2.92777176,
+                    1.88171953,
+                    7.35341276,  # 5th highest value; idx. 9
+                    8.43207833,  # 2nd highest value; idx. 10
+                    -9.85711836,
+                    -5.96209236,
+                    -1.13039161,
+                    -7.1115294,
+                    -0.8369633,
+                    -5.3186408,
+                    7.06427407,
+                    0.81369344,
+                    -0.82023817,
+                    -5.9179796,
+                    0.58813443,
+                    -6.99778438,
+                    4.71551189,
+                    -0.18771637,
+                    7.44020759,  # 4th highest value; idx. 25
+                    9.38450987,  # 1st highest value; idx. 26
+                    2.12662941,
+                    -9.32562038,
+                    2.35652522,
+                ],  # cummulative prob of 5 highest values <= 0.6
+                [
+                    0.58425518,
+                    4.53139238,
+                    -5.57510464,
+                    -6.28030699,
+                    -7.19529503,
+                    -4.02122551,
+                    1.39337037,
+                    -6.06707057,
+                    1.59480517,
+                    -9.643119,
+                    0.03907799,
+                    0.67231762,
+                    -8.88206726,
+                    6.27115922,  # 4th highest value; idx. 13
+                    2.28520723,
+                    4.82767506,
+                    4.30421368,
+                    8.8275313,  # 2nd highest value; idx. 17
+                    5.44029958,  # 5th highest value; idx. 18
+                    -4.4735794,
+                    7.38579536,  # 3rd highest value; idx. 20
+                    -2.91051663,
+                    2.61946077,
+                    -2.5674762,
+                    -9.48959302,
+                    -4.02922645,
+                    -1.35416918,
+                    9.67702323,  # 1st highest value; idx. 27
+                    -5.89478553,
+                    1.85370467,
+                ],  # cummulative prob of 5 highest values <= 0.6
+            ],
+            dtype=tf.float32,
+        )
+
+        non_inf_expected_idx = tf.convert_to_tensor(
+            [[0, 0], [0, 9], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 18], [1, 20], [1, 27]], dtype=tf.int32,
+        )  # expected non filtered idx as noted above
+
+        non_inf_expected_output = tf.convert_to_tensor(
+            [8.222099, 7.3534126, 8.432078, 7.4402075, 9.38451, 6.271159, 8.827531, 5.4402995, 7.3857956, 9.677023],
+            dtype=tf.float32,
+        )  # expected non filtered values as noted above
+
+        output = tf_top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
+
+        non_inf_output = output[output != -float("inf")]
+        non_inf_idx = tf.cast(
+            tf.where(tf.not_equal(output, tf.constant(-float("inf"), dtype=tf.float32))), dtype=tf.int32,
+        )
+
+        tf.debugging.assert_near(non_inf_output, non_inf_expected_output, rtol=1e-12)
+        tf.debugging.assert_equal(non_inf_idx, non_inf_expected_idx)

tests/test_modeling_tf_ctrl.py

@@ -31,6 +31,7 @@ if is_tf_available():
 class TFCTRLModelTest(TFModelTesterMixin, unittest.TestCase):
 
     all_model_classes = (TFCTRLModel, TFCTRLLMHeadModel) if is_tf_available() else ()
+    all_generative_model_classes = (TFCTRLLMHeadModel,) if is_tf_available() else ()
 
     class TFCTRLModelTester(object):
         def __init__(

tests/test_modeling_tf_gpt2.py

@@ -37,7 +37,7 @@ if is_tf_available():
 class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
 
     all_model_classes = (TFGPT2Model, TFGPT2LMHeadModel, TFGPT2DoubleHeadsModel) if is_tf_available() else ()
-    # all_model_classes = (TFGPT2Model, TFGPT2LMHeadModel) if is_tf_available() else ()
+    all_generative_model_classes = (TFGPT2LMHeadModel,) if is_tf_available() else ()
 
     class TFGPT2ModelTester(object):
         def __init__(
@@ -89,6 +89,8 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
             self.num_labels = num_labels
             self.num_choices = num_choices
             self.scope = scope
+            self.bos_token_id = vocab_size - 1
+            self.eos_token_id = vocab_size - 1
 
         def prepare_config_and_inputs(self):
             input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -123,9 +125,11 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
                 # hidden_dropout_prob=self.hidden_dropout_prob,
                 # attention_probs_dropout_prob=self.attention_probs_dropout_prob,
                 n_positions=self.max_position_embeddings,
-                n_ctx=self.max_position_embeddings
+                n_ctx=self.max_position_embeddings,
                 # type_vocab_size=self.type_vocab_size,
                 # initializer_range=self.initializer_range
+                bos_token_id=self.bos_token_id,
+                eos_token_ids=self.eos_token_id,
             )
 
             head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -144,7 +148,11 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
 
         def create_and_check_gpt2_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = TFGPT2Model(config=config)
-            inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+            inputs = {
+                "input_ids": input_ids,
+                "attention_mask": input_mask,
+                "token_type_ids": token_type_ids,
+            }
             sequence_output = model(inputs)[0]
 
             inputs = [input_ids, None, input_mask]  # None is the input for 'past'
@@ -156,18 +164,22 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
                 "sequence_output": sequence_output.numpy(),
             }
             self.parent.assertListEqual(
-                list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
+                list(result["sequence_output"].shape), [self.batch_size, self.seq_length, self.hidden_size],
             )
 
         def create_and_check_gpt2_lm_head(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = TFGPT2LMHeadModel(config=config)
-            inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids}
+            inputs = {
+                "input_ids": input_ids,
+                "attention_mask": input_mask,
+                "token_type_ids": token_type_ids,
+            }
             prediction_scores = model(inputs)[0]
             result = {
                 "prediction_scores": prediction_scores.numpy(),
             }
             self.parent.assertListEqual(
-                list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size]
+                list(result["prediction_scores"].shape), [self.batch_size, self.seq_length, self.vocab_size],
             )
 
         def create_and_check_gpt2_double_head(
@@ -188,7 +200,7 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
             lm_logits, mc_logits = model(inputs)[:2]
             result = {"lm_logits": lm_logits.numpy(), "mc_logits": mc_logits.numpy()}
             self.parent.assertListEqual(
-                list(result["lm_logits"].shape), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size]
+                list(result["lm_logits"].shape), [self.batch_size, self.num_choices, self.seq_length, self.vocab_size],
             )
             self.parent.assertListEqual(list(result["mc_logits"].shape), [self.batch_size, self.num_choices])
 
@@ -207,7 +219,11 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
                 choice_labels,
             ) = config_and_inputs
 
-            inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+            inputs_dict = {
+                "input_ids": input_ids,
+                "token_type_ids": token_type_ids,
+                "attention_mask": input_mask,
+            }
             return config, inputs_dict
 
     def setUp(self):
@@ -234,3 +250,48 @@ class TFGPT2ModelTest(TFModelTesterMixin, unittest.TestCase):
         for model_name in list(TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
             model = TFGPT2Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
+
+
+def prepare_generation_special_tokens():
+    return {"bos_token_id": 50256, "eos_token_id": 50256}
+
+
+class TFGPT2ModelLanguageGenerationTest(unittest.TestCase):
+
+    special_tokens = prepare_generation_special_tokens()
+
+    @slow
+    def test_lm_generate_distilgpt2(self):
+        model = TFGPT2LMHeadModel.from_pretrained("distilgpt2")
+        input_ids = tf.convert_to_tensor([[464, 1893]], dtype=tf.int32)  # The president
+        expected_output_ids = [
+            464,
+            1893,
+            286,
+            262,
+            1578,
+            1829,
+            11,
+            290,
+            262,
+            1893,
+            286,
+            262,
+            1578,
+            7526,
+            11,
+            423,
+            587,
+            287,
+            262,
+            2635,
+        ]  # The president of the United States, and the president of the United Kingdom, have been in the White
+
+        output_ids = model.generate(
+            input_ids,
+            do_sample=False,
+            bos_token_id=self.special_tokens["bos_token_id"],
+            eos_token_ids=self.special_tokens["eos_token_id"],
+        )
+
+        self.assertListEqual(output_ids[0].numpy().tolist(), expected_output_ids)

tests/test_modeling_tf_openai_gpt.py

@@ -39,6 +39,9 @@ class TFOpenAIGPTModelTest(TFModelTesterMixin, unittest.TestCase):
     all_model_classes = (
         (TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TFOpenAIGPTDoubleHeadsModel) if is_tf_available() else ()
     )
+    all_generative_model_classes = (
+        (TFOpenAIGPTLMHeadModel,) if is_tf_available() else ()
+    )  # TODO (PVP): Add Double HeadsModel when generate() function is changed accordingly
 
     class TFOpenAIGPTModelTester(object):
         def __init__(

tests/test_modeling_tf_roberta.py

@@ -222,9 +222,9 @@ class TFRobertaModelIntegrationTest(unittest.TestCase):
         self.assertEqual(list(output.numpy().shape), expected_shape)
         # compare the actual values for a slice.
         expected_slice = tf.constant(
-            [[[33.8843, -4.3107, 22.7779], [4.6533, -2.8099, 13.6252], [1.8222, -3.6898, 8.8600]]]
+            [[[33.8802, -4.3103, 22.7761], [4.6539, -2.8098, 13.6253], [1.8228, -3.6898, 8.8600]]]
         )
-        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-3))
+        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-4))
 
     @slow
     def test_inference_no_head(self):
@@ -234,9 +234,9 @@ class TFRobertaModelIntegrationTest(unittest.TestCase):
         output = model(input_ids)[0]
         # compare the actual values for a slice.
         expected_slice = tf.constant(
-            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0539, -0.0174], [0.0548, 0.0799, 0.1687]]]
+            [[[-0.0231, 0.0782, 0.0074], [-0.1854, 0.0540, -0.0175], [0.0548, 0.0799, 0.1687]]]
         )
-        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-3))
+        self.assertTrue(numpy.allclose(output[:, :3, :3].numpy(), expected_slice.numpy(), atol=1e-4))
 
     @slow
     def test_inference_classification_head(self):
@@ -247,4 +247,4 @@ class TFRobertaModelIntegrationTest(unittest.TestCase):
         expected_shape = [1, 3]
         self.assertEqual(list(output.numpy().shape), expected_shape)
         expected_tensor = tf.constant([[-0.9469, 0.3913, 0.5118]])
-        self.assertTrue(numpy.allclose(output.numpy(), expected_tensor.numpy(), atol=1e-3))
+        self.assertTrue(numpy.allclose(output.numpy(), expected_tensor.numpy(), atol=1e-4))

tests/test_modeling_tf_transfo_xl.py

@@ -37,6 +37,8 @@ if is_tf_available():
 class TFTransfoXLModelTest(TFModelTesterMixin, unittest.TestCase):
 
     all_model_classes = (TFTransfoXLModel, TFTransfoXLLMHeadModel) if is_tf_available() else ()
+    all_generative_model_classes = () if is_tf_available() else ()
+    # TODO: add this test when TFTransfoXLLMHead has a linear output layer implemented
     test_pruning = False
     test_torchscript = False
     test_resize_embeddings = False
@@ -62,6 +64,7 @@ class TFTransfoXLModelTest(TFModelTesterMixin, unittest.TestCase):
             num_hidden_layers=5,
             scope=None,
             seed=1,
+            eos_token_id=0,
         ):
             self.parent = parent
             self.batch_size = batch_size
@@ -82,6 +85,7 @@ class TFTransfoXLModelTest(TFModelTesterMixin, unittest.TestCase):
             self.num_hidden_layers = num_hidden_layers
             self.scope = scope
             self.seed = seed
+            self.eos_token_id = eos_token_id
 
         def prepare_config_and_inputs(self):
             input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -103,6 +107,7 @@ class TFTransfoXLModelTest(TFModelTesterMixin, unittest.TestCase):
                 d_inner=self.d_inner,
                 div_val=self.div_val,
                 n_layer=self.num_hidden_layers,
+                eos_token_ids=self.eos_token_id,
             )
 
             return (config, input_ids_1, input_ids_2, lm_labels)

tests/test_modeling_tf_xlm.py

@@ -43,6 +43,9 @@ class TFXLMModelTest(TFModelTesterMixin, unittest.TestCase):
         if is_tf_available()
         else ()
     )
+    all_generative_model_classes = (
+        (TFXLMWithLMHeadModel,) if is_tf_available() else ()
+    )  # TODO (PVP): Check other models whether language generation is also applicable
 
     class TFXLMModelTester(object):
         def __init__(
@@ -75,6 +78,7 @@ class TFXLMModelTest(TFModelTesterMixin, unittest.TestCase):
             summary_type="last",
             use_proj=True,
             scope=None,
+            bos_token_id=0,
         ):
             self.parent = parent
             self.batch_size = batch_size
@@ -105,6 +109,7 @@ class TFXLMModelTest(TFModelTesterMixin, unittest.TestCase):
             self.num_labels = num_labels
             self.num_choices = num_choices
             self.scope = scope
+            self.bos_token_id = bos_token_id
 
         def prepare_config_and_inputs(self):
             input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -145,6 +150,7 @@ class TFXLMModelTest(TFModelTesterMixin, unittest.TestCase):
                 initializer_range=self.initializer_range,
                 summary_type=self.summary_type,
                 use_proj=self.use_proj,
+                bos_token_id=self.bos_token_id,
             )
 
             return (

tests/test_modeling_tf_xlnet.py

@@ -51,6 +51,9 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
         if is_tf_available()
         else ()
     )
+    all_generative_model_classes = (
+        (TFXLNetLMHeadModel,) if is_tf_available() else ()
+    )  # TODO (PVP): Check other models whether language generation is also applicable
     test_pruning = False
 
     class TFXLNetModelTester(object):
@@ -77,6 +80,9 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
             initializer_range=0.05,
             seed=1,
             type_vocab_size=2,
+            bos_token_id=1,
+            eos_token_id=2,
+            pad_token_id=5,
         ):
             self.parent = parent
             self.batch_size = batch_size
@@ -100,6 +106,9 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
             self.seed = seed
             self.type_vocab_size = type_vocab_size
             self.type_sequence_label_size = type_sequence_label_size
+            self.bos_token_id = bos_token_id
+            self.pad_token_id = pad_token_id
+            self.eos_token_id = eos_token_id
 
         def prepare_config_and_inputs(self):
             input_ids_1 = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
@@ -139,6 +148,9 @@ class TFXLNetModelTest(TFModelTesterMixin, unittest.TestCase):
                 bi_data=self.bi_data,
                 initializer_range=self.initializer_range,
                 num_labels=self.type_sequence_label_size,
+                bos_token_id=self.bos_token_id,
+                pad_token_id=self.pad_token_id,
+                eos_token_id=self.eos_token_id,
             )
 
             return (

tests/test_modeling_xlm_roberta.py

@@ -30,14 +30,13 @@ class XLMRobertaModelIntegrationTest(unittest.TestCase):
     @slow
     def test_xlm_roberta_base(self):
         model = XLMRobertaModel.from_pretrained("xlm-roberta-base")
-        input_ids = torch.tensor([0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]).unsqueeze(
-            0
-        )  # The dog is cute and lives in the garden house
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # The dog is cute and lives in the garden house
 
         expected_output_shape = torch.Size((1, 12, 768))  # batch_size, sequence_length, embedding_vector_dim
         expected_output_values_last_dim = torch.tensor(
-            [-0.0101, 0.1218, -0.0803, 0.0801, 0.1327, 0.0776, -0.1215, 0.2383, 0.3338, 0.3106, 0.0300, 0.0252]
-        ).unsqueeze(0)
+            [[-0.0101, 0.1218, -0.0803, 0.0801, 0.1327, 0.0776, -0.1215, 0.2383, 0.3338, 0.3106, 0.0300, 0.0252]]
+        )
         #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.base')
         #  xlmr.eval()
         #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]
@@ -50,14 +49,13 @@ class XLMRobertaModelIntegrationTest(unittest.TestCase):
     @slow
     def test_xlm_roberta_large(self):
         model = XLMRobertaModel.from_pretrained("xlm-roberta-large")
-        input_ids = torch.tensor([0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]).unsqueeze(
-            0
-        )  # The dog is cute and lives in the garden house
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # The dog is cute and lives in the garden house
 
         expected_output_shape = torch.Size((1, 12, 1024))  # batch_size, sequence_length, embedding_vector_dim
         expected_output_values_last_dim = torch.tensor(
-            [-0.0699, -0.0318, 0.0705, -0.1241, 0.0999, -0.0520, 0.1004, -0.1838, -0.4704, 0.1437, 0.0821, 0.0126]
-        ).unsqueeze(0)
+            [[-0.0699, -0.0318, 0.0705, -0.1241, 0.0999, -0.0520, 0.1004, -0.1838, -0.4704, 0.1437, 0.0821, 0.0126]]
+        )
         #  xlmr = torch.hub.load('pytorch/fairseq', 'xlmr.large')
         #  xlmr.eval()
         #  expected_output_values_last_dim = xlmr.extract_features(input_ids[0])[:, :, -1]

tests/test_pipelines.py

@@ -2,9 +2,16 @@ import unittest
 from typing import Iterable, List, Optional
 
 from transformers import pipeline
-from transformers.pipelines import Pipeline
+from transformers.pipelines import (
+    FeatureExtractionPipeline,
+    FillMaskPipeline,
+    NerPipeline,
+    Pipeline,
+    QuestionAnsweringPipeline,
+    TextClassificationPipeline,
+)
 
-from .utils import require_tf, require_torch
+from .utils import require_tf, require_torch, slow
 
 
 QA_FINETUNED_MODELS = [
@@ -304,3 +311,30 @@ class MultiColumnInputTestCase(unittest.TestCase):
         for tokenizer, model, config in TF_QA_FINETUNED_MODELS:
             nlp = pipeline(task="question-answering", model=model, config=config, tokenizer=tokenizer, framework="tf")
             self._test_multicolumn_pipeline(nlp, valid_samples, invalid_samples, mandatory_output_keys)
+
+
+class PipelineCommonTests(unittest.TestCase):
+
+    pipelines = (
+        NerPipeline,
+        FeatureExtractionPipeline,
+        QuestionAnsweringPipeline,
+        FillMaskPipeline,
+        TextClassificationPipeline,
+    )
+
+    @slow
+    @require_tf
+    def test_tf_defaults(self):
+        # Test that pipelines can be correctly loaded without any argument
+        for default_pipeline in self.pipelines:
+            with self.subTest(msg="Testing Torch defaults with PyTorch and {}".format(default_pipeline.task)):
+                default_pipeline(framework="tf")
+
+    @slow
+    @require_torch
+    def test_pt_defaults(self):
+        # Test that pipelines can be correctly loaded without any argument
+        for default_pipeline in self.pipelines:
+            with self.subTest(msg="Testing Torch defaults with PyTorch and {}".format(default_pipeline.task)):
+                default_pipeline(framework="pt")

tests/test_tokenization_common.py

@@ -449,6 +449,10 @@ class TokenizerTesterMixin:
 
         sequence = "Sequence"
         padding_size = 10
+
+        # check correct behaviour if no pad_token_id exists and add it eventually
+        self._check_no_pad_token_padding(tokenizer, sequence)
+
         padding_idx = tokenizer.pad_token_id
 
         # RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
@@ -490,6 +494,10 @@ class TokenizerTesterMixin:
         tokenizer = self.get_tokenizer()
 
         sequence = "Sequence"
+
+        # check correct behaviour if no pad_token_id exists and add it eventually
+        self._check_no_pad_token_padding(tokenizer, sequence)
+
         padding_size = 10
         padding_idx = tokenizer.pad_token_id
         token_type_padding_idx = tokenizer.pad_token_type_id
@@ -503,6 +511,7 @@ class TokenizerTesterMixin:
 
         # Test right padding
         tokenizer.padding_side = "right"
+
         padded_sequence = tokenizer.encode_plus(
             sequence,
             max_length=sequence_length + padding_size,
@@ -588,10 +597,14 @@ class TokenizerTesterMixin:
 
         maximum_length = len(max([encoded_sequence["input_ids"] for encoded_sequence in encoded_sequences], key=len))
 
+        # check correct behaviour if no pad_token_id exists and add it eventually
+        self._check_no_pad_token_padding(tokenizer, sequences)
+
         encoded_sequences_padded = [
             tokenizer.encode_plus(sequence, pad_to_max_length=True, max_length=maximum_length)
             for sequence in sequences
         ]
+
         encoded_sequences_batch_padded = tokenizer.batch_encode_plus(sequences, pad_to_max_length=True)
         self.assertListEqual(
             encoded_sequences_padded,
@@ -610,6 +623,10 @@ class TokenizerTesterMixin:
         ]
 
         max_length = 100
+
+        # check correct behaviour if no pad_token_id exists and add it eventually
+        self._check_no_pad_token_padding(tokenizer, sequences)
+
         encoded_sequences = [
             tokenizer.encode_plus(sequence, pad_to_max_length=True, max_length=max_length) for sequence in sequences
         ]
@@ -620,6 +637,7 @@ class TokenizerTesterMixin:
 
         # Left padding tests
         tokenizer = self.get_tokenizer()
+
         tokenizer.padding_side = "left"
         sequences = [
             "Testing batch encode plus",
@@ -628,6 +646,10 @@ class TokenizerTesterMixin:
         ]
 
         max_length = 100
+
+        # check correct behaviour if no pad_token_id exists and add it eventually
+        self._check_no_pad_token_padding(tokenizer, sequences)
+
         encoded_sequences = [
             tokenizer.encode_plus(sequence, pad_to_max_length=True, max_length=max_length) for sequence in sequences
         ]
@@ -668,3 +690,15 @@ class TokenizerTesterMixin:
                 encoded_value = encoded_sequences[key]
 
                 self.assertEqual(pytorch_value, tensorflow_value, encoded_value)
+
+    def _check_no_pad_token_padding(self, tokenizer, sequences):
+        # if tokenizer does not have pad_token_id, an error should be thrown
+        if tokenizer.pad_token_id is None:
+            with self.assertRaises(ValueError):
+                if isinstance(sequences, list):
+                    tokenizer.batch_encode_plus(sequences, pad_to_max_length=True)
+                else:
+                    tokenizer.encode_plus(sequences, pad_to_max_length=True)
+
+            # add pad_token_id to pass subsequent tests
+            tokenizer.add_special_tokens({"pad_token": "<PAD>"})

tests/utils.py

@@ -29,8 +29,22 @@ def parse_flag_from_env(key, default=False):
     return _value
 
 
+def parse_int_from_env(key, default=None):
+    try:
+        value = os.environ[key]
+    except KeyError:
+        _value = default
+    else:
+        try:
+            _value = int(value)
+        except ValueError:
+            raise ValueError("If set, {} must be a int.".format(key))
+    return _value
+
+
 _run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
 _run_custom_tokenizers = parse_flag_from_env("RUN_CUSTOM_TOKENIZERS", default=False)
+_tf_gpu_memory_limit = parse_int_from_env("TF_GPU_MEMORY_LIMIT", default=None)
 
 
 def slow(test_case):

utils/link_tester.py

@@ -14,6 +14,9 @@ import requests
 REGEXP_FIND_S3_LINKS = r"""([\"'])(https:\/\/s3)(.*)?\1"""
 
 
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+
+
 def list_python_files_in_repository():
     """ List all python files in the repository.
 
@@ -36,7 +39,7 @@ def find_all_links(file_paths):
     for path in file_paths:
         links += scan_code_for_links(path)
 
-    return links
+    return [link for link in links if link != S3_BUCKET_PREFIX]
 
 
 def scan_code_for_links(source):