Merge branch 'master' into master

9a3b173c · Thomas Wolf · GitHub · ad908686 · 8a628355 · 9a3b173c
Unverified Commit 9a3b173c authored Nov 04, 2019 by Thomas Wolf Committed by GitHub Nov 04, 2019
20 changed files
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -81,7 +81,7 @@ jobs:
            - checkout
            - run: sudo pip install --progress-bar off -r docs/requirements.txt
            - run: sudo pip install --progress-bar off -r requirements.txt
-            - run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+            - run: ./.circleci/deploy.sh
 workflow_filters: &workflow_filters
    filters:
        branches:
@@ -96,4 +96,4 @@ workflows:
            - build_py3_tf
            - build_py2_torch
            - build_py2_tf
            - deploy_doc: *workflow_filters
\ No newline at end of file
--- a/.circleci/deploy.sh
+++ b/.circleci/deploy.sh
+cd docs
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		echo "Pushing version" $2
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
--- a/.github/ISSUE_TEMPLATE/---new-benchmark.md
+++ b/.github/ISSUE_TEMPLATE/---new-benchmark.md
+---
+name: "\U0001F5A5 New Benchmark"
+about: You benchmark a part of this library and would like to share your results
+title: "[Benchmark]"
+labels: ''
+assignees: ''
+---
+# Benchmarking Transformers
+## Benchmark
+Which part of Transformers did you benchmark?
+## Set-up
+What did you run your benchmarks on? Please include details, such as: CPU, GPU? If using multiple GPUs, which parallelization did you use?
+## Results
+Put your results here!
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -62,6 +62,8 @@ Awesome! Please provide the following information:
 If you are willing to contribute the model yourself, let us know so we can best
 guide you.
+We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder.
 ### Do you want a new feature (that is not a model)?
 A world-class feature request addresses the following points:
@@ -81,6 +83,8 @@ A world-class feature request addresses the following points:
 If your issue is well written we're already 80% of the way there by the time you
 post it.
+We have added **templates** to guide you in the process of adding a new example script for training or testing the models in the library. You can find them in the [`templates`](./templates) folder.
 ## Start contributing! (Pull Requests)
 Before writing code, we strongly advise you to search through the exising PRs or

--- a/README.md
+++ b/README.md
@@ -39,7 +39,7 @@ State-of-the-art NLP for everyone
 Lower compute costs, smaller carbon footprint
 - Researchers can share trained models instead of always retraining
 - Practitioners can reduce compute time and production costs
- 8 architectures with over 30 pretrained models, some in more than 100 languages
+- 10 architectures with over 30 pretrained models, some in more than 100 languages
 Choose the right framework for every part of a model's lifetime
 - Train state-of-the-art models in 3 lines of code
@@ -58,7 +58,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
+| [Documentation](https://huggingface.co/transformers/) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) | Full API documentation and more |
 ## Installation
@@ -111,7 +111,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 ## Model architectures
-🤗 Transformers currently provides 8 NLU/NLG architectures:
+🤗 Transformers currently provides 10 NLU/NLG architectures:
 1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
 2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
@@ -122,6 +122,7 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
 8. **[DistilBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation).
 9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
+10. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
@@ -176,10 +177,11 @@ BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNex
 # All the classes for an architecture can be initiated from pretrained weights for this architecture
 # Note that additional weights added for fine-tuning are only initialized
 # and need to be trained on the down-stream task
-tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+pretrained_weights = 'bert-base-uncased'
+tokenizer = BertTokenizer.from_pretrained(pretrained_weights)
 for model_class in BERT_MODEL_CLASSES:
    # Load pretrained model/tokenizer
-    model = model_class.from_pretrained('bert-base-uncased')
+    model = model_class.from_pretrained(pretrained_weights)
    # Models can return full list of hidden-states & attentions weights at each layer
    model = model_class.from_pretrained(pretrained_weights,
@@ -242,8 +244,9 @@ sentence_2 = "His findings were not compatible with this research."
 inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
 inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
-pred_1 = pytorch_model(**inputs_1)[0].argmax().item()
+pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
-pred_2 = pytorch_model(**inputs_2)[0].argmax().item()
+pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
 print("sentence_1 is", "a paraphrase" if pred_1 else "not a paraphrase", "of sentence_0")
 print("sentence_2 is", "a paraphrase" if pred_2 else "not a paraphrase", "of sentence_0")
 ```
@@ -411,7 +414,7 @@ and from the Salesforce CTRL model:
 python ./examples/run_generation.py \
    --model_type=ctrl \
    --length=20 \
-    --model_name_or_path=gpt2 \
+    --model_name_or_path=ctrl \
    --temperature=0 \
    --repetition_penalty=1.2 \
 ```
@@ -547,12 +550,11 @@ for batch in train_data:
 We now have a paper you can cite for the 🤗 Transformers library:
 ```
-@misc{wolf2019transformers,
+@article{Wolf2019HuggingFacesTS,
-    title={Transformers: State-of-the-art Natural Language Processing},
+  title={HuggingFace's Transformers: State-of-the-art Natural Language Processing},
-    author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Jamie Brew},
+  author={Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and R'emi Louf and Morgan Funtowicz and Jamie Brew},
-    year={2019},
+  journal={ArXiv},
-    eprint={1910.03771},
+  year={2019},
-    archivePrefix={arXiv},
+  volume={abs/1910.03771}
-    primaryClass={cs.CL}
 }
 ```
--- a/docs/source/benchmarks.md
+++ b/docs/source/benchmarks.md
+# Benchmarks
+This section is dedicated to the Benchmarks done by the library, both by maintainers, contributors and users. These 
+benchmark will help keep track of the preformance improvements that are brought to our models across versions.
+## Benchmarking all models for inference
+As of version 2.1 we have benchmarked all models for inference, across many different settings: using PyTorch, with
+and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
+TensorFlow XLA) and GPUs.
+The approach is detailed in the [following blogpost](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
+The results are available [here](https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing).
+## TF2 with mixed precision, XLA, Distribution (@tlkh)
+This work was done by [Timothy Liu](https://github.com/tlkh).
+There are very positive results to be gained from the various TensorFlow 2.0 features:
+- Automatic Mixed Precision (AMP)
+- XLA compiler
+- Distribution strategies (multi-GPU)
+The benefits are listed here (tested on CoLA, MRPC, SST-2):
+- AMP: Between 1.4x to 1.6x decrease in overall time without change in batch size
+- AMP+XLA: Up to 2.5x decrease in overall time on SST-2 (larger dataset)
+- Distribution: Between 1.4x to 3.4x decrease in overall time on 4xV100
+- Combined: Up to 5.7x decrease in overall training time, or 9.1x training throughput
+The model quality (measured by the validation accuracy) fluctuates slightly. Taking an average of 4 training runs 
+on a single GPU gives the following results:
+- CoLA: AMP results in slighter lower acc (0.820 vs 0.824)
+- MRPC: AMP results in lower acc (0.823 vs 0.835)
+- SST-2: AMP results in slighter lower acc (0.918 vs 0.922)
+However, in a distributed setting with 4xV100 (4x batch size), AMP can yield in better results:
+CoLA: AMP results in higher acc (0.828 vs 0.812)
+MRPC: AMP results in lower acc (0.817 vs 0.827)
+SST-2: AMP results in slightly lower acc (0.926 vs 0.929)
+The benchmark script is available [here](https://github.com/NVAITC/benchmarking/blob/master/tf2/bert_dist.py).
+Note: on some tasks (e.g. MRPC), the dataset is too small. The overhead due to the model compilation with XLA as well
+as the distribution strategy setup does not speed things up. The XLA compile time is also the reason why although throughput 
+can increase a lot (e.g. 2.7x for single GPU), overall (end-to-end) training speed-up is not as fast (as low as 1.4x)
+The benefits as seen on SST-2 (larger dataset) is much clear.
+All results can be seen on this [Google Sheet](https://docs.google.com/spreadsheets/d/1538MN224EzjbRL239sqSiUy6YY-rAjHyXhTzz_Zptls/edit#gid=960868445).
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -63,6 +63,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
    bertology
    torchscript
    multilingual
+    benchmarks
 .. toctree::
    :maxdepth: 2

--- a/docs/source/model_doc/ctrl.rst
+++ b/docs/source/model_doc/ctrl.rst
 CTRL
 ----------------------------------------------------
+Note: if you fine-tune a CTRL model using the Salesforce code (https://github.com/salesforce/ctrl),
+you'll be able to convert from TF to our HuggingFace/Transformers format using the 
+``convert_tf_to_huggingface_pytorch.py`` script (see `issue #1654 <https://github.com/huggingface/transformers/issues/1654>`_).
 ``CTRLConfig``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

--- a/docs/source/pretrained_models.rst
+++ b/docs/source/pretrained_models.rst
@@ -136,6 +136,10 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``distilgpt2``                                             | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
 |                   |                                                            | | The DistilGPT2 model distilled from the GPT2 model `gpt2` checkpoint.                                                               |
 |                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilroberta-base``                                     | | 6-layer, 768-hidden, 12-heads, 82M parameters                                                                                       |
+|                   |                                                            | | The DistilRoBERTa model distilled from the RoBERTa model `roberta-base` checkpoint.                                                 |
+|                   |                                                            | (see `details <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__)                                     |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | CTRL              | ``ctrl``                                                   | | 48-layer, 1280-hidden, 16-heads, 1.6B parameters                                                                                    |
 |                   |                                                            | | Salesforce's Large-sized CTRL English model                                                                                         |

--- a/examples/README.md
+++ b/examples/README.md
@@ -5,11 +5,36 @@ similar API between the different models.
 | Section                    | Description                                                                                                                                                |
 |----------------------------|------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. 
 | [Language Model fine-tuning](#language-model-fine-tuning) | Fine-tuning the library models for language modeling on a text dataset. Causal language modeling for GPT/GPT-2, masked language modeling for BERT/RoBERTa. |
 | [Language Generation](#language-generation) | Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.                                         |
 | [GLUE](#glue) | Examples running BERT/XLM/XLNet/RoBERTa on the 9 GLUE tasks. Examples feature distributed training as well as half-precision.                              |
 | [SQuAD](#squad) | Using BERT/RoBERTa/XLNet/XLM for question answering, examples with distributed training.                                                                                  |
 | [Multiple Choice](#multiple-choice) | Examples running BERT/XLNet/RoBERTa on the SWAG/RACE/ARC tasks. 
+| [Named Entity Recognition](#named-entity-recognition) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training.                                                                                  |
+| [Abstractive summarization](#abstractive-summarization) | Fine-tuning the library models for abstractive summarization tasks on the CNN/Daily Mail dataset. |
+## TensorFlow 2.0 Bert models on GLUE
+Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_glue.py).
+Fine-tuning the library TensorFlow 2.0 Bert model for sequence classification on the  MRPC task of the GLUE benchmark: [General Language Understanding Evaluation](https://gluebenchmark.com/).
+This script has an option for mixed precision (Automatic Mixed Precision / AMP) to run models on Tensor Cores (NVIDIA Volta/Turing GPUs) and future hardware and an option for XLA, which uses the XLA compiler to reduce model runtime.
+Options are toggled using `USE_XLA` or `USE_AMP` variables in the script.
+These options and the below benchmark are provided by @tlkh.
+Quick benchmarks from the script (no other modifications):
+| GPU    | Mode | Time (2nd epoch) | Val Acc (3 runs) |
+| --------- | -------- | ----------------------- | ----------------------|
+| Titan V | FP32 | 41s | 0.8438/0.8281/0.8333 |
+| Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
+| V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
+| V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
+| 1080 Ti | FP32 | 55s | - | 
+Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
 ## Language model fine-tuning
@@ -77,7 +102,7 @@ python run_lm_finetuning.py \
 Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/run_generation.py).
-Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL and XLNet.
+Conditional text generation using the auto-regressive models of the library: GPT, GPT-2, Transformer-XL, XLNet, CTRL.
 A similar script is used for our official demo [Write With Transfomer](https://transformer.huggingface.co), where you
 can try out the different models available in the library.
@@ -387,7 +412,7 @@ f1 = 93.15
 exact_match = 86.91
 ```
-This fine-tuneds model is available as a checkpoint under the reference
+This fine-tuned model is available as a checkpoint under the reference
 `bert-large-uncased-whole-word-masking-finetuned-squad`.
 #### Fine-tuning XLNet on SQuAD
@@ -427,3 +452,132 @@ Training with the previously defined hyper-parameters yields the following resul
 "HasAns_total": 10570
 }
 ```
+## Named Entity Recognition
+Based on the script [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py).
+This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
+Details and results for the fine-tuning provided by @stefan-it.
+### Data (Download and pre-processing steps)
+Data can be obtained from the [GermEval 2014](https://sites.google.com/site/germeval2014ner/data) shared task page.
+Here are the commands for downloading and pre-processing train, dev and test datasets. The original data format has four (tab-separated) columns, in a pre-processing step only the two relevant columns (token and outer span NER annotation) are extracted:
+```bash
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-train.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > train.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-dev.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > dev.txt.tmp
+curl -L 'https://sites.google.com/site/germeval2014ner/data/NER-de-test.tsv?attredirects=0&d=1' \
+| grep -v "^#" | cut -f 2,3 | tr '\t' ' ' > test.txt.tmp
+```
+The GermEval 2014 dataset contains some strange "control character" tokens like `'\x96', '\u200e', '\x95', '\xad' or '\x80'`. One problem with these tokens is, that `BertTokenizer` returns an empty token for them, resulting in misaligned `InputExample`s. I wrote a script that a) filters these tokens and b) splits longer sentences into smaller ones (once the max. subtoken length is reached).
+```bash
+wget "https://raw.githubusercontent.com/stefan-it/fine-tuned-berts-seq/master/scripts/preprocess.py"
+```
+Let's define some variables that we need for further pre-processing steps and training the model:
+```bash
+export MAX_LENGTH=128
+export BERT_MODEL=bert-base-multilingual-cased
+```
+Run the pre-processing script on training, dev and test datasets:
+```bash
+python3 preprocess.py train.txt.tmp $BERT_MODEL $MAX_LENGTH > train.txt
+python3 preprocess.py dev.txt.tmp $BERT_MODEL $MAX_LENGTH > dev.txt
+python3 preprocess.py test.txt.tmp $BERT_MODEL $MAX_LENGTH > test.txt
+```
+The GermEval 2014 dataset has much more labels than CoNLL-2002/2003 datasets, so an own set of labels must be used:
+```bash
+cat train.txt dev.txt test.txt | cut -d " " -f 2 | grep -v "^$"| sort | uniq > labels.txt
+```
+### Training
+Additional environment variables must be set:
+```bash
+export OUTPUT_DIR=germeval-model
+export BATCH_SIZE=32
+export NUM_EPOCHS=3
+export SAVE_STEPS=750
+export SEED=1
+```
+To start training, just run:
+```bash
+python3 run_ner.py --data_dir ./ \
+--model_type bert \
+--labels ./labels.txt \
+--model_name_or_path $BERT_MODEL \
+--output_dir $OUTPUT_DIR \
+--max_seq_length  $MAX_LENGTH \
+--num_train_epochs $NUM_EPOCHS \
+--per_gpu_train_batch_size $BATCH_SIZE \
+--save_steps $SAVE_STEPS \
+--seed $SEED \
+--do_train \
+--do_eval \
+--do_predict
+```
+If your GPU supports half-precision training, just add the `--fp16` flag. After training, the model will be both evaluated on development and test datasets.
+### Evaluation
+Evaluation on development dataset outputs the following for our example:
+```bash
+10/04/2019 00:42:06 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:06 - INFO - __main__ -     f1 = 0.8623348017621146
+10/04/2019 00:42:06 - INFO - __main__ -     loss = 0.07183869666975543
+10/04/2019 00:42:06 - INFO - __main__ -     precision = 0.8467916366258111
+10/04/2019 00:42:06 - INFO - __main__ -     recall = 0.8784592370979806
+```
+On the test dataset the following results could be achieved:
+```bash
+10/04/2019 00:42:42 - INFO - __main__ -   ***** Eval results  *****
+10/04/2019 00:42:42 - INFO - __main__ -     f1 = 0.8614389652384803
+10/04/2019 00:42:42 - INFO - __main__ -     loss = 0.07064602487454782
+10/04/2019 00:42:42 - INFO - __main__ -     precision = 0.8604651162790697
+10/04/2019 00:42:42 - INFO - __main__ -     recall = 0.8624150210424085
+```
+## Abstractive summarization
+Based on the script
+[`run_summarization_finetuning.py`](https://github.com/huggingface/transformers/blob/master/examples/run_summarization_finetuning.py).
+Before running this script you should download **both** CNN and Daily Mail
+datasets from [Kyunghyun Cho's website](https://cs.nyu.edu/~kcho/DMQA/)  (the
+links next to "Stories") in the same folder. Then uncompress the archives by running:
+```bash
+tar -xvf cnn_stories.tgz && tar -xvf dailymail_stories.tgz
+```
+note that the finetuning script **will not work** if you do not download both
+datasets. We will refer as `$DATA_PATH` the path to where you uncompressed both
+archive.
+```bash
+export DATA_PATH=/path/to/dataset/
+python run_summarization_finetuning.py \
+    --output_dir=output \
+    --model_type=bert2bert \
+    --model_name_or_path=bert2bert \
+    --do_train \
+    --data_path=$DATA_PATH \
+```
--- a/examples/benchmarks.py
+++ b/examples/benchmarks.py
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Benchmarking the library on inference and training """
+# If checking the tensors placement
+# tf.debugging.set_log_device_placement(True)
+from typing import List
+import timeit
+from transformers import is_tf_available, is_torch_available
+from time import time
+import argparse
+import csv
+if is_tf_available():
+    import tensorflow as tf
+    from transformers import TFAutoModel
+if is_torch_available():
+    import torch
+    from transformers import AutoModel
+from transformers import AutoConfig, AutoTokenizer
+input_text = """Bent over their instruments, three hundred Fertilizers were plunged, as 
+the Director of Hatcheries and Conditioning entered the room, in the 
+scarcely breathing silence, the absent-minded, soliloquizing hum or 
+whistle, of absorbed concentration. A troop of newly arrived students, 
+very young, pink and callow, followed nervously, rather abjectly, at the 
+Director's heels. Each of them carried a notebook, in which, whenever 
+the great man spoke, he desperately scribbled. Straight from the 
+horse's mouth. It was a rare privilege. The D. H. C. for Central London 
+always made a point of personally conducting his new students round 
+the various departments. 
+"Just to give you a general idea," he would explain to them. For of 
+course some sort of general idea they must have, if they were to do 
+their work intelligently-though as little of one, if they were to be good 
+and happy members of society, as possible. For particulars, as every 
+one knows, make for virtue and happiness; generalities are intellectu- 
+ally necessary evils. Not philosophers but fret-sawyers and stamp col- 
+lectors compose the backbone of society. 
+"To-morrow," he would add, smiling at them with a slightly menacing 
+geniality, "you'll be settling down to serious work. You won't have time 
+for generalities. Meanwhile ..." 
+Meanwhile, it was a privilege. Straight from the horse's mouth into the 
+notebook. The boys scribbled like mad. 
+Tall and rather thin but upright, the Director advanced into the room. 
+He had a long chin and big rather prominent teeth, just covered, when 
+he was not talking, by his full, floridly curved lips. Old, young? Thirty? 
+Fifty? Fifty-five? It was hard to say. And anyhow the question didn't 
+arise; in this year of stability, A. F. 632, it didn't occur to you to ask it. 
+"I shall begin at the beginning," said the D.H.C. and the more zealous 
+students recorded his intention in their notebooks: Begin at the begin- 
+ning. "These," he waved his hand, "are the incubators." And opening 
+an insulated door he showed them racks upon racks of numbered test- 
+tubes. "The week's supply of ova. Kept," he explained, "at blood heat; 
+whereas the male gametes," and here he opened another door, "they 
+have to be kept at thirty-five instead of thirty-seven. Full blood heat 
+sterilizes." Rams wrapped in theremogene beget no lambs. 
+Still leaning against the incubators he gave them, while the pencils 
+scurried illegibly across the pages, a brief description of the modern 
+fertilizing process; spoke first, of course, of its surgical introduc- 
+tion-"the operation undergone voluntarily for the good of Society, not 
+to mention the fact that it carries a bonus amounting to six months' 
+salary"; continued with some account of the technique for preserving 
+the excised ovary alive and actively developing; passed on to a consid- 
+eration of optimum temperature, salinity, viscosity; referred to the liq- 
+uor in which the detached and ripened eggs were kept; and, leading 
+his charges to the work tables, actually showed them how this liquor 
+was drawn off from the test-tubes; how it was let out drop by drop 
+onto the specially warmed slides of the microscopes; how the eggs 
+which it contained were inspected for abnormalities, counted and 
+transferred to a porous receptacle; how (and he now took them to 
+watch the operation) this receptacle was immersed in a warm bouillon 
+containing free-swimming spermatozoa-at a minimum concentration 
+of one hundred thousand per cubic centimetre, he insisted; and how, 
+after ten minutes, the container was lifted out of the liquor and its 
+contents re-examined; how, if any of the eggs remained unfertilized, it 
+was again immersed, and, if necessary, yet again; how the fertilized 
+ova went back to the incubators; where the Alphas and Betas re- 
+mained until definitely bottled; while the Gammas, Deltas and Epsilons 
+were brought out again, after only thirty-six hours, to undergo Bo- 
+kanovsky's Process. 
+"Bokanovsky's Process," repeated the Director, and the students un- 
+derlined the words in their little notebooks. 
+One egg, one embryo, one adult-normality. But a bokanovskified egg 
+will bud, will proliferate, will divide. From eight to ninety-six buds, and 
+every bud will grow into a perfectly formed embryo, and every embryo 
+into a full-sized adult. Making ninety-six human beings grow where 
+only one grew before. Progress. 
+"Essentially," the D.H.C. concluded, "bokanovskification consists of a 
+series of arrests of development. We check the normal growth and, 
+paradoxically enough, the egg responds by budding." 
+Responds by budding. The pencils were busy. 
+He pointed. On a very slowly moving band a rack-full of test-tubes was 
+entering a large metal box, another, rack-full was emerging. Machinery 
+faintly purred. It took eight minutes for the tubes to go through, he 
+told them. Eight minutes of hard X-rays being about as much as an 
+egg can stand. A few died; of the rest, the least susceptible divided 
+into two; most put out four buds; some eight; all were returned to the 
+incubators, where the buds began to develop; then, after two days, 
+were suddenly chilled, chilled and checked. Two, four, eight, the buds 
+in their turn budded; and having budded were dosed almost to death 
+with alcohol; consequently burgeoned again and having budded-bud 
+out of bud out of bud-were thereafter-further arrest being generally 
+fatal-left to develop in peace. By which time the original egg was in a 
+fair way to becoming anything from eight to ninety-six embryos- a 
+prodigious improvement, you will agree, on nature. Identical twins-but 
+not in piddling twos and threes as in the old viviparous days, when an 
+egg would sometimes accidentally divide; actually by dozens, by 
+scores at a time. 
+"Scores," the Director repeated and flung out his arms, as though he 
+were distributing largesse. "Scores." 
+But one of the students was fool enough to ask where the advantage 
+lay. 
+"My good boy!" The Director wheeled sharply round on him. "Can't you 
+see? Can't you see?" He raised a hand; his expression was solemn. 
+"Bokanovsky's Process is one of the major instruments of social stabil- 
+ity!" 
+Major instruments of social stability. 
+Standard men and women; in uniform batches. The whole of a small 
+factory staffed with the products of a single bokanovskified egg. 
+"Ninety-six identical twins working ninety-six identical machines!" The 
+voice was almost tremulous with enthusiasm. "You really know where 
+you are. For the first time in history." He quoted the planetary motto. 
+"Community, Identity, Stability." Grand words. "If we could bo- 
+kanovskify indefinitely the whole problem would be solved." 
+Solved by standard Gammas, unvarying Deltas, uniform Epsilons. Mil- 
+lions of identical twins. The principle of mass production at last applied 
+to biology. 
+"But, alas," the Director shook his head, "we can't bokanovskify indefi- 
+nitely." 
+Ninety-six seemed to be the limit; seventy-two a good average. From 
+the same ovary and with gametes of the same male to manufacture as 
+many batches of identical twins as possible-that was the best (sadly a 
+second best) that they could do. And even that was difficult. 
+"For in nature it takes thirty years for two hundred eggs to reach ma- 
+turity. But our business is to stabilize the population at this moment, 
+here and now. Dribbling out twins over a quarter of a century-what 
+would be the use of that?" 
+Obviously, no use at all. But Podsnap's Technique had immensely ac- 
+celerated the process of ripening. They could make sure of at least a 
+hundred and fifty mature eggs within two years. Fertilize and bo- 
+kanovskify-in other words, multiply by seventy-two-and you get an 
+average of nearly eleven thousand brothers and sisters in a hundred 
+and fifty batches of identical twins, all within two years of the same 
+age. 
+"And in exceptional cases we can make one ovary yield us over fifteen 
+thousand adult individuals." 
+Beckoning to a fair-haired, ruddy young man who happened to be 
+passing at the moment. "Mr. Foster," he called. The ruddy young man 
+approached. "Can you tell us the record for a single ovary, Mr. Foster?" 
+"Sixteen thousand and twelve in this Centre," Mr. Foster replied with- 
+out hesitation. He spoke very quickly, had a vivacious blue eye, and 
+took an evident pleasure in quoting figures. "Sixteen thousand and 
+twelve; in one hundred and eighty-nine batches of identicals. But of 
+course they've done much better," he rattled on, "in some of the tropi- 
+cal Centres. Singapore has often produced over sixteen thousand five 
+hundred; and Mombasa has actually touched the seventeen thousand 
+mark. But then they have unfair advantages. You should see the way a 
+negro ovary responds to pituitary! It's quite astonishing, when you're 
+used to working with European material. Still," he added, with a laugh 
+(but the light of combat was in his eyes and the lift of his chin was 
+challenging), "still, we mean to beat them if we can. I'm working on a 
+wonderful Delta-Minus ovary at this moment. Only just eighteen 
+months old. Over twelve thousand seven hundred children already, ei- 
+ther decanted or in embryo. And still going strong. We'll beat them 
+yet." 
+"That's the spirit I like!" cried the Director, and clapped Mr. Foster on 
+the shoulder. "Come along with us, and give these boys the benefit of 
+your expert knowledge." 
+Mr. Foster smiled modestly. "With pleasure." They went. 
+In the Bottling Room all was harmonious bustle and ordered activity. 
+Flaps of fresh sow's peritoneum ready cut to the proper size came 
+shooting up in little lifts from the Organ Store in the sub-basement. 
+Whizz and then, click! the lift-hatches hew open; the bottle-liner had 
+only to reach out a hand, take the flap, insert, smooth-down, and be- 
+fore the lined bottle had had time to travel out of reach along the end- 
+less band, whizz, click! another flap of peritoneum had shot up from 
+the depths, ready to be slipped into yet another bottle, the next of that 
+slow interminable procession on the band. 
+Next to the Liners stood the Matriculators. The procession advanced; 
+one by one the eggs were transferred from their test-tubes to the 
+larger containers; deftly the peritoneal lining was slit, the morula 
+dropped into place, the saline solution poured in ... and already the 
+bottle had passed, and it was the turn of the labellers. Heredity, date 
+of fertilization, membership of Bokanovsky Group-details were trans- 
+ferred from test-tube to bottle. No longer anonymous, but named, 
+identified, the procession marched slowly on; on through an opening in 
+the wall, slowly on into the Social Predestination Room. 
+"Eighty-eight cubic metres of card-index," said Mr. Foster with relish, 
+as they entered."""
+def create_setup_and_compute(model_names: List[str],
+                             gpu: bool = True,
+                             tensorflow: bool = False,
+                             average_over: int = 3,
+                             torchscript: bool = False,
+                             xla: bool = False,
+                             amp: bool = False,
+                             fp16: bool = False,
+                             save_to_csv: bool = False,
+                             csv_filename: str = f"results_{round(time())}.csv"):
+    if xla:
+        tf.config.optimizer.set_jit(True)
+    if amp:
+        tf.config.optimizer.set_experimental_options({"auto_mixed_precision": True})
+    if tensorflow:
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_tensorflow(model_names, dictionary, average_over, amp)
+    else:
+        device = 'cuda' if (gpu and torch.cuda.is_available()) else 'cpu'
+        dictionary = {model_name: {} for model_name in model_names}
+        results = _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16)
+    print("=========== RESULTS ===========")
+    for model_name in model_names:
+        print("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
+        for batch_size in results[model_name]["bs"]:
+            print("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
+            for slice_size in results[model_name]["ss"]:
+                result = results[model_name]['results'][batch_size][slice_size]
+                if isinstance(result, str):
+                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
+                          f"{result}")
+                else:
+                    print(f"\t\t{model_name}/{batch_size}/{slice_size}: "
+                          f"{(round(1000 * result) / 1000)}"
+                          f"s")
+    if save_to_csv:
+        with open(csv_filename, mode='w') as csv_file:
+            fieldnames = ['model',
+                          '1x8', '1x64', '1x128', '1x256', '1x512', '1x1024',
+                          '2x8', '2x64', '2x128', '2x256', '2x512', '2x1024',
+                          '4x8', '4x64', '4x128', '4x256', '4x512', '4x1024',
+                          '8x8', '8x64', '8x128', '8x256', '8x512', '8x1024',
+                          ]
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
+            writer.writeheader()
+            for model_name in model_names:
+                model_results = {
+                    f'{bs}x{ss}': results[model_name]['results'][bs][ss]
+                    for bs in results[model_name]["results"]
+                    for ss in results[model_name]['results'][bs]
+                }
+                writer.writerow({'model': model_name, **model_results})
+def _compute_pytorch(model_names, dictionary, average_over, device, torchscript, fp16):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name, torchscript=torchscript)
+        model = AutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+        batch_sizes = [1, 2, 4, 8]
+        slice_sizes = [8, 64, 128, 256, 512, 1024]
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+        for batch_size in batch_sizes:
+            if fp16:
+                model.half()
+            model.to(device)
+            model.eval()
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
+                    try:
+                        if torchscript:
+                            print("Tracing model with sequence size", sequence.shape)
+                            inference = torch.jit.trace(model, sequence)
+                            inference(sequence)
+                        else:
+                            inference = model
+                            inference(sequence)
+                        print("Going through model with sequence of shape", sequence.shape)
+                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
+                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                    except RuntimeError as e:
+                        print("Doesn't fit on GPU.", e)
+                        torch.cuda.empty_cache()
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+    return dictionary
+def _compute_tensorflow(model_names, dictionary, average_over, amp):
+    for c, model_name in enumerate(model_names):
+        print(f"{c + 1} / {len(model_names)}")
+        config = AutoConfig.from_pretrained(model_name)
+        model = TFAutoModel.from_pretrained(model_name, config=config)
+        tokenizer = AutoTokenizer.from_pretrained(model_name)
+        tokenized_sequence = tokenizer.encode(input_text, add_special_tokens=False)
+        max_input_size = tokenizer.max_model_input_sizes[model_name]
+        batch_sizes = [1, 2, 4, 8]
+        slice_sizes = [8, 64, 128, 256, 512, 1024]
+        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "results": {}}
+        dictionary[model_name]["results"] = {i: {} for i in batch_sizes}
+        print("Using model", model)
+        @tf.function
+        def inference(inputs):
+            return model(inputs)
+        for batch_size in batch_sizes:
+            for slice_size in slice_sizes:
+                if max_input_size is not None and slice_size > max_input_size:
+                    dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+                else:
+                    sequence = tf.stack([tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size)
+                    try:
+                        print("Going through model with sequence of shape", sequence.shape)
+                        # To make sure that the model is traced + that the tensors are on the appropriate device
+                        inference(sequence)
+                        runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
+                        average_time = sum(runtimes)/float(len(runtimes)) / 3.0
+                        dictionary[model_name]["results"][batch_size][slice_size] = average_time
+                    except tf.errors.ResourceExhaustedError as e:
+                        print("Doesn't fit on GPU.", e)
+                        torch.cuda.empty_cache()
+                        dictionary[model_name]["results"][batch_size][slice_size] = "N/A"
+    return dictionary
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--models", required=False, type=str, default='all', help="Model checkpoints to be provided "
+                                                                                  "to the AutoModel classes. Leave "
+                                                                                  "blank to benchmark the base version "
+                                                                                  "of all available model "
+                                                                                  "architectures.")
+    parser.add_argument("--torch", required=False, action="store_true", help="Benchmark the Pytorch version of the "
+                                                                             "models")
+    parser.add_argument("--torch_cuda", required=False, action="store_true", help="Pytorch only: run on available "
+                                                                                  "cuda devices")
+    parser.add_argument("--torchscript", required=False, action="store_true", help="Pytorch only: trace the models "
+                                                                                   "using torchscript")
+    parser.add_argument("--tensorflow", required=False, action="store_true", help="Benchmark the TensorFlow version "
+                                                                                  "of the models. Will run on GPU if "
+                                                                                  "the correct dependencies are "
+                                                                                  "installed")
+    parser.add_argument("--xla", required=False, action="store_true", help="TensorFlow only: use XLA acceleration.")
+    parser.add_argument("--amp", required=False, action="store_true", help="TensorFlow only: use automatic mixed precision acceleration.")
+    parser.add_argument("--fp16", required=False, action="store_true", help="PyTorch only: use FP16 to accelerate inference.")
+    parser.add_argument("--keras_predict", required=False, action="store_true", help="Whether to use model.predict "
+                                                                                     "instead of model() to do a "
+                                                                                     "forward pass.")
+    parser.add_argument("--save_to_csv", required=False, action="store_true", help="Save to a CSV file.")
+    parser.add_argument("--csv_filename", required=False, default=None, help="CSV filename used if saving results to csv.")
+    parser.add_argument("--average_over", required=False, default=30, type=int, help="Times an experiment will be run.")
+    args = parser.parse_args()
+    if args.models == 'all':
+        args.models = [
+            "gpt2",
+            "bert-base-cased",
+            "xlnet-base-cased",
+            "xlm-mlm-en-2048",
+            "transfo-xl-wt103",
+            "openai-gpt",
+            "distilbert-base-uncased",
+            "distilgpt2",
+            "roberta-base",
+            "ctrl"
+        ]
+    else:
+        args.models = args.models.split()
+    print("Running with arguments", args)
+    if args.torch:
+        if is_torch_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                tensorflow=False,
+                gpu=args.torch_cuda,
+                torchscript=args.torchscript,
+                fp16=args.fp16,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over
+            )
+        else:
+            raise ImportError("Trying to run a PyTorch benchmark but PyTorch was not found in the environment.")
+    if args.tensorflow:
+        if is_tf_available():
+            create_setup_and_compute(
+                model_names=args.models,
+                tensorflow=True,
+                xla=args.xla,
+                amp=args.amp,
+                save_to_csv=args.save_to_csv,
+                csv_filename=args.csv_filename,
+                average_over=args.average_over
+            )
+        else:
+            raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
+if __name__ == '__main__':
+    main()
--- a/examples/distillation/README.md
+++ b/examples/distillation/README.md
 # Distil*
-This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT and DistilGPT2.
+This folder contains the original code used to train Distil* as well as examples showcasing how to use DistilBERT, DistilRoBERTa and DistilGPT2.
-**2019, October 3rd - 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 23rd, 2019 - Update** We release **DistilRoBERTa**: 95% of `RoBERTa-base`'s performance on GLUE, twice as fast as RoBERTa while being 35% smaller.
+**October 3rd, 2019 - Update** We release our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108) explaining our approach on **DistilBERT**. It includes updated results and further experiments. We applied the same method to GPT2 and release the weights of **DistilGPT2**. DistilGPT2 is two times faster and 33% smaller than GPT2. **The paper superseeds our [previous blogpost](https://medium.com/huggingface/distilbert-8cf3380435b5) with a different distillation loss and better performances. Please use the paper as a reference when comparing/reporting results on DistilBERT.**
+**September 19th, 2019 - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
-**2019, September 19th - Update:** We fixed bugs in the code and released an upadted version of the weights trained with a modification of the distillation loss. DistilBERT now reaches 97% of `BERT-base`'s performance on GLUE, and 86.9 F1 score on SQuAD v1.1 dev set (compared to 88.5 for `BERT-base`). We will publish a formal write-up of our approach in the near future!
 ## What is Distil*
 Distil* is a class of compressed models that started with DistilBERT. DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving 97% of BERT's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
-We have applied the same method to GPT2 and release the weights of the compressed model. On the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for DistilGPT2 (after fine-tuning on the train set).
+We have applied the same method to other Transformer architectures and released the weights:
+- GPT2: on the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 15.0 compared to 18.5 for **DistilGPT2** (after fine-tuning on the train set).
+- RoBERTa: **DistilRoBERTa** reaches 95% of `RoBERTa-base` performance on GLUE while being twice faster and 35% smaller.
+- and more to come! 🤗🤗🤗
 For more information on DistilBERT, please refer to our [NeurIPS workshop paper](https://arxiv.org/abs/1910.01108).
 Here are the results on the dev sets of GLUE:
-| Model      | Macro-score | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI |
+| Model                     | Macro-score                    | CoLA | MNLI | MRPC | QNLI | QQP  | RTE  | SST-2| STS-B| WNLI              |
-| :---:      |    :---:    | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:|
+| :---:                     |    :---:                       | :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:| :---:             |
-| BERT-base  |  **77.6**   | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7 |
+| BERT-base                 |  **77.6**                      | 48.9 | 84.3 | 88.6 | 89.3 | 89.5 | 71.3 | 91.7 | 91.2 | 43.7              |
-| DistilBERT |  **76.8**   | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4 |
+| DistilBERT                |  **76.8**                      | 49.1 | 81.8 | 90.2 | 90.2 | 89.2 | 62.9 | 92.7 | 90.7 | 44.4              |
+| ---                       |    ---                         |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  --- |  ---              |
+| RoBERTa-base (reported)   |  **83.2**/**86.4**<sup>2</sup> | 63.6 | 87.6 | 90.2 | 92.8 | 91.9 | 78.7 | 94.8 | 91.2 | 57.7<sup>3</sup> |
+| DistilRoBERTa<sup>1</sup> |  **79.0**/**82.3**<sup>2</sup> | 59.4 | 83.9 | 86.6 | 90.8 | 89.4 | 67.9 | 92.5 | 88.3 | 52.1              |
+<sup>1</sup> We did not use the MNLI checkpoint for fine-tuning but directy perform transfer learning on the pre-trained DistilRoBERTa.
+<sup>2</sup> Macro-score computed without WNLI.
+<sup>3</sup> We compute this score ourselves for completeness.
 ## Setup
@@ -27,13 +42,15 @@ This part of the library has only be tested with Python3.6+. There are few speci
 **Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).
 ## How to use DistilBERT
 Transformers includes two pre-trained Distil* models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
 - `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
 - `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
- `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 and . The model has 6 layers, 768 dimension and 12 heads, totalizing 82M (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+- `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.
 - and more to come! 🤗🤗🤗
 Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
@@ -47,7 +64,10 @@ outputs = model(input_ids)
 last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 ```
-Similarly, using DistilGPT2 simply consists in calling the GPT2 classes from a different pretrained checkpoint: `model = GPT2Model.from_pretrained('distilgpt2')`.
+Similarly, using the other Distil* models simply consists in calling the base classes with a different pretrained checkpoint:
+- DistilGPT2: `model = GPT2Model.from_pretrained('distilgpt2')`
+- DistilRoBERTa: `model = RobertaModel.from_pretrained('distilroberta-base')`
 ## How to train Distil*
@@ -88,7 +108,7 @@ python train.py \
    --student_config training_configs/distilbert-base-uncased.json \
    --teacher_type bert \
    --teacher_name bert-base-uncased \
-    --alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --mlm \
+    --alpha_ce 5.0 --alpha_mlm 2.0 --alpha_cos 1.0 --alpha_clm 0.0 --mlm \
    --freeze_pos_embs \
    --dump_path serialization_dir/my_first_training \
    --data_file data/binarized_text.bert-base-uncased.pickle \
@@ -124,7 +144,7 @@ python -m torch.distributed.launch \
        --student_config training_configs/distilbert-base-uncased.json \
        --teacher_type bert \
        --teacher_name bert-base-uncased \
-        --alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --mlm \
+        --alpha_ce 0.33 --alpha_mlm 0.33 --alpha_cos 0.33 --alpha_clm 0.0 --mlm \
        --freeze_pos_embs \
        --dump_path serialization_dir/my_first_training \
        --data_file data/binarized_text.bert-base-uncased.pickle \
@@ -146,4 +166,4 @@ If you find the ressource useful, you should cite the following paper:
  booktitle={NeurIPS EMC^2 Workshop},
  year={2019}
 }
 ```
\ No newline at end of file
--- a/examples/distillation/scripts/binarized_data.py
+++ b/examples/distillation/scripts/binarized_data.py
@@ -68,7 +68,7 @@ def main():
    start = time.time()
    for text in data:
        text = f'{bos} {text.strip()} {sep}'
-        token_ids = tokenizer.encode(text)
+        token_ids = tokenizer.encode(text, add_special_tokens=False)
        rslt.append(token_ids)
        iter += 1

--- a/examples/requirements.txt
+++ b/examples/requirements.txt
 tensorboardX
-scikit-learn
+tensorboard
\ No newline at end of file
+scikit-learn
+seqeval
--- a/examples/run_generation.py
+++ b/examples/run_generation.py
@@ -79,13 +79,12 @@ def set_seed(args):
 def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
        Args:
-            logits: logits distribution shape (vocabulary size)
+            logits: logits distribution shape (batch size x vocabulary size)
            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
-    assert logits.dim() == 1  # batch size 1 for now - could be updated for more but the code would be less clear
    top_k = min(top_k, logits.size(-1))  # Safety check
    if top_k > 0:
        # Remove all tokens with a probability less than the last token of the top-k
@@ -102,7 +101,8 @@ def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')
        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
        sorted_indices_to_remove[..., 0] = 0
-        indices_to_remove = sorted_indices[sorted_indices_to_remove]
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
        logits[indices_to_remove] = filter_value
    return logits
@@ -136,18 +136,19 @@ def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=
                inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
-            next_token_logits = outputs[0][0, -1, :] / (temperature if temperature > 0 else 1.)
+            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
-            # reptition penalty from CTRL (https://arxiv.org/abs/1909.05858)
+            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for _ in set(generated):
+            for i in range(num_samples):
-                next_token_logits[_] /= repetition_penalty
+                for _ in set(generated[i].tolist()):
+                    next_token_logits[i, _] /= repetition_penalty
            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: #greedy sampling:
+            if temperature == 0: # greedy sampling:
-                next_token = torch.argmax(filtered_logits).unsqueeze(0)
+                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
            else:
                next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
+            generated = torch.cat((generated, next_token), dim=1)
    return generated
@@ -161,6 +162,7 @@ def main():
    parser.add_argument("--padding_text", type=str, default="")
    parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
    parser.add_argument("--length", type=int, default=20)
+    parser.add_argument("--num_samples", type=int, default=1)
    parser.add_argument("--temperature", type=float, default=1.0,
                        help="temperature of 0 implies greedy sampling")
    parser.add_argument("--repetition_penalty", type=float, default=1.0,
@@ -196,7 +198,7 @@ def main():
    logger.info(args)
    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7 : 
+        if args.temperature > 0.7:
            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
    while True:
@@ -223,10 +225,14 @@ def main():
        if args.model_type in ["transfo-xl", "xlnet"]:
            # Models with memory likes to have a long prompt for short inputs.
            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text)
+        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
+        if args.model_type == "ctrl":
+            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
+                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
        out = sample_sequence(
            model=model,
            context=context_tokens,
+            num_samples=args.num_samples,
            length=args.length,
            temperature=args.temperature,
            top_k=args.top_k,
@@ -238,12 +244,13 @@ def main():
            xlm_lang=xlm_lang,
            device=args.device,
        )
-        out = out[0, len(context_tokens):].tolist()
+        out = out[:, len(context_tokens):].tolist()
+        for o in out:
+            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
+            text = text[: text.find(args.stop_token) if args.stop_token else None]
-        text = tokenizer.decode(out, clean_up_tokenization_spaces=True, skip_special_tokens=True)
+            print(text)
-        text = text[: text.find(args.stop_token) if args.stop_token else None]
-        print(text)
        if args.prompt:
            break
    return text

--- a/examples/run_glue.py
+++ b/examples/run_glue.py
@@ -154,13 +154,16 @@ def train(args, train_dataset, model, tokenizer):
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
-                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
            else:
                loss.backward()
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0 and not args.tpu:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()

--- a/examples/run_lm_finetuning.py
+++ b/examples/run_lm_finetuning.py
@@ -309,10 +309,12 @@ def evaluate(args, model, tokenizer, prefix=""):
    model.eval()
    for batch in tqdm(eval_dataloader, desc="Evaluating"):
-        batch = batch.to(args.device)
+        inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
+        inputs = inputs.to(args.device)
+        labels = labels.to(args.device)
        with torch.no_grad():
-            outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
+            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
            lm_loss = outputs[0]
            eval_loss += lm_loss.mean().item()
        nb_eval_steps += 1

--- a/examples/run_ner.py
+++ b/examples/run_ner.py
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
+from __future__ import absolute_import, division, print_function
+import argparse
+import glob
+import logging
+import os
+import random
+import numpy as np
+import torch
+from seqeval.metrics import precision_score, recall_score, f1_score
+from tensorboardX import SummaryWriter
+from torch.nn import CrossEntropyLoss
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+from utils_ner import convert_examples_to_features, get_labels, read_examples_from_file
+from transformers import AdamW, WarmupLinearSchedule
+from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, BertTokenizer
+from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
+logger = logging.getLogger(__name__)
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig)),
+    ())
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForTokenClassification, BertTokenizer),
+    "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer)
+}
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+def train(args, train_dataset, model, tokenizer, labels, pad_token_label_id):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {"params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
+         "weight_decay": args.weight_decay},
+        {"params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], "weight_decay": 0.0}
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                args.train_batch_size * args.gradient_accumulation_steps * (
+                    torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {"input_ids": batch[0],
+                      "attention_mask": batch[1],
+                      "token_type_ids": batch[2] if args.model_type in ["bert", "xlnet"] else None,
+                      # XLM and RoBERTa don"t use segment_ids
+                      "labels": batch[3]}
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                scheduler.step()  # Update learning rate schedule
+                optimizer.step()
+                model.zero_grad()
+                global_step += 1
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id)
+                        for key, value in results.items():
+                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
+                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
+                    logging_loss = tr_loss
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+    return global_step, tr_loss / global_step
+def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""):
+    eval_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode=mode)
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
+    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+    # Eval!
+    logger.info("***** Running evaluation %s *****", prefix)
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    preds = None
+    out_label_ids = None
+    model.eval()
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        batch = tuple(t.to(args.device) for t in batch)
+        with torch.no_grad():
+            inputs = {"input_ids": batch[0],
+                      "attention_mask": batch[1],
+                      "token_type_ids": batch[2] if args.model_type in ["bert", "xlnet"] else None,
+                      # XLM and RoBERTa don"t use segment_ids
+                      "labels": batch[3]}
+            outputs = model(**inputs)
+            tmp_eval_loss, logits = outputs[:2]
+            if args.n_gpu > 1:
+                tmp_eval_loss = tmp_eval_loss.mean()  # mean() to average on multi-gpu parallel evaluating
+            eval_loss += tmp_eval_loss.item()
+        nb_eval_steps += 1
+        if preds is None:
+            preds = logits.detach().cpu().numpy()
+            out_label_ids = inputs["labels"].detach().cpu().numpy()
+        else:
+            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+            out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
+    eval_loss = eval_loss / nb_eval_steps
+    preds = np.argmax(preds, axis=2)
+    label_map = {i: label for i, label in enumerate(labels)}
+    out_label_list = [[] for _ in range(out_label_ids.shape[0])]
+    preds_list = [[] for _ in range(out_label_ids.shape[0])]
+    for i in range(out_label_ids.shape[0]):
+        for j in range(out_label_ids.shape[1]):
+            if out_label_ids[i, j] != pad_token_label_id:
+                out_label_list[i].append(label_map[out_label_ids[i][j]])
+                preds_list[i].append(label_map[preds[i][j]])
+    results = {
+        "loss": eval_loss,
+        "precision": precision_score(out_label_list, preds_list),
+        "recall": recall_score(out_label_list, preds_list),
+        "f1": f1_score(out_label_list, preds_list)
+    }
+    logger.info("***** Eval results %s *****", prefix)
+    for key in sorted(results.keys()):
+        logger.info("  %s = %s", key, str(results[key]))
+    return results, preds_list
+def load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(args.data_dir, "cached_{}_{}_{}".format(mode,
+        list(filter(None, args.model_name_or_path.split("/"))).pop(),
+        str(args.max_seq_length)))
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", args.data_dir)
+        examples = read_examples_from_file(args.data_dir, mode)
+        features = convert_examples_to_features(examples, labels, args.max_seq_length, tokenizer,
+                                                cls_token_at_end=bool(args.model_type in ["xlnet"]),
+                                                # xlnet has a cls token at the end
+                                                cls_token=tokenizer.cls_token,
+                                                cls_token_segment_id=2 if args.model_type in ["xlnet"] else 0,
+                                                sep_token=tokenizer.sep_token,
+                                                sep_token_extra=bool(args.model_type in ["roberta"]),
+                                                # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+                                                pad_on_left=bool(args.model_type in ["xlnet"]),
+                                                # pad on the left for xlnet
+                                                pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
+                                                pad_token_segment_id=4 if args.model_type in ["xlnet"] else 0,
+                                                pad_token_label_id=pad_token_label_id
+                                                )
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
+    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
+    all_label_ids = torch.tensor([f.label_ids for f in features], dtype=torch.long)
+    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+    return dataset
+def main():
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--data_dir", default=None, type=str, required=True,
+                        help="The input data dir. Should contain the training files for the CoNLL-2003 NER task.")
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+    ## Other parameters
+    parser.add_argument("--labels", default="", type=str,
+                        help="Path to a file containing all labels. If not specified, CoNLL-2003 labels are used.")
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+    parser.add_argument("--max_seq_length", default=128, type=int,
+                        help="The maximum total input sequence length after tokenization. Sequences longer "
+                             "than this will be truncated, sequences shorter will be padded.")
+    parser.add_argument("--do_train", action="store_true",
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true",
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--do_predict", action="store_true",
+                        help="Whether to run predictions on the test set.")
+    parser.add_argument("--evaluate_during_training", action="store_true",
+                        help="Whether to run evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action="store_true",
+                        help="Set this flag if you are using an uncased model.")
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+    parser.add_argument("--logging_steps", type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action="store_true",
+                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+    parser.add_argument("--no_cuda", action="store_true",
+                        help="Avoid using CUDA when available")
+    parser.add_argument("--overwrite_output_dir", action="store_true",
+                        help="Overwrite the content of the output directory")
+    parser.add_argument("--overwrite_cache", action="store_true",
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument("--seed", type=int, default=42,
+                        help="random seed for initialization")
+    parser.add_argument("--fp16", action="store_true",
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument("--fp16_opt_level", type=str, default="O1",
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="For distributed training: local_rank")
+    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
+    args = parser.parse_args()
+    if os.path.exists(args.output_dir) and os.listdir(
+            args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args.output_dir))
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+    # Setup logging
+    logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+                        datefmt="%m/%d/%Y %H:%M:%S",
+                        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                   args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+    # Set seed
+    set_seed(args)
+    # Prepare CONLL-2003 task
+    labels = get_labels(args.labels)
+    num_labels = len(labels)
+    # Use cross entropy ignore index as padding label id so that only real label ids contribute to the loss later
+    pad_token_label_id = CrossEntropyLoss().ignore_index
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          num_labels=num_labels)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+                                                do_lower_case=args.do_lower_case)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path),
+                                        config=config)
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+    model.to(args.device)
+    logger.info("Training/evaluation parameters %s", args)
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, labels, pad_token_label_id, mode="train")
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer, labels, pad_token_label_id)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = model.module if hasattr(model, "module") else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+            result, _ = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="dev", prefix=global_step)
+            if global_step:
+                result = {"{}_{}".format(global_step, k): v for k, v in result.items()}
+            results.update(result)
+        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            for key in sorted(results.keys()):
+                writer.write("{} = {}\n".format(key, str(results[key])))
+    if args.do_predict and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model = model_class.from_pretrained(args.output_dir)
+        model.to(args.device)
+        result, predictions = evaluate(args, model, tokenizer, labels, pad_token_label_id, mode="test")
+        # Save results
+        output_test_results_file = os.path.join(args.output_dir, "test_results.txt")
+        with open(output_test_results_file, "w") as writer:
+            for key in sorted(result.keys()):
+                writer.write("{} = {}\n".format(key, str(result[key])))
+        # Save predictions
+        output_test_predictions_file = os.path.join(args.output_dir, "test_predictions.txt")
+        with open(output_test_predictions_file, "w") as writer:
+            with open(os.path.join(args.data_dir, "test.txt"), "r") as f:
+                example_id = 0
+                for line in f:
+                    if line.startswith("-DOCSTART-") or line == "" or line == "\n":
+                        writer.write(line)
+                        if not predictions[example_id]:
+                            example_id += 1
+                    elif predictions[example_id]:
+                        output_line = line.split()[0] + " " + predictions[example_id].pop(0) + "\n"
+                        writer.write(output_line)
+                    else:
+                        logger.warning("Maximum sequence length exceeded: No prediction for '%s'.", line.split()[0])
+    return results
+if __name__ == "__main__":
+    main()
--- a/examples/run_squad.py
+++ b/examples/run_squad.py
@@ -138,8 +138,8 @@ def train(args, train_dataset, model, tokenizer):
            model.train()
            batch = tuple(t.to(args.device) for t in batch)
            inputs = {'input_ids':       batch[0],
-                      'attention_mask':  batch[1], 
+                      'attention_mask':  batch[1],
-                      'start_positions': batch[3], 
+                      'start_positions': batch[3],
                      'end_positions':   batch[4]}
            if args.model_type != 'distilbert':
                inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
@@ -157,13 +157,16 @@ def train(args, train_dataset, model, tokenizer):
            if args.fp16:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
-                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
            else:
                loss.backward()
-                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
            tr_loss += loss.item()
            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
                optimizer.step()
                scheduler.step()  # Update learning rate schedule
                model.zero_grad()
@@ -485,6 +488,16 @@ def main():
    logger.info("Training/evaluation parameters %s", args)
+    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
+    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
+    # remove the need for this code, but it is still valid.
+    if args.fp16:
+        try:
+            import apex
+            apex.amp.register_half_function(torch, 'einsum')
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
    # Training
    if args.do_train:
        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)

--- a/examples/run_summarization_finetuning.py
+++ b/examples/run_summarization_finetuning.py
+# coding=utf-8
+# Copyright 2019 The HuggingFace Inc. team.
+# Copyright (c) 2019 The HuggingFace Inc.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Finetuning seq2seq models for sequence generation."""
+import argparse
+import functools
+import logging
+import os
+import random
+import sys
+import numpy as np
+from tqdm import tqdm, trange
+import torch
+from torch.optim import Adam
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
+from transformers import (
+    AutoTokenizer,
+    BertForMaskedLM,
+    BertConfig,
+    PreTrainedEncoderDecoder,
+    Model2Model,
+)
+from utils_summarization import (
+    CNNDailyMailDataset,
+    encode_for_summarization,
+    fit_to_block_size,
+    build_lm_labels,
+    build_mask,
+    compute_token_type_ids,
+)
+logger = logging.getLogger(__name__)
+logging.basicConfig(stream=sys.stdout, level=logging.INFO)
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+# ------------
+# Load dataset
+# ------------
+def load_and_cache_examples(args, tokenizer):
+    dataset = CNNDailyMailDataset(tokenizer, data_dir=args.data_dir)
+    return dataset
+def collate(data, tokenizer, block_size):
+    """ List of tuple as an input. """
+    # remove the files with empty an story/summary, encode and fit to block
+    data = filter(lambda x: not (len(x[0]) == 0 or len(x[1]) == 0), data)
+    data = [
+        encode_for_summarization(story, summary, tokenizer) for story, summary in data
+    ]
+    data = [
+        (
+            fit_to_block_size(story, block_size, tokenizer.pad_token_id),
+            fit_to_block_size(summary, block_size, tokenizer.pad_token_id),
+        )
+        for story, summary in data
+    ]
+    stories = torch.tensor([story for story, summary in data])
+    summaries = torch.tensor([summary for story, summary in data])
+    encoder_token_type_ids = compute_token_type_ids(stories, tokenizer.cls_token_id)
+    encoder_mask = build_mask(stories, tokenizer.pad_token_id)
+    decoder_mask = build_mask(summaries, tokenizer.pad_token_id)
+    lm_labels = build_lm_labels(summaries, tokenizer.pad_token_id)
+    return (
+        stories,
+        summaries,
+        encoder_token_type_ids,
+        encoder_mask,
+        decoder_mask,
+        lm_labels,
+    )
+# ----------
+# Optimizers
+# ----------
+class BertSumOptimizer(object):
+    """ Specific optimizer for BertSum.
+    As described in [1], the authors fine-tune BertSum for abstractive
+    summarization using two Adam Optimizers with different warm-up steps and
+    learning rate. They also use a custom learning rate scheduler.
+    [1] Liu, Yang, and Mirella Lapata. "Text summarization with pretrained encoders."
+        arXiv preprint arXiv:1908.08345 (2019).
+    """
+    def __init__(self, model, lr, warmup_steps, beta_1=0.99, beta_2=0.999, eps=1e-8):
+        self.encoder = model.encoder
+        self.decoder = model.decoder
+        self.lr = lr
+        self.warmup_steps = warmup_steps
+        self.optimizers = {
+            "encoder": Adam(
+                model.encoder.parameters(),
+                lr=lr["encoder"],
+                betas=(beta_1, beta_2),
+                eps=eps,
+            ),
+            "decoder": Adam(
+                model.decoder.parameters(),
+                lr=lr["decoder"],
+                betas=(beta_1, beta_2),
+                eps=eps,
+            ),
+        }
+        self._step = 0
+    def _update_rate(self, stack):
+        return self.lr[stack] * min(
+            self._step ** (-0.5), self._step * self.warmup_steps[stack] ** (-0.5)
+        )
+    def zero_grad(self):
+        self.optimizer_decoder.zero_grad()
+        self.optimizer_encoder.zero_grad()
+    def step(self):
+        self._step += 1
+        for stack, optimizer in self.optimizers.items():
+            new_rate = self._update_rate(stack)
+            for param_group in optimizer.param_groups:
+                param_group["lr"] = new_rate
+            optimizer.step()
+# ------------
+# Train
+# ------------
+def train(args, model, tokenizer):
+    """ Fine-tune the pretrained model on the corpus. """
+    set_seed(args)
+    # Load the data
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_dataset = load_and_cache_examples(args, tokenizer)
+    train_sampler = RandomSampler(train_dataset)
+    model_collate_fn = functools.partial(collate, tokenizer=tokenizer, block_size=512)
+    train_dataloader = DataLoader(
+        train_dataset,
+        sampler=train_sampler,
+        batch_size=args.train_batch_size,
+        collate_fn=model_collate_fn,
+    )
+    # Training schedule
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = t_total // (
+            len(train_dataloader) // args.gradient_accumulation_steps + 1
+        )
+    else:
+        t_total = (
+            len(train_dataloader)
+            // args.gradient_accumulation_steps
+            * args.num_train_epochs
+        )
+    # Prepare the optimizer
+    lr = {"encoder": 0.002, "decoder": 0.2}
+    warmup_steps = {"encoder": 20000, "decoder": 10000}
+    optimizer = BertSumOptimizer(model, lr, warmup_steps)
+    # Train
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info(
+        "  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size
+    )
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size * args.gradient_accumulation_steps
+        # * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+    model.zero_grad()
+    train_iterator = trange(args.num_train_epochs, desc="Epoch", disable=True)
+    global_step = 0
+    tr_loss = 0.0
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=True)
+        for step, batch in enumerate(epoch_iterator):
+            source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
+            source = source.to(args.device)
+            target = target.to(args.device)
+            encoder_token_type_ids = encoder_token_type_ids.to(args.device)
+            encoder_mask = encoder_mask.to(args.device)
+            decoder_mask = decoder_mask.to(args.device)
+            lm_labels = lm_labels.to(args.device)
+            model.train()
+            outputs = model(
+                source,
+                target,
+                encoder_token_type_ids=encoder_token_type_ids,
+                encoder_attention_mask=encoder_mask,
+                decoder_attention_mask=decoder_mask,
+                decoder_lm_labels=lm_labels,
+            )
+            loss = outputs[0]
+            print(loss)
+            if args.gradient_accumulation_steps > 1:
+                loss /= args.gradient_accumulation_steps
+            loss.backward()
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                optimizer.step()
+                model.zero_grad()
+                global_step += 1
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+    return global_step, tr_loss / global_step
+# ------------
+# Train
+# ------------
+def evaluate(args, model, tokenizer, prefix=""):
+    set_seed(args)
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
+    eval_sampler = SequentialSampler(eval_dataset)
+    eval_dataloader = DataLoader(
+        eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size
+    )
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    model.eval()
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        source, target, encoder_token_type_ids, encoder_mask, decoder_mask, lm_labels = batch
+        source = source.to(args.device)
+        target = target.to(args.device)
+        encoder_token_type_ids = encoder_token_type_ids.to(args.device)
+        encoder_mask = encoder_mask.to(args.device)
+        decoder_mask = decoder_mask.to(args.device)
+        lm_labels = lm_labels.to(args.device)
+        with torch.no_grad():
+            outputs = model(
+                source,
+                target,
+                encoder_token_type_ids=encoder_token_type_ids,
+                encoder_attention_mask=encoder_mask,
+                decoder_attention_mask=decoder_mask,
+                decoder_lm_labels=lm_labels,
+            )
+            lm_loss = outputs[0]
+            eval_loss += lm_loss.mean().item()
+        nb_eval_steps += 1
+    eval_loss = eval_loss / nb_eval_steps
+    perplexity = torch.exp(torch.tensor(eval_loss))
+    result = {"perplexity": perplexity}
+    # Save the evaluation's results
+    output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+    if not os.path.exists(args.output_dir):
+        os.makedirs(args.output_dir)
+    with open(output_eval_file, "w") as writer:
+        logger.info("***** Eval results {} *****".format(prefix))
+        for key in sorted(result.keys()):
+            logger.info("  %s = %s", key, str(result[key]))
+            writer.write("%s = %s\n" % (key, str(result[key])))
+    return result
+def main():
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--data_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The input training data file (a text file).",
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    # Optional parameters
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--do_evaluate",
+        type=bool,
+        default=False,
+        help="Run model evaluation on out-of-sample data.",
+    )
+    parser.add_argument("--do_train", type=bool, default=False, help="Run training.")
+    parser.add_argument(
+        "--do_overwrite_output_dir",
+        type=bool,
+        default=False,
+        help="Whether to overwrite the output dir.",
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default="bert-base-cased",
+        type=str,
+        help="The model checkpoint to initialize the encoder and decoder's weights with.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default="bert",
+        type=str,
+        help="The decoder architecture to be fine-tuned.",
+    )
+    parser.add_argument(
+        "--max_grad_norm", default=1.0, type=float, help="Max gradient norm."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument(
+        "--to_cpu", default=False, type=bool, help="Whether to force training on CPU."
+    )
+    parser.add_argument(
+        "--num_train_epochs",
+        default=10,
+        type=int,
+        help="Total number of training epochs to perform.",
+    )
+    parser.add_argument(
+        "--per_gpu_train_batch_size",
+        default=4,
+        type=int,
+        help="Batch size per GPU/CPU for training.",
+    )
+    parser.add_argument("--seed", default=42, type=int)
+    args = parser.parse_args()
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.do_overwrite_output_dir
+    ):
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --do_overwrite_output_dir to overwrite.".format(
+                args.output_dir
+            )
+        )
+    # Set up training device
+    if args.to_cpu or not torch.cuda.is_available():
+        args.device = torch.device("cpu")
+        args.n_gpu = 0
+    else:
+        args.device = torch.device("cuda")
+        args.n_gpu = torch.cuda.device_count()
+    # Load pretrained model and tokenizer. The decoder's weights are randomly initialized.
+    tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
+    config = BertConfig.from_pretrained(args.model_name_or_path)
+    decoder_model = BertForMaskedLM(config)
+    model = Model2Model.from_pretrained(
+        args.model_name_or_path, decoder_model=decoder_model
+    )
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        0,
+        args.device,
+        args.n_gpu,
+        False,
+        False,
+    )
+    logger.info("Training/evaluation parameters %s", args)
+    # Train the model
+    model.to(args.device)
+    if args.do_train:
+        global_step, tr_loss = train(args, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+        if not os.path.exists(args.output_dir):
+            os.makedirs(args.output_dir)
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+        torch.save(args, os.path.join(args.output_dir, "training_arguments.bin"))
+    # Evaluate the model
+    results = {}
+    if args.do_evaluate:
+        checkpoints = []
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            encoder_checkpoint = os.path.join(checkpoint, "encoder")
+            decoder_checkpoint = os.path.join(checkpoint, "decoder")
+            model = PreTrainedEncoderDecoder.from_pretrained(
+                encoder_checkpoint, decoder_checkpoint
+            )
+            model.to(args.device)
+            results = "placeholder"
+    return results
+if __name__ == "__main__":
+    main()