Ensure from_pretrained takes a str in unittest.

Signed-off-by: Morgan Funtowicz <morgan@huggingface.co>

.github/workflows/self-scheduled.yml

@@ -31,13 +31,12 @@ jobs:
     - name: Install dependencies
       run: |
         source .env/bin/activate
-        pip install .[sklearn,tf,torch,testing]
+        pip install .[sklearn,torch,testing]
 
     - name: Are GPUs recognized by our DL frameworks
       run: |
         source .env/bin/activate
         python -c "import torch; print(torch.cuda.is_available())"
-        python -c "import tensorflow as tf; print(tf.test.is_built_with_cuda(), tf.config.list_physical_devices('GPU'))"
 
     - name: Run all tests on GPU
       env:

CONTRIBUTING.md

@@ -44,9 +44,16 @@ Did not find it? :( So we can act quickly on it, please follow these steps:
 To get the OS and software versions automatically, you can run the following command:
 
 ```bash
-python transformers-cli env
+transformers-cli env
 ```
 
+or from the root of the repository the following command:
+
+```bash
+python src/transformers/commands/transformers_cli.py env
+```
+
+
 ### Do you want to implement a new model?
 
 Awesome! Please provide the following information:
@@ -198,11 +205,12 @@ Follow these steps to start contributing:
    are useful to avoid duplicated work, and to differentiate it from PRs ready
    to be merged;
 4. Make sure existing tests pass;
-5. Add high-coverage tests. No quality test, no merge. 
+5. Add high-coverage tests. No quality testing = no merge. 
  - If you are adding a new model, make sure that you use `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)`, which triggers the common tests.
  - If you are adding new `@slow` tests, make sure they pass using `RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`. 
+ - If you are adding a new tokenizer, write tests, and make sure `RUN_SLOW=1 python -m pytest tests/test_tokenization_{your_model_name}.py` passes.
 CircleCI does not run them. 
-6. All public methods must have informative docstrings;
+6. All public methods must have informative docstrings that work nicely with sphinx. See `modeling_ctrl.py` for an example.
 
 ### Tests
 

README.md

@@ -165,8 +165,9 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 18. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 19. **[Reformer](https://huggingface.co/transformers/model_doc/reformer.html)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
 20. **[MarianMT](https://huggingface.co/transformers/model_doc/marian.html)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
-21. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
-22. 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.
+21. **[Longformer](https://huggingface.co/transformers/model_doc/longformer.html)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+22. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
+23. 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).
 
@@ -339,8 +340,8 @@ python ./examples/text-classification/run_glue.py \
     --do_eval \
     --data_dir $GLUE_DIR/$TASK_NAME \
     --max_seq_length 128 \
-    --per_gpu_eval_batch_size=8   \
-    --per_gpu_train_batch_size=8   \
+    --per_device_eval_batch_size=8   \
+    --per_device_train_batch_size=8   \
     --learning_rate 2e-5 \
     --num_train_epochs 3.0 \
     --output_dir /tmp/$TASK_NAME/
@@ -366,8 +367,8 @@ python ./examples/text-classification/run_glue.py \
     --data_dir=${GLUE_DIR}/STS-B  \
     --output_dir=./proc_data/sts-b-110   \
     --max_seq_length=128   \
-    --per_gpu_eval_batch_size=8   \
-    --per_gpu_train_batch_size=8   \
+    --per_device_eval_batch_size=8   \
+    --per_device_train_batch_size=8   \
     --gradient_accumulation_steps=1 \
     --max_steps=1200  \
     --model_name=xlnet-large-cased   \
@@ -390,8 +391,8 @@ python -m torch.distributed.launch --nproc_per_node 8 ./examples/text-classifica
     --do_eval   \
     --data_dir $GLUE_DIR/MRPC/   \
     --max_seq_length 128   \
-    --per_gpu_eval_batch_size=8   \
-    --per_gpu_train_batch_size=8   \
+    --per_device_eval_batch_size=8   \
+    --per_device_train_batch_size=8   \
     --learning_rate 2e-5   \
     --num_train_epochs 3.0  \
     --output_dir /tmp/mrpc_output/ \
@@ -427,8 +428,8 @@ python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answer
     --max_seq_length 384 \
     --doc_stride 128 \
     --output_dir ../models/wwm_uncased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=3   \
-    --per_gpu_train_batch_size=3   \
+    --per_device_eval_batch_size=3   \
+    --per_device_train_batch_size=3   \
 ```
 
 Training with these hyper-parameters gave us the following results:

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.9.1'
+release = u'2.10.0'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/examples.md

@@ -1,649 +0,0 @@
-# Examples
-
-In this section a few examples are put together. All of these examples work for several models, making use of the very
-similar API between the different models.
-
-**Important**
-To run the latest versions of the examples, you have to install from source and install some specific requirements for the examples.
-Execute the following steps in a new virtual environment:
-
-```bash
-git clone https://github.com/huggingface/transformers
-cd transformers
-pip install .
-pip install -r ./examples/requirements.txt
-```
-
-| Section                    | Description                                                                                                                                                |
-|----------------------------|------------------------------------------------------------------------------------------------------------------------------------------
-| [TensorFlow 2.0 models on GLUE](#TensorFlow-2.0-Bert-models-on-GLUE) | Examples running BERT TensorFlow 2.0 model on the GLUE tasks. |
-| [Running on TPUs](#running-on-tpus) | Examples on running fine-tuning tasks on Google TPUs to accelerate workloads. |
-| [Language Model training](#language-model-training) | Fine-tuning (or training from scratch) 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](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | Using BERT for Named Entity Recognition (NER) on the CoNLL 2003 dataset, examples with distributed training. |
-| [XNLI](#xnli) | Examples running BERT/XLM on the XNLI benchmark. |
-| [Adversarial evaluation of model performances](#adversarial-evaluation-of-model-performances) | Testing a model with adversarial evaluation of natural language inference on the Heuristic Analysis for NLI Systems (HANS) dataset (McCoy et al., 2019.) |
-
-## TensorFlow 2.0 Bert models on GLUE
-
-Based on the script [`run_tf_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/text-classification/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).
-
-## Running on TPUs
-
-You can accelerate your workloads on Google's TPUs. For information on how to setup your TPU environment refer to this
-[README](https://github.com/pytorch/xla/blob/master/README.md).
-
-The following are some examples of running the `*_tpu.py` finetuning scripts on TPUs. All steps for data preparation are
-identical to your normal GPU + Huggingface setup.
-
-### GLUE
-
-Before running anyone of these GLUE tasks you should download the
-[GLUE data](https://gluebenchmark.com/tasks) by running
-[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
-and unpack it to some directory `$GLUE_DIR`.
-
-For running your GLUE task on MNLI dataset you can run something like the following:
-
-```
-export XRT_TPU_CONFIG="tpu_worker;0;$TPU_IP_ADDRESS:8470"
-export GLUE_DIR=/path/to/glue
-export TASK_NAME=MNLI
-
-python run_glue_tpu.py \
-  --model_type bert \
-  --model_name_or_path bert-base-cased \
-  --task_name $TASK_NAME \
-  --do_train \
-  --do_eval \
-  --data_dir $GLUE_DIR/$TASK_NAME \
-  --max_seq_length 128 \
-  --train_batch_size 32 \
-  --learning_rate 3e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/$TASK_NAME \
-  --overwrite_output_dir \
-  --logging_steps 50 \
-  --save_steps 200 \
-  --num_cores=8 \
-  --only_log_master
-```
-
-
-## Language model training
-
-Based on the script [`run_language_modeling.py`](https://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_language_modeling.py).
-
-Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT and RoBERTa (DistilBERT
-to be added soon). GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa
-are fine-tuned using a masked language modeling (MLM) loss.
-
-Before running the following example, you should get a file that contains text on which the language model will be
-trained or fine-tuned. A good example of such text is the [WikiText-2 dataset](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/).
-
-We will refer to two different files: `$TRAIN_FILE`, which contains text for training, and `$TEST_FILE`, which contains
-text that will be used for evaluation.
-
-### GPT-2/GPT and causal language modeling
-
-The following example fine-tunes GPT-2 on WikiText-2. We're using the raw WikiText-2 (no tokens were replaced before
-the tokenization). The loss here is that of causal language modeling.
-
-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=gpt2 \
-    --model_name_or_path=gpt2 \
-    --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --do_eval \
-    --eval_data_file=$TEST_FILE
-```
-
-This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run. It reaches
-a score of ~20 perplexity once fine-tuned on the dataset.
-
-### RoBERTa/BERT and masked language modeling
-
-The following example fine-tunes RoBERTa on WikiText-2. Here too, we're using the raw WikiText-2. The loss is different
-as BERT/RoBERTa have a bidirectional mechanism; we're therefore using the same loss that was used during their
-pre-training: masked language modeling.
-
-In accordance to the RoBERTa paper, we use dynamic masking rather than static masking. The model may, therefore, converge
-slightly slower (over-fitting takes more epochs).
-
-We use the `--mlm` flag so that the script may change its loss function.
-
-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=roberta \
-    --model_name_or_path=roberta-base \
-    --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --do_eval \
-    --eval_data_file=$TEST_FILE \
-    --mlm
-```
-
-## Language generation
-
-Based on the script [`run_generation.py`](https://github.com/huggingface/transformers/blob/master/examples/text-generation/run_generation.py).
-
-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.
-
-Example usage:
-
-```bash
-python run_generation.py \
-    --model_type=gpt2 \
-    --model_name_or_path=gpt2
-```
-
-## GLUE
-
-Based on the script [`run_glue.py`](https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_glue.py).
-
-Fine-tuning the library models for sequence classification on the GLUE benchmark: [General Language Understanding
-Evaluation](https://gluebenchmark.com/). This script can fine-tune the following models: BERT, XLM, XLNet and RoBERTa.
-
-GLUE is made up of a total of 9 different tasks. We get the following results on the dev set of the benchmark with an
-uncased  BERT base model (the checkpoint `bert-base-uncased`). All experiments ran single V100 GPUs with a total train
-batch sizes between 16 and 64. Some of these tasks have a small dataset and training can lead to high variance in the results
-between different runs. We report the median on 5 runs (with different seeds) for each of the metrics.
-
-| Task  | Metric                       | Result      |
-|-------|------------------------------|-------------|
-| CoLA  | Matthew's corr               | 49.23       |
-| SST-2 | Accuracy                     | 91.97       |
-| MRPC  | F1/Accuracy                  | 89.47/85.29 |
-| STS-B | Person/Spearman corr.        | 83.95/83.70 |
-| QQP   | Accuracy/F1                  | 88.40/84.31 |
-| MNLI  | Matched acc./Mismatched acc. | 80.61/81.08 |
-| QNLI  | Accuracy                     | 87.46       |
-| RTE   | Accuracy                     | 61.73       |
-| WNLI  | Accuracy                     | 45.07       |
-
-Some of these results are significantly different from the ones reported on the test set
-of GLUE benchmark on the website. For QQP and WNLI, please refer to [FAQ #12](https://gluebenchmark.com/faq) on the webite.
-
-Before running any one of these GLUE tasks you should download the
-[GLUE data](https://gluebenchmark.com/tasks) by running
-[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
-and unpack it to some directory `$GLUE_DIR`.
-
-```bash
-export GLUE_DIR=/path/to/glue
-export TASK_NAME=MRPC
-
-python run_glue.py \
-  --model_type bert \
-  --model_name_or_path bert-base-cased \
-  --task_name $TASK_NAME \
-  --do_train \
-  --do_eval \
-  --data_dir $GLUE_DIR/$TASK_NAME \
-  --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/$TASK_NAME/
-```
-
-where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.
-
-The dev set results will be present within the text file `eval_results.txt` in the specified output_dir.
-In case of MNLI, since there are two separate dev sets (matched and mismatched), there will be a separate
-output folder called `/tmp/MNLI-MM/` in addition to `/tmp/MNLI/`.
-
-The code has not been tested with half-precision training with apex on any GLUE task apart from MRPC, MNLI,
-CoLA, SST-2. The following section provides details on how to run half-precision training with MRPC. With that being
-said, there shouldn’t be any issues in running half-precision training with the remaining GLUE tasks as well,
-since the data processor for each task inherits from the base class DataProcessor.
-
-### MRPC
-
-#### Fine-tuning example
-
-The following examples fine-tune BERT on the Microsoft Research Paraphrase Corpus (MRPC) corpus and runs in less
-than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.
-
-Before running any one of these GLUE tasks you should download the
-[GLUE data](https://gluebenchmark.com/tasks) by running
-[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
-and unpack it to some directory `$GLUE_DIR`.
-
-```bash
-export GLUE_DIR=/path/to/glue
-
-python run_glue.py \
-  --model_name_or_path bert-base-cased \
-  --task_name MRPC \
-  --do_train \
-  --do_eval \
-  --data_dir $GLUE_DIR/MRPC/ \
-  --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/mrpc_output/
-```
-
-Our test ran on a few seeds with [the original implementation hyper-
-parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation
-results between 84% and 88%.
-
-#### Using Apex and mixed-precision
-
-Using Apex and 16 bit precision, the fine-tuning on MRPC only takes 27 seconds. First install
-[apex](https://github.com/NVIDIA/apex), then run the following example:
-
-```bash
-export GLUE_DIR=/path/to/glue
-
-python run_glue.py \
-  --model_name_or_path bert-base-cased \
-  --task_name MRPC \
-  --do_train \
-  --do_eval \
-  --data_dir $GLUE_DIR/MRPC/ \
-  --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/mrpc_output/ \
-  --fp16
-```
-
-#### Distributed training
-
-Here is an example using distributed training on 8 V100 GPUs. The model used is the BERT whole-word-masking and it
-reaches F1 > 92 on MRPC.
-
-```bash
-export GLUE_DIR=/path/to/glue
-
-python -m torch.distributed.launch \
-    --nproc_per_node 8 run_glue.py \
-    --model_name_or_path bert-base-cased \
-    --task_name MRPC \
-    --do_train \
-    --do_eval \
-    --data_dir $GLUE_DIR/MRPC/ \
-    --max_seq_length 128 \
-    --per_gpu_train_batch_size 8 \
-    --learning_rate 2e-5 \
-    --num_train_epochs 3.0 \
-    --output_dir /tmp/mrpc_output/
-```
-
-Training with these hyper-parameters gave us the following results:
-
-```bash
-acc = 0.8823529411764706
-acc_and_f1 = 0.901702786377709
-eval_loss = 0.3418912578906332
-f1 = 0.9210526315789473
-global_step = 174
-loss = 0.07231863956341798
-```
-
-### MNLI
-
-The following example uses the BERT-large, uncased, whole-word-masking model and fine-tunes it on the MNLI task.
-
-```bash
-export GLUE_DIR=/path/to/glue
-
-python -m torch.distributed.launch \
-    --nproc_per_node 8 run_glue.py \
-    --model_name_or_path bert-base-cased \
-    --task_name mnli \
-    --do_train \
-    --do_eval \
-    --data_dir $GLUE_DIR/MNLI/ \
-    --max_seq_length 128 \
-    --per_gpu_train_batch_size 8 \
-    --learning_rate 2e-5 \
-    --num_train_epochs 3.0 \
-    --output_dir output_dir \
-```
-
-The results  are the following:
-
-```bash
-***** Eval results *****
-  acc = 0.8679706601466992
-  eval_loss = 0.4911287787382479
-  global_step = 18408
-  loss = 0.04755385363816904
-
-***** Eval results *****
-  acc = 0.8747965825874695
-  eval_loss = 0.45516540421714036
-  global_step = 18408
-  loss = 0.04755385363816904
-```
-
-## Multiple Choice
-
-Based on the script [`run_multiple_choice.py`]().
-
-#### Fine-tuning on SWAG
-Download [swag](https://github.com/rowanz/swagaf/tree/master/data) data
-
-```bash
-#training on 4 tesla V100(16GB) GPUS
-export SWAG_DIR=/path/to/swag_data_dir
-python ./examples/multiple-choice/run_multiple_choice.py \
---task_name swag \
---model_name_or_path roberta-base \
---do_train \
---do_eval \
---data_dir $SWAG_DIR \
---learning_rate 5e-5 \
---num_train_epochs 3 \
---max_seq_length 80 \
---output_dir models_bert/swag_base \
---per_gpu_eval_batch_size=16 \
---per_gpu_train_batch_size=16 \
---gradient_accumulation_steps 2 \
---overwrite_output
-```
-Training with the defined hyper-parameters yields the following results:
-```
-***** Eval results *****
-eval_acc = 0.8338998300509847
-eval_loss = 0.44457291918821606
-```
-
-## SQuAD
-
-Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/question-answering/run_squad.py).
-
-#### Fine-tuning BERT on SQuAD1.0
-
-This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large)
-on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a
-$SQUAD_DIR directory.
-
-* [train-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json)
-* [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
-* [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
-
-And for SQuAD2.0, you need to download:
-
-- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
-- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
-- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
-
-```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-  --model_type bert \
-  --model_name_or_path bert-base-uncased \
-  --do_train \
-  --do_eval \
-  --train_file $SQUAD_DIR/train-v1.1.json \
-  --predict_file $SQUAD_DIR/dev-v1.1.json \
-  --per_gpu_train_batch_size 12 \
-  --learning_rate 3e-5 \
-  --num_train_epochs 2.0 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir /tmp/debug_squad/
-```
-
-Training with the previously defined hyper-parameters yields the following results:
-
-```bash
-f1 = 88.52
-exact_match = 81.22
-```
-
-#### Distributed training
-
-
-Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.1:
-
-```bash
-python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answering/run_squad.py \
-    --model_type bert \
-    --model_name_or_path bert-large-uncased-whole-word-masking \
-    --do_train \
-    --do_eval \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2 \
-    --max_seq_length 384 \
-    --doc_stride 128 \
-    --output_dir ./examples/models/wwm_uncased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=3   \
-    --per_gpu_train_batch_size=3   \
-```
-
-Training with the previously defined hyper-parameters yields the following results:
-
-```bash
-f1 = 93.15
-exact_match = 86.91
-```
-
-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
-
-This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
-
-##### Command for SQuAD1.0:
-
-```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-    --model_type xlnet \
-    --model_name_or_path xlnet-large-cased \
-    --do_train \
-    --do_eval \
-    --train_file $SQUAD_DIR/train-v1.1.json \
-    --predict_file $SQUAD_DIR/dev-v1.1.json \
-    --learning_rate 3e-5 \
-    --num_train_epochs 2 \
-    --max_seq_length 384 \
-    --doc_stride 128 \
-    --output_dir ./wwm_cased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=4  \
-    --per_gpu_train_batch_size=4   \
-    --save_steps 5000
-```
-
-##### Command for SQuAD2.0:
-
-```bash
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-    --model_type xlnet \
-    --model_name_or_path xlnet-large-cased \
-    --do_train \
-    --do_eval \
-    --version_2_with_negative \
-    --train_file $SQUAD_DIR/train-v2.0.json \
-    --predict_file $SQUAD_DIR/dev-v2.0.json \
-    --learning_rate 3e-5 \
-    --num_train_epochs 4 \
-    --max_seq_length 384 \
-    --doc_stride 128 \
-    --output_dir ./wwm_cased_finetuned_squad/ \
-    --per_gpu_eval_batch_size=2  \
-    --per_gpu_train_batch_size=2   \
-    --save_steps 5000
-```
-
-Larger batch size may improve the performance while costing more memory.
-
-##### Results for SQuAD1.0 with the previously defined hyper-parameters:
-
-```python
-{
-"exact": 85.45884578997162,
-"f1": 92.5974600601065,
-"total": 10570,
-"HasAns_exact": 85.45884578997162,
-"HasAns_f1": 92.59746006010651,
-"HasAns_total": 10570
-}
-```
-
-##### Results for SQuAD2.0 with the previously defined hyper-parameters:
-
-```python
-{
-"exact": 80.4177545691906,
-"f1": 84.07154997729623,
-"total": 11873,
-"HasAns_exact": 76.73751686909581,
-"HasAns_f1": 84.05558584352873,
-"HasAns_total": 5928,
-"NoAns_exact": 84.0874684608915,
-"NoAns_f1": 84.0874684608915,
-"NoAns_total": 5945
-}
-```
-
-
-
-
-## XNLI
-
-Based on the script [`run_xnli.py`](https://github.com/huggingface/transformers/blob/master/examples/text-classification/run_xnli.py).
-
-[XNLI](https://www.nyu.edu/projects/bowman/xnli/) is crowd-sourced dataset based on [MultiNLI](http://www.nyu.edu/projects/bowman/multinli/). It is an evaluation benchmark for cross-lingual text representations. Pairs of text are labeled with textual entailment annotations for 15 different languages (including both high-resource language such as English and low-resource languages such as Swahili).
-
-#### Fine-tuning on XNLI
-
-This example code fine-tunes mBERT (multi-lingual BERT) on the XNLI dataset. It runs in 106 mins
-on a single tesla V100 16GB. The data for XNLI can be downloaded with the following links and should be both saved (and un-zipped) in a
-`$XNLI_DIR` directory.
-
-* [XNLI 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-1.0.zip)
-* [XNLI-MT 1.0](https://www.nyu.edu/projects/bowman/xnli/XNLI-MT-1.0.zip)
-
-```bash
-export XNLI_DIR=/path/to/XNLI
-
-python run_xnli.py \
-  --model_type bert \
-  --model_name_or_path bert-base-multilingual-cased \
-  --language de \
-  --train_language en \
-  --do_train \
-  --do_eval \
-  --data_dir $XNLI_DIR \
-  --per_gpu_train_batch_size 32 \
-  --learning_rate 5e-5 \
-  --num_train_epochs 2.0 \
-  --max_seq_length 128 \
-  --output_dir /tmp/debug_xnli/ \
-  --save_steps -1
-```
-
-Training with the previously defined hyper-parameters yields the following results on the **test** set:
-
-```bash
-acc = 0.7093812375249501
-```
-
-## MM-IMDb
-
-Based on the script [`run_mmimdb.py`](https://github.com/huggingface/transformers/blob/master/examples/contrib/mm-imdb/run_mmimdb.py).
-
-[MM-IMDb](http://lisi1.unal.edu.co/mmimdb/) is a Multimodal dataset with around 26,000 movies including images, plots and other metadata.
-
-### Training on MM-IMDb
-
-```
-python run_mmimdb.py \
-    --data_dir /path/to/mmimdb/dataset/ \
-    --model_type bert \
-    --model_name_or_path bert-base-uncased \
-    --output_dir /path/to/save/dir/ \
-    --do_train \
-    --do_eval \
-    --max_seq_len 512 \
-    --gradient_accumulation_steps 20 \
-    --num_image_embeds 3 \
-    --num_train_epochs 100 \
-    --patience 5
-```
-
-## Adversarial evaluation of model performances
-
-Here is an example on evaluating a model using adversarial evaluation of natural language inference with the Heuristic Analysis for NLI Systems (HANS) dataset [McCoy et al., 2019](https://arxiv.org/abs/1902.01007). The example was gracefully provided by [Nafise Sadat Moosavi](https://github.com/ns-moosavi).
-
-The HANS dataset can be downloaded from [this location](https://github.com/tommccoy1/hans).
-
-This is an example of using test_hans.py:
-
-```bash
-export HANS_DIR=path-to-hans
-export MODEL_TYPE=type-of-the-model-e.g.-bert-roberta-xlnet-etc
-export MODEL_PATH=path-to-the-model-directory-that-is-trained-on-NLI-e.g.-by-using-run_glue.py
-
-python examples/hans/test_hans.py \
-        --task_name hans \
-        --model_type $MODEL_TYPE \
-        --do_eval \
-        --data_dir $HANS_DIR \
-        --model_name_or_path $MODEL_PATH \
-        --max_seq_length 128 \
-        --output_dir $MODEL_PATH \
-```
-
-This will create the hans_predictions.txt file in MODEL_PATH, which can then be evaluated using hans/evaluate_heur_output.py from the HANS dataset.
-
-The results of the BERT-base model that is trained on MNLI using batch size 8 and the random seed 42 on the HANS dataset is as follows:
-
-```bash
-Heuristic entailed results:
-lexical_overlap: 0.9702
-subsequence: 0.9942
-constituent: 0.9962
-
-Heuristic non-entailed results:
-lexical_overlap: 0.199
-subsequence: 0.0396
-constituent: 0.118
-```

docs/source/examples.md

@@ -0,0 +1 @@
+../../examples/README.md

docs/source/index.rst

@@ -109,3 +109,4 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     model_doc/dialogpt
     model_doc/reformer
     model_doc/marian
+    model_doc/longformer

docs/source/model_doc/longformer.rst

@@ -0,0 +1,76 @@
+Longformer
+----------------------------------------------------
+**DISCLAIMER:** This model is still a work in progress, if you see something strange,
+file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`_
+
+Overview
+~~~~~
+The Longformer model was presented in `Longformer: The Long-Document Transformer <https://arxiv.org/pdf/2004.05150.pdf>`_ by Iz Beltagy, Matthew E. Peters, Arman Cohan.
+Here the abstract: 
+
+*Transformer-based models are unable to process long sequences due to their self-attention operation, which scales quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or longer. Longformer's attention mechanism is a drop-in replacement for the standard self-attention and combines a local windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on WikiHop and TriviaQA.*
+
+The Authors' code can be found `here <https://github.com/allenai/longformer>`_ .
+
+Longformer Self Attention
+~~~~~~~~~~~~~~~~~~~~
+Longformer self attention employs self attention on both a "local" context and a "global" context.
+Most tokens only attend "locally" to each other meaning that each token attends to its :math:`\frac{1}{2} w` previous tokens and :math:`\frac{1}{2} w` succeding tokens with :math:`w` being the window length as defined in `config.attention_window`. Note that `config.attention_window` can be of type ``list`` to define a different :math:`w` for each layer. 
+A selecetd few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.
+
+Note that "locally" and "globally" attending tokens are projected by different query, key and value matrices.
+Also note that every "locally" attending token not only attends to tokens within its window :math:`w`, but also to all "globally" attending tokens so that global attention is *symmetric*.
+
+The user can define which tokens are masked, which tokens attend "locally" and which tokens attend "globally" by setting the `config.attention_mask` `torch.Tensor` appropriately. In contrast to other models `Longformer` accepts the following values in `config.attention_mask`: `0` - the token is masked and not attended at all (as is done in other models), `1` - the token attends "locally", `2` - token attends "globally". For more information please also refer to :func:`~transformers.LongformerModel.forward` method.
+
+Using Longformer self attention, the memory and time complexity of the query-key matmul operation, which usually represents the memory and time bottleneck, can be reduced from :math:`\mathcal{O}(n_s \times n_s)` to :math:`\mathcal{O}(n_s \times w)`, with :math:`n_s` being the sequence length and :math:`w` being the average window size. It is assumed that the number of "globally" attending tokens is insignificant as compared to the number of "locally" attending tokens.
+
+For more information, please refer to the official `paper <https://arxiv.org/pdf/2004.05150.pdf>`_ .
+
+
+Training
+~~~~~~~~~~~~~~~~~~~~
+``LongformerForMaskedLM`` is trained the exact same way, ``RobertaForMaskedLM`` is trained and 
+should be used as follows:
+
+::
+
+  input_ids = tokenizer.encode('This is a sentence from [MASK] training data', return_tensors='pt')
+  mlm_labels = tokenizer.encode('This is a sentence from the training data', return_tensors='pt')
+
+  loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]
+
+
+LongformerConfig
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerConfig
+    :members:
+
+
+LongformerTokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerTokenizer
+    :members: 
+
+
+LongformerModel
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerModel
+    :members:
+
+
+LongformerForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerForMaskedLM
+    :members:
+
+
+LongformerForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerForQuestionAnswering
+    :members:

docs/source/model_doc/reformer.rst

@@ -5,7 +5,7 @@ file a `Github Issue <https://github.com/huggingface/transformers/issues/new?ass
 
 Overview
 ~~~~~
-The Reformer model was presented in `Reformer: The Efficient Transformer <https://https://arxiv.org/abs/2001.04451.pdf>`_ by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
+The Reformer model was presented in `Reformer: The Efficient Transformer <https://arxiv.org/abs/2001.04451.pdf>`_ by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
 Here the abstract: 
 
 *Large Transformer models routinely achieve state-of-the-art results on a number of tasks but training these models can be prohibitively costly, especially on long sequences. We introduce two techniques to improve the efficiency of Transformers. For one, we replace dot-product attention by one that uses locality-sensitive hashing, changing its complexity from O(L^2) to O(Llog(L)), where L is the length of the sequence. Furthermore, we use reversible residual layers instead of the standard residuals, which allows storing activations only once in the training process instead of N times, where N is the number of layers. The resulting model, the Reformer, performs on par with Transformer models while being much more memory-efficient and much faster on long sequences.*
@@ -62,7 +62,7 @@ For more information, see the `original Paper <https://arxiv.org/abs/2001.04451>
 
 Note that ``config.num_buckets`` can also be factorized into a ``list``:math:`(n_{\text{buckets}}^1, n_{\text{buckets}}^2)`. This way instead of assigning the query key embedding vectors to one of :math:`(1,\ldots, n_{\text{buckets}})` they are assigned to one of :math:`(1-1,\ldots, n_{\text{buckets}}^1-1, \ldots, 1-n_{\text{buckets}}^2, \ldots, n_{\text{buckets}}^1-n_{\text{buckets}}^2)`. This is crucial for very long sequences to save memory.
 
-It is recommended to leave ``config.num_buckets=None``, so that depending on the sequence length, a good value for ``num_buckets`` are calculated on the fly.
+When training a model from scratch, it is recommended to leave ``config.num_buckets=None``, so that depending on the sequence length a good value for ``num_buckets`` is calculated on the fly. This value will then automatically be saved in the config and should be reused for inference.
 
 Using LSH self attention, the memory and time complexity of the query-key matmul operation can be reduced from :math:`\mathcal{O}(n_s \times n_s)` to :math:`\mathcal{O}(n_s \times \log(n_s))`, which usually represents the memory and time bottleneck in a transformer model, with :math:`n_s` being the sequence length.
 

docs/source/pretrained_models.rst

@@ -305,3 +305,9 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 | MarianMT          | ``Helsinki-NLP/opus-mt-{src}-{tgt}``                       | | 12-layer, 512-hidden, 8-heads, ~74M parameter Machine translation models. Parameter counts vary depending on vocab size.            |
 |                   |                                                            | | (see `model list <https://huggingface.co/Helsinki-NLP>`_)                                                                           |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| Longformer        | ``longformer-base-4096``                                   | | 12-layer, 768-hidden, 12-heads, ~149M parameters                                                                                    |
+|                   |                                                            | | Starting from RoBERTa-base checkpoint, trained on documents of max length 4,096                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``longformer-large-4096``                                  | | 24-layer, 1024-hidden, 16-heads, ~435M parameters                                                                                   |
+|                   |                                                            | | Starting from RoBERTa-large checkpoint, trained on documents of max length 4,096                                                    |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

examples/README.md

@@ -16,17 +16,17 @@ This is still a work-in-progress – in particular documentation is still sparse
 
 | Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab
 |---|---|:---:|:---:|:---:|:---:|
-| [**`language-modeling`**](./language-modeling)       | Raw text        | ✅ | -  | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
-| [**`text-classification`**](./text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/trainer/01_text_classification.ipynb)
-| [**`token-classification`**](./token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
-| [**`multiple-choice`**](./multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
-| [**`question-answering`**](./question-answering)     | SQuAD           | -  | ✅ | -  | -
-| [**`text-generation`**](./text-generation)     | -           | -  | - | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
-| [**`distillation`**](./distillation)       | All               | -  | -  | -  | -
-| [**`summarization`**](./summarization)     | CNN/Daily Mail    | -  | -  | -  | -
-| [**`translation`**](./translation)         | WMT               | -  | -  | -  | -
-| [**`bertology`**](./bertology)             | -                 | -  | -  | -  | -
-| [**`adversarial`**](./adversarial)         | HANS              | -  | -  | -  | -
+| [**`language-modeling`**](https://github.com/huggingface/transformers/tree/master/examples/language-modeling)       | Raw text        | ✅ | -  | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb)
+| [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/trainer/01_text_classification.ipynb)
+| [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
+| [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
+| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | -  | ✅ | -  | -
+| [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)     | -           | -  | - | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
+| [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)       | All               | -  | -  | -  | -
+| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/summarization)     | CNN/Daily Mail    | -  | -  | -  | -
+| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/translation)         | WMT               | -  | -  | -  | -
+| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)             | -                 | -  | -  | -  | -
+| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)         | HANS              | -  | -  | -  | -
 
 
 <br>
@@ -57,7 +57,7 @@ When using Tensorflow, TPUs are supported out of the box as a `tf.distribute.Str
 When using PyTorch, we support TPUs thanks to `pytorch/xla`. For more context and information on how to setup your TPU environment refer to Google's documentation and to the
 very detailed [pytorch/xla README](https://github.com/pytorch/xla/blob/master/README.md).
 
-In this repo, we provide a very simple launcher script named [xla_spawn.py](./xla_spawn.py) that lets you run our example scripts on multiple TPU cores without any boilerplate.
+In this repo, we provide a very simple launcher script named [xla_spawn.py](https://github.com/huggingface/transformers/tree/master/examples/xla_spawn.py) that lets you run our example scripts on multiple TPU cores without any boilerplate.
 Just pass a `--num_cores` flag to this script, then your regular training script with its arguments (this is similar to the `torch.distributed.launch` helper for torch.distributed).
 
 For example for `run_glue`:

examples/benchmarking/plot_csv_file.py

@@ -0,0 +1,113 @@
+import csv
+from collections import defaultdict
+from dataclasses import dataclass, field
+from typing import Optional
+
+import numpy as np
+
+import matplotlib.pyplot as plt
+from transformers import HfArgumentParser
+
+
+@dataclass
+class PlotArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    csv_file: str = field(metadata={"help": "The csv file to plot."},)
+    plot_along_batch: bool = field(
+        default=False,
+        metadata={"help": "Whether to plot along batch size or sequence lengh. Defaults to sequence length."},
+    )
+    is_time: bool = field(
+        default=False,
+        metadata={"help": "Whether the csv file has time results or memory results. Defaults to memory results."},
+    )
+    is_train: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether the csv file has training results or inference results. Defaults to inference results."
+        },
+    )
+    figure_png_file: Optional[str] = field(
+        default=None, metadata={"help": "Filename under which the plot will be saved. If unused no plot is saved."},
+    )
+
+
+class Plot:
+    def __init__(self, args):
+        self.args = args
+        self.result_dict = defaultdict(lambda: dict(bsz=[], seq_len=[], result={}))
+
+        with open(self.args.csv_file, newline="") as csv_file:
+            reader = csv.DictReader(csv_file)
+            for row in reader:
+                model_name = row["model"]
+                self.result_dict[model_name]["bsz"].append(int(row["batch_size"]))
+                self.result_dict[model_name]["seq_len"].append(int(row["sequence_length"]))
+                self.result_dict[model_name]["result"][(int(row["batch_size"]), int(row["sequence_length"]))] = row[
+                    "result"
+                ]
+
+    def plot(self):
+        fig, ax = plt.subplots()
+        title_str = "Time usage" if self.args.is_time else "Memory usage"
+        title_str = title_str + " for training" if self.args.is_train else title_str + " for inference"
+
+        for model_name in self.result_dict.keys():
+            batch_sizes = sorted(list(set(self.result_dict[model_name]["bsz"])))
+            sequence_lengths = sorted(list(set(self.result_dict[model_name]["seq_len"])))
+            results = self.result_dict[model_name]["result"]
+
+            (x_axis_array, inner_loop_array) = (
+                (batch_sizes, sequence_lengths) if self.args.plot_along_batch else (sequence_lengths, batch_sizes)
+            )
+
+            plt.xlim(min(x_axis_array), max(x_axis_array))
+
+            for inner_loop_value in inner_loop_array:
+                if self.args.plot_along_batch:
+                    y_axis_array = np.asarray([results[(x, inner_loop_value)] for x in x_axis_array], dtype=np.int)
+                else:
+                    y_axis_array = np.asarray([results[(inner_loop_value, x)] for x in x_axis_array], dtype=np.float32)
+
+                ax.set_xscale("log", basex=2)
+                ax.set_yscale("log", basey=10)
+
+                (x_axis_label, inner_loop_label) = (
+                    ("batch_size", "sequence_length in #tokens")
+                    if self.args.plot_along_batch
+                    else ("sequence_length in #tokens", "batch_size")
+                )
+
+                x_axis_array = np.asarray(x_axis_array, np.int)
+                plt.scatter(x_axis_array, y_axis_array, label=f"{model_name} - {inner_loop_label}: {inner_loop_value}")
+                plt.plot(x_axis_array, y_axis_array, "--")
+
+            title_str += f" {model_name} vs."
+
+        title_str = title_str[:-4]
+        y_axis_label = "Time in s" if self.args.is_time else "Memory in MB"
+
+        # plot
+        plt.title(title_str)
+        plt.xlabel(x_axis_label)
+        plt.ylabel(y_axis_label)
+        plt.legend()
+
+        if self.args.figure_png_file is not None:
+            plt.savefig(self.args.figure_png_file)
+        else:
+            plt.show()
+
+
+def main():
+    parser = HfArgumentParser(PlotArguments)
+    plot_args = parser.parse_args_into_dataclasses()[0]
+    plot = Plot(args=plot_args)
+    plot.plot()
+
+
+if __name__ == "__main__":
+    main()

examples/benchmarking/run_benchmark.py

@@ -0,0 +1,29 @@
+# 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 """
+
+from transformers import HfArgumentParser, PyTorchBenchmark, PyTorchBenchmarkArguments
+
+
+def main():
+    parser = HfArgumentParser(PyTorchBenchmarkArguments)
+    benchmark_args = parser.parse_args_into_dataclasses()[0]
+    benchmark = PyTorchBenchmark(args=benchmark_args)
+    benchmark.run()
+
+
+if __name__ == "__main__":
+    main()

examples/benchmarks.py

@@ -1,710 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The HuggingFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-""" Benchmarking the library on inference and training """
-
-# If checking the tensors placement
-# tf.debugging.set_log_device_placement(True)
-
-import argparse
-import csv
-import logging
-import timeit
-from time import time
-from typing import Callable, List
-
-from transformers import (
-    AutoConfig,
-    AutoTokenizer,
-    MemorySummary,
-    is_tf_available,
-    is_torch_available,
-    start_memory_tracing,
-    stop_memory_tracing,
-)
-
-
-if is_tf_available():
-    import tensorflow as tf
-    from transformers import TFAutoModel
-
-if is_torch_available():
-    import torch
-    from transformers import AutoModel
-
-
-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],
-    batch_sizes: List[int],
-    slice_sizes: List[int],
-    gpu: bool = True,
-    tensorflow: bool = False,
-    average_over: int = 3,
-    no_speed: bool = False,
-    no_memory: bool = False,
-    verbose: bool = False,
-    torchscript: bool = False,
-    xla: bool = False,
-    amp: bool = False,
-    fp16: bool = False,
-    save_to_csv: bool = False,
-    csv_time_filename: str = f"time_{round(time())}.csv",
-    csv_memory_filename: str = f"memory_{round(time())}.csv",
-    print_fn: Callable[[str], None] = print,
-):
-    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,
-            batch_sizes,
-            slice_sizes,
-            dictionary,
-            average_over,
-            amp,
-            no_speed,
-            no_memory,
-            verbose,
-            print_fn,
-        )
-    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,
-            batch_sizes,
-            slice_sizes,
-            dictionary,
-            average_over,
-            device,
-            torchscript,
-            fp16,
-            no_speed,
-            no_memory,
-            verbose,
-            print_fn,
-        )
-
-    print_fn("=========== RESULTS ===========")
-    for model_name in model_names:
-        print_fn("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
-        for batch_size in results[model_name]["bs"]:
-            print_fn("\t\t" + f"===== BATCH SIZE: {batch_size} =====")
-            for slice_size in results[model_name]["ss"]:
-                time = results[model_name]["time"][batch_size][slice_size]
-                memory = results[model_name]["memory"][batch_size][slice_size]
-                if isinstance(time, str):
-                    print_fn(f"\t\t{model_name}/{batch_size}/{slice_size}: " f"{time} " f"{memory}")
-                else:
-                    print_fn(
-                        f"\t\t{model_name}/{batch_size}/{slice_size}: "
-                        f"{(round(1000 * time) / 1000)}"
-                        f"s "
-                        f"{memory}"
-                    )
-
-    if save_to_csv:
-        with open(csv_time_filename, mode="w") as csv_time_file, open(
-            csv_memory_filename, mode="w"
-        ) as csv_memory_file:
-
-            assert len(model_names) > 0, "At least 1 model should be defined, but got {}".format(model_names)
-
-            fieldnames = ["model", "batch_size", "sequence_length"]
-            time_writer = csv.DictWriter(csv_time_file, fieldnames=fieldnames + ["time_in_s"])
-            time_writer.writeheader()
-            memory_writer = csv.DictWriter(csv_memory_file, fieldnames=fieldnames + ["memory"])
-            memory_writer.writeheader()
-
-            for model_name in model_names:
-                time_dict = results[model_name]["time"]
-                memory_dict = results[model_name]["memory"]
-                for bs in time_dict:
-                    for ss in time_dict[bs]:
-                        time_writer.writerow(
-                            {
-                                "model": model_name,
-                                "batch_size": bs,
-                                "sequence_length": ss,
-                                "time_in_s": "{:.4f}".format(time_dict[bs][ss]),
-                            }
-                        )
-
-                for bs in memory_dict:
-                    for ss in time_dict[bs]:
-                        memory_writer.writerow(
-                            {
-                                "model": model_name,
-                                "batch_size": bs,
-                                "sequence_length": ss,
-                                "memory": memory_dict[bs][ss],
-                            }
-                        )
-
-
-def print_summary_statistics(summary: MemorySummary, print_fn: Callable[[str], None]):
-    print_fn(
-        "\nLines by line memory consumption:\n"
-        + "\n".join(
-            f"{state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
-            for state in summary.sequential
-        )
-    )
-    print_fn(
-        "\nLines with top memory consumption:\n"
-        + "\n".join(
-            f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
-            for state in summary.cumulative[:6]
-        )
-    )
-    print_fn(
-        "\nLines with lowest memory consumption:\n"
-        + "\n".join(
-            f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
-            for state in summary.cumulative[-6:]
-        )
-    )
-    print_fn(f"\nTotal memory increase: {summary.total}")
-
-
-def get_print_function(save_print_log, log_filename):
-    if save_print_log:
-        logging.basicConfig(
-            level=logging.DEBUG,
-            filename=log_filename,
-            filemode="a+",
-            format="%(asctime)-15s %(levelname)-8s %(message)s",
-        )
-
-        def print_with_print_log(*args):
-            logging.info(*args)
-            print(*args)
-
-        return print_with_print_log
-    else:
-        return print
-
-
-def _compute_pytorch(
-    model_names,
-    batch_sizes,
-    slice_sizes,
-    dictionary,
-    average_over,
-    device,
-    torchscript,
-    fp16,
-    no_speed,
-    no_memory,
-    verbose,
-    print_fn,
-):
-    for c, model_name in enumerate(model_names):
-        print_fn(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]
-
-        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "time": {}, "memory": {}}
-        dictionary[model_name]["time"] = {i: {} for i in batch_sizes}
-        dictionary[model_name]["memory"] = {i: {} for i in batch_sizes}
-
-        print_fn("Using model {}".format(model))
-        print_fn("Number of all parameters {}".format(model.num_parameters()))
-
-        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]["time"][batch_size][slice_size] = "N/A"
-                else:
-                    sequence = torch.tensor(tokenized_sequence[:slice_size], device=device).repeat(batch_size, 1)
-                    try:
-                        if torchscript:
-                            print_fn("Tracing model with sequence size {}".format(sequence.shape))
-                            inference = torch.jit.trace(model, sequence)
-                            inference(sequence)
-                        else:
-                            inference = model
-                            inference(sequence)
-
-                        if not no_memory:
-                            # model.add_memory_hooks()  # Forward method tracing (only for PyTorch models)
-
-                            # Line by line memory tracing (all code in the module `transformers`) works for all models/arbitrary code
-                            trace = start_memory_tracing("transformers")
-                            inference(sequence)
-                            summary = stop_memory_tracing(trace)
-
-                            if verbose:
-                                print_summary_statistics(summary, print_fn)
-
-                            dictionary[model_name]["memory"][batch_size][slice_size] = str(summary.total)
-                        else:
-                            dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
-
-                        if not no_speed:
-                            print_fn("Going through model with sequence of shape {}".format(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]["time"][batch_size][slice_size] = average_time
-                        else:
-                            dictionary[model_name]["time"][batch_size][slice_size] = "N/A"
-
-                    except RuntimeError as e:
-                        print_fn("Doesn't fit on GPU. {}".format(e))
-                        torch.cuda.empty_cache()
-                        dictionary[model_name]["time"][batch_size][slice_size] = "N/A"
-                        dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
-    return dictionary
-
-
-def _compute_tensorflow(
-    model_names, batch_sizes, slice_sizes, dictionary, average_over, amp, no_speed, no_memory, verbose, print_fn
-):
-    for c, model_name in enumerate(model_names):
-        print_fn(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]
-
-        dictionary[model_name] = {"bs": batch_sizes, "ss": slice_sizes, "time": {}, "memory": {}}
-        dictionary[model_name]["time"] = {i: {} for i in batch_sizes}
-        dictionary[model_name]["memory"] = {i: {} for i in batch_sizes}
-
-        print_fn("Using model {}".format(model))
-        print_fn("Number of all parameters {}".format(model.num_parameters()))
-
-        @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]["time"][batch_size][slice_size] = "N/A"
-                else:
-                    sequence = tf.stack(
-                        [tf.squeeze(tf.constant(tokenized_sequence[:slice_size])[None, :])] * batch_size
-                    )
-
-                    try:
-                        print_fn("Going through model with sequence of shape {}".format(sequence.shape))
-                        # To make sure that the model is traced + that the tensors are on the appropriate device
-                        inference(sequence)
-
-                        if not no_memory:
-                            # Line by line memory tracing (all code in the module `transformers`) works for all models/arbitrary code
-                            trace = start_memory_tracing("transformers")
-                            inference(sequence)
-                            summary = stop_memory_tracing(trace)
-
-                            if verbose:
-                                print_summary_statistics(summary, print_fn)
-
-                            dictionary[model_name]["memory"][batch_size][slice_size] = str(summary.total)
-                        else:
-                            dictionary[model_name]["memory"][batch_size][slice_size] = "N/A"
-
-                        if not no_speed:
-                            runtimes = timeit.repeat(lambda: inference(sequence), repeat=average_over, number=3)
-                            average_time = sum(runtimes) / float(len(runtimes)) / 3.0
-                            dictionary[model_name]["time"][batch_size][slice_size] = average_time
-                        else:
-                            dictionary[model_name]["time"][batch_size][slice_size] = "N/A"
-
-                    except tf.errors.ResourceExhaustedError as e:
-                        print_fn("Doesn't fit on GPU. {}".format(e))
-                        dictionary[model_name]["time"][batch_size][slice_size] = "N/A"
-                        dictionary[model_name]["memory"][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("--verbose", required=False, action="store_true", help="Verbose memory tracing")
-    parser.add_argument("--no_speed", required=False, action="store_true", help="Don't perform speed measurments")
-    parser.add_argument("--no_memory", required=False, action="store_true", help="Don't perform memory measurments")
-    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(
-        "--log_print", required=False, action="store_true", help="Save all print statements in log file."
-    )
-    parser.add_argument(
-        "--csv_time_filename",
-        required=False,
-        default=f"time_{round(time())}.csv",
-        help="CSV filename used if saving time results to csv.",
-    )
-    parser.add_argument(
-        "--csv_memory_filename",
-        required=False,
-        default=f"memory_{round(time())}.csv",
-        help="CSV filename used if saving memory results to csv.",
-    )
-    parser.add_argument(
-        "--log_filename",
-        required=False,
-        default=f"log_{round(time())}.txt",
-        help="Log filename used if print statements are saved in log.",
-    )
-    parser.add_argument(
-        "--average_over", required=False, default=30, type=int, help="Times an experiment will be run."
-    )
-    parser.add_argument("--batch_sizes", nargs="+", type=int, default=[1, 2, 4, 8])
-    parser.add_argument("--slice_sizes", nargs="+", type=int, default=[8, 64, 128, 256, 512, 1024])
-
-    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",
-            "t5-base",
-            "bart-large",
-        ]
-    else:
-        args.models = args.models.split()
-
-    print_fn = get_print_function(args.log_print, args.log_filename)
-    print_fn("Running with arguments: {}".format(args))
-
-    if args.torch:
-        if is_torch_available():
-            create_setup_and_compute(
-                model_names=args.models,
-                batch_sizes=args.batch_sizes,
-                slice_sizes=args.slice_sizes,
-                tensorflow=False,
-                gpu=args.torch_cuda,
-                torchscript=args.torchscript,
-                fp16=args.fp16,
-                save_to_csv=args.save_to_csv,
-                csv_time_filename=args.csv_time_filename,
-                csv_memory_filename=args.csv_memory_filename,
-                average_over=args.average_over,
-                no_speed=args.no_speed,
-                no_memory=args.no_memory,
-                verbose=args.verbose,
-                print_fn=print_fn,
-            )
-        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,
-                batch_sizes=args.batch_sizes,
-                slice_sizes=args.slice_sizes,
-                tensorflow=True,
-                xla=args.xla,
-                amp=args.amp,
-                save_to_csv=args.save_to_csv,
-                csv_time_filename=args.csv_time_filename,
-                csv_memory_filename=args.csv_memory_filename,
-                average_over=args.average_over,
-                no_speed=args.no_speed,
-                no_memory=args.no_memory,
-                verbose=args.verbose,
-                print_fn=print_fn,
-            )
-        else:
-            raise ImportError("Trying to run a TensorFlow benchmark but TensorFlow was not found in the environment.")
-
-
-if __name__ == "__main__":
-    main()

examples/bertology/run_bertology.py

@@ -64,7 +64,7 @@ def print_2d_tensor(tensor):
 
 
 def compute_heads_importance(
-    args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None
+    args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None, actually_pruned=False
 ):
     """ This method shows how to compute:
         - head attention entropy
@@ -77,7 +77,12 @@ def compute_heads_importance(
 
     if head_mask is None:
         head_mask = torch.ones(n_layers, n_heads).to(args.device)
+
     head_mask.requires_grad_(requires_grad=True)
+    # If actually pruned attention multi-head, set head mask to None to avoid shape mismatch
+    if actually_pruned:
+        head_mask = None
+
     preds = None
     labels = None
     tot_tokens = 0.0
@@ -172,6 +177,7 @@ def mask_heads(args, model, eval_dataloader):
         new_head_mask = new_head_mask.view(-1)
         new_head_mask[current_heads_to_mask] = 0.0
         new_head_mask = new_head_mask.view_as(head_mask)
+        new_head_mask = new_head_mask.clone().detach()
         print_2d_tensor(new_head_mask)
 
         # Compute metric and head importance again
@@ -181,7 +187,7 @@ def mask_heads(args, model, eval_dataloader):
         preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
         current_score = glue_compute_metrics(args.task_name, preds, labels)[args.metric_name]
         logger.info(
-            "Masking: current score: %f, remaning heads %d (%.1f percents)",
+            "Masking: current score: %f, remaining heads %d (%.1f percents)",
             current_score,
             new_head_mask.sum(),
             new_head_mask.sum() / new_head_mask.numel() * 100,
@@ -209,14 +215,23 @@ def prune_heads(args, model, eval_dataloader, head_mask):
     original_time = datetime.now() - before_time
 
     original_num_params = sum(p.numel() for p in model.parameters())
-    heads_to_prune = dict((layer, (1 - head_mask[layer].long()).nonzero().tolist()) for layer in range(len(head_mask)))
+    heads_to_prune = dict(
+        (layer, (1 - head_mask[layer].long()).nonzero().squeeze().tolist()) for layer in range(len(head_mask))
+    )
+
     assert sum(len(h) for h in heads_to_prune.values()) == (1 - head_mask.long()).sum().item()
     model.prune_heads(heads_to_prune)
     pruned_num_params = sum(p.numel() for p in model.parameters())
 
     before_time = datetime.now()
     _, _, preds, labels = compute_heads_importance(
-        args, model, eval_dataloader, compute_entropy=False, compute_importance=False, head_mask=None
+        args,
+        model,
+        eval_dataloader,
+        compute_entropy=False,
+        compute_importance=False,
+        head_mask=None,
+        actually_pruned=True,
     )
     preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
     score_pruning = glue_compute_metrics(args.task_name, preds, labels)[args.metric_name]
@@ -404,7 +419,7 @@ def main():
     logger.info("Training/evaluation parameters %s", args)
 
     # Prepare dataset for the GLUE task
-    eval_dataset = GlueDataset(args, tokenizer=tokenizer, evaluate=True)
+    eval_dataset = GlueDataset(args, tokenizer=tokenizer, mode="dev")
     if args.data_subset > 0:
         eval_dataset = Subset(eval_dataset, list(range(min(args.data_subset, len(eval_dataset)))))
     eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)

examples/contrib/run_transfo_xl.py

@@ -80,7 +80,7 @@ def main():
 
     # Load a pre-trained model
     model = TransfoXLLMHeadModel.from_pretrained(args.model_name)
-    model = model.to(device)
+    model.to(device)
 
     logger.info(
         "Evaluating with bsz {} tgt_len {} ext_len {} mem_len {} clamp_len {}".format(

examples/language-modeling/README.md

@@ -3,8 +3,7 @@
 
 Based on the script [`run_language_modeling.py`](https://github.com/huggingface/transformers/blob/master/examples/language-modeling/run_language_modeling.py).
 
-Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT and RoBERTa (DistilBERT
-to be added soon). GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa
+Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT, DistilBERT and RoBERTa. GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT, DistilBERT and RoBERTa
 are fine-tuned using a masked language modeling (MLM) loss.
 
 Before running the following example, you should get a file that contains text on which the language model will be
@@ -35,7 +34,7 @@ python run_language_modeling.py \
 This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run. It reaches
 a score of ~20 perplexity once fine-tuned on the dataset.
 
-### RoBERTa/BERT and masked language modeling
+### RoBERTa/BERT/DistilBERT and masked language modeling
 
 The following example fine-tunes RoBERTa on WikiText-2. Here too, we're using the raw WikiText-2. The loss is different
 as BERT/RoBERTa have a bidirectional mechanism; we're therefore using the same loss that was used during their

examples/multiple-choice/README.md

@@ -19,7 +19,7 @@ python ./examples/multiple-choice/run_multiple_choice.py \
 --max_seq_length 80 \
 --output_dir models_bert/swag_base \
 --per_gpu_eval_batch_size=16 \
---per_gpu_train_batch_size=16 \
+--per_device_train_batch_size=16 \
 --gradient_accumulation_steps 2 \
 --overwrite_output
 ```
@@ -46,7 +46,7 @@ python ./examples/multiple-choice/run_tf_multiple_choice.py \
 --max_seq_length 80 \
 --output_dir models_bert/swag_base \
 --per_gpu_eval_batch_size=16 \
---per_gpu_train_batch_size=16 \
+--per_device_train_batch_size=16 \
 --logging-dir logs \
 --gradient_accumulation_steps 2 \
 --overwrite_output

examples/question-answering/README.md

@@ -28,6 +28,7 @@ python run_squad.py \
   --model_name_or_path bert-base-uncased \
   --do_train \
   --do_eval \
+  --do_lower_case \
   --train_file $SQUAD_DIR/train-v1.1.json \
   --predict_file $SQUAD_DIR/dev-v1.1.json \
   --per_gpu_train_batch_size 12 \
@@ -56,6 +57,7 @@ python -m torch.distributed.launch --nproc_per_node=8 ./examples/question-answer
     --model_name_or_path bert-large-uncased-whole-word-masking \
     --do_train \
     --do_eval \
+    --do_lower_case \
     --train_file $SQUAD_DIR/train-v1.1.json \
     --predict_file $SQUAD_DIR/dev-v1.1.json \
     --learning_rate 3e-5 \

examples/requirements.txt

@@ -6,3 +6,4 @@ sacrebleu
 rouge-score
 tensorflow_datasets
 pytorch-lightning==0.7.3  # April 10, 2020 release
+matplotlib

examples/test_examples.py

@@ -61,8 +61,8 @@ class ExamplesTests(unittest.TestCase):
             --do_train
             --do_eval
             --output_dir ./tests/fixtures/tests_samples/temp_dir
-            --per_gpu_train_batch_size=2
-            --per_gpu_eval_batch_size=1
+            --per_device_train_batch_size=2
+            --per_device_eval_batch_size=1
             --learning_rate=1e-4
             --max_steps=10
             --warmup_steps=2

examples/text-classification/README.md

@@ -68,7 +68,7 @@ python run_glue.py \
   --do_eval \
   --data_dir $GLUE_DIR/$TASK_NAME \
   --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
+  --per_device_train_batch_size 32 \
   --learning_rate 2e-5 \
   --num_train_epochs 3.0 \
   --output_dir /tmp/$TASK_NAME/
@@ -141,7 +141,7 @@ python run_glue.py \
   --do_eval \
   --data_dir $GLUE_DIR/MRPC/ \
   --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
+  --per_device_train_batch_size 32 \
   --learning_rate 2e-5 \
   --num_train_epochs 3.0 \
   --output_dir /tmp/mrpc_output/
@@ -166,7 +166,7 @@ python run_glue.py \
   --do_eval \
   --data_dir $GLUE_DIR/MRPC/ \
   --max_seq_length 128 \
-  --per_gpu_train_batch_size 32 \
+  --per_device_train_batch_size 32 \
   --learning_rate 2e-5 \
   --num_train_epochs 3.0 \
   --output_dir /tmp/mrpc_output/ \
@@ -189,7 +189,7 @@ python -m torch.distributed.launch \
     --do_eval \
     --data_dir $GLUE_DIR/MRPC/ \
     --max_seq_length 128 \
-    --per_gpu_train_batch_size 8 \
+    --per_device_train_batch_size 8 \
     --learning_rate 2e-5 \
     --num_train_epochs 3.0 \
     --output_dir /tmp/mrpc_output/
@@ -221,7 +221,7 @@ python -m torch.distributed.launch \
     --do_eval \
     --data_dir $GLUE_DIR/MNLI/ \
     --max_seq_length 128 \
-    --per_gpu_train_batch_size 8 \
+    --per_device_train_batch_size 8 \
     --learning_rate 2e-5 \
     --num_train_epochs 3.0 \
     --output_dir output_dir \
@@ -280,7 +280,7 @@ python run_xnli.py \
   --do_train \
   --do_eval \
   --data_dir $XNLI_DIR \
-  --per_gpu_train_batch_size 32 \
+  --per_device_train_batch_size 32 \
   --learning_rate 5e-5 \
   --num_train_epochs 2.0 \
   --max_seq_length 128 \

examples/text-classification/run_glue.py

@@ -135,7 +135,8 @@ def main():
 
     # Get datasets
     train_dataset = GlueDataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
-    eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
+    eval_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev") if training_args.do_eval else None
+    test_dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="test") if training_args.do_predict else None
 
     def compute_metrics(p: EvalPrediction) -> Dict:
         if output_mode == "classification":
@@ -165,7 +166,7 @@ def main():
             tokenizer.save_pretrained(training_args.output_dir)
 
     # Evaluation
-    results = {}
+    eval_results = {}
     if training_args.do_eval:
         logger.info("*** Evaluate ***")
 
@@ -173,10 +174,10 @@ def main():
         eval_datasets = [eval_dataset]
         if data_args.task_name == "mnli":
             mnli_mm_data_args = dataclasses.replace(data_args, task_name="mnli-mm")
-            eval_datasets.append(GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, evaluate=True))
+            eval_datasets.append(GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, mode="dev"))
 
         for eval_dataset in eval_datasets:
-            result = trainer.evaluate(eval_dataset=eval_dataset)
+            eval_result = trainer.evaluate(eval_dataset=eval_dataset)
 
             output_eval_file = os.path.join(
                 training_args.output_dir, f"eval_results_{eval_dataset.args.task_name}.txt"
@@ -184,13 +185,38 @@ def main():
             if trainer.is_world_master():
                 with open(output_eval_file, "w") as writer:
                     logger.info("***** Eval results {} *****".format(eval_dataset.args.task_name))
-                    for key, value in result.items():
+                    for key, value in eval_result.items():
                         logger.info("  %s = %s", key, value)
                         writer.write("%s = %s\n" % (key, value))
 
-            results.update(result)
+            eval_results.update(eval_result)
 
-    return results
+    if training_args.do_predict:
+        logging.info("*** Test ***")
+        test_datasets = [test_dataset]
+        if data_args.task_name == "mnli":
+            mnli_mm_data_args = dataclasses.replace(data_args, task_name="mnli-mm")
+            test_datasets.append(GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, mode="test"))
+
+        for test_dataset in test_datasets:
+            predictions = trainer.predict(test_dataset=test_dataset).predictions
+            if output_mode == "classification":
+                predictions = np.argmax(predictions, axis=1)
+
+            output_test_file = os.path.join(
+                training_args.output_dir, f"test_results_{test_dataset.args.task_name}.txt"
+            )
+            if trainer.is_world_master():
+                with open(output_test_file, "w") as writer:
+                    logger.info("***** Test results {} *****".format(test_dataset.args.task_name))
+                    writer.write("index\tprediction\n")
+                    for index, item in enumerate(predictions):
+                        if output_mode == "regression":
+                            writer.write("%d\t%3.3f\n" % (index, item))
+                        else:
+                            item = test_dataset.get_labels()[item]
+                            writer.write("%d\t%s\n" % (index, item))
+    return eval_results
 
 
 def _mp_fn(index):

examples/token-classification/README.md

@@ -69,7 +69,7 @@ python3 run_ner.py --data_dir ./ \
 --output_dir $OUTPUT_DIR \
 --max_seq_length  $MAX_LENGTH \
 --num_train_epochs $NUM_EPOCHS \
---per_gpu_train_batch_size $BATCH_SIZE \
+--per_device_train_batch_size $BATCH_SIZE \
 --save_steps $SAVE_STEPS \
 --seed $SEED \
 --do_train \
@@ -91,7 +91,7 @@ Instead of passing all parameters via commandline arguments, the `run_ner.py` sc
     "output_dir": "germeval-model",
     "max_seq_length": 128,
     "num_train_epochs": 3,
-    "per_gpu_train_batch_size": 32,
+    "per_device_train_batch_size": 32,
     "save_steps": 750,
     "seed": 1,
     "do_train": true,

model_cards/HooshvareLab/bert-base-parsbert-uncased/README.md

@@ -0,0 +1,124 @@
+## ParsBERT: Transformer-based Model for Persian Language Understanding
+
+ParsBERT is a monolingual language model based on Google’s BERT architecture with the same configurations as BERT-Base. 
+
+Paper presenting ParsBERT: [arXiv:2005.12515](https://arxiv.org/abs/2005.12515)
+
+All the models (downstream tasks) are uncased and trained with whole word masking. (coming soon stay tuned)
+
+
+---
+
+## Introduction
+
+This model is pre-trained on a large Persian corpus with various writing styles from numerous subjects (e.g., scientific, novels, news) with more than 2M documents. A large subset of this corpus was crawled manually.
+
+As a part of ParsBERT methodology, an extensive pre-processing combining POS tagging and WordPiece segmentation was carried out to bring the corpus into a proper format. This process produces more than 40M true sentences. 
+
+
+## Evaluation
+
+ParsBERT is evaluated on three NLP downstream tasks: Sentiment Analysis (SA), Text Classification, and Named Entity Recognition (NER). For this matter and due to insufficient resources, two large datasets for SA and two for text classification were manually composed, which are available for public use and benchmarking. ParsBERT outperformed all other language models, including multilingual BERT and other hybrid deep learning models for all tasks, improving the state-of-the-art performance in Persian language modeling.
+
+## Results
+
+The following table summarizes the F1 score obtained by ParsBERT as compared to other models and architectures.
+
+
+
+### Sentiment Analysis (SA) task
+
+|           Dataset          |  ParsBERT | Multilingual BERT | DeepSentiPers |
+|:--------------------------:|:---------:|:-----------------:|:-------------:|
+|   Digikala User Comments   |   81.74*  |       80.74       |       -       |
+|   SnappFood User Comments  |   88.12*  |       87.87       |       -       |
+|   SentiPers (Multi Class)  |   71.11*  |         -         |     69.33     |
+|  SentiPers (Binary Class)  |   92.13*  |         -         |     91.98     |
+
+
+
+### Text Classification (TC) task
+
+|      Dataset      | ParsBERT | Multilingual BERT |
+|:-----------------:|:--------:|:-----------------:|
+| Digikala Magazine |   93.59* |       90.72       |
+|    Persian News   |   97.19* |       95.79       |
+
+
+### Named Entity Recognition (NER) task
+
+| Dataset | ParsBERT | MorphoBERT |  Beheshti-NER  |  LSTM-CRF  |  Rule-Based CRF  |  BiLSTM-CRF  |
+|:-------:|:--------:|:----------:|:--------------:|:----------:|:----------------:|:------------:|
+|  PEYMA  |   98.79* |      -     |      90.59     |      -     |       84.00      |       -      |
+|  ARMAN  |   93.10* |    89.9    |      84.03     |    86.55   |         -        |     77.45    |
+
+
+**If you tested ParsBERT on a public dataset and you want to add your results to the table above, open a pull request or contact us. Also make sure to have your code available online so we can add it as a reference**
+
+## How to use
+
+### TensorFlow 2.0
+
+```python
+from transformers import AutoConfig, AutoTokenizer, TFAutoModel
+
+config = AutoConfig.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+tokenizer = AutoTokenizer.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+model = AutoModel.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+
+text = "ما در هوشواره معتقدیم با انتقال صحیح دانش و آگاهی، همه‌ی افراد می‌توانند از ابزارهای هوشمند استفاده کنند. شعار ما هوش مصنوعی برای همه است."
+tokenizer.tokenize(text)
+
+>>> ['ما', 'در', 'هوش', '##واره', 'معتقدیم', 'با', 'انتقال', 'صحیح', 'دانش', 'و', 'اگاهی', '،', 'همهی', 'افراد', 'میتوانند', 'از', 'ابزارهای', 'هوشمند', 'استفاده', 'کنند', '.', 'شعار', 'ما', 'هوش', 'مصنوعی', 'برای', 'همه', 'است', '.']
+
+```
+
+### Pytorch
+
+```python
+from transformers import AutoConfig, AutoTokenizer, AutoModel
+
+config = AutoConfig.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+tokenizer = AutoTokenizer.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+model = AutoModel.from_pretrained("HooshvareLab/bert-base-parsbert-uncased")
+```
+
+
+## NLP Tasks Tutorial 
+
+Coming soon stay tuned
+
+
+## Cite 
+
+Please cite the following paper in your publication if you are using [ParsBERT](https://arxiv.org/abs/2005.12515) in your research:
+
+```markdown
+@article{ParsBERT,
+    title={ParsBERT: Transformer-based Model for Persian Language Understanding},
+    author={Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani, Mohammad Manthouri},
+    journal={ArXiv},
+    year={2020},
+    volume={abs/2005.12515}
+}
+```
+
+
+## Acknowledgments
+
+We hereby, express our gratitude to the [Tensorflow Research Cloud (TFRC) program](https://tensorflow.org/tfrc) for providing us with the necessary computation resources. We also thank [Hooshvare](https://hooshvare.com) Research Group for facilitating dataset gathering and scraping online text resources.
+
+
+## Contributors
+
+- Mehrdad Farahani: [Linkedin](https://www.linkedin.com/in/m3hrdadfi/), [Twitter](https://twitter.com/m3hrdadfi), [Github](https://github.com/m3hrdadfi)
+- Mohammad Gharachorloo:  [Linkedin](https://www.linkedin.com/in/mohammad-gharachorloo/), [Twitter](https://twitter.com/MGharachorloo), [Github](https://github.com/baarsaam)
+- Marzieh Farahani:  [Linkedin](https://www.linkedin.com/in/marziehphi/), [Twitter](https://twitter.com/marziehphi), [Github](https://github.com/marziehphi)
+- Mohammad Manthouri:  [Linkedin](https://www.linkedin.com/in/mohammad-manthouri-aka-mansouri-07030766/), [Twitter](https://twitter.com/mmanthouri), [Github](https://github.com/mmanthouri)
+- Hooshvare Team:  [Official Website](https://hooshvare.com/), [Linkedin](https://www.linkedin.com/company/hooshvare), [Twitter](https://twitter.com/hooshvare), [Github](https://github.com/hooshvare), [Instagram](https://www.instagram.com/hooshvare/)
+
+
+## Releases
+
+### Release v0.1 (May 27, 2019)
+This is the first version of our ParsBERT based on BERT<sub>BASE</sub>

model_cards/Tereveni-AI/gpt2-124M-uk-fiction/README.md

@@ -0,0 +1,39 @@
+---
+language: ukrainian
+---
+
+Note: **default code snippet above won't work** because we are using `AlbertTokenizer` with `GPT2LMHeadModel`, see [issue](https://github.com/huggingface/transformers/issues/4285).
+
+## GPT2 124M Trained on Ukranian Fiction
+
+### Training details
+
+Model was trained on corpus of 4040 fiction books, 2.77 GiB in total.
+Evaluation on [brown-uk](https://github.com/brown-uk/corpus) gives perplexity of 50.16. 
+
+### Example usage:
+```python
+from transformers import AlbertTokenizer, GPT2LMHeadModel
+
+tokenizer = AlbertTokenizer.from_pretrained("Tereveni-AI/gpt2-124M-uk-fiction")
+model = GPT2LMHeadModel.from_pretrained("Tereveni-AI/gpt2-124M-uk-fiction")
+
+input_ids = tokenizer.encode("Но зла Юнона, суча дочка,", add_special_tokens=False, return_tensors='pt')
+
+outputs = model.generate(
+    input_ids,
+    do_sample=True,
+    num_return_sequences=3,
+    max_length=50
+)
+
+for i, out in enumerate(outputs):
+    print("{}: {}".format(i, tokenizer.decode(out)))
+```
+
+Prints something like this:
+```bash
+0: Но зла Юнона, суча дочка, яка затьмарила всі її таємниці: І хто з'їсть її душу, той помре». І, не дочекавшись гніву богів, посунула в пітьму, щоб не бачити перед собою. Але, за
+1: Но зла Юнона, суча дочка, і довела мене до божевілля. Але він не знав нічого. Після того як я його побачив, мені стало зле. Я втратив рівновагу. Але в мене не було часу на роздуми. Я вже втратив надію
+2: Но зла Юнона, суча дочка, не нарікала нам! — раптом вигукнула Юнона. — Це ти, старий йолопе! — мовила вона, не перестаючи сміятись. — Хіба ти не знаєш, що мені подобається ходити з тобою?
+```

model_cards/activebus/BERT-DK_laptop/README.md

@@ -0,0 +1,43 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.  
+
+`BERT-DK_laptop` is trained from 100MB laptop corpus under `Electronics/Computers & Accessories/Laptops`. 
+
+
+## Model Description
+
+The original model is from `BERT-base-uncased` trained from Wikipedia+BookCorpus.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+`BERT-DK_laptop` is trained from 100MB laptop corpus under `Electronics/Computers & Accessories/Laptops`. 
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT-DK_laptop")
+model = AutoModel.from_pretrained("activebus/BERT-DK_laptop")
+
+```
+
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/activebus/BERT-DK_rest/README.md

@@ -0,0 +1,41 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.
+
+`BERT-DK_rest` is trained from 1G (19 types) restaurants from Yelp.  
+
+## Model Description
+
+The original model is from `BERT-base-uncased` trained from Wikipedia+BookCorpus.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT-DK_rest")
+model = AutoModel.from_pretrained("activebus/BERT-DK_rest")
+
+```
+
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/activebus/BERT-PT_laptop/README.md

@@ -0,0 +1,41 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.  
+
+`BERT-DK_laptop` is trained from 100MB laptop corpus under `Electronics/Computers & Accessories/Laptops`. 
+`BERT-PT_*` addtionally uses SQuAD 1.1.  
+
+## Model Description
+
+The original model is from `BERT-base-uncased` trained from Wikipedia+BookCorpus.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT-PT_laptop")
+model = AutoModel.from_pretrained("activebus/BERT-PT_laptop")
+
+```
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/activebus/BERT-PT_rest/README.md

@@ -0,0 +1,42 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.  
+
+`BERT-DK_rest` is trained from 1G (19 types) restaurants from Yelp.
+`BERT-PT_*` addtionally uses SQuAD 1.1.  
+
+## Model Description
+
+The original model is from `BERT-base-uncased` trained from Wikipedia+BookCorpus.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT-PT_rest")
+model = AutoModel.from_pretrained("activebus/BERT-PT_rest")
+
+```
+
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/activebus/BERT-XD_Review/README.md

@@ -0,0 +1,44 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.  
+Please visit https://github.com/howardhsu/BERT-for-RRC-ABSA for details.  
+
+`BERT-XD_Review` is a cross-domain (beyond just `laptop` and `restaurant`) language model, where each example is from a single product / restaurant with the same rating, post-trained (fine-tuned) on a combination of 5-core Amazon reviews and all Yelp data, expected to be 22 G in total. It is trained for 4 epochs on `bert-base-uncased`.
+The preprocessing code [here](https://github.com/howardhsu/BERT-for-RRC-ABSA/transformers).
+
+## Model Description
+
+The original model is from `BERT-base-uncased`.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT-XD_Review")
+model = AutoModel.from_pretrained("activebus/BERT-XD_Review")
+
+```
+
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+`BERT_Review` is expected to have similar performance on domain-specific tasks (such as aspect extraction) as `BERT-DK`, but much better on general tasks such as aspect sentiment classification (different domains mostly share similar sentiment words).
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/activebus/BERT_Review/README.md

@@ -0,0 +1,44 @@
+# ReviewBERT
+
+BERT (post-)trained from review corpus to understand sentiment, options and various e-commence aspects.  
+
+`BERT_Review` is cross-domain (beyond just `laptop` and `restaurant`) language model with one example from randomly mixed domains, post-trained (fine-tuned) on a combination of 5-core Amazon reviews and all Yelp data, expected to be 22 G in total. It is trained for 4 epochs on `bert-base-uncased`.
+The preprocessing code [here](https://github.com/howardhsu/BERT-for-RRC-ABSA/transformers).
+
+
+## Model Description
+
+The original model is from `BERT-base-uncased` trained from Wikipedia+BookCorpus.  
+Models are post-trained from [Amazon Dataset](http://jmcauley.ucsd.edu/data/amazon/) and [Yelp Dataset](https://www.yelp.com/dataset/challenge/).  
+
+
+## Instructions
+Loading the post-trained weights are as simple as, e.g., 
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("activebus/BERT_Review")
+model = AutoModel.from_pretrained("activebus/BERT_Review")
+
+```
+
+
+## Evaluation Results
+
+Check our [NAACL paper](https://www.aclweb.org/anthology/N19-1242.pdf) 
+`BERT_Review` is expected to have similar performance on domain-specific tasks (such as aspect extraction) as `BERT-DK`, but much better on general tasks such as aspect sentiment classification (different domains mostly share similar sentiment words).
+
+
+## Citation
+If you find this work useful, please cite as following.
+```
+@inproceedings{xu_bert2019,
+    title = "BERT Post-Training for Review Reading Comprehension and Aspect-based Sentiment Analysis",
+    author = "Xu, Hu and Liu, Bing and Shu, Lei and Yu, Philip S.",
+    booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics",
+    month = "jun",
+    year = "2019",
+}
+```

model_cards/allenai/longformer-base-4096-extra.pos.embd.only/README.md

@@ -0,0 +1,20 @@
+
+# longformer-base-4096-extra.pos.embd.only
+
+This model is similar to `longformer-base-4096` but it was pretrained to preserve RoBERTa weights by freezing all RoBERTa weights and only train the additional position embeddings. 
+
+
+### Citing
+
+If you use `Longformer` in your research, please cite [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150).
+```
+@article{Beltagy2020Longformer,
+  title={Longformer: The Long-Document Transformer},
+  author={Iz Beltagy and Matthew E. Peters and Arman Cohan},
+  journal={arXiv:2004.05150},
+  year={2020},
+}
+```
+
+`Longformer` is an open-source project developed by [the Allen Institute for Artificial Intelligence (AI2)](http://www.allenai.org).
+AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.

model_cards/allenai/longformer-base-4096/README.md

@@ -0,0 +1,24 @@
+
+# longformer-base-4096
+[Longformer](https://arxiv.org/abs/2004.05150) is a transformer model for long documents. 
+
+`longformer-base-4096` is a BERT-like model started from the RoBERTa checkpoint and pretrained for MLM on long documents. It supports sequences of length up to 4,096. 
+ 
+Longformer uses a combination of a sliding window (local) attention and global attention. Global attention is user-configured based on the task to allow the model to learn task-specific representations.
+Please refer to the examples in `modeling_longformer.py` and the paper for more details on how to set global attention.
+
+
+### Citing
+
+If you use `Longformer` in your research, please cite [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150).
+```
+@article{Beltagy2020Longformer,
+  title={Longformer: The Long-Document Transformer},
+  author={Iz Beltagy and Matthew E. Peters and Arman Cohan},
+  journal={arXiv:2004.05150},
+  year={2020},
+}
+```
+
+`Longformer` is an open-source project developed by [the Allen Institute for Artificial Intelligence (AI2)](http://www.allenai.org).
+AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.

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

@@ -6,6 +6,17 @@ language: arabic
 
 Pretrained BERT base language model for Arabic
 
+_If you use this model in your work, please cite this paper (to appear in 2020):_
+
+```
+@inproceedings{
+  title={KUISAIL at SemEval-2020 Task 12: BERT-CNN for Offensive Speech Identification in Social Media},
+  author={Safaya, Ali and Abdullatif, Moutasem and Yuret, Deniz},
+  booktitle={Proceedings of the International Workshop on Semantic Evaluation (SemEval)},
+  year={2020}
+}
+```
+
 ## Pretraining Corpus
 
 `arabic-bert-base` model was pretrained on ~8.2 Billion words:

model_cards/aubmindlab/bert-base-arabert/README.md

@@ -3,8 +3,9 @@ language: arabic
 ---
 
 # AraBERT : Pre-training BERT for Arabic Language Understanding
+<img src="https://github.com/aub-mind/arabert/blob/master/arabert_logo.png" width="100" align="left"/>  
 
-**AraBERT** is an Arabic pretrained lanaguage model based on [Google's BERT architechture](https://github.com/google-research/bert). AraBERT uses the same BERT-Base config.
+**AraBERT** is an Arabic pretrained lanaguage model based on [Google's BERT architechture](https://github.com/google-research/bert). AraBERT uses the same BERT-Base config. More details are available in the [AraBERT PAPER](https://arxiv.org/abs/2003.00104v2) and in the [AraBERT Meetup](https://github.com/WissamAntoun/pydata_khobar_meetup)
 
 There are two version off the model AraBERTv0.1 and AraBERTv1, with the difference being that AraBERTv1 uses pre-segmented text where prefixes and suffixes were splitted using the [Farasa Segmenter](http://alt.qcri.org/farasa/segmenter.html).
 
@@ -12,28 +13,34 @@ The model was trained on ~70M sentences or ~23GB of Arabic text with ~3B words.
 
 We evalaute both AraBERT models on different downstream tasks and compare it to [mBERT]((https://github.com/google-research/bert/blob/master/multilingual.md)), and other state of the art models (*To the extent of our knowledge*). The Tasks were Sentiment Analysis on 6 different datasets ([HARD](https://github.com/elnagara/HARD-Arabic-Dataset), [ASTD-Balanced](https://www.aclweb.org/anthology/D15-1299), [ArsenTD-Lev](https://staff.aub.edu.lb/~we07/Publications/ArSentD-LEV_Sentiment_Corpus.pdf), [LABR](https://github.com/mohamedadaly/LABR), [ArSaS](http://lrec-conf.org/workshops/lrec2018/W30/pdf/22_W30.pdf)), Named Entity Recognition with the [ANERcorp](http://curtis.ml.cmu.edu/w/courses/index.php/ANERcorp), and Arabic Question Answering on [Arabic-SQuAD and ARCD](https://github.com/husseinmozannar/SOQAL)
 
+**Update 2 (21/5/2020) :**
+Added support for the farasapy segmenter https://github.com/MagedSaeed/farasapy in the ``preprocess_arabert.py`` which is ~6x faster than the ``py4j.java_gateway``, consider setting ``use_farasapy=True`` when calling preprocess and pass it an instance of ``FarasaSegmenter(interactive=True)`` with interactive set to ``True`` for faster segmentation.
+
+**Update 1 (21/4/2020) :** 
+Fixed an issue with ARCD fine-tuning which drastically improved performance. Initially we didn't account for the change of the ```answer_start``` during preprocessing.
 ## Results (Acc.)
 Task | prev. SOTA | mBERT | AraBERTv0.1 | AraBERTv1
 ---|:---:|:---:|:---:|:---:
-HARD |95.7 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|95.7|96.2|96.1
-ASTD |86.5 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)| 80.1|92.2|92.6
-ArsenTD-Lev|52.4 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|51|58.9|59.4
-AJGT|93 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)| 83.6|94.1|93.8
-LABR|87.5 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)|83|85.9|86.7
-ANERcorp|81.7 (BiLSTM-CRF)|78.4|84.2|81.9
-ARCD|mBERT|EM:34.2 F1: 61.3|EM:30.1 F1:61.2|EM:30.6 F1: 62.7
+HARD |95.7 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|95.7|**96.2**|96.1
+ASTD |86.5 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)| 80.1|92.2|**92.6**
+ArsenTD-Lev|52.4 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|51|58.9|**59.4**
+AJGT|93 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)| 83.6|93.1|**93.8**
+LABR|**87.5** [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)|83|85.9|86.7
+ANERcorp|81.7 (BiLSTM-CRF)|78.4|**84.2**|81.9
+ARCD|mBERT|EM:34.2 F1: 61.3|EM:51.14 F1:82.13|**EM:54.84 F1: 82.15**
 
-*We would be extremly thankful if everyone can contibute to the Results table by adding more scores on different datasets*
+*If you tested AraBERT on a public dataset and you want to add your results to the table above, open a pull request or contact us. Also make sure to have your code available online so we can add it as a reference*
 
 ## How to use
 
-You can easily use AraBERT since it is almost fully compatible with existing codebases (You can use this repo instead of the official BERT one, the only difference is in the ```tokenization.py``` file where we modify the _is_punctuation function to make it compatible with the "+" symbol and the "[" and "]" characters)
-
-To use HuggingFace's Transformer repository you only need to provide a lost of token that forces the model to not split them, also make sure that the text is pre-segmented:
+You can easily use AraBERT since it is almost fully compatible with existing codebases (Use this repo instead of the official BERT one, the only difference is in the ```tokenization.py``` file where we modify the _is_punctuation function to make it compatible with the "+" symbol and the "[" and "]" characters)
 
+To use HuggingFace's Transformer repository you only need to provide a list of token that forces the model to not split them, also make sure that the text is pre-segmented:
+**Not all libraries built on top of transformers support the `never_split` argument**
 ```python
-from transformers import AutoTokenizer
-from preprocess_arabert import never_split_tokens
+from transformers import AutoTokenizer, AutoModel
+from arabert.preprocess_arabert import never_split_tokens, preprocess
+from farasa.segmenter import FarasaSegmenter
 
 arabert_tokenizer = AutoTokenizer.from_pretrained(
     "aubmindlab/bert-base-arabert",
@@ -42,27 +49,75 @@ arabert_tokenizer = AutoTokenizer.from_pretrained(
     never_split=never_split_tokens)
 arabert_model = AutoModel.from_pretrained("aubmindlab/bert-base-arabert")
 
-arabert_tokenizer.tokenize("و+ لن نبالغ إذا قل +نا إن هاتف أو كمبيوتر ال+ مكتب في زمن +نا هذا ضروري")
+#Preprocess the text to make it compatible with AraBERT using farasapy
+farasa_segmenter = FarasaSegmenter(interactive=True)
+
+#or you can use a py4j JavaGateway to the farasa Segmneter .jar but it's slower 
+#(see update 2)
+#from py4j.java_gateway import JavaGateway
+#gateway = JavaGateway.launch_gateway(classpath='./PATH_TO_FARASA/FarasaSegmenterJar.jar')
+#farasa = gateway.jvm.com.qcri.farasa.segmenter.Farasa()
+
+text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"
+text_preprocessed = preprocess( text,
+                                do_farasa_tokenization = True,
+                                farasa = farasa_segmenter,
+                                use_farasapy = True)
+
+>>>text_preprocessed: "و+ لن نبالغ إذا قل +نا إن هاتف أو كمبيوتر ال+ مكتب في زمن +نا هذا ضروري"
+
+arabert_tokenizer.tokenize(text_preprocessed)
 
 >>> ['و+', 'لن', 'نبال', '##غ', 'إذا', 'قل', '+نا', 'إن', 'هاتف', 'أو', 'كمبيوتر', 'ال+', 'مكتب', 'في', 'زمن', '+نا', 'هذا', 'ضروري']
 ```
 
 **AraBERTv0.1 is compatible with all existing libraries, since it needs no pre-segmentation.**
 ```python
-from transformers import AutoTokenizer
-from preprocess_arabert import never_split_tokens
+from transformers import AutoTokenizer, AutoModel
 
 arabert_tokenizer = AutoTokenizer.from_pretrained("aubmindlab/bert-base-arabertv01",do_lower_case=False)
 arabert_model = AutoModel.from_pretrained("aubmindlab/bert-base-arabertv01")
 
-arabert_tokenizer.tokenize("ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري")
+text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"
+arabert_tokenizer.tokenize(text)
 
 >>> ['ولن', 'ن', '##بالغ', 'إذا', 'قلنا', 'إن', 'هاتف', 'أو', 'كمبيوتر', 'المكتب', 'في', 'زمن', '##ن', '##ا', 'هذا', 'ضروري']
 ```
 
 
-The ```araBERT_(initial_Demo_TF)_.ipynb``` Notebook is a small demo using the AJGT dataset using TensorFlow (GPU and TPU compatible).
+The ```araBERT_(Updated_Demo_TF).ipynb``` Notebook is a small demo using the AJGT dataset using TensorFlow (GPU and TPU compatible).
 
+**Coming Soon :** Fine-tunning demo using HuggingFace's Trainer API
+
+**AraBERT on ARCD**
+During the preprocessing step the ```answer_start``` character position needs to be recalculated. You can use the file ```arcd_preprocessing.py``` as shown below to clean, preprocess the ARCD dataset before running ```run_squad.py```. More detailed Colab notebook is available in the [SOQAL repo](https://github.com/husseinmozannar/SOQAL).
+```bash
+python arcd_preprocessing.py \
+    --input_file="/PATH_TO/arcd-test.json" \
+    --output_file="arcd-test-pre.json" \
+    --do_farasa_tokenization=True \
+    --use_farasapy=True \
+```
+```bash
+python SOQAL/bert/run_squad.py \
+  --vocab_file="/PATH_TO_PRETRAINED_TF_CKPT/vocab.txt" \
+  --bert_config_file="/PATH_TO_PRETRAINED_TF_CKPT/config.json" \
+  --init_checkpoint="/PATH_TO_PRETRAINED_TF_CKPT/" \
+  --do_train=True \
+  --train_file=turk_combined_all_pre.json \
+  --do_predict=True \
+  --predict_file=arcd-test-pre.json \
+  --train_batch_size=32 \
+  --predict_batch_size=24 \
+  --learning_rate=3e-5 \
+  --num_train_epochs=4 \
+  --max_seq_length=384 \
+  --doc_stride=128 \
+  --do_lower_case=False\
+  --output_dir="/PATH_TO/OUTPUT_PATH"/ \
+  --use_tpu=True \
+  --tpu_name=$TPU_ADDRESS \
+```
 ## Model Weights and Vocab Download
 Models | AraBERTv0.1 | AraBERTv1
 ---|:---:|:---:
@@ -73,21 +128,17 @@ PyTorch| [Drive_Link](https://drive.google.com/open?id=1-_3te42mQCPD8SxwZ3l-VBL7
 
 ## If you used this model please cite us as:
 ```
-@misc{antoun2020arabert,
-    title={AraBERT: Transformer-based Model for Arabic Language Understanding},
-    author={Wissam Antoun and Fady Baly and Hazem Hajj},
-    year={2020},
-    eprint={2003.00104},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
+@inproceedings{antoun2020arabert,
+  title={AraBERT: Transformer-based Model for Arabic Language Understanding},
+  author={Antoun, Wissam and Baly, Fady and Hajj, Hazem},
+  booktitle={LREC 2020 Workshop Language Resources and Evaluation Conference 11--16 May 2020},
+  pages={9}
 }
 ```
 ## Acknowledgments 
-Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the [AUB MIND Lab](https://sites.aub.edu.lb/mindlab/) Members for the continous support. Also thanks to [Yakshof](https://www.yakshof.com/#/) and Assafir for data and storage access.
+Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the [AUB MIND Lab](https://sites.aub.edu.lb/mindlab/) Members for the continous support. Also thanks to [Yakshof](https://www.yakshof.com/#/) and Assafir for data and storage access. Another thanks for Habib Rahal (https://www.behance.net/rahalhabib), for putting a face to AraBERT.
 
 ## Contacts
 **Wissam Antoun**: [Linkedin](https://www.linkedin.com/in/giulio-ravasio-3a81a9110/) | [Twitter](https://twitter.com/wissam_antoun) | [Github](https://github.com/WissamAntoun) | <wfa07@mail.aub.edu> | <wissam.antoun@gmail.com>
 
-**Fady Baly**: [Linkedin](https://www.linkedin.com/in/fadybaly/) | [Twitter](https://twitter.com/BalyFady) | [Github](https://github.com/fadybaly) | <fgb06@mail.aub.edu> | <baly.fady@gmail.com>
-
-***We are looking for sponsors to train BERT-Large and other Transformer models, the sponsor only needs to cover to data storage and compute cost of the generating the pretraining data***
+**Fady Baly**: [Linkedin](https://www.linkedin.com/in/fadybaly/) | [Twitter](https://twitter.com/fadybaly) | [Github](https://github.com/fadybaly) | <fgb06@mail.aub.edu> | <baly.fady@gmail.com>

model_cards/aubmindlab/bert-base-arabertv01/README.md

@@ -3,8 +3,9 @@ language: arabic
 ---
 
 # AraBERT : Pre-training BERT for Arabic Language Understanding
+<img src="https://github.com/aub-mind/arabert/blob/master/arabert_logo.png" width="100" align="left"/>  
 
-**AraBERT** is an Arabic pretrained lanaguage model based on [Google's BERT architechture](https://github.com/google-research/bert). AraBERT uses the same BERT-Base config.
+**AraBERT** is an Arabic pretrained lanaguage model based on [Google's BERT architechture](https://github.com/google-research/bert). AraBERT uses the same BERT-Base config. More details are available in the [AraBERT PAPER](https://arxiv.org/abs/2003.00104v2) and in the [AraBERT Meetup](https://github.com/WissamAntoun/pydata_khobar_meetup)
 
 There are two version off the model AraBERTv0.1 and AraBERTv1, with the difference being that AraBERTv1 uses pre-segmented text where prefixes and suffixes were splitted using the [Farasa Segmenter](http://alt.qcri.org/farasa/segmenter.html).
 
@@ -12,28 +13,34 @@ The model was trained on ~70M sentences or ~23GB of Arabic text with ~3B words.
 
 We evalaute both AraBERT models on different downstream tasks and compare it to [mBERT]((https://github.com/google-research/bert/blob/master/multilingual.md)), and other state of the art models (*To the extent of our knowledge*). The Tasks were Sentiment Analysis on 6 different datasets ([HARD](https://github.com/elnagara/HARD-Arabic-Dataset), [ASTD-Balanced](https://www.aclweb.org/anthology/D15-1299), [ArsenTD-Lev](https://staff.aub.edu.lb/~we07/Publications/ArSentD-LEV_Sentiment_Corpus.pdf), [LABR](https://github.com/mohamedadaly/LABR), [ArSaS](http://lrec-conf.org/workshops/lrec2018/W30/pdf/22_W30.pdf)), Named Entity Recognition with the [ANERcorp](http://curtis.ml.cmu.edu/w/courses/index.php/ANERcorp), and Arabic Question Answering on [Arabic-SQuAD and ARCD](https://github.com/husseinmozannar/SOQAL)
 
+**Update 2 (21/5/2020) :**
+Added support for the farasapy segmenter https://github.com/MagedSaeed/farasapy in the ``preprocess_arabert.py`` which is ~6x faster than the ``py4j.java_gateway``, consider setting ``use_farasapy=True`` when calling preprocess and pass it an instance of ``FarasaSegmenter(interactive=True)`` with interactive set to ``True`` for faster segmentation.
+
+**Update 1 (21/4/2020) :** 
+Fixed an issue with ARCD fine-tuning which drastically improved performance. Initially we didn't account for the change of the ```answer_start``` during preprocessing.
 ## Results (Acc.)
 Task | prev. SOTA | mBERT | AraBERTv0.1 | AraBERTv1
 ---|:---:|:---:|:---:|:---:
-HARD |95.7 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|95.7|96.2|96.1
-ASTD |86.5 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)| 80.1|92.2|92.6
-ArsenTD-Lev|52.4 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|51|58.9|59.4
-AJGT|93 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)| 83.6|94.1|93.8
-LABR|87.5 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)|83|85.9|86.7
-ANERcorp|81.7 (BiLSTM-CRF)|78.4|84.2|81.9
-ARCD|mBERT|EM:34.2 F1: 61.3|EM:30.1 F1:61.2|EM:30.6 F1: 62.7
+HARD |95.7 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|95.7|**96.2**|96.1
+ASTD |86.5 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)| 80.1|92.2|**92.6**
+ArsenTD-Lev|52.4 [ElJundi et.al.](https://www.aclweb.org/anthology/W19-4608/)|51|58.9|**59.4**
+AJGT|93 [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)| 83.6|93.1|**93.8**
+LABR|**87.5** [Dahou et.al.](https://dl.acm.org/doi/fullHtml/10.1145/3314941)|83|85.9|86.7
+ANERcorp|81.7 (BiLSTM-CRF)|78.4|**84.2**|81.9
+ARCD|mBERT|EM:34.2 F1: 61.3|EM:51.14 F1:82.13|**EM:54.84 F1: 82.15**
 
-*We would be extremly thankful if everyone can contibute to the Results table by adding more scores on different datasets*
+*If you tested AraBERT on a public dataset and you want to add your results to the table above, open a pull request or contact us. Also make sure to have your code available online so we can add it as a reference*
 
 ## How to use
 
-You can easily use AraBERT since it is almost fully compatible with existing codebases (You can use this repo instead of the official BERT one, the only difference is in the ```tokenization.py``` file where we modify the _is_punctuation function to make it compatible with the "+" symbol and the "[" and "]" characters)
-
-To use HuggingFace's Transformer repository you only need to provide a lost of token that forces the model to not split them, also make sure that the text is pre-segmented:
+You can easily use AraBERT since it is almost fully compatible with existing codebases (Use this repo instead of the official BERT one, the only difference is in the ```tokenization.py``` file where we modify the _is_punctuation function to make it compatible with the "+" symbol and the "[" and "]" characters)
 
+To use HuggingFace's Transformer repository you only need to provide a list of token that forces the model to not split them, also make sure that the text is pre-segmented:
+**Not all libraries built on top of transformers support the `never_split` argument**
 ```python
-from transformers import AutoTokenizer
-from preprocess_arabert import never_split_tokens
+from transformers import AutoTokenizer, AutoModel
+from arabert.preprocess_arabert import never_split_tokens, preprocess
+from farasa.segmenter import FarasaSegmenter
 
 arabert_tokenizer = AutoTokenizer.from_pretrained(
     "aubmindlab/bert-base-arabert",
@@ -42,27 +49,75 @@ arabert_tokenizer = AutoTokenizer.from_pretrained(
     never_split=never_split_tokens)
 arabert_model = AutoModel.from_pretrained("aubmindlab/bert-base-arabert")
 
-arabert_tokenizer.tokenize("و+ لن نبالغ إذا قل +نا إن هاتف أو كمبيوتر ال+ مكتب في زمن +نا هذا ضروري")
+#Preprocess the text to make it compatible with AraBERT using farasapy
+farasa_segmenter = FarasaSegmenter(interactive=True)
+
+#or you can use a py4j JavaGateway to the farasa Segmneter .jar but it's slower 
+#(see update 2)
+#from py4j.java_gateway import JavaGateway
+#gateway = JavaGateway.launch_gateway(classpath='./PATH_TO_FARASA/FarasaSegmenterJar.jar')
+#farasa = gateway.jvm.com.qcri.farasa.segmenter.Farasa()
+
+text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"
+text_preprocessed = preprocess( text,
+                                do_farasa_tokenization = True,
+                                farasa = farasa_segmenter,
+                                use_farasapy = True)
+
+>>>text_preprocessed: "و+ لن نبالغ إذا قل +نا إن هاتف أو كمبيوتر ال+ مكتب في زمن +نا هذا ضروري"
+
+arabert_tokenizer.tokenize(text_preprocessed)
 
 >>> ['و+', 'لن', 'نبال', '##غ', 'إذا', 'قل', '+نا', 'إن', 'هاتف', 'أو', 'كمبيوتر', 'ال+', 'مكتب', 'في', 'زمن', '+نا', 'هذا', 'ضروري']
 ```
 
 **AraBERTv0.1 is compatible with all existing libraries, since it needs no pre-segmentation.**
 ```python
-from transformers import AutoTokenizer
-from preprocess_arabert import never_split_tokens
+from transformers import AutoTokenizer, AutoModel
 
 arabert_tokenizer = AutoTokenizer.from_pretrained("aubmindlab/bert-base-arabertv01",do_lower_case=False)
 arabert_model = AutoModel.from_pretrained("aubmindlab/bert-base-arabertv01")
 
-arabert_tokenizer.tokenize("ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري")
+text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"
+arabert_tokenizer.tokenize(text)
 
 >>> ['ولن', 'ن', '##بالغ', 'إذا', 'قلنا', 'إن', 'هاتف', 'أو', 'كمبيوتر', 'المكتب', 'في', 'زمن', '##ن', '##ا', 'هذا', 'ضروري']
 ```
 
 
-The ```araBERT_(initial_Demo_TF)_.ipynb``` Notebook is a small demo using the AJGT dataset using TensorFlow (GPU and TPU compatible).
+The ```araBERT_(Updated_Demo_TF).ipynb``` Notebook is a small demo using the AJGT dataset using TensorFlow (GPU and TPU compatible).
 
+**Coming Soon :** Fine-tunning demo using HuggingFace's Trainer API
+
+**AraBERT on ARCD**
+During the preprocessing step the ```answer_start``` character position needs to be recalculated. You can use the file ```arcd_preprocessing.py``` as shown below to clean, preprocess the ARCD dataset before running ```run_squad.py```. More detailed Colab notebook is available in the [SOQAL repo](https://github.com/husseinmozannar/SOQAL).
+```bash
+python arcd_preprocessing.py \
+    --input_file="/PATH_TO/arcd-test.json" \
+    --output_file="arcd-test-pre.json" \
+    --do_farasa_tokenization=True \
+    --use_farasapy=True \
+```
+```bash
+python SOQAL/bert/run_squad.py \
+  --vocab_file="/PATH_TO_PRETRAINED_TF_CKPT/vocab.txt" \
+  --bert_config_file="/PATH_TO_PRETRAINED_TF_CKPT/config.json" \
+  --init_checkpoint="/PATH_TO_PRETRAINED_TF_CKPT/" \
+  --do_train=True \
+  --train_file=turk_combined_all_pre.json \
+  --do_predict=True \
+  --predict_file=arcd-test-pre.json \
+  --train_batch_size=32 \
+  --predict_batch_size=24 \
+  --learning_rate=3e-5 \
+  --num_train_epochs=4 \
+  --max_seq_length=384 \
+  --doc_stride=128 \
+  --do_lower_case=False\
+  --output_dir="/PATH_TO/OUTPUT_PATH"/ \
+  --use_tpu=True \
+  --tpu_name=$TPU_ADDRESS \
+```
 ## Model Weights and Vocab Download
 Models | AraBERTv0.1 | AraBERTv1
 ---|:---:|:---:
@@ -73,21 +128,17 @@ PyTorch| [Drive_Link](https://drive.google.com/open?id=1-_3te42mQCPD8SxwZ3l-VBL7
 
 ## If you used this model please cite us as:
 ```
-@misc{antoun2020arabert,
-    title={AraBERT: Transformer-based Model for Arabic Language Understanding},
-    author={Wissam Antoun and Fady Baly and Hazem Hajj},
-    year={2020},
-    eprint={2003.00104},
-    archivePrefix={arXiv},
-    primaryClass={cs.CL}
+@inproceedings{antoun2020arabert,
+  title={AraBERT: Transformer-based Model for Arabic Language Understanding},
+  author={Antoun, Wissam and Baly, Fady and Hajj, Hazem},
+  booktitle={LREC 2020 Workshop Language Resources and Evaluation Conference 11--16 May 2020},
+  pages={9}
 }
 ```
 ## Acknowledgments 
-Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the [AUB MIND Lab](https://sites.aub.edu.lb/mindlab/) Members for the continous support. Also thanks to [Yakshof](https://www.yakshof.com/#/) and Assafir for data and storage access.
+Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the [AUB MIND Lab](https://sites.aub.edu.lb/mindlab/) Members for the continous support. Also thanks to [Yakshof](https://www.yakshof.com/#/) and Assafir for data and storage access. Another thanks for Habib Rahal (https://www.behance.net/rahalhabib), for putting a face to AraBERT.
 
 ## Contacts
 **Wissam Antoun**: [Linkedin](https://www.linkedin.com/in/giulio-ravasio-3a81a9110/) | [Twitter](https://twitter.com/wissam_antoun) | [Github](https://github.com/WissamAntoun) | <wfa07@mail.aub.edu> | <wissam.antoun@gmail.com>
 
-**Fady Baly**: [Linkedin](https://www.linkedin.com/in/fadybaly/) | [Twitter](https://twitter.com/BalyFady) | [Github](https://github.com/fadybaly) | <fgb06@mail.aub.edu> | <baly.fady@gmail.com>
-
-***We are looking for sponsors to train BERT-Large and other Transformer models, the sponsor only needs to cover to data storage and compute cost of the generating the pretraining data***
+**Fady Baly**: [Linkedin](https://www.linkedin.com/in/fadybaly/) | [Twitter](https://twitter.com/fadybaly) | [Github](https://github.com/fadybaly) | <fgb06@mail.aub.edu> | <baly.fady@gmail.com>

model_cards/bayartsogt/albert-mongolian/README.md

@@ -0,0 +1,55 @@
+# ALBERT-Mongolian
+[pretraining repo link](https://github.com/bayartsogt-ya/albert-mongolian)
+## Model description
+Here we provide pretrained ALBERT model and trained SentencePiece model for Mongolia text. Training data is the Mongolian wikipedia corpus from Wikipedia Downloads and Mongolian News corpus.
+
+## Evaluation Result:
+```
+loss = 1.7478163
+masked_lm_accuracy = 0.6838185
+masked_lm_loss = 1.6687671
+sentence_order_accuracy = 0.998125
+sentence_order_loss = 0.007942731
+```
+
+## Fine-tuning Result on Eduge Dataset:
+```
+                precision    recall  f1-score   support
+
+  байгал орчин       0.83      0.76      0.80       483
+     боловсрол       0.79      0.75      0.77       420
+         спорт       0.98      0.96      0.97      1391
+     технологи       0.85      0.83      0.84       543
+       улс төр       0.88      0.87      0.87      1336
+    урлаг соёл       0.89      0.94      0.91       726
+         хууль       0.87      0.83      0.85       840
+   эдийн засаг       0.80      0.84      0.82      1265
+    эрүүл мэнд       0.84      0.90      0.87       562
+
+      accuracy                           0.87      7566
+     macro avg       0.86      0.85      0.86      7566
+  weighted avg       0.87      0.87      0.87      7566
+```
+
+## Reference
+1. [ALBERT - official repo](https://github.com/google-research/albert)
+2. [WikiExtrator](https://github.com/attardi/wikiextractor)
+3. [Mongolian BERT](https://github.com/tugstugi/mongolian-bert)
+4. [ALBERT - Japanese](https://github.com/alinear-corp/albert-japanese)
+5. [Mongolian Text Classification](https://github.com/sharavsambuu/mongolian-text-classification)
+6. [You's paper](https://arxiv.org/abs/1904.00962)
+
+## Citation
+```
+@misc{albert-mongolian,
+  author = {Bayartsogt Yadamsuren},
+  title = {ALBERT Pretrained Model on Mongolian Datasets},
+  year = {2020},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/bayartsogt-ya/albert-mongolian/}}
+}
+```
+
+## For More Information
+Please contact by bayartsogtyadamsuren@icloud.com

model_cards/bert-base-german-cased-README.md

@@ -18,13 +18,16 @@ tags:
 **Eval data:** Conll03 (NER), GermEval14 (NER), GermEval18 (Classification), GNAD (Classification)  
 **Infrastructure**: 1x TPU v2  
 **Published**: Jun 14th, 2019
+
+**Update April 3rd, 2020**: we updated the vocabulary file on deepset's s3 to conform with the default tokenization of punctuation tokens. 
+For details see the related [FARM issue](https://github.com/deepset-ai/FARM/issues/60). If you want to use the old vocab we have also uploaded a ["deepset/bert-base-german-cased-oldvocab"](https://huggingface.co/deepset/bert-base-german-cased-oldvocab) model.
  
 ## Details
 - We trained using Google's Tensorflow code on a single cloud TPU v2 with standard settings.
 - We trained 810k steps with a batch size of 1024 for sequence length 128 and 30k steps with sequence length 512. Training took about 9 days.
 - As training data we used the latest German Wikipedia dump (6GB of raw txt files), the OpenLegalData dump (2.4 GB) and news articles (3.6 GB).
 - We cleaned the data dumps with tailored scripts and segmented sentences with spacy v2.1. To create tensorflow records we used the recommended sentencepiece library for creating the word piece vocabulary and tensorflow scripts to convert the text to data usable by BERT.
-- Update April 3rd, 2020: updated the vocab file on deepset s3 to adjust tokenization of punctuation.
+
 
 See https://deepset.ai/german-bert for more details
 

model_cards/deepset/bert-base-german-cased-oldvocab/README.md

@@ -0,0 +1,28 @@
+---
+language: german
+thumbnail: https://static.tildacdn.com/tild6438-3730-4164-b266-613634323466/german_bert.png
+tags:
+- exbert
+---
+
+<a href="https://huggingface.co/exbert/?model=bert-base-german-cased">
+	<img width="300px" src="https://hf-dinosaur.huggingface.co/exbert/button.png">
+</a>
+
+# German BERT with old vocabulary
+For details see the related [FARM issue](https://github.com/deepset-ai/FARM/issues/60).
+
+
+## About us
+![deepset logo](https://raw.githubusercontent.com/deepset-ai/FARM/master/docs/img/deepset_logo.png)
+
+We bring NLP to the industry via open source!  
+Our focus: Industry specific language models & large scale QA systems.  
+  
+Some of our work: 
+- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
+- [FARM](https://github.com/deepset-ai/FARM)
+- [Haystack](https://github.com/deepset-ai/haystack/)
+
+Get in touch:
+[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Website](https://deepset.ai)  

model_cards/dkleczek/bert-base-polish-cased-v1/README.md

@@ -0,0 +1,135 @@
+---
+language: polish
+thumbnail: https://raw.githubusercontent.com/kldarek/polbert/master/img/polbert.png
+---
+
+# Polbert - Polish BERT
+Polish version of BERT language model is here! It is now available in two variants: cased and uncased, both can be downloaded and used via HuggingFace transformers library. I recommend using the cased model, more info on the differences and benchmark results below. 
+
+![PolBERT image](https://raw.githubusercontent.com/kldarek/polbert/master/img/polbert.png)
+
+## Cased and uncased variants
+
+* I initially trained the uncased model, the corpus and training details are referenced below. Here are some issues I found after I published the uncased model:
+    * Some Polish characters and accents are not tokenized correctly through the BERT tokenizer when applying lowercase. This doesn't impact sequence classification much, but may influence token classfication tasks significantly.
+    * I noticed a lot of duplicates in the Open Subtitles dataset, which dominates the training corpus.
+    * I didn't use Whole Word Masking. 
+* The cased model improves on the uncased model in the following ways:
+    * All Polish characters and accents should now be tokenized correctly. 
+    * I removed duplicates from Open Subtitles dataset. The corpus is smaller, but more balanced now. 
+    * The model is trained with Whole Word Masking. 
+
+## Pre-training corpora
+
+Below is the list of corpora used along with the output of `wc` command (counting lines, words and characters). These corpora were divided into sentences with srxsegmenter (see references), concatenated and tokenized with HuggingFace BERT Tokenizer. 
+
+### Uncased
+
+| Tables        | Lines           | Words  | Characters  |
+| ------------- |--------------:| -----:| -----:|
+| [Polish subset of Open Subtitles](http://opus.nlpl.eu/OpenSubtitles-v2018.php)      | 236635408| 1431199601 | 7628097730 |
+| [Polish subset of ParaCrawl](http://opus.nlpl.eu/ParaCrawl.php)     | 8470950      |   176670885 | 1163505275 |
+| [Polish Parliamentary Corpus](http://clip.ipipan.waw.pl/PPC) | 9799859      |    121154785 | 938896963 |
+| [Polish Wikipedia - Feb 2020](https://dumps.wikimedia.org/plwiki/latest/plwiki-latest-pages-articles.xml.bz2) | 8014206      |    132067986 | 1015849191 |
+| Total | 262920423      |    1861093257 | 10746349159 |
+
+### Cased
+
+| Tables        | Lines           | Words  | Characters  |
+| ------------- |--------------:| -----:| -----:|
+| [Polish subset of Open Subtitles (Deduplicated) ](http://opus.nlpl.eu/OpenSubtitles-v2018.php)      | 41998942| 213590656 | 1424873235 |
+| [Polish subset of ParaCrawl](http://opus.nlpl.eu/ParaCrawl.php)     | 8470950      |   176670885 | 1163505275 |
+| [Polish Parliamentary Corpus](http://clip.ipipan.waw.pl/PPC) | 9799859      |    121154785 | 938896963 |
+| [Polish Wikipedia - Feb 2020](https://dumps.wikimedia.org/plwiki/latest/plwiki-latest-pages-articles.xml.bz2) | 8014206      |    132067986 | 1015849191 |
+| Total | 68283960      |    646479197 | 4543124667 |
+
+
+## Pre-training details
+
+### Uncased 
+
+* Polbert was trained with code provided in Google BERT's github repository (https://github.com/google-research/bert)
+* Currently released model follows bert-base-uncased model architecture (12-layer, 768-hidden, 12-heads, 110M parameters)
+* Training set-up: in total 1 million training steps: 
+    * 100.000 steps - 128 sequence length, batch size 512, learning rate 1e-4 (10.000 steps warmup)
+    * 800.000 steps - 128 sequence length, batch size 512, learning rate 5e-5
+    * 100.000 steps - 512 sequence length, batch size 256, learning rate 2e-5
+* The model was trained on a single Google Cloud TPU v3-8 
+
+### Cased
+
+* Same approach as uncased model, with the following differences:
+    * Whole Word Masking
+* Training set-up:
+    * 100.000 steps - 128 sequence length, batch size 2048, learning rate 1e-4 (10.000 steps warmup)
+    * 100.000 steps - 128 sequence length, batch size 2048, learning rate 5e-5
+    * 100.000 steps - 512 sequence length, batch size 256, learning rate 2e-5
+
+
+## Usage
+Polbert is released via [HuggingFace Transformers library](https://huggingface.co/transformers/).
+
+For an example use as language model, see [this notebook](/LM_testing.ipynb) file. 
+
+### Uncased
+
+```python
+from transformers import *
+model = BertForMaskedLM.from_pretrained("dkleczek/bert-base-polish-uncased-v1")
+tokenizer = BertTokenizer.from_pretrained("dkleczek/bert-base-polish-uncased-v1")
+nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
+for pred in nlp(f"Adam Mickiewicz wielkim polskim {nlp.tokenizer.mask_token} był."):
+  print(pred)
+# Output:
+# {'sequence': '[CLS] adam mickiewicz wielkim polskim poeta był. [SEP]', 'score': 0.47196975350379944, 'token': 26596}
+# {'sequence': '[CLS] adam mickiewicz wielkim polskim bohaterem był. [SEP]', 'score': 0.09127858281135559, 'token': 10953}
+# {'sequence': '[CLS] adam mickiewicz wielkim polskim człowiekiem był. [SEP]', 'score': 0.0647173821926117, 'token': 5182}
+# {'sequence': '[CLS] adam mickiewicz wielkim polskim pisarzem był. [SEP]', 'score': 0.05232388526201248, 'token': 24293}
+# {'sequence': '[CLS] adam mickiewicz wielkim polskim politykiem był. [SEP]', 'score': 0.04554257541894913, 'token': 44095}
+```
+
+### Cased
+
+```python
+model = BertForMaskedLM.from_pretrained("dkleczek/bert-base-polish-cased-v1")
+tokenizer = BertTokenizer.from_pretrained("dkleczek/bert-base-polish-cased-v1")
+nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
+for pred in nlp(f"Adam Mickiewicz wielkim polskim {nlp.tokenizer.mask_token} był."):
+  print(pred)
+# Output:
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim pisarzem był. [SEP]', 'score': 0.5391148328781128, 'token': 37120}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim człowiekiem był. [SEP]', 'score': 0.11683262139558792, 'token': 6810}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim bohaterem był. [SEP]', 'score': 0.06021466106176376, 'token': 17709}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim mistrzem był. [SEP]', 'score': 0.051870670169591904, 'token': 14652}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim artystą był. [SEP]', 'score': 0.031787533313035965, 'token': 35680}
+```
+
+See the next section for an example usage of Polbert in downstream tasks. 
+
+## Evaluation
+Thanks to Allegro, we now have the [KLEJ benchmark](https://klejbenchmark.com/leaderboard/), a set of nine evaluation tasks for the Polish language understanding. The following results are achieved by running standard set of evaluation scripts (no tricks!) utilizing both cased and uncased variants of Polbert.
+
+| Model	| Average |	NKJP-NER | CDSC-E |	CDSC-R |	CBD	| PolEmo2.0-IN |	PolEmo2.0-OUT |	DYK	| PSC |	AR |
+| ------------- |--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|
+| Polbert cased | 81.7 |	93.6 |	93.4 |	93.8 |	52.7 |	87.4 |	71.1 |	59.1 |	98.6 |	85.2 |
+| Polbert uncased | 81.4 |	90.1 |	93.9 |	93.5 |	55.0 |	88.1 |	68.8 |	59.4 |	98.8 |	85.4 |
+
+Note how the uncased model performs better than cased on some tasks? My guess this is because of the oversampling of Open Subtitles dataset and its similarity to data in some of these tasks. All these benchmark tasks are sequence classification, so the relative strength of the cased model is not so visible here. 
+
+## Bias
+The data used to train the model is biased. It may reflect stereotypes related to gender, ethnicity etc. Please be careful when using the model for downstream task to consider these biases and mitigate them.  
+
+## Acknowledgements
+* I'd like to express my gratitude to Google [TensorFlow Research Cloud (TFRC)](https://www.tensorflow.org/tfrc) for providing the free TPU credits - thank you!
+* Also appreciate the help from Timo Möller from [deepset](https://deepset.ai) for sharing tips and scripts based on their experience training German BERT model.
+* Big thanks to Allegro for releasing KLEJ Benchmark and specifically to Piotr Rybak for help with the evaluation and pointing out some issues with the tokenization. 
+* Finally, thanks to Rachel Thomas, Jeremy Howard and Sylvain Gugger from [fastai](https://www.fast.ai) for their NLP and Deep Learning courses! 
+
+## Author
+Darek Kłeczek - contact me on Twitter [@dk21](https://twitter.com/dk21)
+
+## References
+* https://github.com/google-research/bert
+* https://github.com/narusemotoki/srx_segmenter
+* SRX rules file for sentence splitting in Polish, written by Marcin Miłkowski: https://raw.githubusercontent.com/languagetool-org/languagetool/master/languagetool-core/src/main/resources/org/languagetool/resource/segment.srx
+* [KLEJ benchmark](https://klejbenchmark.com/leaderboard/)

model_cards/dkleczek/bert-base-polish-uncased-v1/README.md

@@ -4,14 +4,27 @@ thumbnail: https://raw.githubusercontent.com/kldarek/polbert/master/img/polbert.
 ---
 
 # Polbert - Polish BERT
-Polish version of BERT language model is here! While this is still work in progress, I'm happy to share the first model, similar to BERT-Base and trained on a large Polish corpus. If you'd like to contribute to this project, please reach out to me!
+Polish version of BERT language model is here! It is now available in two variants: cased and uncased, both can be downloaded and used via HuggingFace transformers library. I recommend using the cased model, more info on the differences and benchmark results below. 
 
 ![PolBERT image](https://raw.githubusercontent.com/kldarek/polbert/master/img/polbert.png)
 
+## Cased and uncased variants
+
+* I initially trained the uncased model, the corpus and training details are referenced below. Here are some issues I found after I published the uncased model:
+    * Some Polish characters and accents are not tokenized correctly through the BERT tokenizer when applying lowercase. This doesn't impact sequence classification much, but may influence token classfication tasks significantly.
+    * I noticed a lot of duplicates in the Open Subtitles dataset, which dominates the training corpus.
+    * I didn't use Whole Word Masking. 
+* The cased model improves on the uncased model in the following ways:
+    * All Polish characters and accents should now be tokenized correctly. 
+    * I removed duplicates from Open Subtitles dataset. The corpus is smaller, but more balanced now. 
+    * The model is trained with Whole Word Masking. 
+
 ## Pre-training corpora
 
 Below is the list of corpora used along with the output of `wc` command (counting lines, words and characters). These corpora were divided into sentences with srxsegmenter (see references), concatenated and tokenized with HuggingFace BERT Tokenizer. 
 
+### Uncased
+
 | Tables        | Lines           | Words  | Characters  |
 | ------------- |--------------:| -----:| -----:|
 | [Polish subset of Open Subtitles](http://opus.nlpl.eu/OpenSubtitles-v2018.php)      | 236635408| 1431199601 | 7628097730 |
@@ -20,7 +33,21 @@ Below is the list of corpora used along with the output of `wc` command (countin
 | [Polish Wikipedia - Feb 2020](https://dumps.wikimedia.org/plwiki/latest/plwiki-latest-pages-articles.xml.bz2) | 8014206      |    132067986 | 1015849191 |
 | Total | 262920423      |    1861093257 | 10746349159 |
 
+### Cased
+
+| Tables        | Lines           | Words  | Characters  |
+| ------------- |--------------:| -----:| -----:|
+| [Polish subset of Open Subtitles (Deduplicated) ](http://opus.nlpl.eu/OpenSubtitles-v2018.php)      | 41998942| 213590656 | 1424873235 |
+| [Polish subset of ParaCrawl](http://opus.nlpl.eu/ParaCrawl.php)     | 8470950      |   176670885 | 1163505275 |
+| [Polish Parliamentary Corpus](http://clip.ipipan.waw.pl/PPC) | 9799859      |    121154785 | 938896963 |
+| [Polish Wikipedia - Feb 2020](https://dumps.wikimedia.org/plwiki/latest/plwiki-latest-pages-articles.xml.bz2) | 8014206      |    132067986 | 1015849191 |
+| Total | 68283960      |    646479197 | 4543124667 |
+
+
 ## Pre-training details
+
+### Uncased 
+
 * Polbert was trained with code provided in Google BERT's github repository (https://github.com/google-research/bert)
 * Currently released model follows bert-base-uncased model architecture (12-layer, 768-hidden, 12-heads, 110M parameters)
 * Training set-up: in total 1 million training steps: 
@@ -29,10 +56,22 @@ Below is the list of corpora used along with the output of `wc` command (countin
     * 100.000 steps - 512 sequence length, batch size 256, learning rate 2e-5
 * The model was trained on a single Google Cloud TPU v3-8 
 
+### Cased
+
+* Same approach as uncased model, with the following differences:
+    * Whole Word Masking
+* Training set-up:
+    * 100.000 steps - 128 sequence length, batch size 2048, learning rate 1e-4 (10.000 steps warmup)
+    * 100.000 steps - 128 sequence length, batch size 2048, learning rate 5e-5
+    * 100.000 steps - 512 sequence length, batch size 256, learning rate 2e-5
+
+
 ## Usage
 Polbert is released via [HuggingFace Transformers library](https://huggingface.co/transformers/).
 
-For an example use as language model, see [this notebook](https://github.com/kldarek/polbert/blob/master/LM_testing.ipynb) file. 
+For an example use as language model, see [this notebook](/LM_testing.ipynb) file. 
+
+### Uncased
 
 ```python
 from transformers import *
@@ -41,7 +80,6 @@ tokenizer = BertTokenizer.from_pretrained("dkleczek/bert-base-polish-uncased-v1"
 nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
 for pred in nlp(f"Adam Mickiewicz wielkim polskim {nlp.tokenizer.mask_token} był."):
   print(pred)
-
 # Output:
 # {'sequence': '[CLS] adam mickiewicz wielkim polskim poeta był. [SEP]', 'score': 0.47196975350379944, 'token': 26596}
 # {'sequence': '[CLS] adam mickiewicz wielkim polskim bohaterem był. [SEP]', 'score': 0.09127858281135559, 'token': 10953}
@@ -50,23 +88,42 @@ for pred in nlp(f"Adam Mickiewicz wielkim polskim {nlp.tokenizer.mask_token} by
 # {'sequence': '[CLS] adam mickiewicz wielkim polskim politykiem był. [SEP]', 'score': 0.04554257541894913, 'token': 44095}
 ```
 
+### Cased
+
+```python
+model = BertForMaskedLM.from_pretrained("dkleczek/bert-base-polish-cased-v1")
+tokenizer = BertTokenizer.from_pretrained("dkleczek/bert-base-polish-cased-v1")
+nlp = pipeline('fill-mask', model=model, tokenizer=tokenizer)
+for pred in nlp(f"Adam Mickiewicz wielkim polskim {nlp.tokenizer.mask_token} był."):
+  print(pred)
+# Output:
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim pisarzem był. [SEP]', 'score': 0.5391148328781128, 'token': 37120}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim człowiekiem był. [SEP]', 'score': 0.11683262139558792, 'token': 6810}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim bohaterem był. [SEP]', 'score': 0.06021466106176376, 'token': 17709}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim mistrzem był. [SEP]', 'score': 0.051870670169591904, 'token': 14652}
+# {'sequence': '[CLS] Adam Mickiewicz wielkim polskim artystą był. [SEP]', 'score': 0.031787533313035965, 'token': 35680}
+```
+
 See the next section for an example usage of Polbert in downstream tasks. 
 
 ## Evaluation
-I'd love to get some help from the Polish NLP community here! If you feel like evaluating Polbert on some benchmark tasks, it would be great if you can share the results. 
+Thanks to Allegro, we now have the [KLEJ benchmark](https://klejbenchmark.com/leaderboard/), a set of nine evaluation tasks for the Polish language understanding. The following results are achieved by running standard set of evaluation scripts (no tricks!) utilizing both cased and uncased variants of Polbert.
 
-So far, I've compared the performance of Polbert vs Multilingual BERT on PolEmo 2.0 sentiment classification, here are the results. These results are are produced with a linear classification layer on top of pooled output, trained for 10 epochs with learning rate 3e-5. The checkpoint with the lowest loss on validation set is evaluated on the test set. 
+| Model	| Average |	NKJP-NER | CDSC-E |	CDSC-R |	CBD	| PolEmo2.0-IN |	PolEmo2.0-OUT |	DYK	| PSC |	AR |
+| ------------- |--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|--------------:|
+| Polbert cased | 81.7 |	93.6 |	93.4 |	93.8 |	52.7 |	87.4 |	71.1 |	59.1 |	98.6 |	85.2 |
+| Polbert uncased | 81.4 |	90.1 |	93.9 |	93.5 |	55.0 |	88.1 |	68.8 |	59.4 |	98.8 |	85.4 |
 
-| PolEmo 2.0 Sentiment Classifcation | Test Accuracy | 
-| ------------- |--------------:|
-| Multilingual BERT | 0.78 |
-| Polbert | 0.85 |
+Note how the uncased model performs better than cased on some tasks? My guess this is because of the oversampling of Open Subtitles dataset and its similarity to data in some of these tasks. All these benchmark tasks are sequence classification, so the relative strength of the cased model is not so visible here. 
 
 ## Bias
 The data used to train the model is biased. It may reflect stereotypes related to gender, ethnicity etc. Please be careful when using the model for downstream task to consider these biases and mitigate them.  
 
 ## Acknowledgements
-I'd like to express my gratitude to Google [TensorFlow Research Cloud (TFRC)](https://www.tensorflow.org/tfrc) for providing the free TPU credits - thank you! Also appreciate the help from Timo Möller from [deepset](https://deepset.ai) for sharing tips and scripts based on their experience training German BERT model. Finally, thanks to Rachel Thomas, Jeremy Howard and Sylvain Gugger from [fastai](https://www.fast.ai) for their NLP and Deep Learning courses!
+* I'd like to express my gratitude to Google [TensorFlow Research Cloud (TFRC)](https://www.tensorflow.org/tfrc) for providing the free TPU credits - thank you!
+* Also appreciate the help from Timo Möller from [deepset](https://deepset.ai) for sharing tips and scripts based on their experience training German BERT model.
+* Big thanks to Allegro for releasing KLEJ Benchmark and specifically to Piotr Rybak for help with the evaluation and pointing out some issues with the tokenization. 
+* Finally, thanks to Rachel Thomas, Jeremy Howard and Sylvain Gugger from [fastai](https://www.fast.ai) for their NLP and Deep Learning courses! 
 
 ## Author
 Darek Kłeczek - contact me on Twitter [@dk21](https://twitter.com/dk21)
@@ -75,5 +132,4 @@ Darek Kłeczek - contact me on Twitter [@dk21](https://twitter.com/dk21)
 * https://github.com/google-research/bert
 * https://github.com/narusemotoki/srx_segmenter
 * SRX rules file for sentence splitting in Polish, written by Marcin Miłkowski: https://raw.githubusercontent.com/languagetool-org/languagetool/master/languagetool-core/src/main/resources/org/languagetool/resource/segment.srx
-* PolEmo 2.0 Sentiment Analysis Dataset for CoNLL: https://clarin-pl.eu/dspace/handle/11321/710
-
+* [KLEJ benchmark](https://klejbenchmark.com/leaderboard/)

model_cards/facebook/bart-large-cnn/README.md

@@ -0,0 +1,6 @@
+---
+tags:
+- summarization
+
+license: mit
+---

model_cards/huseinzol05/t5-small-bahasa-cased/README.md

@@ -0,0 +1,74 @@
+---
+language: malay
+---
+
+# Bahasa T5 Model
+
+Pretrained T5 small language model for Malay and Indonesian. 
+
+## Pretraining Corpus
+
+`t5-small-bahasa-cased` model was pretrained on multiple tasks. Below is list of tasks we trained on,
+
+1. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [local Wikipedia](https://github.com/huseinzol05/Malaya-Dataset#wikipedia-1).
+2. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [local news](https://github.com/huseinzol05/Malaya-Dataset#public-news).
+3. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [local parliament text](https://github.com/huseinzol05/Malaya-Dataset#parliament).
+4. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [IIUM Confession](https://github.com/huseinzol05/Malaya-Dataset#iium-confession).
+5. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [Wattpad](https://github.com/huseinzol05/Malaya-Dataset#wattpad).
+6. [Unsupervised](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1875) on [Academia PDF](https://github.com/huseinzol05/Malaya-Dataset#academia-pdf).
+7. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [local Wikipedia](https://github.com/huseinzol05/Malaya-Dataset#wikipedia-1).
+8. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [local news](https://github.com/huseinzol05/Malaya-Dataset#public-news).
+9. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [local parliament text](https://github.com/huseinzol05/Malaya-Dataset#parliament).
+10. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [IIUM Confession](https://github.com/huseinzol05/Malaya-Dataset#iium-confession).
+11. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [Wattpad](https://github.com/huseinzol05/Malaya-Dataset#wattpad).
+12. [Next sentence prediction](https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py#L1129) on [Academia PDF](https://github.com/huseinzol05/Malaya-Dataset#academia-pdf).
+13. [Bahasa SNLI](https://github.com/huseinzol05/Malaya-Dataset#snli).
+14. [Bahasa Question Quora](https://github.com/huseinzol05/Malaya-Dataset#quora).
+15. [Bahasa Natural Questions](https://github.com/huseinzol05/Malaya-Dataset#natural-questions).
+16. [News title summarization](https://github.com/huseinzol05/Malaya-Dataset#crawled-news).
+17. [Stemming to original wikipedia](https://github.com/huseinzol05/Malaya/blob/master/pretrained-model/t5/generate-stemming.ipynb).
+18. [Synonym to original wikipedia](https://github.com/huseinzol05/Malaya/blob/master/pretrained-model/t5/generate-synonym.ipynb).
+
+Preprocessing steps can reproduce from here, [Malaya/pretrained-model/preprocess](https://github.com/huseinzol05/Malaya/tree/master/pretrained-model/preprocess).
+
+## Pretraining details
+
+- This model was trained using Google T5's github [repository](https://github.com/google-research/text-to-text-transfer-transformer) on v3-8 TPU.
+- All steps can reproduce from here, [Malaya/pretrained-model/t5](https://github.com/huseinzol05/Malaya/tree/master/pretrained-model/t5).
+
+## Load Pretrained Model
+
+You can use this model by installing `torch` or `tensorflow` and Huggingface library `transformers`. And you can use it directly by initializing it like this:  
+
+```python
+from transformers import T5Tokenizer, T5Model
+
+model = T5Model.from_pretrained('huseinzol05/t5-small-bahasa-cased')
+tokenizer = T5Tokenizer.from_pretrained('huseinzol05/t5-small-bahasa-cased')
+```
+
+## Example using T5ForConditionalGeneration
+
+```python
+from transformers import T5Tokenizer, T5ForConditionalGeneration
+
+tokenizer = T5Tokenizer.from_pretrained('huseinzol05/t5-small-bahasa-cased')
+model = T5ForConditionalGeneration.from_pretrained('huseinzol05/t5-small-bahasa-cased')
+input_ids = tokenizer.encode('soalan: siapakah perdana menteri malaysia?', return_tensors = 'pt')
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
+
+Output is,
+
+```
+'Mahathir Mohamad'
+```
+
+## Results
+
+For further details on the model performance, simply checkout accuracy page from Malaya, https://malaya.readthedocs.io/en/latest/Accuracy.html, we compared with traditional models.
+
+## Acknowledgement
+
+Thanks to [Im Big](https://www.facebook.com/imbigofficial/), [LigBlou](https://www.facebook.com/ligblou), [Mesolitica](https://mesolitica.com/) and [KeyReply](https://www.keyreply.com/) for sponsoring AWS, Google and GPU clouds to train T5 for Bahasa. 

model_cards/mrm8488/RuPERTa-base-finetuned-ner/README.md

@@ -0,0 +1,92 @@
+---
+language: spanish
+thumbnail:
+---
+
+# RuPERTa-base  (Spanish RoBERTa) + NER 🎃🏷
+
+This model is a fine-tuned on [NER-C](https://www.kaggle.com/nltkdata/conll-corpora) version of [RuPERTa-base](https://huggingface.co/mrm8488/RuPERTa-base) for **NER** downstream task.
+
+## Details of the downstream task (NER) - Dataset
+
+- [Dataset:  CONLL Corpora ES](https://www.kaggle.com/nltkdata/conll-corpora) 📚
+
+| Dataset                | # Examples |
+| ---------------------- | ----- |
+| Train                  |  329 K |
+| Dev                    | 40 K |
+
+
+- [Fine-tune on NER script provided by Huggingface](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_ner.py)
+
+- Labels covered:
+
+```
+B-LOC
+B-MISC
+B-ORG
+B-PER
+I-LOC
+I-MISC
+I-ORG
+I-PER
+O
+```
+
+## Metrics on evaluation set 🧾
+
+|                                                      Metric                                                       |  # score  |
+| :------------------------------------------------------------------------------------: | :-------: |
+| F1                                       | **77.55**  
+| Precision                                | **75.53** | 
+| Recall                                   | **79.68** |    
+
+## Model in action 🔨
+
+
+Example of usage:
+
+```python
+import torch
+from transformers import AutoModelForTokenClassification, AutoTokenizer
+
+id2label = {
+    "0": "B-LOC",
+    "1": "B-MISC",
+    "2": "B-ORG",
+    "3": "B-PER",
+    "4": "I-LOC",
+    "5": "I-MISC",
+    "6": "I-ORG",
+    "7": "I-PER",
+    "8": "O"
+}
+
+text ="Julien, CEO de HF, nació en Francia."
+input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)
+
+outputs = model(input_ids)
+last_hidden_states = outputs[0]
+
+for m in last_hidden_states:
+  for index, n in enumerate(m):
+    if(index > 0 and index <= len(text.split(" "))):
+      print(text.split(" ")[index-1] + ": " + id2label[str(torch.argmax(n).item())])
+      
+'''
+Output:
+--------
+Julien,: I-PER
+CEO: O
+de: O
+HF,: B-ORG
+nació: I-PER
+en: I-PER
+Francia.: I-LOC
+'''
+```
+Yeah! Not too bad 🎉
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain

model_cards/mrm8488/RuPERTa-base-finetuned-pos/README.md

@@ -0,0 +1,111 @@
+---
+language: spanish
+thumbnail:
+---
+
+# RuPERTa-base  (Spanish RoBERTa) + POS 🎃🏷
+
+This model is a fine-tuned on [CONLL CORPORA](https://www.kaggle.com/nltkdata/conll-corpora) version of [RuPERTa-base](https://huggingface.co/mrm8488/RuPERTa-base) for **POS** downstream task.
+
+## Details of the downstream task (POS) - Dataset
+
+- [Dataset:  CONLL Corpora ES](https://www.kaggle.com/nltkdata/conll-corpora) 📚
+
+| Dataset                | # Examples |
+| ---------------------- | ----- |
+| Train                  | 445 K |
+| Dev                    | 55 K |
+
+- [Fine-tune on NER script provided by Huggingface](https://github.com/huggingface/transformers/blob/master/examples/token-classification/run_ner.py)
+
+- Labels covered:
+
+```
+ADJ
+ADP
+ADV
+AUX
+CCONJ
+DET
+INTJ
+NOUN
+NUM
+PART
+PRON
+PROPN
+PUNCT
+SCONJ
+SYM
+VERB
+```
+
+## Metrics on evaluation set 🧾
+
+|                                                      Metric                                                       |  # score  |
+| :------------------------------------------------------------------------------------: | :-------: |
+| F1                                       | **97.39**  
+| Precision                                | **97.47** | 
+| Recall                                   | **9732** |    
+
+## Model in action 🔨
+
+
+Example of usage
+
+```python
+import torch
+from transformers import AutoModelForTokenClassification, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained('mrm8488/RuPERTa-base-finetuned-pos')
+model = AutoModelForTokenClassification.from_pretrained('mrm8488/RuPERTa-base-finetuned-pos')
+
+id2label = {
+    "0": "O",
+    "1": "ADJ",
+    "2": "ADP",
+    "3": "ADV",
+    "4": "AUX",
+    "5": "CCONJ",
+    "6": "DET",
+    "7": "INTJ",
+    "8": "NOUN",
+    "9": "NUM",
+    "10": "PART",
+    "11": "PRON",
+    "12": "PROPN",
+    "13": "PUNCT",
+    "14": "SCONJ",
+    "15": "SYM",
+    "16": "VERB"
+}
+
+text ="Mis amigos están pensando viajar a Londres este verano."
+input_ids = torch.tensor(tokenizer.encode(text)).unsqueeze(0)
+
+outputs = model(input_ids)
+last_hidden_states = outputs[0]
+
+for m in last_hidden_states:
+  for index, n in enumerate(m):
+    if(index > 0 and index <= len(text.split(" "))):
+      print(text.split(" ")[index-1] + ": " + id2label[str(torch.argmax(n).item())])
+      
+'''
+Output:
+--------
+Mis: NUM
+amigos: PRON
+están: AUX
+pensando: ADV
+viajar: VERB
+a: ADP
+Londres: PROPN
+este: DET
+verano..: NOUN
+'''
+```
+Yeah! Not too bad 🎉
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain

model_cards/mrm8488/bert-italian-finedtuned-squadv1-it-alfa/README.md

@@ -0,0 +1,110 @@
+---
+language: italian
+thumbnail:
+---
+
+# Italian BERT fine-tuned on SQuAD_it v1
+
+[Italian BERT base cased](https://huggingface.co/dbmdz/bert-base-italian-cased) fine-tuned on [italian SQuAD](https://github.com/crux82/squad-it) for **Q&A** downstream task.
+
+## Details of Italian BERT
+
+The source data for the Italian BERT model consists of a recent Wikipedia dump and various texts from the OPUS corpora collection. The final training corpus has a size of 13GB and 2,050,057,573 tokens.
+
+For sentence splitting, we use NLTK (faster compared to spacy). Our cased and uncased models are training with an initial sequence length of 512 subwords for ~2-3M steps.
+
+For the XXL Italian models, we use the same training data from OPUS and extend it with data from the Italian part of the OSCAR corpus. Thus, the final training corpus has a size of 81GB and 13,138,379,147 tokens.
+More in its official [model card](https://huggingface.co/dbmdz/bert-base-italian-cased)
+
+Created by [Stefan](https://huggingface.co/stefan-it) at [MDZ](https://huggingface.co/dbmdz)
+
+## Details of the downstream task (Q&A) - Dataset 📚 🧐 ❓
+
+[Italian SQuAD v1.1](https://rajpurkar.github.io/SQuAD-explorer/) is derived from the SQuAD dataset and it is obtained through semi-automatic translation of the SQuAD dataset
+into Italian. It represents a large-scale dataset for open question answering processes on factoid questions in Italian.
+**The dataset contains more than 60,000 question/answer pairs derived from the original English dataset.** The dataset is split into training and test sets to support the replicability of the benchmarking of QA systems:
+
+- `SQuAD_it-train.json`: it contains training examples derived from the original SQuAD 1.1 trainig material.
+- `SQuAD_it-test.json`: it contains test/benchmarking examples derived from the origial SQuAD 1.1 development material.
+
+More details about SQuAD-it can be found in [Croce et al. 2018]. The original paper can be found at this [link](https://link.springer.com/chapter/10.1007/978-3-030-03840-3_29).
+
+## Model training 🏋️‍
+
+The model was trained on a Tesla P100 GPU and 25GB of RAM.
+The script for fine tuning can be found [here](https://github.com/huggingface/transformers/blob/master/examples/question-answering/run_squad.py)
+
+## Results 📝
+
+| Metric | # Value   |
+| ------ | --------- |
+| **EM** | **62.51** |
+| **F1** | **74.16** |
+
+### Raw metrics
+
+```json
+{
+  "exact": 62.5180707057432,
+  "f1": 74.16038329042492,
+  "total": 7609,
+  "HasAns_exact": 62.5180707057432,
+  "HasAns_f1": 74.16038329042492,
+  "HasAns_total": 7609,
+  "best_exact": 62.5180707057432,
+  "best_exact_thresh": 0.0,
+  "best_f1": 74.16038329042492,
+  "best_f1_thresh": 0.0
+}
+```
+
+## Comparison ⚖️
+
+| Model                                                                                                                            | EM        | F1 score  |
+| -------------------------------------------------------------------------------------------------------------------------------- | --------- | --------- |
+| [DrQA-it trained on SQuAD-it ](https://github.com/crux82/squad-it/blob/master/README.md#evaluating-a-neural-model-over-squad-it) | 56.1      | 65.9      |
+| This one                                                                                                                         | **62.51** | **74.16** |
+
+## Model in action 🚀
+
+Fast usage with **pipelines** 🧪
+
+```python
+from transformers import pipeline
+
+nlp_qa = pipeline(
+    'question-answering',
+    model='mrm8488/bert-italian-finedtuned-squadv1-it-alfa',
+    tokenizer='mrm8488/bert-italian-finedtuned-squadv1-it-alfa'
+)
+
+nlp_qa(
+    {
+        'question': 'Per quale lingua stai lavorando?',
+        'context': 'Manuel Romero è colaborando attivamente con HF / trasformatori per il trader del poder de las últimas ' +
+       'técnicas di procesamiento de lenguaje natural al idioma español'
+    }
+)
+
+# Output: {'answer': 'español', 'end': 174, 'score': 0.9925341537498156, 'start': 168}
+```
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain
+
+Dataset citation
+
+<details>
+@InProceedings{10.1007/978-3-030-03840-3_29,
+	author="Croce, Danilo and Zelenanska, Alexandra and Basili, Roberto",
+	editor="Ghidini, Chiara and Magnini, Bernardo and Passerini, Andrea and Traverso, Paolo",
+	title="Neural Learning for Question Answering in Italian",
+	booktitle="AI*IA 2018 -- Advances in Artificial Intelligence",
+	year="2018",
+	publisher="Springer International Publishing",
+	address="Cham",
+	pages="389--402",
+	isbn="978-3-030-03840-3"
+}
+</detail>

model_cards/mrm8488/t5-base-finetuned-squadv2/README.md

@@ -0,0 +1,68 @@
+---
+language: english
+thumbnail:
+---
+
+# T5-base fine-tuned on SQuAD v2
+
+[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) fine-tuned on [SQuAD v2](https://rajpurkar.github.io/SQuAD-explorer/) for **Q&A** downstream task.
+
+## Details of T5
+
+The **T5** model was presented in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf) by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* in Here the abstract:
+
+Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
+
+
+## Details of the downstream task (Q&A) - Dataset 📚 🧐 ❓
+
+[SQuAD v2](https://rajpurkar.github.io/SQuAD-explorer/) combines the 100,000 questions in SQuAD1.1 with over 50,000 unanswerable questions written adversarially by crowdworkers to look similar to answerable ones. To do well on SQuAD2.0, systems must not only answer questions when possible, but also determine when no answer is supported by the paragraph and abstain from answering.
+
+| Dataset  | Split | # samples |
+| -------- | ----- | --------- |
+| SQuAD2.0 | train | 130k      |
+| SQuAD2.0 | eval  | 12.3k     |
+
+
+
+## Model fine-tuning 🏋️‍
+
+The training script is a slightly modified version of [this one](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb)
+
+## Results 📝
+
+| Metric | # Value   |
+| ------ | --------- |
+| **EM** | **77.64** |
+| **F1** | **81.32** |
+
+
+
+## Model in Action 🚀
+
+```python
+from transformers import AutoModelWithLMHead, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-squadv2")
+model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-squadv2")
+
+def get_answer(question, context):
+  input_text = "question: %s  context: %s </s>" % (question, context)
+  features = tokenizer.batch_encode_plus([input_text], return_tensors='pt')
+
+  output = model.generate(input_ids=features['input_ids'], 
+               attention_mask=features['attention_mask'])
+  
+  return tokenizer.decode(output[0])
+
+context = "Manuel have created RuPERTa-base with the support of HF-Transformers and Google"
+question = "Who has supported Manuel?"
+
+get_answer(question, context)
+
+# output: 'HF-Transformers and Google'
+```
+
+> Created by [Manuel Romero/@mrm8488](https://twitter.com/mrm8488) | [LinkedIn](https://www.linkedin.com/in/manuel-romero-cs/)
+
+> Made with <span style="color: #e25555;">&hearts;</span> in Spain

model_cards/oliverguhr/german-sentiment-bert/README.md

@@ -34,7 +34,7 @@ The code above will output following list:
 ["negative","negative","positive","positive","neutral", "neutral"]
 ```
 
-## minimal working Sample
+## A minimal working Sample
 
 
 ```python
@@ -90,7 +90,7 @@ print(model.predict_sentiment(texts))
 
 ## Model and Data
 
-If you are interested in code and data that was used to train this model please have a look at [this repository](https://github.com/oliverguhr/german-sentiment) and our [paper](http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.201.pdf). Here is a table of the F1 scores that his model achieves on following datasets. Since we trained this model on a newer version of the transformer library, the results are slightly better than reported in the paper.
+If you are interested in code and data that was used to train this model please have a look at [this repository](https://github.com/oliverguhr/german-sentiment) and our [paper](http://www.lrec-conf.org/proceedings/lrec2020/pdf/2020.lrec-1.202.pdf). Here is a table of the F1 scores that his model achieves on following datasets. Since we trained this model on a newer version of the transformer library, the results are slightly better than reported in the paper.
 
 | Dataset                                                      | F1 micro Score |
 | :----------------------------------------------------------- | -------------: |

model_cards/savasy/bert-base-turkish-sentiment-cased/README.md

@@ -5,12 +5,12 @@ https://huggingface.co/savasy/bert-base-turkish-sentiment-cased
 This model is used for Sentiment Analysis, which is based on BERTurk for Turkish Language https://huggingface.co/dbmdz/bert-base-turkish-cased
 
 
-# Dataset
+## Dataset
 
-The dataset is taken from the studies [2] and [3] and merged.
+The dataset is taken from the studies [[2]](#paper-2) and [[3]](#paper-3), and merged.
 
 * The study [2] gathered movie and product reviews. The products are book, DVD, electronics, and kitchen.
-The movie dataset is taken from a cinema Web page (www.beyazperde.com) with
+The movie dataset is taken from a cinema Web page ([Beyazperde](www.beyazperde.com)) with
 5331 positive and 5331 negative sentences. Reviews in the Web page are marked in
 scale from 0 to 5 by the users who made the reviews. The study considered a review
 sentiment positive if the rating is equal to or bigger than 4, and negative if it is less
@@ -19,9 +19,9 @@ Web page. They constructed benchmark dataset consisting of reviews regarding som
 products (book, DVD, etc.). Likewise, reviews are marked in the range from 1 to 5,
 and majority class of reviews are 5. Each category has 700 positive and 700 negative
 reviews in which average rating of negative reviews is 2.27 and of positive reviews
-is 4.5. This dataset is also used the study [1]
+is 4.5. This dataset is also used by the study [[1]](#paper-1).
 
-* The study[3] collected tweet dataset. They proposed a new approach for automatically classifying the sentiment of microblog messages. The proposed approach is based on utilizing robust feature representation and fusion. 
+* The study [[3]](#paper-3) collected tweet dataset. They proposed a new approach for automatically classifying the sentiment of microblog messages. The proposed approach is based on utilizing robust feature representation and fusion. 
 
 *Merged Dataset* 
 
@@ -32,20 +32,21 @@ is 4.5. This dataset is also used the study [1]
 |  32000 |train.tsv|
 |  *48290* |*total*|
 
+### The dataset is used by following papers
 
-The dataset is used by following papers
- 
-* 1 Yildirim, Savaş. (2020). Comparing Deep Neural Networks to Traditional Models for Sentiment Analysis in Turkish Language. 10.1007/978-981-15-1216-2_12. 
-* 2 Demirtas, Erkin and Mykola Pechenizkiy. 2013. Cross-lingual polarity detection with machine translation. In Proceedings of the Second International Workshop on Issues of Sentiment
+<a id="paper-1">[1]</a> Yildirim, Savaş. (2020). Comparing Deep Neural Networks to Traditional Models for Sentiment Analysis in Turkish Language. 10.1007/978-981-15-1216-2_12. 
+
+<a id="paper-2">[2]</a> Demirtas, Erkin and Mykola Pechenizkiy. 2013. Cross-lingual polarity detection with machine translation. In Proceedings of the Second International Workshop on Issues of Sentiment
 Discovery and Opinion Mining (WISDOM ’13)
-* [3] Hayran, A.,   Sert, M. (2017), "Sentiment Analysis on Microblog Data based on Word Embedding and Fusion Techniques", IEEE 25th Signal Processing and Communications Applications Conference (SIU 2017), Belek, Turkey
 
-# Training
+<a id="paper-3">[3]</a> Hayran, A.,   Sert, M. (2017), "Sentiment Analysis on Microblog Data based on Word Embedding and Fusion Techniques", IEEE 25th Signal Processing and Communications Applications Conference (SIU 2017), Belek, Turkey
 
-```
+
+## Training
+
+```shell
 export GLUE_DIR="./sst-2-newall"
 export TASK_NAME=SST-2
- 
 
 python3 run_glue.py \
   --model_type bert \
@@ -59,88 +60,79 @@ python3 run_glue.py \
   --learning_rate 2e-5 \
   --num_train_epochs 3.0 \
   --output_dir "./model"
-
 ```
 
 
+## Results
+
+> 05/10/2020 17:00:43 - INFO - transformers.trainer -   \*\*\*\*\* Running Evaluation \*\*\*\*\*  
+> 05/10/2020 17:00:43 - INFO - transformers.trainer -     Num examples = 7999  
+> 05/10/2020 17:00:43 - INFO - transformers.trainer -     Batch size = 8  
+> Evaluation: 100% 1000/1000 [00:34<00:00, 29.04it/s]  
+> 05/10/2020 17:01:17 - INFO - \_\_main__ -   \*\*\*\*\* Eval results sst-2 \*\*\*\*\*  
+> 05/10/2020 17:01:17 - INFO - \_\_main__ -     acc = 0.9539942492811602  
+> 05/10/2020 17:01:17 - INFO - \_\_main__ -     loss = 0.16348013816401363
+
+Accuracy is about **95.4%**
 
 
-# Results
+## Code Usage
 
-> 05/10/2020 17:00:43 - INFO - transformers.trainer -   ***** Running Evaluation *****
-
-> 05/10/2020 17:00:43 - INFO - transformers.trainer -     Num examples = 7999
-
-> 05/10/2020 17:00:43 - INFO - transformers.trainer -     Batch size = 8
-
->Evaluation: 100% 1000/1000 [00:34<00:00, 29.04it/s]
-
->05/10/2020 17:01:17 - INFO - __main__ -   ***** Eval results sst-2 *****
-
->05/10/2020 17:01:17 - INFO - __main__ -     acc = 0.9539942492811602
-
->05/10/2020 17:01:17 - INFO - __main__ -     loss = 0.16348013816401363
-
-
-Accuracy is about *%95.4*
-# Code Usage
-
-```
+```python
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
+
 model = AutoModelForSequenceClassification.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
 tokenizer = AutoTokenizer.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
 sa= pipeline("sentiment-analysis", tokenizer=tokenizer, model=model)
 
-p= sa("bu telefon modelleri çok kaliteli , her parçası çok özel bence")
+p = sa("bu telefon modelleri çok kaliteli , her parçası çok özel bence")
 print(p)
-#[{'label': 'LABEL_1', 'score': 0.9871089}]
-print (p[0]['label']=='LABEL_1')
-#True
+# [{'label': 'LABEL_1', 'score': 0.9871089}]
+print(p[0]['label'] == 'LABEL_1')
+# True
 
-
-p= sa("Film çok kötü ve çok sahteydi")
+p = sa("Film çok kötü ve çok sahteydi")
 print(p)
-#[{'label': 'LABEL_0', 'score': 0.9975505}]
-print (p[0]['label']=='LABEL_1')
-#False
+# [{'label': 'LABEL_0', 'score': 0.9975505}]
+print(p[0]['label'] == 'LABEL_1')
+# False
 ```
 
-# Test your data
+
+## Test
+### Data
 
 Suppose your file has lots of lines of comment and label (1 or 0) at the end  (tab seperated)
 
-> comment1 ... \t label
-
-> comment2 ... \t label
- 
+> comment1 ... \t label  
+> comment2 ... \t label  
 > ...
 
+### Code
 
-
-```
+```python
 from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline
 
-f="/path/to/your/file/yourfile.tsv"
 model = AutoModelForSequenceClassification.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
 tokenizer = AutoTokenizer.from_pretrained("savasy/bert-base-turkish-sentiment-cased")
-sa= pipeline("sentiment-analysis", tokenizer=tokenizer, model=model)
+sa = pipeline("sentiment-analysis", tokenizer=tokenizer, model=model)
 
-i,crr=0,0
-for line in open(f):
- lines=line.strip().split("\t")
- if len(lines)==2:
-  i=i+1
-  if i%100==0:
-   print(i)
-  pred= sa(lines[0])
-  pred=pred[0]["label"].split("_")[1]
-  if pred== lines[1]:
-   crr=crr+1
+input_file = "/path/to/your/file/yourfile.tsv"
+
+i, crr = 0, 0
+for line in open(input_file):
+    lines = line.strip().split("\t")
+    if len(lines) == 2:
+        
+        i = i + 1
+        if i%100 == 0:
+            print(i)
+        
+        pred = sa(lines[0])
+        pred = pred[0]["label"].split("_")[1]
+        
+        if pred == lines[1]:
+        crr = crr + 1
 
 print(crr, i, crr/i)
 ```
-
-
-
-
-

model_cards/valhalla/longformer-base-4096-finetuned-squadv1/README.md

@@ -0,0 +1,54 @@
+# LONGFORMER-BASE-4096 fine-tuned on SQuAD v1
+This is longformer-base-4096 model fine-tuned on SQuAD v1 dataset for question answering task. 
+
+[Longformer](https://arxiv.org/abs/2004.05150) model  created by Iz Beltagy, Matthew E. Peters, Arman Coha from AllenAI.  As the paper explains it 
+
+> `Longformer` is a BERT-like model for long documents. 
+
+The pre-trained model can handle sequences with upto 4096 tokens. 
+
+
+## Model Training
+This model was trained on google colab v100 GPU. You can find the fine-tuning colab here [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1zEl5D-DdkBKva-DdreVOmN0hrAfzKG1o?usp=sharing).
+
+Few things to keep in mind while training longformer for QA task,
+by default longformer uses sliding-window local attention on all tokens. But For QA, all question tokens should  have global attention. For more details on this please refer the paper. The `LongformerForQuestionAnswering` model automatically does that for you. To allow it to do that 
+1. The input sequence must have three sep tokens, i.e the sequence should be encoded like this
+   ` <s> question</s></s> context</s>`.  If you encode the question and answer as a input  pair, then the tokenizer already takes care of that, you shouldn't worry about it.
+2. `input_ids` should always be a batch of examples. 
+
+## Results
+|Metric       | # Value |
+|-------------|---------|
+| Exact Match | 85.1466 |
+| F1          | 91.5415 |
+
+## Model in Action  🚀
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForQuestionAnswering,
+
+tokenizer = AutoTokenizer.from_pretrained("valhalla/longformer-base-4096-finetuned-squadv1")
+model = AutoModelForQuestionAnswering.from_pretrained("valhalla/longformer-base-4096-finetuned-squadv1")
+
+text = "Huggingface has democratized NLP. Huge thanks to Huggingface for this."
+question = "What has Huggingface done ?"
+encoding = tokenizer.encode_plus(question, text, return_tensors="pt")
+input_ids = encoding["input_ids"]
+
+# default is local attention everywhere
+# the forward method will automatically set global attention on question tokens
+attention_mask = encoding["attention_mask"]
+
+start_scores, end_scores = model(input_ids, attention_mask=attention_mask)
+all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0].tolist())
+
+answer_tokens = all_tokens[torch.argmax(start_scores) :torch.argmax(end_scores)+1]
+answer = tokenizer.decode(tokenizer.convert_tokens_to_ids(answer_tokens))
+# output => democratized NLP
+```
+
+The `LongformerForQuestionAnswering` isn't yet supported in `pipeline` . I'll update this card once the support has been added.
+
+> Created with ❤️ by Suraj Patil [![Github icon](https://cdn0.iconfinder.com/data/icons/octicons/1024/mark-github-32.png)](https://github.com/patil-suraj/)
+[![Twitter icon](https://cdn0.iconfinder.com/data/icons/shift-logotypes/32/Twitter-32.png)](https://twitter.com/psuraj28)

notebooks/README.md

@@ -23,4 +23,10 @@ Pull Request so it can be included under the Community notebooks.
 
 | Notebook     |      Description      |      Author      |      |
 |:----------|:-------------|:-------------|------:|
-| [Train T5 on TPU](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb)  | How to train T5 on SQUAD with transformers and nlp | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
+| [Train T5 on TPU](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb)  | How to train T5 on SQUAD with Transformers and Nlp | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
+| [Fine-tune T5 for Classification and Multiple Choice](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb)  | How to fine-tune T5 for classification and multiple choice tasks using a text-to-text format with PyTorch Lightning |  [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) |
+| [Fine-tune DialoGPT on New Datasets and Languages](https://github.com/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb)  | How to fine-tune the DialoGPT model on a new dataset for open-dialog conversational chatbots |  [Nathan Cooper](https://github.com/ncoop57) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) |
+| [Long Sequence Modeling with Reformer](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb)  | How to train on sequences as long as 500,000 tokens with Reformer |  [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb)  |
+| [Fine-tune BART for Summarization](https://github.com/ohmeow/ohmeow_website/blob/master/_notebooks/2020-05-23-text-generation-with-blurr.ipynb) | How to fine-tune BART for summarization with fastai using blurr | [Wayde Gilliam](https://ohmeow.com/) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ohmeow/ohmeow_website/blob/master/_notebooks/2020-05-23-text-generation-with-blurr.ipynb) |
+| [Fine-tune a pre-trained Transformer on anyone's tweets](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb)  | How to generate tweets in the style of your favorite Twitter account by fine-tune a GPT-2 model |  [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) |
+| [A Step by Step Guide to Tracking Hugging Face Model Performance](https://colab.research.google.com/drive/1NEiqNPhiouu2pPwDAVeFoN4-vTYMz9F8)  | A quick tutorial for training NLP models with HuggingFace and & visualizing their performance with Weights & Biases |  [Jack Morris](https://github.com/jxmorris12) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1NEiqNPhiouu2pPwDAVeFoN4-vTYMz9F8) |

setup.py

@@ -95,8 +95,8 @@ extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3", "sciki
 
 setup(
     name="transformers",
-    version="2.9.1",
-    author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
+    version="2.10.0",
+    author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",
     long_description=open("README.md", "r", encoding="utf-8").read(),
@@ -111,6 +111,8 @@ setup(
         "tokenizers == 0.7.0",
         # dataclasses for Python versions that don't have it
         "dataclasses;python_version<'3.7'",
+        # utilities from PyPA to e.g. compare versions
+        "packaging",
         # filesystem locks e.g. to prevent parallel downloads
         "filelock",
         # for downloading models over HTTPS
@@ -125,7 +127,9 @@ setup(
         "sacremoses",
     ],
     extras_require=extras,
-    scripts=["transformers-cli"],
+    entry_points={
+        "console_scripts": ["transformers-cli=transformers.commands.transformers_cli:main"]
+    },
     python_requires=">=3.6.0",
     classifiers=[
         "Development Status :: 5 - Production/Stable",

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "2.9.1"
+__version__ = "2.10.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -19,19 +19,6 @@ else:
 
 import logging
 
-# Benchmarking
-from .benchmark_utils import (
-    Frame,
-    Memory,
-    MemoryState,
-    MemorySummary,
-    MemoryTrace,
-    UsedMemoryState,
-    bytes_to_human_readable,
-    start_memory_tracing,
-    stop_memory_tracing,
-)
-
 # Configurations
 from .configuration_albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
 from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, CONFIG_MAPPING, AutoConfig
@@ -44,6 +31,7 @@ from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, Electr
 from .configuration_encoder_decoder import EncoderDecoderConfig
 from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
 from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
+from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
 from .configuration_marian import MarianConfig
 from .configuration_mmbt import MMBTConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
@@ -138,6 +126,7 @@ from .tokenization_distilbert import DistilBertTokenizer, DistilBertTokenizerFas
 from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
+from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_reformer import ReformerTokenizer
 from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
@@ -319,6 +308,7 @@ if is_torch_available():
         ElectraForMaskedLM,
         ElectraForTokenClassification,
         ElectraPreTrainedModel,
+        ElectraForSequenceClassification,
         ElectraModel,
         load_tf_weights_in_electra,
         ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -332,6 +322,15 @@ if is_torch_available():
         REFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
+    from .modeling_longformer import (
+        LongformerModel,
+        LongformerForMaskedLM,
+        LongformerForSequenceClassification,
+        LongformerForTokenClassification,
+        LongformerForQuestionAnswering,
+        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
+    )
+
     # Optimization
     from .optimization import (
         AdamW,
@@ -347,6 +346,9 @@ if is_torch_available():
     from .data.data_collator import DefaultDataCollator, DataCollator, DataCollatorForLanguageModeling
     from .data.datasets import GlueDataset, TextDataset, LineByLineTextDataset, GlueDataTrainingArguments
 
+    # Benchmarks
+    from .benchmark import PyTorchBenchmark, PyTorchBenchmarkArguments
+
 # TensorFlow
 if is_tf_available():
     from .modeling_tf_utils import (

src/transformers/benchmark/__init__.py

@@ -0,0 +1,10 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+from ..file_utils import is_torch_available
+
+
+if is_torch_available():
+    from .benchmark_args import PyTorchBenchmarkArguments
+    from .benchmark import PyTorchBenchmark

src/transformers/benchmark/benchmark.py

@@ -0,0 +1,146 @@
+# 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 in PyTorch.
+"""
+
+
+import inspect
+import logging
+import timeit
+
+from transformers import MODEL_MAPPING, MODEL_WITH_LM_HEAD_MAPPING, PretrainedConfig, is_torch_available
+
+from .benchmark_utils import Benchmark, Memory, start_memory_tracing, stop_memory_tracing
+
+
+if is_torch_available():
+    import torch
+    from .benchmark_args import PyTorchBenchmarkArguments
+
+
+logger = logging.getLogger(__name__)
+
+
+class PyTorchBenchmark(Benchmark):
+
+    args: PyTorchBenchmarkArguments
+    configs: PretrainedConfig
+    framework: str = "PyTorch"
+
+    @property
+    def framework_version(self):
+        return torch.__version__
+
+    def train(self, model_name, batch_size, sequence_length, trace_memory=False):
+        try:
+            config = self.config_dict[model_name]
+            model = MODEL_WITH_LM_HEAD_MAPPING[config.__class__](config)
+            model.to(self.args.device)
+            model.train()
+
+            input_ids = torch.randint(
+                model.config.vocab_size, (batch_size, sequence_length), dtype=torch.long, device=self.args.device
+            )
+
+            def compute_loss_and_backprob():
+                # TODO: Not all models call labels argument labels => this hack using the function signature should be corrected once all models have a common name for labels
+                function_argument_names = inspect.getfullargspec(model.forward).args
+                if "labels" in function_argument_names:
+                    loss = model(input_ids, labels=input_ids)[0]
+                elif "lm_labels" in function_argument_names:
+                    loss = model(input_ids, lm_labels=input_ids)[0]
+                elif "masked_lm_labels" in function_argument_names:
+                    loss = model(input_ids, masked_lm_labels=input_ids)[0]
+                else:
+                    NotImplementedError(f"{model_name} does not seem to allow training with labels")
+
+                loss.backward()
+                model.zero_grad()
+
+            if trace_memory is True:
+                if self.args.trace_memory_line_by_line or self.args.n_gpu == 0:
+                    trace = start_memory_tracing("transformers")
+                else:
+                    # clear cuda cache
+                    torch.cuda.empty_cache()
+                    torch.cuda.reset_peak_memory_stats()
+
+                # calculate loss and do backpropagation
+                compute_loss_and_backprob()
+
+                if self.args.trace_memory_line_by_line or self.args.n_gpu == 0:
+                    summary = stop_memory_tracing(trace)
+                    memory = summary.total
+                else:
+                    memory = Memory(torch.cuda.max_memory_reserved())
+
+                return memory
+            else:
+                # as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
+                runtimes = timeit.repeat(lambda: compute_loss_and_backprob(), repeat=self.args.repeat, number=10,)
+                return min(runtimes) / 10.0
+        except RuntimeError as e:
+            self.print_fn("Doesn't fit on GPU. {}".format(e))
+            return "N/A"
+
+    def inference(self, model_name, batch_size, sequence_length, trace_memory=False):
+        try:
+            config = self.config_dict[model_name]
+            model = MODEL_MAPPING[config.__class__](config)
+            model.to(self.args.device)
+            model.eval()
+
+            input_ids = torch.randint(
+                config.vocab_size, (batch_size, sequence_length), dtype=torch.long, device=self.args.device
+            )
+            if trace_memory is True:
+                if self.args.trace_memory_line_by_line or self.args.n_gpu == 0:
+                    trace = start_memory_tracing("transformers")
+                else:
+                    # clear cuda cache
+                    torch.cuda.empty_cache()
+                    if hasattr(torch.cuda, "max_memory_reserved"):
+                        torch.cuda.reset_peak_memory_stats()
+                    else:
+                        logger.info(
+                            "Please consider updating PyTorch to version 1.4 to get more accuracy on GPU memory usage"
+                        )
+                        torch.cuda.reset_max_memory_cached()
+
+                model(input_ids)
+
+                if self.args.trace_memory_line_by_line or self.args.n_gpu == 0:
+                    summary = stop_memory_tracing(trace)
+                    memory = summary.total
+                else:
+                    if hasattr(torch.cuda, "max_memory_reserved"):
+                        memory = Memory(torch.cuda.max_memory_reserved())
+                    else:
+                        logger.info(
+                            "Please consider updating PyTorch to version 1.4 to get more accuracy on GPU memory usage"
+                        )
+                        memory = Memory(torch.cuda.max_memory_cached())
+
+                return memory
+            else:
+                # as written in https://docs.python.org/2/library/timeit.html#timeit.Timer.repeat, min should be taken rather than the average
+                runtimes = timeit.repeat(lambda: model(input_ids), repeat=self.args.repeat, number=10,)
+                return min(runtimes) / 10.0
+
+        except RuntimeError as e:
+            self.print_fn("Doesn't fit on GPU. {}".format(e))
+            return "N/A"

src/transformers/benchmark/benchmark_args.py

@@ -0,0 +1,78 @@
+# 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.
+
+import logging
+from dataclasses import dataclass, field
+from typing import Tuple
+
+from ..file_utils import cached_property, is_torch_available, torch_required
+from .benchmark_args_utils import BenchmarkArguments
+
+
+if is_torch_available():
+    import torch
+
+try:
+    import torch_xla.core.xla_model as xm
+
+    _has_tpu = True
+except ImportError:
+    _has_tpu = False
+
+
+@torch_required
+def is_tpu_available():
+    return _has_tpu
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class PyTorchBenchmarkArguments(BenchmarkArguments):
+    no_cuda: bool = field(default=False, metadata={"help": "Whether to run on available cuda devices"})
+    torchscript: bool = field(default=False, metadata={"help": "Trace the models using torchscript"})
+    fp16: bool = field(default=False, metadata={"help": "Use FP16 to accelerate inference."})
+
+    @cached_property
+    @torch_required
+    def _setup_devices(self) -> Tuple["torch.device", int]:
+        logger.info("PyTorch: setting up devices")
+        if self.no_cuda:
+            device = torch.device("cpu")
+            n_gpu = 0
+        elif is_tpu_available():
+            device = xm.xla_device()
+            n_gpu = 0
+        else:
+            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+            n_gpu = torch.cuda.device_count()
+        return device, n_gpu
+
+    @property
+    @torch_required
+    def device_idx(self) -> int:
+        return torch.cuda.current_device()
+
+    @property
+    @torch_required
+    def device(self) -> "torch.device":
+        return self._setup_devices[0]
+
+    @property
+    @torch_required
+    def n_gpu(self):
+        return self._setup_devices[1]

src/transformers/benchmark/benchmark_args_utils.py

@@ -0,0 +1,98 @@
+# 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.
+
+import dataclasses
+import json
+from dataclasses import dataclass, field
+from time import time
+from typing import List
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class BenchmarkArguments:
+    """
+    BenchMarkArguments are arguments we use in our benchmark scripts
+    **which relate to the training loop itself**.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    models: List[str] = list_field(
+        default=[],
+        metadata={
+            "help": "Model checkpoints to be provided to the AutoModel classes. Leave blank to benchmark the base version of all available models"
+        },
+    )
+
+    batch_sizes: List[int] = list_field(
+        default=[8], metadata={"help": "List of batch sizes for which memory and time performance will be evaluated"}
+    )
+
+    sequence_lengths: List[int] = list_field(
+        default=[8, 32, 128, 512],
+        metadata={"help": "List of sequence lengths for which memory and time performance will be evaluated"},
+    )
+
+    no_inference: bool = field(default=False, metadata={"help": "Don't benchmark inference of model"})
+    training: bool = field(default=False, metadata={"help": "Benchmark training of model"})
+    verbose: bool = field(default=False, metadata={"help": "Verbose memory tracing"})
+    no_speed: bool = field(default=False, metadata={"help": "Don't perform speed measurments"})
+    no_memory: bool = field(default=False, metadata={"help": "Don't perform memory measurments"})
+    trace_memory_line_by_line: bool = field(default=False, metadata={"help": "Trace memory line by line"})
+    save_to_csv: bool = field(default=False, metadata={"help": "Save result to a CSV file"})
+    log_print: bool = field(default=False, metadata={"help": "Save all print statements in a log file"})
+    no_env_print: bool = field(default=False, metadata={"help": "Don't print environment information"})
+    inference_time_csv_file: str = field(
+        default=f"inference_time_{round(time())}.csv",
+        metadata={"help": "CSV filename used if saving time results to csv."},
+    )
+    inference_memory_csv_file: str = field(
+        default=f"inference_memory_{round(time())}.csv",
+        metadata={"help": "CSV filename used if saving memory results to csv."},
+    )
+    train_time_csv_file: str = field(
+        default=f"train_time_{round(time())}.csv",
+        metadata={"help": "CSV filename used if saving time results to csv for training."},
+    )
+    train_memory_csv_file: str = field(
+        default=f"train_memory_{round(time())}.csv",
+        metadata={"help": "CSV filename used if saving memory results to csv for training."},
+    )
+    env_info_csv_file: str = field(
+        default=f"env_info_{round(time())}.csv",
+        metadata={"help": "CSV filename used if saving environment information."},
+    )
+    log_filename: str = field(
+        default=f"log_{round(time())}.csv",
+        metadata={"help": "Log filename used if print statements are saved in log."},
+    )
+    repeat: int = field(default=3, metadata={"help": "Times an experiment will be run."})
+
+    def to_json_string(self):
+        """
+        Serializes this instance to a JSON string.
+        """
+        return json.dumps(dataclasses.asdict(self), indent=2)
+
+    @property
+    def model_names(self):
+        return self.models

src/transformers/benchmark/benchmark_utils.py

@@ -4,18 +4,28 @@ This file is adapted from the AllenNLP library at https://github.com/allenai/all
 Copyright by the AllenNLP authors.
 """
 
+import copy
+import csv
 import linecache
 import logging
 import os
+import platform
 import sys
-from collections import defaultdict
+from abc import ABC, abstractmethod
+from collections import defaultdict, namedtuple
+from datetime import datetime
 from typing import Iterable, List, NamedTuple, Optional, Union
 
-from .file_utils import is_tf_available, is_torch_available
+from transformers import AutoConfig, PretrainedConfig
+from transformers import __version__ as version
+
+from ..file_utils import is_tf_available, is_torch_available
+from .benchmark_args_utils import BenchmarkArguments
 
 
 if is_torch_available():
     from torch.cuda import empty_cache as torch_empty_cache
+
 if is_tf_available():
     from tensorflow.python.eager import context as tf_context
 
@@ -25,6 +35,10 @@ logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
 _is_memory_tracing_enabled = False
 
+BenchmarkOutput = namedtuple(
+    "BenchmarkOutput", ["time_inference_result", "memory_inference_result", "time_train_result", "memory_train_result"]
+)
+
 
 def is_memory_tracing_enabled():
     global _is_memory_tracing_enabled
@@ -62,14 +76,14 @@ class UsedMemoryState(NamedTuple):
 
 class Memory(NamedTuple):
     """ `Memory` NamedTuple have a single field `bytes` and
-        you can get a human readable string of the number of bytes by calling `__repr__`
+        you can get a human readable str of the number of mega bytes by calling `__repr__`
             - `byte` (integer): number of bytes,
     """
 
     bytes: int
 
     def __repr__(self) -> str:
-        return bytes_to_human_readable(self.bytes)
+        return str(bytes_to_mega_bytes(self.bytes))
 
 
 class MemoryState(NamedTuple):
@@ -99,6 +113,7 @@ class MemorySummary(NamedTuple):
 
     sequential: List[MemoryState]
     cumulative: List[MemoryState]
+    current: List[MemoryState]
     total: Memory
 
 
@@ -234,10 +249,12 @@ def start_memory_tracing(
 
             # Sum used memory for all GPUs
             py3nvml.nvmlInit()
+
             for i in devices:
                 handle = py3nvml.nvmlDeviceGetHandleByIndex(i)
                 meminfo = py3nvml.nvmlDeviceGetMemoryInfo(handle)
                 gpu_mem += meminfo.used
+
             py3nvml.nvmlShutdown()
 
         mem_state = UsedMemoryState(traced_state, cpu_mem, gpu_mem)
@@ -295,8 +312,11 @@ def stop_memory_tracing(
 
     if memory_trace is not None and len(memory_trace) > 1:
         memory_diff_trace = []
+        memory_curr_trace = []
+
         cumulative_memory_dict = defaultdict(lambda: [0, 0, 0])
-        for (frame, cpu_mem, gpu_mem), (next_frame, next_cpu_mem, next_gpu_mem) in zip(
+
+        for ((frame, cpu_mem, gpu_mem), (next_frame, next_cpu_mem, next_gpu_mem),) in zip(
             memory_trace[:-1], memory_trace[1:]
         ):
             cpu_mem_inc = next_cpu_mem - cpu_mem
@@ -307,6 +327,16 @@ def stop_memory_tracing(
                     frame=frame, cpu=Memory(cpu_mem_inc), gpu=Memory(gpu_mem_inc), cpu_gpu=Memory(cpu_gpu_mem_inc),
                 )
             )
+
+            memory_curr_trace.append(
+                MemoryState(
+                    frame=frame,
+                    cpu=Memory(next_cpu_mem),
+                    gpu=Memory(next_gpu_mem),
+                    cpu_gpu=Memory(next_gpu_mem + next_cpu_mem),
+                )
+            )
+
             cumulative_memory_dict[frame][0] += cpu_mem_inc
             cumulative_memory_dict[frame][1] += gpu_mem_inc
             cumulative_memory_dict[frame][2] += cpu_gpu_mem_inc
@@ -321,21 +351,287 @@ def stop_memory_tracing(
             for frame, (cpu_mem_inc, gpu_mem_inc, cpu_gpu_mem_inc) in cumulative_memory
         )
 
+        memory_curr_trace = sorted(memory_curr_trace, key=lambda x: x.cpu_gpu.bytes, reverse=True)
+
         if ignore_released_memory:
             total_memory = sum(max(0, step_trace.cpu_gpu.bytes) for step_trace in memory_diff_trace)
         else:
             total_memory = sum(step_trace.cpu_gpu.bytes for step_trace in memory_diff_trace)
+
         total_memory = Memory(total_memory)
-        return MemorySummary(sequential=memory_diff_trace, cumulative=cumulative_memory, total=total_memory)
+
+        return MemorySummary(
+            sequential=memory_diff_trace, cumulative=cumulative_memory, current=memory_curr_trace, total=total_memory,
+        )
 
     return None
 
 
-def bytes_to_human_readable(memory_amount: int) -> str:
-    """ Utility to convert a number of bytes (int) in a human readable string (with units)
+def bytes_to_mega_bytes(memory_amount: int) -> int:
+    """ Utility to convert a number of bytes (int) into a number of mega bytes (int)
     """
-    for unit in ["B", "KB", "MB", "GB"]:
-        if memory_amount > -1024.0 and memory_amount < 1024.0:
-            return "{:.3f}{}".format(memory_amount, unit)
-        memory_amount /= 1024.0
-    return "{:.3f}TB".format(memory_amount)
+    return memory_amount >> 20
+
+
+class Benchmark(ABC):
+    """
+    Benchmarks is a simple but feature-complete benchmarking script
+    to compare memory and time performance of models in Transformers.
+    """
+
+    args: BenchmarkArguments
+    configs: PretrainedConfig
+    framework: str
+
+    def __init__(self, args: BenchmarkArguments = None, configs: PretrainedConfig = None):
+        self.args = args
+
+        if configs is None:
+            self.config_dict = {
+                model_name: AutoConfig.from_pretrained(model_name) for model_name in self.args.model_names
+            }
+        else:
+            self.config_dict = {model_name: config for model_name, config in zip(self.args.model_names, configs)}
+
+        self._print_fn = None
+        self._framework_version = None
+        self._environment_info = None
+
+    @property
+    def print_fn(self):
+        if self._print_fn is None:
+            if self.args.log_print:
+                logging.basicConfig(
+                    level=logging.DEBUG,
+                    filename=self.args.log_filename,
+                    filemode="a+",
+                    format="%(asctime)-15s %(levelname)-8s %(message)s",
+                )
+
+                def print_and_log(*args):
+                    logging.info(*args)
+                    print(*args)
+
+                self._print_fn = print_and_log
+            else:
+                self._print_fn = print
+        return self._print_fn
+
+    @property
+    def is_gpu(self):
+        return self.args.n_gpu > 0
+
+    @property
+    @abstractmethod
+    def framework_version(self):
+        pass
+
+    @abstractmethod
+    def train(self, model_name, batch_size, sequence_length):
+        pass
+
+    @abstractmethod
+    def inference(self, model_name, batch_size, sequence_length):
+        pass
+
+    def run(self):
+        result_dict = {model_name: {} for model_name in self.args.model_names}
+        inference_result_time = copy.deepcopy(result_dict)
+        inference_result_memory = copy.deepcopy(result_dict)
+        train_result_time = copy.deepcopy(result_dict)
+        train_result_memory = copy.deepcopy(result_dict)
+
+        for c, model_name in enumerate(self.args.model_names):
+            self.print_fn(f"{c + 1} / {len(self.args.model_names)}")
+
+            model_dict = {
+                "bs": self.args.batch_sizes,
+                "ss": self.args.sequence_lengths,
+                "result": {i: {} for i in self.args.batch_sizes},
+            }
+            inference_result_time[model_name] = copy.deepcopy(model_dict)
+            inference_result_memory[model_name] = copy.deepcopy(model_dict)
+            train_result_time[model_name] = copy.deepcopy(model_dict)
+            train_result_memory[model_name] = copy.deepcopy(model_dict)
+
+            for batch_size in self.args.batch_sizes:
+                for sequence_length in self.args.sequence_lengths:
+                    if not self.args.no_inference:
+                        if not self.args.no_memory:
+                            memory = self.inference(model_name, batch_size, sequence_length, trace_memory=True)
+                            inference_result_memory[model_name]["result"][batch_size][sequence_length] = memory
+                        if not self.args.no_speed:
+                            time = self.inference(model_name, batch_size, sequence_length, trace_memory=False)
+                            inference_result_time[model_name]["result"][batch_size][sequence_length] = time
+
+                    if self.args.training:
+                        if not self.args.no_memory:
+                            memory = self.train(model_name, batch_size, sequence_length, trace_memory=True)
+                            train_result_memory[model_name]["result"][batch_size][sequence_length] = memory
+                        if not self.args.no_speed:
+                            time = self.inference(model_name, batch_size, sequence_length, trace_memory=False)
+                            train_result_time[model_name]["result"][batch_size][sequence_length] = time
+
+        if not self.args.no_inference:
+            if not self.args.no_speed:
+                self.print_fn("======= INFERENCE - SPEED - RESULT =======")
+                self.print_results(inference_result_time)
+                self.save_to_csv(inference_result_time, self.args.inference_time_csv_file)
+
+            if not self.args.no_memory:
+                self.print_fn("======= INFERENCE - MEMORY - RESULT =======")
+                self.print_results(inference_result_memory)
+                self.save_to_csv(inference_result_memory, self.args.inference_memory_csv_file)
+
+        if self.args.training:
+            if not self.args.no_speed:
+                self.print_fn("======= TRAIN - SPEED - RESULT =======")
+                self.print_results(train_result_time)
+                self.save_to_csv(train_result_time, self.args.train_time_csv_file)
+
+            if not self.args.no_memory:
+                self.print_fn("======= TRAIN - MEMORY - RESULT =======")
+                self.print_results(train_result_memory)
+                self.save_to_csv(train_result_memory, self.args.train_memory_csv_file)
+
+        if not self.args.no_env_print:
+            self.print_fn("\n======== ENVIRONMENT - INFORMATION ========")
+            self.print_fn(
+                "\n".join(["- {}: {}".format(prop, val) for prop, val in self.environment_info.items()]) + "\n"
+            )
+
+        if self.args.save_to_csv:
+            with open(self.args.env_info_csv_file, mode="w", newline="") as csv_file:
+                writer = csv.writer(csv_file)
+                for key, value in self.environment_info.items():
+                    writer.writerow([key, value])
+
+        return BenchmarkOutput(inference_result_time, inference_result_memory, train_result_time, train_result_memory)
+
+    @property
+    def environment_info(self):
+        if self._environment_info is None:
+            info = {}
+            info["transformers_version"] = version
+            info["framework"] = self.framework
+            info["framework_version"] = self.framework_version
+            info["python_version"] = platform.python_version()
+            info["system"] = platform.system()
+            info["cpu"] = platform.processor()
+            info["architecture"] = platform.architecture()[0]
+            info["date"] = datetime.date(datetime.now())
+            info["time"] = datetime.time(datetime.now())
+
+            try:
+                import psutil
+            except (ImportError):
+                logger.warning(
+                    "Psutil not installed, we won't log available CPU memory."
+                    "Install psutil (pip install psutil) to log available CPU memory."
+                )
+                info["cpu_ram_mb"] = "N/A"
+            else:
+                info["cpu_ram_mb"] = bytes_to_mega_bytes(psutil.virtual_memory().total)
+
+            info["use_gpu"] = self.is_gpu
+            if self.is_gpu:
+                info["num_gpus"] = self.args.n_gpu
+                try:
+                    from py3nvml import py3nvml
+
+                    py3nvml.nvmlInit()
+                    handle = py3nvml.nvmlDeviceGetHandleByIndex(self.args.device_idx)
+                except ImportError:
+                    logger.warning(
+                        "py3nvml not installed, we won't log GPU memory usage. "
+                        "Install py3nvml (pip install py3nvml) to log information about GPU."
+                    )
+                    info["gpu"] = "N/A"
+                    info["gpu_ram_mb"] = "N/A"
+                    info["gpu_power_watts"] = "N/A"
+                    info["gpu_performance_state"] = "N/A"
+                except (OSError, py3nvml.NVMLError):
+                    logger.warning(
+                        "Error while initializing comunication with GPU. " "We won't log information about GPU."
+                    )
+                    info["gpu"] = "N/A"
+                    info["gpu_ram_mb"] = "N/A"
+                    info["gpu_power_watts"] = "N/A"
+                    info["gpu_performance_state"] = "N/A"
+                    py3nvml.nvmlShutdown()
+                else:
+                    info["gpu"] = py3nvml.nvmlDeviceGetName(handle)
+                    info["gpu_ram_mb"] = bytes_to_mega_bytes(py3nvml.nvmlDeviceGetMemoryInfo(handle).total)
+                    info["gpu_power_watts"] = py3nvml.nvmlDeviceGetPowerManagementLimit(handle) / 1000
+                    info["gpu_performance_state"] = py3nvml.nvmlDeviceGetPerformanceState(handle)
+                    py3nvml.nvmlShutdown()
+
+            self._environment_info = info
+        return self._environment_info
+
+    def print_results(self, result_dict):
+        for model_name in self.args.model_names:
+            self.print_fn("\t" + f"======= MODEL CHECKPOINT: {model_name} =======")
+            for batch_size in result_dict[model_name]["bs"]:
+                for sequence_length in result_dict[model_name]["ss"]:
+                    result = result_dict[model_name]["result"][batch_size][sequence_length]
+                    if isinstance(result, float):
+                        self.print_fn(
+                            f"\t\t{model_name}/{batch_size}/{sequence_length}: " f"{(round(1000 * result) / 1000)}s"
+                        )
+                    else:
+                        self.print_fn(f"\t\t{model_name}/{batch_size}/{sequence_length}: " f"{result} MB")
+
+    def print_memory_trace_statistics(self, summary: MemorySummary):
+        self.print_fn(
+            "\nLine by line memory consumption:\n"
+            + "\n".join(
+                f"{state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.sequential
+            )
+        )
+        self.print_fn(
+            "\nLines with top memory consumption:\n"
+            + "\n".join(
+                f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.cumulative[:6]
+            )
+        )
+        self.print_fn(
+            "\nLines with lowest memory consumption:\n"
+            + "\n".join(
+                f"=> {state.frame.filename}:{state.frame.line_number}: mem {state.cpu_gpu}: {state.frame.line_text}"
+                for state in summary.cumulative[-6:]
+            )
+        )
+        self.print_fn(f"\nTotal memory increase: {summary.total}")
+
+    def save_to_csv(self, result_dict, filename):
+        if not self.args.save_to_csv:
+            return
+        self.print_fn("Saving results to csv.")
+        with open(filename, mode="w") as csv_file:
+
+            assert len(self.args.model_names) > 0, "At least 1 model should be defined, but got {}".format(
+                self.model_names
+            )
+
+            fieldnames = ["model", "batch_size", "sequence_length"]
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames + ["result"])
+            writer.writeheader()
+
+            for model_name in self.args.model_names:
+                result_dict_model = result_dict[model_name]["result"]
+                for bs in result_dict_model:
+                    for ss in result_dict_model[bs]:
+                        result_model = result_dict_model[bs][ss]
+                        writer.writerow(
+                            {
+                                "model": model_name,
+                                "batch_size": bs,
+                                "sequence_length": ss,
+                                "result": ("{}" if not isinstance(result_model, float) else "{:.4f}").format(
+                                    result_model
+                                ),
+                            }
+                        )

src/transformers/commands/transformers_cli.py

@@ -8,9 +8,10 @@ from transformers.commands.run import RunCommand
 from transformers.commands.serving import ServeCommand
 from transformers.commands.user import UserCommands
 
-if __name__ == '__main__':
-    parser = ArgumentParser('Transformers CLI tool', usage='transformers-cli <command> [<args>]')
-    commands_parser = parser.add_subparsers(help='transformers-cli command helpers')
+
+def main():
+    parser = ArgumentParser("Transformers CLI tool", usage="transformers-cli <command> [<args>]")
+    commands_parser = parser.add_subparsers(help="transformers-cli command helpers")
 
     # Register commands
     ConvertCommand.register_subcommand(commands_parser)
@@ -23,10 +24,14 @@ if __name__ == '__main__':
     # Let's go
     args = parser.parse_args()
 
-    if not hasattr(args, 'func'):
+    if not hasattr(args, "func"):
         parser.print_help()
         exit(1)
 
     # Run
     service = args.func(args)
     service.run()
+
+
+if __name__ == "__main__":
+    main()

src/transformers/configuration_auto.py

@@ -28,6 +28,7 @@ from .configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, Electr
 from .configuration_encoder_decoder import EncoderDecoderConfig
 from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
 from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
+from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
 from .configuration_marian import MarianConfig
 from .configuration_openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig
 from .configuration_reformer import ReformerConfig
@@ -62,6 +63,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
         FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -77,6 +79,7 @@ CONFIG_MAPPING = OrderedDict(
         ("marian", MarianConfig,),
         ("bart", BartConfig,),
         ("reformer", ReformerConfig,),
+        ("longformer", LongformerConfig,),
         ("roberta", RobertaConfig,),
         ("flaubert", FlaubertConfig,),
         ("bert", BertConfig,),
@@ -133,6 +136,7 @@ class AutoConfig:
             - contains `albert`: :class:`~transformers.AlbertConfig` (ALBERT model)
             - contains `camembert`: :class:`~transformers.CamembertConfig` (CamemBERT model)
             - contains `xlm-roberta`: :class:`~transformers.XLMRobertaConfig` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerConfig` (Longformer model)
             - contains `roberta`: :class:`~transformers.RobertaConfig` (RoBERTa model)
             - contains `reformer`: :class:`~transformers.ReformerConfig` (Reformer model)
             - contains `bert`: :class:`~transformers.BertConfig` (Bert model)
@@ -145,7 +149,6 @@ class AutoConfig:
             - contains `flaubert` : :class:`~transformers.FlaubertConfig` (Flaubert model)
             - contains `electra` : :class:`~transformers.ElectraConfig` (ELECTRA model)
 
-
         Args:
             pretrained_model_name_or_path (:obj:`string`):
                 Is either: \

src/transformers/configuration_distilbert.py

@@ -73,10 +73,10 @@ class DistilBertConfig(PretrainedConfig):
                 The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
             qa_dropout (:obj:`float`, optional, defaults to 0.1):
                 The dropout probabilities used in the question answering model
-                :class:`~tranformers.DistilBertForQuestionAnswering`.
+                :class:`~transformers.DistilBertForQuestionAnswering`.
             seq_classif_dropout (:obj:`float`, optional, defaults to 0.2):
                 The dropout probabilities used in the sequence classification model
-                :class:`~tranformers.DistilBertForSequenceClassification`.
+                :class:`~transformers.DistilBertForSequenceClassification`.
 
         Example::
 

src/transformers/configuration_longformer.py

@@ -0,0 +1,73 @@
+# coding=utf-8
+# Copyright 2020 The Allen Institute for AI team and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Longformer configuration """
+
+import logging
+from typing import List, Union
+
+from .configuration_roberta import RobertaConfig
+
+
+logger = logging.getLogger(__name__)
+
+LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "allenai/longformer-base-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096/config.json",
+    "allenai/longformer-large-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096/config.json",
+    "allenai/longformer-large-4096-finetuned-triviaqa": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-finetuned-triviaqa/config.json",
+    "allenai/longformer-base-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096-extra.pos.embd.only/config.json",
+    "allenai/longformer-large-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-extra.pos.embd.only/config.json",
+}
+
+
+class LongformerConfig(RobertaConfig):
+    r"""
+        This is the configuration class to store the configuration of an :class:`~transformers.LongformerModel`.
+        It is used to instantiate an Longformer model according to the specified arguments, defining the model
+        architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+        the RoBERTa `roberta-base <https://huggingface.co/roberta-base>`__ architecture with a sequence length 4,096.
+
+        The :class:`~transformers.LongformerConfig` class directly inherits :class:`~transformers.RobertaConfig`.
+        It reuses the same defaults. Please check the parent class for more information.
+
+        Args:
+            attention_window (:obj:`int` or :obj:`List[int]`, optional, defaults to 512):
+                Size of an attention window around each token. If :obj:`int`, use the same size for all layers.
+                To specify a different window size for each layer, use a :obj:`List[int]` where
+                ``len(attention_window) == num_hidden_layers``.
+
+        Example::
+
+            from transformers import LongformerConfig, LongformerModel
+
+            # Initializing a Longformer configuration
+            configuration = LongformerConfig()
+
+            # Initializing a model from the configuration
+            model = LongformerModel(configuration)
+
+            # Accessing the model configuration
+            configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
+    """
+    pretrained_config_archive_map = LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP
+    model_type = "longformer"
+
+    def __init__(self, attention_window: Union[List[int], int] = 512, sep_token_id: int = 2, **kwargs):
+        super().__init__(**kwargs)
+        self.attention_window = attention_window
+        self.sep_token_id = sep_token_id

src/transformers/configuration_reformer.py

@@ -110,10 +110,10 @@ class ReformerConfig(PretrainedConfig):
                 Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
             num_attention_heads (:obj:`int`, optional, defaults to 12):
                 Number of attention heads for each attention layer in the Transformer encoder.
-            num_buckets (:obj:`int` or :obj:`list(int)`, optional, defaults to `64`):
+            num_buckets (:obj:`int` or :obj:`list(int)`, optional, defaults to `None`):
                 Number of buckets, the key query vectors can be "hashed into" using the locality sensitive hashing scheme. Each query key vector is hashed into a hash in `1, ..., num_buckets`.
                 The number of buckets can also be factorized into a list for improved memory complexity. In this case, each query key vector is hashed into a hash in `1-1, 1-2, ..., num_buckets[0]-1, ..., num_buckets[0]-num_buckets[1]` if `num_buckets` is factorized into two factors.
-                The number of buckets (or the product the factors) should approximately equal sequence length / lsh_chunk_length.
+                The number of buckets (or the product the factors) should approximately equal sequence length / lsh_chunk_length. If `num_buckets` is set to `None`, a good value for `num_buckets` is calculated on the fly.
             num_hashes (:obj:`int`, optional, defaults to 1):
                 Number of hashing rounds (e.g. number of random rotations) in Local Sensitive Hashing scheme.
                 The higher `num_hashes`, the more accurate the `LSHSelfAttention` becomes, but also the more memory and time intensive the hashing becomes.
@@ -172,7 +172,7 @@ class ReformerConfig(PretrainedConfig):
         lsh_num_chunks_after=0,
         max_position_embeddings=4096,
         num_attention_heads=2,
-        num_buckets=32,
+        num_buckets=None,
         num_hashes=1,
         pad_token_id=0,
         vocab_size=320,

src/transformers/convert_graph_to_onnx.py

@@ -103,7 +103,7 @@ def load_graph_from_args(framework: str, model: str, tokenizer: Optional[str] =
     print("Loading pipeline (model: {}, tokenizer: {})".format(model, tokenizer))
 
     # Allocate tokenizer and model
-    return pipeline("feature-extraction", model=model, framework=framework)
+    return pipeline("feature-extraction", model=model, tokenizer=tokenizer, framework=framework)
 
 
 def convert_pytorch(nlp: Pipeline, opset: int, output: str, use_external_format: bool):

src/transformers/convert_longformer_original_pytorch_lightning_to_pytorch.py

@@ -0,0 +1,86 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert RoBERTa checkpoint."""
+
+
+import argparse
+
+import pytorch_lightning as pl
+import torch
+
+from transformers.modeling_longformer import LongformerForQuestionAnswering, LongformerModel
+
+
+class LightningModel(pl.LightningModule):
+    def __init__(self, model):
+        super().__init__()
+        self.model = model
+        self.num_labels = 2
+        self.qa_outputs = torch.nn.Linear(self.model.config.hidden_size, self.num_labels)
+
+    # implement only because lighning requires to do so
+    def forward(self):
+        pass
+
+
+def convert_longformer_qa_checkpoint_to_pytorch(
+    longformer_model: str, longformer_question_answering_ckpt_path: str, pytorch_dump_folder_path: str
+):
+
+    # load longformer model from model identifier
+    longformer = LongformerModel.from_pretrained(longformer_model)
+    lightning_model = LightningModel(longformer)
+
+    ckpt = torch.load(longformer_question_answering_ckpt_path, map_location=torch.device("cpu"))
+    lightning_model.load_state_dict(ckpt["state_dict"])
+
+    # init longformer question answering model
+    longformer_for_qa = LongformerForQuestionAnswering.from_pretrained(longformer_model)
+
+    # transfer weights
+    longformer_for_qa.longformer.load_state_dict(lightning_model.model.state_dict())
+    longformer_for_qa.qa_outputs.load_state_dict(lightning_model.qa_outputs.state_dict())
+    longformer_for_qa.eval()
+
+    # save model
+    longformer_for_qa.save_pretrained(pytorch_dump_folder_path)
+
+    print("Conversion succesful. Model saved under {}".format(pytorch_dump_folder_path))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--longformer_model",
+        default=None,
+        type=str,
+        required=True,
+        help="model identifier of longformer. Should be either `longformer-base-4096` or `longformer-large-4096`.",
+    )
+    parser.add_argument(
+        "--longformer_question_answering_ckpt_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path the official PyTorch Lighning Checkpoint.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_longformer_qa_checkpoint_to_pytorch(
+        args.longformer_model, args.longformer_question_answering_ckpt_path, args.pytorch_dump_folder_path
+    )

src/transformers/data/datasets/glue.py

@@ -2,7 +2,8 @@ import logging
 import os
 import time
 from dataclasses import dataclass, field
-from typing import List, Optional
+from enum import Enum
+from typing import List, Optional, Union
 
 import torch
 from filelock import FileLock
@@ -47,6 +48,12 @@ class GlueDataTrainingArguments:
         self.task_name = self.task_name.lower()
 
 
+class Split(Enum):
+    train = "train"
+    dev = "dev"
+    test = "test"
+
+
 class GlueDataset(Dataset):
     """
     This will be superseded by a framework-agnostic approach
@@ -62,16 +69,21 @@ class GlueDataset(Dataset):
         args: GlueDataTrainingArguments,
         tokenizer: PreTrainedTokenizer,
         limit_length: Optional[int] = None,
-        evaluate=False,
+        mode: Union[str, Split] = Split.train,
     ):
         self.args = args
-        processor = glue_processors[args.task_name]()
+        self.processor = glue_processors[args.task_name]()
         self.output_mode = glue_output_modes[args.task_name]
+        if isinstance(mode, str):
+            try:
+                mode = Split[mode]
+            except KeyError:
+                raise KeyError("mode is not a valid split name")
         # Load data features from cache or dataset file
         cached_features_file = os.path.join(
             args.data_dir,
             "cached_{}_{}_{}_{}".format(
-                "dev" if evaluate else "train", tokenizer.__class__.__name__, str(args.max_seq_length), args.task_name,
+                mode.value, tokenizer.__class__.__name__, str(args.max_seq_length), args.task_name,
             ),
         )
 
@@ -88,7 +100,7 @@ class GlueDataset(Dataset):
                 )
             else:
                 logger.info(f"Creating features from dataset file at {args.data_dir}")
-                label_list = processor.get_labels()
+                label_list = self.processor.get_labels()
                 if args.task_name in ["mnli", "mnli-mm"] and tokenizer.__class__ in (
                     RobertaTokenizer,
                     RobertaTokenizerFast,
@@ -96,11 +108,12 @@ class GlueDataset(Dataset):
                 ):
                     # HACK(label indices are swapped in RoBERTa pretrained model)
                     label_list[1], label_list[2] = label_list[2], label_list[1]
-                examples = (
-                    processor.get_dev_examples(args.data_dir)
-                    if evaluate
-                    else processor.get_train_examples(args.data_dir)
-                )
+                if mode == Split.dev:
+                    examples = self.processor.get_dev_examples(args.data_dir)
+                elif mode == Split.test:
+                    examples = self.processor.get_test_examples(args.data_dir)
+                else:
+                    examples = self.processor.get_train_examples(args.data_dir)
                 if limit_length is not None:
                     examples = examples[:limit_length]
                 self.features = glue_convert_examples_to_features(
@@ -122,3 +135,6 @@ class GlueDataset(Dataset):
 
     def __getitem__(self, i) -> InputFeatures:
         return self.features[i]
+
+    def get_labels(self):
+        return self.processor.get_labels()

src/transformers/data/processors/glue.py

@@ -126,7 +126,9 @@ def _glue_convert_examples_to_features(
 
     label_map = {label: i for i, label in enumerate(label_list)}
 
-    def label_from_example(example: InputExample) -> Union[int, float]:
+    def label_from_example(example: InputExample) -> Union[int, float, None]:
+        if example.label is None:
+            return None
         if output_mode == "classification":
             return label_map[example.label]
         elif output_mode == "regression":
@@ -180,12 +182,16 @@ class MrpcProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["0", "1"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -193,7 +199,7 @@ class MrpcProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, i)
             text_a = line[3]
             text_b = line[4]
-            label = line[0]
+            label = None if set_type == "test" else line[0]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 
@@ -218,12 +224,16 @@ class MnliProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")), "dev_matched")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test_matched.tsv")), "test_matched")
+
     def get_labels(self):
         """See base class."""
         return ["contradiction", "entailment", "neutral"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -231,7 +241,7 @@ class MnliProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, line[0])
             text_a = line[8]
             text_b = line[9]
-            label = line[-1]
+            label = None if set_type.startswith("test") else line[-1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 
@@ -241,7 +251,11 @@ class MnliMismatchedProcessor(MnliProcessor):
 
     def get_dev_examples(self, data_dir):
         """See base class."""
-        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev_mismatched.tsv")), "dev_matched")
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev_mismatched.tsv")), "dev_mismatched")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test_mismatched.tsv")), "test_mismatched")
 
 
 class ColaProcessor(DataProcessor):
@@ -264,17 +278,25 @@ class ColaProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["0", "1"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
+        test_mode = set_type == "test"
+        if test_mode:
+            lines = lines[1:]
+        text_index = 1 if test_mode else 3
         examples = []
         for (i, line) in enumerate(lines):
             guid = "%s-%s" % (set_type, i)
-            text_a = line[3]
-            label = line[1]
+            text_a = line[text_index]
+            label = None if test_mode else line[1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
         return examples
 
@@ -299,19 +321,23 @@ class Sst2Processor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["0", "1"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
                 continue
             guid = "%s-%s" % (set_type, i)
             text_a = line[0]
-            label = line[1]
+            label = None if set_type == "test" else line[1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
         return examples
 
@@ -336,12 +362,16 @@ class StsbProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return [None]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -349,7 +379,7 @@ class StsbProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, line[0])
             text_a = line[7]
             text_b = line[8]
-            label = line[-1]
+            label = None if set_type == "test" else line[-1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 
@@ -374,21 +404,28 @@ class QqpProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["0", "1"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
+        test_mode = set_type == "test"
+        q1_index = 1 if test_mode else 3
+        q2_index = 2 if test_mode else 4
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
                 continue
             guid = "%s-%s" % (set_type, line[0])
             try:
-                text_a = line[3]
-                text_b = line[4]
-                label = line[5]
+                text_a = line[q1_index]
+                text_b = line[q2_index]
+                label = None if test_mode else line[5]
             except IndexError:
                 continue
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
@@ -413,14 +450,18 @@ class QnliProcessor(DataProcessor):
 
     def get_dev_examples(self, data_dir):
         """See base class."""
-        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev_matched")
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
 
     def get_labels(self):
         """See base class."""
         return ["entailment", "not_entailment"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -428,7 +469,7 @@ class QnliProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, line[0])
             text_a = line[1]
             text_b = line[2]
-            label = line[-1]
+            label = None if set_type == "test" else line[-1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 
@@ -453,12 +494,16 @@ class RteProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["entailment", "not_entailment"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -466,7 +511,7 @@ class RteProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, line[0])
             text_a = line[1]
             text_b = line[2]
-            label = line[-1]
+            label = None if set_type == "test" else line[-1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 
@@ -491,12 +536,16 @@ class WnliProcessor(DataProcessor):
         """See base class."""
         return self._create_examples(self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
 
+    def get_test_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(self._read_tsv(os.path.join(data_dir, "test.tsv")), "test")
+
     def get_labels(self):
         """See base class."""
         return ["0", "1"]
 
     def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
+        """Creates examples for the training, dev and test sets."""
         examples = []
         for (i, line) in enumerate(lines):
             if i == 0:
@@ -504,7 +553,7 @@ class WnliProcessor(DataProcessor):
             guid = "%s-%s" % (set_type, line[0])
             text_a = line[1]
             text_b = line[2]
-            label = line[-1]
+            label = None if set_type == "test" else line[-1]
             examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
         return examples
 

src/transformers/data/processors/utils.py

@@ -98,6 +98,10 @@ class DataProcessor:
         """Gets a collection of `InputExample`s for the dev set."""
         raise NotImplementedError()
 
+    def get_test_examples(self, data_dir):
+        """Gets a collection of `InputExample`s for the test set."""
+        raise NotImplementedError()
+
     def get_labels(self):
         """Gets the list of labels for this data set."""
         raise NotImplementedError()

src/transformers/file_utils.py

@@ -59,6 +59,7 @@ try:
 except (ImportError, AssertionError):
     _tf_available = False  # pylint: disable=invalid-name
 
+
 try:
     from torch.hub import _get_torch_home
 

src/transformers/hf_argparser.py

@@ -4,7 +4,7 @@ import sys
 from argparse import ArgumentParser
 from enum import Enum
 from pathlib import Path
-from typing import Any, Iterable, NewType, Tuple, Union
+from typing import Any, Iterable, List, NewType, Tuple, Union
 
 
 DataClass = NewType("DataClass", Any)
@@ -52,9 +52,13 @@ class HfArgumentParser(ArgumentParser):
                     "We will add compatibility when Python 3.9 is released."
                 )
             typestring = str(field.type)
-            for x in (int, float, str):
-                if typestring == f"typing.Union[{x.__name__}, NoneType]":
-                    field.type = x
+            for prim_type in (int, float, str):
+                for collection in (List,):
+                    if typestring == f"typing.Union[{collection[prim_type]}, NoneType]":
+                        field.type = collection[prim_type]
+                if typestring == f"typing.Union[{prim_type.__name__}, NoneType]":
+                    field.type = prim_type
+
             if isinstance(field.type, type) and issubclass(field.type, Enum):
                 kwargs["choices"] = list(field.type)
                 kwargs["type"] = field.type
@@ -65,6 +69,14 @@ class HfArgumentParser(ArgumentParser):
                 if field.default is True:
                     field_name = f"--no-{field.name}"
                     kwargs["dest"] = field.name
+            elif hasattr(field.type, "__origin__") and issubclass(field.type.__origin__, List):
+                kwargs["nargs"] = "+"
+                kwargs["type"] = field.type.__args__[0]
+                assert all(
+                    x == kwargs["type"] for x in field.type.__args__
+                ), "{} cannot be a List of mixed types".format(field.name)
+                if field.default_factory is not dataclasses.MISSING:
+                    kwargs["default"] = field.default_factory()
             else:
                 kwargs["type"] = field.type
                 if field.default is not dataclasses.MISSING:
@@ -126,6 +138,9 @@ class HfArgumentParser(ArgumentParser):
         if return_remaining_strings:
             return (*outputs, remaining_args)
         else:
+            if remaining_args:
+                raise ValueError(f"Some specified arguments are not used by the HfArgumentParser: {remaining_args}")
+
             return (*outputs,)
 
     def parse_json_file(self, json_file: str) -> Tuple[DataClass, ...]:

src/transformers/modeling_auto.py

@@ -30,6 +30,7 @@ from .configuration_auto import (
     EncoderDecoderConfig,
     FlaubertConfig,
     GPT2Config,
+    LongformerConfig,
     OpenAIGPTConfig,
     ReformerConfig,
     RobertaConfig,
@@ -87,6 +88,7 @@ from .modeling_electra import (
     ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
     ElectraForMaskedLM,
     ElectraForPreTraining,
+    ElectraForSequenceClassification,
     ElectraForTokenClassification,
     ElectraModel,
 )
@@ -99,6 +101,14 @@ from .modeling_flaubert import (
     FlaubertWithLMHeadModel,
 )
 from .modeling_gpt2 import GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2LMHeadModel, GPT2Model
+from .modeling_longformer import (
+    LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
+    LongformerForMaskedLM,
+    LongformerForQuestionAnswering,
+    LongformerForSequenceClassification,
+    LongformerForTokenClassification,
+    LongformerModel,
+)
 from .modeling_marian import MarianMTModel
 from .modeling_openai import OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OpenAIGPTLMHeadModel, OpenAIGPTModel
 from .modeling_reformer import ReformerModel, ReformerModelWithLMHead
@@ -162,6 +172,7 @@ ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict(
         FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
         ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -174,6 +185,7 @@ MODEL_MAPPING = OrderedDict(
         (CamembertConfig, CamembertModel),
         (XLMRobertaConfig, XLMRobertaModel),
         (BartConfig, BartModel),
+        (LongformerConfig, LongformerModel),
         (RobertaConfig, RobertaModel),
         (BertConfig, BertModel),
         (OpenAIGPTConfig, OpenAIGPTModel),
@@ -196,6 +208,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForMaskedLM),
         (XLMRobertaConfig, XLMRobertaForMaskedLM),
         (BartConfig, BartForConditionalGeneration),
+        (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
         (BertConfig, BertForPreTraining),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
@@ -218,6 +231,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (XLMRobertaConfig, XLMRobertaForMaskedLM),
         (MarianConfig, MarianMTModel),
         (BartConfig, BartForConditionalGeneration),
+        (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
         (BertConfig, BertForMaskedLM),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
@@ -240,11 +254,13 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForSequenceClassification),
         (XLMRobertaConfig, XLMRobertaForSequenceClassification),
         (BartConfig, BartForSequenceClassification),
+        (LongformerConfig, LongformerForSequenceClassification),
         (RobertaConfig, RobertaForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
         (FlaubertConfig, FlaubertForSequenceClassification),
         (XLMConfig, XLMForSequenceClassification),
+        (ElectraConfig, ElectraForSequenceClassification),
     ]
 )
 
@@ -252,6 +268,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
     [
         (DistilBertConfig, DistilBertForQuestionAnswering),
         (AlbertConfig, AlbertForQuestionAnswering),
+        (LongformerConfig, LongformerForQuestionAnswering),
         (RobertaConfig, RobertaForQuestionAnswering),
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
@@ -266,6 +283,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForTokenClassification),
         (XLMConfig, XLMForTokenClassification),
         (XLMRobertaConfig, XLMRobertaForTokenClassification),
+        (LongformerConfig, LongformerForTokenClassification),
         (RobertaConfig, RobertaForTokenClassification),
         (BertConfig, BertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
@@ -313,6 +331,7 @@ class AutoModel:
                 The model class to instantiate is selected based on the configuration class:
 
                 - isInstance of `distilbert` configuration class: :class:`~transformers.DistilBertModel` (DistilBERT model)
+                - isInstance of `longformer` configuration class: :class:`~transformers.LongformerModel` (Longformer model)
                 - isInstance of `roberta` configuration class: :class:`~transformers.RobertaModel` (RoBERTa model)
                 - isInstance of `bert` configuration class: :class:`~transformers.BertModel` (Bert model)
                 - isInstance of `openai-gpt` configuration class: :class:`~transformers.OpenAIGPTModel` (OpenAI GPT model)
@@ -355,6 +374,7 @@ class AutoModel:
             - contains `albert`: :class:`~transformers.AlbertModel` (ALBERT model)
             - contains `camembert`: :class:`~transformers.CamembertModel` (CamemBERT model)
             - contains `xlm-roberta`: :class:`~transformers.XLMRobertaModel` (XLM-RoBERTa model)
+            - contains `longformer` :class:`~transformers.LongformerModel` (Longformer model)
             - contains `roberta`: :class:`~transformers.RobertaModel` (RoBERTa model)
             - contains `bert`: :class:`~transformers.BertModel` (Bert model)
             - contains `openai-gpt`: :class:`~transformers.OpenAIGPTModel` (OpenAI GPT model)
@@ -388,7 +408,7 @@ class AutoModel:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
 
@@ -407,7 +427,7 @@ class AutoModel:
                 The proxies are used on each request.
 
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
 
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.
@@ -463,6 +483,7 @@ class AutoModelForPreTraining:
                 The model class to instantiate is selected based on the configuration class:
 
                 - isInstance of `distilbert` configuration class: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
+                - isInstance of `longformer` configuration class: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
                 - isInstance of `roberta` configuration class: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
                 - isInstance of `bert` configuration class: :class:`~transformers.BertForPreTraining` (Bert model)
                 - isInstance of `openai-gpt` configuration class: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
@@ -504,6 +525,7 @@ class AutoModelForPreTraining:
             - contains `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
             - contains `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
             - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
             - contains `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
             - contains `bert`: :class:`~transformers.BertForPreTraining` (Bert model)
             - contains `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
@@ -536,7 +558,7 @@ class AutoModelForPreTraining:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
             cache_dir: (`optional`) string:
@@ -550,7 +572,7 @@ class AutoModelForPreTraining:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.
 
@@ -606,6 +628,7 @@ class AutoModelWithLMHead:
                 The model class to instantiate is selected based on the configuration class:
 
                 - isInstance of `distilbert` configuration class: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
+                - isInstance of `longformer` configuration class: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
                 - isInstance of `roberta` configuration class: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
                 - isInstance of `bert` configuration class: :class:`~transformers.BertForMaskedLM` (Bert model)
                 - isInstance of `openai-gpt` configuration class: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
@@ -648,6 +671,7 @@ class AutoModelWithLMHead:
             - contains `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
             - contains `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
             - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
             - contains `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
             - contains `bert`: :class:`~transformers.BertForMaskedLM` (Bert model)
             - contains `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
@@ -680,7 +704,7 @@ class AutoModelWithLMHead:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
             cache_dir: (`optional`) string:
@@ -694,7 +718,7 @@ class AutoModelWithLMHead:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.
 
@@ -819,7 +843,7 @@ class AutoModelForSequenceClassification:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
 
@@ -838,7 +862,7 @@ class AutoModelForSequenceClassification:
                 The proxies are used on each request.
 
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
 
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.
@@ -961,7 +985,7 @@ class AutoModelForQuestionAnswering:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
 
@@ -977,7 +1001,7 @@ class AutoModelForQuestionAnswering:
                 The proxies are used on each request.
 
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
 
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.
@@ -1106,7 +1130,7 @@ class AutoModelForTokenClassification:
                 - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
 
             state_dict: (`optional`) dict:
-                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                an optional state dictionary for the model to use instead of a state dictionary loaded from saved weights file.
                 This option can be used if you want to create a model from a pretrained configuration but load your own weights.
                 In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
 
@@ -1122,7 +1146,7 @@ class AutoModelForTokenClassification:
                 The proxies are used on each request.
 
             output_loading_info: (`optional`) boolean:
-                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+                Set to ``True`` to also return a dictionary containing missing keys, unexpected keys and error messages.
 
             kwargs: (`optional`) Remaining dictionary of keyword arguments:
                 These arguments will be passed to the configuration and the model.

src/transformers/modeling_bert.py

@@ -19,7 +19,6 @@
 import logging
 import math
 import os
-from functools import partial
 
 import torch
 from torch import nn
@@ -154,7 +153,7 @@ class BertEmbeddings(nn.Module):
         # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
         # any TensorFlow checkpoint file
         self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.dropout = partial(nn.functional.dropout, p=config.hidden_dropout_prob)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
         if input_ids is not None:
@@ -177,7 +176,7 @@ class BertEmbeddings(nn.Module):
 
         embeddings = inputs_embeds + position_embeddings + token_type_embeddings
         embeddings = self.LayerNorm(embeddings)
-        embeddings = self.dropout(embeddings, training=self.training)
+        embeddings = self.dropout(embeddings)
         return embeddings
 
 
@@ -194,26 +193,17 @@ class BertSelfAttention(nn.Module):
         self.num_attention_heads = config.num_attention_heads
         self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
         self.all_head_size = self.num_attention_heads * self.attention_head_size
-        self.scale_factor = 1. / math.sqrt(self.attention_head_size)
 
         self.query = nn.Linear(config.hidden_size, self.all_head_size)
         self.key = nn.Linear(config.hidden_size, self.all_head_size)
         self.value = nn.Linear(config.hidden_size, self.all_head_size)
 
-        self.qkv_weight = nn.Parameter(torch.cat(
-            (self.query.weight, self.key.weight, self.value.weight), dim=0
-        ).t().contiguous())
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
 
-        self.qkv_bias = nn.Parameter(
-            torch.cat((self.query.bias, self.key.bias, self.value.bias), dim=0)
-        )
-
-        self.dropout = partial(nn.functional.dropout, p=config.attention_probs_dropout_prob)
-
-    # def transpose_for_scores(self, x):
-    #     new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
-    #     x = x.view(*new_x_shape)
-    #     return x.permute(0, 2, 1, 3)
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
 
     def forward(
         self,
@@ -223,45 +213,49 @@ class BertSelfAttention(nn.Module):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
     ):
+        mixed_query_layer = self.query(hidden_states)
+
         # If this is instantiated as a cross-attention module, the keys
         # and values come from an encoder; the attention mask needs to be
         # such that the encoder's padding tokens are not attended to.
         if encoder_hidden_states is not None:
-            mixed_query_layer = self.query(hidden_states)
             mixed_key_layer = self.key(encoder_hidden_states)
             mixed_value_layer = self.value(encoder_hidden_states)
             attention_mask = encoder_attention_mask
         else:
-            # Compute qkv_bias + (hidden_states @ qkv_weight)
-            qkv = torch.matmul(hidden_states, self.qkv_weight) + self.qkv_bias
-            mixed_query_layer, mixed_key_layer, mixed_value_layer = qkv.chunk(3, dim=2)
+            mixed_key_layer = self.key(hidden_states)
+            mixed_value_layer = self.value(hidden_states)
 
-        attention_scores = torch.bmm(mixed_query_layer, mixed_key_layer.transpose(2, 1))
-        attention_scores = attention_scores * self.scale_factor
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
 
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
         if attention_mask is not None:
             # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
-            attention_scores = attention_scores + attention_mask.squeeze(2)
+            attention_scores = attention_scores + attention_mask
 
         # Normalize the attention scores to probabilities.
-        attention_probs = torch.softmax(attention_scores, dim=-1)
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
 
         # This is actually dropping out entire tokens to attend to, which might
         # seem a bit unusual, but is taken from the original Transformer paper.
-        attention_probs = self.dropout(attention_probs, training=self.training)
+        attention_probs = self.dropout(attention_probs)
 
         # Mask heads if we want to
         if head_mask is not None:
-            bs = attention_probs.size(0)
-
-            # Split attention_probs
-            attention_probs = attention_probs.view(bs, -1, self.num_attention_heads, self.attention_head_size)
             attention_probs = attention_probs * head_mask
-            attention_probs = attention_probs.flatten(2)
 
-        context_layer = torch.bmm(attention_probs, mixed_value_layer)
+        context_layer = torch.matmul(attention_probs, value_layer)
 
-        return (context_layer, attention_probs) if self.output_attentions else (context_layer,)
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
+        return outputs
 
 
 class BertSelfOutput(nn.Module):
@@ -269,11 +263,11 @@ class BertSelfOutput(nn.Module):
         super().__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
         self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.dropout = partial(nn.functional.dropout, p=config.hidden_dropout_prob)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
         hidden_states = self.dense(hidden_states)
-        hidden_states = self.dropout(hidden_states, training=self.training)
+        hidden_states = self.dropout(hidden_states)
         hidden_states = self.LayerNorm(hidden_states + input_tensor)
         return hidden_states
 
@@ -320,7 +314,8 @@ class BertAttention(nn.Module):
             hidden_states, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask
         )
         attention_output = self.output(self_outputs[0], hidden_states)
-        return (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
 
 
 class BertIntermediate(nn.Module):
@@ -343,11 +338,11 @@ class BertOutput(nn.Module):
         super().__init__()
         self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
         self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
-        self.dropout = partial(nn.functional.dropout, p=config.hidden_dropout_prob)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
         hidden_states = self.dense(hidden_states)
-        hidden_states = self.dropout(hidden_states, training=self.training)
+        hidden_states = self.dropout(hidden_states)
         hidden_states = self.LayerNorm(hidden_states + input_tensor)
         return hidden_states
 
@@ -432,13 +427,15 @@ class BertPooler(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
 
     def forward(self, hidden_states):
         # We "pool" the model by simply taking the hidden state corresponding
         # to the first token.
         first_token_tensor = hidden_states[:, 0]
         pooled_output = self.dense(first_token_tensor)
-        return torch.tanh(pooled_output)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
 
 
 class BertPredictionHeadTransform(nn.Module):

src/transformers/modeling_ctrl.py

@@ -208,9 +208,11 @@ CTRL_START_DOCSTRING = r"""
 
 CTRL_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only input_ids that do not have their past calculated should be passed as input_ids.
 
             Indices can be obtained using :class:`transformers.CTRLTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -220,9 +222,7 @@ CTRL_INPUTS_DOCSTRING = r"""
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids`
-            (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)`
-            instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
+            The input_ids which have their past given to this model should not be passed as input ids as they have already been computed.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -233,7 +233,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -246,7 +245,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
         input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
-            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
             If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
@@ -344,16 +342,6 @@ class CTRLModel(CTRLPreTrainedModel):
 
         """
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:

src/transformers/modeling_electra.py

@@ -3,6 +3,7 @@ import os
 
 import torch
 import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
 
 from .activations import get_activation
 from .configuration_electra import ElectraConfig
@@ -330,6 +331,112 @@ class ElectraModel(ElectraPreTrainedModel):
         return hidden_states
 
 
+class ElectraClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = get_activation("gelu")(x)  # although BERT uses tanh here, it seems Electra authors used gelu here
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+@add_start_docstrings(
+    """ELECTRA Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ELECTRA_START_DOCSTRING,
+)
+class ElectraForSequenceClassification(ElectraPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.electra = ElectraModel(config)
+        self.classifier = ElectraClassificationHead(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(ELECTRA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        from transformers import BertTokenizer, BertForSequenceClassification
+        import torch
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+
+        loss, logits = outputs[:2]
+
+        """
+        discriminator_hidden_states = self.electra(
+            input_ids, attention_mask, token_type_ids, position_ids, head_mask, inputs_embeds
+        )
+
+        sequence_output = discriminator_hidden_states[0]
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + discriminator_hidden_states[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
 @add_start_docstrings(
     """
     Electra model with a binary classification head on top as used during pre-training for identifying generated

src/transformers/modeling_encoder_decoder.py

@@ -125,7 +125,7 @@ class EncoderDecoderModel(PreTrainedModel):
 
         Examples::
 
-            from tranformers import EncoderDecoder
+            from transformers import EncoderDecoder
 
             model = EncoderDecoder.from_encoder_decoder_pretrained('bert-base-uncased', 'bert-base-uncased') # initialize Bert2Bert
         """

src/transformers/modeling_gpt2.py

@@ -286,9 +286,11 @@ GPT2_START_DOCSTRING = r"""
 
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`.
 
             Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -299,7 +301,7 @@ GPT2_INPUTS_DOCSTRING = r"""
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids` (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)` instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
+            The `input_ids` which have their past given to this model should not be passed as `input_ids` as they have already been computed.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -311,8 +313,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
-
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
@@ -324,7 +324,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
         input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
-            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
             If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
@@ -410,16 +409,6 @@ class GPT2Model(GPT2PreTrainedModel):
 
         """
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:

src/transformers/modeling_mmbt.py

@@ -149,7 +149,7 @@ MMBT_INPUTS_DOCSTRING = r"""    Inputs:
     MMBT_START_DOCSTRING,
     MMBT_INPUTS_DOCSTRING,
 )
-class MMBTModel(ModuleUtilsMixin):
+class MMBTModel(nn.Module, ModuleUtilsMixin):
     r"""
         Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
             **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``

src/transformers/modeling_reformer.py

@@ -283,6 +283,8 @@ class EfficientAttentionMixin:
 class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
     def __init__(self, config):
         super().__init__()
+        self.config = config
+
         self.chunk_length = config.lsh_attn_chunk_length
         self.num_hashes = config.num_hashes
         self.num_buckets = config.num_buckets
@@ -532,15 +534,22 @@ class LSHSelfAttention(nn.Module, EfficientAttentionMixin):
         return sorted_bucket_idx, undo_sorted_bucket_idx
 
     def _set_num_buckets(self, sequence_length):
-        # recommended `num_buckets` from paper
-        num_buckets = 2 * sequence_length // self.chunk_length
+        # `num_buckets` should be set to 2 * sequence_length // chunk_length as recommended in paper
+        num_buckets_pow_2 = (2 * (sequence_length // self.chunk_length)).bit_length() - 1
+        # make sure buckets are power of 2
+        num_buckets = 2 ** num_buckets_pow_2
 
         # factorize `num_buckets` if `num_buckets` becomes too large
-        num_buckets_limit = max(int((self.max_position_embeddings // self.chunk_length) ** (0.5)), self.chunk_length,)
-        if num_buckets > 2 * num_buckets_limit:
-            num_buckets = [num_buckets_limit, num_buckets // num_buckets_limit + 1]
+        num_buckets_limit = 2 * max(
+            int((self.max_position_embeddings // self.chunk_length) ** (0.5)), self.chunk_length,
+        )
+        if num_buckets > num_buckets_limit:
+            num_buckets = [2 ** (num_buckets_pow_2 // 2), 2 ** (num_buckets_pow_2 - num_buckets_pow_2 // 2)]
 
         logger.warning("config.num_buckets is not set. Setting config.num_buckets to {}...".format(num_buckets))
+
+        # set num buckets in config to be properly saved
+        self.config.num_buckets = num_buckets
         self.num_buckets = num_buckets
 
     def _attend(

src/transformers/modeling_t5.py

@@ -745,7 +745,7 @@ class T5Stack(T5PreTrainedModel):
                 # layer_outputs = hidden-states, key-value-states (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
                 position_bias = layer_outputs[3 if self.output_attentions else 2]
                 if self.is_decoder and encoder_hidden_states is not None:
-                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 3]
+                    encoder_decoder_position_bias = layer_outputs[5 if self.output_attentions else 3]
             # append next layer key value states
             present_key_value_states = present_key_value_states + (present_key_value_state,)
 

src/transformers/modeling_tf_ctrl.py

@@ -411,9 +411,12 @@ CTRL_START_DOCSTRING = r"""
 
 CTRL_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
+
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only input_ids that do not have their past calculated should be passed as input_ids (see `past`).
 
             Indices can be obtained using :class:`transformers.CTRLTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -423,9 +426,8 @@ CTRL_INPUTS_DOCSTRING = r"""
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids`
-            (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)`
-            instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
+            The token ids which have their past given to this model
+            should not be passed as input ids as they have already been computed.
         attention_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -436,7 +438,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -452,7 +453,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
-            If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
         use_cache (:obj:`bool`):
             If `use_cache` is True, `past` key value states are returned and
             can be used to speed up decoding (see `past`). Defaults to `True`.

src/transformers/modeling_tf_gpt2.py

@@ -284,16 +284,6 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         else:
             input_ids = inputs
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
@@ -431,9 +421,11 @@ GPT2_START_DOCSTRING = r"""
 
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`.
 
             Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -442,8 +434,9 @@ GPT2_INPUTS_DOCSTRING = r"""
             `What are input IDs? <../glossary.html#input-ids>`__
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
+            (see `past` output below). Can be used to speed up sequential decoding.
+            The token ids which have their past given to this model
+            should not be passed as `input_ids` as they have already been computed.
         attention_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -454,7 +447,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -470,7 +462,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
-            If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
@@ -639,7 +630,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
+            should not be passed as `input_ids` as they have already been computed.
         hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.

src/transformers/modeling_tf_t5.py

@@ -682,7 +682,7 @@ class TFT5MainLayer(tf.keras.layers.Layer):
                 # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
                 position_bias = layer_outputs[3 if self.output_attentions else 2]
                 if self.is_decoder and encoder_hidden_states is not None:
-                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 3]
+                    encoder_decoder_position_bias = layer_outputs[5 if self.output_attentions else 3]
             # append next layer key value states
             present_key_value_states = present_key_value_states + (present_key_value_state,)
 

src/transformers/modeling_tf_xlnet.py

@@ -586,8 +586,9 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
 
         if data_mask is not None:
             # all mems can be attended to
-            mems_mask = tf.zeros([shape_list(data_mask)[0], mlen, bsz], dtype=dtype_float)
-            data_mask = tf.concat([mems_mask, data_mask], axis=1)
+            if mlen > 0:
+                mems_mask = tf.zeros([shape_list(data_mask)[0], mlen, bsz], dtype=dtype_float)
+                data_mask = tf.concat([mems_mask, data_mask], axis=1)
             if attn_mask is None:
                 attn_mask = data_mask[:, :, :, None]
             else:
@@ -598,7 +599,8 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
 
         if attn_mask is not None:
             non_tgt_mask = -tf.eye(qlen, dtype=dtype_float)
-            non_tgt_mask = tf.concat([tf.zeros([qlen, mlen], dtype=dtype_float), non_tgt_mask], axis=-1)
+            if mlen > 0:
+                non_tgt_mask = tf.concat([tf.zeros([qlen, mlen], dtype=dtype_float), non_tgt_mask], axis=-1)
             non_tgt_mask = tf.cast((attn_mask + non_tgt_mask[:, :, None, None]) > 0, dtype=dtype_float)
         else:
             non_tgt_mask = None
@@ -621,8 +623,11 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         # Segment embedding
         if token_type_ids is not None:
             # Convert `token_type_ids` to one-hot `seg_mat`
-            mem_pad = tf.zeros([mlen, bsz], dtype=tf.int32)
-            cat_ids = tf.concat([mem_pad, token_type_ids], 0)
+            if mlen > 0:
+                mem_pad = tf.zeros([mlen, bsz], dtype=tf.int32)
+                cat_ids = tf.concat([mem_pad, token_type_ids], 0)
+            else:
+                cat_ids = token_type_ids
 
             # `1` indicates not in the same segment [qlen x klen x bsz]
             seg_mat = tf.cast(tf.logical_not(tf.equal(token_type_ids[:, None], cat_ids[None, :])), tf.int32)
@@ -640,14 +645,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads] (a head_mask for each layer)
         # and head_mask is converted to shape [num_hidden_layers x qlen x klen x bsz x n_head]
         if head_mask is not None:
-            if head_mask.dim() == 1:
-                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0).unsqueeze(0)
-                head_mask = head_mask.expand(self.n_layer, -1, -1, -1, -1)
-            elif head_mask.dim() == 2:
-                head_mask = head_mask.unsqueeze(1).unsqueeze(1).unsqueeze(1)
-            head_mask = head_mask.to(
-                dtype=next(self.parameters()).dtype
-            )  # switch to fload if need + fp16 compatibility
+            raise NotImplementedError
         else:
             head_mask = [None] * self.n_layer
 

src/transformers/modeling_utils.py

@@ -17,7 +17,7 @@
 import inspect
 import logging
 import os
-from typing import Callable, List, Tuple
+from typing import Callable, Dict, Iterable, List, Optional, Tuple
 
 import torch
 from torch import Tensor, device, dtype, nn
@@ -164,7 +164,7 @@ class ModuleUtilsMixin:
 
         return encoder_extended_attention_mask
 
-    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: tuple, device: device):
+    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple, device: device) -> Tensor:
         """Makes broadcastable attention mask and causal mask so that future and maked tokens are ignored.
 
         Arguments:
@@ -208,7 +208,7 @@ class ModuleUtilsMixin:
         extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
         return extended_attention_mask
 
-    def get_head_mask(self, head_mask, num_hidden_layers, is_attention_chunked=False):
+    def get_head_mask(self, head_mask: Tensor, num_hidden_layers: int, is_attention_chunked: bool = False) -> Tensor:
         """
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -302,7 +302,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         else:
             raise NotImplementedError
 
-    def set_input_embeddings(self, value):
+    def set_input_embeddings(self, value: nn.Module):
         """
         Set model's input embeddings
 
@@ -354,7 +354,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
             output_embeddings.out_features = input_embeddings.num_embeddings
 
-    def resize_token_embeddings(self, new_num_tokens=None):
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None):
         """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
         Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
 
@@ -387,18 +387,22 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         self.set_input_embeddings(new_embeddings)
         return self.get_input_embeddings()
 
-    def _get_resized_embeddings(self, old_embeddings, new_num_tokens=None):
+    def _get_resized_embeddings(
+        self, old_embeddings: torch.nn.Embedding, new_num_tokens: Optional[int] = None
+    ) -> torch.nn.Embedding:
         """ Build a resized Embedding Module from a provided token Embedding Module.
             Increasing the size will add newly initialized vectors at the end
             Reducing the size will remove vectors from the end
 
         Args:
+            old_embeddings: ``torch.nn.Embedding``
+                Old embeddings to be resized.
             new_num_tokens: (`optional`) int
                 New number of tokens in the embedding matrix.
                 Increasing the size will add newly initialized vectors at the end
                 Reducing the size will remove vectors from the end
                 If not provided or None: return the provided token Embedding Module.
-        Return: ``torch.nn.Embeddings``
+        Return: ``torch.nn.Embedding``
             Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
         """
         if new_num_tokens is None:
@@ -433,7 +437,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
         # Tie weights if needed
         self.tie_weights()
 
-    def prune_heads(self, heads_to_prune):
+    def prune_heads(self, heads_to_prune: Dict):
         """ Prunes heads of the base model.
 
             Arguments:
@@ -770,7 +774,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
             import torch_xla.core.xla_model as xm
 
             model = xm.send_cpu_data_to_device(model, xm.xla_device())
-            model = model.to(xm.xla_device())
+            model.to(xm.xla_device())
 
         return model
 
@@ -801,28 +805,28 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin):
     @torch.no_grad()
     def generate(
         self,
-        input_ids=None,
-        max_length=None,
-        min_length=None,
-        do_sample=None,
-        early_stopping=None,
-        num_beams=None,
-        temperature=None,
-        top_k=None,
-        top_p=None,
-        repetition_penalty=None,
-        bad_words_ids=None,
-        bos_token_id=None,
-        pad_token_id=None,
-        eos_token_id=None,
-        length_penalty=None,
-        no_repeat_ngram_size=None,
-        num_return_sequences=None,
-        attention_mask=None,
-        decoder_start_token_id=None,
-        use_cache=None,
+        input_ids: Optional[torch.LongTensor] = None,
+        max_length: Optional[int] = None,
+        min_length: Optional[int] = None,
+        do_sample: Optional[bool] = None,
+        early_stopping: Optional[bool] = None,
+        num_beams: Optional[int] = None,
+        temperature: Optional[float] = None,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+        repetition_penalty: Optional[float] = None,
+        bad_words_ids: Optional[Iterable[int]] = None,
+        bos_token_id: Optional[int] = None,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+        length_penalty: Optional[float] = None,
+        no_repeat_ngram_size: Optional[int] = None,
+        num_return_sequences: Optional[int] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_start_token_id: Optional[int] = None,
+        use_cache: Optional[bool] = None,
         **model_specific_kwargs
-    ):
+    ) -> torch.LongTensor:
         r""" Generates sequences for models with a LM head. The method currently supports greedy decoding, beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.
 
         Adapted in part from `Facebook's XLM beam search code`_.
@@ -1606,7 +1610,7 @@ def calc_banned_ngram_tokens(prev_input_ids: Tensor, num_hypos: int, no_repeat_n
     return banned_tokens
 
 
-def calc_banned_bad_words_ids(prev_input_ids, bad_words_ids):
+def calc_banned_bad_words_ids(prev_input_ids: Iterable[int], bad_words_ids: Iterable[int]) -> Iterable[int]:
     banned_tokens = []
 
     def _tokens_match(prev_tokens, tokens):
@@ -1642,7 +1646,13 @@ def calc_banned_bad_words_ids(prev_input_ids, bad_words_ids):
     return banned_tokens
 
 
-def top_k_top_p_filtering(logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens_to_keep=1):
+def top_k_top_p_filtering(
+    logits: Tensor,
+    top_k: int = 0,
+    top_p: float = 1.0,
+    filter_value: float = -float("Inf"),
+    min_tokens_to_keep: int = 1,
+) -> Tensor:
     """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
         Args:
             logits: logits distribution shape (batch size, vocabulary size)

src/transformers/optimization.py

@@ -162,7 +162,7 @@ class AdamW(Optimizer):
                     bias_correction2 = 1.0 - beta2 ** state["step"]
                     step_size = step_size * math.sqrt(bias_correction2) / bias_correction1
 
-                p.data.addcdiv_(-step_size, exp_avg, denom)
+                p.data.addcdiv_(exp_avg, denom, value=-step_size)
 
                 # Just adding the square of the weights to the loss function is *not*
                 # the correct way of using L2 regularization/weight decay with Adam,

src/transformers/optimization_tf.py

@@ -75,7 +75,7 @@ def create_optimizer(init_lr, num_train_steps, num_warmup_steps, end_lr=0.0, opt
         beta_1=0.9,
         beta_2=0.999,
         epsilon=1e-6,
-        exclude_from_weight_decay=["layer_norm", "bias"],
+        exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"],
     )
 
     return optimizer

src/transformers/pipelines.py

@@ -1628,11 +1628,7 @@ SUPPORTED_TASKS = {
         "impl": SummarizationPipeline,
         "tf": TFAutoModelWithLMHead if is_tf_available() else None,
         "pt": AutoModelWithLMHead if is_torch_available() else None,
-        "default": {
-            "model": {"pt": "bart-large-cnn", "tf": "t5-small"},
-            "config": None,
-            "tokenizer": {"pt": ("bart-large-cnn", {"use_fast": False}), "tf": "t5-small"},
-        },
+        "default": {"model": {"pt": "bart-large-cnn", "tf": "t5-small"}, "config": None, "tokenizer": None},
     },
     "translation_en_to_fr": {
         "impl": TranslationPipeline,

src/transformers/tokenization_auto.py

@@ -29,6 +29,7 @@ from .configuration_auto import (
     ElectraConfig,
     FlaubertConfig,
     GPT2Config,
+    LongformerConfig,
     OpenAIGPTConfig,
     ReformerConfig,
     RobertaConfig,
@@ -50,6 +51,7 @@ from .tokenization_distilbert import DistilBertTokenizer, DistilBertTokenizerFas
 from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
+from .tokenization_longformer import LongformerTokenizer
 from .tokenization_marian import MarianTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_reformer import ReformerTokenizer
@@ -73,6 +75,7 @@ TOKENIZER_MAPPING = OrderedDict(
         (XLMRobertaConfig, (XLMRobertaTokenizer, None)),
         (MarianConfig, (MarianTokenizer, None)),
         (BartConfig, (BartTokenizer, None)),
+        (LongformerConfig, (LongformerTokenizer, None)),
         (RobertaConfig, (RobertaTokenizer, RobertaTokenizerFast)),
         (ReformerConfig, (ReformerTokenizer, None)),
         (ElectraConfig, (ElectraTokenizer, ElectraTokenizerFast)),
@@ -105,6 +108,7 @@ class AutoTokenizer:
             - contains `albert`: AlbertTokenizer (ALBERT model)
             - contains `camembert`: CamembertTokenizer (CamemBERT model)
             - contains `xlm-roberta`: XLMRobertaTokenizer (XLM-RoBERTa model)
+            - contains `longformer`: LongformerTokenizer (AllenAI Longformer model)
             - contains `roberta`: RobertaTokenizer (RoBERTa model)
             - contains `bert`: BertTokenizer (Bert model)
             - contains `openai-gpt`: OpenAIGPTTokenizer (OpenAI GPT model)
@@ -136,6 +140,7 @@ class AutoTokenizer:
             - contains `albert`: AlbertTokenizer (ALBERT model)
             - contains `camembert`: CamembertTokenizer (CamemBERT model)
             - contains `xlm-roberta`: XLMRobertaTokenizer (XLM-RoBERTa model)
+            - contains `longformer`: LongformerTokenizer (AllenAI Longformer model)
             - contains `roberta`: RobertaTokenizer (RoBERTa model)
             - contains `bert-base-japanese`: BertJapaneseTokenizer (Bert model)
             - contains `bert`: BertTokenizer (Bert model)

src/transformers/tokenization_bert.py

@@ -672,3 +672,33 @@ class BertTokenizerFast(PreTrainedTokenizerFast):
             output += token_ids_1 + [self.sep_token_id]
 
         return output
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A BERT sequence pair mask has the following format:
+
+        ::
+
+            0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+            | first sequence    | second sequence |
+
+        if token_ids_1 is None, only returns the first portion of the mask (0's).
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of ids.
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
+            sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]

src/transformers/tokenization_longformer.py

@@ -0,0 +1,60 @@
+# coding=utf-8
+# Copyright 2020 The Allen Institute for AI team and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+
+from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
+
+
+logger = logging.getLogger(__name__)
+
+
+# vocab and merges same as roberta
+vocab_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-vocab.json"
+merges_url = "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-merges.txt"
+_all_longformer_models = [
+    "allenai/longformer-base-4096",
+    "allenai/longformer-large-4096",
+    "allenai/longformer-large-4096-finetuned-triviaqa",
+    "allenai/longformer-base-4096-extra.pos.embd.only",
+    "allenai/longformer-large-4096-extra.pos.embd.only",
+]
+
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "allenai/longformer-base-4096": 4096,
+    "allenai/longformer-large-4096": 4096,
+    "allenai/longformer-large-4096-finetuned-triviaqa": 4096,
+    "allenai/longformer-base-4096-extra.pos.embd.only": 4096,
+    "allenai/longformer-large-4096-extra.pos.embd.only": 4096,
+}
+
+
+class LongformerTokenizer(RobertaTokenizer):
+    # merges and vocab same as Roberta
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    pretrained_vocab_files_map = {
+        "vocab_file": {m: vocab_url for m in _all_longformer_models},
+        "merges_file": {m: merges_url for m in _all_longformer_models},
+    }
+
+
+class LongformerTokenizerFast(RobertaTokenizerFast):
+    # merges and vocab same as Roberta
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    pretrained_vocab_files_map = {
+        "vocab_file": {m: vocab_url for m in _all_longformer_models},
+        "merges_file": {m: merges_url for m in _all_longformer_models},
+    }

src/transformers/tokenization_marian.py

@@ -1,7 +1,9 @@
 import json
 import re
 import warnings
-from typing import Dict, List, Optional, Union
+from pathlib import Path
+from shutil import copyfile
+from typing import Dict, List, Optional, Tuple, Union
 
 import sentencepiece
 
@@ -15,7 +17,7 @@ vocab_files_names = {
     "vocab": "vocab.json",
     "tokenizer_config_file": "tokenizer_config.json",
 }
-MODEL_NAMES = ("opus-mt-en-de",)  # TODO(SS): the only required constant is vocab_files_names
+MODEL_NAMES = ("opus-mt-en-de",)  # TODO(SS): delete this, the only required constant is vocab_files_names
 PRETRAINED_VOCAB_FILES_MAP = {
     k: {m: f"{S3_BUCKET_PREFIX}/Helsinki-NLP/{m}/{fname}" for m in MODEL_NAMES}
     for k, fname in vocab_files_names.items()
@@ -55,14 +57,16 @@ class MarianTokenizer(PreTrainedTokenizer):
         eos_token="</s>",
         pad_token="<pad>",
         max_len=512,
+        **kwargs,
     ):
 
         super().__init__(
-            # bos_token=bos_token,
+            # bos_token=bos_token,  unused. Start decoding with config.decoder_start_token_id
             max_len=max_len,
             eos_token=eos_token,
             unk_token=unk_token,
             pad_token=pad_token,
+            **kwargs,
         )
         self.encoder = load_json(vocab)
         if self.unk_token not in self.encoder:
@@ -72,21 +76,23 @@ class MarianTokenizer(PreTrainedTokenizer):
 
         self.source_lang = source_lang
         self.target_lang = target_lang
+        self.supported_language_codes: list = [k for k in self.encoder if k.startswith(">>") and k.endswith("<<")]
+        self.spm_files = [source_spm, target_spm]
 
         # load SentencePiece model for pre-processing
-        self.spm_source = sentencepiece.SentencePieceProcessor()
-        self.spm_source.Load(source_spm)
-
-        self.spm_target = sentencepiece.SentencePieceProcessor()
-        self.spm_target.Load(target_spm)
+        self.spm_source = load_spm(source_spm)
+        self.spm_target = load_spm(target_spm)
+        self.current_spm = self.spm_source
 
         # Multilingual target side: default to using first supported language code.
-        self.supported_language_codes: list = [k for k in self.encoder if k.startswith(">>") and k.endswith("<<")]
 
+        self._setup_normalizer()
+
+    def _setup_normalizer(self):
         try:
             from mosestokenizer import MosesPunctuationNormalizer
 
-            self.punc_normalizer = MosesPunctuationNormalizer(source_lang)
+            self.punc_normalizer = MosesPunctuationNormalizer(self.source_lang)
         except ImportError:
             warnings.warn("Recommended: pip install mosestokenizer")
             self.punc_normalizer = lambda x: x
@@ -176,6 +182,65 @@ class MarianTokenizer(PreTrainedTokenizer):
     def vocab_size(self) -> int:
         return len(self.encoder)
 
+    def save_vocabulary(self, save_directory: str) -> Tuple[str]:
+        """save vocab file to json and copy spm files from their original path."""
+        save_dir = Path(save_directory)
+        assert save_dir.is_dir(), f"{save_directory} should be a directory"
+        save_json(self.encoder, save_dir / self.vocab_files_names["vocab"])
+
+        for f in self.spm_files:
+            dest_path = save_dir / Path(f).name
+            if not dest_path.exists():
+                copyfile(f, save_dir / Path(f).name)
+        return tuple(save_dir / f for f in self.vocab_files_names)
+
+    def get_vocab(self) -> Dict:
+        vocab = self.encoder.copy()
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def __getstate__(self) -> Dict:
+        state = self.__dict__.copy()
+        state.update({k: None for k in ["spm_source", "spm_target", "current_spm", "punc_normalizer"]})
+        return state
+
+    def __setstate__(self, d: Dict) -> None:
+        self.__dict__ = d
+        self.spm_source, self.spm_target = (load_spm(f) for f in self.spm_files)
+        self.current_spm = self.spm_source
+        self._setup_normalizer()
+
+    def num_special_tokens_to_add(self, **unused):
+        """Just EOS"""
+        return 1
+
+    def _special_token_mask(self, seq):
+        all_special_ids = set(self.all_special_ids)  # call it once instead of inside list comp
+        all_special_ids.remove(self.unk_token_id)  # <unk> is only sometimes special
+        return [1 if x in all_special_ids else 0 for x in seq]
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List, token_ids_1: Optional[List] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """Get list where entries are [1] if a token is [eos] or [pad] else 0."""
+        if already_has_special_tokens:
+            return self._special_token_mask(token_ids_0)
+        elif token_ids_1 is None:
+            return self._special_token_mask(token_ids_0) + [1]
+        else:
+            return self._special_token_mask(token_ids_0 + token_ids_1) + [1]
+
+
+def load_spm(path: str) -> sentencepiece.SentencePieceProcessor:
+    spm = sentencepiece.SentencePieceProcessor()
+    spm.Load(path)
+    return spm
+
+
+def save_json(data, path: str) -> None:
+    with open(path, "w") as f:
+        json.dump(data, f, indent=2)
+
 
 def load_json(path: str) -> Union[Dict, List]:
     with open(path, "r") as f:

src/transformers/tokenization_roberta.py

@@ -199,7 +199,7 @@ class RobertaTokenizer(GPT2Tokenizer):
             if token_ids_1 is not None:
                 raise ValueError(
                     "You should not supply a second sequence if the provided sequence of "
-                    "ids is already formated with special tokens for the model."
+                    "ids is already formatted with special tokens for the model."
                 )
             return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
 
@@ -343,3 +343,27 @@ class RobertaTokenizerFast(GPT2TokenizerFast):
             return output
 
         return output + [self.eos_token_id] + token_ids_1 + [self.eos_token_id]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        RoBERTa does not make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of ids.
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of zeros.
+
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]

src/transformers/tokenization_utils.py

@@ -22,8 +22,10 @@ import logging
 import operator
 import os
 import re
+import warnings
 from collections import UserDict, defaultdict
 from contextlib import contextmanager
+from pathlib import Path
 from typing import Any, Dict, List, NamedTuple, Optional, Sequence, Tuple, Union
 
 from tokenizers import AddedToken as AddedTokenFast
@@ -502,6 +504,7 @@ class SpecialTokensMixin:
             if key in self.SPECIAL_TOKENS_ATTRIBUTES:
                 if key == "additional_special_tokens":
                     assert isinstance(value, (list, tuple)) and all(isinstance(t, str) for t in value)
+                    setattr(self, key, value)
                 elif isinstance(value, AddedTokenFast):
                     setattr(self, key, str(value))
                 elif isinstance(value, str):
@@ -771,26 +774,26 @@ class PreTrainedTokenizer(SpecialTokensMixin):
         raise NotImplementedError
 
     @property
-    def is_fast(self):
+    def is_fast(self) -> bool:
         return False
 
     @property
-    def max_len(self):
+    def max_len(self) -> int:
         """ Kept here for backward compatibility.
             Now renamed to `model_max_length` to avoid ambiguity.
         """
         return self.model_max_length
 
     @property
-    def max_len_single_sentence(self):
+    def max_len_single_sentence(self) -> int:
         return self.model_max_length - self.num_special_tokens_to_add(pair=False)
 
     @property
-    def max_len_sentences_pair(self):
+    def max_len_sentences_pair(self) -> int:
         return self.model_max_length - self.num_special_tokens_to_add(pair=True)
 
     @max_len_single_sentence.setter
-    def max_len_single_sentence(self, value):
+    def max_len_single_sentence(self, value) -> int:
         """ For backward compatibility, allow to try to setup 'max_len_single_sentence' """
         if value == self.model_max_length - self.num_special_tokens_to_add(pair=False):
             logger.warning(
@@ -802,7 +805,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
             )
 
     @max_len_sentences_pair.setter
-    def max_len_sentences_pair(self, value):
+    def max_len_sentences_pair(self, value) -> int:
         """ For backward compatibility, allow to try to setup 'max_len_sentences_pair' """
         if value == self.model_max_length - self.num_special_tokens_to_add(pair=True):
             logger.warning(
@@ -822,7 +825,14 @@ class PreTrainedTokenizer(SpecialTokensMixin):
         super().__init__(**kwargs)
 
         # For backward compatibility we fallback to set model_max_length from max_len if provided
-        model_max_length = model_max_length if model_max_length is not None else kwargs.pop("max_len", None)
+        if "max_len" in kwargs:
+            warnings.warn(
+                "Parameter max_len is deprecated and will be removed in a future release. "
+                "Use model_max_length instead.",
+                category=FutureWarning,
+            )
+
+            model_max_length = kwargs.pop("max_len")
         self.model_max_length = model_max_length if model_max_length is not None else VERY_LARGE_INTEGER
 
         # Padding side is right by default and overridden in subclasses. If specified in the kwargs, it is changed.
@@ -1077,7 +1087,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
 
         return tokenizer
 
-    def save_pretrained(self, save_directory):
+    def save_pretrained(self, save_directory: Union[str, Path]):
         """ Save the tokenizer vocabulary files together with:
                 - added tokens,
                 - special-tokens-to-class-attributes-mapping,
@@ -1089,6 +1099,10 @@ class PreTrainedTokenizer(SpecialTokensMixin):
             This method make sure the full tokenizer can then be re-loaded using the
             :func:`~transformers.PreTrainedTokenizer.from_pretrained` class method.
         """
+
+        # Ensure save_directory is a str
+        save_directory = str(save_directory)
+
         if not os.path.isdir(save_directory):
             logger.error("Saving directory ({}) should be a directory".format(save_directory))
             return
@@ -1118,7 +1132,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
 
         return vocab_files + (special_tokens_map_file, added_tokens_file)
 
-    def save_vocabulary(self, save_directory):
+    def save_vocabulary(self, save_directory: Union[str, Path]) -> Tuple[str]:
         """ Save the tokenizer vocabulary to a directory. This method does *NOT* save added tokens
             and special token mappings.
 
@@ -1128,7 +1142,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
         """
         raise NotImplementedError
 
-    def add_tokens(self, new_tokens):
+    def add_tokens(self, new_tokens: Union[str, List[str]]) -> int:
         """
         Add a list of new tokens to the tokenizer class. If the new tokens are not in the
         vocabulary, they are added to it with indices starting from length of the current vocabulary.
@@ -1156,7 +1170,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
         if not isinstance(new_tokens, list):
             new_tokens = [new_tokens]
 
-        to_add_tokens = []
+        tokens_to_add = []
         for token in new_tokens:
             assert isinstance(token, str)
             if self.init_kwargs.get("do_lower_case", False) and token not in self.all_special_tokens:
@@ -1164,18 +1178,18 @@ class PreTrainedTokenizer(SpecialTokensMixin):
             if (
                 token != self.unk_token
                 and self.convert_tokens_to_ids(token) == self.convert_tokens_to_ids(self.unk_token)
-                and token not in to_add_tokens
+                and token not in tokens_to_add
             ):
-                to_add_tokens.append(token)
+                tokens_to_add.append(token)
                 logger.info("Adding %s to the vocabulary", token)
 
-        added_tok_encoder = dict((tok, len(self) + i) for i, tok in enumerate(to_add_tokens))
+        added_tok_encoder = dict((tok, len(self) + i) for i, tok in enumerate(tokens_to_add))
         added_tok_decoder = {v: k for k, v in added_tok_encoder.items()}
         self.added_tokens_encoder.update(added_tok_encoder)
         self.unique_added_tokens_encoder = set(self.added_tokens_encoder.keys()).union(set(self.all_special_tokens))
         self.added_tokens_decoder.update(added_tok_decoder)
 
-        return len(to_add_tokens)
+        return len(tokens_to_add)
 
     def num_special_tokens_to_add(self, pair=False):
         """
@@ -2080,10 +2094,7 @@ class PreTrainedTokenizer(SpecialTokensMixin):
     def build_inputs_with_special_tokens(self, token_ids_0: List, token_ids_1: Optional[List] = None) -> List:
         """
         Build model inputs from a sequence or a pair of sequence for sequence classification tasks
-        by concatenating and adding special tokens.
-        A RoBERTa sequence has the following format:
-            single sequence: <s> X </s>
-            pair of sequences: <s> A </s></s> B </s>
+        by concatenating and adding special tokens. This implementation does not add special tokens.
         """
         if token_ids_1 is None:
             return token_ids_0
@@ -2655,12 +2666,12 @@ class PreTrainedTokenizerFast(PreTrainedTokenizer):
         else:
             return text
 
-    def save_vocabulary(self, save_directory: str) -> Tuple[str]:
+    def save_vocabulary(self, save_directory: Union[str, Path]) -> Tuple[str]:
         if os.path.isdir(save_directory):
-            files = self._tokenizer.save(save_directory)
+            files = self._tokenizer.save(str(save_directory))
         else:
             folder, file = os.path.split(os.path.abspath(save_directory))
-            files = self._tokenizer.save(folder, name=file)
+            files = self._tokenizer.save(str(folder), name=file)
 
         return tuple(files)
 

src/transformers/trainer.py

@@ -7,7 +7,7 @@ import re
 import shutil
 from contextlib import contextmanager
 from pathlib import Path
-from typing import Callable, Dict, List, Optional, Tuple, Union
+from typing import Callable, Dict, List, Optional, Tuple
 
 import numpy as np
 import torch
@@ -242,9 +242,6 @@ class Trainer:
             collate_fn=self.data_collator.collate_batch,
         )
 
-        if is_tpu_available():
-            data_loader = pl.ParallelLoader(data_loader, [self.args.device]).per_device_loader(self.args.device)
-
         return data_loader
 
     def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
@@ -269,9 +266,6 @@ class Trainer:
             collate_fn=self.data_collator.collate_batch,
         )
 
-        if is_tpu_available():
-            data_loader = pl.ParallelLoader(data_loader, [self.args.device]).per_device_loader(self.args.device)
-
         return data_loader
 
     def get_test_dataloader(self, test_dataset: Dataset) -> DataLoader:
@@ -292,9 +286,6 @@ class Trainer:
             collate_fn=self.data_collator.collate_batch,
         )
 
-        if is_tpu_available():
-            data_loader = pl.ParallelLoader(data_loader, [self.args.device]).per_device_loader(self.args.device)
-
         return data_loader
 
     def get_optimizers(
@@ -351,15 +342,11 @@ class Trainer:
                 self.model, log=os.getenv("WANDB_WATCH", "gradients"), log_freq=max(100, self.args.logging_steps)
             )
 
-    def num_examples(self, dataloader: Union[DataLoader, "pl.PerDeviceLoader"]) -> int:
+    def num_examples(self, dataloader: DataLoader) -> int:
         """
         Helper to get num of examples from a DataLoader, by accessing its Dataset.
         """
-        if is_tpu_available():
-            assert isinstance(dataloader, pl.PerDeviceLoader)
-            return len(dataloader._loader._loader.dataset)
-        else:
-            return len(dataloader.dataset)
+        return len(dataloader.dataset)
 
     def train(self, model_path: Optional[str] = None):
         """
@@ -429,7 +416,7 @@ class Trainer:
         logger.info("***** Running training *****")
         logger.info("  Num examples = %d", self.num_examples(train_dataloader))
         logger.info("  Num Epochs = %d", num_train_epochs)
-        logger.info("  Instantaneous batch size per device = %d", self.args.per_gpu_train_batch_size)
+        logger.info("  Instantaneous batch size per device = %d", self.args.per_device_train_batch_size)
         logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d", total_train_batch_size)
         logger.info("  Gradient Accumulation steps = %d", self.args.gradient_accumulation_steps)
         logger.info("  Total optimization steps = %d", t_total)
@@ -466,7 +453,14 @@ class Trainer:
             if isinstance(train_dataloader, DataLoader) and isinstance(train_dataloader.sampler, DistributedSampler):
                 train_dataloader.sampler.set_epoch(epoch)
 
-            epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=not self.is_local_master())
+            if is_tpu_available():
+                parallel_loader = pl.ParallelLoader(train_dataloader, [self.args.device]).per_device_loader(
+                    self.args.device
+                )
+                epoch_iterator = tqdm(parallel_loader, desc="Iteration", disable=not self.is_local_master())
+            else:
+                epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=not self.is_local_master())
+
             for step, inputs in enumerate(epoch_iterator):
 
                 # Skip past any already trained steps if resuming training
@@ -514,24 +508,28 @@ class Trainer:
                         if self.args.evaluate_during_training:
                             self.evaluate()
 
-                    if self.is_world_master():
-                        if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
-                            # In all cases (even distributed/parallel), self.model is always a reference
-                            # to the model we want to save.
-                            if hasattr(model, "module"):
-                                assert model.module is self.model
-                            else:
-                                assert model is self.model
-                            # Save model checkpoint
-                            output_dir = os.path.join(
-                                self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}"
-                            )
+                    if self.args.save_steps > 0 and self.global_step % self.args.save_steps == 0:
+                        # In all cases (even distributed/parallel), self.model is always a reference
+                        # to the model we want to save.
+                        if hasattr(model, "module"):
+                            assert model.module is self.model
+                        else:
+                            assert model is self.model
+                        # Save model checkpoint
+                        output_dir = os.path.join(self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.global_step}")
 
-                            self.save_model(output_dir)
+                        self.save_model(output_dir)
+
+                        if self.is_world_master():
                             self._rotate_checkpoints()
+
+                        if is_tpu_available():
+                            xm.rendezvous("saving_optimizer_states")
+                            xm.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                            xm.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                        elif self.is_world_master():
                             torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
                             torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
-                            logger.info("Saving optimizer and scheduler states to %s", output_dir)
 
                 if self.args.max_steps > 0 and self.global_step > self.args.max_steps:
                     epoch_iterator.close()
@@ -713,6 +711,7 @@ class Trainer:
         In that case, this method will also return metrics, like in evaluate().
         """
         test_dataloader = self.get_test_dataloader(test_dataset)
+
         return self._prediction_loop(test_dataloader, description="Prediction")
 
     def _prediction_loop(
@@ -735,10 +734,7 @@ class Trainer:
         # Note: in torch.distributed mode, there's no point in wrapping the model
         # inside a DistributedDataParallel as we'll be under `no_grad` anyways.
 
-        if is_tpu_available():
-            batch_size = dataloader._loader._loader.batch_size
-        else:
-            batch_size = dataloader.batch_size
+        batch_size = dataloader.batch_size
         logger.info("***** Running %s *****", description)
         logger.info("  Num examples = %d", self.num_examples(dataloader))
         logger.info("  Batch size = %d", batch_size)
@@ -747,6 +743,9 @@ class Trainer:
         label_ids: torch.Tensor = None
         model.eval()
 
+        if is_tpu_available():
+            dataloader = pl.ParallelLoader(dataloader, [self.args.device]).per_device_loader(self.args.device)
+
         for inputs in tqdm(dataloader, desc=description):
             has_labels = any(inputs.get(k) is not None for k in ["labels", "lm_labels", "masked_lm_labels"])
 

src/transformers/trainer_tf.py

@@ -141,7 +141,7 @@ class TFTrainer:
                 self.optimizer = tf.keras.optimizers.get(
                     {"class_name": self.args.optimizer_name, "config": {"learning_rate": self.args.learning_rate}}
                 )
-        logger.info("Created an/a {} optimizer".format(self.optimizer))
+        logger.info("Created an/a {} optimizer".format(self.args.optimizer_name))
 
     def _create_checkpoint_manager(self, max_to_keep: int = 5, load_model: bool = True) -> None:
         """
@@ -335,12 +335,8 @@ class TFTrainer:
             gradient / tf.cast(gradient_scale, gradient.dtype) for gradient in self.gradient_accumulator.gradients
         ]
         gradients = [(tf.clip_by_value(grad, -self.args.max_grad_norm, self.args.max_grad_norm)) for grad in gradients]
-        vars = self.model.trainable_variables
 
-        if self.args.mode in ["token-classification", "question-answering"]:
-            vars = [var for var in self.model.trainable_variables if "pooler" not in var.name]
-
-        self.optimizer.apply_gradients(list(zip(gradients, vars)))
+        self.optimizer.apply_gradients(list(zip(gradients, self.model.trainable_variables)))
         self.gradient_accumulator.reset()
 
     def _accumulate_next_gradients(self):
@@ -375,12 +371,10 @@ class TFTrainer:
     def _forward(self, features, labels):
         """Forwards a training example and accumulates the gradients."""
         per_example_loss, _ = self._run_model(features, labels, True)
-        vars = self.model.trainable_variables
-
-        if self.args.mode in ["token-classification", "question-answering"]:
-            vars = [var for var in self.model.trainable_variables if "pooler" not in var.name]
-
-        gradients = self.optimizer.get_gradients(per_example_loss, vars)
+        gradients = tf.gradients(per_example_loss, self.model.trainable_variables)
+        gradients = [
+            g if g is not None else tf.zeros_like(v) for g, v in zip(gradients, self.model.trainable_variables)
+        ]
 
         self.gradient_accumulator(gradients)
 

src/transformers/training_args.py

@@ -58,8 +58,28 @@ class TrainingArguments:
         default=False, metadata={"help": "Run evaluation during training at each logging step."},
     )
 
-    per_gpu_train_batch_size: int = field(default=8, metadata={"help": "Batch size per GPU/CPU for training."})
-    per_gpu_eval_batch_size: int = field(default=8, metadata={"help": "Batch size per GPU/CPU for evaluation."})
+    per_device_train_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
+    )
+    per_device_eval_batch_size: int = field(
+        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
+    )
+
+    per_gpu_train_batch_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Deprecated, the use of `--per_device_train_batch_size` is preferred. "
+            "Batch size per GPU/TPU core/CPU for training."
+        },
+    )
+    per_gpu_eval_batch_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Deprecated, the use of `--per_device_eval_batch_size` is preferred."
+            "Batch size per GPU/TPU core/CPU for evaluation."
+        },
+    )
+
     gradient_accumulation_steps: int = field(
         default=1,
         metadata={"help": "Number of updates steps to accumulate before performing a backward/update pass."},
@@ -115,11 +135,23 @@ class TrainingArguments:
 
     @property
     def train_batch_size(self) -> int:
-        return self.per_gpu_train_batch_size * max(1, self.n_gpu)
+        if self.per_gpu_train_batch_size:
+            logger.warning(
+                "Using deprecated `--per_gpu_train_batch_size` argument which will be removed in a future "
+                "version. Using `--per_device_train_batch_size` is preferred."
+            )
+        per_device_batch_size = self.per_gpu_train_batch_size or self.per_device_train_batch_size
+        return per_device_batch_size * max(1, self.n_gpu)
 
     @property
     def eval_batch_size(self) -> int:
-        return self.per_gpu_eval_batch_size * max(1, self.n_gpu)
+        if self.per_gpu_eval_batch_size:
+            logger.warning(
+                "Using deprecated `--per_gpu_eval_batch_size` argument which will be removed in a future "
+                "version. Using `--per_device_eval_batch_size` is preferred."
+            )
+        per_device_batch_size = self.per_gpu_eval_batch_size or self.per_device_eval_batch_size
+        return per_device_batch_size * max(1, self.n_gpu)
 
     @cached_property
     @torch_required

tests/test_benchmark.py

@@ -0,0 +1,90 @@
+import os
+import tempfile
+import unittest
+from pathlib import Path
+
+from transformers import GPT2Config, is_torch_available
+
+from .utils import require_torch
+
+
+if is_torch_available():
+    from transformers import (
+        PyTorchBenchmarkArguments,
+        PyTorchBenchmark,
+    )
+
+
+@require_torch
+class BenchmarkTest(unittest.TestCase):
+    def check_results_dict_not_empty(self, results):
+        for model_result in results.values():
+            for batch_size, sequence_length in zip(model_result["bs"], model_result["ss"]):
+                result = model_result["result"][batch_size][sequence_length]
+                self.assertIsNotNone(result)
+
+    def test_inference_no_configs(self):
+        MODEL_ID = "sshleifer/tiny-gpt2"
+        benchmark_args = PyTorchBenchmarkArguments(
+            models=[MODEL_ID], training=False, no_inference=False, sequence_lengths=[8], batch_sizes=[1]
+        )
+        benchmark = PyTorchBenchmark(benchmark_args)
+        results = benchmark.run()
+        self.check_results_dict_not_empty(results.time_inference_result)
+        self.check_results_dict_not_empty(results.memory_inference_result)
+
+    def test_train_no_configs(self):
+        MODEL_ID = "sshleifer/tiny-gpt2"
+        benchmark_args = PyTorchBenchmarkArguments(
+            models=[MODEL_ID], training=True, no_inference=True, sequence_lengths=[8], batch_sizes=[1]
+        )
+        benchmark = PyTorchBenchmark(benchmark_args)
+        results = benchmark.run()
+        self.check_results_dict_not_empty(results.time_train_result)
+        self.check_results_dict_not_empty(results.memory_train_result)
+
+    def test_inference_with_configs(self):
+        MODEL_ID = "sshleifer/tiny-gpt2"
+        config = GPT2Config.from_pretrained(MODEL_ID)
+        benchmark_args = PyTorchBenchmarkArguments(
+            models=[MODEL_ID], training=False, no_inference=False, sequence_lengths=[8], batch_sizes=[1]
+        )
+        benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
+        results = benchmark.run()
+        self.check_results_dict_not_empty(results.time_inference_result)
+        self.check_results_dict_not_empty(results.memory_inference_result)
+
+    def test_train_with_configs(self):
+        MODEL_ID = "sshleifer/tiny-gpt2"
+        config = GPT2Config.from_pretrained(MODEL_ID)
+        benchmark_args = PyTorchBenchmarkArguments(
+            models=[MODEL_ID], training=True, no_inference=True, sequence_lengths=[8], batch_sizes=[1]
+        )
+        benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
+        results = benchmark.run()
+        self.check_results_dict_not_empty(results.time_train_result)
+        self.check_results_dict_not_empty(results.memory_train_result)
+
+    def test_save_csv_files(self):
+        MODEL_ID = "sshleifer/tiny-gpt2"
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            benchmark_args = PyTorchBenchmarkArguments(
+                models=[MODEL_ID],
+                training=True,
+                no_inference=False,
+                save_to_csv=True,
+                sequence_lengths=[8],
+                batch_sizes=[1],
+                inference_time_csv_file=os.path.join(tmp_dir, "inf_time.csv"),
+                train_memory_csv_file=os.path.join(tmp_dir, "train_mem.csv"),
+                inference_memory_csv_file=os.path.join(tmp_dir, "inf_mem.csv"),
+                train_time_csv_file=os.path.join(tmp_dir, "train_time.csv"),
+                env_info_csv_file=os.path.join(tmp_dir, "env.csv"),
+            )
+            benchmark = PyTorchBenchmark(benchmark_args)
+            benchmark.run()
+            self.assertTrue(Path(os.path.join(tmp_dir, "inf_time.csv")).exists())
+            self.assertTrue(Path(os.path.join(tmp_dir, "train_time.csv")).exists())
+            self.assertTrue(Path(os.path.join(tmp_dir, "inf_mem.csv")).exists())
+            self.assertTrue(Path(os.path.join(tmp_dir, "train_mem.csv")).exists())
+            self.assertTrue(Path(os.path.join(tmp_dir, "env.csv")).exists())

tests/test_hf_argparser.py

@@ -3,11 +3,15 @@ import unittest
 from argparse import Namespace
 from dataclasses import dataclass, field
 from enum import Enum
-from typing import Optional
+from typing import List, Optional
 
 from transformers import HfArgumentParser, TrainingArguments
 
 
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
 @dataclass
 class BasicExample:
     foo: int
@@ -43,6 +47,16 @@ class OptionalExample:
     foo: Optional[int] = None
     bar: Optional[float] = field(default=None, metadata={"help": "help message"})
     baz: Optional[str] = None
+    ces: Optional[List[str]] = list_field(default=[])
+    des: Optional[List[int]] = list_field(default=[])
+
+
+@dataclass
+class ListExample:
+    foo_int: List[int] = list_field(default=[])
+    bar_int: List[int] = list_field(default=[1, 2, 3])
+    foo_str: List[str] = list_field(default=["Hallo", "Bonjour", "Hello"])
+    foo_float: List[float] = list_field(default=[0.1, 0.2, 0.3])
 
 
 class HfArgumentParserTest(unittest.TestCase):
@@ -101,6 +115,26 @@ class HfArgumentParserTest(unittest.TestCase):
         args = parser.parse_args(["--foo", "titi"])
         self.assertEqual(args.foo, BasicEnum.titi)
 
+    def test_with_list(self):
+        parser = HfArgumentParser(ListExample)
+
+        expected = argparse.ArgumentParser()
+        expected.add_argument("--foo_int", nargs="+", default=[], type=int)
+        expected.add_argument("--bar_int", nargs="+", default=[1, 2, 3], type=int)
+        expected.add_argument("--foo_str", nargs="+", default=["Hallo", "Bonjour", "Hello"], type=str)
+        expected.add_argument("--foo_float", nargs="+", default=[0.1, 0.2, 0.3], type=float)
+
+        self.argparsersEqual(parser, expected)
+
+        args = parser.parse_args([])
+        self.assertEqual(
+            args,
+            Namespace(foo_int=[], bar_int=[1, 2, 3], foo_str=["Hallo", "Bonjour", "Hello"], foo_float=[0.1, 0.2, 0.3]),
+        )
+
+        args = parser.parse_args("--foo_int 1 --bar_int 2 3 --foo_str a b c --foo_float 0.1 0.7".split())
+        self.assertEqual(args, Namespace(foo_int=[1], bar_int=[2, 3], foo_str=["a", "b", "c"], foo_float=[0.1, 0.7]))
+
     def test_with_optional(self):
         parser = HfArgumentParser(OptionalExample)
 
@@ -108,13 +142,15 @@ class HfArgumentParserTest(unittest.TestCase):
         expected.add_argument("--foo", default=None, type=int)
         expected.add_argument("--bar", default=None, type=float, help="help message")
         expected.add_argument("--baz", default=None, type=str)
+        expected.add_argument("--ces", nargs="+", default=[], type=str)
+        expected.add_argument("--des", nargs="+", default=[], type=int)
         self.argparsersEqual(parser, expected)
 
         args = parser.parse_args([])
-        self.assertEqual(args, Namespace(foo=None, bar=None, baz=None))
+        self.assertEqual(args, Namespace(foo=None, bar=None, baz=None, ces=[], des=[]))
 
-        args = parser.parse_args("--foo 12 --bar 3.14 --baz 42".split())
-        self.assertEqual(args, Namespace(foo=12, bar=3.14, baz="42"))
+        args = parser.parse_args("--foo 12 --bar 3.14 --baz 42 --ces a b c --des 1 2 3".split())
+        self.assertEqual(args, Namespace(foo=12, bar=3.14, baz="42", ces=["a", "b", "c"], des=[1, 2, 3]))
 
     def test_integration_training_args(self):
         parser = HfArgumentParser(TrainingArguments)