Sampling instead of argmax

Utils


Initial



wip


wip

.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:

README.md

@@ -340,8 +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/$TASK_NAME/
@@ -367,8 +367,8 @@ python ./examples/text-classification/run_glue.py \
     --data_dir=${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   \
@@ -391,8 +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/ \
@@ -428,8 +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/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/model_doc/albert.rst

@@ -94,3 +94,17 @@ TFAlbertForSequenceClassification
 
 .. autoclass:: transformers.TFAlbertForSequenceClassification
     :members:
+
+
+TFAlbertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAlbertForMultipleChoice
+    :members:
+
+
+TFAlbertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFAlbertForQuestionAnswering
+    :members:

docs/source/model_doc/longformer.rst

@@ -21,7 +21,7 @@ A selecetd few tokens attend "globally" to all other tokens, as it is convention
 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.
+The user can define which tokens attend "locally" and which tokens attend "globally" by setting the tensor `global_attention_mask` at run-time appropriately. `Longformer` employs the following logic for `global_attention_mask`: `0` - the token attends "locally", `1` - 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.
 
@@ -67,3 +67,25 @@ LongformerForMaskedLM
 
 .. autoclass:: transformers.LongformerForMaskedLM
     :members:
+
+
+LongformerForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerForQuestionAnswering
+    :members:
+
+
+LongformerForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerForMultipleChoice
+    :members:
+
+
+LongformerForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LongformerForTokenClassification
+    :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/model_doc/roberta.rst

@@ -74,6 +74,13 @@ RobertaForSequenceClassification
     :members:
 
 
+RobertaForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.RobertaForMultipleChoice
+    :members:
+
+
 RobertaForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

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/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/language-modeling/run_electra_pretraining.py

@@ -0,0 +1,756 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Pre-training language models using the ELECTRA method.
+"""
+
+
+import logging
+import math
+import multiprocessing
+import os
+import tarfile
+from dataclasses import dataclass, field
+from itertools import chain
+from pathlib import Path
+from typing import Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.utils.data import IterableDataset
+from tqdm import tqdm
+
+from transformers import (
+    AutoConfig,
+    AutoTokenizer,
+    DataCollatorForLanguageModeling,
+    ElectraForMaskedLM,
+    ElectraForPreTraining,
+    EvalPrediction,
+    HfArgumentParser,
+    PreTrainedTokenizer,
+    TextDataset,
+    Trainer,
+    set_seed,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.training_args import TrainingArguments
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    discriminator_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The discriminator checkpoint for weights initialization. Leave None if you want to train a model "
+            "from scratch."
+        },
+    )
+
+    discriminator_config_name: Optional[str] = field(
+        default=None,
+        metadata={"help": "Pretrained config name or path if not the same as the discriminator model_name"},
+    )
+
+    generator_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The generator checkpoint for weights initialization. Leave None if you want to train a model "
+            "from scratch."
+        },
+    )
+
+    generator_config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as the generator model_name"}
+    )
+
+    tokenizer_name: Optional[str] = field(
+        default=None,
+        metadata={"help": "Pretrained tokenizer name or path if not the same as discriminator model_name"},
+    )
+
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    train_data_file: Optional[str] = field(
+        default=None, metadata={"help": "The input training data file (a text file)."}
+    )
+
+    eval_data_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+
+    open_web_text_directory: Optional[str] = field(
+        default=None,
+        metadata={"help": "The directory containing files that will be used for training and evaluation."},
+    )
+
+    block_size: int = field(
+        default=-1,
+        metadata={
+            "help": "Optional input sequence length after tokenization."
+            "The training dataset will be truncated in block of this size for training."
+            "Default to the model max input length for single sentence inputs (take into account special tokens)."
+        },
+    )
+
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+
+    num_dataset_building_processes: int = field(
+        default=1, metadata={"help": "The number of workers that will be used to build the dataset."}
+    )
+
+    num_tensors_per_file: int = field(
+        default=2048,
+        metadata={
+            "help": "The number of tensors that will be stored in each file after tokenization."
+            "The smaller the amount, the smaller the filesize, but the larger the amount"
+            "of files that will be created."
+        },
+    )
+
+    mask_probability: float = field(
+        default=0.15, metadata={"help": "Percentage of the input that will be masked or replaced."}
+    )
+
+    max_predictions_per_sequence: int = field(
+        default=-1, metadata={"help": "Maximum tokens that will be masked in a sequence."},
+    )
+
+
+@dataclass
+class ElectraTrainingArguments(TrainingArguments):
+    max_steps: int = field(
+        default=1_000_000,
+        metadata={"help": "If > 0: set total number of training steps to perform. Override num_train_epochs."},
+    )
+
+    max_eval_steps: int = field(
+        default=100, metadata={"help": "If > 0: set total number of eval steps to perform."},
+    )
+    warmup_steps: int = field(default=10_000, metadata={"help": "Linear warmup over warmup_steps."})
+
+    weight_decay: float = field(default=0.1, metadata={"help": "Weight decay if we apply some."})
+
+    generator_weight: float = field(default=1.0, metadata={"help": "Weight coefficient for the generator loss"})
+
+    discriminator_weight: float = field(
+        default=50.0, metadata={"help": "Weight coefficient for the discriminator loss"}
+    )
+
+
+def get_dataset(
+    data_args: DataTrainingArguments,
+    training_args: TrainingArguments,
+    model_args: ModelArguments,
+    tokenizer: Union[PreTrainedTokenizer, str],
+    evaluate=False,
+    local_rank=-1,
+):
+    if data_args.open_web_text_directory is not None:
+        # Whether to overwrite the cache. We don't want to overwrite the cache when evaluating if we're both training
+        # and evaluating, as the same dataset is used. We don't need to tokenize twice for training
+        # and evaluation.
+        should_overwrite_cache = False
+
+        # If argument is specified and training, then respect the argument
+        if data_args.overwrite_cache and not evaluate:
+            should_overwrite_cache = True
+        if data_args.overwrite_cache and training_args.do_eval and not training_args.do_train:
+            should_overwrite_cache = True
+
+        return OpenWebTextDataset(data_args, model_args, overwrite_cache=should_overwrite_cache)
+    else:
+        file_path = data_args.eval_data_file if evaluate else data_args.train_data_file
+        return TextDataset(
+            tokenizer=tokenizer, file_path=file_path, block_size=data_args.block_size, local_rank=local_rank,
+        )
+
+
+class OpenWebTextDataset(IterableDataset):
+    def __init__(self, data_args, model_args, overwrite_cache=False):
+        self.tokenizer_cache = model_args.cache_dir
+        self.directory = Path(data_args.open_web_text_directory)
+        self.archives = os.listdir(self.directory)
+        self.tokenizer_identifier = model_args.tokenizer_name
+        self.num_tensors_per_file = data_args.num_tensors_per_file
+        self.feature_directory = (
+            self.directory
+            / f"features_{self.tokenizer_identifier.replace('/', '_')}_{data_args.block_size if data_args.block_size is not None else 'no-max-seq'}_{self.num_tensors_per_file}"
+        )
+        self.block_size = data_args.block_size
+
+        # The dataset was already processed
+        if os.path.exists(self.feature_directory) and not overwrite_cache:
+            logger.info(
+                f"Re-using cache from {self.feature_directory}. Warning: we have no way of detecting an "
+                f"incomplete cache. If the tokenization was started but not finished, please use the "
+                f"`--ignore_cache=True` flag."
+            )
+            self.feature_set_paths = [
+                self.feature_directory / feature_set_path for feature_set_path in os.listdir(self.feature_directory)
+            ]
+            return
+
+        logger.info(f"Writing features at {self.feature_directory}")
+        os.makedirs(self.feature_directory, exist_ok=overwrite_cache)
+
+        n_archives_per_job = math.ceil(len(self.archives) / data_args.num_dataset_building_processes)
+        self.job_archives = [
+            self.archives[i * n_archives_per_job : (i + 1) * n_archives_per_job]
+            for i in range(data_args.num_dataset_building_processes)
+        ]
+        # Sanity check: make sure we're not leaving any archive behind.
+        assert sum([len(archive) for archive in self.job_archives]) == len(self.archives)
+
+        if data_args.num_dataset_building_processes == 1:
+            self.feature_set_paths = self._extract_open_web_text()
+        else:
+            pool = multiprocessing.Pool(processes=data_args.num_dataset_building_processes)
+            self.feature_set_paths = pool.map(
+                self._extract_open_web_text, range(data_args.num_dataset_building_processes)
+            )
+            self.feature_set_paths = [file_path for feature_set in self.feature_set_paths for file_path in feature_set]
+
+    def _extract_open_web_text(self, job_id=0):
+        """
+        OpenWebText is saved under the following format:
+
+        openwebtext.zip
+            |-> archive_xxx.zip
+                |-> file_xxx.txt
+                |-> file_xxz.txt
+                ...
+            |-> archive_xxz.zip
+                |-> file_xxy.txt
+                ...
+            ...
+        """
+        tokenizer = AutoTokenizer.from_pretrained(
+            self.tokenizer_identifier, use_fast=True, cache_dir=self.tokenizer_cache
+        )
+
+        # Create openwebtext/tmp directory to store temporary files
+        temporary_directory = self.directory / "tmp" / f"job_{job_id}"
+        feature_index = 0
+        feature_set_paths = []
+
+        os.makedirs(temporary_directory, exist_ok=True)
+
+        # Extract archives and tokenize in directory
+        progress_bar = tqdm(
+            self.job_archives[job_id], desc="Extracting archives", total=len(self.job_archives[0]), disable=job_id != 0
+        )
+
+        features = []
+        for archive in progress_bar:
+            if os.path.isdir(self.directory / archive):
+                logger.info("Ignoring rogue directory.")
+                continue
+            with tarfile.open(self.directory / archive) as t:
+                extracted_archive = temporary_directory / f"{archive}-extracted"
+                t.extractall(extracted_archive)
+
+            files = os.listdir(extracted_archive)
+            for file in files:
+                file_path = extracted_archive / file
+
+                with open(file_path, "r") as f:
+                    text = f.read()
+                    block_size = tokenizer.model_max_length if self.block_size is None else self.block_size
+                    encoding = tokenizer.encode_plus(text, return_overflowing_tokens=True, max_length=block_size)
+
+                    features.append(torch.tensor(encoding["input_ids"]))
+
+                    for overflowing_encoding in encoding.encodings[0].overflowing:
+                        features.append(torch.tensor(overflowing_encoding.ids))
+
+            while len(features) > self.num_tensors_per_file:
+                feature_set_path = self.feature_directory / f"feature_set_{job_id}_{feature_index}.pt"
+                torch.save(features[: self.num_tensors_per_file], feature_set_path)
+                features = features[self.num_tensors_per_file :]
+                feature_index += 1
+                feature_set_paths.append(feature_set_path)
+
+        if len(features) > 0:
+            feature_set_path = self.feature_directory / f"feature_set_{job_id}_{feature_index}.pt"
+            torch.save(features, feature_set_path)
+            feature_set_paths.append(feature_set_path)
+
+        return feature_set_paths
+
+    @staticmethod
+    def parse_file(file_index):
+        try:
+            features = torch.load(file_index)
+            yield from features
+        except RuntimeError:
+            raise RuntimeError(f"Corrupted file {file_index}")
+
+    def __len__(self):
+        return len(self.feature_set_paths) * self.num_tensors_per_file
+
+    def __iter__(self):
+        return chain.from_iterable(map(self.parse_file, self.feature_set_paths))
+
+
+class CombinedModel(nn.Module):
+    def __init__(
+        self,
+        discriminator: PreTrainedModel,
+        generator: PreTrainedModel,
+        tokenizer: PreTrainedTokenizer,
+        training_args: ElectraTrainingArguments,
+        data_args: DataTrainingArguments,
+    ):
+        super().__init__()
+
+        self.discriminator = discriminator
+        self.generator = generator
+
+        # Embeddings are shared
+        self.discriminator.set_input_embeddings(self.generator.get_input_embeddings())
+
+        self.tokenizer = tokenizer
+
+        self.discriminator_weight = training_args.discriminator_weight
+        self.generator_weight = training_args.generator_weight
+        self.mask_probability = data_args.mask_probability
+        self.max_predictions_per_sequence = data_args.max_predictions_per_sequence
+
+        # Original implementation has a default of :(mask_probability + 0.005) * max_sequence_length
+        if self.max_predictions_per_sequence == -1:
+            self.max_predictions_per_sequence = (self.mask_probability + 0.005) * data_args.block_size
+
+        class Config:
+            xla_device: bool = False
+
+        self.config = Config()
+
+    def mask_inputs(
+        self, input_ids: torch.Tensor, tokens_to_ignore, proposal_distribution=1.0,
+    ):
+        input_ids = input_ids.clone()
+        inputs_which_can_be_masked = torch.ones_like(input_ids)
+        for token in tokens_to_ignore:
+            inputs_which_can_be_masked -= torch.eq(input_ids, token).long()
+
+        total_number_of_tokens = input_ids.shape[-1]
+
+        # Identify the number of tokens to be masked, which should be: 1 < num < max_predictions per seq.
+        # It is set to be: n_tokens * mask_probability, but is truncated if it goes beyond bounds.
+        number_of_tokens_to_be_masked = torch.max(
+            torch.tensor(1),
+            torch.min(
+                torch.tensor(self.max_predictions_per_sequence, dtype=torch.long),
+                torch.tensor(int(total_number_of_tokens * self.mask_probability), dtype=torch.long),
+            ),
+        )
+
+        # The probability of each token being masked
+        sample_prob = proposal_distribution * inputs_which_can_be_masked
+        sample_prob /= torch.sum(sample_prob)
+
+        # Sample from the probabilities
+        masked_lm_positions = sample_prob.multinomial(number_of_tokens_to_be_masked)
+
+        # Gather the IDs from the positions
+        masked_lm_ids = input_ids.gather(-1, masked_lm_positions)
+
+        return masked_lm_ids, masked_lm_positions
+
+    @staticmethod
+    def gather_positions(sequence, positions):
+        batch_size, sequence_length, dimension = sequence.shape
+        position_shift = (sequence_length * torch.arange(batch_size, device=sequence.device)).unsqueeze(-1)
+        flat_positions = torch.reshape(positions + position_shift, [-1]).long()
+        flat_sequence = torch.reshape(sequence, [batch_size * sequence_length, dimension])
+        gathered = flat_sequence.index_select(0, flat_positions)
+        return torch.reshape(gathered, [batch_size, -1, dimension])
+
+    def forward(
+        self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, labels=None,
+    ):
+        # get the masked positions as well as their original values
+        masked_input_ids, masked_lm_positions = self.mask_inputs(
+            input_ids, [self.tokenizer.cls_token_id, self.tokenizer.sep_token_id, self.tokenizer.mask_token_id],
+        )
+
+        # only masked values should be counted in the loss; build a tensor containing the true values and -100 otherwise
+        masked_lm_labels = torch.full_like(input_ids, -100)
+        masked_lm_labels.scatter_(-1, masked_lm_positions, masked_input_ids)
+
+        # Create a tensor filled with masks
+        masked_tokens = torch.full_like(masked_input_ids, self.tokenizer.mask_token_id)
+        masked_lm_inputs = input_ids.clone()
+
+        # Of the evaluated tokens, 15% of those will keep their original tokens
+        replace_with_mask_positions = masked_lm_positions * (
+            torch.rand(masked_lm_positions.shape, device=masked_lm_positions.device) < (1 - self.mask_probability)
+        )
+
+        # Scatter the masks at the masked positions
+        masked_lm_inputs.scatter_(-1, replace_with_mask_positions, masked_tokens)
+        masked_lm_inputs[..., 0] = 101
+
+        generator_loss, generator_output = self.generator(
+            masked_lm_inputs,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            position_ids,
+            masked_lm_labels=masked_lm_labels,
+        )[:2]
+
+        # get the generator's predicted value on each masked position
+        fake_logits = self.gather_positions(generator_output, masked_lm_positions)
+        fake_softmaxed = torch.softmax(fake_logits, dim=-1)
+        fake_sampled = fake_softmaxed\
+            .view(fake_logits.shape[0] * fake_logits.shape[1], fake_logits.shape[2])\
+            .multinomial(1)\
+            .view(fake_logits.shape[:-1])
+
+        # create a tensor containing the predicted tokens
+        fake_tokens = input_ids.scatter(-1, masked_lm_positions, fake_sampled)
+        fake_tokens[:, 0] = input_ids[:, 0]
+        discriminator_labels = (labels != fake_tokens).int()
+
+        discriminator_loss, discriminator_output = self.discriminator(
+            fake_tokens,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            position_ids,
+            labels=discriminator_labels,
+        )[:2]
+
+        discriminator_predictions = torch.round((torch.sign(discriminator_output) + 1.0) * 0.5)
+
+        total_loss = (self.discriminator_weight * discriminator_loss) + (self.generator_weight * generator_loss)
+
+        return (
+            total_loss,
+            (generator_output, discriminator_output),
+            (masked_input_ids, fake_sampled),
+            (discriminator_labels, discriminator_predictions),
+        )
+
+    def save_pretrained(self, directory):
+        if self.config.xla_device:
+            self.discriminator.config.xla_device = True
+            self.generator.config.xla_device = True
+        else:
+            self.discriminator.config.xla_device = False
+            self.generator.config.xla_device = False
+
+        generator_path = os.path.join(directory, "generator")
+        discriminator_path = os.path.join(directory, "discriminator")
+
+        if not os.path.exists(generator_path):
+            os.makedirs(generator_path)
+
+        if not os.path.exists(discriminator_path):
+            os.makedirs(discriminator_path)
+
+        self.generator.save_pretrained(generator_path)
+        self.discriminator.save_pretrained(discriminator_path)
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, ElectraTrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if data_args.open_web_text_directory is None and data_args.eval_data_file is None and training_args.do_eval:
+        raise ValueError(
+            "Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
+            "or remove the --do_eval argument."
+        )
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        training_args.local_rank,
+        training_args.device,
+        training_args.n_gpu,
+        bool(training_args.local_rank != -1),
+        training_args.fp16,
+    )
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed
+    set_seed(training_args.seed)
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+
+    if model_args.discriminator_config_name:
+        discriminator_config = AutoConfig.from_pretrained(
+            model_args.discriminator_config_name, cache_dir=model_args.cache_dir
+        )
+    elif model_args.discriminator_name_or_path:
+        discriminator_config = AutoConfig.from_pretrained(
+            model_args.discriminator_name_or_path, cache_dir=model_args.cache_dir
+        )
+    else:
+        raise ValueError("Either --discriminator_config_name or --discriminator_name_or_path should be specified.")
+
+    if model_args.generator_config_name:
+        generator_config = AutoConfig.from_pretrained(model_args.generator_config_name, cache_dir=model_args.cache_dir)
+    elif model_args.generator_name_or_path:
+        generator_config = AutoConfig.from_pretrained(
+            model_args.generator_name_or_path, cache_dir=model_args.cache_dir
+        )
+    else:
+        raise ValueError("Either --generator_config_name or --generator_name_or_path should be specified.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)
+    elif model_args.discriminator_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.discriminator_name_or_path, cache_dir=model_args.cache_dir
+        )
+        model_args.tokenizer_name = model_args.discriminator_name_or_path
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another "
+            "script, save it, and load it from here, using --tokenizer_name"
+        )
+
+    if model_args.discriminator_name_or_path:
+        discriminator = ElectraForPreTraining.from_pretrained(
+            model_args.discriminator_name_or_path,
+            from_tf=bool(".ckpt" in model_args.discriminator_name_or_path),
+            config=discriminator_config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        discriminator = ElectraForPreTraining(discriminator_config)
+
+    if model_args.generator_name_or_path:
+        generator = ElectraForMaskedLM.from_pretrained(
+            model_args.generator_name_or_path,
+            from_tf=bool(".ckpt" in model_args.generator_name_or_path),
+            config=generator_config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        generator = ElectraForMaskedLM(generator_config)
+
+    discriminator.resize_token_embeddings(len(tokenizer))
+    generator.resize_token_embeddings(len(tokenizer))
+
+    if data_args.block_size <= 0:
+        data_args.block_size = tokenizer.max_len
+        # Our input block size will be the max possible for the model
+    else:
+        data_args.block_size = min(data_args.block_size, tokenizer.max_len)
+
+    # Need to update this to something cleaner
+    try:
+        import torch_xla.core.xla_model as xm
+
+        if xm.is_master_ordinal(local=True):
+            get_dataset(data_args, training_args, model_args, tokenizer=tokenizer, local_rank=training_args.local_rank)
+
+        xm.rendezvous("dataset building")
+    except ImportError:
+        logger.info("Not running on TPU")
+
+    # Get datasets
+    train_dataset = (
+        get_dataset(data_args, training_args, model_args, tokenizer=tokenizer, local_rank=training_args.local_rank)
+        if training_args.do_train
+        else None
+    )
+    eval_dataset = (
+        get_dataset(
+            data_args,
+            training_args,
+            model_args,
+            tokenizer=tokenizer,
+            local_rank=training_args.local_rank,
+            evaluate=True,
+        )
+        if training_args.do_eval
+        else None
+    )
+
+    # Masking is done inside the CombinedModel
+    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False, mlm_probability=0)
+
+    model = CombinedModel(discriminator, generator, tokenizer, training_args, data_args)
+
+    def compute_metrics(evaluation_predictions: EvalPrediction) -> Dict[str, int]:
+        predictions: Dict[str, np.ndarray] = evaluation_predictions.predictions
+        labels: Dict[str, np.ndarray] = evaluation_predictions.label_ids
+
+        generator_labels, generator_predictions = labels["generator"], predictions["generator"]
+        discriminator_labels, discriminator_predictions = labels["discriminator"], predictions["discriminator"]
+
+        true_positives = (
+            np.logical_and(
+                np.equal(discriminator_predictions, discriminator_labels), np.equal(discriminator_labels, 1)
+            )
+            .sum()
+            .astype(float)
+        )
+        false_negatives = (
+            np.logical_and(
+                np.not_equal(discriminator_predictions, discriminator_labels), np.equal(discriminator_labels, 1)
+            )
+            .sum()
+            .astype(float)
+        )
+        false_positives = (
+            np.logical_and(
+                np.not_equal(discriminator_predictions, discriminator_labels), np.equal(discriminator_labels, 0)
+            )
+            .sum()
+            .astype(float)
+        )
+
+        generator_accuracy = (
+            np.equal(generator_labels, generator_predictions).sum().astype(float) / generator_predictions.size
+        )
+        discriminator_accuracy = (
+            np.equal(discriminator_labels, discriminator_predictions).sum().astype(float) / discriminator_labels.size
+        )
+        discriminator_precision = true_positives / (true_positives + false_positives)
+        discriminator_recall = true_positives / (true_positives + false_negatives)
+
+        return {
+            "generator_accuracy": generator_accuracy,
+            "discriminator_accuracy": discriminator_accuracy,
+            "discriminator_precision": discriminator_precision,
+            "discriminator_recall": discriminator_recall,
+        }
+
+    def manage_evaluation_predictions(model_outputs: Tuple[torch.Tensor]) -> Dict[str, torch.Tensor]:
+        total_loss, models_output, generator_evaluation_values, discriminator_evaluation_values = model_outputs
+        generator_labels, generator_predictions = generator_evaluation_values
+        discriminator_labels, discriminator_predictions = discriminator_evaluation_values
+
+        return {
+            "generator_labels": generator_labels.detach(),
+            "generator_predictions": generator_predictions.detach(),
+            "discriminator_labels": discriminator_labels.detach(),
+            "discriminator_predictions": discriminator_predictions.detach(),
+        }
+
+    def eval_prediction_mapping(dictionary: Dict[str, np.ndarray]) -> EvalPrediction:
+        labels = {"generator": dictionary["generator_labels"], "discriminator": dictionary["discriminator_labels"]}
+        predictions = {
+            "generator": dictionary["generator_predictions"],
+            "discriminator": dictionary["discriminator_predictions"],
+        }
+
+        return EvalPrediction(labels, predictions)
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        data_collator=data_collator,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        compute_metrics=compute_metrics,
+        manage_evaluation_predictions=manage_evaluation_predictions,
+        eval_prediction_mapping=eval_prediction_mapping,
+        prediction_loss_only=True,
+    )
+
+    # Training
+    if training_args.do_train:
+        model_path = (
+            model_args.discriminator_name_or_path
+            if model_args.discriminator_name_or_path is not None
+            and os.path.isdir(model_args.discriminator_name_or_path)
+            else None
+        )
+        trainer.train(model_path=model_path)
+        trainer.save_model()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval and training_args.local_rank in [-1, 0]:
+        logger.info("*** Evaluate ***")
+
+        result = trainer.evaluate()
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results *****")
+            for key in sorted(result.keys()):
+                logger.info("  %s = %s", key, str(result[key]))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+        results.update(result)
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()

examples/movement-pruning/README.md

@@ -0,0 +1,183 @@
+# Movement Pruning: Adaptive Sparsity by Fine-Tuning
+
+*Magnitude pruning is a widely used strategy for reducing model size in pure supervised learning; however, it is less effective in the transfer learning regime that has become standard for state-of-the-art natural language processing applications. We propose the use of *movement pruning*, a simple, deterministic first-order weight pruning method that is more adaptive to pretrained model fine-tuning. Experiments show that when pruning large pretrained language models, movement pruning shows significant improvements in high-sparsity regimes. When combined with distillation, the approach achieves minimal accuracy loss with down to only 3% of the model parameters:*
+
+| Fine-pruning+Distillation<br>(Teacher=BERT-base fine-tuned) | BERT base<br>fine-tuned | Remaining<br>Weights (%) | Magnitude Pruning      | L0 Regularization      | Movement Pruning       | Soft Movement Pruning          |
+| :---:                                                       | :---:                   | :---:                    | :---:                  | :---:                  | :---:                  | :---:                          |
+| SQuAD - Dev<br>EM/F1                                        | 80.4/88.1               | 10%<br>3%                | 70.2/80.1<br>45.5/59.6 | 72.4/81.9<br>64.3/75.8 | 75.6/84.3<br>67.5/78.0 | **76.6/84.9**<br>**72.7/82.3** |
+| MNLI - Dev<br>acc/MM acc                                    | 84.5/84.9               | 10%<br>3%                | 78.3/79.3<br>69.4/70.6 | 78.7/79.7<br>76.0/76.2 | 80.1/80.4<br>76.5/77.4 | **81.2/81.8**<br>**79.5/80.1** |
+| QQP - Dev<br>acc/F1                                         | 91.4/88.4               | 10%<br>3%                | 79.8/65.0<br>72.4/57.8 | 88.1/82.8<br>87.0/81.9 | 89.7/86.2<br>86.1/81.5 | **90.2/86.8**<br>**89.1/85.5** |
+
+This page contains information on how to fine-prune pre-trained models such as `BERT` to obtain extremely sparse models with movement pruning. In contrast to magnitude pruning which selects weights that are far from 0, movement pruning retains weights that are moving away from 0.
+
+For more information, we invite you to check out [our paper](https://arxiv.org/abs/2005.07683).
+You can also have a look at this fun *Explain Like I'm Five* introductory [slide deck](https://www.slideshare.net/VictorSanh/movement-pruning-explain-like-im-five-234205241).
+
+<div align="center">
+<img src="https://www.seekpng.com/png/detail/166-1669328_how-to-make-emmental-cheese-at-home-icooker.png" width="400">
+</div>
+
+## Extreme sparsity and efficient storage
+
+One promise of extreme pruning is to obtain extremely small models that can be easily sent (and stored) on edge devices. By setting weights to 0., we reduce the amount of information we need to store, and thus decreasing the memory size. We are able to obtain extremely sparse fine-pruned models with movement pruning: ~95% of the dense performance with ~5% of total remaining weights in the BERT encoder.
+
+In [this notebook](https://github.com/huggingface/transformers/blob/master/examples/movement-pruning/Saving_PruneBERT.ipynb), we showcase how we can leverage standard tools that exist out-of-the-box to efficiently store an extremely sparse question answering model (only 6% of total remaining weights in the encoder). We are able to reduce the memory size of the encoder **from the 340MB (the orignal dense BERT) to 11MB**, without any additional training of the model (every operation is performed *post fine-pruning*). It is sufficiently small to store it on a [91' floppy disk](https://en.wikipedia.org/wiki/Floptical) 📎!
+
+While movement pruning does not directly optimize for memory footprint (but rather the number of non-null weights), we hypothetize that further memory compression ratios can be achieved with specific quantization aware trainings (see for instance [Q8BERT](https://arxiv.org/abs/1910.06188), [And the Bit Goes Down](https://arxiv.org/abs/1907.05686) or [Quant-Noise](https://arxiv.org/abs/2004.07320)).
+
+## Fine-pruned models
+
+As examples, we release two English PruneBERT checkpoints (models fine-pruned from a pre-trained `BERT` checkpoint), one on SQuAD and the other on MNLI.
+
+- **`prunebert-base-uncased-6-finepruned-w-distil-squad`**<br/>
+Pre-trained `BERT-base-uncased` fine-pruned with soft movement pruning on SQuAD v1.1. We use an additional distillation signal from `BERT-base-uncased` finetuned on SQuAD. The encoder counts 6% of total non-null weights and reaches 83.8 F1 score. The model can be accessed with: `pruned_bert = BertForQuestionAnswering.from_pretrained("huggingface/prunebert-base-uncased-6-finepruned-w-distil-squad")`
+- **`prunebert-base-uncased-6-finepruned-w-distil-mnli`**<br/>
+Pre-trained `BERT-base-uncased` fine-pruned with soft movement pruning on MNLI. We use an additional distillation signal from `BERT-base-uncased` finetuned on MNLI. The encoder counts 6% of total non-null weights and reaches 80.7 (matched) accuracy. The model can be accessed with: `pruned_bert = BertForSequenceClassification.from_pretrained("huggingface/prunebert-base-uncased-6-finepruned-w-distil-mnli")`
+
+## How to fine-prune?
+
+### Setup
+
+The code relies on the 🤗 Transformers library. In addition to the dependencies listed in the [`examples`](https://github.com/huggingface/transformers/tree/master/examples) folder, you should install a few additional dependencies listed in the `requirements.txt` file: `pip install -r requirements.txt`.
+
+Note that we built our experiments on top of a stabilized version of the library (commit https://github.com/huggingface/transformers/commit/352d5472b0c1dec0f420d606d16747d851b4bda8): we do not guarantee that everything is still compatible with the latest version of the master branch.
+
+### Fine-pruning with movement pruning
+
+Below, we detail how to reproduce the results reported in the paper. We use SQuAD as a running example. Commands (and scripts) can be easily adapted for other tasks.
+
+The following command fine-prunes a pre-trained `BERT-base` on SQuAD using movement pruning towards 15% of remaining weights (85% sparsity). Note that we freeze all the embeddings modules (from their pre-trained value) and only prune the Fully Connected layers in the encoder (12 layers of Transformer Block).
+
+```bash
+SERIALIZATION_DIR=<OUTPUT_DIR>
+SQUAD_DATA=<SQUAD_DATA>
+
+python examples/movement-pruning/masked_run_squad.py \
+    --output_dir $SERIALIZATION_DIR \
+    --data_dir $SQUAD_DATA \
+    --train_file train-v1.1.json \
+    --predict_file dev-v1.1.json \
+    --do_train --do_eval --do_lower_case \
+    --model_type masked_bert \
+    --model_name_or_path bert-base-uncased \
+    --per_gpu_train_batch_size 16 \
+    --warmup_steps 5400 \
+    --num_train_epochs 10 \
+    --learning_rate 3e-5 --mask_scores_learning_rate 1e-2 \
+    --initial_threshold 1 --final_threshold 0.15 \
+    --initial_warmup 1 --final_warmup 2 \
+    --pruning_method topK --mask_init constant --mask_scale 0.
+```
+
+### Fine-pruning with other methods
+
+We can also explore other fine-pruning methods by changing the `pruning_method` parameter:
+
+Soft movement pruning
+```bash
+python examples/movement-pruning/masked_run_squad.py \
+    --output_dir $SERIALIZATION_DIR \
+    --data_dir $SQUAD_DATA \
+    --train_file train-v1.1.json \
+    --predict_file dev-v1.1.json \
+    --do_train --do_eval --do_lower_case \
+    --model_type masked_bert \
+    --model_name_or_path bert-base-uncased \
+    --per_gpu_train_batch_size 16 \
+    --warmup_steps 5400 \
+    --num_train_epochs 10 \
+    --learning_rate 3e-5 --mask_scores_learning_rate 1e-2 \
+    --initial_threshold 0 --final_threshold 0.1 \
+    --initial_warmup 1 --final_warmup 2 \
+    --pruning_method sigmoied_threshold --mask_init constant --mask_scale 0. \
+    --regularization l1 --final_lambda 400.
+```
+
+L0 regularization
+```bash
+python examples/movement-pruning/masked_run_squad.py \
+    --output_dir $SERIALIZATION_DIR \
+    --data_dir $SQUAD_DATA \
+    --train_file train-v1.1.json \
+    --predict_file dev-v1.1.json \
+    --do_train --do_eval --do_lower_case \
+    --model_type masked_bert \
+    --model_name_or_path bert-base-uncased \
+    --per_gpu_train_batch_size 16 \
+    --warmup_steps 5400 \
+    --num_train_epochs 10 \
+    --learning_rate 3e-5 --mask_scores_learning_rate 1e-1 \
+    --initial_threshold 1. --final_threshold 1. \
+    --initial_warmup 1 --final_warmup 1 \
+    --pruning_method l0 --mask_init constant --mask_scale 2.197 \
+    --regularization l0 --final_lambda 125.
+```
+
+Iterative Magnitude Pruning
+```bash
+python examples/movement-pruning/masked_run_squad.py \
+    --output_dir ./dbg \
+    --data_dir examples/distillation/data/squad_data \
+    --train_file train-v1.1.json \
+    --predict_file dev-v1.1.json \
+    --do_train --do_eval --do_lower_case \
+    --model_type masked_bert \
+    --model_name_or_path bert-base-uncased \
+    --per_gpu_train_batch_size 16 \
+    --warmup_steps 5400 \
+    --num_train_epochs 10 \
+    --learning_rate 3e-5 \
+    --initial_threshold 1 --final_threshold 0.15 \
+    --initial_warmup 1 --final_warmup 2 \
+    --pruning_method magnitude
+```
+
+### After fine-pruning
+
+**Counting parameters**
+
+Regularization based pruning methods (soft movement pruning and L0 regularization) rely on the penalty to induce sparsity. The multiplicative coefficient controls the sparsity level.
+To obtain the effective sparsity level in the encoder, we simply count the number of activated (non-null) weights:
+
+```bash
+python examples/movement-pruning/count_parameters.py \
+    --pruning_method sigmoied_threshold \
+    --threshold 0.1 \
+    --serialization_dir $SERIALIZATION_DIR
+```
+
+**Pruning once for all**
+
+Once the model has been fine-pruned, the pruned weights can be set to 0. once for all (reducing the amount of information to store). In our running experiments, we can convert a `MaskedBertForQuestionAnswering` (a BERT model augmented to enable on-the-fly pruning capabilities) to a standard `BertForQuestionAnswering`:
+
+```bash
+python examples/movement-pruning/bertarize.py \
+    --pruning_method sigmoied_threshold \
+    --threshold 0.1 \
+    --model_name_or_path $SERIALIZATION_DIR
+```
+
+## Hyper-parameters
+
+For reproducibility purposes, we share the detailed results presented in the paper. These [tables](https://docs.google.com/spreadsheets/d/17JgRq_OFFTniUrz6BZWW_87DjFkKXpI1kYDSsseT_7g/edit?usp=sharing) exhaustively describe the individual hyper-parameters used for each data point.
+
+## Inference speed
+
+Early experiments show that even though models fine-pruned with (soft) movement pruning are extremely sparse, they do not benefit from significant improvement in terms of inference speed when using the standard PyTorch inference.
+We are currently benchmarking and exploring inference setups specifically for sparse architectures.
+In particular, hardware manufacturers are announcing devices that will speedup inference for sparse networks considerably.
+
+## Citation
+
+If you find this resource useful, please consider citing the following paper:
+
+```
+@article{sanh2020movement,
+    title={Movement Pruning: Adaptive Sparsity by Fine-Tuning},
+    author={Victor Sanh and Thomas Wolf and Alexander M. Rush},
+    year={2020},
+    eprint={2005.07683},
+    archivePrefix={arXiv},
+    primaryClass={cs.CL}
+}
+```

examples/movement-pruning/Saving_PruneBERT.ipynb

@@ -0,0 +1,612 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Saving PruneBERT\n",
+    "\n",
+    "\n",
+    "This notebook aims at showcasing how we can leverage standard tools to save (and load) an extremely sparse model fine-pruned with [movement pruning](https://arxiv.org/abs/2005.07683) (or any other unstructured pruning mehtod).\n",
+    "\n",
+    "In this example, we used BERT (base-uncased, but the procedure described here is not specific to BERT and can be applied to a large variety of models.\n",
+    "\n",
+    "We first obtain an extremely sparse model by fine-pruning with movement pruning on SQuAD v1.1. We then used the following combination of standard tools:\n",
+    "- We reduce the precision of the model with Int8 dynamic quantization using [PyTorch implementation](https://pytorch.org/tutorials/intermediate/dynamic_quantization_bert_tutorial.html). We only quantized the Fully Connected Layers.\n",
+    "- Sparse quantized matrices are converted into the [Compressed Sparse Row format](https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csr_matrix.html).\n",
+    "- We use HDF5 with `gzip` compression to store the weights.\n",
+    "\n",
+    "We experiment with a question answering model with only 6% of total remaining weights in the encoder (previously obtained with movement pruning). **We are able to reduce the memory size of the encoder from 340MB (original dense BERT) to 11MB**, which fits on a [91' floppy disk](https://en.wikipedia.org/wiki/Floptical)!\n",
+    "\n",
+    "<img src=\"https://upload.wikimedia.org/wikipedia/commons/thumb/0/00/Floptical_disk_21MB.jpg/440px-Floptical_disk_21MB.jpg\" width=\"200\">"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Includes\n",
+    "\n",
+    "import h5py\n",
+    "import os\n",
+    "import json\n",
+    "from collections import OrderedDict\n",
+    "\n",
+    "from scipy import sparse\n",
+    "import numpy as np\n",
+    "\n",
+    "import torch\n",
+    "from torch import nn\n",
+    "\n",
+    "from transformers import *\n",
+    "\n",
+    "os.chdir('../../')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Saving"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Dynamic quantization induces little or no loss of performance while significantly reducing the memory footprint."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load fine-pruned model and quantize the model\n",
+    "\n",
+    "model_path = \"serialization_dir/bert-base-uncased/92/squad/l1\"\n",
+    "model_name = \"bertarized_l1_with_distil_0._0.1_1_2_l1_1100._3e-5_1e-2_sigmoied_threshold_constant_0._10_epochs\"\n",
+    "\n",
+    "model = BertForQuestionAnswering.from_pretrained(os.path.join(model_path, model_name))\n",
+    "model.to('cpu')\n",
+    "\n",
+    "quantized_model = torch.quantization.quantize_dynamic(\n",
+    "                    model=model,\n",
+    "                    qconfig_spec = {\n",
+    "                        torch.nn.Linear : torch.quantization.default_dynamic_qconfig,\n",
+    "                    },\n",
+    "                    dtype=torch.qint8,\n",
+    "                )\n",
+    "# print(quantized_model)\n",
+    "\n",
+    "qtz_st = quantized_model.state_dict()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Saving the original (encoder + classifier) in the standard torch.save format\n",
+    "\n",
+    "dense_st = {name: param for name, param in model.state_dict().items() \n",
+    "                            if \"embedding\" not in name and \"pooler\" not in name}\n",
+    "torch.save(dense_st, 'dbg/dense_squad.pt',)\n",
+    "dense_mb_size = os.path.getsize(\"dbg/dense_squad.pt\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Decompose quantization for bert.encoder.layer.0.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.0.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.0.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.0.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.0.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.0.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.1.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.2.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.3.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.4.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.5.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.6.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.7.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.8.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.9.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.10.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.attention.self.query._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.attention.self.key._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.attention.self.value._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.attention.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.intermediate.dense._packed_params.weight\n",
+      "Decompose quantization for bert.encoder.layer.11.output.dense._packed_params.weight\n",
+      "Decompose quantization for bert.pooler.dense._packed_params.weight\n",
+      "Decompose quantization for qa_outputs._packed_params.weight\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Elementary representation: we decompose the quantized tensors into (scale, zero_point, int_repr).\n",
+    "# See https://pytorch.org/docs/stable/quantization.html\n",
+    "\n",
+    "# We further leverage the fact that int_repr is sparse matrix to optimize the storage: we decompose int_repr into\n",
+    "# its CSR representation (data, indptr, indices).\n",
+    "\n",
+    "elementary_qtz_st = {}\n",
+    "for name, param in qtz_st.items():\n",
+    "    if param.is_quantized:\n",
+    "        print(\"Decompose quantization for\", name)\n",
+    "        # We need to extract the scale, the zero_point and the int_repr for the quantized tensor and modules\n",
+    "        scale = param.q_scale()                                # torch.tensor(1,) - float32\n",
+    "        zero_point = param.q_zero_point()                      # torch.tensor(1,) - int32\n",
+    "        elementary_qtz_st[f\"{name}.scale\"] = scale\n",
+    "        elementary_qtz_st[f\"{name}.zero_point\"] = zero_point\n",
+    "\n",
+    "        # We assume the int_repr is sparse and compute its CSR representation\n",
+    "        # Only the FCs in the encoder are actually sparse\n",
+    "        int_repr = param.int_repr()                         # torch.tensor(nb_rows, nb_columns) - int8\n",
+    "        int_repr_cs = sparse.csr_matrix(int_repr)           # scipy.sparse.csr.csr_matrix\n",
+    "\n",
+    "        elementary_qtz_st[f\"{name}.int_repr.data\"] = int_repr_cs.data                  # np.array int8\n",
+    "        elementary_qtz_st[f\"{name}.int_repr.indptr\"] = int_repr_cs.indptr              # np.array int32\n",
+    "        assert max(int_repr_cs.indices) < 65535 # If not, we shall fall back to int32\n",
+    "        elementary_qtz_st[f\"{name}.int_repr.indices\"] = np.uint16(int_repr_cs.indices) # np.array uint16\n",
+    "        elementary_qtz_st[f\"{name}.int_repr.shape\"] = int_repr_cs.shape                # tuple(int, int)\n",
+    "    else:\n",
+    "        elementary_qtz_st[name] = param\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Encoder Size (MB) - Sparse & Quantized - `torch.save`: 21.29\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Saving the pruned (encoder + classifier) in the standard torch.save format\n",
+    "\n",
+    "dense_optimized_st = {name: param for name, param in elementary_qtz_st.items() \n",
+    "                                    if \"embedding\" not in name and \"pooler\" not in name}\n",
+    "torch.save(dense_optimized_st, 'dbg/dense_squad_optimized.pt',)\n",
+    "print(\"Encoder Size (MB) - Sparse & Quantized - `torch.save`:\",\n",
+    "      round(os.path.getsize(\"dbg/dense_squad_optimized.pt\")/1e6, 2))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Skip bert.embeddings.word_embeddings.weight\n",
+      "Skip bert.embeddings.position_embeddings.weight\n",
+      "Skip bert.embeddings.token_type_embeddings.weight\n",
+      "Skip bert.embeddings.LayerNorm.weight\n",
+      "Skip bert.embeddings.LayerNorm.bias\n",
+      "Skip bert.pooler.dense.scale\n",
+      "Skip bert.pooler.dense.zero_point\n",
+      "Skip bert.pooler.dense._packed_params.weight.scale\n",
+      "Skip bert.pooler.dense._packed_params.weight.zero_point\n",
+      "Skip bert.pooler.dense._packed_params.weight.int_repr.data\n",
+      "Skip bert.pooler.dense._packed_params.weight.int_repr.indptr\n",
+      "Skip bert.pooler.dense._packed_params.weight.int_repr.indices\n",
+      "Skip bert.pooler.dense._packed_params.weight.int_repr.shape\n",
+      "Skip bert.pooler.dense._packed_params.bias\n",
+      "\n",
+      "Encoder Size (MB) - Dense:              340.25\n",
+      "Encoder Size (MB) - Sparse & Quantized: 11.27\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Save the decomposed state_dict with an HDF5 file\n",
+    "# Saving only the encoder + QA Head\n",
+    "\n",
+    "with h5py.File('dbg/squad_sparse.h5','w') as hf:\n",
+    "    for name, param in elementary_qtz_st.items():\n",
+    "        if \"embedding\" in name:\n",
+    "            print(f\"Skip {name}\")\n",
+    "            continue\n",
+    "\n",
+    "        if \"pooler\" in name:\n",
+    "            print(f\"Skip {name}\")\n",
+    "            continue\n",
+    "\n",
+    "        if type(param) == torch.Tensor:\n",
+    "            if param.numel() == 1:\n",
+    "                # module scale\n",
+    "                # module zero_point\n",
+    "                hf.attrs[name] = param\n",
+    "                continue\n",
+    "\n",
+    "            if param.requires_grad:\n",
+    "                # LayerNorm\n",
+    "                param = param.detach().numpy()\n",
+    "            hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
+    "\n",
+    "        elif type(param) == float or type(param) == int or type(param) == tuple:\n",
+    "            # float - tensor _packed_params.weight.scale\n",
+    "            # int   - tensor_packed_params.weight.zero_point\n",
+    "            # tuple - tensor _packed_params.weight.shape\n",
+    "            hf.attrs[name] = param\n",
+    "\n",
+    "        else:\n",
+    "            hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
+    "\n",
+    "\n",
+    "with open('dbg/metadata.json', 'w') as f:\n",
+    "    f.write(json.dumps(qtz_st._metadata))  \n",
+    "\n",
+    "size = os.path.getsize(\"dbg/squad_sparse.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
+    "print(\"\")\n",
+    "print(\"Encoder Size (MB) - Dense:             \", round(dense_mb_size/1e6, 2))\n",
+    "print(\"Encoder Size (MB) - Sparse & Quantized:\", round(size/1e6, 2))\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "Size (MB): 99.39\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Save the decomposed state_dict to HDF5 storage\n",
+    "# Save everything in the architecutre (embedding + encoder + QA Head)\n",
+    "\n",
+    "with h5py.File('dbg/squad_sparse_with_embs.h5','w') as hf:\n",
+    "    for name, param in elementary_qtz_st.items():\n",
+    "#         if \"embedding\" in name:\n",
+    "#             print(f\"Skip {name}\")\n",
+    "#             continue\n",
+    "\n",
+    "#         if \"pooler\" in name:\n",
+    "#             print(f\"Skip {name}\")\n",
+    "#             continue\n",
+    "\n",
+    "        if type(param) == torch.Tensor:\n",
+    "            if param.numel() == 1:\n",
+    "                # module scale\n",
+    "                # module zero_point\n",
+    "                hf.attrs[name] = param\n",
+    "                continue\n",
+    "\n",
+    "            if param.requires_grad:\n",
+    "                # LayerNorm\n",
+    "                param = param.detach().numpy()\n",
+    "            hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
+    "\n",
+    "        elif type(param) == float or type(param) == int or type(param) == tuple:\n",
+    "            # float - tensor _packed_params.weight.scale\n",
+    "            # int   - tensor _packed_params.weight.zero_point\n",
+    "            # tuple - tensor _packed_params.weight.shape\n",
+    "            hf.attrs[name] = param\n",
+    "\n",
+    "        else:\n",
+    "            hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
+    "\n",
+    "\n",
+    "with open('dbg/metadata.json', 'w') as f:\n",
+    "    f.write(json.dumps(qtz_st._metadata))   \n",
+    "\n",
+    "size = os.path.getsize(\"dbg/squad_sparse_with_embs.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
+    "print('\\nSize (MB):', round(size/1e6, 2))\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Reconstruct the elementary state dict\n",
+    "\n",
+    "reconstructed_elementary_qtz_st = {}\n",
+    "\n",
+    "hf = h5py.File('dbg/squad_sparse_with_embs.h5','r')\n",
+    "\n",
+    "for attr_name, attr_param in hf.attrs.items():\n",
+    "    if 'shape' in attr_name:\n",
+    "        attr_param = tuple(attr_param)\n",
+    "    elif \".scale\" in attr_name:\n",
+    "        if \"_packed_params\" in attr_name:\n",
+    "            attr_param = float(attr_param)\n",
+    "        else:\n",
+    "            attr_param = torch.tensor(attr_param)\n",
+    "    elif \".zero_point\" in attr_name:\n",
+    "        if \"_packed_params\" in attr_name:\n",
+    "            attr_param = int(attr_param)\n",
+    "        else:\n",
+    "            attr_param = torch.tensor(attr_param)\n",
+    "    reconstructed_elementary_qtz_st[attr_name] = attr_param\n",
+    "    # print(f\"Unpack {attr_name}\")\n",
+    "    \n",
+    "# Get the tensors/arrays\n",
+    "for data_name, data_param in hf.items():\n",
+    "    if \"LayerNorm\" in data_name or \"_packed_params.bias\" in data_name:\n",
+    "        reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n",
+    "    elif \"embedding\" in data_name:\n",
+    "        reconstructed_elementary_qtz_st[data_name] = torch.from_numpy(np.array(data_param))\n",
+    "    else: # _packed_params.weight.int_repr.data, _packed_params.weight.int_repr.indices and _packed_params.weight.int_repr.indptr\n",
+    "        data_param = np.array(data_param)\n",
+    "        if \"indices\" in data_name:\n",
+    "            data_param = np.array(data_param, dtype=np.int32)\n",
+    "        reconstructed_elementary_qtz_st[data_name] = data_param\n",
+    "    # print(f\"Unpack {data_name}\")\n",
+    "    \n",
+    "\n",
+    "hf.close()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Sanity checks\n",
+    "\n",
+    "for name, param in reconstructed_elementary_qtz_st.items():\n",
+    "    assert name in elementary_qtz_st\n",
+    "for name, param in elementary_qtz_st.items():\n",
+    "    assert name in reconstructed_elementary_qtz_st, name\n",
+    "\n",
+    "for name, param in reconstructed_elementary_qtz_st.items():\n",
+    "    assert type(param) == type(elementary_qtz_st[name]), name\n",
+    "    if type(param) == torch.Tensor:\n",
+    "        assert torch.all(torch.eq(param, elementary_qtz_st[name])), name\n",
+    "    elif type(param) == np.ndarray:\n",
+    "        assert (param == elementary_qtz_st[name]).all(), name\n",
+    "    else:\n",
+    "        assert param == elementary_qtz_st[name], name"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Re-assemble the sparse int_repr from the CSR format\n",
+    "\n",
+    "reconstructed_qtz_st = {}\n",
+    "\n",
+    "for name, param in reconstructed_elementary_qtz_st.items():\n",
+    "    if \"weight.int_repr.indptr\" in name:\n",
+    "        prefix_ = name[:-16]\n",
+    "        data    = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.data\"]\n",
+    "        indptr  = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indptr\"]\n",
+    "        indices = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indices\"]\n",
+    "        shape   = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.shape\"]\n",
+    "\n",
+    "        int_repr = sparse.csr_matrix(arg1=(data, indices, indptr),\n",
+    "                                     shape=shape)\n",
+    "        int_repr = torch.tensor(int_repr.todense())\n",
+    "\n",
+    "        scale = reconstructed_elementary_qtz_st[f\"{prefix_}.scale\"]\n",
+    "        zero_point = reconstructed_elementary_qtz_st[f\"{prefix_}.zero_point\"]\n",
+    "        weight = torch._make_per_tensor_quantized_tensor(int_repr,\n",
+    "                                                         scale,\n",
+    "                                                         zero_point)\n",
+    "\n",
+    "        reconstructed_qtz_st[f\"{prefix_}\"] = weight\n",
+    "    elif \"int_repr.data\" in name or \"int_repr.shape\" in name or \"int_repr.indices\" in name or \\\n",
+    "         \"weight.scale\" in name or \"weight.zero_point\" in name:\n",
+    "        continue\n",
+    "    else:\n",
+    "        reconstructed_qtz_st[name] = param\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Sanity checks\n",
+    "\n",
+    "for name, param in reconstructed_qtz_st.items():\n",
+    "    assert name in qtz_st\n",
+    "for name, param in qtz_st.items():\n",
+    "    assert name in reconstructed_qtz_st, name\n",
+    "\n",
+    "for name, param in reconstructed_qtz_st.items():\n",
+    "    assert type(param) == type(qtz_st[name]), name\n",
+    "    if type(param) == torch.Tensor:\n",
+    "        assert torch.all(torch.eq(param, qtz_st[name])), name\n",
+    "    elif type(param) == np.ndarray:\n",
+    "        assert (param == qtz_st[name]).all(), name\n",
+    "    else:\n",
+    "        assert param == qtz_st[name], name"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Sanity checks"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<All keys matched successfully>"
+      ]
+     },
+     "execution_count": 12,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# Load the re-constructed state dict into a model\n",
+    "\n",
+    "dummy_model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')\n",
+    "dummy_model.to('cpu')\n",
+    "\n",
+    "reconstructed_qtz_model = torch.quantization.quantize_dynamic(\n",
+    "                            model=dummy_model,\n",
+    "                            qconfig_spec = None,\n",
+    "                            dtype=torch.qint8,\n",
+    "                          )\n",
+    "\n",
+    "reconstructed_qtz_st = OrderedDict(reconstructed_qtz_st)\n",
+    "with open('dbg/metadata.json', 'r') as read_file:\n",
+    "    metadata = json.loads(read_file.read())\n",
+    "reconstructed_qtz_st._metadata = metadata\n",
+    "\n",
+    "reconstructed_qtz_model.load_state_dict(reconstructed_qtz_st)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Sanity check passed\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Sanity checks on the infernce\n",
+    "\n",
+    "N = 32\n",
+    "\n",
+    "for _ in range(25):\n",
+    "    inputs = torch.randint(low=0, high=30000, size=(N, 128))\n",
+    "    mask = torch.ones(size=(N, 128))\n",
+    "\n",
+    "    y_reconstructed = reconstructed_qtz_model(input_ids=inputs, attention_mask=mask)[0]\n",
+    "    y               = quantized_model(input_ids=inputs, attention_mask=mask)[0]\n",
+    "    \n",
+    "    assert torch.all(torch.eq(y, y_reconstructed))\n",
+    "print(\"Sanity check passed\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.6.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

examples/movement-pruning/bertarize.py

@@ -0,0 +1,132 @@
+# Copyright 2020-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Once a model has been fine-pruned, the weights that are masked during the forward pass can be pruned once for all.
+For instance, once the a model from the :class:`~emmental.MaskedBertForSequenceClassification` is trained, it can be saved (and then loaded)
+as a standard :class:`~transformers.BertForSequenceClassification`.
+"""
+
+import argparse
+import os
+import shutil
+
+import torch
+
+from emmental.modules import MagnitudeBinarizer, ThresholdBinarizer, TopKBinarizer
+
+
+def main(args):
+    pruning_method = args.pruning_method
+    threshold = args.threshold
+
+    model_name_or_path = args.model_name_or_path.rstrip("/")
+    target_model_path = args.target_model_path
+
+    print(f"Load fine-pruned model from {model_name_or_path}")
+    model = torch.load(os.path.join(model_name_or_path, "pytorch_model.bin"))
+    pruned_model = {}
+
+    for name, tensor in model.items():
+        if "embeddings" in name or "LayerNorm" in name or "pooler" in name:
+            pruned_model[name] = tensor
+            print(f"Copied layer {name}")
+        elif "classifier" in name or "qa_output" in name:
+            pruned_model[name] = tensor
+            print(f"Copied layer {name}")
+        elif "bias" in name:
+            pruned_model[name] = tensor
+            print(f"Copied layer {name}")
+        else:
+            if pruning_method == "magnitude":
+                mask = MagnitudeBinarizer.apply(inputs=tensor, threshold=threshold)
+                pruned_model[name] = tensor * mask
+                print(f"Pruned layer {name}")
+            elif pruning_method == "topK":
+                if "mask_scores" in name:
+                    continue
+                prefix_ = name[:-6]
+                scores = model[f"{prefix_}mask_scores"]
+                mask = TopKBinarizer.apply(scores, threshold)
+                pruned_model[name] = tensor * mask
+                print(f"Pruned layer {name}")
+            elif pruning_method == "sigmoied_threshold":
+                if "mask_scores" in name:
+                    continue
+                prefix_ = name[:-6]
+                scores = model[f"{prefix_}mask_scores"]
+                mask = ThresholdBinarizer.apply(scores, threshold, True)
+                pruned_model[name] = tensor * mask
+                print(f"Pruned layer {name}")
+            elif pruning_method == "l0":
+                if "mask_scores" in name:
+                    continue
+                prefix_ = name[:-6]
+                scores = model[f"{prefix_}mask_scores"]
+                l, r = -0.1, 1.1
+                s = torch.sigmoid(scores)
+                s_bar = s * (r - l) + l
+                mask = s_bar.clamp(min=0.0, max=1.0)
+                pruned_model[name] = tensor * mask
+                print(f"Pruned layer {name}")
+            else:
+                raise ValueError("Unknown pruning method")
+
+    if target_model_path is None:
+        target_model_path = os.path.join(
+            os.path.dirname(model_name_or_path), f"bertarized_{os.path.basename(model_name_or_path)}"
+        )
+
+    if not os.path.isdir(target_model_path):
+        shutil.copytree(model_name_or_path, target_model_path)
+        print(f"\nCreated folder {target_model_path}")
+
+    torch.save(pruned_model, os.path.join(target_model_path, "pytorch_model.bin"))
+    print("\nPruned model saved! See you later!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--pruning_method",
+        choices=["l0", "magnitude", "topK", "sigmoied_threshold"],
+        type=str,
+        required=True,
+        help="Pruning Method (l0 = L0 regularization, magnitude = Magnitude pruning, topK = Movement pruning, sigmoied_threshold = Soft movement pruning)",
+    )
+    parser.add_argument(
+        "--threshold",
+        type=float,
+        required=False,
+        help="For `magnitude` and `topK`, it is the level of remaining weights (in %) in the fine-pruned model."
+        "For `sigmoied_threshold`, it is the threshold \tau against which the (sigmoied) scores are compared."
+        "Not needed for `l0`",
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        type=str,
+        required=True,
+        help="Folder containing the model that was previously fine-pruned",
+    )
+    parser.add_argument(
+        "--target_model_path",
+        default=None,
+        type=str,
+        required=False,
+        help="Folder containing the model that was previously fine-pruned",
+    )
+
+    args = parser.parse_args()
+
+    main(args)

examples/movement-pruning/counts_parameters.py

@@ -0,0 +1,92 @@
+# Copyright 2020-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Count remaining (non-zero) weights in the encoder (i.e. the transformer layers).
+Sparsity and remaining weights levels are equivalent: sparsity % = 100 - remaining weights %.
+"""
+import argparse
+import os
+
+import torch
+
+from emmental.modules import ThresholdBinarizer, TopKBinarizer
+
+
+def main(args):
+    serialization_dir = args.serialization_dir
+    pruning_method = args.pruning_method
+    threshold = args.threshold
+
+    st = torch.load(os.path.join(serialization_dir, "pytorch_model.bin"), map_location="cpu")
+
+    remaining_count = 0  # Number of remaining (not pruned) params in the encoder
+    encoder_count = 0  # Number of params in the encoder
+
+    print("name".ljust(60, " "), "Remaining Weights %", "Remaning Weight")
+    for name, param in st.items():
+        if "encoder" not in name:
+            continue
+
+        if "mask_scores" in name:
+            if pruning_method == "topK":
+                mask_ones = TopKBinarizer.apply(param, threshold).sum().item()
+            elif pruning_method == "sigmoied_threshold":
+                mask_ones = ThresholdBinarizer.apply(param, threshold, True).sum().item()
+            elif pruning_method == "l0":
+                l, r = -0.1, 1.1
+                s = torch.sigmoid(param)
+                s_bar = s * (r - l) + l
+                mask = s_bar.clamp(min=0.0, max=1.0)
+                mask_ones = (mask > 0.0).sum().item()
+            else:
+                raise ValueError("Unknown pruning method")
+            remaining_count += mask_ones
+            print(name.ljust(60, " "), str(round(100 * mask_ones / param.numel(), 3)).ljust(20, " "), str(mask_ones))
+        else:
+            encoder_count += param.numel()
+            if "bias" in name or "LayerNorm" in name:
+                remaining_count += param.numel()
+
+    print("")
+    print("Remaining Weights (global) %: ", 100 * remaining_count / encoder_count)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--pruning_method",
+        choices=["l0", "topK", "sigmoied_threshold"],
+        type=str,
+        required=True,
+        help="Pruning Method (l0 = L0 regularization, topK = Movement pruning, sigmoied_threshold = Soft movement pruning)",
+    )
+    parser.add_argument(
+        "--threshold",
+        type=float,
+        required=False,
+        help="For `topK`, it is the level of remaining weights (in %) in the fine-pruned model."
+        "For `sigmoied_threshold`, it is the threshold \tau against which the (sigmoied) scores are compared."
+        "Not needed for `l0`",
+    )
+    parser.add_argument(
+        "--serialization_dir",
+        type=str,
+        required=True,
+        help="Folder containing the model that was previously fine-pruned",
+    )
+
+    args = parser.parse_args()
+
+    main(args)

examples/movement-pruning/emmental/__init__.py

@@ -0,0 +1,10 @@
+# flake8: noqa
+from .configuration_bert_masked import MaskedBertConfig
+from .modeling_bert_masked import (
+    MaskedBertForMultipleChoice,
+    MaskedBertForQuestionAnswering,
+    MaskedBertForSequenceClassification,
+    MaskedBertForTokenClassification,
+    MaskedBertModel,
+)
+from .modules import *

examples/movement-pruning/emmental/configuration_bert_masked.py

@@ -0,0 +1,73 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Masked BERT model configuration. It replicates the class `~transformers.BertConfig`
+and adapts it to the specificities of MaskedBert (`pruning_method`, `mask_init` and `mask_scale`."""
+
+
+import logging
+
+from transformers.configuration_bert import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from transformers.configuration_utils import PretrainedConfig
+
+
+logger = logging.getLogger(__name__)
+
+
+class MaskedBertConfig(PretrainedConfig):
+    """
+    A class replicating the `~transformers.BertConfig` with additional parameters for pruning/masking configuration.
+    """
+
+    pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+    model_type = "masked_bert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        pruning_method="topK",
+        mask_init="constant",
+        mask_scale=0.0,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.pruning_method = pruning_method
+        self.mask_init = mask_init
+        self.mask_scale = mask_scale

examples/movement-pruning/emmental/modules/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+from .binarizer import MagnitudeBinarizer, ThresholdBinarizer, TopKBinarizer
+from .masked_nn import MaskedLinear

examples/movement-pruning/emmental/modules/binarizer.py

@@ -0,0 +1,144 @@
+# coding=utf-8
+# Copyright 2020-present, AllenAI Authors, University of Illinois Urbana-Champaign,
+# Intel Nervana Systems 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.
+"""
+Binarizers take a (real value) matrice as input and produce a binary (values in {0,1}) mask of the same shape.
+"""
+
+import torch
+from torch import autograd
+
+
+class ThresholdBinarizer(autograd.Function):
+    """
+    Thresholdd binarizer.
+    Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j} > \tau`
+    where `\tau` is a real value threshold.
+
+    Implementation is inspired from:
+        https://github.com/arunmallya/piggyback
+        Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights
+        Arun Mallya, Dillon Davis, Svetlana Lazebnik
+    """
+
+    @staticmethod
+    def forward(ctx, inputs: torch.tensor, threshold: float, sigmoid: bool):
+        """
+        Args:
+            inputs (`torch.FloatTensor`)
+                The input matrix from which the binarizer computes the binary mask.
+            threshold (`float`)
+                The threshold value (in R).
+            sigmoid (`bool`)
+                If set to ``True``, we apply the sigmoid function to the `inputs` matrix before comparing to `threshold`.
+                In this case, `threshold` should be a value between 0 and 1.
+        Returns:
+            mask (`torch.FloatTensor`)
+                Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
+                retained, 0 - the associated weight is pruned).
+        """
+        nb_elems = inputs.numel()
+        nb_min = int(0.005 * nb_elems) + 1
+        if sigmoid:
+            mask = (torch.sigmoid(inputs) > threshold).type(inputs.type())
+        else:
+            mask = (inputs > threshold).type(inputs.type())
+        if mask.sum() < nb_min:
+            # We limit the pruning so that at least 0.5% (half a percent) of the weights are remaining
+            k_threshold = inputs.flatten().kthvalue(max(nb_elems - nb_min, 1)).values
+            mask = (inputs > k_threshold).type(inputs.type())
+        return mask
+
+    @staticmethod
+    def backward(ctx, gradOutput):
+        return gradOutput, None, None
+
+
+class TopKBinarizer(autograd.Function):
+    """
+    Top-k Binarizer.
+    Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j}`
+    is among the k% highest values of S.
+
+    Implementation is inspired from:
+        https://github.com/allenai/hidden-networks
+        What's hidden in a randomly weighted neural network?
+        Vivek Ramanujan*, Mitchell Wortsman*, Aniruddha Kembhavi, Ali Farhadi, Mohammad Rastegari
+    """
+
+    @staticmethod
+    def forward(ctx, inputs: torch.tensor, threshold: float):
+        """
+        Args:
+            inputs (`torch.FloatTensor`)
+                The input matrix from which the binarizer computes the binary mask.
+            threshold (`float`)
+                The percentage of weights to keep (the rest is pruned).
+                `threshold` is a float between 0 and 1.
+        Returns:
+            mask (`torch.FloatTensor`)
+                Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
+                retained, 0 - the associated weight is pruned).
+        """
+        # Get the subnetwork by sorting the inputs and using the top threshold %
+        mask = inputs.clone()
+        _, idx = inputs.flatten().sort(descending=True)
+        j = int(threshold * inputs.numel())
+
+        # flat_out and mask access the same memory.
+        flat_out = mask.flatten()
+        flat_out[idx[j:]] = 0
+        flat_out[idx[:j]] = 1
+        return mask
+
+    @staticmethod
+    def backward(ctx, gradOutput):
+        return gradOutput, None
+
+
+class MagnitudeBinarizer(object):
+    """
+    Magnitude Binarizer.
+    Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j}`
+    is among the k% highest values of |S| (absolute value).
+
+    Implementation is inspired from https://github.com/NervanaSystems/distiller/blob/2291fdcc2ea642a98d4e20629acb5a9e2e04b4e6/distiller/pruning/automated_gradual_pruner.py#L24
+    """
+
+    @staticmethod
+    def apply(inputs: torch.tensor, threshold: float):
+        """
+        Args:
+            inputs (`torch.FloatTensor`)
+                The input matrix from which the binarizer computes the binary mask.
+                This input marix is typically the weight matrix.
+            threshold (`float`)
+                The percentage of weights to keep (the rest is pruned).
+                `threshold` is a float between 0 and 1.
+        Returns:
+            mask (`torch.FloatTensor`)
+                Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
+                retained, 0 - the associated weight is pruned).
+        """
+        # Get the subnetwork by sorting the inputs and using the top threshold %
+        mask = inputs.clone()
+        _, idx = inputs.abs().flatten().sort(descending=True)
+        j = int(threshold * inputs.numel())
+
+        # flat_out and mask access the same memory.
+        flat_out = mask.flatten()
+        flat_out[idx[j:]] = 0
+        flat_out[idx[:j]] = 1
+        return mask

examples/movement-pruning/emmental/modules/masked_nn.py

@@ -0,0 +1,107 @@
+# coding=utf-8
+# Copyright 2020-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Masked Linear module: A fully connected layer that computes an adaptive binary mask on the fly.
+The mask (binary or not) is computed at each forward pass and multiplied against
+the weight matrix to prune a portion of the weights.
+The pruned weight matrix is then multiplied against the inputs (and if necessary, the bias is added).
+"""
+
+import math
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.nn import init
+
+from .binarizer import MagnitudeBinarizer, ThresholdBinarizer, TopKBinarizer
+
+
+class MaskedLinear(nn.Linear):
+    """
+    Fully Connected layer with on the fly adaptive mask.
+    If needed, a score matrix is created to store the importance of each associated weight.
+    """
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        bias: bool = True,
+        mask_init: str = "constant",
+        mask_scale: float = 0.0,
+        pruning_method: str = "topK",
+    ):
+        """
+        Args:
+            in_features (`int`)
+                Size of each input sample
+            out_features (`int`)
+                Size of each output sample
+            bias (`bool`)
+                If set to ``False``, the layer will not learn an additive bias.
+                Default: ``True``
+            mask_init (`str`)
+                The initialization method for the score matrix if a score matrix is needed.
+                Choices: ["constant", "uniform", "kaiming"]
+                Default: ``constant``
+            mask_scale (`float`)
+                The initialization parameter for the chosen initialization method `mask_init`.
+                Default: ``0.``
+            pruning_method (`str`)
+                Method to compute the mask.
+                Choices: ["topK", "threshold", "sigmoied_threshold", "magnitude", "l0"]
+                Default: ``topK``
+        """
+        super(MaskedLinear, self).__init__(in_features=in_features, out_features=out_features, bias=bias)
+        assert pruning_method in ["topK", "threshold", "sigmoied_threshold", "magnitude", "l0"]
+        self.pruning_method = pruning_method
+
+        if self.pruning_method in ["topK", "threshold", "sigmoied_threshold", "l0"]:
+            self.mask_scale = mask_scale
+            self.mask_init = mask_init
+            self.mask_scores = nn.Parameter(torch.Tensor(self.weight.size()))
+            self.init_mask()
+
+    def init_mask(self):
+        if self.mask_init == "constant":
+            init.constant_(self.mask_scores, val=self.mask_scale)
+        elif self.mask_init == "uniform":
+            init.uniform_(self.mask_scores, a=-self.mask_scale, b=self.mask_scale)
+        elif self.mask_init == "kaiming":
+            init.kaiming_uniform_(self.mask_scores, a=math.sqrt(5))
+
+    def forward(self, input: torch.tensor, threshold: float):
+        # Get the mask
+        if self.pruning_method == "topK":
+            mask = TopKBinarizer.apply(self.mask_scores, threshold)
+        elif self.pruning_method in ["threshold", "sigmoied_threshold"]:
+            sig = "sigmoied" in self.pruning_method
+            mask = ThresholdBinarizer.apply(self.mask_scores, threshold, sig)
+        elif self.pruning_method == "magnitude":
+            mask = MagnitudeBinarizer.apply(self.weight, threshold)
+        elif self.pruning_method == "l0":
+            l, r, b = -0.1, 1.1, 2 / 3
+            if self.training:
+                u = torch.zeros_like(self.mask_scores).uniform_().clamp(0.0001, 0.9999)
+                s = torch.sigmoid((u.log() - (1 - u).log() + self.mask_scores) / b)
+            else:
+                s = torch.sigmoid(self.mask_scores)
+            s_bar = s * (r - l) + l
+            mask = s_bar.clamp(min=0.0, max=1.0)
+        # Mask weights with computed mask
+        weight_thresholded = mask * self.weight
+        # Compute output (linear layer) with masked weights
+        return F.linear(input, weight_thresholded, self.bias)

examples/movement-pruning/masked_run_glue.py

@@ -0,0 +1,926 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-pruning Masked BERT on sequence classification on GLUE."""
+
+import argparse
+import glob
+import json
+import logging
+import os
+import random
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from emmental import MaskedBertConfig, MaskedBertForSequenceClassification
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    BertConfig,
+    BertForSequenceClassification,
+    BertTokenizer,
+    get_linear_schedule_with_warmup,
+)
+from transformers import glue_compute_metrics as compute_metrics
+from transformers import glue_convert_examples_to_features as convert_examples_to_features
+from transformers import glue_output_modes as output_modes
+from transformers import glue_processors as processors
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter
+
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig,)), (),)
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
+    "masked_bert": (MaskedBertConfig, MaskedBertForSequenceClassification, BertTokenizer),
+}
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def schedule_threshold(
+    step: int,
+    total_step: int,
+    warmup_steps: int,
+    initial_threshold: float,
+    final_threshold: float,
+    initial_warmup: int,
+    final_warmup: int,
+    final_lambda: float,
+):
+    if step <= initial_warmup * warmup_steps:
+        threshold = initial_threshold
+    elif step > (total_step - final_warmup * warmup_steps):
+        threshold = final_threshold
+    else:
+        spars_warmup_steps = initial_warmup * warmup_steps
+        spars_schedu_steps = (final_warmup + initial_warmup) * warmup_steps
+        mul_coeff = 1 - (step - spars_warmup_steps) / (total_step - spars_schedu_steps)
+        threshold = final_threshold + (initial_threshold - final_threshold) * (mul_coeff ** 3)
+    regu_lambda = final_lambda * threshold / final_threshold
+    return threshold, regu_lambda
+
+
+def regularization(model: nn.Module, mode: str):
+    regu, counter = 0, 0
+    for name, param in model.named_parameters():
+        if "mask_scores" in name:
+            if mode == "l1":
+                regu += torch.norm(torch.sigmoid(param), p=1) / param.numel()
+            elif mode == "l0":
+                regu += torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1)).sum() / param.numel()
+            else:
+                ValueError("Don't know this mode.")
+            counter += 1
+    return regu / counter
+
+
+def train(args, train_dataset, model, tokenizer, teacher=None):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter(log_dir=args.output_dir)
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if "mask_score" in n and p.requires_grad],
+            "lr": args.mask_scores_learning_rate,
+        },
+        {
+            "params": [
+                p
+                for n, p in model.named_parameters()
+                if "mask_score" not in n and p.requires_grad and not any(nd in n for nd in no_decay)
+            ],
+            "lr": args.learning_rate,
+            "weight_decay": args.weight_decay,
+        },
+        {
+            "params": [
+                p
+                for n, p in model.named_parameters()
+                if "mask_score" not in n and p.requires_grad and any(nd in n for nd in no_decay)
+            ],
+            "lr": args.learning_rate,
+            "weight_decay": 0.0,
+        },
+    ]
+
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
+    )
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
+        os.path.join(args.model_name_or_path, "scheduler.pt")
+    ):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
+
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True,
+        )
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size
+        * args.gradient_accumulation_steps
+        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+    # Distillation
+    if teacher is not None:
+        logger.info("  Training with distillation")
+
+    global_step = 0
+    # Global TopK
+    if args.global_topk:
+        threshold_mem = None
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to global_step of last saved checkpoint from model path
+        try:
+            global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
+        except ValueError:
+            global_step = 0
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(
+        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0],
+    )
+    set_seed(args)  # Added here for reproductibility
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            threshold, regu_lambda = schedule_threshold(
+                step=global_step,
+                total_step=t_total,
+                warmup_steps=args.warmup_steps,
+                final_threshold=args.final_threshold,
+                initial_threshold=args.initial_threshold,
+                final_warmup=args.final_warmup,
+                initial_warmup=args.initial_warmup,
+                final_lambda=args.final_lambda,
+            )
+            # Global TopK
+            if args.global_topk:
+                if threshold == 1.0:
+                    threshold = -1e2  # Or an indefinitely low quantity
+                else:
+                    if (threshold_mem is None) or (global_step % args.global_topk_frequency_compute == 0):
+                        # Sort all the values to get the global topK
+                        concat = torch.cat(
+                            [param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
+                        )
+                        n = concat.numel()
+                        kth = max(n - (int(n * threshold) + 1), 1)
+                        threshold_mem = concat.kthvalue(kth).values.item()
+                        threshold = threshold_mem
+                    else:
+                        threshold = threshold_mem
+            inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = (
+                    batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
+                )  # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
+
+            if "masked" in args.model_type:
+                inputs["threshold"] = threshold
+
+            outputs = model(**inputs)
+            loss, logits_stu = outputs  # model outputs are always tuple in transformers (see doc)
+
+            # Distillation loss
+            if teacher is not None:
+                if "token_type_ids" not in inputs:
+                    inputs["token_type_ids"] = None if args.teacher_type == "xlm" else batch[2]
+                with torch.no_grad():
+                    (logits_tea,) = teacher(
+                        input_ids=inputs["input_ids"],
+                        token_type_ids=inputs["token_type_ids"],
+                        attention_mask=inputs["attention_mask"],
+                    )
+
+                loss_logits = F.kl_div(
+                    input=F.log_softmax(logits_stu / args.temperature, dim=-1),
+                    target=F.softmax(logits_tea / args.temperature, dim=-1),
+                    reduction="batchmean",
+                ) * (args.temperature ** 2)
+
+                loss = args.alpha_distil * loss_logits + args.alpha_ce * loss
+
+            # Regularization
+            if args.regularization is not None:
+                regu_ = regularization(model=model, mode=args.regularization)
+                loss = loss + regu_lambda * regu_
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0 or (
+                # last step in epoch but step is always smaller than gradient_accumulation_steps
+                len(epoch_iterator) <= args.gradient_accumulation_steps
+                and (step + 1) == len(epoch_iterator)
+            ):
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    tb_writer.add_scalar("threshold", threshold, global_step)
+                    for name, param in model.named_parameters():
+                        if not param.requires_grad:
+                            continue
+                        tb_writer.add_scalar("parameter_mean/" + name, param.data.mean(), global_step)
+                        tb_writer.add_scalar("parameter_std/" + name, param.data.std(), global_step)
+                        tb_writer.add_scalar("parameter_min/" + name, param.data.min(), global_step)
+                        tb_writer.add_scalar("parameter_max/" + name, param.data.max(), global_step)
+                        tb_writer.add_scalar("grad_mean/" + name, param.grad.data.mean(), global_step)
+                        tb_writer.add_scalar("grad_std/" + name, param.grad.data.std(), global_step)
+                        if args.regularization is not None and "mask_scores" in name:
+                            if args.regularization == "l1":
+                                perc = (torch.sigmoid(param) > threshold).sum().item() / param.numel()
+                            elif args.regularization == "l0":
+                                perc = (torch.sigmoid(param - 2 / 3 * np.log(0.1 / 1.1))).sum().item() / param.numel()
+                            tb_writer.add_scalar("retained_weights_perc/" + name, perc, global_step)
+
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    logs = {}
+                    if (
+                        args.local_rank == -1 and args.evaluate_during_training
+                    ):  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            eval_key = "eval_{}".format(key)
+                            logs[eval_key] = value
+
+                    loss_scalar = (tr_loss - logging_loss) / args.logging_steps
+                    learning_rate_scalar = scheduler.get_lr()
+                    logs["learning_rate"] = learning_rate_scalar[0]
+                    if len(learning_rate_scalar) > 1:
+                        for idx, lr in enumerate(learning_rate_scalar[1:]):
+                            logs[f"learning_rate/{idx+1}"] = lr
+                    logs["loss"] = loss_scalar
+                    if teacher is not None:
+                        logs["loss/distil"] = loss_logits.item()
+                    if args.regularization is not None:
+                        logs["loss/regularization"] = regu_.item()
+                    if (teacher is not None) or (args.regularization is not None):
+                        if (teacher is not None) and (args.regularization is not None):
+                            logs["loss/instant_ce"] = (
+                                loss.item()
+                                - regu_lambda * logs["loss/regularization"]
+                                - args.alpha_distil * logs["loss/distil"]
+                            ) / args.alpha_ce
+                        elif teacher is not None:
+                            logs["loss/instant_ce"] = (
+                                loss.item() - args.alpha_distil * logs["loss/distil"]
+                            ) / args.alpha_ce
+                        else:
+                            logs["loss/instant_ce"] = loss.item() - regu_lambda * logs["loss/regularization"]
+                    logging_loss = tr_loss
+
+                    for key, value in logs.items():
+                        tb_writer.add_scalar(key, value, global_step)
+                    print(json.dumps({**logs, **{"step": global_step}}))
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = (
+                        model.module if hasattr(model, "module") else model
+                    )  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    # Loop to handle MNLI double evaluation (matched, mis-matched)
+    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
+    eval_outputs_dirs = (args.output_dir, args.output_dir + "/MM") if args.task_name == "mnli" else (args.output_dir,)
+
+    results = {}
+    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
+        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
+
+        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(eval_output_dir)
+
+        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+        # Note that DistributedSampler samples randomly
+        eval_sampler = SequentialSampler(eval_dataset)
+        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+        # multi-gpu eval
+        if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
+            model = torch.nn.DataParallel(model)
+
+        # Eval!
+        logger.info("***** Running evaluation {} *****".format(prefix))
+        logger.info("  Num examples = %d", len(eval_dataset))
+        logger.info("  Batch size = %d", args.eval_batch_size)
+        eval_loss = 0.0
+        nb_eval_steps = 0
+        preds = None
+        out_label_ids = None
+
+        # Global TopK
+        if args.global_topk:
+            threshold_mem = None
+
+        for batch in tqdm(eval_dataloader, desc="Evaluating"):
+            model.eval()
+            batch = tuple(t.to(args.device) for t in batch)
+
+            with torch.no_grad():
+                inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
+                if args.model_type != "distilbert":
+                    inputs["token_type_ids"] = (
+                        batch[2] if args.model_type in ["bert", "masked_bert", "xlnet", "albert"] else None
+                    )  # XLM, DistilBERT, RoBERTa, and XLM-RoBERTa don't use segment_ids
+                if "masked" in args.model_type:
+                    inputs["threshold"] = args.final_threshold
+                    if args.global_topk:
+                        if threshold_mem is None:
+                            concat = torch.cat(
+                                [param.view(-1) for name, param in model.named_parameters() if "mask_scores" in name]
+                            )
+                            n = concat.numel()
+                            kth = max(n - (int(n * args.final_threshold) + 1), 1)
+                            threshold_mem = concat.kthvalue(kth).values.item()
+                        inputs["threshold"] = threshold_mem
+                outputs = model(**inputs)
+                tmp_eval_loss, logits = outputs[:2]
+
+                eval_loss += tmp_eval_loss.mean().item()
+            nb_eval_steps += 1
+            if preds is None:
+                preds = logits.detach().cpu().numpy()
+                out_label_ids = inputs["labels"].detach().cpu().numpy()
+            else:
+                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+                out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
+
+        eval_loss = eval_loss / nb_eval_steps
+        if args.output_mode == "classification":
+            from scipy.special import softmax
+
+            probs = softmax(preds, axis=-1)
+            entropy = np.exp((-probs * np.log(probs)).sum(axis=-1).mean())
+            preds = np.argmax(preds, axis=1)
+        elif args.output_mode == "regression":
+            preds = np.squeeze(preds)
+        result = compute_metrics(eval_task, preds, out_label_ids)
+        results.update(result)
+        if entropy is not None:
+            result["eval_avg_entropy"] = entropy
+
+        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results {} *****".format(prefix))
+            for key in sorted(result.keys()):
+                logger.info("  %s = %s", key, str(result[key]))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return results
+
+
+def load_and_cache_examples(args, task, tokenizer, evaluate=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    processor = processors[task]()
+    output_mode = output_modes[task]
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(
+        args.data_dir,
+        "cached_{}_{}_{}_{}".format(
+            "dev" if evaluate else "train",
+            list(filter(None, args.model_name_or_path.split("/"))).pop(),
+            str(args.max_seq_length),
+            str(task),
+        ),
+    )
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", args.data_dir)
+        label_list = processor.get_labels()
+        if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta", "xlmroberta"]:
+            # 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)
+        )
+        features = convert_examples_to_features(
+            examples, tokenizer, max_length=args.max_seq_length, label_list=label_list, output_mode=output_mode,
+        )
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
+    all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
+    if output_mode == "classification":
+        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
+    elif output_mode == "regression":
+        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
+
+    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Required parameters
+    parser.add_argument(
+        "--data_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
+    )
+    parser.add_argument(
+        "--task_name",
+        default=None,
+        type=str,
+        required=True,
+        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    # Other parameters
+    parser.add_argument(
+        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        default="",
+        type=str,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        default="",
+        type=str,
+        help="Where do you want to store the pre-trained models downloaded from s3",
+    )
+    parser.add_argument(
+        "--max_seq_length",
+        default=128,
+        type=int,
+        help="The maximum total input sequence length after tokenization. Sequences longer "
+        "than this will be truncated, sequences shorter will be padded.",
+    )
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step.",
+    )
+    parser.add_argument(
+        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model.",
+    )
+
+    parser.add_argument(
+        "--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.",
+    )
+    parser.add_argument(
+        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.",
+    )
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+
+    # Pruning parameters
+    parser.add_argument(
+        "--mask_scores_learning_rate",
+        default=1e-2,
+        type=float,
+        help="The Adam initial learning rate of the mask scores.",
+    )
+    parser.add_argument(
+        "--initial_threshold", default=1.0, type=float, help="Initial value of the threshold (for scheduling)."
+    )
+    parser.add_argument(
+        "--final_threshold", default=0.7, type=float, help="Final value of the threshold (for scheduling)."
+    )
+    parser.add_argument(
+        "--initial_warmup",
+        default=1,
+        type=int,
+        help="Run `initial_warmup` * `warmup_steps` steps of threshold warmup during which threshold stays"
+        "at its `initial_threshold` value (sparsity schedule).",
+    )
+    parser.add_argument(
+        "--final_warmup",
+        default=2,
+        type=int,
+        help="Run `final_warmup` * `warmup_steps` steps of threshold cool-down during which threshold stays"
+        "at its final_threshold value (sparsity schedule).",
+    )
+
+    parser.add_argument(
+        "--pruning_method",
+        default="topK",
+        type=str,
+        help="Pruning Method (l0 = L0 regularization, magnitude = Magnitude pruning, topK = Movement pruning, sigmoied_threshold = Soft movement pruning).",
+    )
+    parser.add_argument(
+        "--mask_init",
+        default="constant",
+        type=str,
+        help="Initialization method for the mask scores. Choices: constant, uniform, kaiming.",
+    )
+    parser.add_argument(
+        "--mask_scale", default=0.0, type=float, help="Initialization parameter for the chosen initialization method."
+    )
+
+    parser.add_argument("--regularization", default=None, help="Add L0 or L1 regularization to the mask scores.")
+    parser.add_argument(
+        "--final_lambda",
+        default=0.0,
+        type=float,
+        help="Regularization intensity (used in conjunction with `regulariation`.",
+    )
+
+    parser.add_argument("--global_topk", action="store_true", help="Global TopK on the Scores.")
+    parser.add_argument(
+        "--global_topk_frequency_compute",
+        default=25,
+        type=int,
+        help="Frequency at which we compute the TopK global threshold.",
+    )
+
+    # Distillation parameters (optional)
+    parser.add_argument(
+        "--teacher_type",
+        default=None,
+        type=str,
+        help="Teacher type. Teacher tokenizer and student (model) tokenizer must output the same tokenization. Only for distillation.",
+    )
+    parser.add_argument(
+        "--teacher_name_or_path",
+        default=None,
+        type=str,
+        help="Path to the already fine-tuned teacher model. Only for distillation.",
+    )
+    parser.add_argument(
+        "--alpha_ce", default=0.5, type=float, help="Cross entropy loss linear weight. Only for distillation."
+    )
+    parser.add_argument(
+        "--alpha_distil", default=0.5, type=float, help="Distillation loss linear weight. Only for distillation."
+    )
+    parser.add_argument(
+        "--temperature", default=2.0, type=float, help="Distillation temperature. Only for distillation."
+    )
+
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument(
+        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.",
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+
+    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
+    parser.add_argument(
+        "--eval_all_checkpoints",
+        action="store_true",
+        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
+    )
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
+    parser.add_argument(
+        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory",
+    )
+    parser.add_argument(
+        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets",
+    )
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
+    parser.add_argument(
+        "--fp16_opt_level",
+        type=str,
+        default="O1",
+        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+        "See details at https://nvidia.github.io/apex/amp.html",
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    args = parser.parse_args()
+
+    # Regularization
+    if args.regularization == "null":
+        args.regularization = None
+
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        args.local_rank,
+        device,
+        args.n_gpu,
+        bool(args.local_rank != -1),
+        args.fp16,
+    )
+
+    # Set seed
+    set_seed(args)
+
+    # Prepare GLUE task
+    args.task_name = args.task_name.lower()
+    if args.task_name not in processors:
+        raise ValueError("Task not found: %s" % (args.task_name))
+    processor = processors[args.task_name]()
+    args.output_mode = output_modes[args.task_name]
+    label_list = processor.get_labels()
+    num_labels = len(label_list)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(
+        args.config_name if args.config_name else args.model_name_or_path,
+        num_labels=num_labels,
+        finetuning_task=args.task_name,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+        pruning_method=args.pruning_method,
+        mask_init=args.mask_init,
+        mask_scale=args.mask_scale,
+    )
+    tokenizer = tokenizer_class.from_pretrained(
+        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+        do_lower_case=args.do_lower_case,
+    )
+    model = model_class.from_pretrained(
+        args.model_name_or_path,
+        from_tf=bool(".ckpt" in args.model_name_or_path),
+        config=config,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+
+    if args.teacher_type is not None:
+        assert args.teacher_name_or_path is not None
+        assert args.alpha_distil > 0.0
+        assert args.alpha_distil + args.alpha_ce > 0.0
+        teacher_config_class, teacher_model_class, _ = MODEL_CLASSES[args.teacher_type]
+        teacher_config = teacher_config_class.from_pretrained(args.teacher_name_or_path)
+        teacher = teacher_model_class.from_pretrained(
+            args.teacher_name_or_path,
+            from_tf=False,
+            config=teacher_config,
+            cache_dir=args.cache_dir if args.cache_dir else None,
+        )
+        teacher.to(args.device)
+    else:
+        teacher = None
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer, teacher=teacher)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(
+                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
+            )
+            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
+
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+            result = evaluate(args, model, tokenizer, prefix=prefix)
+            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
+            results.update(result)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/movement-pruning/requirements.txt

@@ -0,0 +1,6 @@
+torch>=1.4.0
+-e git+https://github.com/huggingface/transformers.git@352d5472b0c1dec0f420d606d16747d851b4bda8#egg=transformers
+knockknock>=0.1.8.1
+h5py>=2.10.0
+numpy>=1.18.2
+scipy>=1.4.1

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/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-armanner-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)
+
+
+## Persian NER [ARMAN, PEYMA, ARMAN+PEYMA]
+
+This task aims to extract named entities in the text, such as names and label with appropriate `NER` classes such as locations, organizations, etc. The datasets used for this task contain sentences that are marked with `IOB` format. In this format, tokens that are not part of an entity are tagged as `”O”` the `”B”`tag corresponds to the first word of an object, and the `”I”` tag corresponds to the rest of the terms of the same entity. Both `”B”` and `”I”` tags are followed by a hyphen (or underscore), followed by the entity category. Therefore, the NER task is a multi-class token classification problem that labels the tokens upon being fed a raw text. There are two primary datasets used in Persian NER, `ARMAN`, and `PEYMA`. In ParsBERT, we prepared ner for both datasets as well as a combination of both datasets.
+
+
+
+### PEYMA
+
+PEYMA dataset includes 7,145 sentences with a total of 302,530 tokens from which 41,148 tokens are tagged with seven different classes.
+
+1. Organization
+2. Money
+3. Location
+4. Date
+5. Time
+6. Person
+7. Percent
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 16964 |
+|     Money    |  2037 |
+|   Location   |  8782 |
+|     Date     |  4259 |
+|     Time     |  732  |
+|    Person    |  7675 |
+|    Percent   |  699  |
+
+
+
+**Download**
+You can download the dataset from [here](http://nsurl.org/tasks/task-7-named-entity-recognition-ner-for-farsi/)
+
+---
+
+### ARMAN
+
+ARMAN dataset holds 7,682 sentences with 250,015 sentences tagged over six different classes.
+
+1. Organization
+2. Location
+3. Facility
+4. Event
+5. Product
+6. Person
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 30108 |
+|   Location   | 12924 |
+|   Facility   |  4458 |
+|     Event    |  7557 |
+|    Product   |  4389 |
+|    Person    | 15645 |
+
+
+
+**Download**
+You can download the dataset from [here](https://github.com/HaniehP/PersianNER)
+
+
+
+## Results
+
+The following table summarizes the F1 score obtained by ParsBERT as compared to other models and architectures.
+
+| Dataset         | ParsBERT | MorphoBERT |  Beheshti-NER  |  LSTM-CRF  |  Rule-Based CRF  |  BiLSTM-CRF  |
+|:---------------:|:--------:|:----------:|:--------------:|:----------:|:----------------:|:------------:|
+|  ARMAN + PEYMA  |   95.13* |      -     |        -       |      -     |         -        |       -      |
+|  PEYMA          |   98.79* |      -     |      90.59     |      -     |       84.00      |       -      |
+|  ARMAN          |   93.10* |    89.9    |      84.03     |    86.55   |         -        |     77.45    |
+
+
+## How to use :hugs:
+| Notebook     |      Description      |   |
+|:----------|:-------------|------:|
+| [How to use Pipelines](https://github.com/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb)  | Simple and efficient way to use State-of-the-Art models on downstream tasks through transformers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb) |
+
+
+## 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/)
+
++ And a special thanks to Sara Tabrizi for her fantastic poster design. Follow her on: [Linkedin](https://www.linkedin.com/in/sara-tabrizi-64548b79/), [Behance](https://www.behance.net/saratabrizi), [Instagram](https://www.instagram.com/sara_b_tabrizi/)
+
+## Releases
+
+### Release v0.1 (May 29, 2019)
+This is the first version of our ParsBERT NER!

model_cards/HooshvareLab/bert-base-parsbert-ner-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)
+
+
+## Persian NER [ARMAN, PEYMA, ARMAN+PEYMA]
+
+This task aims to extract named entities in the text, such as names and label with appropriate `NER` classes such as locations, organizations, etc. The datasets used for this task contain sentences that are marked with `IOB` format. In this format, tokens that are not part of an entity are tagged as `”O”` the `”B”`tag corresponds to the first word of an object, and the `”I”` tag corresponds to the rest of the terms of the same entity. Both `”B”` and `”I”` tags are followed by a hyphen (or underscore), followed by the entity category. Therefore, the NER task is a multi-class token classification problem that labels the tokens upon being fed a raw text. There are two primary datasets used in Persian NER, `ARMAN`, and `PEYMA`. In ParsBERT, we prepared ner for both datasets as well as a combination of both datasets.
+
+
+
+### PEYMA
+
+PEYMA dataset includes 7,145 sentences with a total of 302,530 tokens from which 41,148 tokens are tagged with seven different classes.
+
+1. Organization
+2. Money
+3. Location
+4. Date
+5. Time
+6. Person
+7. Percent
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 16964 |
+|     Money    |  2037 |
+|   Location   |  8782 |
+|     Date     |  4259 |
+|     Time     |  732  |
+|    Person    |  7675 |
+|    Percent   |  699  |
+
+
+
+**Download**
+You can download the dataset from [here](http://nsurl.org/tasks/task-7-named-entity-recognition-ner-for-farsi/)
+
+---
+
+### ARMAN
+
+ARMAN dataset holds 7,682 sentences with 250,015 sentences tagged over six different classes.
+
+1. Organization
+2. Location
+3. Facility
+4. Event
+5. Product
+6. Person
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 30108 |
+|   Location   | 12924 |
+|   Facility   |  4458 |
+|     Event    |  7557 |
+|    Product   |  4389 |
+|    Person    | 15645 |
+
+
+
+**Download**
+You can download the dataset from [here](https://github.com/HaniehP/PersianNER)
+
+
+
+## Results
+
+The following table summarizes the F1 score obtained by ParsBERT as compared to other models and architectures.
+
+| Dataset         | ParsBERT | MorphoBERT |  Beheshti-NER  |  LSTM-CRF  |  Rule-Based CRF  |  BiLSTM-CRF  |
+|:---------------:|:--------:|:----------:|:--------------:|:----------:|:----------------:|:------------:|
+|  ARMAN + PEYMA  |   95.13* |      -     |        -       |      -     |         -        |       -      |
+|  PEYMA          |   98.79* |      -     |      90.59     |      -     |       84.00      |       -      |
+|  ARMAN          |   93.10* |    89.9    |      84.03     |    86.55   |         -        |     77.45    |
+
+
+## How to use :hugs:
+| Notebook     |      Description      |   |
+|:----------|:-------------|------:|
+| [How to use Pipelines](https://github.com/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb)  | Simple and efficient way to use State-of-the-Art models on downstream tasks through transformers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb) |
+
+
+## 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/)
+
++ And a special thanks to Sara Tabrizi for her fantastic poster design. Follow her on: [Linkedin](https://www.linkedin.com/in/sara-tabrizi-64548b79/), [Behance](https://www.behance.net/saratabrizi), [Instagram](https://www.instagram.com/sara_b_tabrizi/)
+
+## Releases
+
+### Release v0.1 (May 29, 2019)
+This is the first version of our ParsBERT NER!

model_cards/HooshvareLab/bert-base-parsbert-peymaner-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)
+
+
+## Persian NER [ARMAN, PEYMA, ARMAN+PEYMA]
+
+This task aims to extract named entities in the text, such as names and label with appropriate `NER` classes such as locations, organizations, etc. The datasets used for this task contain sentences that are marked with `IOB` format. In this format, tokens that are not part of an entity are tagged as `”O”` the `”B”`tag corresponds to the first word of an object, and the `”I”` tag corresponds to the rest of the terms of the same entity. Both `”B”` and `”I”` tags are followed by a hyphen (or underscore), followed by the entity category. Therefore, the NER task is a multi-class token classification problem that labels the tokens upon being fed a raw text. There are two primary datasets used in Persian NER, `ARMAN`, and `PEYMA`. In ParsBERT, we prepared ner for both datasets as well as a combination of both datasets.
+
+
+
+### PEYMA
+
+PEYMA dataset includes 7,145 sentences with a total of 302,530 tokens from which 41,148 tokens are tagged with seven different classes.
+
+1. Organization
+2. Money
+3. Location
+4. Date
+5. Time
+6. Person
+7. Percent
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 16964 |
+|     Money    |  2037 |
+|   Location   |  8782 |
+|     Date     |  4259 |
+|     Time     |  732  |
+|    Person    |  7675 |
+|    Percent   |  699  |
+
+
+
+**Download**
+You can download the dataset from [here](http://nsurl.org/tasks/task-7-named-entity-recognition-ner-for-farsi/)
+
+---
+
+### ARMAN
+
+ARMAN dataset holds 7,682 sentences with 250,015 sentences tagged over six different classes.
+
+1. Organization
+2. Location
+3. Facility
+4. Event
+5. Product
+6. Person
+
+
+|     Label    |   #   |
+|:------------:|:-----:|
+| Organization | 30108 |
+|   Location   | 12924 |
+|   Facility   |  4458 |
+|     Event    |  7557 |
+|    Product   |  4389 |
+|    Person    | 15645 |
+
+
+
+**Download**
+You can download the dataset from [here](https://github.com/HaniehP/PersianNER)
+
+
+
+## Results
+
+The following table summarizes the F1 score obtained by ParsBERT as compared to other models and architectures.
+
+| Dataset         | ParsBERT | MorphoBERT |  Beheshti-NER  |  LSTM-CRF  |  Rule-Based CRF  |  BiLSTM-CRF  |
+|:---------------:|:--------:|:----------:|:--------------:|:----------:|:----------------:|:------------:|
+|  ARMAN + PEYMA  |   95.13* |      -     |        -       |      -     |         -        |       -      |
+|  PEYMA          |   98.79* |      -     |      90.59     |      -     |       84.00      |       -      |
+|  ARMAN          |   93.10* |    89.9    |      84.03     |    86.55   |         -        |     77.45    |
+
+
+## How to use :hugs:
+| Notebook     |      Description      |   |
+|:----------|:-------------|------:|
+| [How to use Pipelines](https://github.com/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb)  | Simple and efficient way to use State-of-the-Art models on downstream tasks through transformers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/hooshvare/parsbert-ner/blob/master/persian-ner-pipeline.ipynb) |
+
+
+## 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/)
+
++ And a special thanks to Sara Tabrizi for her fantastic poster design. Follow her on: [Linkedin](https://www.linkedin.com/in/sara-tabrizi-64548b79/), [Behance](https://www.behance.net/saratabrizi), [Instagram](https://www.instagram.com/sara_b_tabrizi/)
+
+## Releases
+
+### Release v0.1 (May 29, 2019)
+This is the first version of our ParsBERT NER!

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 | mBERT | 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 | mBERT |
+|:-----------------:|:--------:|:-----:|
+| Digikala Magazine |   93.59* | 90.72 |
+|    Persian News   |   97.19* | 95.79 |
+
+
+### Named Entity Recognition (NER) task
+
+| Dataset | ParsBERT |  mBERT   | MorphoBERT |  Beheshti-NER  |  LSTM-CRF  |  Rule-Based CRF  |  BiLSTM-CRF  |
+|:-------:|:--------:|:--------:|:----------:|:--------------:|:----------:|:----------------:|:------------:|
+|  PEYMA  |   93.10* |   86.64  |      -     |      90.59     |      -     |       84.00      |       -      |
+|  ARMAN  |   98.79* |   95.89  |    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، همه 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', '،', 'همه', '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

@@ -1,15 +1,24 @@
+---
+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
 
-Example usage:
+### 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')
+input_ids = tokenizer.encode("Но зла Юнона, суча дочка,", add_special_tokens=False, return_tensors='pt')
 
 outputs = model.generate(
     input_ids,
@@ -19,5 +28,12 @@ outputs = model.generate(
 )
 
 for i, out in enumerate(outputs):
-    print('{}: {}'.format(i, tokenizer.decode(out)))
+    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/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/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/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/longformer-base-4096-finetuned-squadv2/README.md

@@ -0,0 +1,65 @@
+---
+language: english
+thumbnail:
+---
+
+# Longformer-base-4096 fine-tuned on SQuAD v2
+
+[Longformer-base-4096 model](https://huggingface.co/allenai/longformer-base-4096) fine-tuned on [SQuAD v2](https://rajpurkar.github.io/SQuAD-explorer/) for **Q&A** downstream task.
+
+## 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.
+
+## 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/drive/1zEl5D-DdkBKva-DdreVOmN0hrAfzKG1o?usp=sharing)
+
+
+
+## Model in Action 🚀
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForQuestionAnswering
+
+tokenizer = AutoTokenizer.from_pretrained("mrm8488/longformer-base-4096-finetuned-squadv2")
+model = AutoModelForQuestionAnswering.from_pretrained("mrm8488/longformer-base-4096-finetuned-squadv2")
+
+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
+```
+If given the same context we ask something that is not there, the output for **no answer** will be ```<s>```
+
+> 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/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/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

@@ -25,4 +25,11 @@ Pull Request so it can be included under the Community notebooks.
 |:----------|:-------------|:-------------|------:|
 | [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)
+| [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) |
+| [Pretrain Longformer](https://github.com/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb)  | How to build a "long" version of existing pretrained models |  [Iz Beltagy](https://beltagy.net) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) |
+| [Fine-tune Longformer for QA](https://github.com/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) | How to fine-tune longformer model for QA task | [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/Notebooks/blob/master/longformer_qa_training.ipynb) |
+| [Evaluate Model with 🤗nlp](https://github.com/patrickvonplaten/notebooks/blob/master/How_to_evaluate_Longformer_on_TriviaQA_using_NLP.ipynb) | How to evaluate longformer on TriviaQA with `nlp` | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing) |

setup.py

@@ -71,13 +71,13 @@ extras["sklearn"] = ["scikit-learn"]
 # keras2onnx and onnxconverter-common version is specific through a commit until 1.7.0 lands on pypi
 extras["tf"] = [
     "tensorflow",
-    "onnxconverter-common",
-    "keras2onnx"
+    "onnxconverter-common @ git+git://github.com/microsoft/onnxconverter-common.git@f64ca15989b6dc95a1f3507ff6e4c395ba12dff5#egg=onnxconverter-common",
+    "keras2onnx @ git+git://github.com/onnx/keras-onnx.git@cbdc75cb950b16db7f0a67be96a278f8d2953b48#egg=keras2onnx"
 ]
 extras["tf-cpu"] = [
     "tensorflow-cpu",
-    "onnxconverter-common",
-    "keras2onnx"
+    "onnxconverter-common @ git+git://github.com/microsoft/onnxconverter-common.git@f64ca15989b6dc95a1f3507ff6e4c395ba12dff5#egg=onnxconverter-common",
+    "keras2onnx @ git+git://github.com/onnx/keras-onnx.git@cbdc75cb950b16db7f0a67be96a278f8d2953b48#egg=keras2onnx"
 ]
 extras["torch"] = ["torch"]
 
@@ -88,7 +88,7 @@ extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator"]
 extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme"]
 extras["quality"] = [
     "black",
-    "isort",
+    "isort @ git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort",
     "flake8",
 ]
 extras["dev"] = extras["testing"] + extras["quality"] + ["mecab-python3", "scikit-learn", "tensorflow", "torch"]
@@ -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

@@ -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
@@ -139,7 +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
+from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
 from .tokenization_reformer import ReformerTokenizer
 from .tokenization_roberta import RobertaTokenizer, RobertaTokenizerFast
@@ -335,7 +322,15 @@ if is_torch_available():
         REFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
-    from .modeling_longformer import LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP, LongformerModel, LongformerForMaskedLM
+    from .modeling_longformer import (
+        LongformerModel,
+        LongformerForMaskedLM,
+        LongformerForSequenceClassification,
+        LongformerForMultipleChoice,
+        LongformerForTokenClassification,
+        LongformerForQuestionAnswering,
+        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
+    )
 
     # Optimization
     from .optimization import (
@@ -352,6 +347,9 @@ 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_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

@@ -23,8 +23,11 @@ from .configuration_roberta import RobertaConfig
 logger = logging.getLogger(__name__)
 
 LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "longformer-base-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096/config.json",
-    "longformer-large-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096/config.json",
+    "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",
 }
 
 
@@ -64,6 +67,7 @@ class LongformerConfig(RobertaConfig):
     pretrained_config_archive_map = LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "longformer"
 
-    def __init__(self, attention_window: Union[List[int], int] = 512, **kwargs):
+    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/configuration_utils.py

@@ -89,6 +89,7 @@ class PretrainedConfig(object):
         self.id2label = kwargs.pop("id2label", None)
         self.label2id = kwargs.pop("label2id", None)
         if self.id2label is not None:
+            kwargs.pop("num_labels", None)
             self.id2label = dict((int(key), value) for key, value in self.id2label.items())
             # Keys are always strings in JSON so convert ids to int here.
         else:

src/transformers/convert_graph_to_onnx.py

@@ -1,5 +1,4 @@
 from argparse import ArgumentParser
-from itertools import takewhile
 from os import listdir, makedirs
 from os.path import abspath, dirname, exists
 from typing import Dict, List, Optional, Tuple
@@ -38,14 +37,17 @@ def ensure_valid_input(model, tokens, input_names):
 
     """
     model_args_name = model.forward.__code__.co_varnames
-    model_args_pos = [(model_args_name.index(name) - 1, name) for name in input_names]
-    model_args = [None] * (max(map(lambda x: x[0], model_args_pos)) + 1)
 
-    for arg_pos, arg_name in model_args_pos:
-        model_args[arg_pos] = tokens[arg_name]
+    ordered_input_names = []
+    model_args = []
+    for arg_name in model_args_name[1:]:  # start at index 1 to skip "self" argument
+        if arg_name in input_names:
+            ordered_input_names.append(arg_name)
+            model_args.append(tokens[arg_name])
+        else:
+            break
 
-    model_args = tuple(model_args)  # Need to be ordered
-    return tuple(takewhile(lambda arg: arg is not None, model_args))
+    return ordered_input_names, tuple(model_args)
 
 
 def infer_shapes(nlp: Pipeline, framework: str) -> Tuple[List[str], List[str], Dict, BatchEncoding]:
@@ -117,13 +119,13 @@ def convert_pytorch(nlp: Pipeline, opset: int, output: str, use_external_format:
 
     with torch.no_grad():
         input_names, output_names, dynamic_axes, tokens = infer_shapes(nlp, "pt")
-        model_args = ensure_valid_input(nlp.model, tokens, input_names)
+        ordered_input_names, model_args = ensure_valid_input(nlp.model, tokens, input_names)
 
         export(
             nlp.model,
             model_args,
             f=output,
-            input_names=input_names,
+            input_names=ordered_input_names,
             output_names=output_names,
             dynamic_axes=dynamic_axes,
             do_constant_folding=True,

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

@@ -86,6 +86,15 @@ class GlueDataset(Dataset):
                 mode.value, tokenizer.__class__.__name__, str(args.max_seq_length), args.task_name,
             ),
         )
+        label_list = self.processor.get_labels()
+        if args.task_name in ["mnli", "mnli-mm"] and tokenizer.__class__ in (
+            RobertaTokenizer,
+            RobertaTokenizerFast,
+            XLMRobertaTokenizer,
+        ):
+            # HACK(label indices are swapped in RoBERTa pretrained model)
+            label_list[1], label_list[2] = label_list[2], label_list[1]
+        self.label_list = label_list
 
         # Make sure only the first process in distributed training processes the dataset,
         # and the others will use the cache.
@@ -100,14 +109,7 @@ class GlueDataset(Dataset):
                 )
             else:
                 logger.info(f"Creating features from dataset file at {args.data_dir}")
-                label_list = self.processor.get_labels()
-                if args.task_name in ["mnli", "mnli-mm"] and tokenizer.__class__ in (
-                    RobertaTokenizer,
-                    RobertaTokenizerFast,
-                    XLMRobertaTokenizer,
-                ):
-                    # HACK(label indices are swapped in RoBERTa pretrained model)
-                    label_list[1], label_list[2] = label_list[2], label_list[1]
+
                 if mode == Split.dev:
                     examples = self.processor.get_dev_examples(args.data_dir)
                 elif mode == Split.test:
@@ -137,4 +139,4 @@ class GlueDataset(Dataset):
         return self.features[i]
 
     def get_labels(self):
-        return self.processor.get_labels()
+        return self.label_list

src/transformers/data/processors/glue.py

@@ -332,11 +332,12 @@ class Sst2Processor(DataProcessor):
     def _create_examples(self, lines, set_type):
         """Creates examples for the training, dev and test sets."""
         examples = []
+        text_index = 1 if set_type == "test" else 0
         for (i, line) in enumerate(lines):
             if i == 0:
                 continue
             guid = "%s-%s" % (set_type, i)
-            text_a = line[0]
+            text_a = line[text_index]
             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

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_albert.py

@@ -425,7 +425,7 @@ ALBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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`):
+        inputs_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.

src/transformers/modeling_auto.py

@@ -101,7 +101,15 @@ 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, LongformerModel
+from .modeling_longformer import (
+    LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
+    LongformerForMaskedLM,
+    LongformerForMultipleChoice,
+    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
@@ -247,6 +255,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForSequenceClassification),
         (XLMRobertaConfig, XLMRobertaForSequenceClassification),
         (BartConfig, BartForSequenceClassification),
+        (LongformerConfig, LongformerForSequenceClassification),
         (RobertaConfig, RobertaForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
@@ -260,6 +269,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
     [
         (DistilBertConfig, DistilBertForQuestionAnswering),
         (AlbertConfig, AlbertForQuestionAnswering),
+        (LongformerConfig, LongformerForQuestionAnswering),
         (RobertaConfig, RobertaForQuestionAnswering),
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
@@ -274,6 +284,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForTokenClassification),
         (XLMConfig, XLMForTokenClassification),
         (XLMRobertaConfig, XLMRobertaForTokenClassification),
+        (LongformerConfig, LongformerForTokenClassification),
         (RobertaConfig, RobertaForTokenClassification),
         (BertConfig, BertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
@@ -287,6 +298,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
     [
         (CamembertConfig, CamembertForMultipleChoice),
         (XLMRobertaConfig, XLMRobertaForMultipleChoice),
+        (LongformerConfig, LongformerForMultipleChoice),
         (RobertaConfig, RobertaForMultipleChoice),
         (BertConfig, BertForMultipleChoice),
         (XLNetConfig, XLNetForMultipleChoice),
@@ -316,6 +328,11 @@ class AutoModel:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -398,7 +415,7 @@ 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.
 
@@ -417,7 +434,7 @@ 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.
@@ -468,6 +485,11 @@ class AutoModelForPreTraining:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -548,7 +570,7 @@ 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:
@@ -562,7 +584,7 @@ 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.
 
@@ -613,6 +635,11 @@ class AutoModelWithLMHead:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -694,7 +721,7 @@ 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:
@@ -708,7 +735,7 @@ 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.
 
@@ -759,6 +786,11 @@ class AutoModelForSequenceClassification:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -833,7 +865,7 @@ 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.
 
@@ -852,7 +884,7 @@ 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.
@@ -906,6 +938,11 @@ class AutoModelForQuestionAnswering:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -975,7 +1012,7 @@ 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.
 
@@ -991,7 +1028,7 @@ 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.
@@ -1046,6 +1083,11 @@ class AutoModelForTokenClassification:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -1120,7 +1162,7 @@ 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.
 
@@ -1136,7 +1178,7 @@ 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_bart.py

@@ -904,7 +904,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
         **unused
     ):
         r"""
-        masked_lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for computing the masked language modeling loss.
             Indices should either be in ``[0, ..., config.vocab_size]`` or -100 (see ``input_ids`` docstring).
             Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens
@@ -913,7 +913,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
 
     Returns:
         :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        masked_lm_loss (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+        masked_lm_loss (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
             Masked language modeling loss.
         prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).

src/transformers/modeling_bert.py

@@ -543,7 +543,7 @@ BERT_START_DOCSTRING = r"""
 
 BERT_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.BertTokenizer`.
@@ -551,19 +551,19 @@ BERT_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             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
 
             `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`):
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -632,7 +632,7 @@ class BertModel(BertPreTrainedModel):
         for layer, heads in heads_to_prune.items():
             self.encoder.layer[layer].attention.prune_heads(heads)
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -759,7 +759,7 @@ class BertForPreTraining(BertPreTrainedModel):
     def get_output_embeddings(self):
         return self.cls.predictions.decoder
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -859,7 +859,7 @@ class BertForMaskedLM(BertPreTrainedModel):
     def get_output_embeddings(self):
         return self.cls.predictions.decoder
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -992,7 +992,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -1036,11 +1036,12 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForNextSentencePrediction.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
-        outputs = model(input_ids)
-
-        seq_relationship_scores = outputs[0]
+        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        next_sentence = "The sky is blue due to the shorter wavelength of blue light."
+        encoding = tokenizer.encode_plus(prompt, next_sentence, return_tensors='pt')
 
+        loss, logits = model(**encoding, next_sentence_label=torch.LongTensor([1]))
+        assert logits[0, 0] < logits[0, 1] # next sentence was random
         """
 
         outputs = self.bert(
@@ -1081,7 +1082,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -1177,7 +1178,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -1191,7 +1192,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
             Labels for computing the multiple choice classification loss.
-            Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
+            Indices should be in ``[0, ..., num_choices-1]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
 
     Returns:
@@ -1221,14 +1222,17 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForMultipleChoice.from_pretrained('bert-base-uncased')
-        choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
 
-        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
-        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
-        outputs = model(input_ids, labels=labels)
+        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        choice0 = "It is eaten with a fork and a knife."
+        choice1 = "It is eaten while held in the hand."
+        labels = torch.tensor(0) # choice0 is correct (according to Wikipedia ;))
 
-        loss, classification_scores = outputs[:2]
+        encoding = tokenizer.batch_encode_plus([[prompt, choice0], [prompt, choice1]], return_tensors='pt', pad_to_max_length=True)
+        outputs = model(**{k: v.unsqueeze(0) for k,v in encoding.items()}, labels=labels) # batch size is 1
 
+        # the linear classifier still needs to be trained
+        loss, logits = outputs[:2]
         """
         num_choices = input_ids.shape[1]
 
@@ -1278,7 +1282,7 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -1375,7 +1379,7 @@ class BertForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,

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`):
@@ -245,11 +244,10 @@ CTRL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
             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`).
+            If `past` is used, optionally only the last `inputs_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`.
@@ -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

@@ -132,7 +132,7 @@ class ElectraDiscriminatorPredictions(nn.Module):
     def forward(self, discriminator_hidden_states, attention_mask):
         hidden_states = self.dense(discriminator_hidden_states)
         hidden_states = get_activation(self.config.hidden_act)(hidden_states)
-        logits = self.dense_prediction(hidden_states).squeeze()
+        logits = self.dense_prediction(hidden_states).squeeze(-1)
 
         return logits
 

src/transformers/modeling_encoder_decoder.py

@@ -35,6 +35,7 @@ class EncoderDecoderModel(PreTrainedModel):
         class method for the encoder and `AutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` class method for the decoder.
     """
     config_class = EncoderDecoderConfig
+    base_model_prefix = "encoder_decoder"
 
     def __init__(
         self,
@@ -125,7 +126,7 @@ 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
         """
@@ -158,12 +159,26 @@ class EncoderDecoderModel(PreTrainedModel):
             ), "If `decoder_model` is not defined as an argument, a `decoder_pretrained_model_name_or_path` has to be defined"
             from .modeling_auto import AutoModelWithLMHead
 
+            if "config" not in kwargs_decoder:
+                from transformers import AutoConfig
+
+                decoder_config = AutoConfig.from_pretrained(decoder_pretrained_model_name_or_path)
+                if decoder_config.is_decoder is False:
+                    logger.info(
+                        f"Initializing {decoder_pretrained_model_name_or_path} as a decoder model. Cross attention layers are added to {decoder_pretrained_model_name_or_path} and randomly initialized if {decoder_pretrained_model_name_or_path}'s architecture allows for cross attention layers."
+                    )
+                    decoder_config.is_decoder = True
+
+                kwargs_decoder["config"] = decoder_config
+
+            if kwargs_decoder["config"].is_decoder is False:
+                logger.warning(
+                    f"Decoder model {decoder_pretrained_model_name_or_path} is not initialized as a decoder. In order to initialize {decoder_pretrained_model_name_or_path} as a decoder, make sure that the attribute `is_decoder` of `decoder_config` passed to `.from_encoder_decoder_pretrained(...)` is set to `True` or do not pass a `decoder_config` to `.from_encoder_decoder_pretrained(...)`"
+                )
+
             decoder = AutoModelWithLMHead.from_pretrained(decoder_pretrained_model_name_or_path, **kwargs_decoder)
-        decoder.config.is_decoder = True
 
-        model = cls(encoder=encoder, decoder=decoder)
-
-        return model
+        return cls(encoder=encoder, decoder=decoder)
 
     def forward(
         self,

src/transformers/modeling_flaubert.py

@@ -95,7 +95,7 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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`):
+        inputs_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.

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.
@@ -323,11 +323,10 @@ GPT2_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
             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`).
+            If `past` is used, optionally only the last `inputs_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`.
 """
@@ -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:
@@ -565,7 +554,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-100, 0, ..., config.vocab_size]``
             All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``

src/transformers/modeling_longformer.py

@@ -19,7 +19,7 @@ import math
 
 import torch
 import torch.nn as nn
-from torch.nn import CrossEntropyLoss
+from torch.nn import CrossEntropyLoss, MSELoss
 from torch.nn import functional as F
 
 from .configuration_longformer import LongformerConfig
@@ -31,11 +31,52 @@ from .modeling_roberta import RobertaLMHead, RobertaModel
 logger = logging.getLogger(__name__)
 
 LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP = {
-    "longformer-base-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096/pytorch_model.bin",
-    "longformer-large-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096/pytorch_model.bin",
+    "allenai/longformer-base-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096/pytorch_model.bin",
+    "allenai/longformer-large-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096/pytorch_model.bin",
+    "allenai/longformer-large-4096-finetuned-triviaqa": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-finetuned-triviaqa/pytorch_model.bin",
+    "allenai/longformer-base-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096-extra.pos.embd.only/pytorch_model.bin",
+    "allenai/longformer-large-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-extra.pos.embd.only/pytorch_model.bin",
 }
 
 
+def _get_question_end_index(input_ids, sep_token_id):
+    """
+        Computes the index of the first occurance of `sep_token_id`.
+    """
+
+    sep_token_indices = (input_ids == sep_token_id).nonzero()
+    batch_size = input_ids.shape[0]
+
+    assert sep_token_indices.shape[1] == 2, "`input_ids` should have two dimensions"
+    assert (
+        sep_token_indices.shape[0] == 3 * batch_size
+    ), f"There should be exactly three separator tokens: {sep_token_id} in every sample for questions answering. You might also consider to set `global_attention_mask` manually in the forward function to avoid this error."
+
+    return sep_token_indices.view(batch_size, 3, 2)[:, 0, 1]
+
+
+def _compute_global_attention_mask(input_ids, sep_token_id, before_sep_token=True):
+    """
+        Computes global attention mask by putting attention on all tokens
+        before `sep_token_id` if `before_sep_token is True` else after
+        `sep_token_id`.
+    """
+
+    question_end_index = _get_question_end_index(input_ids, sep_token_id)
+    question_end_index = question_end_index.unsqueeze(dim=1)  # size: batch_size x 1
+    # bool attention mask with True in locations of global attention
+    attention_mask = torch.arange(input_ids.shape[1], device=input_ids.device)
+    if before_sep_token is True:
+        attention_mask = (attention_mask.expand_as(input_ids) < question_end_index).to(torch.uint8)
+    else:
+        # last token is separation token and should not be counted and in the middle are two separation tokens
+        attention_mask = (attention_mask.expand_as(input_ids) > (question_end_index + 1)).to(torch.uint8) * (
+            attention_mask.expand_as(input_ids) < input_ids.shape[-1]
+        ).to(torch.uint8)
+
+    return attention_mask
+
+
 class LongformerSelfAttention(nn.Module):
     def __init__(self, config, layer_id):
         super().__init__()
@@ -307,9 +348,7 @@ class LongformerSelfAttention(nn.Module):
             selected_v[selection_padding_mask_nonzeros] = v[extra_attention_mask_nonzeros]
             # use `matmul` because `einsum` crashes sometimes with fp16
             # attn = torch.einsum('blhs,bshd->blhd', (selected_attn_probs, selected_v))
-            attn = torch.matmul(
-                selected_attn_probs.transpose(1, 2), selected_v.transpose(1, 2).type_as(selected_attn_probs)
-            ).transpose(1, 2)
+            attn = torch.matmul(selected_attn_probs.transpose(1, 2), selected_v.transpose(1, 2)).transpose(1, 2)
             attn_probs = attn_probs.narrow(
                 -1, max_num_extra_indices_per_batch, attn_probs.size(-1) - max_num_extra_indices_per_batch
             ).contiguous()
@@ -373,7 +412,7 @@ class LongformerSelfAttention(nn.Module):
             ]
             attn[extra_attention_mask_nonzeros[::-1]] = nonzero_selected_attn.view(
                 len(selection_padding_mask_nonzeros[0]), -1
-            ).type_as(hidden_states)
+            )
 
         context_layer = attn.transpose(0, 1)
         if self.output_attentions:
@@ -408,7 +447,7 @@ LONGFORMER_START_DOCSTRING = r"""
 
 LONGFORMER_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.LonmgformerTokenizer`.
@@ -416,25 +455,30 @@ LONGFORMER_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
-            Mask to decide the attention given on each token, local attention, global attenion, or no attention (for padding tokens).
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+
+        global_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            Mask to decide the attention given on each token, local attention or global attenion.
             Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for
             task-specific finetuning because it makes the model more flexible at representing the task. For example,
             for classification, the <s> token should be given global attention. For QA, all question tokens should also have
             global attention. Please refer to the Longformer paper https://arxiv.org/abs/2004.05150 for more details.
-            Mask values selected in ``[0, 1, 2]``:
-            ``0`` for no attention (padding tokens),
-            ``1`` for local attention (a sliding window attention),
-            ``2`` for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
+            Mask values selected in ``[0, 1]``:
+            ``0`` for local attention (a sliding window attention),
+            ``1`` for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
 
-            `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             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
 
             `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`):
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -534,11 +578,12 @@ class LongformerModel(RobertaModel):
 
         return padding_len, input_ids, attention_mask, token_type_ids, position_ids, inputs_embeds
 
-    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
         attention_mask=None,
+        global_attention_mask=None,
         token_type_ids=None,
         position_ids=None,
         inputs_embeds=None,
@@ -569,8 +614,8 @@ class LongformerModel(RobertaModel):
         import torch
         from transformers import LongformerModel, LongformerTokenizer
 
-        model = LongformerModel.from_pretrained('longformer-base-4096')
-        tokenizer = LongformerTokenizer.from_pretrained('longformer-base-4096')
+        model = LongformerModel.from_pretrained('allenai/longformer-base-4096')
+        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
 
         SAMPLE_TEXT = ' '.join(['Hello world! '] * 1000)  # long input document
         input_ids = torch.tensor(tokenizer.encode(SAMPLE_TEXT)).unsqueeze(0)  # batch of size 1
@@ -590,6 +635,19 @@ class LongformerModel(RobertaModel):
             if isinstance(self.config.attention_window, int)
             else max(self.config.attention_window)
         )
+
+        # merge `global_attention_mask` and `attention_mask`
+        if global_attention_mask is not None:
+            # longformer self attention expects attention mask to have 0 (no attn), 1 (local attn), 2 (global attn)
+            # (global_attention_mask + 1) => 1 for local attention, 2 for global attention
+            # => final attention_mask => 0 for no attention, 1 for local attention 2 for global attention
+            if attention_mask is not None:
+                attention_mask = attention_mask * (global_attention_mask + 1)
+            else:
+                # simply use `global_attention_mask` as `attention_mask`
+                # if no `attention_mask` is given
+                attention_mask = global_attention_mask + 1
+
         padding_len, input_ids, attention_mask, token_type_ids, position_ids, inputs_embeds = self._pad_to_window_size(
             input_ids=input_ids,
             attention_mask=attention_mask,
@@ -638,11 +696,12 @@ class LongformerForMaskedLM(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
         attention_mask=None,
+        global_attention_mask=None,
         token_type_ids=None,
         position_ids=None,
         inputs_embeds=None,
@@ -678,8 +737,8 @@ class LongformerForMaskedLM(BertPreTrainedModel):
         import torch
         from transformers import LongformerForMaskedLM, LongformerTokenizer
 
-        model = LongformerForMaskedLM.from_pretrained('longformer-base-4096')
-        tokenizer = LongformerTokenizer.from_pretrained('longformer-base-4096')
+        model = LongformerForMaskedLM.from_pretrained('allenai/longformer-base-4096')
+        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
 
         SAMPLE_TEXT = ' '.join(['Hello world! '] * 1000)  # long input document
         input_ids = torch.tensor(tokenizer.encode(SAMPLE_TEXT)).unsqueeze(0)  # batch of size 1
@@ -692,6 +751,7 @@ class LongformerForMaskedLM(BertPreTrainedModel):
         outputs = self.longformer(
             input_ids,
             attention_mask=attention_mask,
+            global_attention_mask=global_attention_mask,
             token_type_ids=token_type_ids,
             position_ids=position_ids,
             inputs_embeds=inputs_embeds,
@@ -707,3 +767,466 @@ class LongformerForMaskedLM(BertPreTrainedModel):
             outputs = (masked_lm_loss,) + outputs
 
         return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Longformer Model transformer with a sequence classification/regression head on top (a linear layer
+    on top of the pooled output) e.g. for GLUE tasks. """,
+    LONGFORMER_START_DOCSTRING,
+)
+class LongformerForSequenceClassification(BertPreTrainedModel):
+    config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "longformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.longformer = LongformerModel(config)
+        self.classifier = LongformerClassificationHead(config)
+
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        global_attention_mask=None,
+        token_type_ids=None,
+        position_ids=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.LongformerConfig`) 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 LongformerTokenizer, LongformerForSequenceClassification
+        import torch
+
+        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
+        model = LongformerForSequenceClassification.from_pretrained('allenai/longformer-base-4096')
+        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]
+
+        """
+
+        if global_attention_mask is None:
+            logger.info("Initializing global attention on CLS token...")
+            global_attention_mask = torch.zeros_like(input_ids)
+            # global attention on cls token
+            global_attention_mask[:, 0] = 1
+
+        outputs = self.longformer(
+            input_ids,
+            attention_mask=attention_mask,
+            global_attention_mask=global_attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+        )
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]
+        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)
+
+
+class LongformerClassificationHead(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, hidden_states, **kwargs):
+        hidden_states = hidden_states[:, 0, :]  # take <s> token (equiv. to [CLS])
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.dense(hidden_states)
+        hidden_states = torch.tanh(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        output = self.out_proj(hidden_states)
+        return output
+
+
+@add_start_docstrings(
+    """Longformer Model with a span classification head on top for extractive question-answering tasks like SQuAD / TriviaQA (a linear layers on top of
+    the hidden-states output to compute `span start logits` and `span end logits`). """,
+    LONGFORMER_START_DOCSTRING,
+)
+class LongformerForQuestionAnswering(BertPreTrainedModel):
+    config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "longformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.longformer = LongformerModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids,
+        attention_mask=None,
+        global_attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.LongformerConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end 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 LongformerTokenizer, LongformerForQuestionAnswering
+        import torch
+
+        tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa")
+        model = LongformerForQuestionAnswering.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa")
+
+        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        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)) # remove space prepending space token
+
+        """
+
+        # set global attention on question tokens
+        if global_attention_mask is None:
+            logger.info("Initializing global attention on question tokens...")
+            # put global attention on all tokens until `config.sep_token_id` is reached
+            global_attention_mask = _compute_global_attention_mask(input_ids, self.config.sep_token_id)
+
+        outputs = self.longformer(
+            input_ids,
+            attention_mask=attention_mask,
+            global_attention_mask=global_attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Longformer Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    LONGFORMER_START_DOCSTRING,
+)
+class LongformerForTokenClassification(BertPreTrainedModel):
+    config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "longformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.longformer = LongformerModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        global_attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        labels=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.LongformerConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        from transformers import LongformerTokenizer, LongformerForTokenClassification
+        import torch
+
+        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
+        model = LongformerForTokenClassification.from_pretrained('allenai/longformer-base-4096')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
+
+        """
+
+        outputs = self.longformer(
+            input_ids,
+            attention_mask=attention_mask,
+            global_attention_mask=global_attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Longformer Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    LONGFORMER_START_DOCSTRING,
+)
+class LongformerForMultipleChoice(BertPreTrainedModel):
+    config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "longformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.longformer = LongformerModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        token_type_ids=None,
+        attention_mask=None,
+        global_attention_mask=None,
+        labels=None,
+        position_ids=None,
+        inputs_embeds=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the multiple choice classification loss.
+            Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above)
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor`` of shape ``(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        classification_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        from transformers import LongformerTokenizer, LongformerForMultipleChoice
+        import torch
+
+        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
+        model = LongformerForMultipleChoice.from_pretrained('allenai/longformer-base-4096')
+        # context = "The dog is cute" | choice = "the dog" / "the cat"
+        choices = [("The dog is cute", "the dog"), ("The dog is cute", "the cat")]
+        input_ids = torch.tensor([tokenizer.encode(s[0], s[1], add_special_tokens=True) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
+
+        # global attention is automatically put on "the dog" and "the cat"
+        outputs = model(input_ids, labels=labels)
+        loss, classification_scores = outputs[:2]
+
+        """
+        num_choices = input_ids.shape[1]
+
+        # set global attention on question tokens
+        if global_attention_mask is None:
+            logger.info("Initializing global attention on multiple choice...")
+            # put global attention on all tokens after `config.sep_token_id`
+            global_attention_mask = torch.stack(
+                [
+                    _compute_global_attention_mask(input_ids[:, i], self.config.sep_token_id, before_sep_token=False)
+                    for i in range(num_choices)
+                ],
+                dim=1,
+            )
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1))
+        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        flat_global_attention_mask = (
+            global_attention_mask.view(-1, global_attention_mask.size(-1))
+            if global_attention_mask is not None
+            else None
+        )
+
+        outputs = self.longformer(
+            flat_input_ids,
+            position_ids=flat_position_ids,
+            token_type_ids=flat_token_type_ids,
+            attention_mask=flat_attention_mask,
+            global_attention_mask=flat_global_attention_mask,
+        )
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)

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_openai.py

@@ -313,7 +313,7 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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`):
+        inputs_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.
@@ -491,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-100, 0, ..., config.vocab_size]``
             All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_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(
@@ -1653,7 +1662,7 @@ class ReformerModel(ReformerPreTrainedModel):
                 padded_position_ids = position_ids.unsqueeze(0).expand(input_shape[0], padding_length)
                 position_ids = torch.cat([position_ids, padded_position_ids], dim=-1)
 
-        # Extend `input_embeds` with padding to match least common multiple chunk_length
+        # Extend `inputs_embeds` with padding to match least common multiple chunk_length
         if inputs_embeds is not None:
             padded_inputs_embeds = self.embeddings(padded_input_ids, position_ids)
             inputs_embeds = torch.cat([inputs_embeds, padded_inputs_embeds], dim=-2)

src/transformers/modeling_roberta.py

@@ -95,7 +95,7 @@ ROBERTA_START_DOCSTRING = r"""
 
 ROBERTA_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.RobertaTokenizer`.
@@ -103,19 +103,19 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             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
 
             `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`):
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -175,7 +175,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head.decoder
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -286,7 +286,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
         self.roberta = RobertaModel(config)
         self.classifier = RobertaClassificationHead(config)
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -379,7 +379,7 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -479,7 +479,7 @@ class RobertaForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids=None,
@@ -598,7 +598,7 @@ class RobertaForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def forward(
         self,
         input_ids,

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_albert.py

@@ -628,7 +628,7 @@ ALBERT_START_DOCSTRING = r"""
 
 ALBERT_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:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.AlbertTokenizer`.
@@ -636,19 +636,19 @@ ALBERT_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional, defaults to :obj:`None`):
+        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             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
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
-        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -657,7 +657,7 @@ ALBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` 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.
@@ -676,7 +676,7 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
         super().__init__(config, *inputs, **kwargs)
         self.albert = TFAlbertMainLayer(config, name="albert")
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
         Returns:
@@ -734,7 +734,7 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel):
     def get_output_embeddings(self):
         return self.albert.embeddings
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -795,7 +795,7 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel):
     def get_output_embeddings(self):
         return self.albert.embeddings
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Returns:
@@ -852,7 +852,7 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Returns:
@@ -908,7 +908,7 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
         )
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -983,7 +983,7 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel):
         """
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
-    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
     def call(
         self,
         inputs,

src/transformers/modeling_tf_auto.py

@@ -238,6 +238,12 @@ class TFAutoModel(object):
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `distilbert` configuration class: TFDistilBertModel (DistilBERT model)
@@ -378,6 +384,11 @@ class TFAutoModelForPreTraining(object):
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
         Args:
             config (:class:`~transformers.PretrainedConfig`):
                 The model class to instantiate is selected based on the configuration class:
@@ -544,6 +555,12 @@ class TFAutoModelWithLMHead(object):
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
@@ -699,6 +716,12 @@ class TFAutoModelForMultipleChoice:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `albert` configuration class: AlbertModel (Albert model)
@@ -849,6 +872,12 @@ class TFAutoModelForSequenceClassification(object):
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
@@ -1006,6 +1035,12 @@ class TFAutoModelForQuestionAnswering(object):
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
@@ -1143,6 +1178,12 @@ class TFAutoModelForTokenClassification:
         r""" Instantiates one of the base model classes of the library
         from a configuration.
 
+        Note:
+            Loading a model from its configuration file does **not** load the model weights.
+            It only affects the model's configuration. Use :func:`~transformers.AutoModel.from_pretrained` to load
+            the model weights
+
+        Args:
             config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
                 The model class to instantiate is selected based on the configuration class:
                     - isInstance of `bert` configuration class: BertModel (Bert model)

src/transformers/modeling_tf_bert.py

@@ -621,7 +621,7 @@ BERT_START_DOCSTRING = r"""
 
 BERT_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:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.BertTokenizer`.
@@ -629,19 +629,19 @@ BERT_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             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
 
             `What are token type IDs? <../glossary.html#token-type-ids>`__
-        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -669,7 +669,7 @@ class TFBertModel(TFBertPreTrainedModel):
         super().__init__(config, *inputs, **kwargs)
         self.bert = TFBertMainLayer(config, name="bert")
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Returns:
@@ -726,7 +726,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
     def get_output_embeddings(self):
         return self.bert.embeddings
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -782,7 +782,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
     def get_output_embeddings(self):
         return self.bert.embeddings
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -832,7 +832,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
         self.bert = TFBertMainLayer(config, name="bert")
         self.nsp = TFBertNSPHead(config, name="nsp___cls")
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -857,10 +857,13 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForNextSentencePrediction.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
-        outputs = model(input_ids)
-        seq_relationship_scores = outputs[0]
 
+        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        next_sentence = "The sky is blue due to the shorter wavelength of blue light."
+        encoding = tokenizer.encode_plus(prompt, next_sentence, return_tensors='tf')
+
+        logits = model(encoding['input_ids'], token_type_ids=encoding['token_type_ids'])[0]
+        assert logits[0][0] < logits[0][1] # the next sentence was random
         """
         outputs = self.bert(inputs, **kwargs)
 
@@ -888,7 +891,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -954,7 +957,7 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel):
         """
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
     def call(
         self,
         inputs,
@@ -990,11 +993,15 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForMultipleChoice.from_pretrained('bert-base-uncased')
-        choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
-        input_ids = tf.constant([tokenizer.encode(s) for s in choices])[None, :]  # Batch size 1, 2 choices
-        outputs = model(input_ids)
-        classification_scores = outputs[0]
 
+        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        choice0 = "It is eaten with a fork and a knife."
+        choice1 = "It is eaten while held in the hand."
+        encoding = tokenizer.batch_encode_plus([[prompt, choice0], [prompt, choice1]], return_tensors='tf', pad_to_max_length=True)
+
+        # linear classifier on the output is not yet trained
+        outputs = model(encoding['input_ids'][None, :])
+        logits = outputs[0]
         """
         if isinstance(inputs, (tuple, list)):
             input_ids = inputs[0]
@@ -1065,7 +1072,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -1122,7 +1129,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     def call(self, inputs, **kwargs):
         r"""
     Return:

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`):
@@ -448,11 +449,10 @@ CTRL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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`).
         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_flaubert.py

@@ -91,7 +91,7 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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.

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`):
@@ -466,11 +458,10 @@ GPT2_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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`).
         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_openai.py

@@ -411,7 +411,7 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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.

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_transfo_xl.py

@@ -679,7 +679,7 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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.

src/transformers/modeling_tf_xlm.py

@@ -560,7 +560,7 @@ XLM_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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.

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
 
@@ -781,7 +779,7 @@ XLNET_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        input_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`tf.Tensor` or :obj:`Numpy array` 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.

src/transformers/modeling_transfo_xl.py

@@ -538,7 +538,7 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             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`):
+        inputs_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.
@@ -852,7 +852,7 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         r"""
         labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             Indices are selected in ``[-100, 0, ..., config.vocab_size]``
             All labels set to ``-100`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``