better error message

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

CONTRIBUTING.md

@@ -44,16 +44,9 @@ 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
-transformers-cli env
+python 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

@@ -63,7 +63,7 @@ Choose the right framework for every part of a model's lifetime
 
 ## Installation
 
-This repo is tested on Python 3.6+, PyTorch 1.0.0+ (PyTorch 1.3.1+ for examples) and TensorFlow 2.0.
+This repo is tested on Python 3.6+, PyTorch 1.0.0+ and TensorFlow 2.0.
 
 You should install 🤗 Transformers in a [virtual environment](https://docs.python.org/3/library/venv.html). If you're unfamiliar with Python virtual environments, check out the [user guide](https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/).
 
@@ -340,8 +340,8 @@ python ./examples/text-classification/run_glue.py \
     --do_eval \
     --data_dir $GLUE_DIR/$TASK_NAME \
     --max_seq_length 128 \
-    --per_device_eval_batch_size=8   \
-    --per_device_train_batch_size=8   \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_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_device_eval_batch_size=8   \
-    --per_device_train_batch_size=8   \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_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_device_eval_batch_size=8   \
-    --per_device_train_batch_size=8   \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_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_device_eval_batch_size=3   \
-    --per_device_train_batch_size=3   \
+    --per_gpu_eval_batch_size=3   \
+    --per_gpu_train_batch_size=3   \
 ```
 
 Training with these hyper-parameters gave us the following results:

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.11.0'
+release = u'2.10.0'
 
 
 # -- General configuration ---------------------------------------------------

docs/source/examples.md

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

docs/source/examples.md

@@ -0,0 +1,649 @@
+# 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/model_doc/albert.rst

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

docs/source/model_doc/bart.rst

@@ -22,7 +22,7 @@ Implementation Notes
 - The forward pass of ``BartModel`` will create decoder inputs (using the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``)  if they are not passed. This is different than some other modeling APIs.
 - Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to ``fairseq.encode`` starts with a space.
 - ``BartForConditionalGeneration.generate`` should be used for conditional generation tasks like summarization, see the example in that docstrings
-- Models that load the ``"facebook/bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
+- Models that load the ``"bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
 
 
 

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 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.
+The user can define which tokens are masked, which tokens attend "locally" and which tokens attend "globally" by setting the `config.attention_mask` `torch.Tensor` appropriately. In contrast to other models `Longformer` accepts the following values in `config.attention_mask`: `0` - the token is masked and not attended at all (as is done in other models), `1` - the token attends "locally", `2` - token attends "globally". For more information please also refer to :func:`~transformers.LongformerModel.forward` method.
 
 Using Longformer self attention, the memory and time complexity of the query-key matmul operation, which usually represents the memory and time bottleneck, can be reduced from :math:`\mathcal{O}(n_s \times n_s)` to :math:`\mathcal{O}(n_s \times w)`, with :math:`n_s` being the sequence length and :math:`w` being the average window size. It is assumed that the number of "globally" attending tokens is insignificant as compared to the number of "locally" attending tokens.
 
@@ -74,18 +74,3 @@ LongformerForQuestionAnswering
 
 .. autoclass:: transformers.LongformerForQuestionAnswering
     :members:
-
-
-LongformerForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.LongformerForMultipleChoice
-    :members:
-
-
-LongformerForTokenClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.LongformerForTokenClassification
-    :members:
-

docs/source/model_doc/roberta.rst

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

docs/source/pretrained_models.rst

@@ -63,33 +63,33 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | | Trained on uncased German text by DBMDZ                                                                                             |
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``cl-tohoku/bert-base-japanese``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-japanese``                                     | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece.                                                               |
 |                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``cl-tohoku/bert-base-japanese-whole-word-masking``        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-japanese-whole-word-masking``                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                                      |
 |                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``cl-tohoku/bert-base-japanese-char``                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-japanese-char``                                | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on Japanese text. Text is tokenized into characters.                                                                        |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``cl-tohoku/bert-base-japanese-char-whole-word-masking``   | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-japanese-char-whole-word-masking``             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters.                                               |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``TurkuNLP/bert-base-finnish-cased-v1``                    | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-finnish-cased-v1``                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on cased Finnish text.                                                                                                      |
 |                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``TurkuNLP/bert-base-finnish-uncased-v1``                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-finnish-uncased-v1``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on uncased Finnish text.                                                                                                    |
 |                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``wietsedv/bert-base-dutch-cased``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   | ``bert-base-dutch-cased``                                  | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on cased Dutch text.                                                                                                        |
 |                   |                                                            | (see `details on wietsedv repository <https://github.com/wietsedv/bertje/>`__).                                                       |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -259,32 +259,32 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   | ``xlm-roberta-large``                                      | | ~355M parameters with 24-layers, 1027-hidden-state, 4096 feed-forward hidden-state, 16-heads,                                       |
 |                   |                                                            | | Trained on 2.5 TB of newly created clean CommonCrawl data in 100 languages                                                          |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| FlauBERT          | ``flaubert/flaubert_small_cased``                          | | 6-layer, 512-hidden, 8-heads, 54M parameters                                                                                        |
+| FlauBERT          | ``flaubert-small-cased``                                   | | 6-layer, 512-hidden, 8-heads, 54M parameters                                                                                        |
 |                   |                                                            | | FlauBERT small architecture                                                                                                         |
 |                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``flaubert/flaubert_base_uncased``                         | | 12-layer, 768-hidden, 12-heads, 137M parameters                                                                                     |
+|                   | ``flaubert-base-uncased``                                  | | 12-layer, 768-hidden, 12-heads, 137M parameters                                                                                     |
 |                   |                                                            | | FlauBERT base architecture with uncased vocabulary                                                                                  |
 |                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``flaubert/flaubert_base_cased``                           | | 12-layer, 768-hidden, 12-heads, 138M parameters                                                                                     |
+|                   | ``flaubert-base-cased``                                    | | 12-layer, 768-hidden, 12-heads, 138M parameters                                                                                     |
 |                   |                                                            | | FlauBERT base architecture with cased vocabulary                                                                                    |
 |                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``flaubert/flaubert_large_cased``                          | | 24-layer, 1024-hidden, 16-heads, 373M parameters                                                                                    |
+|                   | ``flaubert-large-cased``                                   | | 24-layer, 1024-hidden, 16-heads, 373M parameters                                                                                    |
 |                   |                                                            | | FlauBERT large architecture                                                                                                         |
 |                   |                                                            | (see `details <https://github.com/getalp/Flaubert>`__)                                                                                |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| Bart              | ``facebook/bart-large``                                    | | 24-layer, 1024-hidden, 16-heads, 406M parameters                                                                                    |
+| Bart              | ``bart-large``                                             | | 24-layer, 1024-hidden, 16-heads, 406M parameters                                                                                    |
 |                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_)                                                       |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``facebook/bart-large-mnli``                               | | Adds a 2 layer classification head with 1 million parameters                                                                        |
+|                   | ``bart-large-mnli``                                        | | Adds a 2 layer classification head with 1 million parameters                                                                        |
 |                   |                                                            | | bart-large base architecture with a classification head, finetuned on MNLI                                                          |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``facebook/bart-large-cnn``                                | | 12-layer, 1024-hidden, 16-heads, 406M parameters       (same as base)                                                               |
+|                   | ``bart-large-cnn``                                         | | 12-layer, 1024-hidden, 16-heads, 406M parameters       (same as base)                                                               |
 |                   |                                                            | | bart-large base architecture finetuned on cnn summarization task                                                                    |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``facebook/mbart-large-en-ro``                             | | 12-layer, 1024-hidden, 16-heads, 880M parameters                                                                                    |
+|                   | ``mbart-large-en-ro``                                      | | 12-layer, 1024-hidden, 16-heads, 880M parameters                                                                                    |
 |                   |                                                            | | bart-large architecture pretrained on cc25 multilingual data , finetuned on WMT english romanian translation.                       |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | DialoGPT          | ``DialoGPT-small``                                         | | 12-layer, 768-hidden, 12-heads, 124M parameters                                                                                     |
@@ -305,9 +305,9 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 | MarianMT          | ``Helsinki-NLP/opus-mt-{src}-{tgt}``                       | | 12-layer, 512-hidden, 8-heads, ~74M parameter Machine translation models. Parameter counts vary depending on vocab size.            |
 |                   |                                                            | | (see `model list <https://huggingface.co/Helsinki-NLP>`_)                                                                           |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| Longformer        | ``allenai/longformer-base-4096``                           | | 12-layer, 768-hidden, 12-heads, ~149M parameters                                                                                    |
+| Longformer        | ``longformer-base-4096``                                   | | 12-layer, 768-hidden, 12-heads, ~149M parameters                                                                                    |
 |                   |                                                            | | Starting from RoBERTa-base checkpoint, trained on documents of max length 4,096                                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                   | ``allenai/longformer-large-4096``                          | | 24-layer, 1024-hidden, 16-heads, ~435M parameters                                                                                   |
+|                   | ``longformer-large-4096``                                  | | 24-layer, 1024-hidden, 16-heads, ~435M parameters                                                                                   |
 |                   |                                                            | | Starting from RoBERTa-large checkpoint, trained on documents of max length 4,096                                                    |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

examples/README.md

@@ -1,7 +1,6 @@
 ## Examples
 
 Version 2.9 of `transformers` introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.0+.
 
 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
@@ -17,17 +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`**](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              | -  | -  | -  | -
+| [**`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              | -  | -  | -  | -
 
 
 <br>
@@ -58,7 +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](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.
+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.
 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/adversarial/test_hans.py

@@ -65,6 +65,13 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+ALL_MODELS = sum(
+    (
+        tuple(conf.pretrained_config_archive_map.keys())
+        for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig, DistilBertConfig)
+    ),
+    (),
+)
 
 MODEL_CLASSES = {
     "bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
@@ -382,7 +389,7 @@ def main():
         default=None,
         type=str,
         required=True,
-        help="Path to pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--task_name",

examples/contrib/mm-imdb/run_mmimdb.py

@@ -34,11 +34,26 @@ from tqdm import tqdm, trange
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
-    AutoConfig,
-    AutoModel,
-    AutoTokenizer,
+    AlbertConfig,
+    AlbertModel,
+    AlbertTokenizer,
+    BertConfig,
+    BertModel,
+    BertTokenizer,
+    DistilBertConfig,
+    DistilBertModel,
+    DistilBertTokenizer,
     MMBTConfig,
     MMBTForClassification,
+    RobertaConfig,
+    RobertaModel,
+    RobertaTokenizer,
+    XLMConfig,
+    XLMModel,
+    XLMTokenizer,
+    XLNetConfig,
+    XLNetModel,
+    XLNetTokenizer,
     get_linear_schedule_with_warmup,
 )
 from utils_mmimdb import ImageEncoder, JsonlDataset, collate_fn, get_image_transforms, get_mmimdb_labels
@@ -52,6 +67,23 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+ALL_MODELS = sum(
+    (
+        tuple(conf.pretrained_config_archive_map.keys())
+        for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig, DistilBertConfig)
+    ),
+    (),
+)
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertModel, BertTokenizer),
+    "xlnet": (XLNetConfig, XLNetModel, XLNetTokenizer),
+    "xlm": (XLMConfig, XLMModel, XLMTokenizer),
+    "roberta": (RobertaConfig, RobertaModel, RobertaTokenizer),
+    "distilbert": (DistilBertConfig, DistilBertModel, DistilBertTokenizer),
+    "albert": (AlbertConfig, AlbertModel, AlbertTokenizer),
+}
+
 
 def set_seed(args):
     random.seed(args.seed)
@@ -319,12 +351,19 @@ def main():
         required=True,
         help="The input data dir. Should contain the .jsonl files for MMIMDB.",
     )
+    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 pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--output_dir",
@@ -346,7 +385,7 @@ def main():
     )
     parser.add_argument(
         "--cache_dir",
-        default=None,
+        default="",
         type=str,
         help="Where do you want to store the pre-trained models downloaded from s3",
     )
@@ -487,14 +526,18 @@ def main():
     # Setup model
     labels = get_mmimdb_labels()
     num_labels = len(labels)
-    transformer_config = AutoConfig.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
-    tokenizer = AutoTokenizer.from_pretrained(
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    transformer_config = config_class.from_pretrained(
+        args.config_name if args.config_name else args.model_name_or_path
+    )
+    tokenizer = tokenizer_class.from_pretrained(
         args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
         do_lower_case=args.do_lower_case,
-        cache_dir=args.cache_dir,
+        cache_dir=args.cache_dir if args.cache_dir else None,
     )
-    transformer = AutoModel.from_pretrained(
-        args.model_name_or_path, config=transformer_config, cache_dir=args.cache_dir
+    transformer = model_class.from_pretrained(
+        args.model_name_or_path, config=transformer_config, cache_dir=args.cache_dir if args.cache_dir else None
     )
     img_encoder = ImageEncoder(args)
     config = MMBTConfig(transformer_config, num_labels=num_labels)
@@ -540,12 +583,13 @@ def main():
         # Load a trained model and vocabulary that you have fine-tuned
         model = MMBTForClassification(config, transformer, img_encoder)
         model.load_state_dict(torch.load(os.path.join(args.output_dir, WEIGHTS_NAME)))
-        tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
         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(

examples/contrib/run_swag.py

@@ -31,8 +31,14 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 from torch.utils.data.distributed import DistributedSampler
 from tqdm import tqdm, trange
 
-from transformers import WEIGHTS_NAME, AdamW, AutoConfig, AutoTokenizer, get_linear_schedule_with_warmup
-from transformers.modeling_auto import AutoModelForMultipleChoice
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    BertConfig,
+    BertForMultipleChoice,
+    BertTokenizer,
+    get_linear_schedule_with_warmup,
+)
 
 
 try:
@@ -43,6 +49,12 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in [BertConfig]), ())
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForMultipleChoice, BertTokenizer),
+}
+
 
 class SwagExample(object):
     """A single training/test example for the SWAG dataset."""
@@ -480,12 +492,19 @@ def main():
         required=True,
         help="SWAG csv for predictions. E.g., val.csv or test.csv",
     )
+    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 pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--output_dir",
@@ -517,6 +536,9 @@ def main():
     parser.add_argument(
         "--evaluate_during_training", action="store_true", help="Rul 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(
@@ -630,9 +652,13 @@ def main():
     if args.local_rank not in [-1, 0]:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
-    config = AutoConfig.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
-    tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,)
-    model = AutoModelForMultipleChoice.from_pretrained(
+    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)
+    tokenizer = tokenizer_class.from_pretrained(
+        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case
+    )
+    model = model_class.from_pretrained(
         args.model_name_or_path, from_tf=bool(".ckpt" in args.model_name_or_path), config=config
     )
 
@@ -668,8 +694,8 @@ def main():
         torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = AutoModelForMultipleChoice.from_pretrained(args.output_dir)
-        tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
         model.to(args.device)
 
     # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
@@ -692,8 +718,8 @@ def main():
         for checkpoint in checkpoints:
             # Reload the model
             global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
-            model = AutoModelForMultipleChoice.from_pretrained(checkpoint)
-            tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+            model = model_class.from_pretrained(checkpoint)
+            tokenizer = tokenizer_class.from_pretrained(checkpoint)
             model.to(args.device)
 
             # Evaluate

examples/distillation/run_squad_w_distillation.py

@@ -67,6 +67,9 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig)), ()
+)
 
 MODEL_CLASSES = {
     "bert": (BertConfig, BertForQuestionAnswering, BertTokenizer),
@@ -502,7 +505,7 @@ def main():
         default=None,
         type=str,
         required=True,
-        help="Path to pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--output_dir",

examples/movement-pruning/README.md

@@ -1,183 +0,0 @@
-# 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

@@ -1,612 +0,0 @@
-{
- "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

@@ -1,132 +0,0 @@
-# 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

@@ -1,92 +0,0 @@
-# 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

@@ -1,10 +0,0 @@
-# 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

@@ -1,71 +0,0 @@
-# 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_utils import PretrainedConfig
-
-
-logger = logging.getLogger(__name__)
-
-
-class MaskedBertConfig(PretrainedConfig):
-    """
-    A class replicating the `~transformers.BertConfig` with additional parameters for pruning/masking configuration.
-    """
-
-    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

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

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

@@ -1,144 +0,0 @@
-# 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

@@ -1,107 +0,0 @@
-# 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

@@ -1,924 +0,0 @@
-# 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__)
-
-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 pretrained model or model identifier from huggingface.co/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

@@ -1,6 +0,0 @@
-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_device_train_batch_size=16 \
+--per_gpu_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_device_train_batch_size=16 \
+--per_gpu_train_batch_size=16 \
 --logging-dir logs \
 --gradient_accumulation_steps 2 \
 --overwrite_output

examples/question-answering/run_squad.py

@@ -58,6 +58,8 @@ logger = logging.getLogger(__name__)
 MODEL_CONFIG_CLASSES = list(MODEL_FOR_QUESTION_ANSWERING_MAPPING.keys())
 MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
 
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in MODEL_CONFIG_CLASSES), (),)
+
 
 def set_seed(args):
     random.seed(args.seed)
@@ -489,7 +491,7 @@ def main():
         default=None,
         type=str,
         required=True,
-        help="Path to pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--output_dir",

examples/summarization/bart/evaluate_cnn.py

@@ -21,7 +21,7 @@ def generate_summaries(
 ):
     fout = Path(out_file).open("w")
     model = BartForConditionalGeneration.from_pretrained(model_name).to(device)
-    tokenizer = BartTokenizer.from_pretrained("facebook/bart-large")
+    tokenizer = BartTokenizer.from_pretrained("bart-large")
 
     max_length = 140
     min_length = 55
@@ -54,7 +54,7 @@ def run_generate():
         "output_path", type=str, help="where to save summaries",
     )
     parser.add_argument(
-        "model_name", type=str, default="facebook/bart-large-cnn", help="like bart-large-cnn",
+        "model_name", type=str, default="bart-large-cnn", help="like bart-large-cnn",
     )
     parser.add_argument(
         "--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.",

examples/summarization/bart/test_bart_examples.py

@@ -129,7 +129,7 @@ class TestBartExamples(unittest.TestCase):
         summaries = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
         _dump_articles((tmp_dir / "train.source"), articles)
         _dump_articles((tmp_dir / "train.target"), summaries)
-        tokenizer = BartTokenizer.from_pretrained("facebook/bart-large")
+        tokenizer = BartTokenizer.from_pretrained("bart-large")
         max_len_source = max(len(tokenizer.encode(a)) for a in articles)
         max_len_target = max(len(tokenizer.encode(a)) for a in summaries)
         trunc_target = 4

examples/summarization/bertabs/configuration_bertabs.py

@@ -61,6 +61,7 @@ class BertAbsConfig(PretrainedConfig):
             the decoder.
     """
 
+    pretrained_config_archive_map = BERTABS_FINETUNED_CONFIG_MAP
     model_type = "bertabs"
 
     def __init__(

examples/summarization/bertabs/modeling_bertabs.py

@@ -33,13 +33,14 @@ from transformers import BertConfig, BertModel, PreTrainedModel
 
 MAX_SIZE = 5000
 
-BERTABS_FINETUNED_MODEL_ARCHIVE_LIST = [
-    "remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization",
-]
+BERTABS_FINETUNED_MODEL_MAP = {
+    "bertabs-finetuned-cnndm": "https://cdn.huggingface.co/remi/bertabs-finetuned-cnndm-extractive-abstractive-summarization/pytorch_model.bin",
+}
 
 
 class BertAbsPreTrainedModel(PreTrainedModel):
     config_class = BertAbsConfig
+    pretrained_model_archive_map = BERTABS_FINETUNED_MODEL_MAP
     load_tf_weights = False
     base_model_prefix = "bert"
 

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_device_train_batch_size=2
-            --per_device_eval_batch_size=1
+            --per_gpu_train_batch_size=2
+            --per_gpu_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_device_train_batch_size 32 \
+  --per_gpu_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_device_train_batch_size 32 \
+  --per_gpu_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_device_train_batch_size 32 \
+  --per_gpu_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_device_train_batch_size 8 \
+    --per_gpu_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_device_train_batch_size 8 \
+    --per_gpu_train_batch_size 8 \
     --learning_rate 2e-5 \
     --num_train_epochs 3.0 \
     --output_dir output_dir \
@@ -258,7 +258,7 @@ TEST RESULTS {'val_loss': tensor(0.0707), 'precision': 0.852427800698191, 'recal
 
 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 a 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).
+[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
 
@@ -273,13 +273,14 @@ on a single tesla V100 16GB. The data for XNLI can be downloaded with the follow
 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_device_train_batch_size 32 \
+  --per_gpu_train_batch_size 32 \
   --learning_rate 5e-5 \
   --num_train_epochs 2.0 \
   --max_seq_length 128 \

examples/text-classification/run_xnli.py

@@ -13,7 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Finetuning multi-lingual models on XNLI (e.g. Bert, DistilBERT, XLM).
+""" Finetuning multi-lingual models on XNLI (Bert, DistilBERT, XLM).
     Adapted from `examples/text-classification/run_glue.py`"""
 
 
@@ -32,9 +32,15 @@ from tqdm import tqdm, trange
 from transformers import (
     WEIGHTS_NAME,
     AdamW,
-    AutoConfig,
-    AutoModelForSequenceClassification,
-    AutoTokenizer,
+    BertConfig,
+    BertForSequenceClassification,
+    BertTokenizer,
+    DistilBertConfig,
+    DistilBertForSequenceClassification,
+    DistilBertTokenizer,
+    XLMConfig,
+    XLMForSequenceClassification,
+    XLMTokenizer,
     get_linear_schedule_with_warmup,
 )
 from transformers import glue_convert_examples_to_features as convert_examples_to_features
@@ -51,6 +57,16 @@ except ImportError:
 
 logger = logging.getLogger(__name__)
 
+ALL_MODELS = sum(
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, DistilBertConfig, XLMConfig)), ()
+)
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, BertForSequenceClassification, BertTokenizer),
+    "xlm": (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
+    "distilbert": (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
+}
+
 
 def set_seed(args):
     random.seed(args.seed)
@@ -361,12 +377,19 @@ def main():
         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 pretrained model or model identifier from huggingface.co/models",
+        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS),
     )
     parser.add_argument(
         "--language",
@@ -398,7 +421,7 @@ def main():
     )
     parser.add_argument(
         "--cache_dir",
-        default=None,
+        default="",
         type=str,
         help="Where do you want to store the pre-trained models downloaded from s3",
     )
@@ -539,23 +562,24 @@ def main():
     if args.local_rank not in [-1, 0]:
         torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
 
-    config = AutoConfig.from_pretrained(
+    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,
+        cache_dir=args.cache_dir if args.cache_dir else None,
     )
-    args.model_type = config.model_type
-    tokenizer = AutoTokenizer.from_pretrained(
+    tokenizer = tokenizer_class.from_pretrained(
         args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
         do_lower_case=args.do_lower_case,
-        cache_dir=args.cache_dir,
+        cache_dir=args.cache_dir if args.cache_dir else None,
     )
-    model = AutoModelForSequenceClassification.from_pretrained(
+    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,
+        cache_dir=args.cache_dir if args.cache_dir else None,
     )
 
     if args.local_rank == 0:
@@ -590,13 +614,14 @@ def main():
         torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
 
         # Load a trained model and vocabulary that you have fine-tuned
-        model = AutoModelForSequenceClassification.from_pretrained(args.output_dir)
-        tokenizer = AutoTokenizer.from_pretrained(args.output_dir)
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
         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(
@@ -608,7 +633,7 @@ def main():
             global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
             prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
 
-            model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
+            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())

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_device_train_batch_size $BATCH_SIZE \
+--per_gpu_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_device_train_batch_size": 32,
+    "per_gpu_train_batch_size": 32,
     "save_steps": 750,
     "seed": 1,
     "do_train": true,

examples/token-classification/run_ner.py

@@ -13,7 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Fine-tuning the library models for named entity recognition on CoNLL-2003. """
+""" Fine-tuning the library models for named entity recognition on CoNLL-2003 (Bert or Roberta). """
 
 
 import logging

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

@@ -1,124 +0,0 @@
-## 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

@@ -1,124 +0,0 @@
-## 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

@@ -1,124 +0,0 @@
-## 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

@@ -1,124 +0,0 @@
-## 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/allenai/longformer-base-4096-extra.pos.embd.only/README.md

@@ -1,20 +0,0 @@
-
-# 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

@@ -1,24 +0,0 @@
-
-# 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/dkleczek/bert-base-polish-cased-v1/README.md

@@ -1,135 +0,0 @@
----
-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

@@ -4,27 +4,14 @@ thumbnail: https://raw.githubusercontent.com/kldarek/polbert/master/img/polbert.
 ---
 
 # 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. 
+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!
 
 ![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 |
@@ -33,21 +20,7 @@ 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: 
@@ -56,22 +29,10 @@ 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](/LM_testing.ipynb) file. 
-
-### Uncased
+For an example use as language model, see [this notebook](https://github.com/kldarek/polbert/blob/master/LM_testing.ipynb) file. 
 
 ```python
 from transformers import *
@@ -80,6 +41,7 @@ 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}
@@ -88,42 +50,23 @@ 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
-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.
+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. 
 
-| 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 |
+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. 
 
-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. 
+| PolEmo 2.0 Sentiment Classifcation | Test Accuracy | 
+| ------------- |--------------:|
+| Multilingual BERT | 0.78 |
+| Polbert | 0.85 |
 
 ## 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! 
+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!
 
 ## Author
 Darek Kłeczek - contact me on Twitter [@dk21](https://twitter.com/dk21)
@@ -132,4 +75,5 @@ 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
-* [KLEJ benchmark](https://klejbenchmark.com/leaderboard/)
+* PolEmo 2.0 Sentiment Analysis Dataset for CoNLL: https://clarin-pl.eu/dspace/handle/11321/710
+

model_cards/mrm8488/longformer-base-4096-finetuned-squadv2/README.md

@@ -1,65 +0,0 @@
----
-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/valhalla/longformer-base-4096-finetuned-squadv1/README.md

@@ -1,54 +0,0 @@
-# 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

@@ -25,12 +25,6 @@ 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) |
-| [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 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) |
-| [Fine-tune T5 for Sentiment Span Extraction](https://github.com/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb)  | How to fine-tune T5 for sentiment span extraction using a text-to-text format with PyTorch Lightning |  [Lorenzo Ampil](https://github.com/enzoampil) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) |

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"]
 
@@ -88,14 +88,14 @@ 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"]
 
 setup(
     name="transformers",
-    version="2.11.0",
+    version="2.10.0",
     author="Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Sam Shleifer, Patrick von Platen, Google AI Language Team Authors, Open AI team Authors, Facebook AI Authors, Carnegie Mellon University Authors",
     author_email="thomas@huggingface.co",
     description="State-of-the-art Natural Language Processing for TensorFlow 2.0 and PyTorch",

src/transformers/__init__.py

@@ -2,7 +2,7 @@
 # There's no way to ignore "F401 '...' imported but unused" warnings in this
 # module, but to preserve other warnings. So, don't check this module at all.
 
-__version__ = "2.11.0"
+__version__ = "2.10.0"
 
 # Work around to update TensorFlow's absl.logging threshold which alters the
 # default Python logging output behavior when present.
@@ -159,6 +159,7 @@ if is_torch_available():
         AutoModelWithLMHead,
         AutoModelForTokenClassification,
         AutoModelForMultipleChoice,
+        ALL_PRETRAINED_MODEL_ARCHIVE_MAP,
         MODEL_MAPPING,
         MODEL_FOR_PRETRAINING_MAPPING,
         MODEL_WITH_LM_HEAD_MAPPING,
@@ -179,7 +180,7 @@ if is_torch_available():
         BertForTokenClassification,
         BertForQuestionAnswering,
         load_tf_weights_in_bert,
-        BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         BertLayer,
     )
     from .modeling_openai import (
@@ -188,7 +189,7 @@ if is_torch_available():
         OpenAIGPTLMHeadModel,
         OpenAIGPTDoubleHeadsModel,
         load_tf_weights_in_openai_gpt,
-        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_transfo_xl import (
         TransfoXLPreTrainedModel,
@@ -196,7 +197,7 @@ if is_torch_available():
         TransfoXLLMHeadModel,
         AdaptiveEmbedding,
         load_tf_weights_in_transfo_xl,
-        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_gpt2 import (
         GPT2PreTrainedModel,
@@ -204,9 +205,9 @@ if is_torch_available():
         GPT2LMHeadModel,
         GPT2DoubleHeadsModel,
         load_tf_weights_in_gpt2,
-        GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
-    from .modeling_ctrl import CTRLPreTrainedModel, CTRLModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_LIST
+    from .modeling_ctrl import CTRLPreTrainedModel, CTRLModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
     from .modeling_xlnet import (
         XLNetPreTrainedModel,
         XLNetModel,
@@ -217,7 +218,7 @@ if is_torch_available():
         XLNetForQuestionAnsweringSimple,
         XLNetForQuestionAnswering,
         load_tf_weights_in_xlnet,
-        XLNET_PRETRAINED_MODEL_ARCHIVE_LIST,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_xlm import (
         XLMPreTrainedModel,
@@ -227,13 +228,13 @@ if is_torch_available():
         XLMForTokenClassification,
         XLMForQuestionAnswering,
         XLMForQuestionAnsweringSimple,
-        XLM_PRETRAINED_MODEL_ARCHIVE_LIST,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_bart import (
         BartForSequenceClassification,
         BartModel,
         BartForConditionalGeneration,
-        BART_PRETRAINED_MODEL_ARCHIVE_LIST,
+        BART_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_marian import MarianMTModel
     from .tokenization_marian import MarianTokenizer
@@ -244,7 +245,7 @@ if is_torch_available():
         RobertaForMultipleChoice,
         RobertaForTokenClassification,
         RobertaForQuestionAnswering,
-        ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_distilbert import (
         DistilBertPreTrainedModel,
@@ -253,7 +254,7 @@ if is_torch_available():
         DistilBertForSequenceClassification,
         DistilBertForQuestionAnswering,
         DistilBertForTokenClassification,
-        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_camembert import (
         CamembertForMaskedLM,
@@ -262,7 +263,7 @@ if is_torch_available():
         CamembertForMultipleChoice,
         CamembertForTokenClassification,
         CamembertForQuestionAnswering,
-        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_encoder_decoder import EncoderDecoderModel
     from .modeling_t5 import (
@@ -270,7 +271,7 @@ if is_torch_available():
         T5Model,
         T5ForConditionalGeneration,
         load_tf_weights_in_t5,
-        T5_PRETRAINED_MODEL_ARCHIVE_LIST,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_albert import (
         AlbertPreTrainedModel,
@@ -281,7 +282,7 @@ if is_torch_available():
         AlbertForQuestionAnswering,
         AlbertForTokenClassification,
         load_tf_weights_in_albert,
-        ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_xlm_roberta import (
         XLMRobertaForMaskedLM,
@@ -289,7 +290,7 @@ if is_torch_available():
         XLMRobertaForMultipleChoice,
         XLMRobertaForSequenceClassification,
         XLMRobertaForTokenClassification,
-        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
     from .modeling_mmbt import ModalEmbeddings, MMBTModel, MMBTForClassification
 
@@ -299,7 +300,7 @@ if is_torch_available():
         FlaubertForSequenceClassification,
         FlaubertForQuestionAnswering,
         FlaubertForQuestionAnsweringSimple,
-        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_electra import (
@@ -310,7 +311,7 @@ if is_torch_available():
         ElectraForSequenceClassification,
         ElectraModel,
         load_tf_weights_in_electra,
-        ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_reformer import (
@@ -318,17 +319,14 @@ if is_torch_available():
         ReformerLayer,
         ReformerModel,
         ReformerModelWithLMHead,
-        REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        REFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_longformer import (
         LongformerModel,
         LongformerForMaskedLM,
-        LongformerForSequenceClassification,
-        LongformerForMultipleChoice,
-        LongformerForTokenClassification,
         LongformerForQuestionAnswering,
-        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     # Optimization
@@ -366,6 +364,7 @@ if is_tf_available():
         TFAutoModelForQuestionAnswering,
         TFAutoModelWithLMHead,
         TFAutoModelForTokenClassification,
+        TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP,
         TF_MODEL_MAPPING,
         TF_MODEL_FOR_PRETRAINING_MAPPING,
         TF_MODEL_WITH_LM_HEAD_MAPPING,
@@ -386,7 +385,7 @@ if is_tf_available():
         TFBertForMultipleChoice,
         TFBertForTokenClassification,
         TFBertForQuestionAnswering,
-        TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_gpt2 import (
@@ -395,7 +394,7 @@ if is_tf_available():
         TFGPT2Model,
         TFGPT2LMHeadModel,
         TFGPT2DoubleHeadsModel,
-        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_openai import (
@@ -404,7 +403,7 @@ if is_tf_available():
         TFOpenAIGPTModel,
         TFOpenAIGPTLMHeadModel,
         TFOpenAIGPTDoubleHeadsModel,
-        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_transfo_xl import (
@@ -412,7 +411,7 @@ if is_tf_available():
         TFTransfoXLMainLayer,
         TFTransfoXLModel,
         TFTransfoXLLMHeadModel,
-        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
         TFAdaptiveEmbedding,
     )
 
@@ -424,7 +423,7 @@ if is_tf_available():
         TFXLNetForSequenceClassification,
         TFXLNetForTokenClassification,
         TFXLNetForQuestionAnsweringSimple,
-        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_xlm import (
@@ -434,7 +433,7 @@ if is_tf_available():
         TFXLMWithLMHeadModel,
         TFXLMForSequenceClassification,
         TFXLMForQuestionAnsweringSimple,
-        TF_XLM_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_xlm_roberta import (
@@ -442,7 +441,7 @@ if is_tf_available():
         TFXLMRobertaModel,
         TFXLMRobertaForSequenceClassification,
         TFXLMRobertaForTokenClassification,
-        TF_XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_roberta import (
@@ -453,7 +452,7 @@ if is_tf_available():
         TFRobertaForSequenceClassification,
         TFRobertaForTokenClassification,
         TFRobertaForQuestionAnswering,
-        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_camembert import (
@@ -461,14 +460,14 @@ if is_tf_available():
         TFCamembertForMaskedLM,
         TFCamembertForSequenceClassification,
         TFCamembertForTokenClassification,
-        TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_flaubert import (
         TFFlaubertModel,
         TFFlaubertWithLMHeadModel,
         TFFlaubertForSequenceClassification,
-        TF_FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_distilbert import (
@@ -479,14 +478,14 @@ if is_tf_available():
         TFDistilBertForSequenceClassification,
         TFDistilBertForTokenClassification,
         TFDistilBertForQuestionAnswering,
-        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_ctrl import (
         TFCTRLPreTrainedModel,
         TFCTRLModel,
         TFCTRLLMHeadModel,
-        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_albert import (
@@ -498,14 +497,14 @@ if is_tf_available():
         TFAlbertForMultipleChoice,
         TFAlbertForSequenceClassification,
         TFAlbertForQuestionAnswering,
-        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_t5 import (
         TFT5PreTrainedModel,
         TFT5Model,
         TFT5ForConditionalGeneration,
-        TF_T5_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     from .modeling_tf_electra import (
@@ -514,7 +513,7 @@ if is_tf_available():
         TFElectraForPreTraining,
         TFElectraForMaskedLM,
         TFElectraForTokenClassification,
-        TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
     )
 
     # Optimization

src/transformers/benchmark/benchmark.py

@@ -94,7 +94,8 @@ class PyTorchBenchmark(Benchmark):
                 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))
+            self.print_fn("{} with batch size {} and sequence length {} doesn't fit on GPU for training.".format(model_name, batch_size, sequence_length))
+            logger.info(e)
             return "N/A"
 
     def inference(self, model_name, batch_size, sequence_length, trace_memory=False):
@@ -142,5 +143,6 @@ class PyTorchBenchmark(Benchmark):
                 return min(runtimes) / 10.0
 
         except RuntimeError as e:
-            self.print_fn("Doesn't fit on GPU. {}".format(e))
+            self.print_fn("{} with batch size {} and sequence length {} doesn't fit on GPU for inference.".format(model_name, batch_size, sequence_length))
+            logger.info(e)
             return "N/A"

src/transformers/configuration_albert.py

@@ -32,7 +32,7 @@ ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 class AlbertConfig(PretrainedConfig):
     r"""
-        This is the configuration class to store the configuration of a :class:`~transformers.AlbertModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.AlbertModel`.
         It is used to instantiate an ALBERT model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the ALBERT `xxlarge <https://huggingface.co/albert-xxlarge-v2>`__ architecture.
@@ -97,8 +97,13 @@ class AlbertConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "albert"
 
     def __init__(

src/transformers/configuration_auto.py

@@ -113,12 +113,12 @@ class AutoConfig:
         )
 
     @classmethod
-    def for_model(cls, model_type: str, *args, **kwargs):
-        if model_type in CONFIG_MAPPING:
-            config_class = CONFIG_MAPPING[model_type]
-            return config_class(*args, **kwargs)
+    def for_model(cls, model_type, *args, **kwargs):
+        for pattern, config_class in CONFIG_MAPPING.items():
+            if pattern in model_type:
+                return config_class(*args, **kwargs)
         raise ValueError(
-            "Unrecognized model identifier: {}. Should contain one of {}".format(
+            "Unrecognized model identifier in {}. Should contain one of {}".format(
                 model_type, ", ".join(CONFIG_MAPPING.keys())
             )
         )
@@ -130,24 +130,24 @@ class AutoConfig:
 
         The configuration class to instantiate is selected
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: :class:`~transformers.T5Config` (T5 model)
-            - `distilbert`: :class:`~transformers.DistilBertConfig` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertConfig` (ALBERT model)
-            - `camembert`: :class:`~transformers.CamembertConfig` (CamemBERT model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaConfig` (XLM-RoBERTa model)
-            - `longformer`: :class:`~transformers.LongformerConfig` (Longformer model)
-            - `roberta`: :class:`~transformers.RobertaConfig` (RoBERTa model)
-            - `reformer`: :class:`~transformers.ReformerConfig` (Reformer model)
-            - `bert`: :class:`~transformers.BertConfig` (Bert model)
-            - `openai-gpt`: :class:`~transformers.OpenAIGPTConfig` (OpenAI GPT model)
-            - `gpt2`: :class:`~transformers.GPT2Config` (OpenAI GPT-2 model)
-            - `transfo-xl`: :class:`~transformers.TransfoXLConfig` (Transformer-XL model)
-            - `xlnet`: :class:`~transformers.XLNetConfig` (XLNet model)
-            - `xlm`: :class:`~transformers.XLMConfig` (XLM model)
-            - `ctrl` : :class:`~transformers.CTRLConfig` (CTRL model)
-            - `flaubert` : :class:`~transformers.FlaubertConfig` (Flaubert model)
-            - `electra` : :class:`~transformers.ElectraConfig` (ELECTRA model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+            - contains `t5`: :class:`~transformers.T5Config` (T5 model)
+            - contains `distilbert`: :class:`~transformers.DistilBertConfig` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertConfig` (ALBERT model)
+            - contains `camembert`: :class:`~transformers.CamembertConfig` (CamemBERT model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaConfig` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerConfig` (Longformer model)
+            - contains `roberta`: :class:`~transformers.RobertaConfig` (RoBERTa model)
+            - contains `reformer`: :class:`~transformers.ReformerConfig` (Reformer model)
+            - contains `bert`: :class:`~transformers.BertConfig` (Bert model)
+            - contains `openai-gpt`: :class:`~transformers.OpenAIGPTConfig` (OpenAI GPT model)
+            - contains `gpt2`: :class:`~transformers.GPT2Config` (OpenAI GPT-2 model)
+            - contains `transfo-xl`: :class:`~transformers.TransfoXLConfig` (Transformer-XL model)
+            - contains `xlnet`: :class:`~transformers.XLNetConfig` (XLNet model)
+            - contains `xlm`: :class:`~transformers.XLMConfig` (XLM model)
+            - contains `ctrl` : :class:`~transformers.CTRLConfig` (CTRL model)
+            - contains `flaubert` : :class:`~transformers.FlaubertConfig` (Flaubert model)
+            - contains `electra` : :class:`~transformers.ElectraConfig` (ELECTRA model)
 
         Args:
             pretrained_model_name_or_path (:obj:`string`):
@@ -193,7 +193,9 @@ class AutoConfig:
             assert unused_kwargs == {'foo': False}
 
         """
-        config_dict, _ = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        config_dict, _ = PretrainedConfig.get_config_dict(
+            pretrained_model_name_or_path, pretrained_config_archive_map=ALL_PRETRAINED_CONFIG_ARCHIVE_MAP, **kwargs
+        )
 
         if "model_type" in config_dict:
             config_class = CONFIG_MAPPING[config_dict["model_type"]]

src/transformers/configuration_bart.py

@@ -23,11 +23,11 @@ from .configuration_utils import PretrainedConfig
 logger = logging.getLogger(__name__)
 
 BART_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "facebook/bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/config.json",
-    "facebook/bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/config.json",
-    "facebook/bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/config.json",
-    "facebook/bart-large-xsum": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-xsum/config.json",
-    "facebook/mbart-large-en-ro": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/mbart-large-en-ro/config.json",
+    "bart-large": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large/config.json",
+    "bart-large-mnli": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-mnli/config.json",
+    "bart-large-cnn": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-cnn/config.json",
+    "bart-large-xsum": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/bart-large-xsum/config.json",
+    "mbart-large-en-ro": "https://s3.amazonaws.com/models.huggingface.co/bert/facebook/mbart-large-en-ro/config.json",
 }
 
 
@@ -36,6 +36,7 @@ class BartConfig(PretrainedConfig):
         Configuration class for Bart. Parameters are renamed from the fairseq implementation
     """
     model_type = "bart"
+    pretrained_config_archive_map = BART_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(
         self,

src/transformers/configuration_bert.py

@@ -39,14 +39,13 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     "bert-base-cased-finetuned-mrpc": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
     "bert-base-german-dbmdz-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
     "bert-base-german-dbmdz-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
-    "cl-tohoku/bert-base-japanese": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese/config.json",
-    "cl-tohoku/bert-base-japanese-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking/config.json",
-    "cl-tohoku/bert-base-japanese-char": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char/config.json",
-    "cl-tohoku/bert-base-japanese-char-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking/config.json",
-    "TurkuNLP/bert-base-finnish-cased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
-    "TurkuNLP/bert-base-finnish-uncased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
-    "wietsedv/bert-base-dutch-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/wietsedv/bert-base-dutch-cased/config.json",
-    # See all BERT models at https://huggingface.co/models?filter=bert
+    "bert-base-japanese": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese/config.json",
+    "bert-base-japanese-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking/config.json",
+    "bert-base-japanese-char": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char/config.json",
+    "bert-base-japanese-char-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking/config.json",
+    "bert-base-finnish-cased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
+    "bert-base-finnish-uncased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
+    "bert-base-dutch-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/wietsedv/bert-base-dutch-cased/config.json",
 }
 
 
@@ -103,7 +102,12 @@ class BertConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "bert"
 
     def __init__(

src/transformers/configuration_camembert.py

@@ -36,4 +36,5 @@ class CamembertConfig(RobertaConfig):
     superclass for the appropriate documentation alongside usage examples.
     """
 
+    pretrained_config_archive_map = CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "camembert"

src/transformers/configuration_ctrl.py

@@ -22,12 +22,12 @@ from .configuration_utils import PretrainedConfig
 
 logger = logging.getLogger(__name__)
 
-CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP = {"ctrl": "https://s3.amazonaws.com/models.huggingface.co/bert/ctrl-config.json"}
+CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP = {"ctrl": "https://storage.googleapis.com/sf-ctrl/pytorch/ctrl-config.json"}
 
 
 class CTRLConfig(PretrainedConfig):
     """
-        This is the configuration class to store the configuration of a :class:`~transformers.CTRLModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.CTRLModel`.
         It is used to instantiate an CTRL model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the `ctrl <https://huggingface.co/ctrl>`__ architecture from SalesForce.
@@ -76,8 +76,13 @@ class CTRLConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "ctrl"
 
     def __init__(

src/transformers/configuration_distilbert.py

@@ -90,7 +90,12 @@ class DistilBertConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "distilbert"
 
     def __init__(

src/transformers/configuration_electra.py

@@ -89,7 +89,12 @@ class ElectraConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "electra"
 
     def __init__(

src/transformers/configuration_flaubert.py

@@ -23,10 +23,10 @@ from .configuration_xlm import XLMConfig
 logger = logging.getLogger(__name__)
 
 FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "flaubert/flaubert_small_cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_small_cased/config.json",
-    "flaubert/flaubert_base_uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_base_uncased/config.json",
-    "flaubert/flaubert_base_cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_base_cased/config.json",
-    "flaubert/flaubert_large_cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_large_cased/config.json",
+    "flaubert-small-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_small_cased/config.json",
+    "flaubert-base-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_base_uncased/config.json",
+    "flaubert-base-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_base_cased/config.json",
+    "flaubert-large-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/flaubert/flaubert_large_cased/config.json",
 }
 
 
@@ -142,6 +142,7 @@ class FlaubertConfig(XLMConfig):
                 text in a given language.
     """
 
+    pretrained_config_archive_map = FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "flaubert"
 
     def __init__(self, layerdrop=0.0, pre_norm=False, pad_token_id=2, bos_token_id=0, **kwargs):

src/transformers/configuration_gpt2.py

@@ -110,8 +110,13 @@ class GPT2Config(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "gpt2"
 
     def __init__(

src/transformers/configuration_longformer.py

@@ -23,17 +23,15 @@ from .configuration_roberta import RobertaConfig
 logger = logging.getLogger(__name__)
 
 LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "allenai/longformer-base-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096/config.json",
-    "allenai/longformer-large-4096": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096/config.json",
-    "allenai/longformer-large-4096-finetuned-triviaqa": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-finetuned-triviaqa/config.json",
-    "allenai/longformer-base-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-base-4096-extra.pos.embd.only/config.json",
-    "allenai/longformer-large-4096-extra.pos.embd.only": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-extra.pos.embd.only/config.json",
+    "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",
+    "longformer-large-4096-finetuned-triviaqa": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-finetuned-triviaqa/config.json",
 }
 
 
 class LongformerConfig(RobertaConfig):
     r"""
-        This is the configuration class to store the configuration of a :class:`~transformers.LongformerModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.LongformerModel`.
         It is used to instantiate an Longformer model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the RoBERTa `roberta-base <https://huggingface.co/roberta-base>`__ architecture with a sequence length 4,096.
@@ -59,7 +57,12 @@ class LongformerConfig(RobertaConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "longformer"
 
     def __init__(self, attention_window: Union[List[int], int] = 512, sep_token_id: int = 2, **kwargs):

src/transformers/configuration_marian.py

@@ -18,9 +18,10 @@ from .configuration_bart import BartConfig
 
 
 PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "Helsinki-NLP/opus-mt-en-de": "https://s3.amazonaws.com/models.huggingface.co/bert/Helsinki-NLP/opus-mt-en-de/config.json",
+    "marian-en-de": "https://s3.amazonaws.com/models.huggingface.co/bert/Helsinki-NLP/opus-mt-en-de/config.json",
 }
 
 
 class MarianConfig(BartConfig):
     model_type = "marian"
+    pretrained_config_archive_map = PRETRAINED_CONFIG_ARCHIVE_MAP

src/transformers/configuration_openai.py

@@ -30,7 +30,7 @@ OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 class OpenAIGPTConfig(PretrainedConfig):
     """
-        This is the configuration class to store the configuration of a :class:`~transformers.OpenAIGPTModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.OpenAIGPTModel`.
         It is used to instantiate an GPT model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the `GPT <https://huggingface.co/openai-gpt>`__ architecture from OpenAI.
@@ -108,8 +108,13 @@ class OpenAIGPTConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "openai-gpt"
 
     def __init__(

src/transformers/configuration_reformer.py

@@ -135,7 +135,12 @@ class ReformerConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = REFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "reformer"
 
     def __init__(

src/transformers/configuration_roberta.py

@@ -35,7 +35,7 @@ ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 class RobertaConfig(BertConfig):
     r"""
-        This is the configuration class to store the configuration of a :class:`~transformers.RobertaModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.RobertaModel`.
         It is used to instantiate an RoBERTa model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the BERT `bert-base-uncased <https://huggingface.co/bert-base-uncased>`__ architecture.
@@ -59,7 +59,12 @@ class RobertaConfig(BertConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
+    pretrained_config_archive_map = ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "roberta"
 
     def __init__(self, pad_token_id=1, bos_token_id=0, eos_token_id=2, **kwargs):

src/transformers/configuration_t5.py

@@ -59,6 +59,7 @@ class T5Config(PretrainedConfig):
             initializer_factor: A factor for initializing all weight matrices (should be kept to 1.0, used for initialization testing).
             layer_norm_eps: The epsilon used by LayerNorm.
     """
+    pretrained_config_archive_map = T5_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "t5"
 
     def __init__(

src/transformers/configuration_transfo_xl.py

@@ -30,7 +30,7 @@ TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
 
 class TransfoXLConfig(PretrainedConfig):
     """
-        This is the configuration class to store the configuration of a :class:`~transformers.TransfoXLModel`.
+        This is the configuration class to store the configuration of an :class:`~transformers.TransfoXLModel`.
         It is used to instantiate a Transformer XL model according to the specified arguments, defining the model
         architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
         the `Transformer XL <https://huggingface.co/transfo-xl-wt103>`__ architecture.
@@ -110,8 +110,13 @@ class TransfoXLConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "transfo-xl"
 
     def __init__(

src/transformers/configuration_utils.py

@@ -20,7 +20,7 @@ import copy
 import json
 import logging
 import os
-from typing import Dict, Tuple
+from typing import Dict, Optional, Tuple
 
 from .file_utils import CONFIG_NAME, cached_path, hf_bucket_url, is_remote_url
 
@@ -37,6 +37,7 @@ class PretrainedConfig(object):
             It only affects the model's configuration.
 
         Class attributes (overridden by derived classes):
+            - ``pretrained_config_archive_map``: a python ``dict`` with `shortcut names` (string) as keys and `url` (string) of associated pretrained model configurations as values.
             - ``model_type``: a string that identifies the model type, that we serialize into the JSON file, and that we use to recreate the correct object in :class:`~transformers.AutoConfig`.
 
         Args:
@@ -51,6 +52,7 @@ class PretrainedConfig(object):
             torchscript (:obj:`bool`, `optional`, defaults to :obj:`False`):
                 Is the model used with Torchscript (for PyTorch models).
     """
+    pretrained_config_archive_map: Dict[str, str] = {}
     model_type: str = ""
 
     def __init__(self, **kwargs):
@@ -87,7 +89,6 @@ 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:
@@ -202,7 +203,9 @@ class PretrainedConfig(object):
         return cls.from_dict(config_dict, **kwargs)
 
     @classmethod
-    def get_config_dict(cls, pretrained_model_name_or_path: str, **kwargs) -> Tuple[Dict, Dict]:
+    def get_config_dict(
+        cls, pretrained_model_name_or_path: str, pretrained_config_archive_map: Optional[Dict] = None, **kwargs
+    ) -> Tuple[Dict, Dict]:
         """
         From a `pretrained_model_name_or_path`, resolve to a dictionary of parameters, to be used
         for instantiating a Config using `from_dict`.
@@ -210,6 +213,8 @@ class PretrainedConfig(object):
         Parameters:
             pretrained_model_name_or_path (:obj:`string`):
                 The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.
+            pretrained_config_archive_map: (:obj:`Dict[str, str]`, `optional`) Dict:
+                A map of `shortcut names` to `url`. By default, will use the current class attribute.
 
         Returns:
             :obj:`Tuple[Dict, Dict]`: The dictionary that will be used to instantiate the configuration object.
@@ -221,7 +226,12 @@ class PretrainedConfig(object):
         proxies = kwargs.pop("proxies", None)
         local_files_only = kwargs.pop("local_files_only", False)
 
-        if os.path.isdir(pretrained_model_name_or_path):
+        if pretrained_config_archive_map is None:
+            pretrained_config_archive_map = cls.pretrained_config_archive_map
+
+        if pretrained_model_name_or_path in pretrained_config_archive_map:
+            config_file = pretrained_config_archive_map[pretrained_model_name_or_path]
+        elif os.path.isdir(pretrained_model_name_or_path):
             config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
         elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
             config_file = pretrained_model_name_or_path
@@ -244,11 +254,21 @@ class PretrainedConfig(object):
             config_dict = cls._dict_from_json_file(resolved_config_file)
 
         except EnvironmentError:
-            msg = (
-                f"Can't load config for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
-                f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
-                f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a {CONFIG_NAME} file\n\n"
-            )
+            if pretrained_model_name_or_path in pretrained_config_archive_map:
+                msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
+                    config_file
+                )
+            else:
+                msg = (
+                    "Can't load '{}'. Make sure that:\n\n"
+                    "- '{}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                    "- or '{}' is the correct path to a directory containing a '{}' file\n\n".format(
+                        pretrained_model_name_or_path,
+                        pretrained_model_name_or_path,
+                        pretrained_model_name_or_path,
+                        CONFIG_NAME,
+                    )
+                )
             raise EnvironmentError(msg)
 
         except json.JSONDecodeError:

src/transformers/configuration_xlm.py

@@ -152,8 +152,13 @@ class XLMConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "xlm"
 
     def __init__(

src/transformers/configuration_xlm_roberta.py

@@ -39,4 +39,5 @@ class XLMRobertaConfig(RobertaConfig):
     superclass for the appropriate documentation alongside usage examples.
     """
 
+    pretrained_config_archive_map = XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "xlm-roberta"

src/transformers/configuration_xlnet.py

@@ -122,8 +122,13 @@ class XLNetConfig(PretrainedConfig):
 
             # Accessing the model configuration
             configuration = model.config
+
+        Attributes:
+            pretrained_config_archive_map (Dict[str, str]):
+                A dictionary containing all the available pre-trained checkpoints.
     """
 
+    pretrained_config_archive_map = XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
     model_type = "xlnet"
 
     def __init__(

src/transformers/convert_bart_original_pytorch_checkpoint_to_pytorch.py

@@ -124,7 +124,7 @@ def convert_bart_checkpoint(checkpoint_path, pytorch_dump_folder_path, hf_checkp
         remove_ignore_keys_(state_dict)
         state_dict["shared.weight"] = state_dict["decoder.embed_tokens.weight"]
         fairseq_output = bart.extract_features(tokens)
-        if hf_checkpoint_name == "facebook/bart-large":
+        if hf_checkpoint_name == "bart-large":
             model = BartModel(config).eval()
             model.load_state_dict(state_dict)
             new_model_outputs = model(tokens).model[0]

src/transformers/convert_graph_to_onnx.py

@@ -1,4 +1,5 @@
 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
@@ -37,17 +38,14 @@ 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)
 
-    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
+    for arg_pos, arg_name in model_args_pos:
+        model_args[arg_pos] = tokens[arg_name]
 
-    return ordered_input_names, tuple(model_args)
+    model_args = tuple(model_args)  # Need to be ordered
+    return tuple(takewhile(lambda arg: arg is not None, model_args))
 
 
 def infer_shapes(nlp: Pipeline, framework: str) -> Tuple[List[str], List[str], Dict, BatchEncoding]:
@@ -119,13 +117,13 @@ 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")
-        ordered_input_names, model_args = ensure_valid_input(nlp.model, tokens, input_names)
+        model_args = ensure_valid_input(nlp.model, tokens, input_names)
 
         export(
             nlp.model,
             model_args,
             f=output,
-            input_names=ordered_input_names,
+            input_names=input_names,
             output_names=output_names,
             dynamic_axes=dynamic_axes,
             do_constant_folding=True,

src/transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -32,7 +32,6 @@ from transformers import (
     ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
     TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
-    WEIGHTS_NAME,
     XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
     XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -71,7 +70,6 @@ from transformers import (
     XLMRobertaConfig,
     XLNetConfig,
     cached_path,
-    hf_bucket_url,
     is_torch_available,
     load_pytorch_checkpoint_in_tf2_model,
 )
@@ -84,103 +82,261 @@ if is_torch_available():
         BertForPreTraining,
         BertForQuestionAnswering,
         BertForSequenceClassification,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         GPT2LMHeadModel,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
         XLNetLMHeadModel,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
         XLMWithLMHeadModel,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
         XLMRobertaForMaskedLM,
         TransfoXLLMHeadModel,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
         OpenAIGPTLMHeadModel,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
         RobertaForMaskedLM,
         RobertaForSequenceClassification,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
         CamembertForMaskedLM,
+        CamembertForSequenceClassification,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         FlaubertWithLMHeadModel,
         DistilBertForMaskedLM,
         DistilBertForQuestionAnswering,
+        DistilBertForSequenceClassification,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         CTRLLMHeadModel,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
         AlbertForPreTraining,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         T5ForConditionalGeneration,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
         ElectraForPreTraining,
+        ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
+    )
+else:
+    (
+        BertForPreTraining,
+        BertForQuestionAnswering,
+        BertForSequenceClassification,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        GPT2LMHeadModel,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLNetLMHeadModel,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLMWithLMHeadModel,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLMRobertaForMaskedLM,
+        TransfoXLLMHeadModel,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        OpenAIGPTLMHeadModel,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        RobertaForMaskedLM,
+        RobertaForSequenceClassification,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CamembertForMaskedLM,
+        CamembertForSequenceClassification,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        FlaubertWithLMHeadModel,
+        DistilBertForMaskedLM,
+        DistilBertForSequenceClassification,
+        DistilBertForQuestionAnswering,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CTRLLMHeadModel,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        AlbertForPreTraining,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        T5ForConditionalGeneration,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ElectraForPreTraining,
+        ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
+    ) = (
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
     )
 
 
 logging.basicConfig(level=logging.INFO)
 
 MODEL_CLASSES = {
-    "bert": (BertConfig, TFBertForPreTraining, BertForPreTraining, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,),
+    "bert": (
+        BertConfig,
+        TFBertForPreTraining,
+        BertForPreTraining,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
     "bert-large-uncased-whole-word-masking-finetuned-squad": (
         BertConfig,
         TFBertForQuestionAnswering,
         BertForQuestionAnswering,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "bert-large-cased-whole-word-masking-finetuned-squad": (
         BertConfig,
         TFBertForQuestionAnswering,
         BertForQuestionAnswering,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "bert-base-cased-finetuned-mrpc": (
         BertConfig,
         TFBertForSequenceClassification,
         BertForSequenceClassification,
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
-    "gpt2": (GPT2Config, TFGPT2LMHeadModel, GPT2LMHeadModel, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,),
-    "xlnet": (XLNetConfig, TFXLNetLMHeadModel, XLNetLMHeadModel, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,),
-    "xlm": (XLMConfig, TFXLMWithLMHeadModel, XLMWithLMHeadModel, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,),
+    "gpt2": (
+        GPT2Config,
+        TFGPT2LMHeadModel,
+        GPT2LMHeadModel,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "xlnet": (
+        XLNetConfig,
+        TFXLNetLMHeadModel,
+        XLNetLMHeadModel,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "xlm": (
+        XLMConfig,
+        TFXLMWithLMHeadModel,
+        XLMWithLMHeadModel,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
     "xlm-roberta": (
         XLMRobertaConfig,
         TFXLMRobertaForMaskedLM,
         XLMRobertaForMaskedLM,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
         XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "transfo-xl": (
         TransfoXLConfig,
         TFTransfoXLLMHeadModel,
         TransfoXLLMHeadModel,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
         TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "openai-gpt": (
         OpenAIGPTConfig,
         TFOpenAIGPTLMHeadModel,
         OpenAIGPTLMHeadModel,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
         OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
-    "roberta": (RobertaConfig, TFRobertaForMaskedLM, RobertaForMaskedLM, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,),
+    "roberta": (
+        RobertaConfig,
+        TFRobertaForMaskedLM,
+        RobertaForMaskedLM,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
     "roberta-large-mnli": (
         RobertaConfig,
         TFRobertaForSequenceClassification,
         RobertaForSequenceClassification,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
         ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "camembert": (
         CamembertConfig,
         TFCamembertForMaskedLM,
         CamembertForMaskedLM,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "flaubert": (
         FlaubertConfig,
         TFFlaubertWithLMHeadModel,
         FlaubertWithLMHeadModel,
+        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "distilbert": (
         DistilBertConfig,
         TFDistilBertForMaskedLM,
         DistilBertForMaskedLM,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
     "distilbert-base-distilled-squad": (
         DistilBertConfig,
         TFDistilBertForQuestionAnswering,
         DistilBertForQuestionAnswering,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
         DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
-    "ctrl": (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,),
-    "albert": (AlbertConfig, TFAlbertForPreTraining, AlbertForPreTraining, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,),
-    "t5": (T5Config, TFT5ForConditionalGeneration, T5ForConditionalGeneration, T5_PRETRAINED_CONFIG_ARCHIVE_MAP,),
-    "electra": (ElectraConfig, TFElectraForPreTraining, ElectraForPreTraining, ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,),
+    "ctrl": (
+        CTRLConfig,
+        TFCTRLLMHeadModel,
+        CTRLLMHeadModel,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "albert": (
+        AlbertConfig,
+        TFAlbertForPreTraining,
+        AlbertForPreTraining,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "t5": (
+        T5Config,
+        TFT5ForConditionalGeneration,
+        T5ForConditionalGeneration,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "electra": (
+        ElectraConfig,
+        TFElectraForPreTraining,
+        ElectraForPreTraining,
+        ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
 }
 
 
@@ -190,7 +346,7 @@ def convert_pt_checkpoint_to_tf(
     if model_type not in MODEL_CLASSES:
         raise ValueError("Unrecognized model type, should be one of {}.".format(list(MODEL_CLASSES.keys())))
 
-    config_class, model_class, pt_model_class, aws_config_map = MODEL_CLASSES[model_type]
+    config_class, model_class, pt_model_class, aws_model_maps, aws_config_map = MODEL_CLASSES[model_type]
 
     # Initialise TF model
     if config_file in aws_config_map:
@@ -202,9 +358,10 @@ def convert_pt_checkpoint_to_tf(
     tf_model = model_class(config)
 
     # Load weights from tf checkpoint
-    if pytorch_checkpoint_path in aws_config_map.keys():
-        pytorch_checkpoint_url = hf_bucket_url(pytorch_checkpoint_path, filename=WEIGHTS_NAME)
-        pytorch_checkpoint_path = cached_path(pytorch_checkpoint_url, force_download=not use_cached_models)
+    if pytorch_checkpoint_path in aws_model_maps:
+        pytorch_checkpoint_path = cached_path(
+            aws_model_maps[pytorch_checkpoint_path], force_download=not use_cached_models
+        )
     # Load PyTorch checkpoint in tf2 model:
     tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
 

src/transformers/data/datasets/glue.py

@@ -86,15 +86,6 @@ 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.
@@ -109,7 +100,14 @@ 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:
@@ -139,4 +137,4 @@ class GlueDataset(Dataset):
         return self.features[i]
 
     def get_labels(self):
-        return self.label_list
+        return self.processor.get_labels()

src/transformers/data/processors/glue.py

@@ -332,12 +332,11 @@ 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[text_index]
+            text_a = line[0]
             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/hf_argparser.py

@@ -138,9 +138,6 @@ 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

@@ -31,17 +31,16 @@ from .modeling_utils import PreTrainedModel
 logger = logging.getLogger(__name__)
 
 
-ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "albert-base-v1",
-    "albert-large-v1",
-    "albert-xlarge-v1",
-    "albert-xxlarge-v1",
-    "albert-base-v2",
-    "albert-large-v2",
-    "albert-xlarge-v2",
-    "albert-xxlarge-v2",
-    # See all ALBERT models at https://huggingface.co/models?filter=albert
-]
+ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "albert-base-v1": "https://cdn.huggingface.co/albert-base-v1-pytorch_model.bin",
+    "albert-large-v1": "https://cdn.huggingface.co/albert-large-v1-pytorch_model.bin",
+    "albert-xlarge-v1": "https://cdn.huggingface.co/albert-xlarge-v1-pytorch_model.bin",
+    "albert-xxlarge-v1": "https://cdn.huggingface.co/albert-xxlarge-v1-pytorch_model.bin",
+    "albert-base-v2": "https://cdn.huggingface.co/albert-base-v2-pytorch_model.bin",
+    "albert-large-v2": "https://cdn.huggingface.co/albert-large-v2-pytorch_model.bin",
+    "albert-xlarge-v2": "https://cdn.huggingface.co/albert-xlarge-v2-pytorch_model.bin",
+    "albert-xxlarge-v2": "https://cdn.huggingface.co/albert-xxlarge-v2-pytorch_model.bin",
+}
 
 
 def load_tf_weights_in_albert(model, config, tf_checkpoint_path):
@@ -366,6 +365,7 @@ class AlbertPreTrainedModel(PreTrainedModel):
     """
 
     config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "albert"
 
     def _init_weights(self, module):
@@ -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**.
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
@@ -439,6 +439,7 @@ ALBERT_INPUTS_DOCSTRING = r"""
 class AlbertModel(AlbertPreTrainedModel):
 
     config_class = AlbertConfig
+    pretrained_model_archive_map = ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_albert
     base_model_prefix = "albert"
 

src/transformers/modeling_auto.py

@@ -43,6 +43,7 @@ from .configuration_auto import (
 from .configuration_marian import MarianConfig
 from .configuration_utils import PretrainedConfig
 from .modeling_albert import (
+    ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     AlbertForMaskedLM,
     AlbertForPreTraining,
     AlbertForQuestionAnswering,
@@ -50,8 +51,14 @@ from .modeling_albert import (
     AlbertForTokenClassification,
     AlbertModel,
 )
-from .modeling_bart import BartForConditionalGeneration, BartForSequenceClassification, BartModel
+from .modeling_bart import (
+    BART_PRETRAINED_MODEL_ARCHIVE_MAP,
+    BartForConditionalGeneration,
+    BartForSequenceClassification,
+    BartModel,
+)
 from .modeling_bert import (
+    BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     BertForMaskedLM,
     BertForMultipleChoice,
     BertForPreTraining,
@@ -61,14 +68,16 @@ from .modeling_bert import (
     BertModel,
 )
 from .modeling_camembert import (
+    CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     CamembertForMaskedLM,
     CamembertForMultipleChoice,
     CamembertForSequenceClassification,
     CamembertForTokenClassification,
     CamembertModel,
 )
-from .modeling_ctrl import CTRLLMHeadModel, CTRLModel
+from .modeling_ctrl import CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRLLMHeadModel, CTRLModel
 from .modeling_distilbert import (
+    DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     DistilBertForMaskedLM,
     DistilBertForQuestionAnswering,
     DistilBertForSequenceClassification,
@@ -76,6 +85,7 @@ from .modeling_distilbert import (
     DistilBertModel,
 )
 from .modeling_electra import (
+    ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
     ElectraForMaskedLM,
     ElectraForPreTraining,
     ElectraForSequenceClassification,
@@ -84,24 +94,24 @@ from .modeling_electra import (
 )
 from .modeling_encoder_decoder import EncoderDecoderModel
 from .modeling_flaubert import (
+    FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     FlaubertForQuestionAnsweringSimple,
     FlaubertForSequenceClassification,
     FlaubertModel,
     FlaubertWithLMHeadModel,
 )
-from .modeling_gpt2 import GPT2LMHeadModel, GPT2Model
+from .modeling_gpt2 import GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, GPT2LMHeadModel, GPT2Model
 from .modeling_longformer import (
+    LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
     LongformerForMaskedLM,
-    LongformerForMultipleChoice,
     LongformerForQuestionAnswering,
-    LongformerForSequenceClassification,
-    LongformerForTokenClassification,
     LongformerModel,
 )
 from .modeling_marian import MarianMTModel
-from .modeling_openai import OpenAIGPTLMHeadModel, OpenAIGPTModel
+from .modeling_openai import OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, OpenAIGPTLMHeadModel, OpenAIGPTModel
 from .modeling_reformer import ReformerModel, ReformerModelWithLMHead
 from .modeling_roberta import (
+    ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     RobertaForMaskedLM,
     RobertaForMultipleChoice,
     RobertaForQuestionAnswering,
@@ -109,9 +119,10 @@ from .modeling_roberta import (
     RobertaForTokenClassification,
     RobertaModel,
 )
-from .modeling_t5 import T5ForConditionalGeneration, T5Model
-from .modeling_transfo_xl import TransfoXLLMHeadModel, TransfoXLModel
+from .modeling_t5 import T5_PRETRAINED_MODEL_ARCHIVE_MAP, T5ForConditionalGeneration, T5Model
+from .modeling_transfo_xl import TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP, TransfoXLLMHeadModel, TransfoXLModel
 from .modeling_xlm import (
+    XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
     XLMForQuestionAnsweringSimple,
     XLMForSequenceClassification,
     XLMForTokenClassification,
@@ -119,6 +130,7 @@ from .modeling_xlm import (
     XLMWithLMHeadModel,
 )
 from .modeling_xlm_roberta import (
+    XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     XLMRobertaForMaskedLM,
     XLMRobertaForMultipleChoice,
     XLMRobertaForSequenceClassification,
@@ -126,6 +138,7 @@ from .modeling_xlm_roberta import (
     XLMRobertaModel,
 )
 from .modeling_xlnet import (
+    XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
     XLNetForMultipleChoice,
     XLNetForQuestionAnsweringSimple,
     XLNetForSequenceClassification,
@@ -138,6 +151,30 @@ from .modeling_xlnet import (
 logger = logging.getLogger(__name__)
 
 
+ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict(
+    (key, value)
+    for pretrained_map in [
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        BART_PRETRAINED_MODEL_ARCHIVE_MAP,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP,
+    ]
+    for key, value, in pretrained_map.items()
+)
+
 MODEL_MAPPING = OrderedDict(
     [
         (T5Config, T5Model),
@@ -215,7 +252,6 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForSequenceClassification),
         (XLMRobertaConfig, XLMRobertaForSequenceClassification),
         (BartConfig, BartForSequenceClassification),
-        (LongformerConfig, LongformerForSequenceClassification),
         (RobertaConfig, RobertaForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
@@ -244,7 +280,6 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (CamembertConfig, CamembertForTokenClassification),
         (XLMConfig, XLMForTokenClassification),
         (XLMRobertaConfig, XLMRobertaForTokenClassification),
-        (LongformerConfig, LongformerForTokenClassification),
         (RobertaConfig, RobertaForTokenClassification),
         (BertConfig, BertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
@@ -258,7 +293,6 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
     [
         (CamembertConfig, CamembertForMultipleChoice),
         (XLMRobertaConfig, XLMRobertaForMultipleChoice),
-        (LongformerConfig, LongformerForMultipleChoice),
         (RobertaConfig, RobertaForMultipleChoice),
         (BertConfig, BertForMultipleChoice),
         (XLNetConfig, XLNetForMultipleChoice),
@@ -288,11 +322,6 @@ 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:
@@ -332,26 +361,29 @@ class AutoModel:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: :class:`~transformers.T5Model` (T5 model)
-            - `distilbert`: :class:`~transformers.DistilBertModel` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertModel` (ALBERT model)
-            - `camembert`: :class:`~transformers.CamembertModel` (CamemBERT model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaModel` (XLM-RoBERTa model)
-            - `longformer` :class:`~transformers.LongformerModel` (Longformer model)
-            - `roberta`: :class:`~transformers.RobertaModel` (RoBERTa model)
-            - `bert`: :class:`~transformers.BertModel` (Bert model)
-            - `openai-gpt`: :class:`~transformers.OpenAIGPTModel` (OpenAI GPT model)
-            - `gpt2`: :class:`~transformers.GPT2Model` (OpenAI GPT-2 model)
-            - `transfo-xl`: :class:`~transformers.TransfoXLModel` (Transformer-XL model)
-            - `xlnet`: :class:`~transformers.XLNetModel` (XLNet model)
-            - `xlm`: :class:`~transformers.XLMModel` (XLM model)
-            - `ctrl`: :class:`~transformers.CTRLModel` (Salesforce CTRL  model)
-            - `flaubert`: :class:`~transformers.FlaubertModel` (Flaubert  model)
-            - `electra`: :class:`~transformers.ElectraModel` (Electra  model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
 
-        The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
-        To train the model, you should first set it back in training mode with `model.train()`
+        The base model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: :class:`~transformers.T5Model` (T5 model)
+            - contains `distilbert`: :class:`~transformers.DistilBertModel` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertModel` (ALBERT model)
+            - contains `camembert`: :class:`~transformers.CamembertModel` (CamemBERT model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaModel` (XLM-RoBERTa model)
+            - contains `longformer` :class:`~transformers.LongformerModel` (Longformer model)
+            - contains `roberta`: :class:`~transformers.RobertaModel` (RoBERTa model)
+            - contains `bert`: :class:`~transformers.BertModel` (Bert model)
+            - contains `openai-gpt`: :class:`~transformers.OpenAIGPTModel` (OpenAI GPT model)
+            - contains `gpt2`: :class:`~transformers.GPT2Model` (OpenAI GPT-2 model)
+            - contains `transfo-xl`: :class:`~transformers.TransfoXLModel` (Transformer-XL model)
+            - contains `xlnet`: :class:`~transformers.XLNetModel` (XLNet model)
+            - contains `xlm`: :class:`~transformers.XLMModel` (XLM model)
+            - contains `ctrl`: :class:`~transformers.CTRLModel` (Salesforce CTRL  model)
+            - contains `flaubert`: :class:`~transformers.FlaubertModel` (Flaubert  model)
+            - contains `electra`: :class:`~transformers.ElectraModel` (Electra  model)
+
+            The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
+            To train the model, you should first set it back in training mode with `model.train()`
 
         Args:
             pretrained_model_name_or_path: either:
@@ -442,11 +474,6 @@ 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:
@@ -485,23 +512,26 @@ class AutoModelForPreTraining:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: :class:`~transformers.T5ModelWithLMHead` (T5 model)
-            - `distilbert`: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
-            - `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
-            - `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
-            - `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
-            - `bert`: :class:`~transformers.BertForPreTraining` (Bert model)
-            - `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
-            - `gpt2`: :class:`~transformers.GPT2LMHeadModel` (OpenAI GPT-2 model)
-            - `transfo-xl`: :class:`~transformers.TransfoXLLMHeadModel` (Transformer-XL model)
-            - `xlnet`: :class:`~transformers.XLNetLMHeadModel` (XLNet model)
-            - `xlm`: :class:`~transformers.XLMWithLMHeadModel` (XLM model)
-            - `ctrl`: :class:`~transformers.CTRLLMHeadModel` (Salesforce CTRL model)
-            - `flaubert`: :class:`~transformers.FlaubertWithLMHeadModel` (Flaubert model)
-            - `electra`: :class:`~transformers.ElectraForPreTraining` (Electra model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: :class:`~transformers.T5ModelWithLMHead` (T5 model)
+            - contains `distilbert`: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
+            - contains `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
+            - contains `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
+            - contains `bert`: :class:`~transformers.BertForPreTraining` (Bert model)
+            - contains `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
+            - contains `gpt2`: :class:`~transformers.GPT2LMHeadModel` (OpenAI GPT-2 model)
+            - contains `transfo-xl`: :class:`~transformers.TransfoXLLMHeadModel` (Transformer-XL model)
+            - contains `xlnet`: :class:`~transformers.XLNetLMHeadModel` (XLNet model)
+            - contains `xlm`: :class:`~transformers.XLMWithLMHeadModel` (XLM model)
+            - contains `ctrl`: :class:`~transformers.CTRLLMHeadModel` (Salesforce CTRL model)
+            - contains `flaubert`: :class:`~transformers.FlaubertWithLMHeadModel` (Flaubert model)
+            - contains `electra`: :class:`~transformers.ElectraForPreTraining` (Electra model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -589,11 +619,6 @@ 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:
@@ -633,23 +658,26 @@ class AutoModelWithLMHead:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: :class:`~transformers.T5ModelWithLMHead` (T5 model)
-            - `distilbert`: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
-            - `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
-            - `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
-            - `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
-            - `bert`: :class:`~transformers.BertForMaskedLM` (Bert model)
-            - `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
-            - `gpt2`: :class:`~transformers.GPT2LMHeadModel` (OpenAI GPT-2 model)
-            - `transfo-xl`: :class:`~transformers.TransfoXLLMHeadModel` (Transformer-XL model)
-            - `xlnet`: :class:`~transformers.XLNetLMHeadModel` (XLNet model)
-            - `xlm`: :class:`~transformers.XLMWithLMHeadModel` (XLM model)
-            - `ctrl`: :class:`~transformers.CTRLLMHeadModel` (Salesforce CTRL model)
-            - `flaubert`: :class:`~transformers.FlaubertWithLMHeadModel` (Flaubert model)
-            - `electra`: :class:`~transformers.ElectraForMaskedLM` (Electra model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: :class:`~transformers.T5ModelWithLMHead` (T5 model)
+            - contains `distilbert`: :class:`~transformers.DistilBertForMaskedLM` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertForMaskedLM` (ALBERT model)
+            - contains `camembert`: :class:`~transformers.CamembertForMaskedLM` (CamemBERT model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForMaskedLM` (XLM-RoBERTa model)
+            - contains `longformer`: :class:`~transformers.LongformerForMaskedLM` (Longformer model)
+            - contains `roberta`: :class:`~transformers.RobertaForMaskedLM` (RoBERTa model)
+            - contains `bert`: :class:`~transformers.BertForMaskedLM` (Bert model)
+            - contains `openai-gpt`: :class:`~transformers.OpenAIGPTLMHeadModel` (OpenAI GPT model)
+            - contains `gpt2`: :class:`~transformers.GPT2LMHeadModel` (OpenAI GPT-2 model)
+            - contains `transfo-xl`: :class:`~transformers.TransfoXLLMHeadModel` (Transformer-XL model)
+            - contains `xlnet`: :class:`~transformers.XLNetLMHeadModel` (XLNet model)
+            - contains `xlm`: :class:`~transformers.XLMWithLMHeadModel` (XLM model)
+            - contains `ctrl`: :class:`~transformers.CTRLLMHeadModel` (Salesforce CTRL model)
+            - contains `flaubert`: :class:`~transformers.FlaubertWithLMHeadModel` (Flaubert model)
+            - contains `electra`: :class:`~transformers.ElectraForMaskedLM` (Electra model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -737,11 +765,6 @@ 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:
@@ -781,15 +804,18 @@ class AutoModelForSequenceClassification:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: :class:`~transformers.DistilBertForSequenceClassification` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertForSequenceClassification` (ALBERT model)
-            - `camembert`: :class:`~transformers.CamembertForSequenceClassification` (CamemBERT model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaForSequenceClassification` (XLM-RoBERTa model)
-            - `roberta`: :class:`~transformers.RobertaForSequenceClassification` (RoBERTa model)
-            - `bert`: :class:`~transformers.BertForSequenceClassification` (Bert model)
-            - `xlnet`: :class:`~transformers.XLNetForSequenceClassification` (XLNet model)
-            - `flaubert`: :class:`~transformers.FlaubertForSequenceClassification` (Flaubert model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: :class:`~transformers.DistilBertForSequenceClassification` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertForSequenceClassification` (ALBERT model)
+            - contains `camembert`: :class:`~transformers.CamembertForSequenceClassification` (CamemBERT model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForSequenceClassification` (XLM-RoBERTa model)
+            - contains `roberta`: :class:`~transformers.RobertaForSequenceClassification` (RoBERTa model)
+            - contains `bert`: :class:`~transformers.BertForSequenceClassification` (Bert model)
+            - contains `xlnet`: :class:`~transformers.XLNetForSequenceClassification` (XLNet model)
+            - contains `flaubert`: :class:`~transformers.FlaubertForSequenceClassification` (Flaubert model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -886,11 +912,6 @@ 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:
@@ -927,13 +948,16 @@ class AutoModelForQuestionAnswering:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: :class:`~transformers.DistilBertForQuestionAnswering` (DistilBERT model)
-            - `albert`: :class:`~transformers.AlbertForQuestionAnswering` (ALBERT model)
-            - `bert`: :class:`~transformers.BertForQuestionAnswering` (Bert model)
-            - `xlnet`: :class:`~transformers.XLNetForQuestionAnswering` (XLNet model)
-            - `xlm`: :class:`~transformers.XLMForQuestionAnswering` (XLM model)
-            - `flaubert`: :class:`~transformers.FlaubertForQuestionAnswering` (XLM model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: :class:`~transformers.DistilBertForQuestionAnswering` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.AlbertForQuestionAnswering` (ALBERT model)
+            - contains `bert`: :class:`~transformers.BertForQuestionAnswering` (Bert model)
+            - contains `xlnet`: :class:`~transformers.XLNetForQuestionAnswering` (XLNet model)
+            - contains `xlm`: :class:`~transformers.XLMForQuestionAnswering` (XLM model)
+            - contains `flaubert`: :class:`~transformers.FlaubertForQuestionAnswering` (XLM model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -1028,11 +1052,6 @@ 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:
@@ -1072,15 +1091,18 @@ class AutoModelForTokenClassification:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: :class:`~transformers.DistilBertForTokenClassification` (DistilBERT model)
-            - `xlm`: :class:`~transformers.XLMForTokenClassification` (XLM model)
-            - `xlm-roberta`: :class:`~transformers.XLMRobertaForTokenClassification` (XLM-RoBERTa?Para model)
-            - `camembert`: :class:`~transformers.CamembertForTokenClassification` (Camembert model)
-            - `bert`: :class:`~transformers.BertForTokenClassification` (Bert model)
-            - `xlnet`: :class:`~transformers.XLNetForTokenClassification` (XLNet model)
-            - `roberta`: :class:`~transformers.RobertaForTokenClassification` (Roberta model)
-            - `electra`: :class:`~transformers.ElectraForTokenClassification` (Electra model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: :class:`~transformers.DistilBertForTokenClassification` (DistilBERT model)
+            - contains `xlm`: :class:`~transformers.XLMForTokenClassification` (XLM model)
+            - contains `xlm-roberta`: :class:`~transformers.XLMRobertaForTokenClassification` (XLM-RoBERTa?Para model)
+            - contains `camembert`: :class:`~transformers.CamembertForTokenClassification` (Camembert model)
+            - contains `bert`: :class:`~transformers.BertForTokenClassification` (Bert model)
+            - contains `xlnet`: :class:`~transformers.XLNetForTokenClassification` (XLNet model)
+            - contains `roberta`: :class:`~transformers.RobertaForTokenClassification` (Roberta model)
+            - contains `electra`: :class:`~transformers.ElectraForTokenClassification` (Electra model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`

src/transformers/modeling_bart.py

@@ -32,15 +32,13 @@ from .modeling_utils import PreTrainedModel, create_position_ids_from_input_ids
 logger = logging.getLogger(__name__)
 
 
-BART_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "facebook/bart-large",
-    "facebook/bart-large-mnli",
-    "facebook/bart-large-cnn",
-    "facebook/bart-large-xsum",
-    "facebook/mbart-large-en-ro",
-    # See all BART models at https://huggingface.co/models?filter=bart
-]
-
+BART_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "bart-large": "https://cdn.huggingface.co/facebook/bart-large/pytorch_model.bin",
+    "bart-large-mnli": "https://cdn.huggingface.co/facebook/bart-large-mnli/pytorch_model.bin",
+    "bart-large-cnn": "https://cdn.huggingface.co/facebook/bart-large-cnn/pytorch_model.bin",
+    "bart-large-xsum": "https://cdn.huggingface.co/facebook/bart-large-xsum/pytorch_model.bin",
+    "mbart-large-en-ro": "https://cdn.huggingface.co/facebook/mbart-large-en-ro/pytorch_model.bin",
+}
 
 BART_START_DOCSTRING = r"""
 
@@ -120,6 +118,7 @@ def _prepare_bart_decoder_inputs(
 class PretrainedBartModel(PreTrainedModel):
     config_class = BartConfig
     base_model_prefix = "model"
+    pretrained_model_archive_map = BART_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def _init_weights(self, module):
         std = self.config.init_std
@@ -905,7 +904,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
         **unused
     ):
         r"""
-        lm_labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+        masked_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
@@ -914,7 +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 ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+        masked_lm_loss (`optional`, returned when ``masked_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

@@ -32,31 +32,30 @@ from .modeling_utils import PreTrainedModel, prune_linear_layer
 
 logger = logging.getLogger(__name__)
 
-BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "bert-base-uncased",
-    "bert-large-uncased",
-    "bert-base-cased",
-    "bert-large-cased",
-    "bert-base-multilingual-uncased",
-    "bert-base-multilingual-cased",
-    "bert-base-chinese",
-    "bert-base-german-cased",
-    "bert-large-uncased-whole-word-masking",
-    "bert-large-cased-whole-word-masking",
-    "bert-large-uncased-whole-word-masking-finetuned-squad",
-    "bert-large-cased-whole-word-masking-finetuned-squad",
-    "bert-base-cased-finetuned-mrpc",
-    "bert-base-german-dbmdz-cased",
-    "bert-base-german-dbmdz-uncased",
-    "cl-tohoku/bert-base-japanese",
-    "cl-tohoku/bert-base-japanese-whole-word-masking",
-    "cl-tohoku/bert-base-japanese-char",
-    "cl-tohoku/bert-base-japanese-char-whole-word-masking",
-    "TurkuNLP/bert-base-finnish-cased-v1",
-    "TurkuNLP/bert-base-finnish-uncased-v1",
-    "wietsedv/bert-base-dutch-cased",
-    # See all BERT models at https://huggingface.co/models?filter=bert
-]
+BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "bert-base-uncased": "https://cdn.huggingface.co/bert-base-uncased-pytorch_model.bin",
+    "bert-large-uncased": "https://cdn.huggingface.co/bert-large-uncased-pytorch_model.bin",
+    "bert-base-cased": "https://cdn.huggingface.co/bert-base-cased-pytorch_model.bin",
+    "bert-large-cased": "https://cdn.huggingface.co/bert-large-cased-pytorch_model.bin",
+    "bert-base-multilingual-uncased": "https://cdn.huggingface.co/bert-base-multilingual-uncased-pytorch_model.bin",
+    "bert-base-multilingual-cased": "https://cdn.huggingface.co/bert-base-multilingual-cased-pytorch_model.bin",
+    "bert-base-chinese": "https://cdn.huggingface.co/bert-base-chinese-pytorch_model.bin",
+    "bert-base-german-cased": "https://cdn.huggingface.co/bert-base-german-cased-pytorch_model.bin",
+    "bert-large-uncased-whole-word-masking": "https://cdn.huggingface.co/bert-large-uncased-whole-word-masking-pytorch_model.bin",
+    "bert-large-cased-whole-word-masking": "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-pytorch_model.bin",
+    "bert-large-uncased-whole-word-masking-finetuned-squad": "https://cdn.huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad-pytorch_model.bin",
+    "bert-large-cased-whole-word-masking-finetuned-squad": "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad-pytorch_model.bin",
+    "bert-base-cased-finetuned-mrpc": "https://cdn.huggingface.co/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
+    "bert-base-german-dbmdz-cased": "https://cdn.huggingface.co/bert-base-german-dbmdz-cased-pytorch_model.bin",
+    "bert-base-german-dbmdz-uncased": "https://cdn.huggingface.co/bert-base-german-dbmdz-uncased-pytorch_model.bin",
+    "bert-base-japanese": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese/pytorch_model.bin",
+    "bert-base-japanese-whole-word-masking": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-whole-word-masking/pytorch_model.bin",
+    "bert-base-japanese-char": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-char/pytorch_model.bin",
+    "bert-base-japanese-char-whole-word-masking": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-char-whole-word-masking/pytorch_model.bin",
+    "bert-base-finnish-cased-v1": "https://cdn.huggingface.co/TurkuNLP/bert-base-finnish-cased-v1/pytorch_model.bin",
+    "bert-base-finnish-uncased-v1": "https://cdn.huggingface.co/TurkuNLP/bert-base-finnish-uncased-v1/pytorch_model.bin",
+    "bert-base-dutch-cased": "https://cdn.huggingface.co/wietsedv/bert-base-dutch-cased/pytorch_model.bin",
+}
 
 
 def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
@@ -514,6 +513,7 @@ class BertPreTrainedModel(PreTrainedModel):
     """
 
     config_class = BertConfig
+    pretrained_model_archive_map = BERT_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_bert
     base_model_prefix = "bert"
 
@@ -543,7 +543,7 @@ BERT_START_DOCSTRING = r"""
 
 BERT_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
             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:`{0}`, `optional`, defaults to :obj:`None`):
+        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]``:
             ``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:`{0}`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `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:`{0}`, `optional`, defaults to :obj:`None`):
+        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.
             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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -992,7 +992,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -1036,12 +1036,11 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
 
-        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')
+        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]
 
-        loss, logits = model(**encoding, next_sentence_label=torch.LongTensor([1]))
-        assert logits[0, 0] < logits[0, 1] # next sentence was random
         """
 
         outputs = self.bert(
@@ -1082,7 +1081,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -1178,7 +1177,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -1192,7 +1191,7 @@ 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-1]`` where `num_choices` is the size of the second dimension
+            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:
@@ -1222,17 +1221,14 @@ 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"]
 
-        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 ;))
+        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)
 
-        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
+        loss, classification_scores = outputs[:2]
 
-        # the linear classifier still needs to be trained
-        loss, logits = outputs[:2]
         """
         num_choices = input_ids.shape[1]
 
@@ -1282,7 +1278,7 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -1379,7 +1375,7 @@ class BertForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,

src/transformers/modeling_camembert.py

@@ -31,12 +31,11 @@ from .modeling_roberta import (
 
 logger = logging.getLogger(__name__)
 
-CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "camembert-base",
-    "Musixmatch/umberto-commoncrawl-cased-v1",
-    "Musixmatch/umberto-wikipedia-uncased-v1",
-    # See all CamemBERT models at https://huggingface.co/models?filter=camembert
-]
+CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "camembert-base": "https://cdn.huggingface.co/camembert-base-pytorch_model.bin",
+    "umberto-commoncrawl-cased-v1": "https://cdn.huggingface.co/Musixmatch/umberto-commoncrawl-cased-v1/pytorch_model.bin",
+    "umberto-wikipedia-uncased-v1": "https://cdn.huggingface.co/Musixmatch/umberto-wikipedia-uncased-v1/pytorch_model.bin",
+}
 
 CAMEMBERT_START_DOCSTRING = r"""
 
@@ -63,6 +62,7 @@ class CamembertModel(RobertaModel):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
 
 @add_start_docstrings(
@@ -75,6 +75,7 @@ class CamembertForMaskedLM(RobertaForMaskedLM):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
 
 @add_start_docstrings(
@@ -89,6 +90,7 @@ class CamembertForSequenceClassification(RobertaForSequenceClassification):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
 
 @add_start_docstrings(
@@ -103,6 +105,7 @@ class CamembertForMultipleChoice(RobertaForMultipleChoice):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
 
 @add_start_docstrings(
@@ -117,6 +120,7 @@ class CamembertForTokenClassification(RobertaForTokenClassification):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
 
 @add_start_docstrings(
@@ -131,3 +135,4 @@ class CamembertForQuestionAnswering(RobertaForQuestionAnswering):
     """
 
     config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP

src/transformers/modeling_ctrl.py

@@ -30,10 +30,7 @@ from .modeling_utils import Conv1D, PreTrainedModel
 
 logger = logging.getLogger(__name__)
 
-CTRL_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "ctrl"
-    # See all CTRL models at https://huggingface.co/models?filter=ctrl
-]
+CTRL_PRETRAINED_MODEL_ARCHIVE_MAP = {"ctrl": "https://storage.googleapis.com/sf-ctrl/pytorch/seqlen256_v1.bin"}
 
 
 def angle_defn(pos, i, d_model_size):
@@ -181,6 +178,7 @@ class CTRLPreTrainedModel(PreTrainedModel):
     """
 
     config_class = CTRLConfig
+    pretrained_model_archive_map = CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "transformer"
 
     def _init_weights(self, module):
@@ -246,10 +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**.
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_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 `inputs_embeds` have to be input (see `past`).
+            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_distilbert.py

@@ -36,16 +36,15 @@ from .modeling_utils import PreTrainedModel, prune_linear_layer
 logger = logging.getLogger(__name__)
 
 
-DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "distilbert-base-uncased",
-    "distilbert-base-uncased-distilled-squad",
-    "distilbert-base-cased",
-    "distilbert-base-cased-distilled-squad",
-    "distilbert-base-german-cased",
-    "distilbert-base-multilingual-cased",
-    "distilbert-base-uncased-finetuned-sst-2-english",
-    # See all DistilBERT models at https://huggingface.co/models?filter=distilbert
-]
+DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "distilbert-base-uncased": "https://cdn.huggingface.co/distilbert-base-uncased-pytorch_model.bin",
+    "distilbert-base-uncased-distilled-squad": "https://cdn.huggingface.co/distilbert-base-uncased-distilled-squad-pytorch_model.bin",
+    "distilbert-base-cased": "https://cdn.huggingface.co/distilbert-base-cased-pytorch_model.bin",
+    "distilbert-base-cased-distilled-squad": "https://cdn.huggingface.co/distilbert-base-cased-distilled-squad-pytorch_model.bin",
+    "distilbert-base-german-cased": "https://cdn.huggingface.co/distilbert-base-german-cased-pytorch_model.bin",
+    "distilbert-base-multilingual-cased": "https://cdn.huggingface.co/distilbert-base-multilingual-cased-pytorch_model.bin",
+    "distilbert-base-uncased-finetuned-sst-2-english": "https://cdn.huggingface.co/distilbert-base-uncased-finetuned-sst-2-english-pytorch_model.bin",
+}
 
 
 # UTILS AND BUILDING BLOCKS OF THE ARCHITECTURE #
@@ -328,6 +327,7 @@ class DistilBertPreTrainedModel(PreTrainedModel):
     """
 
     config_class = DistilBertConfig
+    pretrained_model_archive_map = DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = None
     base_model_prefix = "distilbert"
 

src/transformers/modeling_electra.py

@@ -14,15 +14,14 @@ from .modeling_bert import BertEmbeddings, BertEncoder, BertLayerNorm, BertPreTr
 logger = logging.getLogger(__name__)
 
 
-ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "google/electra-small-generator",
-    "google/electra-base-generator",
-    "google/electra-large-generator",
-    "google/electra-small-discriminator",
-    "google/electra-base-discriminator",
-    "google/electra-large-discriminator",
-    # See all ELECTRA models at https://huggingface.co/models?filter=electra
-]
+ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "google/electra-small-generator": "https://cdn.huggingface.co/google/electra-small-generator/pytorch_model.bin",
+    "google/electra-base-generator": "https://cdn.huggingface.co/google/electra-base-generator/pytorch_model.bin",
+    "google/electra-large-generator": "https://cdn.huggingface.co/google/electra-large-generator/pytorch_model.bin",
+    "google/electra-small-discriminator": "https://cdn.huggingface.co/google/electra-small-discriminator/pytorch_model.bin",
+    "google/electra-base-discriminator": "https://cdn.huggingface.co/google/electra-base-discriminator/pytorch_model.bin",
+    "google/electra-large-discriminator": "https://cdn.huggingface.co/google/electra-large-discriminator/pytorch_model.bin",
+}
 
 
 def load_tf_weights_in_electra(model, config, tf_checkpoint_path, discriminator_or_generator="discriminator"):
@@ -161,6 +160,7 @@ class ElectraPreTrainedModel(BertPreTrainedModel):
     """
 
     config_class = ElectraConfig
+    pretrained_model_archive_map = ELECTRA_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_electra
     base_model_prefix = "electra"
 

src/transformers/modeling_encoder_decoder.py

@@ -35,7 +35,6 @@ 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,
@@ -159,26 +158,12 @@ 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
 
-        return cls(encoder=encoder, decoder=decoder)
+        model = cls(encoder=encoder, decoder=decoder)
+
+        return model
 
     def forward(
         self,

src/transformers/modeling_flaubert.py

@@ -35,13 +35,12 @@ from .modeling_xlm import (
 
 logger = logging.getLogger(__name__)
 
-FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "flaubert/flaubert_small_cased",
-    "flaubert/flaubert_base_uncased",
-    "flaubert/flaubert_base_cased",
-    "flaubert/flaubert_large_cased",
-    # See all Flaubert models at https://huggingface.co/models?filter=flaubert
-]
+FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "flaubert-small-cased": "https://cdn.huggingface.co/flaubert/flaubert_small_cased/pytorch_model.bin",
+    "flaubert-base-uncased": "https://cdn.huggingface.co/flaubert/flaubert_base_uncased/pytorch_model.bin",
+    "flaubert-base-cased": "https://cdn.huggingface.co/flaubert/flaubert_base_cased/pytorch_model.bin",
+    "flaubert-large-cased": "https://cdn.huggingface.co/flaubert/flaubert_large_cased/pytorch_model.bin",
+}
 
 
 FLAUBERT_START_DOCSTRING = r"""
@@ -96,7 +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**.
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
@@ -110,6 +109,7 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
 class FlaubertModel(XLMModel):
 
     config_class = FlaubertConfig
+    pretrained_model_archive_map = FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def __init__(self, config):  # , dico, is_encoder, with_output):
         super().__init__(config)
@@ -304,6 +304,7 @@ class FlaubertWithLMHeadModel(XLMWithLMHeadModel):
     """
 
     config_class = FlaubertConfig
+    pretrained_model_archive_map = FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def __init__(self, config):
         super().__init__(config)
@@ -323,6 +324,7 @@ class FlaubertForSequenceClassification(XLMForSequenceClassification):
     """
 
     config_class = FlaubertConfig
+    pretrained_model_archive_map = FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def __init__(self, config):
         super().__init__(config)
@@ -342,6 +344,7 @@ class FlaubertForQuestionAnsweringSimple(XLMForQuestionAnsweringSimple):
     """
 
     config_class = FlaubertConfig
+    pretrained_model_archive_map = FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def __init__(self, config):
         super().__init__(config)
@@ -361,6 +364,7 @@ class FlaubertForQuestionAnswering(XLMForQuestionAnswering):
     """
 
     config_class = FlaubertConfig
+    pretrained_model_archive_map = FLAUBERT_PRETRAINED_MODEL_ARCHIVE_MAP
 
     def __init__(self, config):
         super().__init__(config)

src/transformers/modeling_gpt2.py

@@ -31,14 +31,13 @@ from .modeling_utils import Conv1D, PreTrainedModel, SequenceSummary, prune_conv
 
 logger = logging.getLogger(__name__)
 
-GPT2_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "gpt2",
-    "gpt2-medium",
-    "gpt2-large",
-    "gpt2-xl",
-    "distilgpt2",
-    # See all GPT-2 models at https://huggingface.co/models?filter=gpt2
-]
+GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "gpt2": "https://cdn.huggingface.co/gpt2-pytorch_model.bin",
+    "gpt2-medium": "https://cdn.huggingface.co/gpt2-medium-pytorch_model.bin",
+    "gpt2-large": "https://cdn.huggingface.co/gpt2-large-pytorch_model.bin",
+    "gpt2-xl": "https://cdn.huggingface.co/gpt2-xl-pytorch_model.bin",
+    "distilgpt2": "https://cdn.huggingface.co/distilgpt2-pytorch_model.bin",
+}
 
 
 def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
@@ -252,6 +251,7 @@ class GPT2PreTrainedModel(PreTrainedModel):
     """
 
     config_class = GPT2Config
+    pretrained_model_archive_map = GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_gpt2
     base_model_prefix = "transformer"
 
@@ -323,10 +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**.
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_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 `inputs_embeds` have to be input (see `past`).
+            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`.
 """
@@ -554,7 +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 ``labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_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, MSELoss
+from torch.nn import CrossEntropyLoss
 from torch.nn import functional as F
 
 from .configuration_longformer import LongformerConfig
@@ -30,52 +30,11 @@ from .modeling_roberta import RobertaLMHead, RobertaModel
 
 logger = logging.getLogger(__name__)
 
-LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "allenai/longformer-base-4096",
-    "allenai/longformer-large-4096",
-    "allenai/longformer-large-4096-finetuned-triviaqa",
-    "allenai/longformer-base-4096-extra.pos.embd.only",
-    "allenai/longformer-large-4096-extra.pos.embd.only",
-    # See all Longformer models at https://huggingface.co/models?filter=longformer
-]
-
-
-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
+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",
+    "longformer-large-4096-finetuned-triviaqa": "https://s3.amazonaws.com/models.huggingface.co/bert/allenai/longformer-large-4096-finetuned-triviaqa/pytorch_model.bin",
+}
 
 
 class LongformerSelfAttention(nn.Module):
@@ -349,7 +308,9 @@ 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)).transpose(1, 2)
+            attn = torch.matmul(
+                selected_attn_probs.transpose(1, 2), selected_v.transpose(1, 2).type_as(selected_attn_probs)
+            ).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()
@@ -413,7 +374,7 @@ 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:
@@ -448,7 +409,7 @@ LONGFORMER_START_DOCSTRING = r"""
 
 LONGFORMER_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.LonmgformerTokenizer`.
@@ -456,30 +417,25 @@ 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:`{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.
+        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).
             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]``:
-            ``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).
+            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).
 
-        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            `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`):
             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:`{0}`, `optional`, defaults to :obj:`None`):
+        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.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -514,6 +470,7 @@ class LongformerModel(RobertaModel):
     """
 
     config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "longformer"
 
     def __init__(self, config):
@@ -578,12 +535,11 @@ 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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
         attention_mask=None,
-        global_attention_mask=None,
         token_type_ids=None,
         position_ids=None,
         inputs_embeds=None,
@@ -614,8 +570,8 @@ class LongformerModel(RobertaModel):
         import torch
         from transformers import LongformerModel, LongformerTokenizer
 
-        model = LongformerModel.from_pretrained('allenai/longformer-base-4096')
-        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
+        model = LongformerModel.from_pretrained('longformer-base-4096')
+        tokenizer = LongformerTokenizer.from_pretrained('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
@@ -635,19 +591,6 @@ 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,
@@ -685,6 +628,7 @@ class LongformerModel(RobertaModel):
 @add_start_docstrings("""Longformer Model with a `language modeling` head on top. """, LONGFORMER_START_DOCSTRING)
 class LongformerForMaskedLM(BertPreTrainedModel):
     config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "longformer"
 
     def __init__(self, config):
@@ -695,12 +639,11 @@ class LongformerForMaskedLM(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
         attention_mask=None,
-        global_attention_mask=None,
         token_type_ids=None,
         position_ids=None,
         inputs_embeds=None,
@@ -736,8 +679,8 @@ class LongformerForMaskedLM(BertPreTrainedModel):
         import torch
         from transformers import LongformerForMaskedLM, LongformerTokenizer
 
-        model = LongformerForMaskedLM.from_pretrained('allenai/longformer-base-4096')
-        tokenizer = LongformerTokenizer.from_pretrained('allenai/longformer-base-4096')
+        model = LongformerForMaskedLM.from_pretrained('longformer-base-4096')
+        tokenizer = LongformerTokenizer.from_pretrained('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
@@ -750,7 +693,6 @@ 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,
@@ -768,122 +710,6 @@ class LongformerForMaskedLM(BertPreTrainedModel):
         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
-    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`). """,
@@ -891,6 +717,7 @@ class LongformerClassificationHead(nn.Module):
 )
 class LongformerForQuestionAnswering(BertPreTrainedModel):
     config_class = LongformerConfig
+    pretrained_model_archive_map = LONGFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "longformer"
 
     def __init__(self, config):
@@ -902,12 +729,32 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def _compute_global_attention_mask(self, input_ids):
+        question_end_index = self._get_question_end_index(input_ids)
+        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)
+        attention_mask = attention_mask.expand_as(input_ids) < question_end_index
+
+        attention_mask = attention_mask.int() + 1  # True => global attention; False => local attention
+        return attention_mask.long()
+
+    def _get_question_end_index(self, input_ids):
+        sep_token_indices = (input_ids == self.config.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: {self.config.sep_token_id} in every sample for questions answering"
+
+        return sep_token_indices.view(batch_size, 3, 2)[:, 0, 1]
+
+    @add_start_docstrings_to_callable(LONGFORMER_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids,
         attention_mask=None,
-        global_attention_mask=None,
         token_type_ids=None,
         position_ids=None,
         inputs_embeds=None,
@@ -946,8 +793,8 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
         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")
+        tokenizer = LongformerTokenizer.from_pretrained("longformer-large-4096-finetuned-triviaqa")
+        model = LongformerForQuestionAnswering.from_pretrained("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")
@@ -966,15 +813,17 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
         """
 
         # 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)
+        global_attention_mask = self._compute_global_attention_mask(input_ids)
+        if attention_mask is None:
+            attention_mask = global_attention_mask
+        else:
+            # combine global_attention_mask with attention_mask
+            # global attention on question tokens, no attention on padding tokens
+            attention_mask = global_attention_mask * attention_mask
 
         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,
@@ -1006,222 +855,3 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
             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
-    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
-    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_marian.py

@@ -18,11 +18,6 @@
 from transformers.modeling_bart import BartForConditionalGeneration
 
 
-MARIAN_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    # See all Marian models at https://huggingface.co/models?search=Helsinki-NLP
-]
-
-
 class MarianMTModel(BartForConditionalGeneration):
     r"""
     Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
@@ -46,6 +41,8 @@ class MarianMTModel(BartForConditionalGeneration):
 
     """
 
+    pretrained_model_archive_map = {}  # see https://huggingface.co/models?search=Helsinki-NLP
+
     def prepare_logits_for_generation(self, logits, cur_len, max_length):
         logits[:, self.config.pad_token_id] = float("-inf")
         if cur_len == max_length - 1 and self.config.eos_token_id is not None:

src/transformers/modeling_openai.py

@@ -33,10 +33,7 @@ from .modeling_utils import Conv1D, PreTrainedModel, SequenceSummary, prune_conv
 
 logger = logging.getLogger(__name__)
 
-OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "openai-gpt",
-    # See all OpenAI GPT models at https://huggingface.co/models?filter=openai-gpt
-]
+OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP = {"openai-gpt": "https://cdn.huggingface.co/openai-gpt-pytorch_model.bin"}
 
 
 def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
@@ -255,6 +252,7 @@ class OpenAIGPTPreTrainedModel(PreTrainedModel):
     """
 
     config_class = OpenAIGPTConfig
+    pretrained_model_archive_map = OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_openai_gpt
     base_model_prefix = "transformer"
 
@@ -315,7 +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**.
-        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
@@ -493,7 +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 ``labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_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

@@ -35,11 +35,10 @@ from .modeling_utils import PreTrainedModel, apply_chunking_to_forward
 
 logger = logging.getLogger(__name__)
 
-REFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "google/reformer-crime-and-punishment",
-    "google/reformer-enwik8",
-    # See all Reformer models at https://huggingface.co/models?filter=reformer
-]
+REFORMER_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "google/reformer-crime-and-punishment": "https://cdn.huggingface.co/google/reformer-crime-and-punishment/pytorch_model.bin",
+    "google/reformer-enwik8": "https://cdn.huggingface.co/google/reformer-enwik8/pytorch_model.bin",
+}
 
 
 def mish(x):
@@ -1374,6 +1373,7 @@ class ReformerPreTrainedModel(PreTrainedModel):
     """
 
     config_class = ReformerConfig
+    pretrained_model_archive_map = REFORMER_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "reformer"
 
     @property
@@ -1662,7 +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 `inputs_embeds` with padding to match least common multiple chunk_length
+        # Extend `input_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

@@ -30,15 +30,14 @@ from .modeling_utils import create_position_ids_from_input_ids
 
 logger = logging.getLogger(__name__)
 
-ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "roberta-base",
-    "roberta-large",
-    "roberta-large-mnli",
-    "distilroberta-base",
-    "roberta-base-openai-detector",
-    "roberta-large-openai-detector",
-    # See all RoBERTa models at https://huggingface.co/models?filter=roberta
-]
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "roberta-base": "https://cdn.huggingface.co/roberta-base-pytorch_model.bin",
+    "roberta-large": "https://cdn.huggingface.co/roberta-large-pytorch_model.bin",
+    "roberta-large-mnli": "https://cdn.huggingface.co/roberta-large-mnli-pytorch_model.bin",
+    "distilroberta-base": "https://cdn.huggingface.co/distilroberta-base-pytorch_model.bin",
+    "roberta-base-openai-detector": "https://cdn.huggingface.co/roberta-base-openai-detector-pytorch_model.bin",
+    "roberta-large-openai-detector": "https://cdn.huggingface.co/roberta-large-openai-detector-pytorch_model.bin",
+}
 
 
 class RobertaEmbeddings(BertEmbeddings):
@@ -96,7 +95,7 @@ ROBERTA_START_DOCSTRING = r"""
 
 ROBERTA_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.RobertaTokenizer`.
@@ -104,19 +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:`{0}`, `optional`, defaults to :obj:`None`):
+        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]``:
             ``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:`{0}`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `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:`{0}`, `optional`, defaults to :obj:`None`):
+        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.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
@@ -143,6 +142,7 @@ class RobertaModel(BertModel):
     """
 
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -161,6 +161,7 @@ class RobertaModel(BertModel):
 @add_start_docstrings("""RoBERTa Model with a `language modeling` head on top. """, ROBERTA_START_DOCSTRING)
 class RobertaForMaskedLM(BertPreTrainedModel):
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -174,7 +175,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
     def get_output_embeddings(self):
         return self.lm_head.decoder
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -275,6 +276,7 @@ class RobertaLMHead(nn.Module):
 )
 class RobertaForSequenceClassification(BertPreTrainedModel):
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -284,7 +286,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
         self.roberta = RobertaModel(config)
         self.classifier = RobertaClassificationHead(config)
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -365,6 +367,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
 )
 class RobertaForMultipleChoice(BertPreTrainedModel):
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -376,7 +379,7 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -463,6 +466,7 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
 )
 class RobertaForTokenClassification(BertPreTrainedModel):
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -475,7 +479,7 @@ class RobertaForTokenClassification(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids=None,
@@ -582,6 +586,7 @@ class RobertaClassificationHead(nn.Module):
 )
 class RobertaForQuestionAnswering(BertPreTrainedModel):
     config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "roberta"
 
     def __init__(self, config):
@@ -593,7 +598,7 @@ class RobertaForQuestionAnswering(BertPreTrainedModel):
 
         self.init_weights()
 
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids,

src/transformers/modeling_t5.py

@@ -36,14 +36,13 @@ logger = logging.getLogger(__name__)
 # This dict contrains shortcut names and associated url
 # for the pretrained weights provided with the models
 ####################################################
-T5_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "t5-small",
-    "t5-base",
-    "t5-large",
-    "t5-3b",
-    "t5-11b",
-    # See all T5 models at https://huggingface.co/models?filter=t5
-]
+T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "t5-small": "https://cdn.huggingface.co/t5-small-pytorch_model.bin",
+    "t5-base": "https://cdn.huggingface.co/t5-base-pytorch_model.bin",
+    "t5-large": "https://cdn.huggingface.co/t5-large-pytorch_model.bin",
+    "t5-3b": "https://cdn.huggingface.co/t5-3b-pytorch_model.bin",
+    "t5-11b": "https://cdn.huggingface.co/t5-11b-pytorch_model.bin",
+}
 
 
 ####################################################
@@ -556,6 +555,7 @@ class T5PreTrainedModel(PreTrainedModel):
     """
 
     config_class = T5Config
+    pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
     load_tf_weights = load_tf_weights_in_t5
     base_model_prefix = "transformer"
 

src/transformers/modeling_tf_albert.py

@@ -29,17 +29,16 @@ from .tokenization_utils import BatchEncoding
 
 logger = logging.getLogger(__name__)
 
-TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "albert-base-v1",
-    "albert-large-v1",
-    "albert-xlarge-v1",
-    "albert-xxlarge-v1",
-    "albert-base-v2",
-    "albert-large-v2",
-    "albert-xlarge-v2",
-    "albert-xxlarge-v2",
-    # See all ALBERT models at https://huggingface.co/models?filter=albert
-]
+TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "albert-base-v1": "https://cdn.huggingface.co/albert-base-v1-with-prefix-tf_model.h5",
+    "albert-large-v1": "https://cdn.huggingface.co/albert-large-v1-with-prefix-tf_model.h5",
+    "albert-xlarge-v1": "https://cdn.huggingface.co/albert-xlarge-v1-with-prefix-tf_model.h5",
+    "albert-xxlarge-v1": "https://cdn.huggingface.co/albert-xxlarge-v1-with-prefix-tf_model.h5",
+    "albert-base-v2": "https://cdn.huggingface.co/albert-base-v2-with-prefix-tf_model.h5",
+    "albert-large-v2": "https://cdn.huggingface.co/albert-large-v2-with-prefix-tf_model.h5",
+    "albert-xlarge-v2": "https://cdn.huggingface.co/albert-xlarge-v2-with-prefix-tf_model.h5",
+    "albert-xxlarge-v2": "https://cdn.huggingface.co/albert-xxlarge-v2-with-prefix-tf_model.h5",
+}
 
 
 class TFAlbertEmbeddings(tf.keras.layers.Layer):
@@ -441,6 +440,7 @@ class TFAlbertPreTrainedModel(TFPreTrainedModel):
     """
 
     config_class = AlbertConfig
+    pretrained_model_archive_map = TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "albert"
 
 
@@ -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:`{0}`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
             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:`{0}`, `optional, defaults to :obj:`None`):
+        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` 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]``:
             ``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:`{0}`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `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:`{0}`, `optional`, defaults to :obj:`None`):
+        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `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**.
-        inputs_embeds (:obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+        input_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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     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.format("(batch_size, num_choices, sequence_length)"))
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
     def call(
         self,
         inputs,

src/transformers/modeling_tf_auto.py

@@ -34,6 +34,7 @@ from .configuration_auto import (
 )
 from .configuration_utils import PretrainedConfig
 from .modeling_tf_albert import (
+    TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFAlbertForMaskedLM,
     TFAlbertForMultipleChoice,
     TFAlbertForPreTraining,
@@ -42,6 +43,7 @@ from .modeling_tf_albert import (
     TFAlbertModel,
 )
 from .modeling_tf_bert import (
+    TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFBertForMaskedLM,
     TFBertForMultipleChoice,
     TFBertForPreTraining,
@@ -50,32 +52,40 @@ from .modeling_tf_bert import (
     TFBertForTokenClassification,
     TFBertModel,
 )
-from .modeling_tf_ctrl import TFCTRLLMHeadModel, TFCTRLModel
+from .modeling_tf_ctrl import TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, TFCTRLLMHeadModel, TFCTRLModel
 from .modeling_tf_distilbert import (
+    TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFDistilBertForMaskedLM,
     TFDistilBertForQuestionAnswering,
     TFDistilBertForSequenceClassification,
     TFDistilBertForTokenClassification,
     TFDistilBertModel,
 )
-from .modeling_tf_gpt2 import TFGPT2LMHeadModel, TFGPT2Model
-from .modeling_tf_openai import TFOpenAIGPTLMHeadModel, TFOpenAIGPTModel
+from .modeling_tf_gpt2 import TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP, TFGPT2LMHeadModel, TFGPT2Model
+from .modeling_tf_openai import TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP, TFOpenAIGPTLMHeadModel, TFOpenAIGPTModel
 from .modeling_tf_roberta import (
+    TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFRobertaForMaskedLM,
     TFRobertaForQuestionAnswering,
     TFRobertaForSequenceClassification,
     TFRobertaForTokenClassification,
     TFRobertaModel,
 )
-from .modeling_tf_t5 import TFT5ForConditionalGeneration, TFT5Model
-from .modeling_tf_transfo_xl import TFTransfoXLLMHeadModel, TFTransfoXLModel
+from .modeling_tf_t5 import TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP, TFT5ForConditionalGeneration, TFT5Model
+from .modeling_tf_transfo_xl import (
+    TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+    TFTransfoXLLMHeadModel,
+    TFTransfoXLModel,
+)
 from .modeling_tf_xlm import (
+    TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFXLMForQuestionAnsweringSimple,
     TFXLMForSequenceClassification,
     TFXLMModel,
     TFXLMWithLMHeadModel,
 )
 from .modeling_tf_xlnet import (
+    TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
     TFXLNetForQuestionAnsweringSimple,
     TFXLNetForSequenceClassification,
     TFXLNetForTokenClassification,
@@ -87,6 +97,24 @@ from .modeling_tf_xlnet import (
 logger = logging.getLogger(__name__)
 
 
+TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict(
+    (key, value)
+    for pretrained_map in [
+        TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+    ]
+    for key, value, in pretrained_map.items()
+)
+
 TF_MODEL_MAPPING = OrderedDict(
     [
         (T5Config, TFT5Model),
@@ -180,17 +208,20 @@ class TFAutoModel(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: TFT5Model (T5 model)
-            - `distilbert`: TFDistilBertModel (DistilBERT model)
-            - `roberta`: TFRobertaModel (RoBERTa model)
-            - `bert`: TFBertModel (Bert model)
-            - `openai-gpt`: TFOpenAIGPTModel (OpenAI GPT model)
-            - `gpt2`: TFGPT2Model (OpenAI GPT-2 model)
-            - `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
-            - `xlnet`: TFXLNetModel (XLNet model)
-            - `xlm`: TFXLMModel (XLM model)
-            - `ctrl`: TFCTRLModel (CTRL model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The base model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
+            - contains `distilbert`: TFDistilBertModel (DistilBERT model)
+            - contains `roberta`: TFRobertaModel (RoBERTa model)
+            - contains `bert`: TFBertModel (Bert model)
+            - contains `openai-gpt`: TFOpenAIGPTModel (OpenAI GPT model)
+            - contains `gpt2`: TFGPT2Model (OpenAI GPT-2 model)
+            - contains `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
+            - contains `xlnet`: TFXLNetModel (XLNet model)
+            - contains `xlm`: TFXLMModel (XLM model)
+            - contains `ctrl`: TFCTRLModel (CTRL model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -207,12 +238,6 @@ 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)
@@ -245,18 +270,17 @@ class TFAutoModel(object):
         r""" Instantiates one of the base model classes of the library
         from a pre-trained model configuration.
 
-        The `from_pretrained()` method takes care of returning the correct model class instance
-        based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: TFT5Model (T5 model)
-            - `distilbert`: TFDistilBertModel (DistilBERT model)
-            - `roberta`: TFRobertaModel (RoBERTa model)
-            - `bert`: TFTFBertModel (Bert model)
-            - `openai-gpt`: TFOpenAIGPTModel (OpenAI GPT model)
-            - `gpt2`: TFGPT2Model (OpenAI GPT-2 model)
-            - `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
-            - `xlnet`: TFXLNetModel (XLNet model)
-            - `ctrl`: TFCTRLModel (CTRL model)
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
+            - contains `distilbert`: TFDistilBertModel (DistilBERT model)
+            - contains `roberta`: TFRobertaModel (RoBERTa model)
+            - contains `bert`: TFTFBertModel (Bert model)
+            - contains `openai-gpt`: TFOpenAIGPTModel (OpenAI GPT model)
+            - contains `gpt2`: TFGPT2Model (OpenAI GPT-2 model)
+            - contains `transfo-xl`: TFTransfoXLModel (Transformer-XL model)
+            - contains `xlnet`: TFXLNetModel (XLNet model)
+            - contains `ctrl`: TFCTRLModel (CTRL model)
 
         Params:
             pretrained_model_name_or_path: either:
@@ -354,11 +378,6 @@ 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:
@@ -394,18 +413,21 @@ class TFAutoModelForPreTraining(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: :class:`~transformers.TFT5ModelWithLMHead` (T5 model)
-            - `distilbert`: :class:`~transformers.TFDistilBertForMaskedLM` (DistilBERT model)
-            - `albert`: :class:`~transformers.TFAlbertForPreTraining` (ALBERT model)
-            - `roberta`: :class:`~transformers.TFRobertaForMaskedLM` (RoBERTa model)
-            - `bert`: :class:`~transformers.TFBertForPreTraining` (Bert model)
-            - `openai-gpt`: :class:`~transformers.TFOpenAIGPTLMHeadModel` (OpenAI GPT model)
-            - `gpt2`: :class:`~transformers.TFGPT2LMHeadModel` (OpenAI GPT-2 model)
-            - `transfo-xl`: :class:`~transformers.TFTransfoXLLMHeadModel` (Transformer-XL model)
-            - `xlnet`: :class:`~transformers.TFXLNetLMHeadModel` (XLNet model)
-            - `xlm`: :class:`~transformers.TFXLMWithLMHeadModel` (XLM model)
-            - `ctrl`: :class:`~transformers.TFCTRLLMHeadModel` (Salesforce CTRL model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: :class:`~transformers.TFT5ModelWithLMHead` (T5 model)
+            - contains `distilbert`: :class:`~transformers.TFDistilBertForMaskedLM` (DistilBERT model)
+            - contains `albert`: :class:`~transformers.TFAlbertForPreTraining` (ALBERT model)
+            - contains `roberta`: :class:`~transformers.TFRobertaForMaskedLM` (RoBERTa model)
+            - contains `bert`: :class:`~transformers.TFBertForPreTraining` (Bert model)
+            - contains `openai-gpt`: :class:`~transformers.TFOpenAIGPTLMHeadModel` (OpenAI GPT model)
+            - contains `gpt2`: :class:`~transformers.TFGPT2LMHeadModel` (OpenAI GPT-2 model)
+            - contains `transfo-xl`: :class:`~transformers.TFTransfoXLLMHeadModel` (Transformer-XL model)
+            - contains `xlnet`: :class:`~transformers.TFXLNetLMHeadModel` (XLNet model)
+            - contains `xlm`: :class:`~transformers.TFXLMWithLMHeadModel` (XLM model)
+            - contains `ctrl`: :class:`~transformers.TFCTRLLMHeadModel` (Salesforce CTRL model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -492,17 +514,20 @@ class TFAutoModelWithLMHead(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: TFT5ForConditionalGeneration (T5 model)
-            - `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
-            - `roberta`: TFRobertaForMaskedLM (RoBERTa model)
-            - `bert`: TFBertForMaskedLM (Bert model)
-            - `openai-gpt`: TFOpenAIGPTLMHeadModel (OpenAI GPT model)
-            - `gpt2`: TFGPT2LMHeadModel (OpenAI GPT-2 model)
-            - `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
-            - `xlnet`: TFXLNetLMHeadModel (XLNet model)
-            - `xlm`: TFXLMWithLMHeadModel (XLM model)
-            - `ctrl`: TFCTRLLMHeadModel (CTRL model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5ForConditionalGeneration (T5 model)
+            - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
+            - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
+            - contains `bert`: TFBertForMaskedLM (Bert model)
+            - contains `openai-gpt`: TFOpenAIGPTLMHeadModel (OpenAI GPT model)
+            - contains `gpt2`: TFGPT2LMHeadModel (OpenAI GPT-2 model)
+            - contains `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
+            - contains `xlnet`: TFXLNetLMHeadModel (XLNet model)
+            - contains `xlm`: TFXLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: TFCTRLLMHeadModel (CTRL model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -519,12 +544,6 @@ 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)
@@ -559,17 +578,20 @@ class TFAutoModelWithLMHead(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `t5`: TFT5ForConditionalGeneration (T5 model)
-            - `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
-            - `roberta`: TFRobertaForMaskedLM (RoBERTa model)
-            - `bert`: TFBertForMaskedLM (Bert model)
-            - `openai-gpt`: TFOpenAIGPTLMHeadModel (OpenAI GPT model)
-            - `gpt2`: TFGPT2LMHeadModel (OpenAI GPT-2 model)
-            - `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
-            - `xlnet`: TFXLNetLMHeadModel (XLNet model)
-            - `xlm`: TFXLMWithLMHeadModel (XLM model)
-            - `ctrl`: TFCTRLLMHeadModel (CTRL model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5ForConditionalGeneration (T5 model)
+            - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
+            - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
+            - contains `bert`: TFBertForMaskedLM (Bert model)
+            - contains `openai-gpt`: TFOpenAIGPTLMHeadModel (OpenAI GPT model)
+            - contains `gpt2`: TFGPT2LMHeadModel (OpenAI GPT-2 model)
+            - contains `transfo-xl`: TFTransfoXLLMHeadModel (Transformer-XL model)
+            - contains `xlnet`: TFXLNetLMHeadModel (XLNet model)
+            - contains `xlm`: TFXLMWithLMHeadModel (XLM model)
+            - contains `ctrl`: TFCTRLLMHeadModel (CTRL model)
 
         Params:
             pretrained_model_name_or_path: either:
@@ -655,9 +677,12 @@ class TFAutoModelForMultipleChoice:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `albert`: TFAlbertForMultipleChoice (Albert model)
-            - `bert`: TFBertForMultipleChoice (Bert model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `albert`: TFAlbertForMultipleChoice (Albert model)
+            - contains `bert`: TFBertForMultipleChoice (Bert model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -674,12 +699,6 @@ 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)
@@ -709,9 +728,12 @@ class TFAutoModelForMultipleChoice:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `albert`: TFRobertaForMultiple (Albert model)
-            - `bert`: TFBertForMultipleChoice (Bert model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `albert`: TFRobertaForMultiple (Albert model)
+            - contains `bert`: TFBertForMultipleChoice (Bert model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -802,12 +824,15 @@ class TFAutoModelForSequenceClassification(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: TFDistilBertForSequenceClassification (DistilBERT model)
-            - `roberta`: TFRobertaForSequenceClassification (RoBERTa model)
-            - `bert`: TFBertForSequenceClassification (Bert model)
-            - `xlnet`: TFXLNetForSequenceClassification (XLNet model)
-            - `xlm`: TFXLMForSequenceClassification (XLM model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: TFDistilBertForSequenceClassification (DistilBERT model)
+            - contains `roberta`: TFRobertaForSequenceClassification (RoBERTa model)
+            - contains `bert`: TFBertForSequenceClassification (Bert model)
+            - contains `xlnet`: TFXLNetForSequenceClassification (XLNet model)
+            - contains `xlm`: TFXLMForSequenceClassification (XLM model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -824,12 +849,6 @@ 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)
@@ -862,12 +881,15 @@ class TFAutoModelForSequenceClassification(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: TFDistilBertForSequenceClassification (DistilBERT model)
-            - `roberta`: TFRobertaForSequenceClassification (RoBERTa model)
-            - `bert`: TFBertForSequenceClassification (Bert model)
-            - `xlnet`: TFXLNetForSequenceClassification (XLNet model)
-            - `xlm`: TFXLMForSequenceClassification (XLM model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: TFDistilBertForSequenceClassification (DistilBERT model)
+            - contains `roberta`: TFRobertaForSequenceClassification (RoBERTa model)
+            - contains `bert`: TFBertForSequenceClassification (Bert model)
+            - contains `xlnet`: TFXLNetForSequenceClassification (XLNet model)
+            - contains `xlm`: TFXLMForSequenceClassification (XLM model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -958,13 +980,16 @@ class TFAutoModelForQuestionAnswering(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: TFDistilBertForQuestionAnswering (DistilBERT model)
-            - `albert`: TFAlbertForQuestionAnswering (ALBERT model)
-            - `roberta`: TFRobertaForQuestionAnswering (RoBERTa model)
-            - `bert`: TFBertForQuestionAnswering (Bert model)
-            - `xlnet`: TFXLNetForQuestionAnswering (XLNet model)
-            - `xlm`: TFXLMForQuestionAnswering (XLM model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: TFDistilBertForQuestionAnswering (DistilBERT model)
+            - contains `albert`: TFAlbertForQuestionAnswering (ALBERT model)
+            - contains `roberta`: TFRobertaForQuestionAnswering (RoBERTa model)
+            - contains `bert`: TFBertForQuestionAnswering (Bert model)
+            - contains `xlnet`: TFXLNetForQuestionAnswering (XLNet model)
+            - contains `xlm`: TFXLMForQuestionAnswering (XLM model)
 
         This class cannot be instantiated using `__init__()` (throws an error).
     """
@@ -981,12 +1006,6 @@ 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)
@@ -1020,13 +1039,16 @@ class TFAutoModelForQuestionAnswering(object):
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `distilbert`: TFDistilBertForQuestionAnswering (DistilBERT model)
-            - `albert`: TFAlbertForQuestionAnswering (ALBERT model)
-            - `roberta`: TFRobertaForQuestionAnswering (RoBERTa model)
-            - `bert`: TFBertForQuestionAnswering (Bert model)
-            - `xlnet`: TFXLNetForQuestionAnswering (XLNet model)
-            - `xlm`: TFXLMForQuestionAnswering (XLM model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: TFDistilBertForQuestionAnswering (DistilBERT model)
+            - contains `albert`: TFAlbertForQuestionAnswering (ALBERT model)
+            - contains `roberta`: TFRobertaForQuestionAnswering (RoBERTa model)
+            - contains `bert`: TFBertForQuestionAnswering (Bert model)
+            - contains `xlnet`: TFXLNetForQuestionAnswering (XLNet model)
+            - contains `xlm`: TFXLMForQuestionAnswering (XLM model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`
@@ -1121,12 +1143,6 @@ 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)
@@ -1158,11 +1174,14 @@ class TFAutoModelForTokenClassification:
 
         The `from_pretrained()` method takes care of returning the correct model class instance
         based on the `model_type` property of the config object, or when it's missing,
-        falling back to using pattern matching on the `pretrained_model_name_or_path` string:
-            - `bert`: BertForTokenClassification (Bert model)
-            - `xlnet`: XLNetForTokenClassification (XLNet model)
-            - `distilbert`: DistilBertForTokenClassification (DistilBert model)
-            - `roberta`: RobertaForTokenClassification (Roberta model)
+        falling back to using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `bert`: BertForTokenClassification (Bert model)
+            - contains `xlnet`: XLNetForTokenClassification (XLNet model)
+            - contains `distilbert`: DistilBertForTokenClassification (DistilBert model)
+            - contains `roberta`: RobertaForTokenClassification (Roberta model)
 
         The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated)
         To train the model, you should first set it back in training mode with `model.train()`

src/transformers/modeling_tf_bert.py

@@ -30,29 +30,28 @@ from .tokenization_utils import BatchEncoding
 logger = logging.getLogger(__name__)
 
 
-TF_BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "bert-base-uncased",
-    "bert-large-uncased",
-    "bert-base-cased",
-    "bert-large-cased",
-    "bert-base-multilingual-uncased",
-    "bert-base-multilingual-cased",
-    "bert-base-chinese",
-    "bert-base-german-cased",
-    "bert-large-uncased-whole-word-masking",
-    "bert-large-cased-whole-word-masking",
-    "bert-large-uncased-whole-word-masking-finetuned-squad",
-    "bert-large-cased-whole-word-masking-finetuned-squad",
-    "bert-base-cased-finetuned-mrpc",
-    "cl-tohoku/bert-base-japanese",
-    "cl-tohoku/bert-base-japanese-whole-word-masking",
-    "cl-tohoku/bert-base-japanese-char",
-    "cl-tohoku/bert-base-japanese-char-whole-word-masking",
-    "TurkuNLP/bert-base-finnish-cased-v1",
-    "TurkuNLP/bert-base-finnish-uncased-v1",
-    "wietsedv/bert-base-dutch-cased",
-    # See all BERT models at https://huggingface.co/models?filter=bert
-]
+TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    "bert-base-uncased": "https://cdn.huggingface.co/bert-base-uncased-tf_model.h5",
+    "bert-large-uncased": "https://cdn.huggingface.co/bert-large-uncased-tf_model.h5",
+    "bert-base-cased": "https://cdn.huggingface.co/bert-base-cased-tf_model.h5",
+    "bert-large-cased": "https://cdn.huggingface.co/bert-large-cased-tf_model.h5",
+    "bert-base-multilingual-uncased": "https://cdn.huggingface.co/bert-base-multilingual-uncased-tf_model.h5",
+    "bert-base-multilingual-cased": "https://cdn.huggingface.co/bert-base-multilingual-cased-tf_model.h5",
+    "bert-base-chinese": "https://cdn.huggingface.co/bert-base-chinese-tf_model.h5",
+    "bert-base-german-cased": "https://cdn.huggingface.co/bert-base-german-cased-tf_model.h5",
+    "bert-large-uncased-whole-word-masking": "https://cdn.huggingface.co/bert-large-uncased-whole-word-masking-tf_model.h5",
+    "bert-large-cased-whole-word-masking": "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-tf_model.h5",
+    "bert-large-uncased-whole-word-masking-finetuned-squad": "https://cdn.huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad-tf_model.h5",
+    "bert-large-cased-whole-word-masking-finetuned-squad": "https://cdn.huggingface.co/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
+    "bert-base-cased-finetuned-mrpc": "https://cdn.huggingface.co/bert-base-cased-finetuned-mrpc-tf_model.h5",
+    "bert-base-japanese": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese/tf_model.h5",
+    "bert-base-japanese-whole-word-masking": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-whole-word-masking/tf_model.h5",
+    "bert-base-japanese-char": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-char/tf_model.h5",
+    "bert-base-japanese-char-whole-word-masking": "https://cdn.huggingface.co/cl-tohoku/bert-base-japanese-char-whole-word-masking/tf_model.h5",
+    "bert-base-finnish-cased-v1": "https://cdn.huggingface.co/TurkuNLP/bert-base-finnish-cased-v1/tf_model.h5",
+    "bert-base-finnish-uncased-v1": "https://cdn.huggingface.co/TurkuNLP/bert-base-finnish-uncased-v1/tf_model.h5",
+    "bert-base-dutch-cased": "https://cdn.huggingface.co/wietsedv/bert-base-dutch-cased/tf_model.h5",
+}
 
 
 def gelu(x):
@@ -586,6 +585,7 @@ class TFBertPreTrainedModel(TFPreTrainedModel):
     """
 
     config_class = BertConfig
+    pretrained_model_archive_map = TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
     base_model_prefix = "bert"
 
 
@@ -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:`{0}`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
             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:`{0}`, `optional`, defaults to :obj:`None`):
+        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` 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]``:
             ``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:`{0}`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `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:`{0}`, `optional`, defaults to :obj:`None`):
+        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`, `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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     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.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -857,13 +857,10 @@ 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)
 
@@ -891,7 +888,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -957,7 +954,7 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel):
         """
         return {"input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS)}
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def call(
         self,
         inputs,
@@ -993,15 +990,11 @@ 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]
@@ -1072,7 +1065,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="classifier"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""
     Return:
@@ -1129,7 +1122,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
             config.num_labels, kernel_initializer=get_initializer(config.initializer_range), name="qa_outputs"
         )
 
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
     def call(self, inputs, **kwargs):
         r"""
     Return: