Have a trace of the working torch versions

* change TokenClassificationTask class methods to static methods

Since we do not require self in the class methods of TokenClassificationTask we should probably switch to static methods. Also, since the class TokenClassificationTask does not contain a constructor it is currently unusable as is. By switching to static methods this fixes the issue of having to document the intent of the broken class.

Also, since the get_labels and read_examples_from_file methods are ought to be implemented. Static method definitions are unchanged even after inheritance, which means that it can be overridden, similar to other class methods.

* Trigger Build

Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

.circleci/config.yml

@@ -114,6 +114,115 @@ jobs:
             - store_artifacts:
                   path: ~/transformers/reports
 
+    run_tests_torch_1_3:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install .[sklearn,testing]
+            - run: pip install torch==1.3.0
+            - save_cache:
+                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
+    run_tests_torch_1_4:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install .[sklearn,testing]
+            - run: pip install torch==1.4.0
+            - save_cache:
+                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
+    run_tests_torch_1_5:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install .[sklearn,testing]
+            - run: pip install torch==1.5.1
+            - save_cache:
+                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
+    run_tests_torch_1_6:
+        working_directory: ~/transformers
+        docker:
+            - image: circleci/python:3.7
+        environment:
+            OMP_NUM_THREADS: 1
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - restore_cache:
+                  keys:
+                      - v0.4-torch-{{ checksum "setup.py" }}
+                      - v0.4-{{ checksum "setup.py" }}
+            - run: pip install --upgrade pip
+            - run: pip install .[sklearn,testing]
+            - run: pip install torch==1.6.0
+            - save_cache:
+                  key: v0.4-torch-{{ checksum "setup.py" }}
+                  paths:
+                      - '~/.cache/pip'
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=tests_torch ./tests/ | tee tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
+
     run_tests_tf:
         working_directory: ~/transformers
         docker:
@@ -393,6 +502,10 @@ workflows:
             - run_tests_custom_tokenizers
             - run_tests_torch_and_tf
             - run_tests_torch
+            - run_tests_torch_1_3
+            - run_tests_torch_1_4
+            - run_tests_torch_1_5
+            - run_tests_torch_1_6
             - run_tests_tf
             - run_tests_flax
             - run_tests_pipelines_torch

.github/ISSUE_TEMPLATE/bug-report.md

@@ -36,13 +36,15 @@ assignees: ''
  examples/distillation: @VictorSanh
  nlp datasets: [different repo](https://github.com/huggingface/nlp)
  rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
- Text Generation: @TevenLeScao
+ Text Generation: @patrickvonplaten @TevenLeScao
  blenderbot: @mariamabarham
  Bart: @sshleifer
  Marian: @sshleifer
  T5: @patrickvonplaten
  Longformer/Reformer: @patrickvonplaten
- TransfoXL/XLNet: @TevenLeScao 
+ TransfoXL/XLNet: @TevenLeScao
+ RAG: @patrickvonplaten, @lhoestq
+ FSTM: @stas00
  examples/seq2seq: @sshleifer
  examples/bert-loses-patience: @JetRunner
  tensorflow: @jplu

.github/PULL_REQUEST_TEMPLATE.md

@@ -60,4 +60,5 @@ members/contributors which may be interested in your PR.
  tensorflow: @jplu
  examples/token-classification: @stefan-it
  documentation: @sgugger
+ FSTM: @stas00
  -->

examples/README.md

@@ -68,7 +68,7 @@ For example for `run_glue`:
 
 ```bash
 python examples/xla_spawn.py --num_cores 8 \
-	examples/text-classification/run_glue.py
+	examples/text-classification/run_glue.py \
 	--model_name_or_path bert-base-cased \
 	--task_name mnli \
 	--data_dir ./data/glue_data/MNLI \

examples/language-modeling/README.md

@@ -10,7 +10,7 @@ These scripts leverage the 🤗 Datasets library and the Trainer API. You can ea
 need extra processing on your datasets.
 
 **Note:** The old script `run_language_modeling.py` is still available
-[here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/language-modeling/run_language_modeling.py).
+[here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/run_language_modeling.py).
 
 The following examples, will run on a datasets hosted on our [hub](https://huggingface.co/datasets) or with your own
 text files for training and validation. We give examples of both below.
@@ -68,13 +68,13 @@ python run_mlm.py \
 To run on your own training and validation files, use the following command:
 
 ```bash
-python run_clm.py \
+python run_mlm.py \
     --model_name_or_path roberta-base \
     --train_file path_to_train_file \
     --validation_file path_to_validation_file \
     --do_train \
     --do_eval \
-    --output_dir /tmp/test-clm
+    --output_dir /tmp/test-mlm
 ```
 
 If your dataset is organized with one sample per line, you can use the `--line_by_line` flag (otherwise the script

examples/language-modeling/run_clm.py

@@ -190,7 +190,7 @@ def main():
         if data_args.train_file is not None:
             data_files["train"] = data_args.train_file
         if data_args.validation_file is not None:
-            data_files["validation"] = data_args.train_file
+            data_files["validation"] = data_args.validation_file
         extension = data_args.train_file.split(".")[-1]
         if extension == "txt":
             extension = "text"

examples/language-modeling/run_mlm.py

@@ -201,7 +201,7 @@ def main():
         if data_args.train_file is not None:
             data_files["train"] = data_args.train_file
         if data_args.validation_file is not None:
-            data_files["validation"] = data_args.train_file
+            data_files["validation"] = data_args.validation_file
         extension = data_args.train_file.split(".")[-1]
         if extension == "txt":
             extension = "text"
@@ -264,7 +264,15 @@ def main():
         def tokenize_function(examples):
             # Remove empty lines
             examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
-            return tokenizer(examples["text"], padding=padding, truncation=True, max_length=data_args.max_seq_length)
+            return tokenizer(
+                examples["text"],
+                padding=padding,
+                truncation=True,
+                max_length=data_args.max_seq_length,
+                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
+                # receives the `special_tokens_mask`.
+                return_special_tokens_mask=True,
+            )
 
         tokenized_datasets = datasets.map(
             tokenize_function,
@@ -275,8 +283,10 @@ def main():
         )
     else:
         # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
+        # efficient when it receives the `special_tokens_mask`.
         def tokenize_function(examples):
-            return tokenizer(examples[text_column_name])
+            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)
 
         tokenized_datasets = datasets.map(
             tokenize_function,

examples/language-modeling/run_mlm_wwm.py

@@ -204,7 +204,7 @@ def main():
     if data_args.train_file is not None:
         data_files["train"] = data_args.train_file
     if data_args.validation_file is not None:
-        data_files["validation"] = data_args.train_file
+        data_files["validation"] = data_args.validation_file
     extension = data_args.train_file.split(".")[-1]
     if extension == "txt":
         extension = "text"

examples/language-modeling/run_plm.py

@@ -198,7 +198,7 @@ def main():
         if data_args.train_file is not None:
             data_files["train"] = data_args.train_file
         if data_args.validation_file is not None:
-            data_files["validation"] = data_args.train_file
+            data_files["validation"] = data_args.validation_file
         extension = data_args.train_file.split(".")[-1]
         if extension == "txt":
             extension = "text"

examples/token-classification/utils_ner.py

@@ -66,14 +66,16 @@ class Split(Enum):
 
 
 class TokenClassificationTask:
-    def read_examples_from_file(self, data_dir, mode: Union[Split, str]) -> List[InputExample]:
+    @staticmethod
+    def read_examples_from_file(data_dir, mode: Union[Split, str]) -> List[InputExample]:
         raise NotImplementedError
 
-    def get_labels(self, path: str) -> List[str]:
+    @staticmethod
+    def get_labels(path: str) -> List[str]:
         raise NotImplementedError
 
+    @staticmethod
     def convert_examples_to_features(
-        self,
         examples: List[InputExample],
         label_list: List[str],
         max_seq_length: int,

model_cards/bionlp/bluebert_pubmed_uncased_L-12_H-768_A-12/README.md

@@ -0,0 +1,60 @@
+---
+language: 
+- en
+tags:
+- bluebert
+license: 
+- PUBLIC DOMAIN NOTICE
+datasets:
+- pubmed
+
+---
+
+# BlueBert-Base, Uncased, PubMed
+
+## Model description
+
+A BERT model pre-trained on PubMed abstracts
+
+## Intended uses & limitations
+
+#### How to use
+
+Please see https://github.com/ncbi-nlp/bluebert
+
+## Training data
+
+We provide [preprocessed PubMed texts](https://ftp.ncbi.nlm.nih.gov/pub/lu/Suppl/NCBI-BERT/pubmed_uncased_sentence_nltk.txt.tar.gz) that were used to pre-train the BlueBERT models. 
+The corpus contains ~4000M words extracted from the [PubMed ASCII code version](https://www.ncbi.nlm.nih.gov/research/bionlp/APIs/BioC-PubMed/). 
+
+Pre-trained model: https://huggingface.co/bert-base-uncased
+
+## Training procedure
+
+*  lowercasing the text
+*  removing speical chars `\x00`-`\x7F`
+*  tokenizing the text using the [NLTK Treebank tokenizer](https://www.nltk.org/_modules/nltk/tokenize/treebank.html)
+
+Below is a code snippet for more details.
+
+```python
+value = value.lower()
+value = re.sub(r'[\r\n]+', ' ', value)
+value = re.sub(r'[^\x00-\x7F]+', ' ', value)
+
+tokenized = TreebankWordTokenizer().tokenize(value)
+sentence = ' '.join(tokenized)
+sentence = re.sub(r"\s's\b", "'s", sentence)
+```
+
+### BibTeX entry and citation info
+
+```bibtex
+@InProceedings{peng2019transfer,
+  author    = {Yifan Peng and Shankai Yan and Zhiyong Lu},
+  title     = {Transfer Learning in Biomedical Natural Language Processing: An Evaluation of BERT and ELMo on Ten Benchmarking Datasets},
+  booktitle = {Proceedings of the 2019 Workshop on Biomedical Natural Language Processing (BioNLP 2019)},
+  year      = {2019},
+  pages     = {58--65},
+}
+```

model_cards/deepset/roberta-base-squad2-v2/README.md

@@ -0,0 +1,117 @@
+---
+datasets:
+- squad_v2
+---
+
+# roberta-base for QA 
+
+## Overview
+**Language model:** roberta-base  
+**Language:** English  
+**Downstream-task:** Extractive QA  
+**Training data:** SQuAD 2.0  
+**Eval data:** SQuAD 2.0  
+**Code:**  See [example](https://github.com/deepset-ai/FARM/blob/master/examples/question_answering.py) in [FARM](https://github.com/deepset-ai/FARM/blob/master/examples/question_answering.py)  
+**Infrastructure**: 4x Tesla v100
+
+## Hyperparameters
+
+```
+batch_size = 96
+n_epochs = 2
+base_LM_model = "roberta-base"
+max_seq_len = 386
+learning_rate = 3e-5
+lr_schedule = LinearWarmup
+warmup_proportion = 0.2
+doc_stride=128
+max_query_length=64
+``` 
+
+## Performance
+Evaluated on the SQuAD 2.0 dev set with the [official eval script](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/).
+
+```
+"exact": 79.97136359807968
+"f1": 83.00449234495325
+
+"total": 11873
+"HasAns_exact": 78.03643724696356
+"HasAns_f1": 84.11139298441825
+"HasAns_total": 5928
+"NoAns_exact": 81.90075693860386
+"NoAns_f1": 81.90075693860386
+"NoAns_total": 5945
+```
+
+## Usage
+
+### In Transformers
+```python
+from transformers.pipelines import pipeline
+from transformers.modeling_auto import AutoModelForQuestionAnswering
+from transformers.tokenization_auto import AutoTokenizer
+
+model_name = "deepset/roberta-base-squad2-v2"
+
+# a) Get predictions
+nlp = pipeline('question-answering', model=model_name, tokenizer=model_name)
+QA_input = {
+    'question': 'Why is model conversion important?',
+    'context': 'The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks.'
+}
+res = nlp(QA_input)
+
+# b) Load model & tokenizer
+model = AutoModelForQuestionAnswering.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+```
+
+### In FARM
+
+```python
+from farm.modeling.adaptive_model import AdaptiveModel
+from farm.modeling.tokenization import Tokenizer
+from farm.infer import Inferencer
+
+model_name = "deepset/roberta-base-squad2-v2"
+
+# a) Get predictions
+nlp = Inferencer.load(model_name, task_type="question_answering")
+QA_input = [{"questions": ["Why is model conversion important?"],
+             "text": "The option to convert models between FARM and transformers gives freedom to the user and let people easily switch between frameworks."}]
+res = nlp.inference_from_dicts(dicts=QA_input, rest_api_schema=True)
+
+# b) Load model & tokenizer
+model = AdaptiveModel.convert_from_transformers(model_name, device="cpu", task_type="question_answering")
+tokenizer = Tokenizer.load(model_name)
+```
+
+### In haystack
+For doing QA at scale (i.e. many docs instead of single paragraph), you can load the model also in [haystack](https://github.com/deepset-ai/haystack/):
+```python
+reader = FARMReader(model_name_or_path="deepset/roberta-base-squad2")
+# or 
+reader = TransformersReader(model="deepset/roberta-base-squad2",tokenizer="deepset/roberta-base-squad2")
+```
+
+
+## Authors
+Branden Chan: `branden.chan [at] deepset.ai`
+Timo Möller: `timo.moeller [at] deepset.ai`
+Malte Pietsch: `malte.pietsch [at] deepset.ai`
+Tanay Soni: `tanay.soni [at] deepset.ai`
+
+## About us
+![deepset logo](https://raw.githubusercontent.com/deepset-ai/FARM/master/docs/img/deepset_logo.png)
+
+We bring NLP to the industry via open source!  
+Our focus: Industry specific language models & large scale QA systems.  
+  
+Some of our work: 
+- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
+- [FARM](https://github.com/deepset-ai/FARM)
+- [Haystack](https://github.com/deepset-ai/haystack/)
+
+Get in touch:
+[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Website](https://deepset.ai)

model_cards/deepset/roberta-base-squad2/README.md

@@ -5,6 +5,8 @@ datasets:
 
 # roberta-base for QA 
 
+NOTE: This model has been superseded by deepset/roberta-base-squad2-v2. For an explanation of why, see [this github issue](https://github.com/deepset-ai/FARM/issues/552) from the FARM repository.
+
 ## Overview
 **Language model:** roberta-base  
 **Language:** English  
@@ -106,4 +108,3 @@ Some of our work:
 
 Get in touch:
 [Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Website](https://deepset.ai)
-

model_cards/distilbert-base-cased-README.md

@@ -1,3 +1,40 @@
 ---
+language: en
 license: apache-2.0
+datasets:
+- bookcorpus
+- wikipedia
 ---
+
+# DistilBERT base model (cased)
+
+This model is a distilled version of the [BERT base model](https://huggingface.co/bert-base-cased).
+It was introduced in [this paper](https://arxiv.org/abs/1910.01108).
+The code for the distillation process can be found
+[here](https://github.com/huggingface/transformers/tree/master/examples/distillation).
+This model is cased: it does make a difference between english and English.
+
+All the training details on the pre-training, the uses, limitations and potential biases are the same as for [DistilBERT-base-uncased](https://huggingface.co/distilbert-base-uncased).
+We highly encourage to check it if you want to know more.
+
+## Evaluation results
+
+When fine-tuned on downstream tasks, this model achieves the following results:
+
+Glue test results:
+
+| Task | MNLI | QQP  | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE  |
+|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|
+|      | 81.5 | 87.8 | 88.2 | 90.4  | 47.2 | 85.5  | 85.6 | 60.6 |
+
+### BibTeX entry and citation info
+
+```bibtex
+@article{Sanh2019DistilBERTAD,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.01108}
+}
+```

model_cards/distilbert-base-cased-distilled-squad-README.md

@@ -6,3 +6,8 @@ metrics:
 - squad
 license: apache-2.0
 ---
+
+# DistilBERT base cased distilled SQuAD
+
+This model is a fine-tune checkpoint of [DistilBERT-base-cased](https://huggingface.co/distilbert-base-cased), fine-tuned using (a second step of) knowledge distillation on SQuAD v1.1.
+This model reaches a F1 score of 87.1 on the dev set (for comparison, BERT bert-base-cased version reaches a F1 score of 88.7).

model_cards/distilbert-base-multilingual-cased-README.md

@@ -1,4 +1,35 @@
 ---
 language: multilingual
 license: apache-2.0
+datasets:
+- wikipedia
 ---
+
+# DistilBERT base multilingual model (cased)
+
+This model is a distilled version of the [BERT base multilingual model](bert-base-multilingual-cased). The code for the distillation process can be found
+[here](https://github.com/huggingface/transformers/tree/master/examples/distillation). This model is cased: it does make a difference between english and English.
+
+The model is trained on the concatenation of Wikipedia in 104 different languages listed [here](https://github.com/google-research/bert/blob/master/multilingual.md#list-of-languages).
+The model has 6 layers, 768 dimension and 12 heads, totalizing 134M parameters (compared to 177M parameters for mBERT-base).
+On average DistilmBERT is twice as fast as mBERT-base.
+
+We encourage to check [BERT base multilingual model](bert-base-multilingual-cased) to know more about usage, limitations and potential biases.
+
+| Model                        | English | Spanish | Chinese | German | Arabic  | Urdu |
+| :---:                        | :---:   | :---:   | :---:   | :---:  | :---:   | :---:|
+| mBERT base cased (computed)  | 82.1    | 74.6    | 69.1    | 72.3   | 66.4    | 58.5 |
+| mBERT base uncased (reported)| 81.4    | 74.3    | 63.8    | 70.5   | 62.1    | 58.3 |
+| DistilmBERT                  | 78.2    | 69.1    | 64.0    | 66.3   | 59.1    | 54.7 |
+
+### BibTeX entry and citation info
+
+```bibtex
+@article{Sanh2019DistilBERTAD,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.01108}
+}
+```

model_cards/distilbert-base-uncased-README.md

@@ -10,7 +10,7 @@ datasets:
 
 # DistilBERT base model (uncased)
 
-This model is a distilled version of the [BERT base mode](https://huggingface.co/distilbert-base-uncased). It was
+This model is a distilled version of the [BERT base model](https://huggingface.co/bert-base-uncased). It was
 introduced in [this paper](https://arxiv.org/abs/1910.01108). The code for the distillation process can be found
 [here](https://github.com/huggingface/transformers/tree/master/examples/distillation). This model is uncased: it does
 not make a difference between english and English.
@@ -102,7 +102,7 @@ output = model(encoded_input)
 
 Even if the training data used for this model could be characterized as fairly neutral, this model can have biased
 predictions. It also inherits some of
-[the bias of its teacher model](https://huggingface.co/bert-base-uncased#limitations-and-bias). 
+[the bias of its teacher model](https://huggingface.co/bert-base-uncased#limitations-and-bias).
 
 ```python
 >>> from transformers import pipeline
@@ -196,9 +196,9 @@ When fine-tuned on downstream tasks, this model achieves the following results:
 
 Glue test results:
 
-| Task | MNLI | QQP  | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE  | Average |
-|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|:-------:|
-|      | 82.2 | 88.5 | 89.2 | 91.3  | 51.3 | 85.8  | 87.5 | 59.9 | 77.0    |
+| Task | MNLI | QQP  | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE  |
+|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|
+|      | 82.2 | 88.5 | 89.2 | 91.3  | 51.3 | 85.8  | 87.5 | 59.9 |
 
 
 ### BibTeX entry and citation info

model_cards/distilbert-base-uncased-distilled-squad-README.md

@@ -1,4 +1,5 @@
 ---
+language: en
 datasets:
 - squad
 widget:
@@ -8,3 +9,8 @@ widget:
   context: "The Amazon rainforest (Portuguese: Floresta Amazônica or Amazônia; Spanish: Selva Amazónica, Amazonía or usually Amazonia; French: Forêt amazonienne; Dutch: Amazoneregenwoud), also known in English as Amazonia or the Amazon Jungle, is a moist broadleaf forest that covers most of the Amazon basin of South America. This basin encompasses 7,000,000 square kilometres (2,700,000 sq mi), of which 5,500,000 square kilometres (2,100,000 sq mi) are covered by the rainforest. This region includes territory belonging to nine nations. The majority of the forest is contained within Brazil, with 60% of the rainforest, followed by Peru with 13%, Colombia with 10%, and with minor amounts in Venezuela, Ecuador, Bolivia, Guyana, Suriname and French Guiana. States or departments in four nations contain \"Amazonas\" in their names. The Amazon represents over half of the planet's remaining rainforests, and comprises the largest and most biodiverse tract of tropical rainforest in the world, with an estimated 390 billion individual trees divided into 16,000 species."
 license: apache-2.0
 ---
+
+# DistilBERT base uncased distilled SQuAD
+
+This model is a fine-tune checkpoint of [DistilBERT-base-uncased](https://huggingface.co/distilbert-base-uncased), fine-tuned using (a second step of) knowledge distillation on SQuAD v1.1.
+This model reaches a F1 score of 86.9 on the dev set (for comparison, Bert bert-base-uncased version reaches a F1 score of 88.5).

model_cards/distilbert-base-uncased-finetuned-sst-2-english-README.md

@@ -1,3 +1,19 @@
 ---
+language: en
 license: apache-2.0
+datasets:
+- sst-2
 ---
+
+# DistilBERT base uncased finetuned SST-2
+
+This model is a fine-tune checkpoint of [DistilBERT-base-uncased](https://huggingface.co/distilbert-base-uncased), fine-tuned on SST-2.
+This model reaches an accuracy of 91.3 on the dev set (for comparison, Bert bert-base-uncased version reaches an accuracy of 92.7).
+
+# Fine-tuning hyper-parameters
+
+- learning_rate = 1e-5
+- batch_size = 32
+- warmup = 600
+- max_seq_length = 128
+- num_train_epochs = 3.0

model_cards/distilgpt2-README.md

@@ -1,10 +1,21 @@
 ---
+language: en
 tags:
 - exbert
 
 license: apache-2.0
+datasets:
+- openwebtext
 ---
 
+# DistilGPT2
+
+DistilGPT2 English language model pretrained with the supervision of [GPT2](https://huggingface.co/gpt2) (the smallest version of GPT2) on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset. The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 124M parameters for GPT2). On average, DistilGPT2 is two times faster than GPT2.
+
+On the [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/) benchmark, GPT2 reaches a perplexity on the test set of 16.3 compared to 21.1 for DistilGPT2 (after fine-tuning on the train set).
+
+We encourage to check [GPT2](https://huggingface.co/gpt2) to know more about usage, limitations and potential biases.
+
 <a href="https://huggingface.co/exbert/?model=distilgpt2">
 	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>

model_cards/distilroberta-base-README.md

@@ -1,10 +1,50 @@
 ---
+language: en
 tags:
 - exbert
 
 license: apache-2.0
+datasets:
+- openwebtext
 ---
 
+# DistilRoBERTa base model
+
+This model is a distilled version of the [RoBERTa-base model](https://huggingface.co/roberta-base). It follows the same training procedure as [DistilBERT](https://huggingface.co/distilbert-base-uncased).
+The code for the distillation process can be found [here](https://github.com/huggingface/transformers/tree/master/examples/distillation).
+This model is case-sensitive: it makes a difference between english and English.
+
+The model has 6 layers, 768 dimension and 12 heads, totalizing 82M parameters (compared to 125M parameters for RoBERTa-base).
+On average DistilRoBERTa is twice as fast as Roberta-base.
+
+We encourage to check [RoBERTa-base model](https://huggingface.co/roberta-base) to know more about usage, limitations and potential biases.
+
+## Training data
+
+DistilRoBERTa was pre-trained on [OpenWebTextCorpus](https://skylion007.github.io/OpenWebTextCorpus/), a reproduction of OpenAI's WebText dataset (it is ~4 times less training data than the teacher RoBERTa).
+
+## Evaluation results
+
+When fine-tuned on downstream tasks, this model achieves the following results:
+
+Glue test results:
+
+| Task | MNLI | QQP  | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE  |
+|:----:|:----:|:----:|:----:|:-----:|:----:|:-----:|:----:|:----:|
+|      | 84.0 | 89.4 | 90.8 | 92.5  | 59.3 | 88.3  | 86.6 | 67.9 |
+
+### BibTeX entry and citation info
+
+```bibtex
+@article{Sanh2019DistilBERTAD,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
+  journal={ArXiv},
+  year={2019},
+  volume={abs/1910.01108}
+}
+```
+
 <a href="https://huggingface.co/exbert/?model=distilroberta-base">
 	<img width="300px" src="https://cdn-media.huggingface.co/exbert/button.png">
 </a>

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

@@ -0,0 +1,66 @@
+---
+language: en
+datasets:
+- qasc
+---
+
+# T5-base fine-tuned on QASC 
+
+[Google's T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) fine-tuned on [QASC](https://allenai.org/data/qasc) for **QA** (via *sentence composition*) downstream task.
+
+## Details of T5
+
+The **T5** model was presented in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf) by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu* in Here the abstract:
+
+Transfer learning, where a model is first pre-trained on a data-rich task before being fine-tuned on a downstream task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration with scale and our new “Colossal Clean Crawled Corpus”, we achieve state-of-the-art results on many benchmarks covering summarization, question answering, text classification, and more. To facilitate future work on transfer learning for NLP, we release our dataset, pre-trained models, and code.
+
+![model image](https://i.imgur.com/jVFMMWR.png)
+
+
+## Details of the dataset 📚 
+
+**Question Answering via Sentence Composition** (QASC) is a question-answering dataset with a focus on sentence composition. It consists of 9,980 8-way multiple-choice questions about grade school science (8,134 train, 926 dev, 920 test), and comes with a corpus of 17M sentences.
+
+
+## Model fine-tuning 🏋️‍
+
+The training script is a slightly modified version of [this  awesome one](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb) by [Suraj Patil](https://twitter.com/psuraj28). The **context** passed to the *encoder* is the combination of the 2 *facts* (`fact1` and `fact2`). The **question** is just the `formatted_question` field. The **answer** passed to the *decoder* is the`text` right answer instead of the `label` (A, B, C... See `choices` field). More details about the dataset format/fields [here](https://huggingface.co/nlp/viewer/?dataset=qasc)
+
+## Metrics on validation set 📋
+
+| Metric | Score |
+|--------|-------|
+|Accuracy (EM) | **97.73**|
+
+
+## Model in Action 🚀
+
+```python
+from transformers import AutoModelWithLMHead, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("mrm8488/t5-base-finetuned-qasc")
+model = AutoModelWithLMHead.from_pretrained("mrm8488/t5-base-finetuned-qasc")
+
+def get_response(question, context, max_length=64):
+  input_text = 'question: %s  context: %s' % (question, context)
+  features = tokenizer([input_text], return_tensors='pt')
+
+  output = model.generate(input_ids=features['input_ids'], 
+               attention_mask=features['attention_mask'],
+               max_length=max_length)
+
+  return tokenizer.decode(output[0])
+  
+fact_1 = 'a watch is used for measuring time'
+fact_2 = 'Times are measured in seconds.'
+context = fact_1 + ' ' + fact_2
+question = 'What can be used to measure seconds? (A) Watch (B) seconds (C) fluid (D) Ruler (E) goggles (F) glasses (G) Drill (H) Scale'
+
+get_response(question, context)
+
+# output: 'Watch'
+```
+
+> 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

setup.py

@@ -105,7 +105,7 @@ extras["sentencepiece"] = ["sentencepiece==0.1.91"]
 extras["retrieval"] = ["faiss-cpu", "datasets"]
 extras["testing"] = ["pytest", "pytest-xdist", "timeout-decorator", "parameterized", "psutil"] + extras["retrieval"]
 # sphinx-rtd-theme==0.5.0 introduced big changes in the style.
-extras["docs"] = ["recommonmark", "sphinx", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3", "sphinx-copybutton"]
+extras["docs"] = ["recommonmark", "sphinx==3.2.1", "sphinx-markdown-tables", "sphinx-rtd-theme==0.4.3", "sphinx-copybutton"]
 extras["quality"] = ["black >= 20.8b1", "isort >= 5.5.4", "flake8 >= 3.8.3"]
 
 

src/transformers/__init__.py

@@ -281,7 +281,6 @@ if is_torch_available():
     from .data.data_collator import (
         DataCollator,
         DataCollatorForLanguageModeling,
-        DataCollatorForNextSentencePrediction,
         DataCollatorForPermutationLanguageModeling,
         DataCollatorForSOP,
         DataCollatorForTokenClassification,

src/transformers/data/data_collator.py

@@ -1,4 +1,5 @@
 import random
+import warnings
 from dataclasses import dataclass
 from typing import Any, Callable, Dict, List, NewType, Optional, Tuple, Union
 
@@ -175,72 +176,111 @@ class DataCollatorForTokenClassification:
         return batch
 
 
+def _collate_batch(examples, tokenizer):
+    """Collate `examples` into a batch, using the information in `tokenizer` for padding if necessary."""
+    # Tensorize if necessary.
+    if isinstance(examples[0], (list, tuple)):
+        examples = [torch.tensor(e, dtype=torch.long) for e in examples]
+
+    # Check if padding is necessary.
+    length_of_first = examples[0].size(0)
+    are_tensors_same_length = all(x.size(0) == length_of_first for x in examples)
+    if are_tensors_same_length:
+        return torch.stack(examples, dim=0)
+
+    # If yes, check if we have a `pad_token`.
+    if tokenizer._pad_token is None:
+        raise ValueError(
+            "You are attempting to pad samples but the tokenizer you are using"
+            f" ({tokenizer.__class__.__name__}) does not have a pad token."
+        )
+
+    # Creating the full tensor and filling it with our data.
+    max_length = max(x.size(0) for x in examples)
+    result = examples[0].new_full([len(examples), max_length], tokenizer.pad_token_id)
+    for i, example in enumerate(examples):
+        if tokenizer.padding_side == "right":
+            result[i, : example.shape[0]] = example
+        else:
+            result[i, -example.shape[0] :] = example
+    return result
+
+
 @dataclass
 class DataCollatorForLanguageModeling:
     """
-    Data collator used for language modeling.
+    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
+    are not all of the same length.
 
-    - collates batches of tensors, honoring their tokenizer's pad_token
-    - preprocesses batches for masked language modeling
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        mlm (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not to use masked language modeling. If set to :obj:`False`, the labels are the same as the
+            inputs with the padding tokens ignored (by setting them to -100). Otherwise, the labels are -100 for
+            non-masked tokens and the value to predict for the masked token.
+        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
+            The probability with which to (randomly) mask tokens in the input, when :obj:`mlm` is set to :obj:`True`.
+
+    .. note::
+
+        For best performance, this data collator should be used with a dataset having items that are dictionaries or
+        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
+        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
+        argument :obj:`return_special_tokens_mask=True`.
     """
 
     tokenizer: PreTrainedTokenizerBase
     mlm: bool = True
     mlm_probability: float = 0.15
 
+    def __post_init__(self):
+        if self.mlm and self.tokenizer.mask_token is None:
+            raise ValueError(
+                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
+                "You should pass `mlm=False` to train on causal language modeling instead."
+            )
+
     def __call__(
         self, examples: List[Union[List[int], torch.Tensor, Dict[str, torch.Tensor]]]
     ) -> Dict[str, torch.Tensor]:
+        # Handle dict or lists with proper padding and conversion to tensor.
         if isinstance(examples[0], (dict, BatchEncoding)):
-            examples = [e["input_ids"] for e in examples]
-        batch = self._tensorize_batch(examples)
-        if self.mlm:
-            inputs, labels = self.mask_tokens(batch)
-            return {"input_ids": inputs, "labels": labels}
+            batch = self.tokenizer.pad(examples, return_tensors="pt")
         else:
-            labels = batch.clone().detach()
+            batch = {"input_ids": _collate_batch(examples, self.tokenizer)}
+
+        # If special token mask has been preprocessed, pop it from the dict.
+        special_tokens_mask = batch.pop("special_tokens_mask", None)
+        if self.mlm:
+            batch["input_ids"], batch["labels"] = self.mask_tokens(
+                batch["input_ids"], special_tokens_mask=special_tokens_mask
+            )
+        else:
+            labels = batch["input_ids"]
             if self.tokenizer.pad_token_id is not None:
                 labels[labels == self.tokenizer.pad_token_id] = -100
-            return {"input_ids": batch, "labels": labels}
+            batch["labels"] = labels
+        return batch
 
-    def _tensorize_batch(
-        self, examples: List[Union[List[int], torch.Tensor, Dict[str, torch.Tensor]]]
-    ) -> torch.Tensor:
-        # In order to accept both lists of lists and lists of Tensors
-        if isinstance(examples[0], (list, tuple)):
-            examples = [torch.tensor(e, dtype=torch.long) for e in examples]
-        length_of_first = examples[0].size(0)
-        are_tensors_same_length = all(x.size(0) == length_of_first for x in examples)
-        if are_tensors_same_length:
-            return torch.stack(examples, dim=0)
-        else:
-            if self.tokenizer._pad_token is None:
-                raise ValueError(
-                    "You are attempting to pad samples but the tokenizer you are using"
-                    f" ({self.tokenizer.__class__.__name__}) does not have one."
-                )
-            return pad_sequence(examples, batch_first=True, padding_value=self.tokenizer.pad_token_id)
-
-    def mask_tokens(self, inputs: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def mask_tokens(
+        self, inputs: torch.Tensor, special_tokens_mask: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
         """
         Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
         """
-
-        if self.tokenizer.mask_token is None:
-            raise ValueError(
-                "This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the --mlm flag if you want to use this tokenizer."
-            )
-
         labels = inputs.clone()
-        # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
+        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
         probability_matrix = torch.full(labels.shape, self.mlm_probability)
-        special_tokens_mask = [
-            self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
-        ]
-        probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
-        if self.tokenizer._pad_token is not None:
-            padding_mask = labels.eq(self.tokenizer.pad_token_id)
-            probability_matrix.masked_fill_(padding_mask, value=0.0)
+        if special_tokens_mask is None:
+            special_tokens_mask = [
+                self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
+            ]
+            special_tokens_mask = torch.tensor(special_tokens_mask, dtype=torch.bool)
+        else:
+            special_tokens_mask = special_tokens_mask.bool()
+
+        probability_matrix.masked_fill_(special_tokens_mask, value=0.0)
         masked_indices = torch.bernoulli(probability_matrix).bool()
         labels[~masked_indices] = -100  # We only compute loss on masked tokens
 
@@ -385,9 +425,16 @@ class DataCollatorForSOP(DataCollatorForLanguageModeling):
     - preprocesses batches for both masked language modeling and sentence order prediction
     """
 
+    def __init__(self, *args, **kwargs):
+        warnings.warn(
+            "DataCollatorForSOP is deprecated and will be removed in a future version, you can now use "
+            "DataCollatorForLanguageModeling instead.",
+            FutureWarning,
+        )
+
     def __call__(self, examples: List[Dict[str, torch.Tensor]]) -> Dict[str, torch.Tensor]:
         input_ids = [example["input_ids"] for example in examples]
-        input_ids = self._tensorize_batch(input_ids)
+        input_ids = _collate_batch(input_ids, self.tokenizer)
         input_ids, labels, attention_mask = self.mask_tokens(input_ids)
 
         token_type_ids = [example["token_type_ids"] for example in examples]
@@ -582,136 +629,3 @@ class DataCollatorForPermutationLanguageModeling:
             ) & masked_indices[i]
 
         return inputs.long(), perm_mask, target_mapping, labels.long()
-
-
-@dataclass
-class DataCollatorForNextSentencePrediction:
-    """
-    Data collator used for next sentence prediction. - collates examples which contains pre-generated negative examples
-    - preprocesses batches for masked language modeling
-    """
-
-    tokenizer: PreTrainedTokenizerBase
-    mlm: bool = True
-    block_size: int = 512
-    short_seq_probability: float = 0.1
-    nsp_probability: float = 0.5
-    mlm_probability: float = 0.15
-
-    def __call__(self, examples: List[Dict[str, torch.Tensor]]) -> Dict[str, torch.Tensor]:
-        """
-        The input should contain negative examples, :class:`~transformers.DataCollatorForNextSentencePrediction` will
-        not generate any negative examples
-
-        Args:
-            examples (:obj:`List[Dict]`): Each dictionary should have the following keys:
-
-                  - ``tokens_a``: A sequence of tokens, which should appear before ``tokens_b`` in the text.
-                  - ``tokens_b``: A sequence of tokens, which should appear after ``tokens_a`` in the text.
-                  - ``is_random_next``: 1 if this pair is generated randomly, else 0.
-        """
-
-        tokens_a = [e["tokens_a"] for e in examples]
-        tokens_b = [e["tokens_b"] for e in examples]
-        nsp_labels = [1 if e["is_random_next"] else 0 for e in examples]
-
-        input_ids = []
-        segment_ids = []
-        attention_masks = []
-
-        assert len(tokens_a) == len(tokens_b)
-        for i in range(len(tokens_a)):
-            input_id, attention_mask, segment_id = self.create_features_from_example(tokens_a[i], tokens_b[i])
-            input_ids.append(input_id)
-            segment_ids.append(segment_id)
-            attention_masks.append(attention_mask)
-        if self.mlm:
-            input_ids, mlm_labels = self.mask_tokens(self._tensorize_batch(input_ids))
-        else:
-            input_ids = self._tensorize_batch(input_ids)
-
-        result = {
-            "input_ids": input_ids,
-            "attention_mask": self._tensorize_batch(attention_masks),
-            "token_type_ids": self._tensorize_batch(segment_ids),
-            "labels": mlm_labels if self.mlm else None,
-            "next_sentence_label": torch.tensor(nsp_labels),
-        }
-        return result
-
-    def _tensorize_batch(self, examples: List[torch.Tensor]) -> torch.Tensor:
-        length_of_first = examples[0].size(0)
-        are_tensors_same_length = all(x.size(0) == length_of_first for x in examples)
-        if are_tensors_same_length:
-            return torch.stack(examples, dim=0)
-        else:
-            if self.tokenizer._pad_token is None:
-                raise ValueError(
-                    "You are attempting to pad samples but the tokenizer you are using"
-                    f" ({self.tokenizer.__class__.__name__}) does not have one."
-                )
-            return pad_sequence(examples, batch_first=True, padding_value=self.tokenizer.pad_token_id)
-
-    def create_features_from_example(self, tokens_a, tokens_b):
-        """Creates examples for a single document."""
-
-        max_num_tokens = self.block_size - self.tokenizer.num_special_tokens_to_add(pair=True)
-
-        tokens_a, tokens_b, _ = self.tokenizer.truncate_sequences(
-            tokens_a,
-            tokens_b,
-            num_tokens_to_remove=len(tokens_a) + len(tokens_b) - max_num_tokens,
-            truncation_strategy="longest_first",
-        )
-
-        input_id = self.tokenizer.build_inputs_with_special_tokens(tokens_a, tokens_b)
-        attention_mask = [1] * len(input_id)
-        segment_id = self.tokenizer.create_token_type_ids_from_sequences(tokens_a, tokens_b)
-        assert len(input_id) <= self.block_size
-
-        # pad
-        while len(input_id) < self.block_size:
-            input_id.append(0)
-            attention_mask.append(0)
-            segment_id.append(0)
-
-        input_id = torch.tensor(input_id)
-        attention_mask = torch.tensor(attention_mask)
-        segment_id = torch.tensor(segment_id)
-
-        return input_id, attention_mask, segment_id
-
-    def mask_tokens(self, inputs: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
-        """
-        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
-        """
-
-        if self.tokenizer.mask_token is None:
-            raise ValueError(
-                "This tokenizer does not have a mask token which is necessary for masked language modeling. Remove the --mlm flag if you want to use this tokenizer."
-            )
-
-        labels = inputs.clone()
-        # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
-        probability_matrix = torch.full(labels.shape, self.mlm_probability)
-        special_tokens_mask = [
-            self.tokenizer.get_special_tokens_mask(val, already_has_special_tokens=True) for val in labels.tolist()
-        ]
-        probability_matrix.masked_fill_(torch.tensor(special_tokens_mask, dtype=torch.bool), value=0.0)
-        if self.tokenizer._pad_token is not None:
-            padding_mask = labels.eq(self.tokenizer.pad_token_id)
-            probability_matrix.masked_fill_(padding_mask, value=0.0)
-        masked_indices = torch.bernoulli(probability_matrix).bool()
-        labels[~masked_indices] = -100  # We only compute loss on masked tokens
-
-        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
-        indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
-        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)
-
-        # 10% of the time, we replace masked input tokens with random word
-        indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
-        random_words = torch.randint(len(self.tokenizer), labels.shape, dtype=torch.long)
-        inputs[indices_random] = random_words[indices_random]
-
-        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
-        return inputs, labels

src/transformers/data/datasets/language_modeling.py

@@ -3,6 +3,7 @@ import os
 import pickle
 import random
 import time
+import warnings
 from typing import Dict, List, Optional
 
 import torch
@@ -17,6 +18,11 @@ from ...utils import logging
 logger = logging.get_logger(__name__)
 
 
+DEPRECATION_WARNING = (
+    "This dataset will be removed from the library soon, preprocessing should be handled with the 🤗 Datasets library."
+)
+
+
 class TextDataset(Dataset):
     """
     This will be superseded by a framework-agnostic approach soon.
@@ -30,6 +36,7 @@ class TextDataset(Dataset):
         overwrite_cache=False,
         cache_dir: Optional[str] = None,
     ):
+        warnings.warn(DEPRECATION_WARNING, FutureWarning)
         assert os.path.isfile(file_path), f"Input file path {file_path} not found"
 
         block_size = block_size - tokenizer.num_special_tokens_to_add(pair=False)
@@ -94,6 +101,7 @@ class LineByLineTextDataset(Dataset):
     """
 
     def __init__(self, tokenizer: PreTrainedTokenizer, file_path: str, block_size: int):
+        warnings.warn(DEPRECATION_WARNING, FutureWarning)
         assert os.path.isfile(file_path), f"Input file path {file_path} not found"
         # Here, we do not cache the features, operating under the assumption
         # that we will soon use fast multithreaded tokenizers from the
@@ -120,6 +128,7 @@ class LineByLineWithRefDataset(Dataset):
     """
 
     def __init__(self, tokenizer: PreTrainedTokenizer, file_path: str, block_size: int, ref_path: str):
+        warnings.warn(DEPRECATION_WARNING, FutureWarning)
         assert os.path.isfile(file_path), f"Input file path {file_path} not found"
         assert os.path.isfile(ref_path), f"Ref file path {file_path} not found"
         # Here, we do not cache the features, operating under the assumption
@@ -156,6 +165,7 @@ class LineByLineWithSOPTextDataset(Dataset):
     """
 
     def __init__(self, tokenizer: PreTrainedTokenizer, file_dir: str, block_size: int):
+        warnings.warn(DEPRECATION_WARNING, FutureWarning)
         assert os.path.isdir(file_dir)
         logger.info(f"Creating features from dataset file folder at {file_dir}")
         self.examples = []
@@ -305,6 +315,7 @@ class TextDatasetForNextSentencePrediction(Dataset):
         short_seq_probability=0.1,
         nsp_probability=0.5,
     ):
+        warnings.warn(DEPRECATION_WARNING, FutureWarning)
         assert os.path.isfile(file_path), f"Input file path {file_path} not found"
 
         self.block_size = block_size - tokenizer.num_special_tokens_to_add(pair=True)
@@ -449,9 +460,18 @@ class TextDatasetForNextSentencePrediction(Dataset):
                     assert len(tokens_a) >= 1
                     assert len(tokens_b) >= 1
 
-                    self.examples.append(
-                        {"tokens_a": tokens_a, "tokens_b": tokens_b, "is_random_next": is_random_next}
-                    )
+                    # add special tokens
+                    input_ids = self.tokenizer.build_inputs_with_special_tokens(tokens_a, tokens_b)
+                    # add token type ids, 0 for sentence a, 1 for sentence b
+                    token_type_ids = self.tokenizer.create_token_type_ids_from_sequences(tokens_a, tokens_b)
+
+                    example = {
+                        "input_ids": torch.tensor(input_ids, dtype=torch.long),
+                        "token_type_ids": torch.tensor(token_type_ids, dtype=torch.long),
+                        "next_sentence_label": torch.tensor(1 if is_random_next else 0, dtype=torch.long),
+                    }
+
+                    self.examples.append(example)
 
                 current_chunk = []
                 current_length = 0

src/transformers/pipelines.py

@@ -22,6 +22,7 @@ import sys
 import uuid
 import warnings
 from abc import ABC, abstractmethod
+from collections.abc import Iterable
 from contextlib import contextmanager
 from os.path import abspath, exists
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
@@ -1597,55 +1598,52 @@ class QuestionAnsweringArgumentHandler(ArgumentHandler):
     command-line supplied arguments.
     """
 
+    def normalize(self, item):
+        if isinstance(item, SquadExample):
+            return item
+        elif isinstance(item, dict):
+            for k in ["question", "context"]:
+                if k not in item:
+                    raise KeyError("You need to provide a dictionary with keys {question:..., context:...}")
+                elif item[k] is None:
+                    raise ValueError("`{}` cannot be None".format(k))
+                elif isinstance(item[k], str) and len(item[k]) == 0:
+                    raise ValueError("`{}` cannot be empty".format(k))
+
+            return QuestionAnsweringPipeline.create_sample(**item)
+        raise ValueError("{} argument needs to be of type (SquadExample, dict)".format(item))
+
     def __call__(self, *args, **kwargs):
-        # Position args, handling is sensibly the same as X and data, so forwarding to avoid duplicating
+        # Detect where the actual inputs are
         if args is not None and len(args) > 0:
             if len(args) == 1:
-                kwargs["X"] = args[0]
+                inputs = args[0]
+            elif len(args) == 2 and {type(el) for el in args} == {str}:
+                inputs = [{"question": args[0], "context": args[1]}]
             else:
-                kwargs["X"] = list(args)
-
+                inputs = list(args)
         # Generic compatibility with sklearn and Keras
         # Batched data
-        if "X" in kwargs or "data" in kwargs:
-            inputs = kwargs["X"] if "X" in kwargs else kwargs["data"]
-
-            if isinstance(inputs, dict):
-                inputs = [inputs]
-            else:
-                # Copy to avoid overriding arguments
-                inputs = [i for i in inputs]
-
-            for i, item in enumerate(inputs):
-                if isinstance(item, dict):
-                    if any(k not in item for k in ["question", "context"]):
-                        raise KeyError("You need to provide a dictionary with keys {question:..., context:...}")
-
-                    inputs[i] = QuestionAnsweringPipeline.create_sample(**item)
-
-                elif not isinstance(item, SquadExample):
-                    raise ValueError(
-                        "{} argument needs to be of type (list[SquadExample | dict], SquadExample, dict)".format(
-                            "X" if "X" in kwargs else "data"
-                        )
-                    )
-
-            # Tabular input
+        elif "X" in kwargs:
+            inputs = kwargs["X"]
+        elif "data" in kwargs:
+            inputs = kwargs["data"]
         elif "question" in kwargs and "context" in kwargs:
-            if isinstance(kwargs["question"], str):
-                kwargs["question"] = [kwargs["question"]]
-
-            if isinstance(kwargs["context"], str):
-                kwargs["context"] = [kwargs["context"]]
-
-            inputs = [
-                QuestionAnsweringPipeline.create_sample(q, c) for q, c in zip(kwargs["question"], kwargs["context"])
-            ]
+            inputs = [{"question": kwargs["question"], "context": kwargs["context"]}]
         else:
             raise ValueError("Unknown arguments {}".format(kwargs))
 
-        if not isinstance(inputs, list):
+        # Normalize inputs
+        if isinstance(inputs, dict):
             inputs = [inputs]
+        elif isinstance(inputs, Iterable):
+            # Copy to avoid overriding arguments
+            inputs = [i for i in inputs]
+        else:
+            raise ValueError("Invalid arguments {}".format(inputs))
+
+        for i, item in enumerate(inputs):
+            inputs[i] = self.normalize(item)
 
         return inputs
 

src/transformers/tokenization_blenderbot.py

@@ -160,7 +160,7 @@ class BlenderbotSmallTokenizer(PreTrainedTokenizer):
             return self.cache[token]
         token = re.sub("([.,!?()])", r" \1", token)
         token = re.sub("(')", r" \1 ", token)
-        token = re.sub("\s{2,}", " ", token)
+        token = re.sub(r"\s{2,}", " ", token)
         if "\n" in token:
             token = token.replace("\n", " __newln__")
 

src/transformers/utils/dummy_pt_objects.py

@@ -26,11 +26,6 @@ class DataCollatorForLanguageModeling:
         requires_pytorch(self)
 
 
-class DataCollatorForNextSentencePrediction:
-    def __init__(self, *args, **kwargs):
-        requires_pytorch(self)
-
-
 class DataCollatorForPermutationLanguageModeling:
     def __init__(self, *args, **kwargs):
         requires_pytorch(self)

templates/adding_a_new_example_script/{{cookiecutter.directory_name}}/run_{{cookiecutter.example_shortcut}}.py

@@ -205,7 +205,7 @@ def main():
         if data_args.train_file is not None:
             data_files["train"] = data_args.train_file
         if data_args.validation_file is not None:
-            data_files["validation"] = data_args.train_file
+            data_files["validation"] = data_args.validation_file
         extension = data_args.train_file.split(".")[-1]
         if extension == "txt":
             extension = "text"

tests/test_data_collator.py

@@ -12,9 +12,7 @@ if is_torch_available():
 
     from transformers import (
         DataCollatorForLanguageModeling,
-        DataCollatorForNextSentencePrediction,
         DataCollatorForPermutationLanguageModeling,
-        DataCollatorForSOP,
         DataCollatorForTokenClassification,
         DataCollatorWithPadding,
         default_data_collator,
@@ -201,13 +199,16 @@ class DataCollatorIntegrationTest(unittest.TestCase):
 
     def test_nsp(self):
         tokenizer = BertTokenizer(self.vocab_file)
-        features = [{"tokens_a": [0, 1, 2, 3, 4], "tokens_b": [0, 1, 2, 3, 4], "is_random_next": i} for i in range(2)]
-        data_collator = DataCollatorForNextSentencePrediction(tokenizer)
+        features = [
+            {"input_ids": [0, 1, 2, 3, 4], "token_type_ids": [0, 1, 2, 3, 4], "next_sentence_label": i}
+            for i in range(2)
+        ]
+        data_collator = DataCollatorForLanguageModeling(tokenizer)
         batch = data_collator(features)
 
-        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
-        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
+        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
         self.assertEqual(batch["next_sentence_label"].shape, torch.Size((2,)))
 
     def test_sop(self):
@@ -216,11 +217,11 @@ class DataCollatorIntegrationTest(unittest.TestCase):
             {
                 "input_ids": torch.tensor([0, 1, 2, 3, 4]),
                 "token_type_ids": torch.tensor([0, 1, 2, 3, 4]),
-                "sentence_order_label": torch.tensor(i),
+                "sentence_order_label": i,
             }
             for i in range(2)
         ]
-        data_collator = DataCollatorForSOP(tokenizer)
+        data_collator = DataCollatorForLanguageModeling(tokenizer)
         batch = data_collator(features)
 
         self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))

tests/test_generation_utils.py

@@ -140,10 +140,6 @@ class GenerationTesterMixin:
             # check `generate()` and `greedy_search()` are equal
             kwargs = {}
             if model.config.is_encoder_decoder:
-                encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
-                    model, input_ids, attention_mask
-                )
-                kwargs["encoder_outputs"] = encoder_outputs
                 max_length = 4
 
             output_ids_generate = model.generate(
@@ -154,6 +150,13 @@ class GenerationTesterMixin:
                 max_length=max_length,
                 **logits_process_kwargs,
             )
+
+            if model.config.is_encoder_decoder:
+                encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
+                    model, input_ids, attention_mask
+                )
+                kwargs["encoder_outputs"] = encoder_outputs
+
             with torch.no_grad():
                 output_ids_greedy = model.greedy_search(
                     input_ids,

tests/test_pipelines_question_answering.py

@@ -1,6 +1,7 @@
 import unittest
 
-from transformers.pipelines import Pipeline
+from transformers.data.processors.squad import SquadExample
+from transformers.pipelines import Pipeline, QuestionAnsweringArgumentHandler
 
 from .test_pipelines_common import CustomInputPipelineCommonMixin
 
@@ -43,5 +44,116 @@ class QAPipelineTests(CustomInputPipelineCommonMixin, unittest.TestCase):
             for key in output_keys:
                 self.assertIn(key, result)
         for bad_input in invalid_inputs:
-            self.assertRaises(Exception, nlp, bad_input)
-        self.assertRaises(Exception, nlp, invalid_inputs)
+            self.assertRaises(ValueError, nlp, bad_input)
+        self.assertRaises(ValueError, nlp, invalid_inputs)
+
+    def test_argument_handler(self):
+        qa = QuestionAnsweringArgumentHandler()
+
+        Q = "Where was HuggingFace founded ?"
+        C = "HuggingFace was founded in Paris"
+
+        normalized = qa(Q, C)
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa(question=Q, context=C)
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa(question=Q, context=C)
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa({"question": Q, "context": C})
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa([{"question": Q, "context": C}])
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa([{"question": Q, "context": C}, {"question": Q, "context": C}])
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 2)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa(X={"question": Q, "context": C})
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa(X=[{"question": Q, "context": C}])
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+        normalized = qa(data={"question": Q, "context": C})
+        self.assertEqual(type(normalized), list)
+        self.assertEqual(len(normalized), 1)
+        self.assertEqual({type(el) for el in normalized}, {SquadExample})
+
+    def test_argument_handler_error_handling(self):
+        qa = QuestionAnsweringArgumentHandler()
+
+        Q = "Where was HuggingFace founded ?"
+        C = "HuggingFace was founded in Paris"
+
+        with self.assertRaises(KeyError):
+            qa({"context": C})
+        with self.assertRaises(KeyError):
+            qa({"question": Q})
+        with self.assertRaises(KeyError):
+            qa([{"context": C}])
+        with self.assertRaises(ValueError):
+            qa(None, C)
+        with self.assertRaises(ValueError):
+            qa("", C)
+        with self.assertRaises(ValueError):
+            qa(Q, None)
+        with self.assertRaises(ValueError):
+            qa(Q, "")
+
+        with self.assertRaises(ValueError):
+            qa(question=None, context=C)
+        with self.assertRaises(ValueError):
+            qa(question="", context=C)
+        with self.assertRaises(ValueError):
+            qa(question=Q, context=None)
+        with self.assertRaises(ValueError):
+            qa(question=Q, context="")
+
+        with self.assertRaises(ValueError):
+            qa({"question": None, "context": C})
+        with self.assertRaises(ValueError):
+            qa({"question": "", "context": C})
+        with self.assertRaises(ValueError):
+            qa({"question": Q, "context": None})
+        with self.assertRaises(ValueError):
+            qa({"question": Q, "context": ""})
+
+        with self.assertRaises(ValueError):
+            qa([{"question": Q, "context": C}, {"question": None, "context": C}])
+        with self.assertRaises(ValueError):
+            qa([{"question": Q, "context": C}, {"question": "", "context": C}])
+
+        with self.assertRaises(ValueError):
+            qa([{"question": Q, "context": C}, {"question": Q, "context": None}])
+        with self.assertRaises(ValueError):
+            qa([{"question": Q, "context": C}, {"question": Q, "context": ""}])
+
+    def test_argument_handler_error_handling_odd(self):
+        qa = QuestionAnsweringArgumentHandler()
+        with self.assertRaises(ValueError):
+            qa(None)
+
+        with self.assertRaises(ValueError):
+            qa(Y=None)
+
+        with self.assertRaises(ValueError):
+            qa(1)