Gradient checkpointing

* Try -j option

* Try other thing

* Bigger machine

* Test lower sphinx version

* Remove trailing space

.circleci/config.yml

@@ -77,16 +77,17 @@ jobs:
                       - v0.4-torch_and_tf-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: pip install .[sklearn,tf-cpu,torch,testing]
             - save_cache:
                 key: v0.4-{{ checksum "setup.py" }}
                 paths:
                     - '~/.cache/pip'
-            - run: RUN_PT_TF_CROSS_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pt_tf_cross_test --durations=0 | tee output.txt
+            - run: RUN_PT_TF_CROSS_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_torch_and_tf ./tests/ -m is_pt_tf_cross_test --durations=0 | tee tests_output.txt
             - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_tests_torch:
         working_directory: ~/transformers
         docker:
@@ -102,13 +103,12 @@ jobs:
                       - v0.4-torch-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: pip install .[sklearn,torch,testing]
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s --make_reports=tests ./tests/ | tee tests_output.txt
+            - 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:
@@ -129,16 +129,17 @@ jobs:
                       - v0.4-tf-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: pip install .[sklearn,tf-cpu,testing]
             - save_cache:
                   key: v0.4-tf-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_tf ./tests/ | tee tests_output.txt
             - store_artifacts:
-               path: ~/transformers/output.txt
-               destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_tests_flax:
         working_directory: ~/transformers
         docker:
@@ -154,16 +155,17 @@ jobs:
                     - v0.4-flax-{{ checksum "setup.py" }}
                     - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: sudo pip install .[flax,sklearn,torch,testing]
             - save_cache:
                   key: v0.4-flax-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ | tee output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_flax ./tests/ | tee tests_output.txt
             - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_tests_pipelines_torch:
         working_directory: ~/transformers
         docker:
@@ -179,16 +181,17 @@ jobs:
                       - v0.4-torch-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: pip install .[sklearn,torch,testing]
             - save_cache:
                   key: v0.4-torch-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_torch -m is_pipeline_test ./tests/ | tee tests_output.txt
             - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_tests_pipelines_tf:
         working_directory: ~/transformers
         docker:
@@ -204,16 +207,17 @@ jobs:
                       - v0.4-tf-{{ checksum "setup.py" }}
                       - v0.4-{{ checksum "setup.py" }}
             - run: pip install --upgrade pip
-            - run: pip install git+https://github.com/huggingface/datasets
             - run: pip install .[sklearn,tf-cpu,testing]
             - save_cache:
                   key: v0.4-tf-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s ./tests/ -m is_pipeline_test | tee output.txt
+            - run: RUN_PIPELINE_TESTS=1 python -m pytest -n 8 --dist=loadfile -rA -s --make-reports=tests_pipelines_tf ./tests/ -m is_pipeline_test | tee tests_output.txt
             - store_artifacts:
-               path: ~/transformers/output.txt
-               destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_tests_custom_tokenizers:
         working_directory: ~/transformers
         docker:
@@ -233,10 +237,12 @@ jobs:
                   key: v0.4-custom_tokenizers-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
+            - run: python -m pytest -s --make-reports=tests_custom_tokenizers ./tests/test_tokenization_bert_japanese.py | tee tests_output.txt
             - store_artifacts:
-                path: ~/transformers/output.txt
-                destination: test_output.txt
+                  path: ~/transformers/tests_output.txt
+            - store_artifacts:
+                  path: ~/transformers/reports
+
     run_examples_torch:
         working_directory: ~/transformers
         docker:
@@ -258,11 +264,12 @@ jobs:
                   key: v0.4-torch_examples-{{ checksum "setup.py" }}
                   paths:
                       - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s --make_reports=examples ./examples/ | tee examples_output.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s --make-reports=examples_torch ./examples/ | tee examples_output.txt
             - store_artifacts:
                   path: ~/transformers/examples_output.txt
             - store_artifacts:
                   path: ~/transformers/reports
+
     build_doc:
         working_directory: ~/transformers
         docker:
@@ -282,6 +289,7 @@ jobs:
             - run: cd docs && make html SPHINXOPTS="-W"
             - store_artifacts:
                 path: ./docs/_build
+
     deploy_doc:
         working_directory: ~/transformers
         docker:
@@ -301,6 +309,7 @@ jobs:
                   paths:
                       - '~/.cache/pip'
             - run: ./.circleci/deploy.sh
+
     check_code_quality:
         working_directory: ~/transformers
         docker:
@@ -327,6 +336,7 @@ jobs:
             - run: python utils/check_copies.py
             - run: python utils/check_dummies.py
             - run: python utils/check_repo.py
+
     check_repository_consistency:
         working_directory: ~/transformers
         docker:
@@ -357,6 +367,7 @@ jobs:
             - setup_remote_docker
             - *build_push_docker
             - *deploy_cluster
+
     cleanup-gke-jobs:
         docker:
             - image: circleci/python:3.6
@@ -366,6 +377,7 @@ jobs:
                   cluster: $GKE_CLUSTER
                   perform-login: true
             - *delete_gke_jobs
+
 workflow_filters: &workflow_filters
     filters:
         branches:

.github/workflows/self-push.yml

@@ -61,7 +61,19 @@ jobs:
           CUDA_VISIBLE_DEVICES: 0
         run: |
           source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_torch_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_gpu_failures_short.txt
+        
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_gpu_test_reports
+          path: reports
+                  
 
   run_tests_tf_gpu:
     runs-on: [self-hosted, single-gpu]
@@ -110,7 +122,18 @@ jobs:
           CUDA_VISIBLE_DEVICES: 0
         run: |
           source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_tf_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_gpu_failures_short.txt
+        
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_tf_gpu_test_reports
+          path: reports
 
   run_tests_torch_multiple_gpu:
     runs-on: [self-hosted, multi-gpu]
@@ -158,7 +181,18 @@ jobs:
           OMP_NUM_THREADS: 1
         run: |
           source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_torch_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_multiple_gpu_failures_short.txt          
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_multi_gpu_test_reports
+          path: reports
 
   run_tests_tf_multiple_gpu:
     runs-on: [self-hosted, multi-gpu]
@@ -206,4 +240,16 @@ jobs:
           OMP_NUM_THREADS: 1
         run: |
           source .env/bin/activate
-          python -m pytest -n 2 --dist=loadfile -s ./tests/
+          python -m pytest -n 2 --dist=loadfile -s --make-reports=tests_tf_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_multiple_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_tf_multi_gpu_test_reports
+          path: reports
+          

.github/workflows/self-scheduled.yml

@@ -60,11 +60,11 @@ jobs:
           RUN_SLOW: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s --make_reports=tests_torch tests
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_torch_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_test_torch_failures_short.txt
+        run: cat reports/tests_torch_gpu_failures_short.txt
         
       - name: Run examples tests on GPU
         if: ${{ always() }}
@@ -74,11 +74,11 @@ jobs:
         run: |
           source .env/bin/activate
           pip install -r examples/requirements.txt
-          python -m pytest -n 1 --dist=loadfile -s --make_reports=examples_torch examples
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=examples_torch_gpu examples
 
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_examples_torch_failures_short.txt
+        run: cat reports/examples_torch_gpu_failures_short.txt
 
       - name: Run all pipeline tests on GPU
         if: ${{ always() }}
@@ -89,11 +89,11 @@ jobs:
           RUN_PIPELINE_TESTS: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make_reports=tests_torch_pipeline tests
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_torch_pipeline_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_tests_torch_pipeline_failures_short.txt
+        run: cat reports/tests_torch_pipeline_gpu_failures_short.txt
 
       - name: Test suite reports artifacts
         if: ${{ always() }}
@@ -154,11 +154,26 @@ jobs:
           RUN_SLOW: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s --make_reports=tests_tf tests
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_tf_gpu tests
           
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_test_tf_failures_short.txt
+        run: cat reports/tests_tf_gpu_failures_short.txt
+
+      - name: Run all pipeline tests on GPU
+        if: ${{ always() }}
+        env:
+          TF_FORCE_GPU_ALLOW_GROWTH: "true"
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+          RUN_PIPELINE_TESTS: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_tf_pipelines_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_pipelines_gpu_failures_short.txt
 
       - name: Test suite reports artifacts
         if: ${{ always() }}
@@ -177,7 +192,7 @@ jobs:
         id: cache
         with:
           path: .env
-          key: v0.1-slow_tests_torch_multi_gpu-${{ hashFiles('setup.py') }}
+          key: v1.1-slow_tests_torch_multi_gpu-${{ hashFiles('setup.py') }}
 
       - name: Python version
         run: |
@@ -218,26 +233,12 @@ jobs:
           RUN_SLOW: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s --make_reports=tests_torch tests
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_torch_multiple_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_test_torch_failures_short.txt
+        run: cat reports/tests_torch_multiple_gpu_failures_short.txt
         
-      - name: Run examples tests on GPU
-        if: ${{ always() }}
-        env:
-          OMP_NUM_THREADS: 1
-          RUN_SLOW: yes
-        run: |
-          source .env/bin/activate
-          pip install -r examples/requirements.txt
-          python -m pytest -n 1 --dist=loadfile -s --make_reports=examples_torch examples
-
-      - name: Failure short reports
-        if: ${{ always() }}
-        run: cat reports/report_examples_torch_failures_short.txt
-
       - name: Run all pipeline tests on GPU
         if: ${{ always() }}
         env:
@@ -247,11 +248,11 @@ jobs:
           RUN_PIPELINE_TESTS: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make_reports=tests_torch_pipeline tests
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_torch_pipeline_multiple_gpu tests
 
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_tests_torch_pipeline_failures_short.txt
+        run: cat reports/tests_torch_pipeline_multiple_gpu_failures_short.txt
 
       - name: Test suite reports artifacts
         if: ${{ always() }}
@@ -270,7 +271,7 @@ jobs:
         id: cache
         with:
           path: .env
-          key: v0.1-slow_tests_tf_multi_gpu-${{ hashFiles('setup.py') }}
+          key: v1.1-slow_tests_tf_multi_gpu-${{ hashFiles('setup.py') }}
 
       - name: Python version
         run: |
@@ -311,9 +312,14 @@ jobs:
           RUN_SLOW: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ --durations=0
+          python -m pytest -n 1 --dist=loadfile -s --make-reports=tests_tf_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_tf_multiple_gpu_failures_short.txt
 
       - name: Run all pipeline tests on GPU
+        if: ${{ always() }}
         env:
           TF_FORCE_GPU_ALLOW_GROWTH: "true"
           OMP_NUM_THREADS: 1
@@ -321,11 +327,11 @@ jobs:
           RUN_PIPELINE_TESTS: yes
         run: |
           source .env/bin/activate
-          python -m pytest -n 1 --dist=loadfile -s ./tests/ -m is_pipeline_test --durations=50
+          python -m pytest -n 1 --dist=loadfile -s -m is_pipeline_test --make-reports=tests_tf_pipelines_multiple_gpu tests
           
       - name: Failure short reports
         if: ${{ always() }}
-        run: cat reports/report_test_tf_failures_short.txt
+        run: cat reports/tests_tf_multiple_gpu_pipelines_failures_short.txt
 
       - name: Test suite reports artifacts
         if: ${{ always() }}
@@ -333,4 +339,4 @@ jobs:
         with:
           name: run_all_tests_tf_multi_gpu_test_reports
           path: reports
-          
+          

docs/source/index.rst

@@ -272,3 +272,4 @@ conversion utilities for the following models:
     internal/pipelines_utils
     internal/tokenization_utils
     internal/trainer_utils
+    internal/generation_utils

docs/source/internal/generation_utils.rst

@@ -0,0 +1,50 @@
+Utilities for Generation
+-----------------------------------------------------------------------------------------------------------------------
+
+This page lists all the utility functions used by :meth:`~transformers.PretrainedModel.generate`,
+:meth:`~transformers.PretrainedModel.greedy_search`, :meth:`~transformers.PretrainedModel.sample`,
+:meth:`~transformers.PretrainedModel.beam_search`, and :meth:`~transformers.PretrainedModel.beam_sample`.
+
+Most of those are only useful if you are studying the code of the generate methods in the library.
+
+LogitsProcessor
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+A :class:`~transformers.LogitsProcessor` can be used to modify the prediction scores of a language model head for
+generation.
+
+.. autoclass:: transformers.LogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.LogitsProcessorList
+    :members: __call__
+
+.. autoclass:: transformers.MinLengthLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.TemperatureLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.RepetitionPenaltyLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.TopPLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.TopKLogitsWarper
+    :members: __call__
+
+.. autoclass:: transformers.NoRepeatNGramLogitsProcessor
+    :members: __call__
+
+.. autoclass:: transformers.NoBadWordsLogitsProcessor
+    :members: __call__
+
+BeamSearch
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BeamScorer
+    :members: process, finalize
+
+.. autoclass:: transformers.BeamSearchScorer
+    :members: process, finalize

docs/source/main_classes/model.rst

@@ -45,7 +45,7 @@ TFModelUtilsMixin
     :members:
 
 
-Generative models
+Generation
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.generation_utils.GenerationMixin

examples/README.md

@@ -5,7 +5,9 @@ Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.
 
 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
-- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might just lack some features),
+- with information on whether they are **built on top of `Trainer`/`TFTrainer`** (if not, they still work, they might
+  just lack some features),
+- whether or not they leverage the [🤗 Datasets](https://github.com/huggingface/datasets) library.
 - links to **Colab notebooks** to walk through the scripts and run them easily,
 - links to **Cloud deployments** to be able to deploy large-scale trainings in the Cloud with little to no setup.
 
@@ -31,19 +33,19 @@ git checkout tags/v3.4.0
 
 ## The Big Table of Tasks
 
-| Task | Example datasets | Trainer support | TFTrainer support | 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)           | -               | n/a | n/a | [![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/seq2seq)                     | CNN/Daily Mail  | ✅  | - | -
-| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | - | -
-| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | - | -
-| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | - | -
+| Task | Example datasets | Trainer support | TFTrainer support | 🤗 Datasets | 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://github.com/huggingface/notebooks/blob/master/examples/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)           | -               | n/a | n/a | - | [![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/seq2seq)                     | CNN/Daily Mail  | ✅  | - | - | -
+| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)                       | WMT             | ✅  | - | - | -
+| [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)                       | -               | - | - | - | -
+| [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)                   | HANS            | ✅ | - | - | -
 
 
 <br>

examples/bert-loses-patience/pabee/modeling_pabee_albert.py

@@ -65,7 +65,7 @@ class AlbertModelWithPabee(AlbertModel):
 
         self.encoder = AlbertTransformerWithPabee(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
         self.patience = 0
         self.inference_instances_num = 0
         self.inference_layers_num = 0
@@ -228,7 +228,7 @@ class AlbertForSequenceClassificationWithPabee(AlbertPreTrainedModel):
             [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
         )
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING)
     def forward(

examples/bert-loses-patience/pabee/modeling_pabee_bert.py

@@ -70,7 +70,7 @@ class BertModelWithPabee(BertModel):
 
         self.encoder = BertEncoderWithPabee(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
         self.patience = 0
         self.inference_instances_num = 0
         self.inference_layers_num = 0
@@ -252,7 +252,7 @@ class BertForSequenceClassificationWithPabee(BertPreTrainedModel):
             [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
         )
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
     def forward(

examples/conftest.py

@@ -17,17 +17,14 @@ warnings.simplefilter(action="ignore", category=FutureWarning)
 
 
 def pytest_addoption(parser):
-    parser.addoption(
-        "--make_reports",
-        action="store",
-        default=False,
-        help="generate report files - the value will be used as a `report_`+val+`reportname.txt`",
-    )
+    from transformers.testing_utils import pytest_addoption_shared
+
+    pytest_addoption_shared(parser)
 
 
 def pytest_terminal_summary(terminalreporter):
     from transformers.testing_utils import pytest_terminal_summary_main
 
-    make_reports = terminalreporter.config.getoption("--make_reports")
+    make_reports = terminalreporter.config.getoption("--make-reports")
     if make_reports:
         pytest_terminal_summary_main(terminalreporter, id=make_reports)

examples/deebert/src/modeling_highway_bert.py

@@ -115,7 +115,7 @@ class DeeBertModel(BertPreTrainedModel):
         self.encoder = DeeBertEncoder(config)
         self.pooler = BertPooler(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def init_highway_pooler(self):
         self.encoder.init_highway_pooler(self.pooler)
@@ -286,7 +286,7 @@ class DeeBertForSequenceClassification(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING)
     def forward(

examples/deebert/src/modeling_highway_roberta.py

@@ -23,7 +23,7 @@ class DeeRobertaModel(DeeBertModel):
         super().__init__(config)
 
         self.embeddings = RobertaEmbeddings(config)
-        self.init_weights()
+        self.init_weights_and_layers()
 
 
 @add_start_docstrings(

examples/language-modeling/README.md

@@ -1,16 +1,19 @@
-
 ## 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,
+ALBERT, BERT, DistilBERT, RoBERTa, XLNet... GPT and GPT-2 are trained or fine-tuned using a causal language modeling
+(CLM) loss while ALBERT, BERT, DistilBERT and RoBERTa are trained or fine-tuned using a masked language modeling (MLM)
+loss. XLNet uses permutation language modeling (PLM), you can find more information about the differences between those
+objectives in our [model summary](https://huggingface.co/transformers/model_summary.html).
 
-Fine-tuning (or training from scratch) the library models for language modeling on a text dataset for GPT, GPT-2, BERT, DistilBERT and RoBERTa. GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT, DistilBERT and RoBERTa
-are fine-tuned using a masked language modeling (MLM) loss.
+These scripts leverage the 🤗 Datasets library and the Trainer API. You can easily customize them to your needs if you
+need extra processing on your datasets.
 
-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/).
+**Note:** The old script `run_language_modeling.py` is still available
+[here](https://github.com/huggingface/transformers/blob/master/examples/contrib/legacy/run_language_modeling.py).
 
-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.
+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.
 
 ### GPT-2/GPT and causal language modeling
 
@@ -18,66 +21,105 @@ The following example fine-tunes GPT-2 on WikiText-2. We're using the raw WikiTe
 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 \
+python run_clm.py \
+    --model_name_or_path gpt2 \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
     --do_train \
-    --train_data_file=$TRAIN_FILE \
     --do_eval \
-    --eval_data_file=$TEST_FILE
+    --output_dir /tmp/test-clm
 ```
 
 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.
 
+To run on your own training and validation files, use the following command:
+
+```bash
+python run_clm.py \
+    --model_name_or_path gpt2 \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-clm
+```
+
+
 ### RoBERTa/BERT/DistilBERT and masked language modeling
 
 The following example fine-tunes RoBERTa on WikiText-2. Here too, we're using the raw WikiText-2. The loss is different
 as BERT/RoBERTa have a bidirectional mechanism; we're therefore using the same loss that was used during their
 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.
-
-If using whole-word masking, use both the`--mlm` and `--wwm` flags.
+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).
 
 ```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 \
+python run_mlm.py \
+    --model_name_or_path roberta-base \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
     --do_train \
-    --train_data_file=$TRAIN_FILE \
     --do_eval \
-    --eval_data_file=$TEST_FILE \
-    --mlm \
-    --whole_word_mask
+    --output_dir /tmp/test-mlm
 ```
 
-For Chinese models, it's same with English model with only `--mlm`. If using whole-word masking, we need to generate a reference files, because it's char level.
+To run on your own training and validation files, use the following command:
 
-**Q :** Why ref file ?
+```bash
+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-mlm
+```
 
-**A :** Suppose we have a Chinese sentence like : `我喜欢你` The original Chinese-BERT will tokenize it as `['我','喜','欢','你']` in char level.
-Actually, `喜欢` is a whole word. For whole word mask proxy, We need res like `['我','喜','##欢','你']`.
-So we need a ref file to tell model which pos of BERT original token should be added `##`.
+If your dataset is organized with one sample per line, you can use the `--line_by_line` flag (otherwise the script
+concatenates all texts and then splits them in blocks of the same length).
+
+**Note:** On TPU, you should use the flag `--pad_to_max_length` in conjunction with the `--line_by_line` flag to make
+sure all your batches have the same length.
+
+### Whole word masking
+
+The BERT authors released a new version of BERT using Whole Word Masking in May 2019. Instead of masking randomly
+selected tokens (which may be part of words), they mask randomly selected words (masking all the tokens corresponding
+to that word). This technique has been refined for Chinese in [this paper](https://arxiv.org/abs/1906.08101).
+
+To fine-tune a model using whole word masking, use the following script:
+
+python run_mlm_wwm.py \
+    --model_name_or_path roberta-base \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-mlm-wwm
+```
+
+For Chinese models, we need to generate a reference files (which requires the ltp library), because it's tokenized at
+the character level.
+
+**Q :** Why a reference file?
+
+**A :** Suppose we have a Chinese sentence like: `我喜欢你` The original Chinese-BERT will tokenize it as
+`['我','喜','欢','你']` (character level). But `喜欢` is a whole word. For whole word masking proxy, we need a result
+like `['我','喜','##欢','你']`, so we need a reference file to tell the model which position of the BERT original token
+should be added `##`.
 
 **Q :** Why LTP ?
 
-**A :** Cause the best known Chinese WWM BERT is [Chinese-BERT-wwm](https://github.com/ymcui/Chinese-BERT-wwm) by HIT. It works well on so many Chines Task like CLUE (Chinese GLUE).
-They use LTP, so if we want to fine-tune their model, we need LTP.
+**A :** Cause the best known Chinese WWM BERT is [Chinese-BERT-wwm](https://github.com/ymcui/Chinese-BERT-wwm) by HIT.
+It works well on so many Chines Task like CLUE (Chinese GLUE). They use LTP, so if we want to fine-tune their model,
+we need LTP.
 
 Now LTP only only works well on `transformers==3.2.0`. So we don't add it to requirements.txt.
-You need to check to `3.2.0` for `run_chinese_ref.py`. And the code  could be found in `examples/contrib`.
+You need to create a separate environment with this version of Transformers to run the `run_chinese_ref.py` script that
+will create the reference files. The script is in `examples/contrib`. Once in the proper environment, run the
+following:
 
 
 ```bash
@@ -87,33 +129,29 @@ export BERT_RESOURCE=/path/to/bert/tokenizer
 export SAVE_PATH=/path/to/data/ref.txt
 
 python examples/contrib/run_chinese_ref.py \
-    --file_name=$TRAIN_FILE \
-    --ltp=$LTP_RESOURCE \
-    --bert=$BERT_RESOURCE \
-    --save_path=$SAVE_PATH 
+    --file_name=path_to_train_or_eval_file \
+    --ltp=path_to_ltp_tokenizer \
+    --bert=path_to_bert_tokenizer \
+    --save_path=path_to_reference_file
 ```
-Now Chinese Ref is only supported by `LineByLineWithRefDataset` Class, so we need add `line_by_line` flag: 
+
+Then you can run the script like this: 
 
 
 ```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-export REF_FILE=/path/to/ref.txt
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_type=roberta \
-    --model_name_or_path=roberta-base \
+python run_mlm_wwm.py \
+    --model_name_or_path roberta-base \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --train_ref_file path_to_train_chinese_ref_file \
+    --validation_ref_file path_to_validation_chinese_ref_file \
     --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --chinese_ref_file=$REF_FILE \
     --do_eval \
-    --eval_data_file=$TEST_FILE \
-    --mlm \
-    --line_by_line \
-    --whole_word_mask
+    --output_dir /tmp/test-mlm-wwm
 ```
 
+**Note:** On TPU, you should the flag `--pad_to_max_length` to make sure all your batches have the same length.
+
 ### XLNet and permutation language modeling
 
 XLNet uses a different training objective, which is permutation language modeling. It is an autoregressive method 
@@ -126,15 +164,32 @@ context length for permutation language modeling.
 The `--max_span_length` flag may also be used to limit the length of a span of masked tokens used 
 for permutation language modeling.
 
-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
+Here is how to fine-tun XLNet on wikitext-2:
 
-python run_language_modeling.py \
-    --output_dir=output \
+```bash
+python run_plm.py \
     --model_name_or_path=xlnet-base-cased \
+    --dataset_name wikitext \
+    --dataset_config_name wikitext-2-raw-v1 \
     --do_train \
-    --train_data_file=$TRAIN_FILE \
     --do_eval \
-    --eval_data_file=$TEST_FILE \
+    --output_dir /tmp/test-plm
 ```
+
+To fine-tune it on your own training and validation file, run:
+
+```bash
+python run_plm.py \
+    --model_name_or_path=xlnet-base-cased \
+    --train_file path_to_train_file \
+    --validation_file path_to_validation_file \
+    --do_train \
+    --do_eval \
+    --output_dir /tmp/test-plm
+```
+
+If your dataset is organized with one sample per line, you can use the `--line_by_line` flag (otherwise the script
+concatenates all texts and then splits them in blocks of the same length).
+
+**Note:** On TPU, you should use the flag `--pad_to_max_length` in conjunction with the `--line_by_line` flag to make
+sure all your batches have the same length.

examples/language-modeling/run_clm.py

@@ -175,10 +175,10 @@ def main():
 
     # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
     # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
-    # (the dataset will be downloaded automatically from the datasets Hub
+    # (the dataset will be downloaded automatically from the datasets Hub).
     #
-    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
-    # behavior (see below)
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
     #
     # In distributed training, the load_dataset function guarantee that only one local process can concurrently
     # download the dataset.
@@ -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"
@@ -259,14 +259,14 @@ def main():
     )
 
     if data_args.block_size <= 0:
-        block_size = tokenizer.max_len
+        block_size = tokenizer.model_max_length
     else:
-        if data_args.block_size > tokenizer.max_len:
+        if data_args.block_size > tokenizer.model_max_length:
             logger.warn(
                 f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model"
-                f"({tokenizer.max_len}). Using block_size={tokenizer.max_len}."
+                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
             )
-        block_size = min(data_args.block_size, tokenizer.max_len)
+        block_size = min(data_args.block_size, tokenizer.model_max_length)
 
     # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
     def group_texts(examples):

examples/language-modeling/run_mlm.py

@@ -116,6 +116,17 @@ class DataTrainingArguments:
     mlm_probability: float = field(
         default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
     )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
 
     def __post_init__(self):
         if self.dataset_name is None and self.train_file is None and self.validation_file is None:
@@ -190,7 +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"
@@ -246,18 +257,73 @@ def main():
         column_names = datasets["validation"].column_names
     text_column_name = "text" if "text" in column_names else column_names[0]
 
-    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"], truncation=True, max_length=data_args.max_seq_length)
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
 
-    tokenized_datasets = datasets.map(
-        tokenize_function,
-        batched=True,
-        num_proc=data_args.preprocessing_num_workers,
-        remove_columns=[text_column_name],
-        load_from_cache_file=not data_args.overwrite_cache,
-    )
+        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)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name])
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        if data_args.max_seq_length is None:
+            max_seq_length = tokenizer.model_max_length
+        else:
+            if data_args.max_seq_length > tokenizer.model_max_length:
+                logger.warn(
+                    f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
+                    f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
+                )
+            max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
 
     # Data collator
     # This one will take care of randomly masking the tokens.

examples/language-modeling/run_mlm_wwm.py

@@ -0,0 +1,334 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Team All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
+text file or a dataset.
+
+Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
+https://huggingface.co/models?filter=masked-lm
+"""
+# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
+
+import json
+import logging
+import math
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from datasets import Dataset, load_dataset
+
+import transformers
+from transformers import (
+    CONFIG_MAPPING,
+    MODEL_FOR_MASKED_LM_MAPPING,
+    AutoConfig,
+    AutoModelForMaskedLM,
+    AutoTokenizer,
+    DataCollatorForWholeWordMask,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    set_seed,
+)
+from transformers.trainer_utils import is_main_process
+
+
+logger = logging.getLogger(__name__)
+MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
+    """
+
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The model checkpoint for weights initialization."
+            "Don't set if you want to train a model from scratch."
+        },
+    )
+    model_type: Optional[str] = field(
+        default=None,
+        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+    use_fast_tokenizer: bool = field(
+        default=True,
+        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
+    validation_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
+    )
+    train_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
+    )
+    validation_ref_file: Optional[str] = field(
+        default=None,
+        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization. Sequences longer "
+            "than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
+
+    def __post_init__(self):
+        if self.train_file is not None:
+            extension = self.train_file.split(".")[-1]
+            assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
+        if self.validation_file is not None:
+            extension = self.validation_file.split(".")[-1]
+            assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+
+
+def add_chinese_references(dataset, ref_file):
+    with open(ref_file, "r", encoding="utf-8") as f:
+        refs = [json.loads(line) for line in f.read().splitlines() if (len(line) > 0 and not line.isspace())]
+    assert len(dataset) == len(refs)
+
+    dataset_dict = {c: dataset[c] for c in dataset.column_names}
+    dataset_dict["chinese_ref"] = refs
+    return Dataset.from_dict(dataset_dict)
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty."
+            "Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if is_main_process(training_args.local_rank) else logging.WARN,
+    )
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
+    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
+    # (the dataset will be downloaded automatically from the datasets Hub).
+    #
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    #
+    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
+    # download the dataset.
+    data_files = {}
+    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.validation_file
+    extension = data_args.train_file.split(".")[-1]
+    if extension == "txt":
+        extension = "text"
+    datasets = load_dataset(extension, data_files=data_files)
+    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
+    # https://huggingface.co/docs/datasets/loading_datasets.html.
+
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+    if model_args.config_name:
+        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
+    elif model_args.model_name_or_path:
+        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    if model_args.tokenizer_name:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = AutoTokenizer.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
+        )
+    else:
+        raise ValueError(
+            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
+            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
+        )
+
+    if model_args.model_name_or_path:
+        model = AutoModelForMaskedLM.from_pretrained(
+            model_args.model_name_or_path,
+            from_tf=bool(".ckpt" in model_args.model_name_or_path),
+            config=config,
+            cache_dir=model_args.cache_dir,
+        )
+    else:
+        logger.info("Training new model from scratch")
+        model = AutoModelForMaskedLM.from_config(config)
+
+    model.resize_token_embeddings(len(tokenizer))
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+    if training_args.do_train:
+        column_names = datasets["train"].column_names
+    else:
+        column_names = datasets["validation"].column_names
+    text_column_name = "text" if "text" in column_names else column_names[0]
+
+    padding = "max_length" if data_args.pad_to_max_length else False
+
+    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)
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True,
+        num_proc=data_args.preprocessing_num_workers,
+        remove_columns=[text_column_name],
+        load_from_cache_file=not data_args.overwrite_cache,
+    )
+
+    # Add the chinese references if provided
+    if data_args.train_ref_file is not None:
+        tokenized_datasets["train"] = add_chinese_references(tokenized_datasets["train"], data_args.train_ref_file)
+    if data_args.valid_ref_file is not None:
+        tokenized_datasets["validation"] = add_chinese_references(
+            tokenized_datasets["validation"], data_args.validation_ref_file
+        )
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = DataCollatorForWholeWordMask(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)
+
+    # Initialize our Trainer
+    trainer = Trainer(
+        model=model,
+        args=training_args,
+        train_dataset=tokenized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=tokenized_datasets["validation"] if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+    )
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()  # Saves the tokenizer too for easy upload
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+
+        eval_output = trainer.evaluate()
+
+        perplexity = math.exp(eval_output["eval_loss"])
+        results["perplexity"] = perplexity
+
+        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm_wwm.txt")
+        if trainer.is_world_process_zero():
+            with open(output_eval_file, "w") as writer:
+                logger.info("***** Eval results *****")
+                for key, value in results.items():
+                    logger.info(f"  {key} = {value}")
+                    writer.write(f"{key} = {value}\n")
+
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()

examples/language-modeling/run_plm.py

@@ -96,7 +96,7 @@ class DataTrainingArguments:
         default=None,
         metadata={
             "help": "The maximum total input sequence length after tokenization. Sequences longer "
-            "than this will be truncated."
+            "than this will be truncated. Default to the max input length of the model."
         },
     )
     preprocessing_num_workers: Optional[int] = field(
@@ -113,6 +113,17 @@ class DataTrainingArguments:
     max_span_length: int = field(
         default=5, metadata={"help": "Maximum length of a span of masked tokens for permutation language modeling."}
     )
+    line_by_line: bool = field(
+        default=False,
+        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
+    )
+    pad_to_max_length: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to pad all samples to `max_seq_length`. "
+            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
+        },
+    )
 
     def __post_init__(self):
         if self.dataset_name is None and self.train_file is None and self.validation_file is None:
@@ -172,10 +183,10 @@ def main():
 
     # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
     # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
-    # (the dataset will be downloaded automatically from the datasets Hub
+    # (the dataset will be downloaded automatically from the datasets Hub).
     #
-    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
-    # behavior (see below)
+    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
     #
     # In distributed training, the load_dataset function guarantee that only one local process can concurrently
     # download the dataset.
@@ -187,7 +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"
@@ -243,18 +254,73 @@ def main():
         column_names = datasets["validation"].column_names
     text_column_name = "text" if "text" in column_names else column_names[0]
 
-    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"], truncation=True, max_length=data_args.max_seq_length)
+    if data_args.line_by_line:
+        # When using line_by_line, we just tokenize each nonempty line.
+        padding = "max_length" if data_args.pad_to_max_length else False
 
-    tokenized_datasets = datasets.map(
-        tokenize_function,
-        batched=True,
-        num_proc=data_args.preprocessing_num_workers,
-        remove_columns=[text_column_name],
-        load_from_cache_file=not data_args.overwrite_cache,
-    )
+        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)
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+    else:
+        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+        def tokenize_function(examples):
+            return tokenizer(examples[text_column_name])
+
+        tokenized_datasets = datasets.map(
+            tokenize_function,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            remove_columns=[text_column_name],
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
+
+        if data_args.max_seq_length is None:
+            max_seq_length = tokenizer.model_max_length
+        else:
+            if data_args.max_seq_length > tokenizer.model_max_length:
+                logger.warn(
+                    f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
+                    f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
+                )
+            max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
+        # max_seq_length.
+        def group_texts(examples):
+            # Concatenate all texts.
+            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
+            total_length = len(concatenated_examples[list(examples.keys())[0]])
+            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
+            # customize this part to your needs.
+            total_length = (total_length // max_seq_length) * max_seq_length
+            # Split by chunks of max_len.
+            result = {
+                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
+                for k, t in concatenated_examples.items()
+            }
+            return result
+
+        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
+        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
+        # might be slower to preprocess.
+        #
+        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
+        tokenized_datasets = tokenized_datasets.map(
+            group_texts,
+            batched=True,
+            num_proc=data_args.preprocessing_num_workers,
+            load_from_cache_file=not data_args.overwrite_cache,
+        )
 
     # Data collator
     data_collator = DataCollatorForPermutationLanguageModeling(

examples/lightning_base.py

@@ -7,6 +7,7 @@ from typing import Any, Dict
 import pytorch_lightning as pl
 from pytorch_lightning.utilities import rank_zero_info
 
+import pkg_resources
 from transformers import (
     AdamW,
     AutoConfig,
@@ -32,6 +33,15 @@ from transformers.optimization import (
 
 logger = logging.getLogger(__name__)
 
+try:
+    pkg = "pytorch_lightning"
+    min_ver = "1.0.4"
+    pkg_resources.require(f"{pkg}>={min_ver}")
+except pkg_resources.VersionConflict:
+    logger.warning(
+        f"{pkg}>={min_ver} is required for a normal functioning of this module, but found {pkg}=={pkg_resources.get_distribution(pkg).version}. Try pip install -r examples/requirements.txt"
+    )
+
 
 MODEL_MODES = {
     "base": AutoModel,

examples/movement-pruning/emmental/modeling_bert_masked.py

@@ -482,7 +482,7 @@ class MaskedBertModel(MaskedBertPreTrainedModel):
         self.encoder = BertEncoder(config)
         self.pooler = BertPooler(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -669,7 +669,7 @@ class MaskedBertForSequenceClassification(MaskedBertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
     def forward(
@@ -754,7 +754,7 @@ class MaskedBertForMultipleChoice(MaskedBertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, 1)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
     def forward(
@@ -844,7 +844,7 @@ class MaskedBertForTokenClassification(MaskedBertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
     def forward(
@@ -930,7 +930,7 @@ class MaskedBertForQuestionAnswering(MaskedBertPreTrainedModel):
         self.bert = MaskedBertModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(MASKED_BERT_INPUTS_DOCSTRING)
     def forward(

examples/seq2seq/finetune_trainer.py

@@ -4,16 +4,9 @@ import sys
 from dataclasses import dataclass, field
 from typing import Optional
 
-from seq2seq_trainer import Seq2SeqTrainer, arg_to_scheduler_choices
-from transformers import (
-    AutoConfig,
-    AutoModelForSeq2SeqLM,
-    AutoTokenizer,
-    HfArgumentParser,
-    MBartTokenizer,
-    TrainingArguments,
-    set_seed,
-)
+from seq2seq_trainer import Seq2SeqTrainer
+from seq2seq_training_args import Seq2SeqTrainingArguments
+from transformers import AutoConfig, AutoModelForSeq2SeqLM, AutoTokenizer, HfArgumentParser, MBartTokenizer, set_seed
 from transformers.trainer_utils import EvaluationStrategy
 from utils import (
     Seq2SeqDataCollator,
@@ -33,41 +26,6 @@ from utils import (
 logger = logging.getLogger(__name__)
 
 
-@dataclass
-class Seq2SeqTrainingArguments(TrainingArguments):
-    """
-    Parameters:
-        label_smoothing (:obj:`float`, `optional`, defaults to 0):
-            The label smoothing epsilon to apply (if not zero).
-        sortish_sampler (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to SortishSamler or not. It sorts the inputs according to lenghts in-order to minimizing the padding size.
-        predict_with_generate (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to use generate to calculate generative metrics (ROUGE, BLEU).
-    """
-
-    label_smoothing: Optional[float] = field(
-        default=0.0, metadata={"help": "The label smoothing epsilon to apply (if not zero)."}
-    )
-    sortish_sampler: bool = field(default=False, metadata={"help": "Whether to SortishSamler or not."})
-    predict_with_generate: bool = field(
-        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
-    )
-    adafactor: bool = field(default=False, metadata={"help": "whether to use adafactor"})
-    encoder_layerdrop: Optional[float] = field(
-        default=None, metadata={"help": "Encoder layer dropout probability. Goes into model.config."}
-    )
-    decoder_layerdrop: Optional[float] = field(
-        default=None, metadata={"help": "Decoder layer dropout probability. Goes into model.config."}
-    )
-    dropout: Optional[float] = field(default=None, metadata={"help": "Dropout probability. Goes into model.config."})
-    attention_dropout: Optional[float] = field(
-        default=None, metadata={"help": "Attention dropout probability. Goes into model.config."}
-    )
-    lr_scheduler: Optional[str] = field(
-        default="linear", metadata={"help": f"Which lr scheduler to use. Selected in {arg_to_scheduler_choices}"}
-    )
-
-
 @dataclass
 class ModelArguments:
     """

examples/seq2seq/seq2seq_trainer.py

@@ -20,12 +20,6 @@ from transformers.optimization import (
 from transformers.trainer_pt_utils import get_tpu_sampler
 
 
-try:
-    from .utils import label_smoothed_nll_loss
-except ImportError:
-    from utils import label_smoothed_nll_loss
-
-
 logger = logging.get_logger(__name__)
 
 arg_to_scheduler = {
@@ -36,7 +30,6 @@ arg_to_scheduler = {
     "constant": get_constant_schedule,
     "constant_w_warmup": get_constant_schedule_with_warmup,
 }
-arg_to_scheduler_choices = sorted(arg_to_scheduler.keys())
 
 
 class Seq2SeqTrainer(Trainer):
@@ -64,6 +57,14 @@ class Seq2SeqTrainer(Trainer):
                 f"The `config.pad_token_id` is `None`. Using `config.eos_token_id` = {self.config.eos_token_id} for padding.."
             )
 
+        if self.args.label_smoothing == 0:
+            self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
+        else:
+            # dynamically import label_smoothed_nll_loss
+            from utils import label_smoothed_nll_loss
+
+            self.loss_fn = label_smoothed_nll_loss
+
     def create_optimizer_and_scheduler(self, num_training_steps: int):
         """
         Setup the optimizer and the learning rate scheduler.
@@ -135,9 +136,7 @@ class Seq2SeqTrainer(Trainer):
             if self.data_args is not None and self.data_args.ignore_pad_token_for_loss:
                 # force training to ignore pad token
                 logits = model(**inputs, use_cache=False)[0]
-
-                loss_fct = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
-                loss = loss_fct(logits.view(-1, logits.shape[-1]), labels.view(-1))
+                loss = self.loss_fn(logits.view(-1, logits.shape[-1]), labels.view(-1))
             else:
                 # compute usual loss via models
                 loss, logits = model(**inputs, labels=labels, use_cache=False)[:2]
@@ -145,9 +144,7 @@ class Seq2SeqTrainer(Trainer):
             # compute label smoothed loss
             logits = model(**inputs, use_cache=False)[0]
             lprobs = torch.nn.functional.log_softmax(logits, dim=-1)
-            loss, _ = label_smoothed_nll_loss(
-                lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id
-            )
+            loss, _ = self.loss_fn(lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id)
         return loss, logits
 
     def compute_loss(self, model, inputs):

examples/seq2seq/seq2seq_training_args.py

@@ -0,0 +1,45 @@
+import logging
+from dataclasses import dataclass, field
+from typing import Optional
+
+from seq2seq_trainer import arg_to_scheduler
+from transformers import TrainingArguments
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Seq2SeqTrainingArguments(TrainingArguments):
+    """
+    Parameters:
+        label_smoothing (:obj:`float`, `optional`, defaults to 0):
+            The label smoothing epsilon to apply (if not zero).
+        sortish_sampler (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to SortishSamler or not. It sorts the inputs according to lenghts in-order to minimizing the padding size.
+        predict_with_generate (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to use generate to calculate generative metrics (ROUGE, BLEU).
+    """
+
+    label_smoothing: Optional[float] = field(
+        default=0.0, metadata={"help": "The label smoothing epsilon to apply (if not zero)."}
+    )
+    sortish_sampler: bool = field(default=False, metadata={"help": "Whether to SortishSamler or not."})
+    predict_with_generate: bool = field(
+        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
+    )
+    adafactor: bool = field(default=False, metadata={"help": "whether to use adafactor"})
+    encoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Encoder layer dropout probability. Goes into model.config."}
+    )
+    decoder_layerdrop: Optional[float] = field(
+        default=None, metadata={"help": "Decoder layer dropout probability. Goes into model.config."}
+    )
+    dropout: Optional[float] = field(default=None, metadata={"help": "Dropout probability. Goes into model.config."})
+    attention_dropout: Optional[float] = field(
+        default=None, metadata={"help": "Attention dropout probability. Goes into model.config."}
+    )
+    lr_scheduler: Optional[str] = field(
+        default="linear",
+        metadata={"help": f"Which lr scheduler to use. Selected in {sorted(arg_to_scheduler.keys())}"},
+    )

examples/seq2seq/test_tatoeba_conversion.py

@@ -1,11 +1,13 @@
+import os
 import tempfile
 import unittest
 
-from transformers.convert_marian_tatoeba_to_pytorch import TatoebaConverter
+from transformers.convert_marian_tatoeba_to_pytorch import DEFAULT_REPO, TatoebaConverter
 from transformers.file_utils import cached_property
 from transformers.testing_utils import slow
 
 
+@unittest.skipUnless(os.path.exists(DEFAULT_REPO), "Tatoeba directory does not exist.")
 class TatoebaConversionTester(unittest.TestCase):
     @cached_property
     def resolver(self):

examples/text-classification/run_glue.py

@@ -177,7 +177,7 @@ def main():
     set_seed(training_args.seed)
 
     # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
-    # or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub
+    # or specify a GLUE benchmark task (the dataset will be downloaded automatically from the datasets Hub).
     #
     # For CSV/JSON files, this script will use as labels the column called 'label' and as pair of sentences the
     # sentences in columns called 'sentence1' and 'sentence2' if such column exists or the first two columns not named

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/huawei-noah/DynaBERT_MNLI/README.md

@@ -8,13 +8,13 @@ allowing both adaptive width and depth using knowledge distillation.
 * This code is modified based on the repository developed by Hugging Face: [Transformers v2.1.1](https://github.com/huggingface/transformers/tree/v2.1.1), and is released in [GitHub](https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/DynaBERT).
 
 ### Reference
-Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Qun Liu.
+Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, Qun Liu.
 [DynaBERT: Dynamic BERT with Adaptive Width and Depth](https://arxiv.org/abs/2004.04037).
 ```
 @inproceedings{hou2020dynabert,
   title = {DynaBERT: Dynamic BERT with Adaptive Width and Depth},
-  author = {Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Qun Liu},  
-  booktitle = {NeurIPS},
+  author = {Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, Qun Liu},  
+  booktitle = {Advances in Neural Information Processing Systems},
   year = {2020}
 }
 ```

model_cards/huawei-noah/DynaBERT_SST-2/README.md

@@ -8,13 +8,13 @@ allowing both adaptive width and depth using knowledge distillation.
 * This code is modified based on the repository developed by Hugging Face: [Transformers v2.1.1](https://github.com/huggingface/transformers/tree/v2.1.1), and is released in [GitHub](https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/DynaBERT).
 
 ### Reference
-Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Qun Liu.
+Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, Qun Liu.
 [DynaBERT: Dynamic BERT with Adaptive Width and Depth](https://arxiv.org/abs/2004.04037).
 ```
 @inproceedings{hou2020dynabert,
   title = {DynaBERT: Dynamic BERT with Adaptive Width and Depth},
-  author = {Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Qun Liu},  
-  booktitle = {NeurIPS},
+  author = {Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, Qun Liu},  
+  booktitle = {Advances in Neural Information Processing Systems},
   year = {2020}
 }
 ```

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

model_cards/neuralspace-reverie/indic-transformers-bn-bert/README.md

@@ -0,0 +1,25 @@
+---
+language: 
+- bn 
+tags:
+- MaskedLM
+- Bengali
+---
+# Indic-Transformers Bengali BERT
+## Model description
+This is a BERT language model pre-trained on ~3 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-bn-bert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-bn-bert')
+text = "আপনি কেমন আছেন?"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 6, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-bn-distilbert/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- bn 
+tags:
+- MaskedLM
+- Bengali
+- DistilBERT
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Bengali DistilBERT
+## Model description
+This is a DistilBERT language model pre-trained on ~6 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-bn-distilbert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-bn-distilbert')
+text = "আপনি কেমন আছেন?"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-bn-roberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- bn 
+tags:
+- MaskedLM
+- Bengali
+- RoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Bengali RoBERTa
+## Model description
+This is a RoBERTa language model pre-trained on ~6 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-bn-roberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-bn-roberta')
+text = "আপনি কেমন আছেন?"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 10, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-bn-xlmroberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- bn 
+tags:
+- MaskedLM
+- Bengali
+- XLMRoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Bengali XLMRoBERTa
+## Model description
+This is a XLMRoBERTa language model pre-trained on ~3 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-bn-xlmroberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-bn-xlmroberta')
+text = "আপনি কেমন আছেন?"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-hi-bert/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- hi 
+tags:
+- MaskedLM
+- Hindi
+- BERT
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Hindi BERT
+## Model description
+This is a BERT language model pre-trained on ~3 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-hi-bert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-hi-bert')
+text = "आपका स्वागत हैं"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-hi-distilbert/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- hi 
+tags:
+- MaskedLM
+- Hindi
+- DistilBERT
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Hindi DistilBERT
+## Model description
+This is a DistilBERT language model pre-trained on ~10 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-hi-distilbert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-hi-distilbert')
+text = "आपका स्वागत हैं"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-hi-roberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- hi 
+tags:
+- MaskedLM
+- Hindi
+- RoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Hindi RoBERTa
+## Model description
+This is a RoBERTa language model pre-trained on ~10 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-hi-roberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-hi-roberta')
+text = "आपका स्वागत हैं"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 11, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-hi-xlmroberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- hi 
+tags:
+- MaskedLM
+- Hindi
+- XLMRoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Hindi XLMRoBERTa
+## Model description
+This is a XLMRoBERTa language model pre-trained on ~3 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-hi-xlmroberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-hi-xlmroberta')
+text = "आपका स्वागत हैं"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-te-bert/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- te
+tags:
+- MaskedLM
+- Telugu
+- BERT
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Telugu BERT
+## Model description
+This is a BERT language model pre-trained on ~1.6 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-te-bert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-te-bert')
+text = "మీరు ఎలా ఉన్నారు"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-te-distilbert/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- te
+tags:
+- MaskedLM
+- Telugu
+- DistilBERT
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Telugu DistilBERT
+## Model description
+This is a DistilBERT language model pre-trained on ~2 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-te-distilbert')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-te-distilbert')
+text = "మీరు ఎలా ఉన్నారు"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-te-roberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- te
+tags:
+- MaskedLM
+- Telugu
+- RoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Telugu RoBERTa
+## Model description
+This is a RoBERTa language model pre-trained on ~2 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-te-roberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-te-roberta')
+text = "మీరు ఎలా ఉన్నారు"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 14, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/neuralspace-reverie/indic-transformers-te-xlmroberta/README.md

@@ -0,0 +1,29 @@
+---
+language: 
+- te
+tags:
+- MaskedLM
+- Telugu
+- XLMRoBERTa
+- Question-Answering
+- Token Classification
+- Text Classification
+---
+# Indic-Transformers Telugu XLMRoBERTa
+## Model description
+This is a XLMRoBERTa language model pre-trained on ~1.6 GB of monolingual training corpus. The pre-training data was majorly taken from [OSCAR](https://oscar-corpus.com/).
+This model can be fine-tuned on various downstream tasks like text-classification, POS-tagging, question-answering, etc. Embeddings from this model can also be used for feature-based training.
+## Intended uses & limitations
+#### How to use
+```
+from transformers import AutoTokenizer, AutoModel
+tokenizer = AutoTokenizer.from_pretrained('neuralspace-reverie/indic-transformers-te-xlmroberta')
+model = AutoModel.from_pretrained('neuralspace-reverie/indic-transformers-te-xlmroberta')
+text = "మీరు ఎలా ఉన్నారు"
+input_ids = tokenizer(text, return_tensors='pt')['input_ids']
+out = model(input_ids)[0]
+print(out.shape)
+# out = [1, 5, 768] 
+```
+#### Limitations and bias
+The original language model has been trained using `PyTorch` and hence the use of `pytorch_model.bin` weights file is recommended. The h5 file for `Tensorflow` has been generated manually by commands suggested [here](https://huggingface.co/transformers/model_sharing.html).

model_cards/patrickvonplaten/bert2bert_cnn_daily_mail/README.md

@@ -0,0 +1,16 @@
+---
+language: en
+license: apache-2.0
+datasets:
+- cnn_dailymail
+tags:
+- summarization
+---
+
+Bert2Bert Summarization with 🤗EncoderDecoder Framework
+This model is a warm-started *BERT2BERT* model fine-tuned on the *CNN/Dailymail* summarization dataset.
+
+The model achieves a **18.22** ROUGE-2 score on *CNN/Dailymail*'s test dataset.
+
+For more details on how the model was fine-tuned, please refer to 
+[this](https://colab.research.google.com/drive/1Ekd5pUeCX7VOrMx94_czTkwNtLN32Uyu?usp=sharing) notebook.

model_cards/patrickvonplaten/roberta_shared_bbc_xsum/README.md

@@ -0,0 +1,16 @@
+---
+language: en
+license: apache-2.0
+datasets:
+- xsum
+tags:
+- summarization
+---
+
+Shared RoBERTa2RoBERTa Summarization with 🤗EncoderDecoder Framework
+This model is a warm-started *RoBERTaShared* model fine-tuned on the *BBC XSum* summarization dataset.
+
+The model achieves a **16.89** ROUGE-2 score on *BBC XSUM*'s test dataset.
+
+For more details on how the model was fine-tuned, please refer to 
+[this](https://colab.research.google.com/drive/1Ekd5pUeCX7VOrMx94_czTkwNtLN32Uyu?usp=sharing) notebook.

notebooks/README.md

@@ -26,7 +26,7 @@ Pull Request so it can be included under the Community notebooks.
 
 | Notebook     |      Description      |      Author      |      |
 |:----------|:-------------|:-------------|------:|
-| [Train T5 in Tensoflow 2 ](https://github.com/snapthat/TF-T5-text-to-text) | How to train T5 for any task using Tensorflow 2. This notebook demonstrates a Question & Answer task implemented in Tensorflow 2 using SQUAD | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
+| [Train T5 in Tensorflow 2 ](https://github.com/snapthat/TF-T5-text-to-text) | How to train T5 for any task using Tensorflow 2. This notebook demonstrates a Question & Answer task implemented in Tensorflow 2 using SQUAD | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
 | [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) |
@@ -51,3 +51,5 @@ Pull Request so it can be included under the Community notebooks.
 |[Fine-tune Roberta for sentiment analysis](https://github.com/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb) | How to fine-tune an Roberta model for sentiment analysis | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)|
 |[Evaluating Question Generation Models](https://github.com/flexudy-pipe/qugeev) | How accurate are the answers to questions generated by your seq2seq transformer model? | [Pascal Zoleko](https://github.com/zolekode) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1bpsSqCQU-iw_5nNoRm_crPq6FRuJthq_?usp=sharing)|
 |[Classify text with DistilBERT and Tensorflow](https://github.com/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb) | How to fine-tune DistilBERT for text classification in TensorFlow | [Peter Bayerle](https://github.com/peterbayerle) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb)|
+|[Leverage BERT for Encoder-Decoder Summarization on CNN/Dailymail](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb) | How to warm-start a *EncoderDecoderModel* with a *bert-base-uncased* checkpoint for summarization on CNN/Dailymail | [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/BERT2BERT_for_CNN_Dailymail.ipynb)|
+|[Leverage RoBERTa for Encoder-Decoder Summarization on BBC XSum](https://github.com/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb) | How to warm-start a shared *EncoderDecoderModel* with a *roberta-base* checkpoint for summarization on BBC/XSum | [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/RoBERTaShared_for_BBC_XSum.ipynb)|

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

@@ -284,6 +284,7 @@ if is_torch_available():
         DataCollatorForNextSentencePrediction,
         DataCollatorForPermutationLanguageModeling,
         DataCollatorForSOP,
+        DataCollatorForTokenClassification,
         DataCollatorForWholeWordMask,
         DataCollatorWithPadding,
         default_data_collator,
@@ -299,6 +300,19 @@ if is_torch_available():
         TextDataset,
         TextDatasetForNextSentencePrediction,
     )
+    from .generation_beam_search import BeamScorer, BeamSearchScorer
+    from .generation_logits_process import (
+        LogitsProcessor,
+        LogitsProcessorList,
+        LogitsWarper,
+        MinLengthLogitsProcessor,
+        NoBadWordsLogitsProcessor,
+        NoRepeatNGramLogitsProcessor,
+        RepetitionPenaltyLogitsProcessor,
+        TemperatureLogitsWarper,
+        TopKLogitsWarper,
+        TopPLogitsWarper,
+    )
     from .generation_utils import top_k_top_p_filtering
     from .modeling_albert import (
         ALBERT_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/activations.py

@@ -2,6 +2,7 @@ import math
 
 import torch
 import torch.nn.functional as F
+from packaging import version
 
 from .utils import logging
 
@@ -9,29 +10,25 @@ from .utils import logging
 logger = logging.get_logger(__name__)
 
 
-def swish(x):
-    return x * torch.sigmoid(x)
-
-
 def _gelu_python(x):
     """
-    Original Implementation of the gelu activation function in Google Bert repo when initially created. For
-    information: OpenAI GPT's gelu is slightly different (and gives slightly different results): 0.5 * x * (1 +
+    Original Implementation of the GELU activation function in Google BERT repo when initially created. For
+    information: OpenAI GPT's GELU is slightly different (and gives slightly different results): 0.5 * x * (1 +
     torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) This is now written in C in
-    torch.nn.functional Also see https://arxiv.org/abs/1606.08415
+    torch.nn.functional Also see the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
     """
     return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
 
 
 def gelu_new(x):
     """
-    Implementation of the gelu activation function currently in Google Bert repo (identical to OpenAI GPT). Also see
-    https://arxiv.org/abs/1606.08415
+    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
+    the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
     """
     return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
 
 
-if torch.__version__ < "1.4.0":
+if version.parse(torch.__version__) < version.parse("1.4"):
     gelu = _gelu_python
 else:
     gelu = F.gelu
@@ -41,6 +38,23 @@ def gelu_fast(x):
     return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x)))
 
 
+def _silu_python(x):
+    """
+    See Gaussian Error Linear Units (Hendrycks et al., https://arxiv.org/abs/1606.08415) where the SiLU (Sigmoid Linear
+    Unit) was originally introduced and coined, and see Sigmoid-Weighted Linear Units for Neural Network Function
+    Approximation in Reinforcement Learning (Elfwing et al., https://arxiv.org/abs/1702.03118) and Swish: a Self-Gated
+    Activation Function (Ramachandran et al., https://arxiv.org/abs/1710.05941v1) where the SiLU was experimented with
+    later.
+    """
+    return x * torch.sigmoid(x)
+
+
+if version.parse(torch.__version__) < version.parse("1.7"):
+    silu = _silu_python
+else:
+    silu = F.silu
+
+
 def mish(x):
     return x * torch.tanh(torch.nn.functional.softplus(x))
 
@@ -51,7 +65,8 @@ def linear_act(x):
 
 ACT2FN = {
     "relu": F.relu,
-    "swish": swish,
+    "silu": silu,
+    "swish": silu,
     "gelu": gelu,
     "tanh": torch.tanh,
     "gelu_new": gelu_new,

src/transformers/activations_tf.py

@@ -52,6 +52,7 @@ ACT2FN = {
     "gelu": tf.keras.layers.Activation(gelu),
     "relu": tf.keras.activations.relu,
     "swish": tf.keras.activations.swish,
+    "silu": tf.keras.activations.swish,
     "gelu_new": tf.keras.layers.Activation(gelu_new),
     "mish": tf.keras.layers.Activation(mish),
     "tanh": tf.keras.activations.tanh,

src/transformers/benchmark/benchmark.py

@@ -254,6 +254,7 @@ class PyTorchBenchmark(Benchmark):
             else:
                 # cpu
                 memory_bytes = measure_peak_memory_cpu(func)
+                print("PEAK", memory_bytes)
                 memory = Memory(memory_bytes) if isinstance(memory_bytes, int) else memory_bytes
 
             if self.args.trace_memory_line_by_line:
@@ -261,6 +262,7 @@ class PyTorchBenchmark(Benchmark):
             else:
                 summary = None
 
+            print(memory, summary)
             return memory, summary
         except RuntimeError as e:
             self.print_fn("Doesn't fit on GPU. {}".format(e))

src/transformers/benchmark/benchmark_utils.py

@@ -76,12 +76,7 @@ def separate_process_wrapper_fn(func: Callable[[], None], do_multi_processing: b
         # run function in an individual
         # process to get correct memory
         def wrapper_func(queue: Queue, *args):
-            try:
-                result = func(*args)
-            except Exception as e:
-                logger.error(e)
-                print(e)
-                result = "N/A"
+            result = func(*args)
             queue.put(result)
 
         queue = Queue()
@@ -291,13 +286,13 @@ def measure_peak_memory_cpu(function: Callable[[], None], interval=0.5, device_i
                 # receive memory and num measurements
                 max_memory = parent_connection.recv()
                 num_measurements = parent_connection.recv()
-            except Exception:
+            except Exception as e:
                 # kill process in a clean way
                 parent = psutil.Process(os.getpid())
                 for child in parent.children(recursive=True):
                     os.kill(child.pid, SIGKILL)
                 mem_process.join(0)
-                raise RuntimeError("Process killed. Error in Process")
+                raise RuntimeError(f"Process killed. Error in Process: {e}")
 
             # run process at least 20 * interval or until it finishes
             mem_process.join(20 * interval)
@@ -692,7 +687,8 @@ class Benchmark(ABC):
                 for sequence_length in self.args.sequence_lengths:
                     if self.args.inference:
                         if self.args.memory:
-                            memory, inference_summary = self.inference_memory(model_name, batch_size, sequence_length)
+                            outputs = self.inference_memory(model_name, batch_size, sequence_length)
+                            memory, inference_summary = outputs if type(outputs) == tuple else (outputs, None)
                             inference_result_memory[model_name]["result"][batch_size][sequence_length] = memory
                         if self.args.speed:
                             time = self.inference_speed(model_name, batch_size, sequence_length)
@@ -700,7 +696,8 @@ class Benchmark(ABC):
 
                     if self.args.training:
                         if self.args.memory:
-                            memory, train_summary = self.train_memory(model_name, batch_size, sequence_length)
+                            outputs = self.train_memory(model_name, batch_size, sequence_length)
+                            memory, train_summary = outputs if type(outputs) == tuple else (outputs, None)
                             train_result_memory[model_name]["result"][batch_size][sequence_length] = memory
                         if self.args.speed:
                             time = self.train_speed(model_name, batch_size, sequence_length)

src/transformers/configuration_albert.py

@@ -61,7 +61,7 @@ class AlbertConfig(PretrainedConfig):
             The number of inner repetition of attention and ffn.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu_new"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0):

src/transformers/configuration_bart.py

@@ -59,7 +59,7 @@ class BartConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_bert.py

@@ -74,7 +74,7 @@ class BertConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_bert_generation.py

@@ -40,7 +40,7 @@ class BertGenerationConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often called feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_blenderbot.py

@@ -56,7 +56,7 @@ class BlenderbotConfig(BartConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_deberta.py

@@ -52,7 +52,7 @@ class DebertaConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"`, :obj:`"gelu"`, :obj:`"tanh"`, :obj:`"gelu_fast"`,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"`, :obj:`"gelu"`, :obj:`"tanh"`, :obj:`"gelu_fast"`,
             :obj:`"mish"`, :obj:`"linear"`, :obj:`"sigmoid"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.

src/transformers/configuration_distilbert.py

@@ -66,7 +66,7 @@ class DistilBertConfig(PretrainedConfig):
             The dropout ratio for the attention probabilities.
         activation (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         initializer_range (:obj:`float`, `optional`, defaults to 0.02):
             The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
         qa_dropout (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_dpr.py

@@ -55,7 +55,7 @@ class DPRConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_electra.py

@@ -60,7 +60,7 @@ class ElectraConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_encoder_decoder.py

@@ -68,7 +68,7 @@ class EncoderDecoderConfig(PretrainedConfig):
         >>> encoder_decoder_config = EncoderDecoderConfig.from_pretrained('my-model')
         >>> model = EncoderDecoderModel.from_pretrained('my-model', config=encoder_decoder_config)
     """
-    model_type = "encoder_decoder"
+    model_type = "encoder-decoder"
     is_composition = True
 
     def __init__(self, **kwargs):

src/transformers/configuration_fsmt.py

@@ -71,7 +71,7 @@ class FSMTConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"relu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_funnel.py

@@ -66,7 +66,7 @@ class FunnelConfig(PretrainedConfig):
             Inner dimension in the feed-forward blocks.
         hidden_act (:obj:`str` or :obj:`callable`, `optional`, defaults to :obj:`"gelu_new"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_gpt2.py

@@ -60,7 +60,7 @@ class GPT2Config(PretrainedConfig):
         n_inner (:obj:`int`, `optional`, defaults to None):
             Dimensionality of the inner feed-forward layers. :obj:`None` will set it to 4 times n_embd
         activation_function (:obj:`str`, `optional`, defaults to :obj:`"gelu"`):
-            Activation function, to be selected in the list :obj:`["relu", "swish", "gelu", "tanh", "gelu_new"]`.
+            Activation function, to be selected in the list :obj:`["relu", "silu", "gelu", "tanh", "gelu_new"]`.
         resid_pdrop (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         embd_pdrop (:obj:`int`, `optional`, defaults to 0.1):

src/transformers/configuration_layoutlm.py

@@ -52,7 +52,7 @@ class LayoutLMConfig(BertConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_lxmert.py

@@ -55,7 +55,7 @@ class LxmertConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_marian.py

@@ -50,7 +50,7 @@ class MarianConfig(BartConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_mbart.py

@@ -55,7 +55,7 @@ class MBartConfig(BartConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_mobilebert.py

@@ -48,7 +48,7 @@ class MobileBertConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"relu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.0):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_openai.py

@@ -54,7 +54,7 @@ class OpenAIGPTConfig(PretrainedConfig):
             Number of attention heads for each attention layer in the Transformer encoder.
         afn (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         resid_pdrop (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         embd_pdrop (:obj:`int`, `optional`, defaults to 0.1):

src/transformers/configuration_pegasus.py

@@ -94,7 +94,7 @@ class PegasusConfig(BartConfig):
             Dimensionality of the "intermediate" (i.e., feed-forward) layer in decoder.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         dropout (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_dropout (:obj:`float`, `optional`, defaults to 0.0):

src/transformers/configuration_prophetnet.py

@@ -39,7 +39,7 @@ class ProphetNetConfig(PretrainedConfig):
             The dropout ratio for activations inside the fully connected layer.
         activation_function (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         vocab_size (:obj:`int`, `optional`, defaults to 30522):
             Vocabulary size of the ProphetNET model. Defines the number of different tokens that can be represented by
             the :obj:`inputs_ids` passed when calling :class:`~transformers.ProphetNetModel`.

src/transformers/configuration_reformer.py

@@ -80,7 +80,7 @@ class ReformerConfig(PretrainedConfig):
             :obj:`None` to ensure fully random rotations in local sensitive hashing scheme.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"relu"`):
             The non-linear activation function (function or string) in the feed forward layer in the residual attention
-            block. If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            block. If string, :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.05):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         hidden_size (:obj:`int`, `optional`, defaults to 256):

src/transformers/configuration_retribert.py

@@ -49,7 +49,7 @@ class RetriBertConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`function`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_squeezebert.py

@@ -50,7 +50,7 @@ class SqueezeBertConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the encoder and pooler. If string,
-            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
             The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
         attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):

src/transformers/configuration_utils.py

@@ -170,6 +170,7 @@ class PretrainedConfig(object):
         self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
         self.pruned_heads = kwargs.pop("pruned_heads", {})
+        self.gradient_checkpointing = kwargs.pop("gradient_checkpointing", False)
         self.tie_word_embeddings = kwargs.pop(
             "tie_word_embeddings", True
         )  # Whether input and output word embeddings should be tied for all MLM, LM and Seq2Seq models.

src/transformers/configuration_xlnet.py

@@ -54,7 +54,7 @@ class XLNetConfig(PretrainedConfig):
             Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
         ff_activation (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
             The non-linear activation function (function or string) in the If string, :obj:`"gelu"`, :obj:`"relu"`,
-            :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+            :obj:`"silu"` and :obj:`"gelu_new"` are supported.
         untie_r (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether or not to untie relative position biases
         attn_type (:obj:`str`, `optional`, defaults to :obj:`"bi"`):

src/transformers/data/data_collator.py

@@ -114,6 +114,67 @@ class DataCollatorWithPadding:
         return batch
 
 
+@dataclass
+class DataCollatorForTokenClassification:
+    """
+    Data collator that will dynamically pad the inputs received, as well as the labels.
+
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+        label_pad_token_id (:obj:`int`, `optional`, defaults to -100):
+            The id to use when padding the labels (-100 will be automatically ignore by PyTorch loss functions).
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    padding: Union[bool, str, PaddingStrategy] = True
+    max_length: Optional[int] = None
+    pad_to_multiple_of: Optional[int] = None
+    label_pad_token_id: int = -100
+
+    def __call__(self, features):
+        label_name = "label" if "label" in features[0].keys() else "labels"
+        labels = [feature[label_name] for feature in features] if label_name in features[0].keys() else None
+        batch = self.tokenizer.pad(
+            features,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            # Conversion to tensors will fail if we have labels as they are not of the same length yet.
+            return_tensors="pt" if labels is None else None,
+        )
+
+        if labels is None:
+            return batch
+
+        sequence_length = torch.tensor(batch["input_ids"]).shape[1]
+        padding_side = self.tokenizer.padding_side
+        if padding_side == "right":
+            batch["labels"] = [label + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels]
+        else:
+            batch["labels"] = [[self.label_pad_token_id] * (sequence_length - len(label)) + label for label in labels]
+
+        batch = {k: torch.tensor(v, dtype=torch.int64) for k, v in batch.items()}
+        return batch
+
+
 @dataclass
 class DataCollatorForLanguageModeling:
     """

src/transformers/generation_beam_search.py

@@ -0,0 +1,357 @@
+# coding=utf-8
+# Copyright 2020 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.
+
+from abc import ABC, abstractmethod
+from collections import UserDict
+from typing import Optional, Tuple
+
+import torch
+
+from .file_utils import add_start_docstrings
+
+
+PROCESS_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PretrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        next_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Current scores of the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_tokens (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            :obj:`input_ids` of the tokens corresponding to the top :obj:`2 * num_beams` non-finished beam hypotheses.
+        next_indices (:obj:`torch.LongTensor` of shape :obj:`(batch_size, 2 * num_beams)`):
+            Beam indices indicating to which beam hypothesis the :obj:`next_tokens` correspond.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`UserDict`: A dictionary composed of the fields as defined above:
+
+            - **next_beam_scores** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Updated
+              scores of all non-finished beams.
+            - **next_beam_tokens** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Next tokens
+              to be added to the non-finished beam_hypotheses.
+            - **next_beam_indices** (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`) -- Beam indices
+              indicating to which beam the next tokens shall be added.
+
+"""
+
+FINALIZE_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size * num_beams, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using any class inheriting from :class:`~transformers.PretrainedTokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        final_beam_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The final scores of all non-finished beams.
+        final_beam_tokens (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The last tokens to be added to the non-finished beam_hypotheses.
+        final_beam_indices (:obj:`torch.FloatTensor` of shape :obj:`(batch_size * num_beams)`):
+            The beam indices indicating to which beam the :obj:`final_beam_tokens` shall be added.
+        pad_token_id (:obj:`int`, `optional`):
+            The id of the `padding` token.
+        eos_token_id (:obj:`int`, `optional`):
+            The id of the `end-of-sequence` token.
+
+    Return:
+        :obj:`torch.LongTensor` of shape :obj:`(batch_size * num_return_sequences, sequence_length)`: The generated
+        sequences. The second dimension (sequence_length) is either equal to :obj:`max_length` or shorter if all
+        batches finished early due to the :obj:`eos_token_id`.
+
+"""
+
+
+class BeamScorer(ABC):
+    """
+    Abstract base class for all beam scorers that are used for :meth:`~transformers.PretrainedModel.beam_search` and
+    :meth:`~transformers.PretrainedModel.beam_sample`.
+    """
+
+    @abstractmethod
+    @add_start_docstrings(PROCESS_INPUTS_DOCSTRING)
+    def process(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        **kwargs
+    ) -> Tuple[torch.Tensor]:
+        raise NotImplementedError("This is an abstract method.")
+
+    @abstractmethod
+    @add_start_docstrings(FINALIZE_INPUTS_DOCSTRING)
+    def finalize(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        **kwargs
+    ) -> torch.LongTensor:
+        raise NotImplementedError("This is an abstract method.")
+
+
+class BeamSearchScorer(BeamScorer):
+    r"""
+    :class:`transformers.BeamScorer` implementing standard beam search decoding.
+
+    Adapted in part from `Facebook's XLM beam search code
+    <https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529>`__.
+
+    Args:
+        batch_size (:obj:`int`):
+            Batch Size of :obj:`input_ids` for which beam search decoding is run in parallel.
+        max_length (:obj:`int`):
+            The maximum length of the sequence to be generated.
+        num_beams (:obj:`int`):
+            Number of beams for beam search.
+        device (:obj:`torch.device`):
+            Defines the device type (*e.g.*, :obj:`"cpu"` or :obj:`"cuda"`) on which this instance of
+            :obj:`BeamSearchScorer` will be allocated.
+        length_penalty (:obj:`float`, `optional`, defaults to 1.0):
+            Exponential penalty to the length. 1.0 means no penalty. Set to values < 1.0 in order to encourage the
+            model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer
+            sequences.
+        do_early_stopping (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to stop the beam search when at least ``num_beams`` sentences are finished per batch or not.
+        num_beam_hyps_to_keep (:obj:`int`, `optional`, defaults to 1):
+            The number of beam hypotheses that shall be returned upon calling
+            :meth:`~transformer.BeamSearchScorer.finalize`.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        max_length: int,
+        num_beams: int,
+        device: torch.device,
+        length_penalty: Optional[float] = 1.0,
+        do_early_stopping: Optional[bool] = False,
+        num_beam_hyps_to_keep: Optional[int] = 1,
+    ):
+        self.max_length = max_length
+        self.num_beams = num_beams
+        self.device = device
+        self.length_penalty = length_penalty
+        self.do_early_stopping = do_early_stopping
+        self.num_beam_hyps_to_keep = num_beam_hyps_to_keep
+
+        self._is_init = False
+        self._beam_hyps = [
+            BeamHypotheses(
+                num_beams=self.num_beams,
+                max_length=self.max_length,
+                length_penalty=self.length_penalty,
+                early_stopping=self.do_early_stopping,
+            )
+            for _ in range(batch_size)
+        ]
+        self._done = torch.tensor([False for _ in range(batch_size)], dtype=torch.bool, device=self.device)
+
+        if not isinstance(num_beams, int) or num_beams <= 1:
+            raise ValueError(
+                f"`num_beams` has to be an integer strictly greater than 1, but is {num_beams}. For `num_beams` == 1, one should make use of `greedy_search` instead."
+            )
+
+    @property
+    def is_done(self) -> bool:
+        return self._done.all()
+
+    def process(
+        self,
+        input_ids: torch.LongTensor,
+        next_scores: torch.FloatTensor,
+        next_tokens: torch.LongTensor,
+        next_indices: torch.LongTensor,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> Tuple[torch.Tensor]:
+        cur_len = input_ids.shape[-1]
+        batch_size = len(self._beam_hyps)
+        assert batch_size == (input_ids.shape[0] // self.num_beams)
+
+        device = input_ids.device
+        next_beam_scores = torch.zeros((batch_size, self.num_beams), dtype=next_scores.dtype, device=device)
+        next_beam_tokens = torch.zeros((batch_size, self.num_beams), dtype=next_tokens.dtype, device=device)
+        next_beam_indices = torch.zeros((batch_size, self.num_beams), dtype=next_indices.dtype, device=device)
+
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                assert (
+                    len(beam_hyp) >= self.num_beams
+                ), "Batch can only be done if at least {} beams have been generated".format(self.num_beams)
+                assert (
+                    eos_token_id is not None and pad_token_id is not None
+                ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
+                # pad the batch
+                next_beam_scores[batch_idx, :] = 0
+                next_beam_tokens[batch_idx, :] = pad_token_id
+                next_beam_indices[batch_idx, :] = 0
+                continue
+
+            # next tokens for this sentence
+            beam_idx = 0
+            for beam_token_rank, (next_token, next_score, next_index) in enumerate(
+                zip(next_tokens[batch_idx], next_scores[batch_idx], next_indices[batch_idx])
+            ):
+                batch_beam_idx = batch_idx * self.num_beams + next_index
+                # add to generated hypotheses if end of sentence
+                if (eos_token_id is not None) and (next_token.item() == eos_token_id):
+                    # if beam_token does not belong to top num_beams tokens, it should not be added
+                    is_beam_token_worse_than_top_num_beams = beam_token_rank >= self.num_beams
+                    if is_beam_token_worse_than_top_num_beams:
+                        continue
+                    beam_hyp.add(
+                        input_ids[batch_beam_idx].clone(),
+                        next_score.item(),
+                    )
+                else:
+                    # add next predicted token since it is not eos_token
+                    next_beam_scores[batch_idx, beam_idx] = next_score
+                    next_beam_tokens[batch_idx, beam_idx] = next_token
+                    next_beam_indices[batch_idx, beam_idx] = batch_beam_idx
+                    beam_idx += 1
+
+                # once the beam for next step is full, don't add more tokens to it.
+                if beam_idx == self.num_beams:
+                    break
+
+            if beam_idx < self.num_beams:
+                raise ValueError(
+                    f"At most {self.num_beams} tokens in {next_tokens[batch_idx]} can be equal to `eos_token_id: {eos_token_id}`. Make sure {next_tokens[batch_idx]} are corrected."
+                )
+
+            # Check if we are done so that we can save a pad step if all(done)
+            self._done[batch_idx] = self._done[batch_idx] or beam_hyp.is_done(
+                next_scores[batch_idx].max().item(), cur_len
+            )
+
+        return UserDict(
+            {
+                "next_beam_scores": next_beam_scores.view(-1),
+                "next_beam_tokens": next_beam_tokens.view(-1),
+                "next_beam_indices": next_beam_indices.view(-1),
+            }
+        )
+
+    def finalize(
+        self,
+        input_ids: torch.LongTensor,
+        final_beam_scores: torch.FloatTensor,
+        final_beam_tokens: torch.LongTensor,
+        final_beam_indices: torch.LongTensor,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+    ) -> torch.LongTensor:
+        batch_size = len(self._beam_hyps)
+
+        # finalize all open beam hypotheses and add to generated hypotheses
+        for batch_idx, beam_hyp in enumerate(self._beam_hyps):
+            if self._done[batch_idx]:
+                continue
+
+            # need to add best num_beams hypotheses to generated hyps
+            for beam_id in range(self.num_beams):
+                batch_beam_idx = batch_idx * self.num_beams + beam_id
+                final_score = final_beam_scores[batch_beam_idx].item()
+                final_tokens = input_ids[batch_beam_idx]
+                beam_hyp.add(final_tokens, final_score)
+
+        # select the best hypotheses
+        sent_lengths = input_ids.new(batch_size * self.num_beam_hyps_to_keep)
+        best = []
+
+        # retrieve best hypotheses
+        for i, beam_hyp in enumerate(self._beam_hyps):
+            sorted_hyps = sorted(beam_hyp.beams, key=lambda x: x[0])
+            for j in range(self.num_beam_hyps_to_keep):
+                best_hyp = sorted_hyps.pop()[1]
+                sent_lengths[self.num_beam_hyps_to_keep * i + j] = len(best_hyp)
+                best.append(best_hyp)
+
+        # prepare for adding eos
+        sent_max_len = min(sent_lengths.max().item() + 1, self.max_length)
+        decoded: torch.LongTensor = input_ids.new(batch_size * self.num_beam_hyps_to_keep, sent_max_len)
+        # shorter batches are padded if needed
+        if sent_lengths.min().item() != sent_lengths.max().item():
+            assert pad_token_id is not None, "`pad_token_id` has to be defined"
+            decoded.fill_(pad_token_id)
+
+        # fill with hypotheses and eos_token_id if the latter fits in
+        for i, hypo in enumerate(best):
+            decoded[i, : sent_lengths[i]] = hypo
+            if sent_lengths[i] < self.max_length:
+                decoded[i, sent_lengths[i]] = eos_token_id
+        return decoded
+
+
+class BeamHypotheses:
+    def __init__(self, num_beams: int, max_length: int, length_penalty: float, early_stopping: bool):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.max_length = max_length - 1  # ignoring bos_token
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.num_beams = num_beams
+        self.beams = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.beams)
+
+    def add(self, hyp: torch.LongTensor, sum_logprobs: float):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / (hyp.shape[-1] ** self.length_penalty)
+        if len(self) < self.num_beams or score > self.worst_score:
+            self.beams.append((score, hyp))
+            if len(self) > self.num_beams:
+                sorted_next_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
+                del self.beams[sorted_next_scores[0][1]]
+                self.worst_score = sorted_next_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs: float, cur_len: int) -> bool:
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated can become better than the worst
+        one in the heap, then we are done with this sentence.
+        """
+
+        if len(self) < self.num_beams:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret

src/transformers/generation_logits_process.py

@@ -0,0 +1,374 @@
+# coding=utf-8
+# Copyright 2020 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.
+
+from abc import ABC
+from typing import Iterable, List
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+from .file_utils import add_start_docstrings
+
+
+LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
+    Args:
+        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`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax
+            or scores for each vocabulary token after SoftMax.
+
+    Return:
+        :obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.vocab_size)`: The processed prediction scores.
+
+"""
+
+
+class LogitsProcessor(ABC):
+    """Abstract base class for all logit processors that can be applied during generation."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        """Torch method for processing logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsWarper(ABC):
+    """Abstract base class for all logit warpers that can be applied during generation with multinomial sampling."""
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        """Torch method for warping logits."""
+        raise NotImplementedError(
+            f"{self.__class__} is an abstract class. Only classes inheriting this class can be called."
+        )
+
+
+class LogitsProcessorList(list):
+    """
+    This class can be used to create a list of :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsWarper` to subsequently process a :obj:`scores` input tensor. This class inherits from
+    list and adds a specific `__call__` method to apply each :class:`~transformers.LogitsProcessor` or
+    :class:`~transformers.LogitsProcessor` to the inputs.
+    """
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        for processor in self:
+            scores = processor(input_ids, scores)
+        return scores
+
+
+class MinLengthLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing a min-length by setting EOS probability to 0.
+
+    Args:
+        min_length (:obj:`int`):
+            The minimum length below which the score of :obj:`eos_token_id` is set to :obj:`-float("Inf")`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, min_length: int, eos_token_id: int):
+        if not isinstance(min_length, int) or min_length < 0:
+            raise ValueError(f"`min_length` has to be a positive integer, but is {min_length}")
+
+        if not isinstance(eos_token_id, int) or eos_token_id < 0:
+            raise ValueError(f"`eos_token_id` has to be a positive integer, but is {eos_token_id}")
+
+        self.min_length = min_length
+        self.eos_token_id = eos_token_id
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        cur_len = input_ids.shape[-1]
+        if cur_len < self.min_length:
+            scores[:, self.eos_token_id] = -float("inf")
+        return scores
+
+
+class TemperatureLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` for temperature (exponential scaling output probability distribution).
+
+    Args:
+        temperature (:obj:`float`):
+            The value used to module the logits distribution.
+    """
+
+    def __init__(self, temperature: float):
+        if not isinstance(temperature, float) or not (temperature > 0):
+            raise ValueError(f"`temperature` has to be a strictly positive float, but is {temperature}")
+
+        self.temperature = temperature
+
+    def __call__(self, input_ids: torch.Tensor, scores: torch.Tensor) -> torch.Tensor:
+        scores = scores / self.temperature
+        return scores
+
+
+class RepetitionPenaltyLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` enforcing an exponential penalty on repeated sequences.
+
+    Args:
+        repetition_penalty (:obj:`float`):
+            The parameter for repetition penalty. 1.0 means no penalty. See `this paper
+            <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details.
+    """
+
+    def __init__(self, penalty: float):
+        if not isinstance(penalty, float) or not (penalty > 0):
+            raise ValueError(f"`penalty` has to be a strictly positive float, but is {penalty}")
+
+        self.penalty = penalty
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        for i in range(scores.shape[0]):
+            for previous_token in set(input_ids[i].tolist()):
+                # if score < 0 then repetition penalty has to be multiplied to reduce the previous token probability
+                if scores[i, previous_token] < 0:
+                    scores[i, previous_token] *= self.penalty
+                else:
+                    scores[i, previous_token] /= self.penalty
+        return scores
+
+
+class TopPLogitsWarper(LogitsWarper):
+    """
+    :class:`transformers.LogitsWarper` that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <=
+    prob_cut_off.
+
+    Args:
+        top_p (:obj:`float`):
+            If set to < 1, only the most probable tokens with probabilities that add up to :obj:`top_p` or higher are
+            kept for generation.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
+            raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+
+        self.top_p = top_p
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        sorted_logits, sorted_indices = torch.sort(scores, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > self.top_p
+        if self.min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., : self.min_tokens_to_keep - 1] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        scores[indices_to_remove] = self.filter_value
+        return scores
+
+
+class TopKLogitsWarper(LogitsWarper):
+    r"""
+    :class:`transformers.LogitsWarper` that performs top-k, i.e. restricting to the k highest probability elements.
+
+    Args:
+        top_k (:obj:`int`):
+            The number of highest probability vocabulary tokens to keep for top-k-filtering.
+        filter_value (:obj:`float`, `optional`, defaults to :obj:`-float("Inf")`):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (:obj:`int`, `optional`, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+    """
+
+    def __init__(self, top_k: int, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not isinstance(top_k, int) or top_k <= 0:
+            raise ValueError(f"`top_k` has to be a strictly positive integer, but is {top_k}")
+
+        self.top_k = top_k
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        top_k = min(max(self.top_k, self.min_tokens_to_keep), scores.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = scores < torch.topk(scores, top_k)[0][..., -1, None]
+        scores[indices_to_remove] = self.filter_value
+        return scores
+
+
+class NoRepeatNGramLogitsProcessor(LogitsProcessor):
+    r"""
+    :class:`transformers.LogitsProcessor` that enforces no repetition of n-grams. See `Fairseq
+    <https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345>`__.
+
+    Args:
+        ngram_size (:obj:`int`):
+            All ngrams of size :obj:`ngram_size` can only occur once.
+    """
+
+    def __init__(self, ngram_size: int):
+        if not isinstance(ngram_size, int) or ngram_size <= 0:
+            raise ValueError(f"`ngram_size` has to be a strictly positive integer, but is {ngram_size}")
+        self.ngram_size = ngram_size
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        num_batch_hypotheses = scores.shape[0]
+        cur_len = input_ids.shape[-1]
+        banned_batch_tokens = self._calc_banned_ngram_tokens(input_ids, num_batch_hypotheses, cur_len)
+
+        for i, banned_tokens in enumerate(banned_batch_tokens):
+            scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+    def _calc_banned_ngram_tokens(
+        self, prev_input_ids: torch.Tensor, num_hypos: int, cur_len: int
+    ) -> List[Iterable[int]]:
+        """Copied from fairseq for no_repeat_ngram in beam_search"""
+        if cur_len + 1 < self.ngram_size:
+            # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+            return [[] for _ in range(num_hypos)]
+        generated_ngrams = [{} for _ in range(num_hypos)]
+        for idx in range(num_hypos):
+            gen_tokens = prev_input_ids[idx].tolist()
+            generated_ngram = generated_ngrams[idx]
+            for ngram in zip(*[gen_tokens[i:] for i in range(self.ngram_size)]):
+                prev_ngram_tuple = tuple(ngram[:-1])
+                generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
+
+        def _get_generated_ngrams(hypo_idx):
+            # Before decoding the next token, prevent decoding of ngrams that have already appeared
+            start_idx = cur_len + 1 - self.ngram_size
+            ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].tolist())
+            return generated_ngrams[hypo_idx].get(ngram_idx, [])
+
+        banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
+        return banned_tokens
+
+
+class NoBadWordsLogitsProcessor(LogitsProcessor):
+    """
+    :class:`transformers.LogitsProcessor` that enforces that specified sequences will never be sampled.
+
+    Args:
+        bad_words_ids (:obj:`List[List[int]]`):
+            List of list of token ids that are not allowed to be generated. In order to get the tokens of the words
+            that should not appear in the generated text, use :obj:`tokenizer(bad_word,
+            add_prefix_space=True).input_ids`.
+        eos_token_id (:obj:`int`):
+            The id of the `end-of-sequence` token.
+    """
+
+    def __init__(self, bad_words_ids: Iterable[Iterable[int]], eos_token_id: int):
+
+        if not isinstance(bad_words_ids, List) or len(bad_words_ids) == 0:
+            raise ValueError(f"`bad_words_ids` has to be a non-emtpy list, but is {bad_words_ids}.")
+        if any(not isinstance(bad_word_ids, list) for bad_word_ids in bad_words_ids):
+            raise ValueError(f"`bad_words_ids` has to be a list of lists, but is {bad_words_ids}.")
+        if any(
+            any((not isinstance(token_id, (int, np.integer)) or token_id < 0) for token_id in bad_word_ids)
+            for bad_word_ids in bad_words_ids
+        ):
+            raise ValueError(
+                f"Each list in `bad_words_ids` has to be a list of positive integers, but is {bad_words_ids}."
+            )
+
+        self.bad_words_ids = list(filter(lambda bad_token_seq: bad_token_seq != [eos_token_id], bad_words_ids))
+
+        for banned_token_seq in self.bad_words_ids:
+            assert len(banned_token_seq) > 0, "Banned words token sequences {} cannot have an empty list".format(
+                bad_words_ids
+            )
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        banned_tokens = self._calc_banned_bad_words_ids(input_ids)
+        scores = self._set_scores_to_inf_for_banned_tokens(scores, banned_tokens)
+
+        return scores
+
+    def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        elif len(tokens) > len(prev_tokens):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+        elif prev_tokens[-len(tokens) :].tolist() == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    def _calc_banned_bad_words_ids(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
+        banned_tokens = []
+        for prev_input_ids_slice in prev_input_ids:
+            banned_tokens_slice = []
+            for banned_token_seq in self.bad_words_ids:
+                if self._tokens_match(prev_input_ids_slice, banned_token_seq[:-1]) is False:
+                    # if tokens do not match continue
+                    continue
+
+                banned_tokens_slice.append(banned_token_seq[-1])
+
+            banned_tokens.append(banned_tokens_slice)
+
+        return banned_tokens
+
+    def _set_scores_to_inf_for_banned_tokens(self, scores: torch.Tensor, banned_tokens: List[List[int]]) -> None:
+        """
+        Modifies the scores in place by setting the banned token positions to `-inf`. Banned token is expected to be a
+        list of list of banned tokens to ban in the format [[batch index, vocabulary position],...
+
+        Args:
+            scores: logits distribution of shape (batch size, vocabulary size)
+            banned_tokens: list of list of tokens to ban of length (batch_size)
+        """
+        banned_mask_list = []
+        for idx, batch_banned_tokens in enumerate(banned_tokens):
+            for token in batch_banned_tokens:
+                banned_mask_list.append([idx, token])
+        if not banned_mask_list:
+            return scores
+
+        banned_mask = torch.LongTensor(banned_mask_list)
+        indices = torch.ones(len(banned_mask))
+        # A sparse tensor is generated from a list of coordinates: [[0, 1], [0, 2], [2, 0]]. A conversion to dense tensor generates:
+        # [ 0  1  1 ]
+        # [ 0  0  0 ]
+        # [ 1  0  0 ]
+
+        banned_mask = (
+            torch.sparse.LongTensor(banned_mask.t(), indices, scores.size()).to(scores.device).to_dense().bool()
+        )
+        scores = scores.masked_fill(banned_mask, -float("inf"))
+        return scores

src/transformers/integrations.py

@@ -29,7 +29,8 @@ try:
     wandb.ensure_configured()
     if wandb.api.api_key is None:
         _has_wandb = False
-        wandb.termwarn("W&B installed but not logged in.  Run `wandb login` or set the WANDB_API_KEY env variable.")
+        if os.getenv("WANDB_DISABLED"):
+            logger.warning("W&B installed but not logged in. Run `wandb login` or set the WANDB_API_KEY env variable.")
     else:
         _has_wandb = False if os.getenv("WANDB_DISABLED") else True
 except (ImportError, AttributeError):
@@ -281,7 +282,9 @@ class TensorBoardCallback(TrainerCallback):
                 if hasattr(model, "config") and model.config is not None:
                     model_config_json = model.config.to_json_string()
                     self.tb_writer.add_text("model_config", model_config_json)
-            self.tb_writer.add_hparams(args.to_sanitized_dict(), metric_dict={})
+            # Version of TensorBoard coming from tensorboardX does not have this method.
+            if hasattr(self.tb_writer, "add_hparams"):
+                self.tb_writer.add_hparams(args.to_sanitized_dict(), metric_dict={})
 
     def on_log(self, args, state, control, logs=None, **kwargs):
         if state.is_world_process_zero:

src/transformers/modeling_albert.py

@@ -478,7 +478,7 @@ class AlbertPreTrainedModel(PreTrainedModel):
 @dataclass
 class AlbertForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.AlbertForPreTrainingModel`.
+    Output type of :class:`~transformers.AlbertForPreTraining`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -600,7 +600,7 @@ class AlbertModel(AlbertPreTrainedModel):
             self.pooler = None
             self.pooler_activation = None
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -608,6 +608,9 @@ class AlbertModel(AlbertPreTrainedModel):
     def set_input_embeddings(self, value):
         self.embeddings.word_embeddings = value
 
+    def get_layers(self):
+        return [layer for layer_group in self.encoder.albert_layer_groups for layer in layer_group.albert_layers]
+
     def _resize_token_embeddings(self, new_num_tokens):
         old_embeddings = self.embeddings.word_embeddings
         new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
@@ -650,11 +653,13 @@ class AlbertModel(AlbertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
     ):
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -719,7 +724,7 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
         self.predictions = AlbertMLMHead(config)
         self.sop_classifier = AlbertSOPHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.predictions.decoder
@@ -871,7 +876,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
         self.albert = AlbertModel(config, add_pooling_layer=False)
         self.predictions = AlbertMLMHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.predictions.decoder
@@ -965,7 +970,7 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
         self.dropout = nn.Dropout(config.classifier_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1053,7 +1058,7 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1141,7 +1146,7 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
         self.albert = AlbertModel(config, add_pooling_layer=False)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1240,7 +1245,7 @@ class AlbertForMultipleChoice(AlbertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, 1)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ALBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/modeling_auto.py

@@ -494,8 +494,8 @@ AUTO_MODEL_PRETRAINED_DOCSTRING = r"""
                     - The model is a model provided by the library (loaded with the `shortcut name` string of a
                       pretrained model).
                     - The model was saved using :meth:`~transformers.PreTrainedModel.save_pretrained` and is reloaded
-                      by suppling the save directory.
-                    - The model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a
+                      by supplying the save directory.
+                    - The model is loaded by supplying a local directory as ``pretrained_model_name_or_path`` and a
                       configuration JSON file named `config.json` is found in the directory.
             state_dict (`Dict[str, torch.Tensor]`, `optional`):
                 A state dictionary to use instead of a state dictionary loaded from saved weights file.

src/transformers/modeling_bart.py

@@ -842,7 +842,7 @@ class BartModel(PretrainedBartModel):
         self.encoder = BartEncoder(config, self.shared)
         self.decoder = BartDecoder(config, self.shared)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -875,8 +875,10 @@ class BartModel(PretrainedBartModel):
         if decoder_input_ids is None:
             use_cache = False
 
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         use_cache = use_cache if use_cache is not None else self.config.use_cache
@@ -1084,7 +1086,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
         )
 
     def prepare_inputs_for_generation(
-        self, decoder_input_ids, past, attention_mask, use_cache, encoder_outputs, **kwargs
+        self, decoder_input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
     ):
         return {
             "input_ids": None,  # encoder_outputs is defined. input_ids not needed
@@ -1328,8 +1330,6 @@ class SinusoidalPositionalEmbedding(nn.Embedding):
 
     def __init__(self, num_positions, embedding_dim, padding_idx=None):
         super().__init__(num_positions, embedding_dim)
-        if embedding_dim % 2 != 0:
-            raise NotImplementedError(f"odd embedding_dim {embedding_dim} not supported")
         self.weight = self._init_weight(self.weight)
 
     @staticmethod
@@ -1342,10 +1342,11 @@ class SinusoidalPositionalEmbedding(nn.Embedding):
         position_enc = np.array(
             [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
         )
-        out[:, 0 : dim // 2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))  # This line breaks for odd n_pos
-        out[:, dim // 2 :] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+        out.requires_grad = False  # set early to avoid an error in pytorch-1.8+
+        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
+        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
         out.detach_()
-        out.requires_grad = False
         return out
 
     @torch.no_grad()

src/transformers/modeling_bert.py

@@ -455,32 +455,14 @@ class BertEncoder(nn.Module):
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
-
-            if getattr(self.config, "gradient_checkpointing", False):
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
-                    hidden_states,
-                    attention_mask,
-                    layer_head_mask,
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                )
-            else:
-                layer_outputs = layer_module(
-                    hidden_states,
-                    attention_mask,
-                    layer_head_mask,
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                    output_attentions,
-                )
+            layer_outputs = layer_module(
+                hidden_states,
+                attention_mask,
+                layer_head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions,
+            )
             hidden_states = layer_outputs[0]
             if output_attentions:
                 all_attentions = all_attentions + (layer_outputs[1],)
@@ -606,7 +588,7 @@ class BertPreTrainedModel(PreTrainedModel):
 @dataclass
 class BertForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.BertForPreTrainingModel`.
+    Output type of :class:`~transformers.BertForPreTraining`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -732,7 +714,7 @@ class BertModel(BertPreTrainedModel):
 
         self.pooler = BertPooler(config) if add_pooling_layer else None
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -740,6 +722,9 @@ class BertModel(BertPreTrainedModel):
     def set_input_embeddings(self, value):
         self.embeddings.word_embeddings = value
 
+    def get_layers(self):
+        return self.encoder.layer
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
@@ -780,11 +765,13 @@ class BertModel(BertPreTrainedModel):
             - 1 for tokens that are **not masked**,
             - 0 for tokens that are **masked**.
         """
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -865,7 +852,7 @@ class BertForPreTraining(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.cls = BertPreTrainingHeads(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.cls.predictions.decoder
@@ -978,7 +965,7 @@ class BertLMHeadModel(BertPreTrainedModel):
         self.bert = BertModel(config, add_pooling_layer=False)
         self.cls = BertOnlyMLMHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.cls.predictions.decoder
@@ -1098,7 +1085,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         self.bert = BertModel(config, add_pooling_layer=False)
         self.cls = BertOnlyMLMHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.cls.predictions.decoder
@@ -1204,7 +1191,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.cls = BertOnlyNSPHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=NextSentencePredictorOutput, config_class=_CONFIG_FOR_DOC)
@@ -1298,7 +1285,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1382,7 +1369,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, 1)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(
@@ -1477,7 +1464,7 @@ class BertForTokenClassification(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1567,7 +1554,7 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         self.bert = BertModel(config, add_pooling_layer=False)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/modeling_bert_generation.py

@@ -277,7 +277,7 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
         self.embeddings = BertGenerationEmbeddings(config)
         self.encoder = BertEncoder(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -285,6 +285,9 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
     def set_input_embeddings(self, value):
         self.embeddings.word_embeddings = value
 
+    def get_layers(self):
+        return self.encoder.layer
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
@@ -322,11 +325,13 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
             the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``: ``1`` for
             tokens that are NOT MASKED, ``0`` for MASKED tokens.
         """
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -416,7 +421,7 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
         self.bert = BertGenerationEncoder(config)
         self.lm_head = BertGenerationOnlyLMHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.lm_head.decoder

src/transformers/modeling_ctrl.py

@@ -149,7 +149,7 @@ class MultiHeadAttention(torch.nn.Module):
             k = torch.cat((past_key, k), dim=-2)
             v = torch.cat((past_value, v), dim=-2)
 
-        if use_cache is True:
+        if use_cache:
             present = torch.stack((k, v))
         else:
             present = (None,)
@@ -334,7 +334,7 @@ class CTRLModel(CTRLPreTrainedModel):
         )
         self.layernorm = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.w
@@ -342,6 +342,9 @@ class CTRLModel(CTRLPreTrainedModel):
     def set_input_embeddings(self, new_embeddings):
         self.w = new_embeddings
 
+    def get_layers(self):
+        return self.h
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
@@ -379,9 +382,11 @@ class CTRLModel(CTRLPreTrainedModel):
             past_key_values = kwargs.pop("past")
         assert kwargs == {}, f"Unexpected keyword arguments: {list(kwargs.keys())}."
 
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        use_cache = torch.tensor(use_cache if use_cache is not None else self.config.use_cache)
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
@@ -459,17 +464,18 @@ class CTRLModel(CTRLPreTrainedModel):
         for i, (h, layer_past) in enumerate(zip(self.h, past_key_values)):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
+            layer_head_mask = head_mask[i] if head_mask is not None else None
             outputs = h(
                 hidden_states,
                 mask,
-                layer_past=layer_past,
-                attention_mask=attention_mask,
-                head_mask=head_mask[i],
-                use_cache=use_cache,
-                output_attentions=output_attentions,
+                layer_past,
+                attention_mask,
+                layer_head_mask,
+                use_cache,
+                output_attentions,
             )
             hidden_states, present = outputs[:2]
-            if use_cache is True:
+            if use_cache:
                 presents = presents + (present,)
 
             if output_attentions:
@@ -509,17 +515,17 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         self.transformer = CTRLModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=True)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.lm_head
 
-    def prepare_inputs_for_generation(self, input_ids, past, **kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past=None, use_cache=None, **kwargs):
         # only last token for inputs_ids if past is defined in kwargs
         if past:
             input_ids = input_ids[:, -1].unsqueeze(-1)
 
-        return {"input_ids": input_ids, "past_key_values": past, "use_cache": kwargs["use_cache"]}
+        return {"input_ids": input_ids, "past_key_values": past, "use_cache": use_cache}
 
     @add_start_docstrings_to_model_forward(CTRL_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/modeling_deberta.py

@@ -406,9 +406,9 @@ class DebertaEncoder(nn.Module):
                 next_kv,
                 attention_mask,
                 output_attentions,
-                query_states=query_states,
-                relative_pos=relative_pos,
-                rel_embeddings=rel_embeddings,
+                query_states,
+                relative_pos,
+                rel_embeddings,
             )
             if output_attentions:
                 hidden_states, att_m = hidden_states
@@ -843,7 +843,7 @@ class DebertaModel(DebertaPreTrainedModel):
         self.encoder = DebertaEncoder(config)
         self.z_steps = 0
         self.config = config
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -851,6 +851,9 @@ class DebertaModel(DebertaPreTrainedModel):
     def set_input_embeddings(self, new_embeddings):
         self.embeddings.word_embeddings = new_embeddings
 
+    def get_layers(self):
+        return self.encoder.layer
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
@@ -876,11 +879,13 @@ class DebertaModel(DebertaPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
     ):
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -968,7 +973,7 @@ class DebertaForSequenceClassification(DebertaPreTrainedModel):
         drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
         self.dropout = StableDropout(drop_out)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.deberta.get_input_embeddings()

src/transformers/modeling_distilbert.py

@@ -306,9 +306,8 @@ class Transformer(nn.Module):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_state,)
 
-            layer_outputs = layer_module(
-                x=hidden_state, attn_mask=attn_mask, head_mask=head_mask[i], output_attentions=output_attentions
-            )
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            layer_outputs = layer_module(hidden_state, attn_mask, layer_head_mask, output_attentions)
             hidden_state = layer_outputs[-1]
 
             if output_attentions:
@@ -420,7 +419,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
         self.embeddings = Embeddings(config)  # Embeddings
         self.transformer = Transformer(config)  # Encoder
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -428,6 +427,9 @@ class DistilBertModel(DistilBertPreTrainedModel):
     def set_input_embeddings(self, new_embeddings):
         self.embeddings.word_embeddings = new_embeddings
 
+    def get_layers(self):
+        return self.transformer.layer
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
@@ -454,11 +456,13 @@ class DistilBertModel(DistilBertPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
     ):
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -502,7 +506,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         self.vocab_layer_norm = nn.LayerNorm(config.dim, eps=1e-12)
         self.vocab_projector = nn.Linear(config.dim, config.vocab_size)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
         self.mlm_loss_fct = nn.CrossEntropyLoss()
 
@@ -593,7 +597,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
         self.classifier = nn.Linear(config.dim, config.num_labels)
         self.dropout = nn.Dropout(config.seq_classif_dropout)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DISTILBERT_INPUTS_DOCSTRING.format("batch_size, num_choices"))
     @add_code_sample_docstrings(
@@ -674,7 +678,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
         assert config.num_labels == 2
         self.dropout = nn.Dropout(config.qa_dropout)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DISTILBERT_INPUTS_DOCSTRING.format("batch_size, num_choices"))
     @add_code_sample_docstrings(
@@ -770,7 +774,7 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
         self.dropout = nn.Dropout(config.dropout)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DISTILBERT_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -854,7 +858,7 @@ class DistilBertForMultipleChoice(DistilBertPreTrainedModel):
         self.classifier = nn.Linear(config.dim, 1)
         self.dropout = nn.Dropout(config.seq_classif_dropout)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(
         DISTILBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")

src/transformers/modeling_dpr.py

@@ -157,7 +157,7 @@ class DPREncoder(PreTrainedModel):
         self.projection_dim = config.projection_dim
         if self.projection_dim > 0:
             self.encode_proj = nn.Linear(self.bert_model.config.hidden_size, config.projection_dim)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def forward(
         self,
@@ -199,8 +199,8 @@ class DPREncoder(PreTrainedModel):
             return self.encode_proj.out_features
         return self.bert_model.config.hidden_size
 
-    def init_weights(self):
-        self.bert_model.init_weights()
+    def init_weights_and_layers(self):
+        self.bert_model.init_weights_and_layers()
         if self.projection_dim > 0:
             self.encode_proj.apply(self.bert_model._init_weights)
 
@@ -214,7 +214,7 @@ class DPRSpanPredictor(PreTrainedModel):
         self.encoder = DPREncoder(config)
         self.qa_outputs = nn.Linear(self.encoder.embeddings_size, 2)
         self.qa_classifier = nn.Linear(self.encoder.embeddings_size, 1)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def forward(
         self,
@@ -261,8 +261,8 @@ class DPRSpanPredictor(PreTrainedModel):
             attentions=outputs.attentions,
         )
 
-    def init_weights(self):
-        self.encoder.init_weights()
+    def init_weights_and_layers(self):
+        self.encoder.init_weights_and_layers()
 
 
 ##################
@@ -281,8 +281,8 @@ class DPRPretrainedContextEncoder(PreTrainedModel):
     base_model_prefix = "ctx_encoder"
     authorized_missing_keys = [r"position_ids"]
 
-    def init_weights(self):
-        self.ctx_encoder.init_weights()
+    def init_weights_and_layers(self):
+        self.ctx_encoder.init_weights_and_layers()
 
 
 class DPRPretrainedQuestionEncoder(PreTrainedModel):
@@ -296,8 +296,8 @@ class DPRPretrainedQuestionEncoder(PreTrainedModel):
     base_model_prefix = "question_encoder"
     authorized_missing_keys = [r"position_ids"]
 
-    def init_weights(self):
-        self.question_encoder.init_weights()
+    def init_weights_and_layers(self):
+        self.question_encoder.init_weights_and_layers()
 
 
 class DPRPretrainedReader(PreTrainedModel):
@@ -311,8 +311,8 @@ class DPRPretrainedReader(PreTrainedModel):
     base_model_prefix = "span_predictor"
     authorized_missing_keys = [r"position_ids"]
 
-    def init_weights(self):
-        self.span_predictor.encoder.init_weights()
+    def init_weights_and_layers(self):
+        self.span_predictor.encoder.init_weights_and_layers()
         self.span_predictor.qa_classifier.apply(self.span_predictor.encoder.bert_model._init_weights)
         self.span_predictor.qa_outputs.apply(self.span_predictor.encoder.bert_model._init_weights)
 
@@ -434,7 +434,7 @@ class DPRContextEncoder(DPRPretrainedContextEncoder):
         super().__init__(config)
         self.config = config
         self.ctx_encoder = DPREncoder(config)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DPR_ENCODERS_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=DPRContextEncoderOutput, config_class=_CONFIG_FOR_DOC)
@@ -460,8 +460,10 @@ class DPRContextEncoder(DPRPretrainedContextEncoder):
             >>> embeddings = model(input_ids).pooler_output
         """
 
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
@@ -512,7 +514,7 @@ class DPRQuestionEncoder(DPRPretrainedQuestionEncoder):
         super().__init__(config)
         self.config = config
         self.question_encoder = DPREncoder(config)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DPR_ENCODERS_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=DPRQuestionEncoderOutput, config_class=_CONFIG_FOR_DOC)
@@ -537,8 +539,10 @@ class DPRQuestionEncoder(DPRPretrainedQuestionEncoder):
             >>> input_ids = tokenizer("Hello, is my dog cute ?", return_tensors='pt')["input_ids"]
             >>> embeddings = model(input_ids).pooler_output
         """
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
@@ -589,7 +593,7 @@ class DPRReader(DPRPretrainedReader):
         super().__init__(config)
         self.config = config
         self.span_predictor = DPRSpanPredictor(config)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(DPR_READER_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=DPRReaderOutput, config_class=_CONFIG_FOR_DOC)
@@ -622,8 +626,10 @@ class DPRReader(DPRPretrainedReader):
             >>> relevance_logits = outputs.relevance_logits
 
         """
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict

src/transformers/modeling_electra.py

@@ -451,32 +451,14 @@ class ElectraEncoder(nn.Module):
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
-
-            if getattr(self.config, "gradient_checkpointing", False):
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
-                    hidden_states,
-                    attention_mask,
-                    layer_head_mask,
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                )
-            else:
-                layer_outputs = layer_module(
-                    hidden_states,
-                    attention_mask,
-                    layer_head_mask,
-                    encoder_hidden_states,
-                    encoder_attention_mask,
-                    output_attentions,
-                )
+            layer_outputs = layer_module(
+                hidden_states,
+                attention_mask,
+                layer_head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions,
+            )
             hidden_states = layer_outputs[0]
             if output_attentions:
                 all_attentions = all_attentions + (layer_outputs[1],)
@@ -555,7 +537,7 @@ class ElectraPreTrainedModel(PreTrainedModel):
 @dataclass
 class ElectraForPreTrainingOutput(ModelOutput):
     """
-    Output type of :class:`~transformers.ElectraForPreTrainingModel`.
+    Output type of :class:`~transformers.ElectraForPreTraining`.
 
     Args:
         loss (`optional`, returned when ``labels`` is provided, ``torch.FloatTensor`` of shape :obj:`(1,)`):
@@ -677,7 +659,7 @@ class ElectraModel(ElectraPreTrainedModel):
 
         self.encoder = ElectraEncoder(config)
         self.config = config
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_input_embeddings(self):
         return self.embeddings.word_embeddings
@@ -685,6 +667,9 @@ class ElectraModel(ElectraPreTrainedModel):
     def set_input_embeddings(self, value):
         self.embeddings.word_embeddings = value
 
+    def get_layers(self):
+        return self.encoder.layer
+
     def _prune_heads(self, heads_to_prune):
         """
         Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
@@ -712,11 +697,13 @@ class ElectraModel(ElectraPreTrainedModel):
         output_hidden_states=None,
         return_dict=None,
     ):
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
+        output_attentions = torch.tensor(
+            output_attentions if output_attentions is not None else self.config.output_attentions
+        )
+        output_hidden_states = torch.tensor(
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        return_dict = torch.tensor(return_dict if return_dict is not None else self.config.use_return_dict)
 
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
@@ -789,7 +776,7 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
         self.electra = ElectraModel(config)
         self.classifier = ElectraClassificationHead(config)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -871,7 +858,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
 
         self.electra = ElectraModel(config)
         self.discriminator_predictions = ElectraDiscriminatorPredictions(config)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=ElectraForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -966,7 +953,7 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
         self.generator_predictions = ElectraGeneratorPredictions(config)
 
         self.generator_lm_head = nn.Linear(config.embedding_size, config.vocab_size)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     def get_output_embeddings(self):
         return self.generator_lm_head
@@ -1058,7 +1045,7 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
         self.electra = ElectraModel(config)
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1145,7 +1132,7 @@ class ElectraForQuestionAnswering(ElectraPreTrainedModel):
         self.electra = ElectraModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1246,7 +1233,7 @@ class ElectraForMultipleChoice(ElectraPreTrainedModel):
         self.sequence_summary = SequenceSummary(config)
         self.classifier = nn.Linear(config.hidden_size, 1)
 
-        self.init_weights()
+        self.init_weights_and_layers()
 
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
     @add_code_sample_docstrings(