Fix some of NielsRogge's comments

TAPAS
Style
2020-11-12 17:07:27 -05:00 · 2020-11-11 22:52:59 -05:00 · 2020-11-11 21:21:04 -05:00 · 2020-11-04 10:42:05 +01:00 · 2020-11-04 10:36:59 +01:00 · 2020-11-04 10:36:59 +01:00
32 changed files with 7909 additions and 270 deletions
@@ -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:
@@ -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
+          
@@ -4,14 +4,186 @@ on:
  push:
    branches:
      - ci_*
-      - trigger-scheduled
  repository_dispatch:
  schedule:
    - cron: "0 0 * * *"

 jobs:
+  run_all_tests_torch_gpu:
+    runs-on: [self-hosted, single-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v  1.1-slow_tests_torch_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[torch,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+          pip list
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          python -c "import torch; print('Cuda available:', torch.cuda.is_available())"
+          python -c "import torch; print('Number of GPUs available:', torch.cuda.device_count())"
+
+      - name: Run all tests on GPU
+        env:
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --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: 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_gpu examples
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/examples_torch_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_torch_pipeline_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_pipeline_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_gpu_test_reports
+          path: reports
+
+
+  run_all_tests_tf_gpu:
+    runs-on: [self-hosted, single-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-slow_tests_tf_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[tf,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+          pip list
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
+          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
+
+      - name: Run all tests on GPU
+        env:
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          python -m pytest -n 1 --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: 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() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_tf_gpu_test_reports
+          path: reports
+          
  run_all_tests_torch_multiple_gpu:
-    runs-on: [self-hosted, multi-gpu-tests]
+    runs-on: [self-hosted, multi-gpu]
    steps:
      - uses: actions/checkout@v2

@@ -61,11 +233,90 @@ 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_tests_torch_failures_short.txt
+        run: cat reports/tests_torch_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
+          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_torch_pipeline_multiple_gpu tests
+
+      - name: Failure short reports
+        if: ${{ always() }}
+        run: cat reports/tests_torch_pipeline_multiple_gpu_failures_short.txt
+
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v2
+        with:
+          name: run_all_tests_torch_multi_gpu_test_reports
+          path: reports
+
+  run_all_tests_tf_multiple_gpu:
+    runs-on: [self-hosted, multi-gpu]
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Loading cache.
+        uses: actions/cache@v2
+        id: cache
+        with:
+          path: .env
+          key: v1.1-slow_tests_tf_multi_gpu-${{ hashFiles('setup.py') }}
+
+      - name: Python version
+        run: |
+          which python
+          python --version
+          pip --version
+
+      - name: Current dir
+        run: pwd
+      - run: nvidia-smi
+
+      - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .env
+          source .env/bin/activate
+          which python
+          python --version
+          pip --version
+
+      - name: Install dependencies
+        run: |
+          source .env/bin/activate
+          pip install --upgrade pip
+          pip install .[tf,sklearn,testing,onnxruntime]
+          pip install git+https://github.com/huggingface/datasets
+          pip list
+
+      - name: Are GPUs recognized by our DL frameworks
+        run: |
+          source .env/bin/activate
+          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('TF GPUs available:', bool(tf.config.list_physical_devices('GPU')))"
+          TF_CPP_MIN_LOG_LEVEL=3 python -c "import tensorflow as tf; print('Number of TF GPUs available:', len(tf.config.list_physical_devices('GPU')))"
+
+      - name: Run all tests on GPU
+        env:
+          OMP_NUM_THREADS: 1
+          RUN_SLOW: yes
+        run: |
+          source .env/bin/activate
+          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() }}
@@ -76,16 +327,16 @@ 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_tf_pipelines_multiple_gpu tests
+          
      - name: Failure short reports
        if: ${{ always() }}
-        run: cat reports/report_tests_torch_pipeline_failures_short.txt
+        run: cat reports/tests_tf_multiple_gpu_pipelines_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
+          name: run_all_tests_tf_multi_gpu_test_reports
          path: reports
-
+          
@@ -188,6 +188,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 ultilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
 1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+1. **[TAPAS](https://huggingface.co/transformers/master/model_doc/tapas.html)** released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
@@ -222,4 +223,4 @@ We now have a [paper](https://arxiv.org/abs/1910.03771) you can cite for the
  year={2019},
  volume={abs/1910.03771}
 }
-```
+```
@@ -145,22 +145,25 @@ conversion utilities for the following models:
 27. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
    Unified Text-to-Text Transformer <https://arxiv.org/abs/1910.10683>`__ by Colin Raffel and Noam Shazeer and Adam
    Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
-28. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+28. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
+    Pre-training <https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller,
+    Francesco Piccinno and Julian Martin Eisenschlos.
+29. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
    Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
    Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-29. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+30. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
    Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis Conneau.
-30. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+31. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
    Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
    Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-31. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+32. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
    Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
    Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
    Zettlemoyer and Veselin Stoyanov.
-32. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
+33. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `XLNet: Generalized Autoregressive
    Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
    Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
-33. `Other community models <https://huggingface.co/models>`__, contributed by the `community
+34. `Other community models <https://huggingface.co/models>`__, contributed by the `community
    <https://huggingface.co/users>`__.

 .. toctree::
@@ -258,6 +261,7 @@ conversion utilities for the following models:
    model_doc/roberta
    model_doc/squeezebert
    model_doc/t5
+    model_doc/tapas
    model_doc/transformerxl
    model_doc/xlm
    model_doc/xlmprophetnet
@@ -0,0 +1,162 @@
+TAPAS
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The TAPAS model was proposed in `TAPAS: Weakly Supervised Table Parsing via Pre-training
+<https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and
+Julian Martin Eisenschlos. It's a BERT-based model specifically designed (and pre-trained) for answering questions
+about tabular data. Compared to BERT, TAPAS uses relative position embeddings and has 7 token types that encode tabular
+structure. TAPAS is pre-trained on the masked language modeling (MLM) objective on a large dataset comprising millions
+of tables from English Wikipedia and corresponding texts. For question answering, TAPAS has 2 heads on top: a cell
+selection head and an aggregation head, for (optionally) performing aggregations (such as counting or summing) among
+selected cells. TAPAS has been fine-tuned on several datasets: SQA (Sequential Question Answering by Microsoft), WTQ
+(Wiki Table Questions by Stanford University) and WikiSQL (by Salesforce). It achieves state-of-the-art on both SQA and
+WTQ, while having comparable performance to SOTA on WikiSQL, with a much simpler architecture.
+
+The abstract from the paper is the following:
+
+*Answering natural language questions over tables is usually seen as a semantic parsing task. To alleviate the
+collection cost of full logical forms, one popular approach focuses on weak supervision consisting of denotations
+instead of logical forms. However, training semantic parsers from weak supervision poses difficulties, and in addition,
+the generated logical forms are only used as an intermediate step prior to retrieving the denotation. In this paper, we
+present TAPAS, an approach to question answering over tables without generating logical forms. TAPAS trains from weak
+supervision, and predicts the denotation by selecting table cells and optionally applying a corresponding aggregation
+operator to such selection. TAPAS extends BERT's architecture to encode tables as input, initializes from an effective
+joint pre-training of text segments and tables crawled from Wikipedia, and is trained end-to-end. We experiment with
+three different semantic parsing datasets, and find that TAPAS outperforms or rivals semantic parsing models by
+improving state-of-the-art accuracy on SQA from 55.1 to 67.2 and performing on par with the state-of-the-art on WIKISQL
+and WIKITQ, but with a simpler model architecture. We additionally find that transfer learning, which is trivial in our
+setting, from WIKISQL to WIKITQ, yields 48.7 accuracy, 4.2 points above the state-of-the-art.*
+
+In addition, the authors have further pre-trained TAPAS to recognize table entailment, by creating a balanced dataset
+of millions of automatically created training examples which are learned in an intermediate step prior to fine-tuning.
+The authors of TAPAS call this further pre-training intermediate pre-training (since TAPAS is first pre-trained on MLM,
+and then on another dataset). They found that intermediate pre-training further improves performance on SQA, achieving
+a new state-of-the-art as well as state-of-the-art on TabFact, a large-scale dataset with 16k Wikipedia tables for
+table entailment (a binary classification task). For more details, see their new paper: `Understanding tables with
+intermediate pre-training <https://arxiv.org/abs/2010.00571>`__ by Julian Martin Eisenschlos, Syrine Krichene and
+Thomas Müller.
+
+The original code can be found `here <https://github.com/google-research/tapas>`__.
+
+Tips:
+
+- TAPAS is a model that uses relative position embeddings by default (restarting the position embeddings at every cell
+  of the table). According to the authors, this usually results in a slightly better performance, and allows you to
+  encode longer sequences without running out of embeddings. If you don't want this, you can set the
+  `reset_position_index_per_cell` parameter of :class:`~transformers.TapasConfig` to False.
+- TAPAS has checkpoints fine-tuned on SQA, which are capable of answering questions related to a table in a
+  conversational set-up. This means that you can ask follow-up questions such as "what is his age?" related to the
+  previous question. Note that the forward pass of TAPAS is a bit different in case of a conversational set-up: in that
+  case, you have to feed every training example one by one to the model, such that the `prev_label_ids` token type ids
+  can be overwritten by the predicted `label_ids` of the model to the previous question.
+- TAPAS is similar to BERT and therefore relies on the masked language modeling (MLM) objective. It is therefore
+  efficient at predicting masked tokens and at NLU in general, but is not optimal for text generation. Models trained
+  with a causal language modeling (CLM) objective are better in that regard.
+
+
+Usage
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+If you just want to perform inference (i.e. making predictions) in a non-conversational setup, you can do the
+following:
+
+.. code-block::
+
+        >>> from transformers import TapasTokenizer, TapasForQuestionAnswering
+        >>> import pandas as pd 
+
+        >>> model_name = 'tapas-base-finetuned-wtq'
+        >>> model = TapasForQuestionAnswering.from_pretrained(model_name)
+        >>> tokenizer = TapasTokenizer.from_pretrained(model_name)
+
+        >>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
+        >>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
+        >>> table = pd.Dataframe(data)
+        >>> inputs = tokenizer(table, queries, return_tensors='pt')
+        >>> logits, logits_agg = model(**inputs)
+        >>> answer_coordinates_batch, aggregation_predictions = tokenizer.convert_logits_to_predictions(inputs, logits, logits_agg)
+
+        >>> # let's print out the results:
+        >>> id2aggregation = {0: "NONE", 1: "SUM", 2: "AVERAGE", 3:"COUNT"}
+        >>> aggregation_predictions_string = [id2aggregation[x] for x in aggregation_predictions]
+
+        >>> answers = []
+        >>> for coordinates in answer_coordinates_batch:
+        ...   if len(coordinates) == 1:
+        ...     # only a single cell:
+        ...     answers.append(df.iat[coordinates[0]])
+        ...   else:
+        ...     # multiple cells
+        ...     cell_values = []
+        ...     for coordinate in coordinates:
+        ...        cell_values.append(df.iat[coordinate])
+        ...     answers.append(", ".join(cell_values))
+
+        >>> display(df)
+        >>> print("")
+        >>> for query, answer, predicted_agg in zip(queries, answers, aggregation_predictions_string):
+        ...   print(query)
+        ...   if predicted_agg == "NONE":
+        ...     print("Predicted answer: " + answer)
+        ...   else:
+        ...     print("Predicted answer: " + predicted_agg + " > " + answer)    
+        When was Brad Pitt born?
+        Predicted answer: 18 december 1963
+        Which actor appeared in the least number of movies?
+        Predicted answer: Leonardo Di Caprio
+        What is the average number of movies?
+        Predicted answer: AVERAGE > 87, 53, 69
+
+
+Tapas specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_tapas.TableQuestionAnsweringOutput
+    :members:
+
+
+TapasConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasConfig
+    :members:
+
+
+TapasTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasTokenizer
+    :members: convert_logits_to_predictions, save_vocabulary
+
+
+TapasModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasModel
+    :members:
+
+
+TapasForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForMaskedLM
+    :members:
+
+
+TapasForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForSequenceClassification
+    :members: forward
+
+
+TapasForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForQuestionAnswering
+    :members:
+
+
@@ -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)
@@ -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:
    """
@@ -30,7 +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):
@@ -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())}"},
+    )
@@ -61,6 +61,7 @@ from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, Re
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
 from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
+from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
@@ -190,6 +191,7 @@ from .tokenization_rag import RagTokenizer
 from .tokenization_retribert import RetriBertTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_squeezebert import SqueezeBertTokenizer
+from .tokenization_tapas import TapasTokenizer
 from .tokenization_transfo_xl import TransfoXLCorpus, TransfoXLTokenizer
 from .tokenization_utils import PreTrainedTokenizer
 from .tokenization_utils_base import (
@@ -284,6 +286,7 @@ if is_torch_available():
        DataCollatorForNextSentencePrediction,
        DataCollatorForPermutationLanguageModeling,
        DataCollatorForSOP,
+        DataCollatorForTokenClassification,
        DataCollatorForWholeWordMask,
        DataCollatorWithPadding,
        default_data_collator,
@@ -557,6 +560,14 @@ if is_torch_available():
        T5PreTrainedModel,
        load_tf_weights_in_t5,
    )
+    from .modeling_tapas import (
+        TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TapasForMaskedLM,
+        TapasForQuestionAnswering,
+        TapasForSequenceClassification,
+        TapasModel,
+        load_tf_weights_in_tapas,
+    )
    from .modeling_transfo_xl import (
        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
        AdaptiveEmbedding,
@@ -130,6 +130,13 @@ class ConvertCommand(BaseTransformersCLICommand):
                raise ImportError(IMPORT_ERROR_MESSAGE)

            convert_gpt2_checkpoint_to_pytorch(self._tf_checkpoint, self._config, self._pytorch_dump_output)
+        elif self._model_type == "tapas":
+            try:
+                from transformers.convert_tapas_original_tf_checkpoint_to_pytorch import (
+                    convert_tf_checkpoint_to_pytorch,
+                )
+            except ImportError:
+                raise ImportError(IMPORT_ERROR_MESSAGE)
        elif self._model_type == "xlnet":
            try:
                from transformers.convert_xlnet_original_tf_checkpoint_to_pytorch import (
@@ -48,6 +48,7 @@ from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, Re
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
 from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
+from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
@@ -88,6 +89,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
        SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
        XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
        PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP,
    ]
    for key, value, in pretrained_map.items()
 )
@@ -131,6 +133,7 @@ CONFIG_MAPPING = OrderedDict(
        ("dpr", DPRConfig),
        ("layoutlm", LayoutLMConfig),
        ("rag", RagConfig),
+        ("tapas", TapasConfig),
    ]
 )

@@ -172,6 +175,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
        ("rag", "RAG"),
        ("xlm-prophetnet", "XLMProphetNet"),
        ("prophetnet", "ProphetNet"),
+        ("tapas", "TAPAS"),
    ]
 )

@@ -0,0 +1,209 @@
+# coding=utf-8
+# Copyright 2020 Google Research and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TAPAS configuration. Adds additional hyperparameters to the configuration of BERT."""
+
+
+from .configuration_utils import PretrainedConfig
+
+
+TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP = {"tapas-base": "", "tapas-large": ""}  # to be added  # to be added
+
+
+class TapasConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.TapasModel`. It is used to
+    instantiate a TAPAS model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the TAPAS `tapas-base-finetuned-sqa`
+    architecture. Configuration objects inherit from :class:`~transformers.PreTrainedConfig` and can be used to control
+    the model outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+    Hyperparameters additional to BERT are taken from run_task_main.py and hparam_utils.py of the original
+    implementation. Original implementation available at https://github.com/google-research/tapas/tree/master.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 30522):
+            Vocabulary size of the TAPAS model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.TapasModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 3072):
+            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.
+        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):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_sizes (:obj:`List[int]`, `optional`, defaults to [3, 256, 256, 2, 256, 256, 10]):
+            The vocabulary sizes of the :obj:`token_type_ids` passed when calling :class:`~transformers.TapasModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        positive_label_weight (:obj:`float`, `optional`, defaults to 10.0):
+            Weight for positive labels.
+        num_aggregation_labels (:obj:`int`, `optional`, defaults to 0):
+            The number of aggregation operators to predict.
+        aggregation_loss_weight (:obj:`float`, `optional`, defaults to 1.0):
+            Importance weight for the aggregation loss.
+        use_answer_as_supervision (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            Whether to use the answer as the only supervision for aggregation examples.
+        answer_loss_importance (:obj:`float`, `optional`, defaults to 1.0):
+            Importance weight for the regression loss.
+        use_normalized_answer_loss (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Normalize loss by max of predicted and expected value.
+        huber_loss_delta: (:obj:`float`, `optional`, defaults to None):
+            Delta parameter used to calculate the regression loss.
+        temperature: (:obj:`float`, `optional`, defaults to 1.0):
+            Scales cell logits to control the skewness of probabilities.
+        aggregation_temperature: (:obj:`float`, `optional`, defaults to 1.0):
+            Scales aggregation logits to control the skewness of probabilities.
+        use_gumbel_for_cells: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Applies Gumbel-Softmax to cell selection.
+        use_gumbel_for_aggregation: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Applies Gumbel-Softmax to aggregation selection.
+        average_approximation_function: (:obj:`string`, `optional`, defaults to :obj:`"ratio"`):
+            Method to calculate expected average of cells in the relaxed case.
+        cell_selection_preference: (:obj:`float`, `optional`, defaults to None):
+            Preference for cell selection in ambiguous cases. Only applicable in case of weak supervision for
+            aggregation (WTQ, WikiSQL). If the total mass of the aggregation probabilities (excluding the "NONE"
+            operator) is higher than this hyperparameter, then aggregation is predicted for an example.
+        answer_loss_cutoff: (:obj:`float`, `optional`, defaults to None):
+            Ignore examples with answer loss larger than cutoff.
+        max_num_rows: (:obj:`int`, `optional`, defaults to 64):
+            Maximum number of rows.
+        max_num_columns: (:obj:`int`, `optional`, defaults to 32):
+            Maximum number of columns.
+        average_logits_per_cell: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to average logits per cell.
+        select_one_column: (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to constrain the model to only select cells from a single column.
+        allow_empty_column_selection: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Allow not to select any column.
+        init_cell_selection_weights_to_zero: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to initialize cell selection weights to 0 so that the initial probabilities are 50%.
+        reset_position_index_per_cell: (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Restart position indexes at every cell.
+        disable_per_token_loss: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Disable any (strong or weak) supervision on cells.
+        span_prediction: (:obj:`string`, `optional`, defaults to :obj:`"none"`):
+            Span selection mode to use. Currently only "none" is supported.
+
+    Example::
+
+        >>> from transformers import TapasModel, TapasConfig
+        >>> # Initializing a Tapas configuration
+        >>> configuration = TapasConfig()
+        >>> # Initializing a model from the configuration
+        >>> model = TapasModel(configuration)
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "tapas"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=1024,
+        type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10],
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        gradient_checkpointing=False,
+        positive_label_weight=10.0,
+        num_aggregation_labels=0,
+        aggregation_loss_weight=1.0,
+        use_answer_as_supervision=None,
+        answer_loss_importance=1.0,
+        use_normalized_answer_loss=False,
+        huber_loss_delta=None,
+        temperature=1.0,
+        aggregation_temperature=1.0,
+        use_gumbel_for_cells=False,
+        use_gumbel_for_aggregation=False,
+        average_approximation_function="ratio",
+        cell_selection_preference=None,
+        answer_loss_cutoff=None,
+        max_num_rows=64,
+        max_num_columns=32,
+        average_logits_per_cell=False,
+        select_one_column=True,
+        allow_empty_column_selection=False,
+        init_cell_selection_weights_to_zero=False,
+        reset_position_index_per_cell=True,
+        disable_per_token_loss=False,
+        span_prediction="none",
+        **kwargs
+    ):
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        # BERT hyperparameters (with updated max_position_embeddings and type_vocab_sizes)
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_sizes = type_vocab_sizes
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.gradient_checkpointing = gradient_checkpointing
+
+        # Fine-tuning task hyperparameters
+        self.positive_label_weight = positive_label_weight
+        self.num_aggregation_labels = num_aggregation_labels
+        self.aggregation_loss_weight = aggregation_loss_weight
+        self.use_answer_as_supervision = use_answer_as_supervision
+        self.answer_loss_importance = answer_loss_importance
+        self.use_normalized_answer_loss = use_normalized_answer_loss
+        self.huber_loss_delta = huber_loss_delta
+        self.temperature = temperature
+        self.aggregation_temperature = aggregation_temperature
+        self.use_gumbel_for_cells = use_gumbel_for_cells
+        self.use_gumbel_for_aggregation = use_gumbel_for_aggregation
+        self.average_approximation_function = average_approximation_function
+        self.cell_selection_preference = cell_selection_preference
+        self.answer_loss_cutoff = answer_loss_cutoff
+        self.max_num_rows = max_num_rows
+        self.max_num_columns = max_num_columns
+        self.average_logits_per_cell = average_logits_per_cell
+        self.select_one_column = select_one_column
+        self.allow_empty_column_selection = allow_empty_column_selection
+        self.init_cell_selection_weights_to_zero = init_cell_selection_weights_to_zero
+        self.reset_position_index_per_cell = reset_position_index_per_cell
+        self.disable_per_token_loss = disable_per_token_loss
+        self.span_prediction = span_prediction
@@ -0,0 +1,84 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert TAPAS checkpoint."""
+
+
+import argparse
+
+import torch
+
+from transformers import (
+    TapasConfig,
+    TapasForQuestionAnswering,
+    TapasForSequenceClassification,
+    load_tf_weights_in_tapas,
+)
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, tapas_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+
+    # WTQ config
+    # config = TapasConfig(# run_task_main.py hparams
+    #         num_aggregation_labels = 4,
+    #         use_answer_as_supervision = True,
+    #         # hparam_utils.py hparams
+    #         answer_loss_cutoff = 0.664694,
+    #         cell_select_pref = 0.207951,
+    #         huber_loss_delta = 0.121194,
+    #         init_cell_selection_weights_to_zero = True,
+    #         select_one_column = True,
+    #         allow_empty_column_selection = False,
+    #         temperature = 0.0352513)
+
+    # SQA config
+    config = TapasConfig()
+
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    # model = TapasForMaskedLM(config)
+    model = TapasForQuestionAnswering(config)
+    # model = TapasForSequenceClassification(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_tapas(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
+    )
+    parser.add_argument(
+        "--tapas_config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The config json file corresponding to the pre-trained TAPAS model. \n"
+        "This specifies the model architecture.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.tapas_config_file, args.pytorch_dump_path)
@@ -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:
    """
@@ -193,6 +193,20 @@ except ImportError:
    _tokenizers_available = False


+try:
+    import torch_scatter
+
+    # Check we're not importing a "torch_scatter" directory somewhere
+    _scatter_available = hasattr(torch_scatter, "__version__") and hasattr(torch_scatter, "scatter")
+    if _scatter_available:
+        logger.debug(f"Succesfully imported torch-scatter version {torch_scatter.__version__}")
+    else:
+        logger.debug("Imported a torch_scatter object but this doesn't seem to be the torch-scatter library.")
+
+except ImportError:
+    _scatter_available = False
+
+
 default_cache_path = os.path.join(torch_cache_home, "transformers")


@@ -289,6 +303,14 @@ def torch_only_method(fn):


 # docstyle-ignore
+def is_sklearn_available():
+    return _has_sklearn
+
+
+def is_scatter_available():
+    return _scatter_available
+
+
 DATASETS_IMPORT_ERROR = """
 {0} requires the 🤗 Datasets library but it was not found in your environment. You can install it with:
 ```
@@ -368,6 +390,12 @@ FLAX_IMPORT_ERROR = """
 installation page: https://github.com/google/flax and follow the ones that match your environment.
 """

+SCATTER_IMPORT_ERROR = """
+{0} requires the torch-scatter library but it was not found in your environment. You can install it with pip as
+explained here: https://github.com/rusty1s/pytorch_scatter.
+
+"""
+

 def requires_datasets(obj):
    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
@@ -417,6 +445,12 @@ def requires_sentencepiece(obj):
        raise ImportError(SENTENCEPIECE_IMPORT_ERROR.format(name))


+def requires_scatter(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_scatter_available():
+        raise ImportError(SCATTER_IMPORT_ERROR.format(name))
+
+
 def add_start_docstrings(*docstr):
    def docstring_decorator(fn):
        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
@@ -49,6 +49,7 @@ from .configuration_auto import (
    RobertaConfig,
    SqueezeBertConfig,
    T5Config,
+    TapasConfig,
    TransfoXLConfig,
    XLMConfig,
    XLMProphetNetConfig,
@@ -188,6 +189,7 @@ from .modeling_squeezebert import (
    SqueezeBertModel,
 )
 from .modeling_t5 import T5ForConditionalGeneration, T5Model
+from .modeling_tapas import TapasForMaskedLM, TapasForQuestionAnswering, TapasForSequenceClassification, TapasModel
 from .modeling_transfo_xl import TransfoXLLMHeadModel, TransfoXLModel
 from .modeling_xlm import (
    XLMForMultipleChoice,
@@ -229,6 +231,7 @@ MODEL_MAPPING = OrderedDict(
    [
        (RetriBertConfig, RetriBertModel),
        (T5Config, T5Model),
+        (TapasConfig, TapasModel),
        (DistilBertConfig, DistilBertModel),
        (AlbertConfig, AlbertModel),
        (CamembertConfig, CamembertModel),
@@ -265,6 +268,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
        (LayoutLMConfig, LayoutLMForMaskedLM),
        (RetriBertConfig, RetriBertModel),
        (T5Config, T5ForConditionalGeneration),
+        (TapasConfig, TapasForMaskedLM),
        (DistilBertConfig, DistilBertForMaskedLM),
        (AlbertConfig, AlbertForPreTraining),
        (CamembertConfig, CamembertForMaskedLM),
@@ -292,6 +296,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
    [
        (LayoutLMConfig, LayoutLMForMaskedLM),
        (T5Config, T5ForConditionalGeneration),
+        (TapasConfig, TapasForMaskedLM),
        (DistilBertConfig, DistilBertForMaskedLM),
        (AlbertConfig, AlbertForMaskedLM),
        (CamembertConfig, CamembertForMaskedLM),
@@ -351,6 +356,7 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
        (LongformerConfig, LongformerForMaskedLM),
        (RobertaConfig, RobertaForMaskedLM),
        (SqueezeBertConfig, SqueezeBertForMaskedLM),
+        (TapasConfig, TapasForMaskedLM),
        (BertConfig, BertForMaskedLM),
        (MobileBertConfig, MobileBertForMaskedLM),
        (FlaubertConfig, FlaubertWithLMHeadModel),
@@ -396,6 +402,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
        (DebertaConfig, DebertaForSequenceClassification),
        (GPT2Config, GPT2ForSequenceClassification),
        (OpenAIGPTConfig, OpenAIGPTForSequenceClassification),
+        (TapasConfig, TapasForSequenceClassification),
    ]
 )

@@ -410,6 +417,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
        (RobertaConfig, RobertaForQuestionAnswering),
        (SqueezeBertConfig, SqueezeBertForQuestionAnswering),
        (BertConfig, BertForQuestionAnswering),
+        (TapasConfig, TapasForQuestionAnswering),
        (XLNetConfig, XLNetForQuestionAnsweringSimple),
        (FlaubertConfig, FlaubertForQuestionAnsweringSimple),
        (MobileBertConfig, MobileBertForQuestionAnswering),
@@ -1324,6 +1324,29 @@ class FillMaskPipeline(Pipeline):
        return results


+class TokenClassificationArgumentHandler(ArgumentHandler):
+    """
+    Handles arguments for token classification.
+    """
+
+    def __call__(self, *args, **kwargs):
+
+        if args is not None and len(args) > 0:
+            if isinstance(args, str):
+                inputs = [args]
+            else:
+                inputs = args
+            batch_size = len(inputs)
+
+        offset_mapping = kwargs.get("offset_mapping", None)
+        if offset_mapping:
+            if isinstance(offset_mapping, list) and isinstance(offset_mapping[0], tuple):
+                offset_mapping = [offset_mapping]
+            if len(offset_mapping) != batch_size:
+                raise ("offset_mapping should have the same batch size as the input")
+        return inputs, offset_mapping
+
+
@add_end_docstrings(
    PIPELINE_INIT_ARGS,
    r"""
@@ -1361,13 +1384,14 @@ class TokenClassificationPipeline(Pipeline):
        ignore_labels=["O"],
        task: str = "",
        grouped_entities: bool = False,
+        ignore_subwords: bool = True,
    ):
        super().__init__(
            model=model,
            tokenizer=tokenizer,
            modelcard=modelcard,
            framework=framework,
-            args_parser=args_parser,
+            args_parser=TokenClassificationArgumentHandler(),
            device=device,
            binary_output=binary_output,
            task=task,
@@ -1382,6 +1406,7 @@ class TokenClassificationPipeline(Pipeline):
        self._basic_tokenizer = BasicTokenizer(do_lower_case=False)
        self.ignore_labels = ignore_labels
        self.grouped_entities = grouped_entities
+        self.ignore_subwords = ignore_subwords

    def __call__(self, inputs: Union[str, List[str]], **kwargs):
        """
@@ -1402,10 +1427,15 @@ class TokenClassificationPipeline(Pipeline):
            - **index** (:obj:`int`, only present when ``self.grouped_entities=False``) -- The index of the
              corresponding token in the sentence.
        """
+
        if isinstance(inputs, str):
            inputs = [inputs]
+
+        offset_mappings = kwargs.get("offset_mappings")
+
        answers = []
-        for sentence in inputs:
+
+        for i, sentence in enumerate(inputs):

            # Manage correct placement of the tensors
            with self.device_placement():
@@ -1415,7 +1445,18 @@ class TokenClassificationPipeline(Pipeline):
                    return_attention_mask=False,
                    return_tensors=self.framework,
                    truncation=True,
+                    return_special_tokens_mask=True,
+                    return_offsets_mapping=self.tokenizer.is_fast,
                )
+                if self.tokenizer.is_fast:
+                    offset_mapping = tokens["offset_mapping"].cpu().numpy()[0]
+                    del tokens["offset_mapping"]
+                elif offset_mappings:
+                    offset_mapping = offset_mappings[i]
+                else:
+                    raise Exception("To decode [UNK] tokens use a fast tokenizer or provide offset_mapping parameter")
+                special_tokens_mask = tokens["special_tokens_mask"].cpu().numpy()[0]
+                del tokens["special_tokens_mask"]

                # Forward
                if self.framework == "tf":
@@ -1432,24 +1473,35 @@ class TokenClassificationPipeline(Pipeline):

            entities = []
            # Filter to labels not in `self.ignore_labels`
+            # Filter special_tokens
            filtered_labels_idx = [
                (idx, label_idx)
                for idx, label_idx in enumerate(labels_idx)
-                if self.model.config.id2label[label_idx] not in self.ignore_labels
+                if (self.model.config.id2label[label_idx] not in self.ignore_labels) and not special_tokens_mask[idx]
            ]

            for idx, label_idx in filtered_labels_idx:
+                start_ind, end_ind = offset_mapping[idx]
+                word_ref = sentence[start_ind:end_ind]
+                word = self.tokenizer.convert_ids_to_tokens([int(input_ids[idx])])[0]
+                is_subword = len(word_ref) != len(word)
+
+                if int(input_ids[idx]) == self.tokenizer.unk_token_id:
+                    word = word_ref
+                    is_subword = False

                entity = {
-                    "word": self.tokenizer.convert_ids_to_tokens(int(input_ids[idx])),
+                    "word": word,
                    "score": score[idx][label_idx].item(),
                    "entity": self.model.config.id2label[label_idx],
                    "index": idx,
                }

+                if self.grouped_entities and self.ignore_subwords:
+                    entity["is_subword"] = is_subword
+
                entities += [entity]

-            # Append grouped entities
            if self.grouped_entities:
                answers += [self.group_entities(entities)]
            # Append ungrouped entities
@@ -1468,8 +1520,8 @@ class TokenClassificationPipeline(Pipeline):
            entities (:obj:`dict`): The entities predicted by the pipeline.
        """
        # Get the first entity in the entity group
-        entity = entities[0]["entity"]
-        scores = np.mean([entity["score"] for entity in entities])
+        entity = entities[0]["entity"].split("-")[-1]
+        scores = np.nanmean([entity["score"] for entity in entities])
        tokens = [entity["word"] for entity in entities]

        entity_group = {
@@ -1494,7 +1546,9 @@ class TokenClassificationPipeline(Pipeline):
            last_idx = entities[-1]["index"]

        for entity in entities:
+
            is_last_idx = entity["index"] == last_idx
+            is_subword = self.ignore_subwords and entity["is_subword"]
            if not entity_group_disagg:
                entity_group_disagg += [entity]
                if is_last_idx:
@@ -1503,10 +1557,19 @@ class TokenClassificationPipeline(Pipeline):

            # If the current entity is similar and adjacent to the previous entity, append it to the disaggregated entity group
            # The split is meant to account for the "B" and "I" suffixes
+            # Shouldn't merge if both entities are B-type
            if (
-                entity["entity"].split("-")[-1] == entity_group_disagg[-1]["entity"].split("-")[-1]
+                (
+                    entity["entity"].split("-")[-1] == entity_group_disagg[-1]["entity"].split("-")[-1]
+                    and entity["entity"].split("-")[0] != "B"
+                )
                and entity["index"] == entity_group_disagg[-1]["index"] + 1
-            ):
+            ) or is_subword:
+                # Modify subword type to be previous_type
+                if is_subword:
+                    entity["entity"] = entity_group_disagg[-1]["entity"].split("-")[-1]
+                    entity["score"] = np.nan  # set ignored scores to nan and use np.nanmean
+
                entity_group_disagg += [entity]
                # Group the entities at the last entity
                if is_last_idx:
@@ -695,6 +695,31 @@ def mockenv(**kwargs):
    return unittest.mock.patch.dict(os.environ, kwargs)


+# --- pytest conf functions --- #
+
+# to avoid multiple invocation from tests/conftest.py and examples/conftest.py - make sure it's called only once
+pytest_opt_registered = {}
+
+
+def pytest_addoption_shared(parser):
+    """
+    This function is to be called from `conftest.py` via `pytest_addoption` wrapper that has to be defined there.
+
+    It allows loading both `conftest.py` files at once without causing a failure due to adding the same `pytest`
+    option.
+
+    """
+    option = "--make-reports"
+    if option not in pytest_opt_registered:
+        parser.addoption(
+            option,
+            action="store",
+            default=False,
+            help="generate report files. The value of this option is used as a prefix to report names",
+        )
+        pytest_opt_registered[option] = 1
+
+
 def pytest_terminal_summary_main(tr, id):
    """
    Generate multiple reports at the end of test suite run - each report goes into a dedicated file in the current
@@ -728,7 +753,7 @@ def pytest_terminal_summary_main(tr, id):
    dir = "reports"
    Path(dir).mkdir(parents=True, exist_ok=True)
    report_files = {
-        k: f"{dir}/report_{id}_{k}.txt"
+        k: f"{dir}/{id}_{k}.txt"
        for k in [
            "durations",
            "errors",
@@ -824,7 +849,7 @@ def pytest_terminal_summary_main(tr, id):
    config.option.tbstyle = orig_tbstyle


-# the following code deals with async io between processes
+# --- distributed testing functions --- #

 # adapted from https://stackoverflow.com/a/59041913/9201239
 import asyncio  # noqa
@@ -50,6 +50,7 @@ from .configuration_auto import (
    RobertaConfig,
    SqueezeBertConfig,
    T5Config,
+    TapasConfig,
    TransfoXLConfig,
    XLMConfig,
    XLMProphetNetConfig,
@@ -85,6 +86,7 @@ from .tokenization_rag import RagTokenizer
 from .tokenization_retribert import RetriBertTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_squeezebert import SqueezeBertTokenizer
+from .tokenization_tapas import TapasTokenizer
 from .tokenization_transfo_xl import TransfoXLTokenizer
 from .tokenization_xlm import XLMTokenizer
 from .utils import logging
@@ -210,6 +212,7 @@ TOKENIZER_MAPPING = OrderedDict(
        (RagConfig, (RagTokenizer, None)),
        (XLMProphetNetConfig, (XLMProphetNetTokenizer, None)),
        (ProphetNetConfig, (ProphetNetTokenizer, None)),
+        (TapasConfig, (TapasTokenizer, None)),
    ]
 )

@@ -48,12 +48,19 @@ class TrainingArguments:
            If :obj:`True`, overwrite the content of the output directory. Use this to continue training if
            :obj:`output_dir` points to a checkpoint directory.
        do_train (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to run training or not.
+            Whether to run training or not. This argument is not directly used by :class:`~transformers.Trainer`, it's
+            intended to be used by your training/evaluation scripts instead. See the `example scripts
+            <https://github.com/huggingface/transformers/tree/master/examples>`__ for more details.
        do_eval (:obj:`bool`, `optional`):
-            Whether to run evaluation on the dev set or not. Will default to :obj:`evaluation_strategy` different from
-            :obj:`"no"`.
+            Whether to run evaluation on the dev set or not. Will be set to :obj:`True` if :obj:`evaluation_strategy`
+            is different from :obj:`"no"`. This argument is not directly used by :class:`~transformers.Trainer`, it's
+            intended to be used by your training/evaluation scripts instead. See the `example scripts
+            <https://github.com/huggingface/transformers/tree/master/examples>`__ for more details.
        do_predict (:obj:`bool`, `optional`, defaults to :obj:`False`):
-            Whether to run predictions on the test set or not.
+            Whether to run predictions on the test set or not. This argument is not directly used by
+            :class:`~transformers.Trainer`, it's intended to be used by your training/evaluation scripts instead. See
+            the `example scripts <https://github.com/huggingface/transformers/tree/master/examples>`__ for more
+            details.
        evaluation_strategy (:obj:`str` or :class:`~transformers.trainer_utils.EvaluationStrategy`, `optional`, defaults to :obj:`"no"`):
            The evaluation strategy to adopt during training. Possible values are:

@@ -45,6 +45,15 @@ class DataCollatorForSOP:
        requires_pytorch(self)


+class DataCollatorForTokenClassification:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
 class DataCollatorForWholeWordMask:
    def __init__(self, *args, **kwargs):
        requires_pytorch(self)
@@ -24,17 +24,14 @@ def pytest_configure(config):


 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)
@@ -1,7 +1,10 @@
+import os
+import shutil
+import tempfile
 import unittest

-from transformers import AutoTokenizer, is_torch_available
-from transformers.testing_utils import require_torch, slow
+from transformers import BertTokenizer, is_torch_available, set_seed
+from transformers.testing_utils import require_torch


 if is_torch_available():
@@ -12,22 +15,25 @@ if is_torch_available():
        DataCollatorForNextSentencePrediction,
        DataCollatorForPermutationLanguageModeling,
        DataCollatorForSOP,
-        GlueDataset,
-        GlueDataTrainingArguments,
-        LineByLineTextDataset,
-        LineByLineWithSOPTextDataset,
-        TextDataset,
-        TextDatasetForNextSentencePrediction,
+        DataCollatorForTokenClassification,
+        DataCollatorWithPadding,
        default_data_collator,
    )


-PATH_SAMPLE_TEXT = "./tests/fixtures/sample_text.txt"
-PATH_SAMPLE_TEXT_DIR = "./tests/fixtures/tests_samples/wiki_text"
-
-
@require_torch
 class DataCollatorIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
+        self.vocab_file = os.path.join(self.tmpdirname, "vocab.txt")
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
    def test_default_with_dict(self):
        features = [{"label": i, "inputs": [0, 1, 2, 3, 4, 5]} for i in range(8)]
        batch = default_data_collator(features)
@@ -57,6 +63,17 @@ class DataCollatorIntegrationTest(unittest.TestCase):
        self.assertEqual(batch["labels"].dtype, torch.long)
        self.assertEqual(batch["inputs"].shape, torch.Size([8, 10]))

+    def test_default_classification_and_regression(self):
+        data_collator = default_data_collator
+
+        features = [{"input_ids": [0, 1, 2, 3, 4], "label": i} for i in range(4)]
+        batch = data_collator(features)
+        self.assertEqual(batch["labels"].dtype, torch.long)
+
+        features = [{"input_ids": [0, 1, 2, 3, 4], "label": float(i)} for i in range(4)]
+        batch = data_collator(features)
+        self.assertEqual(batch["labels"].dtype, torch.float)
+
    def test_default_with_no_labels(self):
        features = [{"label": None, "inputs": [0, 1, 2, 3, 4, 5]} for i in range(8)]
        batch = default_data_collator(features)
@@ -69,128 +86,144 @@ class DataCollatorIntegrationTest(unittest.TestCase):
        self.assertTrue("labels" not in batch)
        self.assertEqual(batch["inputs"].shape, torch.Size([8, 6]))

-    @slow
-    def test_default_classification(self):
-        MODEL_ID = "bert-base-cased-finetuned-mrpc"
-        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
-        data_args = GlueDataTrainingArguments(
-            task_name="mrpc", data_dir="./tests/fixtures/tests_samples/MRPC", overwrite_cache=True
-        )
-        dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev")
-        data_collator = default_data_collator
-        batch = data_collator(dataset.features)
-        self.assertEqual(batch["labels"].dtype, torch.long)
+    def test_data_collator_with_padding(self):
+        tokenizer = BertTokenizer(self.vocab_file)
+        features = [{"input_ids": [0, 1, 2]}, {"input_ids": [0, 1, 2, 3, 4, 5]}]

-    @slow
-    def test_default_regression(self):
-        MODEL_ID = "distilroberta-base"
-        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
-        data_args = GlueDataTrainingArguments(
-            task_name="sts-b", data_dir="./tests/fixtures/tests_samples/STS-B", overwrite_cache=True
-        )
-        dataset = GlueDataset(data_args, tokenizer=tokenizer, mode="dev")
-        data_collator = default_data_collator
-        batch = data_collator(dataset.features)
-        self.assertEqual(batch["labels"].dtype, torch.float)
+        data_collator = DataCollatorWithPadding(tokenizer)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 6]))
+        self.assertEqual(batch["input_ids"][0].tolist(), [0, 1, 2] + [tokenizer.pad_token_id] * 3)
+
+        data_collator = DataCollatorWithPadding(tokenizer, padding="max_length", max_length=10)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 10]))
+
+        data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 8]))
+
+    def test_data_collator_for_token_classification(self):
+        tokenizer = BertTokenizer(self.vocab_file)
+        features = [
+            {"input_ids": [0, 1, 2], "labels": [0, 1, 2]},
+            {"input_ids": [0, 1, 2, 3, 4, 5], "labels": [0, 1, 2, 3, 4, 5]},
+        ]
+
+        data_collator = DataCollatorForTokenClassification(tokenizer)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 6]))
+        self.assertEqual(batch["input_ids"][0].tolist(), [0, 1, 2] + [tokenizer.pad_token_id] * 3)
+        self.assertEqual(batch["labels"].shape, torch.Size([2, 6]))
+        self.assertEqual(batch["labels"][0].tolist(), [0, 1, 2] + [-100] * 3)
+
+        data_collator = DataCollatorForTokenClassification(tokenizer, padding="max_length", max_length=10)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 10]))
+        self.assertEqual(batch["labels"].shape, torch.Size([2, 10]))
+
+        data_collator = DataCollatorForTokenClassification(tokenizer, pad_to_multiple_of=8)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 8]))
+        self.assertEqual(batch["labels"].shape, torch.Size([2, 8]))
+
+        data_collator = DataCollatorForTokenClassification(tokenizer, label_pad_token_id=-1)
+        batch = data_collator(features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size([2, 6]))
+        self.assertEqual(batch["input_ids"][0].tolist(), [0, 1, 2] + [tokenizer.pad_token_id] * 3)
+        self.assertEqual(batch["labels"].shape, torch.Size([2, 6]))
+        self.assertEqual(batch["labels"][0].tolist(), [0, 1, 2] + [-1] * 3)
+
+    def test_data_collator_for_language_modeling(self):
+        tokenizer = BertTokenizer(self.vocab_file)
+        no_pad_features = [{"input_ids": list(range(10))}, {"input_ids": list(range(10))}]
+        pad_features = [{"input_ids": list(range(5))}, {"input_ids": list(range(10))}]

-    @slow
-    def test_lm_tokenizer_without_padding(self):
-        tokenizer = AutoTokenizer.from_pretrained("gpt2")
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
-        # ^ causal lm
+        batch = data_collator(no_pad_features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

-        dataset = LineByLineTextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512)
-        examples = [dataset[i] for i in range(len(dataset))]
+        batch = data_collator(pad_features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))
+
+        tokenizer._pad_token = None
+        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        with self.assertRaises(ValueError):
-            # Expect error due to padding token missing on gpt2:
-            data_collator(examples)
+            # Expect error due to padding token missing
+            data_collator(pad_features)

-        dataset = TextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512, overwrite_cache=True)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
-        self.assertIsInstance(batch, dict)
-        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))
-
-    @slow
-    def test_lm_tokenizer_with_padding(self):
-        tokenizer = AutoTokenizer.from_pretrained("distilroberta-base")
+        set_seed(42)  # For reproducibility
+        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer)
-        # ^ masked lm
+        batch = data_collator(no_pad_features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

-        dataset = LineByLineTextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
-        self.assertIsInstance(batch, dict)
-        self.assertEqual(batch["input_ids"].shape, torch.Size((31, 107)))
-        self.assertEqual(batch["labels"].shape, torch.Size((31, 107)))
+        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
+        self.assertTrue(torch.any(masked_tokens))
+        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

-        dataset = TextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512, overwrite_cache=True)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
-        self.assertIsInstance(batch, dict)
-        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))
+        batch = data_collator(pad_features)
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))
+
+        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
+        self.assertTrue(torch.any(masked_tokens))
+        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

-    @slow
    def test_plm(self):
-        tokenizer = AutoTokenizer.from_pretrained("xlnet-base-cased")
+        tokenizer = BertTokenizer(self.vocab_file)
+        no_pad_features = [{"input_ids": list(range(10))}, {"input_ids": list(range(10))}]
+        pad_features = [{"input_ids": list(range(5))}, {"input_ids": list(range(10))}]
+
        data_collator = DataCollatorForPermutationLanguageModeling(tokenizer)
-        # ^ permutation lm

-        dataset = LineByLineTextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
+        batch = data_collator(pad_features)
        self.assertIsInstance(batch, dict)
-        self.assertEqual(batch["input_ids"].shape, torch.Size((31, 112)))
-        self.assertEqual(batch["perm_mask"].shape, torch.Size((31, 112, 112)))
-        self.assertEqual(batch["target_mapping"].shape, torch.Size((31, 112, 112)))
-        self.assertEqual(batch["labels"].shape, torch.Size((31, 112)))
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["perm_mask"].shape, torch.Size((2, 10, 10)))
+        self.assertEqual(batch["target_mapping"].shape, torch.Size((2, 10, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

-        dataset = TextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512, overwrite_cache=True)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
+        batch = data_collator(no_pad_features)
        self.assertIsInstance(batch, dict)
-        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
-        self.assertEqual(batch["perm_mask"].shape, torch.Size((2, 512, 512)))
-        self.assertEqual(batch["target_mapping"].shape, torch.Size((2, 512, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
+        self.assertEqual(batch["perm_mask"].shape, torch.Size((2, 10, 10)))
+        self.assertEqual(batch["target_mapping"].shape, torch.Size((2, 10, 10)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        example = [torch.randint(5, [5])]
        with self.assertRaises(ValueError):
            # Expect error due to odd sequence length
            data_collator(example)

-    @slow
    def test_nsp(self):
-        tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
+        tokenizer = BertTokenizer(self.vocab_file)
+        features = [{"tokens_a": [0, 1, 2, 3, 4], "tokens_b": [0, 1, 2, 3, 4], "is_random_next": i} for i in range(2)]
        data_collator = DataCollatorForNextSentencePrediction(tokenizer)
+        batch = data_collator(features)

-        dataset = TextDatasetForNextSentencePrediction(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
-        self.assertIsInstance(batch, dict)
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
+        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 512)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))
+        self.assertEqual(batch["next_sentence_label"].shape, torch.Size((2,)))

-        # Since there are randomly generated false samples, the total number of samples is not fixed.
-        total_samples = batch["input_ids"].shape[0]
-        self.assertEqual(batch["input_ids"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["token_type_ids"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["next_sentence_label"].shape, torch.Size((total_samples,)))
-
-    @slow
    def test_sop(self):
-        tokenizer = AutoTokenizer.from_pretrained("albert-base-v2")
+        tokenizer = BertTokenizer(self.vocab_file)
+        features = [
+            {
+                "input_ids": torch.tensor([0, 1, 2, 3, 4]),
+                "token_type_ids": torch.tensor([0, 1, 2, 3, 4]),
+                "sentence_order_label": torch.tensor(i),
+            }
+            for i in range(2)
+        ]
        data_collator = DataCollatorForSOP(tokenizer)
+        batch = data_collator(features)

-        dataset = LineByLineWithSOPTextDataset(tokenizer, file_dir=PATH_SAMPLE_TEXT_DIR, block_size=512)
-        examples = [dataset[i] for i in range(len(dataset))]
-        batch = data_collator(examples)
-        self.assertIsInstance(batch, dict)
-
-        # Since there are randomly generated false samples, the total number of samples is not fixed.
-        total_samples = batch["input_ids"].shape[0]
-        self.assertEqual(batch["input_ids"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["token_type_ids"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["labels"].shape, torch.Size((total_samples, 512)))
-        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((total_samples,)))
+        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
+        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
+        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
+        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))
@@ -0,0 +1,612 @@
+# coding=utf-8
+# Copyright 2020 Google Research and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+import numpy as np
+
+from transformers import is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TapasConfig,
+        TapasForMaskedLM,
+        TapasForQuestionAnswering,
+        TapasForSequenceClassification,
+        TapasModel,
+    )
+
+
+class TapasModelTester:
+    """You can also import this e.g from .test_modeling_tapas import TapasModelTester """
+
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        initializer_range=0.02,
+        max_position_embeddings=512,
+        type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10],
+        type_sequence_label_size=2,
+        positive_weight=10.0,
+        num_aggregation_labels=4,
+        num_labels=2,
+        aggregation_loss_importance=0.8,
+        use_answer_as_supervision=True,
+        answer_loss_importance=0.001,
+        use_normalized_answer_loss=False,
+        huber_loss_delta=25.0,
+        temperature=1.0,
+        agg_temperature=1.0,
+        use_gumbel_for_cells=False,
+        use_gumbel_for_agg=False,
+        average_approximation_function="ratio",
+        cell_select_pref=0.5,
+        answer_loss_cutoff=100,
+        max_num_rows=64,
+        max_num_columns=32,
+        average_logits_per_cell=True,
+        select_one_column=True,
+        allow_empty_column_selection=False,
+        init_cell_selection_weights_to_zero=False,
+        reset_position_index_per_cell=False,
+        disable_per_token_loss=False,
+        span_prediction="none",
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_sizes = type_vocab_sizes
+        self.type_sequence_label_size = type_sequence_label_size
+        self.positive_weight = positive_weight
+        self.num_aggregation_labels = num_aggregation_labels
+        self.num_labels = num_labels
+        self.aggregation_loss_importance = aggregation_loss_importance
+        self.use_answer_as_supervision = use_answer_as_supervision
+        self.answer_loss_importance = answer_loss_importance
+        self.use_normalized_answer_loss = use_normalized_answer_loss
+        self.huber_loss_delta = huber_loss_delta
+        self.temperature = temperature
+        self.agg_temperature = agg_temperature
+        self.use_gumbel_for_cells = use_gumbel_for_cells
+        self.use_gumbel_for_agg = use_gumbel_for_agg
+        self.average_approximation_function = average_approximation_function
+        self.cell_select_pref = cell_select_pref
+        self.answer_loss_cutoff = answer_loss_cutoff
+        self.max_num_rows = max_num_rows
+        self.max_num_columns = max_num_columns
+        self.average_logits_per_cell = average_logits_per_cell
+        self.select_one_column = select_one_column
+        self.allow_empty_column_selection = allow_empty_column_selection
+        self.init_cell_selection_weights_to_zero = init_cell_selection_weights_to_zero
+        self.reset_position_index_per_cell = reset_position_index_per_cell
+        self.disable_per_token_loss = disable_per_token_loss
+        self.span_prediction = span_prediction
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = []
+        for type_vocab_size in self.type_vocab_sizes:
+            token_type_ids.append(ids_tensor(shape=[self.batch_size, self.seq_length], vocab_size=type_vocab_size))
+        token_type_ids = torch.stack(token_type_ids, dim=2)
+
+        sequence_labels = None
+        token_labels = None
+        label_ids = None
+        answer = None
+        numeric_values = None
+        numeric_values_scale = None
+        aggregation_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            label_ids = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+            answer = floats_tensor([self.batch_size])
+            numeric_values = floats_tensor([self.batch_size, self.seq_length])
+            numeric_values_scale = floats_tensor([self.batch_size, self.seq_length])
+            aggregation_labels = ids_tensor([self.batch_size], self.num_aggregation_labels)
+
+        config = TapasConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_sizes=self.type_vocab_sizes,
+            initializer_range=self.initializer_range,
+            positive_weight=self.positive_weight,
+            num_aggregation_labels=self.num_aggregation_labels,
+            num_labels=self.num_labels,
+            aggregation_loss_importance=self.aggregation_loss_importance,
+            use_answer_as_supervision=self.use_answer_as_supervision,
+            answer_loss_importance=self.answer_loss_importance,
+            use_normalized_answer_loss=self.use_normalized_answer_loss,
+            huber_loss_delta=self.huber_loss_delta,
+            temperature=self.temperature,
+            agg_temperature=self.agg_temperature,
+            use_gumbel_for_cells=self.use_gumbel_for_cells,
+            use_gumbel_for_agg=self.use_gumbel_for_agg,
+            average_approximation_function=self.average_approximation_function,
+            cell_select_pref=self.cell_select_pref,
+            answer_loss_cutoff=self.answer_loss_cutoff,
+            max_num_rows=self.max_num_rows,
+            max_num_columns=self.max_num_columns,
+            average_logits_per_cell=self.average_logits_per_cell,
+            select_one_column=self.select_one_column,
+            allow_empty_column_selection=self.allow_empty_column_selection,
+            init_cell_selection_weights_to_zero=self.init_cell_selection_weights_to_zero,
+            reset_position_index_per_cell=self.reset_position_index_per_cell,
+            disable_per_token_loss=self.disable_per_token_loss,
+            span_prediction=self.span_prediction,
+            return_dict=True,
+        )
+
+        return (
+            config,
+            input_ids,
+            input_mask,
+            token_type_ids,
+            sequence_labels,
+            token_labels,
+            label_ids,
+            answer,
+            numeric_values,
+            numeric_values_scale,
+            aggregation_labels,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_for_masked_lm(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_question_answering(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            label_ids=label_ids,
+            answer=answer,
+            numeric_values=numeric_values,
+            numeric_values_scale=numeric_values_scale,
+            aggregation_labels=aggregation_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.logits_aggregation.shape, (self.batch_size, self.num_aggregation_labels))
+
+    def create_and_check_for_sequence_classification(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        config.num_labels = self.num_labels
+        model = TapasForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_mask,
+            token_type_ids,
+            sequence_labels,
+            token_labels,
+            label_ids,
+            answer,
+            numeric_values,
+            numeric_values_scale,
+            aggregation_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class TapasModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            TapasModel,
+            TapasForMaskedLM,
+            TapasForQuestionAnswering,
+            TapasForSequenceClassification,
+        )
+        if is_torch_available()
+        else None
+    )
+    test_pruning = False
+    test_torchscript = True
+    test_resize_embeddings = True
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = TapasModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=TapasConfig, dim=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    # @slow
+    # def test_lm_outputs_same_as_reference_model(self):
+    #     """Write something that could help someone fixing this here."""
+    #     checkpoint_path = "XXX/bart-large"
+    #     model = self.big_model
+    #     tokenizer = AutoTokenizer.from_pretrained(
+    #         checkpoint_path
+    #     )  # same with AutoTokenizer (see tokenization_auto.py). This is not mandatory
+    #     # MODIFY THIS DEPENDING ON YOUR MODELS RELEVANT TASK.
+    #     batch = tokenizer(["I went to the <mask> yesterday"]).to(torch_device)
+    #     desired_mask_result = tokenizer.decode("store")  # update this
+    #     logits = model(**batch).logits
+    #     masked_index = (batch.input_ids == self.tokenizer.mask_token_id).nonzero()
+    #     assert model.num_parameters() == 175e9  # a joke
+    #     mask_entry_logits = logits[0, masked_index.item(), :]
+    #     probs = mask_entry_logits.softmax(dim=0)
+    #     _, predictions = probs.topk(1)
+    #     self.assertEqual(tokenizer.decode(predictions), desired_mask_result)
+
+    # @cached_property
+    # def big_model(self):
+    #     """Cached property means this code will only be executed once."""
+    #     checkpoint_path = "XXX/bart-large"
+    #     model = AutoModelForMaskedLM.from_pretrained(checkpoint_path).to(
+    #         torch_device
+    #     )  # test whether AutoModel can determine your model_class from checkpoint name
+    #     if torch_device == "cuda":
+    #         model.half()
+
+    # optional: do more testing! This will save you time later!
+    # @slow
+    # def test_that_XXX_can_be_used_in_a_pipeline(self):
+    #     """We can use self.big_model here without calling __init__ again."""
+    #     pass
+
+    # def test_XXX_loss_doesnt_change_if_you_add_padding(self):
+    #     pass
+
+    # def test_XXX_bad_args(self):
+    #     pass
+
+    # def test_XXX_backward_pass_reduces_loss(self):
+    #     """Test loss/gradients same as reference implementation, for example."""
+    #     pass
+
+    # @require_torch_and_cuda
+    # def test_large_inputs_in_fp16_dont_cause_overflow(self):
+    #     pass
+
+
+# Below: tests for Tapas utilities, based on segmented_tensor_test.py of the original implementation.
+# These test the operations on segmented tensors.
+class TapasUtilitiesTest(unittest.TestCase):
+    def _prepare_tables(self):
+        """Prepares two tables, both with three distinct rows.
+        The first table has two columns:
+        1.0, 2.0 | 3.0
+        2.0, 0.0 | 1.0
+        1.0, 3.0 | 4.0
+        The second table has three columns:
+        1.0 | 2.0 | 3.0
+        2.0 | 0.0 | 1.0
+        1.0 | 3.0 | 4.0
+        Returns:
+        SegmentedTensors with the tables.
+        """
+        values = torch.tensor(
+            [
+                [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]],
+                [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]],
+            ]
+        )
+        row_index = utils.IndexMap(
+            indices=torch.tensor(
+                [
+                    [[0, 0, 0], [1, 1, 1], [2, 2, 2]],
+                    [[0, 0, 0], [1, 1, 1], [2, 2, 2]],
+                ]
+            ),
+            num_segments=3,
+            batch_dims=1,
+        )
+        col_index = utils.IndexMap(
+            indices=torch.tensor(
+                [
+                    [[0, 0, 1], [0, 0, 1], [0, 0, 1]],
+                    [[0, 1, 2], [0, 1, 2], [0, 1, 2]],
+                ]
+            ),
+            num_segments=3,
+            batch_dims=1,
+        )
+        return values, row_index, col_index
+
+    def test_product_index(self):
+        _, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_index_proj = cell_index.project_outer(cell_index)
+        col_index_proj = cell_index.project_inner(cell_index)
+
+        ind = cell_index.indices
+        self.assertEqual(cell_index.num_segments, 9)
+
+        # Projections should give back the original indices.
+        # we use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(row_index.indices.numpy(), row_index_proj.indices.numpy())
+        self.assertEqual(row_index.num_segments, row_index_proj.num_segments)
+        self.assertEqual(row_index.batch_dims, row_index_proj.batch_dims)
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(col_index.indices.numpy(), col_index_proj.indices.numpy())
+        self.assertEqual(col_index.batch_dims, col_index_proj.batch_dims)
+
+        # The first and second "column" are identified in the first table.
+        for i in range(3):
+            self.assertEqual(ind[0, i, 0], ind[0, i, 1])
+            self.assertNotEqual(ind[0, i, 0], ind[0, i, 2])
+
+        # All rows are distinct in the first table.
+        for i, i_2 in zip(range(3), range(3)):
+            for j, j_2 in zip(range(3), range(3)):
+                if i != i_2 and j != j_2:
+                    self.assertNotEqual(ind[0, i, j], ind[0, i_2, j_2])
+
+        # All cells are distinct in the second table.
+        for i, i_2 in zip(range(3), range(3)):
+            for j, j_2 in zip(range(3), range(3)):
+                if i != i_2 or j != j_2:
+                    self.assertNotEqual(ind[1, i, j], ind[1, i_2, j_2])
+
+    def test_flatten(self):
+        _, row_index, col_index = self._prepare_tables()
+        row_index_flat = utils.flatten(row_index)
+        col_index_flat = utils.flatten(col_index)
+
+        shape = [3, 4, 5]
+        batched_index = utils.IndexMap(indices=torch.zeros(shape).type(torch.LongTensor), num_segments=1, batch_dims=3)
+        batched_index_flat = utils.flatten(batched_index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(
+            row_index_flat.indices.numpy(), [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5]
+        )
+        np.testing.assert_array_equal(
+            col_index_flat.indices.numpy(), [0, 0, 1, 0, 0, 1, 0, 0, 1, 3, 4, 5, 3, 4, 5, 3, 4, 5]
+        )
+        self.assertEqual(batched_index_flat.num_segments.numpy(), np.prod(shape))
+        np.testing.assert_array_equal(batched_index_flat.indices.numpy(), range(np.prod(shape)))
+
+    def test_range_index_map(self):
+        batch_shape = [3, 4]
+        num_segments = 5
+        index = utils.range_index_map(batch_shape, num_segments)
+
+        self.assertEqual(num_segments, index.num_segments)
+        self.assertEqual(2, index.batch_dims)
+        indices = index.indices
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(list(indices.size()), [3, 4, 5])
+        for i in range(batch_shape[0]):
+            for j in range(batch_shape[1]):
+                # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+                np.testing.assert_array_equal(indices[i, j, :].numpy(), range(num_segments))
+
+    def test_reduce_sum(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_sum, _ = utils.reduce_sum(values, row_index)
+        col_sum, _ = utils.reduce_sum(values, col_index)
+        cell_sum, _ = utils.reduce_sum(values, cell_index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(row_sum.numpy(), [[6.0, 3.0, 8.0], [6.0, 3.0, 8.0]])
+        np.testing.assert_allclose(col_sum.numpy(), [[9.0, 8.0, 0.0], [4.0, 5.0, 8.0]])
+        np.testing.assert_allclose(
+            cell_sum.numpy(),
+            [[3.0, 3.0, 0.0, 2.0, 1.0, 0.0, 4.0, 4.0, 0.0], [1.0, 2.0, 3.0, 2.0, 0.0, 1.0, 1.0, 3.0, 4.0]],
+        )
+
+    def test_reduce_mean(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_mean, _ = utils.reduce_mean(values, row_index)
+        col_mean, _ = utils.reduce_mean(values, col_index)
+        cell_mean, _ = utils.reduce_mean(values, cell_index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(
+            row_mean.numpy(), [[6.0 / 3.0, 3.0 / 3.0, 8.0 / 3.0], [6.0 / 3.0, 3.0 / 3.0, 8.0 / 3.0]]
+        )
+        np.testing.assert_allclose(col_mean.numpy(), [[9.0 / 6.0, 8.0 / 3.0, 0.0], [4.0 / 3.0, 5.0 / 3.0, 8.0 / 3.0]])
+        np.testing.assert_allclose(
+            cell_mean.numpy(),
+            [
+                [3.0 / 2.0, 3.0, 0.0, 2.0 / 2.0, 1.0, 0.0, 4.0 / 2.0, 4.0, 0.0],
+                [1.0, 2.0, 3.0, 2.0, 0.0, 1.0, 1.0, 3.0, 4.0],
+            ],
+        )
+
+    def test_reduce_max(self):
+        values = torch.as_tensor([2.0, 1.0, 0.0, 3.0])
+        index = utils.IndexMap(indices=torch.as_tensor([0, 1, 0, 1]), num_segments=2)
+        maximum, _ = utils.reduce_max(values, index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(maximum.numpy(), [2, 3])
+
+    def test_reduce_sum_vectorized(self):
+        values = torch.as_tensor([[1.0, 2.0, 3.0], [2.0, 3.0, 4.0], [3.0, 4.0, 5.0]])
+        index = utils.IndexMap(indices=torch.as_tensor([0, 0, 1]), num_segments=2, batch_dims=0)
+        sums, new_index = utils.reduce_sum(values, index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(sums.numpy(), [[3.0, 5.0, 7.0], [3.0, 4.0, 5.0]])
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(new_index.indices.numpy(), [0, 1])
+        np.testing.assert_array_equal(new_index.num_segments.numpy(), 2)
+        np.testing.assert_array_equal(new_index.batch_dims, 0)
+
+    def test_gather(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+
+        # Compute sums and then gather. The result should have the same shape as
+        # the original table and each element should contain the sum the values in
+        # its cell.
+        sums, _ = utils.reduce_sum(values, cell_index)
+        cell_sum = utils.gather(sums, cell_index)
+        assert cell_sum.size() == values.size()
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_allclose(
+            cell_sum.numpy(),
+            [[[3.0, 3.0, 3.0], [2.0, 2.0, 1.0], [4.0, 4.0, 4.0]], [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]]],
+        )
+
+    def test_gather_vectorized(self):
+        values = torch.as_tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
+        index = utils.IndexMap(indices=torch.as_tensor([[0, 1], [1, 0]]), num_segments=2, batch_dims=1)
+        result = utils.gather(values, index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(result.numpy(), [[[1, 2], [3, 4]], [[7, 8], [5, 6]]])
@@ -1,8 +1,8 @@
 import unittest

-from transformers import pipeline
+from transformers import AutoTokenizer, pipeline
 from transformers.pipelines import Pipeline
-from transformers.testing_utils import require_tf
+from transformers.testing_utils import require_tf, require_torch

 from .test_pipelines_common import CustomInputPipelineCommonMixin

@@ -19,38 +19,54 @@ class NerPipelineTests(CustomInputPipelineCommonMixin, unittest.TestCase):

    def _test_pipeline(self, nlp: Pipeline):
        output_keys = {"entity", "word", "score"}
+        if nlp.grouped_entities:
+            output_keys = {"entity_group", "word", "score"}

        ungrouped_ner_inputs = [
            [
-                {"entity": "B-PER", "index": 1, "score": 0.9994944930076599, "word": "Cons"},
-                {"entity": "B-PER", "index": 2, "score": 0.8025449514389038, "word": "##uelo"},
-                {"entity": "I-PER", "index": 3, "score": 0.9993102550506592, "word": "Ara"},
-                {"entity": "I-PER", "index": 4, "score": 0.9993743896484375, "word": "##új"},
-                {"entity": "I-PER", "index": 5, "score": 0.9992871880531311, "word": "##o"},
-                {"entity": "I-PER", "index": 6, "score": 0.9993029236793518, "word": "No"},
-                {"entity": "I-PER", "index": 7, "score": 0.9981776475906372, "word": "##guera"},
-                {"entity": "B-PER", "index": 15, "score": 0.9998136162757874, "word": "Andrés"},
-                {"entity": "I-PER", "index": 16, "score": 0.999740719795227, "word": "Pas"},
-                {"entity": "I-PER", "index": 17, "score": 0.9997414350509644, "word": "##tran"},
-                {"entity": "I-PER", "index": 18, "score": 0.9996136426925659, "word": "##a"},
-                {"entity": "B-ORG", "index": 28, "score": 0.9989739060401917, "word": "Far"},
-                {"entity": "I-ORG", "index": 29, "score": 0.7188422083854675, "word": "##c"},
+                {"entity": "B-PER", "index": 1, "score": 0.9994944930076599, "is_subword": False, "word": "Cons"},
+                {"entity": "B-PER", "index": 2, "score": 0.8025449514389038, "is_subword": True, "word": "##uelo"},
+                {"entity": "I-PER", "index": 3, "score": 0.9993102550506592, "is_subword": False, "word": "Ara"},
+                {"entity": "I-PER", "index": 4, "score": 0.9993743896484375, "is_subword": True, "word": "##új"},
+                {"entity": "I-PER", "index": 5, "score": 0.9992871880531311, "is_subword": True, "word": "##o"},
+                {"entity": "I-PER", "index": 6, "score": 0.9993029236793518, "is_subword": False, "word": "No"},
+                {"entity": "I-PER", "index": 7, "score": 0.9981776475906372, "is_subword": True, "word": "##guera"},
+                {"entity": "B-PER", "index": 15, "score": 0.9998136162757874, "is_subword": False, "word": "Andrés"},
+                {"entity": "I-PER", "index": 16, "score": 0.999740719795227, "is_subword": False, "word": "Pas"},
+                {"entity": "I-PER", "index": 17, "score": 0.9997414350509644, "is_subword": True, "word": "##tran"},
+                {"entity": "I-PER", "index": 18, "score": 0.9996136426925659, "is_subword": True, "word": "##a"},
+                {"entity": "B-ORG", "index": 28, "score": 0.9989739060401917, "is_subword": False, "word": "Far"},
+                {"entity": "I-ORG", "index": 29, "score": 0.7188422083854675, "is_subword": True, "word": "##c"},
            ],
            [
-                {"entity": "I-PER", "index": 1, "score": 0.9968166351318359, "word": "En"},
-                {"entity": "I-PER", "index": 2, "score": 0.9957635998725891, "word": "##zo"},
-                {"entity": "I-ORG", "index": 7, "score": 0.9986497163772583, "word": "UN"},
+                {"entity": "I-PER", "index": 1, "score": 0.9968166351318359, "is_subword": False, "word": "En"},
+                {"entity": "I-PER", "index": 2, "score": 0.9957635998725891, "is_subword": True, "word": "##zo"},
+                {"entity": "I-ORG", "index": 7, "score": 0.9986497163772583, "is_subword": False, "word": "UN"},
            ],
        ]
+
        expected_grouped_ner_results = [
            [
-                {"entity_group": "B-PER", "score": 0.9710702640669686, "word": "Consuelo Araújo Noguera"},
-                {"entity_group": "B-PER", "score": 0.9997273534536362, "word": "Andrés Pastrana"},
-                {"entity_group": "B-ORG", "score": 0.8589080572128296, "word": "Farc"},
+                {"entity_group": "PER", "score": 0.999369223912557, "word": "Consuelo Araújo Noguera"},
+                {"entity_group": "PER", "score": 0.9997771680355072, "word": "Andrés Pastrana"},
+                {"entity_group": "ORG", "score": 0.9989739060401917, "word": "Farc"},
            ],
            [
-                {"entity_group": "I-PER", "score": 0.9962901175022125, "word": "Enzo"},
-                {"entity_group": "I-ORG", "score": 0.9986497163772583, "word": "UN"},
+                {"entity_group": "PER", "score": 0.9968166351318359, "word": "Enzo"},
+                {"entity_group": "ORG", "score": 0.9986497163772583, "word": "UN"},
+            ],
+        ]
+
+        expected_grouped_ner_results_w_subword = [
+            [
+                {"entity_group": "PER", "score": 0.9994944930076599, "word": "Cons"},
+                {"entity_group": "PER", "score": 0.9663328925768534, "word": "##uelo Araújo Noguera"},
+                {"entity_group": "PER", "score": 0.9997273534536362, "word": "Andrés Pastrana"},
+                {"entity_group": "ORG", "score": 0.8589080572128296, "word": "Farc"},
+            ],
+            [
+                {"entity_group": "PER", "score": 0.9962901175022125, "word": "Enzo"},
+                {"entity_group": "ORG", "score": 0.9986497163772583, "word": "UN"},
            ],
        ]

@@ -77,12 +93,80 @@ class NerPipelineTests(CustomInputPipelineCommonMixin, unittest.TestCase):
            for key in output_keys:
                self.assertIn(key, result)

-        for ungrouped_input, grouped_result in zip(ungrouped_ner_inputs, expected_grouped_ner_results):
-            self.assertEqual(nlp.group_entities(ungrouped_input), grouped_result)
+        if nlp.grouped_entities:
+            if nlp.ignore_subwords:
+                for ungrouped_input, grouped_result in zip(ungrouped_ner_inputs, expected_grouped_ner_results):
+                    self.assertEqual(nlp.group_entities(ungrouped_input), grouped_result)
+            else:
+                for ungrouped_input, grouped_result in zip(
+                    ungrouped_ner_inputs, expected_grouped_ner_results_w_subword
+                ):
+                    self.assertEqual(nlp.group_entities(ungrouped_input), grouped_result)

    @require_tf
    def test_tf_only(self):
        model_name = "Narsil/small"  # This model only has a TensorFlow version
        # We test that if we don't specificy framework='tf', it gets detected automatically
-        nlp = pipeline(task="ner", model=model_name, tokenizer=model_name)
+        tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+        nlp = pipeline(task="ner", model=model_name, tokenizer=tokenizer)
        self._test_pipeline(nlp)
+
+    #         offset=tokenizer(VALID_INPUTS[0],return_offsets_mapping=True)['offset_mapping']
+    #         pipeline_running_kwargs = {"offset_mapping"}  # Additional kwargs to run the pipeline with
+
+    @require_tf
+    def test_tf_defaults(self):
+        for model_name in self.small_models:
+            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+            nlp = pipeline(task="ner", model=model_name, tokenizer=tokenizer, framework="tf")
+        self._test_pipeline(nlp)
+
+    @require_tf
+    def test_tf_small(self):
+        for model_name in self.small_models:
+            print(model_name)
+            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+            nlp = pipeline(
+                task="ner",
+                model=model_name,
+                tokenizer=tokenizer,
+                framework="tf",
+                grouped_entities=True,
+                ignore_subwords=True,
+            )
+            self._test_pipeline(nlp)
+
+            for model_name in self.small_models:
+                tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+                nlp = pipeline(
+                    task="ner",
+                    model=model_name,
+                    tokenizer=tokenizer,
+                    framework="tf",
+                    grouped_entities=True,
+                    ignore_subwords=False,
+                )
+                self._test_pipeline(nlp)
+
+    @require_torch
+    def test_pt_defaults(self):
+        for model_name in self.small_models:
+            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+            nlp = pipeline(task="ner", model=model_name, tokenizer=tokenizer)
+            self._test_pipeline(nlp)
+
+    @require_torch
+    def test_torch_small(self):
+        for model_name in self.small_models:
+            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+            nlp = pipeline(
+                task="ner", model=model_name, tokenizer=tokenizer, grouped_entities=True, ignore_subwords=True
+            )
+            self._test_pipeline(nlp)
+
+        for model_name in self.small_models:
+            tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True)
+            nlp = pipeline(
+                task="ner", model=model_name, tokenizer=tokenizer, grouped_entities=True, ignore_subwords=False
+            )
+            self._test_pipeline(nlp)
@@ -18,13 +18,13 @@ import os
 import tempfile
 import unittest

-import datasets
 import numpy as np

 from transformers import AutoTokenizer, EvaluationStrategy, PretrainedConfig, TrainingArguments, is_torch_available
 from transformers.file_utils import WEIGHTS_NAME
 from transformers.testing_utils import (
    get_tests_dir,
+    require_datasets,
    require_optuna,
    require_sentencepiece,
    require_tokenizers,
@@ -340,7 +340,10 @@ class TrainerIntegrationTest(unittest.TestCase):
        self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
        self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))

+    @require_datasets
    def test_trainer_with_datasets(self):
+        import datasets
+
        np.random.seed(42)
        x = np.random.normal(size=(64,)).astype(np.float32)
        y = 2.0 * x + 3.0 + np.random.normal(scale=0.1, size=(64,))
@@ -658,15 +661,17 @@ class TrainerHyperParameterIntegrationTest(unittest.TestCase):
        def hp_name(trial):
            return MyTrialShortNamer.shortname(trial.params)

-        trainer = get_regression_trainer(
-            learning_rate=0.1,
-            logging_steps=1,
-            evaluation_strategy=EvaluationStrategy.EPOCH,
-            num_train_epochs=4,
-            disable_tqdm=True,
-            load_best_model_at_end=True,
-            logging_dir="runs",
-            run_name="test",
-            model_init=model_init,
-        )
-        trainer.hyperparameter_search(direction="minimize", hp_space=hp_space, hp_name=hp_name, n_trials=4)
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            trainer = get_regression_trainer(
+                output_dir=tmp_dir,
+                learning_rate=0.1,
+                logging_steps=1,
+                evaluation_strategy=EvaluationStrategy.EPOCH,
+                num_train_epochs=4,
+                disable_tqdm=True,
+                load_best_model_at_end=True,
+                logging_dir="runs",
+                run_name="test",
+                model_init=model_init,
+            )
+            trainer.hyperparameter_search(direction="minimize", hp_space=hp_space, hp_name=hp_name, n_trials=4)
@@ -46,8 +46,6 @@ if is_torch_available():
 class TestTrainerDistributed(TestCasePlus):
    @require_torch_multigpu
    def test_trainer(self):
-        import os
-        print(os.system("nvidia-smi"))

        distributed_args = f"""
            -m torch.distributed.launch
Author	SHA1	Message	Date
LysandreJik	eb0c17a3f8	Fix some of NielsRogge's comments	2020-11-12 17:07:27 -05:00
Lysandre	e00e47fbc8	TAPAS	2020-11-11 22:52:59 -05:00
Lysandre	e3e23045b1	Style	2020-11-11 21:21:04 -05:00
Rogge Niels	da59db93ad	Fix important typo	2020-11-04 10:42:05 +01:00
Rogge Niels	37d8870048	Address @LysandreJik's comments	2020-11-04 10:36:59 +01:00
Rogge Niels	11c0a0d905	Address PatrickVonPlaten's comments	2020-11-04 10:36:59 +01:00
Rogge Niels	7ee992e98c	Added tokenization_tapas_utilities to tokenization_tapas	2020-11-04 10:36:59 +01:00
Rogge Niels	c236808b72	Added missing imports to tokenization_tapas	2020-11-04 10:36:59 +01:00
Rogge Niels	9418015c27	Make TapasTokenizer inherit from PreTrainedTokenizer rather than BertTokenizer and remove modeling_tapas_utilities	2020-11-04 10:36:59 +01:00
Rogge Niels	2f44197b38	Added tapas utilities to modeling_tapas	2020-11-04 10:36:59 +01:00
Rogge Niels	45641355a3	Make TapasConfig inherit from PretrainedConfig instead of BertConfig and rename type_vocab_size to type_vocab_sizes	2020-11-04 10:36:59 +01:00
Rogge Niels	538a67eb46	Fixed docs comments LysandreJik and replaced add_start_docstrings_to_callable by add_start_docstrings_to_model_forward	2020-11-04 10:36:59 +01:00
Rogge Niels	42d5d27247	Fixed README	2020-11-04 10:36:59 +01:00
Rogge Niels	e25958043d	Added data utilities which were also not included	2020-11-04 10:36:53 +01:00
Rogge Niels	eb9910798b	Added herbert since this was not included	2020-11-04 10:35:45 +01:00
Rogge Niels	86aa165fcd	Fix numbering of models	2020-11-04 10:35:45 +01:00
Rogge Niels	58ff811920	Minor fix	2020-11-04 10:35:45 +01:00
Rogge Niels	78fc0f44c0	Improved testing	2020-11-04 10:35:45 +01:00
Rogge Niels	da1d0e06b7	Simplified aggregation head using nn.Linear	2020-11-04 10:35:45 +01:00
Rogge Niels	3114c8cd17	Bug fix with TapasForSequenceClassification	2020-11-04 10:35:45 +01:00
Rogge Niels	8f80607e9a	First attempt at using nn.linear rather than nn.parameter for TapasForSequenceClassification	2020-11-04 10:35:45 +01:00
Rogge Niels	13c7cdac47	Added fix for conversion of TapasForSequenceClassification	2020-11-04 10:35:45 +01:00
Rogge Niels	13cadf7a02	Some more improved documentation	2020-11-04 10:35:45 +01:00
Rogge Niels	3383e79fc0	Improved documentation	2020-11-04 10:35:45 +01:00
Rogge Niels	e6eeed9622	Improved loss calculation and documentation of TapasForQuestionAnswering	2020-11-04 10:35:45 +01:00
Rogge Niels	ac313f350f	Improved documentation	2020-11-04 10:35:45 +01:00
Rogge Niels	9bb00602a2	Added TapasForSequenceClassification	2020-11-04 10:35:44 +01:00
Rogge Niels	11429609fc	First commit	2020-11-04 10:35:38 +01:00
Ceyda Cinarel Ceyda Cinarel Lysandre Debut Lysandre	29b536a73a	[WIP] Ner pipeline grouped_entities fixes (#5970 ) * Bug fix: NER pipeline shouldn't group separate entities of same type * style fix * [Bug Fix] Shouldn't group entities that are both 'B' even if they are same type (B-type1 B-type1) != (B-type1 I-type1) [Bug Fix] add an option `ignore_subwords` to ignore subsequent ##wordpieces in predictions. Because some models train on only the first token of a word and not on the subsequent wordpieces (BERT NER default). So it makes sense doing the same thing at inference time. The simplest fix is to just group the subwords with the first wordpiece. [TODO] how to handle ignored scores? just set them to 0 and calculate zero invariant mean ? [TODO] handle different wordpiece_prefix ## ? possible approaches: get it from tokenizer? but currently most tokenizers dont have a wordpiece_prefix property? have an _is_subword(token) [Feature add] added option to `skip_special_tokens`. Cause It was harder to remove them after grouping. [Additional Changes] remove B/I prefix on returned grouped_entities [Feature Request/TODO] Return indexes? [Bug TODO] can't use fast tokenizer with grouped_entities ('BertTokenizerFast' object has no attribute 'convert_tokens_to_string') * use offset_mapping to fix [UNK] token problem * ignore score for subwords * modify ner_pipeline test * modify ner_pipeline test * modify ner_pipeline test * ner_pipeline change ignore_subwords default to true * add ner_pipeline ignore_subword=False test case * fix offset_mapping index * fix style again duh * change is_subword and convert_tokens_to_string logic * merge tests with new test structure * change test names * remove old tests * ner tests for fast tokenizer * fast tokenizers have convert_tokens_to_string * Fix the incorrect merge Co-authored-by: Ceyda Cinarel <snu-ceyda@users.noreply.github.com> Co-authored-by: Lysandre Debut <lysandre@huggingface.co> Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>	2020-11-03 17:21:04 -05:00
Stas Bekman	1bb4bba53c	[CIs] Better reports everywhere (#8275 ) * make it possible to invoke testconf.py in both test suites without crashing on having the same option added * perl -pi -e 's\|--make_reports\|--make-reports\|' to be consistent with other opts * add `pytest --make-reports` to all CIs (and artifacts) * fix	2020-11-03 16:57:12 -05:00
Sylvain Gugger	7f556d2e39	Data collator for token classification (#8274 ) * Add DataCollatorForTokenClassification and clean tests * Make quality	2020-11-03 16:33:27 -05:00
Philip MayandSylvain Gugger	6a064447f2	improve documentation of training_args.py (#8270 ) * improve documentation of training_args.py - do_train - do_eval - do_predict * fix line too long * fix style with black on training_args.py * Update src/transformers/training_args.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/training_args.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/training_args.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * fix line length with utils/style_doc * black reformatting Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>	2020-11-03 15:57:17 -05:00
Sylvain Gugger	4c19f3baab	Clean Trainer tests and datasets dep (#8268 )	2020-11-03 15:50:55 -05:00
Patrick von Platen	068e6b5edd	make files independent (#8267 )	2020-11-03 21:13:33 +01:00