boom boom

2020-06-25 02:10:13 +00:00 · 2020-06-24 20:44:27 +00:00 · 2020-06-23 16:10:00 -04:00 · 2020-06-23 14:35:54 -04:00 · 2020-06-19 14:46:51 +00:00 · 2020-06-19 09:43:20 -04:00
718 changed files with 19624 additions and 52581 deletions
@@ -1,66 +1,4 @@
-version: 2.1
-orbs:
-    gcp-gke: circleci/gcp-gke@1.0.4
-    go: circleci/go@1.3.0
-
-# TPU REFERENCES
-references:
-    checkout_ml_testing: &checkout_ml_testing
-        run:
-            name: Checkout ml-testing-accelerators
-            command: |
-                git clone https://github.com/GoogleCloudPlatform/ml-testing-accelerators.git
-                cd ml-testing-accelerators
-                git fetch origin 5e88ac24f631c27045e62f0e8d5dfcf34e425e25:stable
-                git checkout stable
-    build_push_docker: &build_push_docker
-        run:
-            name: Configure Docker
-            command: |
-                gcloud --quiet auth configure-docker
-                cd docker/transformers-pytorch-tpu
-                if [ -z "$CIRCLE_PR_NUMBER" ]; then docker build --tag "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" -f Dockerfile --build-arg "TEST_IMAGE=1" . ; else docker build --tag "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" -f Dockerfile --build-arg "TEST_IMAGE=1" --build-arg "GITHUB_REF=pull/$CIRCLE_PR_NUMBER/head" . ; fi
-                docker push "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID"
-    deploy_cluster: &deploy_cluster
-        run:
-            name: Deploy the job on the kubernetes cluster
-            command: |
-                go get github.com/google/go-jsonnet/cmd/jsonnet && \
-                export PATH=$PATH:$HOME/go/bin && \
-                kubectl create -f docker/transformers-pytorch-tpu/dataset.yaml || true && \
-                job_name=$(jsonnet -J ml-testing-accelerators/ docker/transformers-pytorch-tpu/bert-base-cased.jsonnet --ext-str image=$GCR_IMAGE_PATH --ext-str image-tag=$CIRCLE_WORKFLOW_JOB_ID | kubectl create -f -) && \
-                job_name=${job_name#job.batch/} && \
-                job_name=${job_name% created} && \
-                echo "Waiting on kubernetes job: $job_name" && \
-                i=0 && \
-                # 30 checks spaced 30s apart = 900s total.
-                max_checks=30 && \
-                status_code=2 && \
-                # Check on the job periodically. Set the status code depending on what
-                # happened to the job in Kubernetes. If we try max_checks times and
-                # still the job hasn't finished, give up and return the starting
-                # non-zero status code.
-                while [ $i -lt $max_checks ]; do ((i++)); if kubectl get jobs $job_name -o jsonpath='Failed:{.status.failed}' | grep "Failed:1"; then status_code=1 && break; elif kubectl get jobs $job_name -o jsonpath='Succeeded:{.status.succeeded}' | grep "Succeeded:1" ; then status_code=0 && break; else echo "Job not finished yet"; fi; sleep 30; done && \
-                echo "Done waiting. Job status code: $status_code" && \
-                pod_name=$(kubectl get po -l controller-uid=`kubectl get job $job_name -o "jsonpath={.metadata.labels.controller-uid}"` | awk 'match($0,!/NAME/) {print $1}') && \
-                echo "GKE pod name: $pod_name" && \
-                kubectl logs -f $pod_name --container=train
-                echo "Done with log retrieval attempt." && \
-                gcloud container images delete "$GCR_IMAGE_PATH:$CIRCLE_WORKFLOW_JOB_ID" --force-delete-tags && \
-                exit $status_code
-    delete_gke_jobs: &delete_gke_jobs
-        run:
-            name: Delete GKE Jobs
-            command: |
-                # Match jobs whose age matches patterns like '1h' or '1d', i.e. any job
-                # that has been around longer than 1hr. First print all columns for
-                # matches, then execute the delete.
-                kubectl get job | awk 'match($4,/[0-9]+[dh]/) {print $0}'
-                kubectl delete job $(kubectl get job | awk 'match($4,/[0-9]+[dh]/) {print $1}')
-
-
-
-
+version: 2
 jobs:
    run_tests_torch_and_tf:
        working_directory: ~/transformers
@@ -72,22 +10,11 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-torch_and_tf-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,torch,testing]
-            - run: pip install codecov pytest-cov
-            - save_cache:
-                key: v0.3-{{ checksum "setup.py" }}
-                paths:
-                    - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ --cov  | tee output.txt
+            - run: sudo pip install .[sklearn,tf-cpu,torch,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
            - run: codecov
-            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+
    run_tests_torch:
        working_directory: ~/transformers
        docker:
@@ -98,20 +25,10 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-torch-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing]
-            - save_cache:
-                  key: v0.3-torch-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ | tee output.txt
-            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+            - run: sudo pip install .[sklearn,torch,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
+            - run: codecov
    run_tests_tf:
        working_directory: ~/transformers
        docker:
@@ -122,20 +39,10 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-tf-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,tf-cpu,testing]
-            - save_cache:
-                  key: v0.3-tf-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ | tee output.txt
-            - store_artifacts:
-               path: ~/transformers/output.txt
-               destination: test_output.txt
+            - run: sudo pip install .[sklearn,tf-cpu,testing]
+            - run: sudo pip install codecov pytest-cov
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
+            - run: codecov
    run_tests_custom_tokenizers:
        working_directory: ~/transformers
        docker:
@@ -144,20 +51,8 @@ jobs:
            RUN_CUSTOM_TOKENIZERS: yes
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-custom_tokenizers-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[ja,testing]
-            - save_cache:
-                  key: v0.3-custom_tokenizers-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
-            - store_artifacts:
-                path: ~/transformers/output.txt
-                destination: test_output.txt
+            - run: sudo pip install .[mecab,testing]
+            - run: python -m pytest -sv ./tests/test_tokenization_bert_japanese.py
    run_examples_torch:
        working_directory: ~/transformers
        docker:
@@ -168,37 +63,16 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-torch_examples-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[sklearn,torch,testing]
-            - run: pip install -r examples/requirements.txt
-            - save_cache:
-                  key: v0.3-torch_examples-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
-            - run: python -m pytest -n 8 --dist=loadfile -rA -s ./examples/ | tee output.txt
-            - store_artifacts:
-                  path: ~/transformers/output.txt
-                  destination: test_output.txt
+            - run: sudo pip install .[sklearn,torch,testing]
+            - run: sudo pip install -r examples/requirements.txt
+            - run: python -m pytest -n 8 --dist=loadfile -s -v ./examples/
    build_doc:
        working_directory: ~/transformers
        docker:
            - image: circleci/python:3.6
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-build_doc-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
-            - run: pip install .[tf,torch,docs]
-            - save_cache:
-                  key: v0.3-build_doc-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
+            - run: sudo pip install .[tf,torch,docs]
            - run: cd docs && make html SPHINXOPTS="-W"
            - store_artifacts:
                path: ./docs/_build
@@ -211,15 +85,7 @@ jobs:
                fingerprints:
                    - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-deploy_doc-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install .[tf,torch,docs]
-            - save_cache:
-                  key: v0.3-deploy_doc-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
+            - run: sudo pip install .[tf,torch,docs]
            - run: ./.circleci/deploy.sh
    check_code_quality:
        working_directory: ~/transformers
@@ -229,22 +95,12 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - restore_cache:
-                  keys:
-                      - v0.3-code_quality-{{ checksum "setup.py" }}
-                      - v0.3-{{ checksum "setup.py" }}
-            - run: pip install --upgrade pip
            # we need a version of isort with https://github.com/timothycrosley/isort/pull/1000
-            - run: pip install git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
-            - run: pip install .[tf,torch,quality]
-            - save_cache:
-                  key: v0.3-code_quality-{{ checksum "setup.py" }}
-                  paths:
-                      - '~/.cache/pip'
+            - run: sudo pip install git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
+            - run: sudo pip install .[tf,torch,quality]
            - run: black --check --line-length 119 --target-version py35 examples templates tests src utils
            - run: isort --check-only --recursive examples templates tests src utils
            - run: flake8 examples templates tests src utils
-            - run: python utils/check_repo.py
    check_repository_consistency:
        working_directory: ~/transformers
        docker:
@@ -253,37 +109,8 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: pip install requests
+            - run: sudo pip install requests
            - run: python ./utils/link_tester.py
-
-# TPU JOBS
-    run_examples_tpu:
-        docker:
-            - image: circleci/python:3.6
-        environment:
-            OMP_NUM_THREADS: 1
-        resource_class: xlarge
-        parallelism: 1
-        steps:
-            - checkout
-            - go/install
-            - *checkout_ml_testing
-            - gcp-gke/install
-            - gcp-gke/update-kubeconfig-with-credentials:
-                  cluster: $GKE_CLUSTER
-                  perform-login: true
-            - setup_remote_docker
-            - *build_push_docker
-            - *deploy_cluster
-    cleanup-gke-jobs:
-        docker:
-            - image: circleci/python:3.6
-        steps:
-            - gcp-gke/install
-            - gcp-gke/update-kubeconfig-with-credentials:
-                  cluster: $GKE_CLUSTER
-                  perform-login: true
-            - *delete_gke_jobs
 workflow_filters: &workflow_filters
    filters:
        branches:
@@ -302,15 +129,3 @@ workflows:
            - run_tests_tf
            - build_doc
            - deploy_doc: *workflow_filters
-    tpu_testing_jobs:
-        triggers:
-            - schedule:
-                # Set to run at the first minute of every hour.
-                cron: "0 8 * * *"
-                filters:
-                    branches:
-                        only:
-                            - master
-        jobs:
-            - cleanup-gke-jobs
-            - run_examples_tpu
@@ -5,26 +5,15 @@ function deploy_doc(){
 	git checkout $1
 	if [ ! -z "$2" ]
 	then
-		if [ "$2" == "master" ]; then
-		    echo "Pushing master"
-			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir/$2/
-			cp -r _build/html/_static .
-		elif ssh -oStrictHostKeyChecking=no $doc "[ -d $dir/$2 ]"; then
+		if [ -d "$dir/$2" ]; then
 			echo "Directory" $2 "already exists"
-			scp -r -oStrictHostKeyChecking=no _static/* $doc:$dir/$2/_static/
 		else
 			echo "Pushing version" $2
-			make clean && make html
-			rm -rf _build/html/_static
-			cp -r _static _build/html
-			scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
 		fi
 	else
-		echo "Pushing stable"
-		make clean && make html
-		rm -rf _build/html/_static
-		cp -r _static _build/html
-		scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
 	fi
 }

@@ -46,5 +35,4 @@ deploy_doc "11c3257" v2.8.0
 deploy_doc "e7cfc1a" v2.9.0
 deploy_doc "7cb203f" v2.9.1
 deploy_doc "10d7239" v2.10.0 
-deploy_doc "b42586e" v2.11.0
-deploy_doc "7fb8bdf" #v3.0.2 Latest stable release
+deploy_doc "b42586e" #v2.11.0 Latest stable release
@@ -7,51 +7,14 @@ assignees: ''

 ---

-
-## Environment info
-<!-- You can run the command `transformers-cli env` and copy-and-paste its output below.
-     Don't forget to fill out the missing fields in that output! -->
-     
- `transformers` version:
- Platform:
- Python version:
- PyTorch version (GPU?):
- Tensorflow version (GPU?):
- Using GPU in script?:
- Using distributed or parallel set-up in script?:
-
-### Who can help
-<!-- Your issue will be replied to more quickly if you can figure out the right person to tag with @
- If you know how to use git blame, that is the easiest way, otherwise, here is a rough guide of **who to tag**.
- Please tag fewer than 3 people.
- 
- albert, bert, GPT2, XLM: @LysandreJik 
- tokenizers: @mfuntowicz
- Trainer: @sgugger
- Speed and Memory Benchmarks: @patrickvonplaten
- Model Cards: @julien-c
- Translation: @sshleifer
- Summarization: @sshleifer
- TextGeneration: @TevenLeScao 
- examples/distillation: @VictorSanh
- nlp datasets: [different repo](https://github.com/huggingface/nlp)
- rust tokenizers: [different repo](https://github.com/huggingface/tokenizers)
- Text Generation: @TevenLeScao
- blenderbot: @mariamabarham
- Bart: @sshleifer
- Marian: @sshleifer
- T5: @patrickvonplaten
- Longformer/Reformer: @patrickvonplaten
- TransfoXL/XLNet: @TevenLeScao 
- examples/seq2seq: @sshleifer
- tensorflow: @jplu 
-documentation: @sgugger
- -->
+# 🐛 Bug

 ## Information

 Model I am using (Bert, XLNet ...):

+Language I am using the model on (English, Chinese ...):
+
 The problem arises when using:
 * [ ] the official example scripts: (give details below)
 * [ ] my own modified scripts: (give details below)
@@ -75,3 +38,15 @@ Steps to reproduce the behavior:
 ## Expected behavior

 <!-- A clear and concise description of what you would expect to happen. -->
+
+## Environment info
+<!-- You can run the command `transformers-cli env` and copy-and-paste its output below.
+     Don't forget to fill out the missing fields in that output! -->
+     
+- `transformers` version:
+- Platform:
+- Python version:
+- PyTorch version (GPU?):
+- Tensorflow version (GPU?):
+- Using GPU in script?:
+- Using distributed or parallel set-up in script?:
@@ -1,6 +1,6 @@
 ---
 name: "❓ Questions & Help"
-about: Post your general questions on the Hugging Face forum or Stack Overflow tagged huggingface-transformers
+about: Post your general questions on Stack Overflow tagged huggingface-transformers
 title: ''
 labels: ''
 assignees: ''
@@ -11,17 +11,19 @@ assignees: ''

 <!-- The GitHub issue tracker is primarly intended for bugs, feature requests,
     new models and benchmarks, and migration questions. For all other questions,
-     we direct you to the Hugging Face forum: https://discuss.huggingface.co/ .
-     You can also try Stack Overflow (SO) where a whole community of PyTorch and
-     Tensorflow enthusiast can help you out. In this case, make sure to tag your
-     question with the right deep learning framework as well as the
-     huggingface-transformers tag: 
+     we direct you to Stack Overflow (SO) where a whole community of PyTorch and
+     Tensorflow enthusiast can help you out. Make sure to tag your question with the
+     right deep learning framework as well as the huggingface-transformers tag: 
     https://stackoverflow.com/questions/tagged/huggingface-transformers 
+     
+     If your question wasn't answered after a period of time on Stack Overflow, you
+     can always open a question on GitHub. You should then link to the SO question 
+     that you posted.
     -->

 ## Details
 <!-- Description of your issue -->

-<!-- You should first ask your question on the forum or SO, and only if
+<!-- You should first ask your question on SO, and only if
     you didn't get an answer ask it here on GitHub. -->
-**A link to original question on the forum/Stack Overflow**:
+**A link to original question on Stack Overflow**:
@@ -0,0 +1,19 @@
+name: GitHub-hosted runner
+
+on: push
+
+jobs:
+  check_code_quality:
+    runs-on: ubuntu-18.04
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v1
+      with:
+        python-version: 3.7
+    # - name: Install dependencies
+    #   run: |
+    #     pip install .[tf,torch,quality]
+
+
+
@@ -18,17 +18,8 @@ jobs:
      uses: actions/setup-python@v1
      with:
        python-version: 3.7
-
-    - name: Loading cache
-      uses: actions/cache@v2
-      id: cache
-      with:
-        path: ~/.cache/pip
-        key: v0-torch_hub-${{ hashFiles('setup.py') }}
-
    - name: Install dependencies
      run: |
-        pip install --upgrade pip
        pip install torch
        pip install numpy tokenizers filelock requests tqdm regex sentencepiece sacremoses packaging

@@ -25,14 +25,6 @@ jobs:
    - name: Current dir
      run: pwd
    - run: nvidia-smi
-
-    - name: Loading cache.
-      uses: actions/cache@v2
-      id: cache
-      with:
-        path: .env
-        key: v0-tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
-
    - name: Create new python env (on self-hosted runners we have to handle isolation ourselves)
      run: |
        python -m venv .env
@@ -43,8 +35,7 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install --upgrade pip
-        pip install torch!=1.6.0
+        pip install torch
        pip install .[sklearn,testing]

    - name: Are GPUs recognized by our DL frameworks
@@ -60,4 +51,4 @@ jobs:
        USE_CUDA: yes
      run: |
        source .env/bin/activate
-        python -m pytest -n 2 --dist=loadfile -s ./tests/
+        python -m pytest -n 2 --dist=loadfile -s -v ./tests/
@@ -13,14 +13,6 @@ jobs:
    runs-on: self-hosted
    steps:
    - uses: actions/checkout@v2
-
-    - name: Loading cache.
-      uses: actions/cache@v2
-      id: cache
-      with:
-        path: .env
-        key: v0-slow_tests_tf_torch_gpu-${{ hashFiles('setup.py') }}
-
    - name: Python version
      run: |
        which python
@@ -30,7 +22,6 @@ jobs:
      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
@@ -40,9 +31,7 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install --upgrade pip
-        pip install torch!=1.6.0
-        pip install .[sklearn,testing]
+        pip install .[sklearn,torch,testing]

    - name: Are GPUs recognized by our DL frameworks
      run: |
@@ -57,14 +46,5 @@ jobs:
        USE_CUDA: yes
      run: |
        source .env/bin/activate
-        python -m pytest -n 1 --dist=loadfile -s ./tests/
-    - name: Run examples tests on GPU
-      env:
-        TF_FORCE_GPU_ALLOW_GROWTH: "true"
-        OMP_NUM_THREADS: 1
-        RUN_SLOW: yes
-        USE_CUDA: yes
-      run: |
-        source .env/bin/activate
-        pip install -r examples/requirements.txt
-        python -m pytest -n 1 --dist=loadfile -s examples
+        python -m pytest -n 1 --dist=loadfile -s -v ./tests/
+        
@@ -66,7 +66,7 @@ If you are willing to contribute the model yourself, let us know so we can best
 guide you.

 We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them 
-in the [`templates`](https://github.com/huggingface/transformers/tree/master/templates) folder.
+in the [`templates`](https://github.com/huggingface/transformers/templates) folder.

 ### Do you want a new feature (that is not a model)?

@@ -88,7 +88,7 @@ If your issue is well written we're already 80% of the way there by the time you
 post it.

 We have added **templates** to guide you in the process of adding a new example script for training or testing the 
-models in the library. You can find them in the [`templates`](https://github.com/huggingface/transformers/tree/master/templates) 
+models in the library. You can find them in the [`templates`](https://github.com/huggingface/transformers/templates) 
 folder.

 ## Start contributing! (Pull Requests)
@@ -215,7 +215,7 @@ Follow these steps to start contributing:
     `RUN_SLOW=1 python -m pytest tests/test_my_new_model.py`. 
   - If you are adding a new tokenizer, write tests, and make sure 
     `RUN_SLOW=1 python -m pytest tests/test_tokenization_{your_model_name}.py` passes.
-   CircleCI does not run the slow tests, but github actions does every night!
+   CircleCI does not run the slow tests. 
 6. All public methods must have informative docstrings that work nicely with sphinx. See `modeling_ctrl.py` for an 
   example.

@@ -238,7 +238,8 @@ and for the examples:
 $ pip install -r examples/requirements.txt  # only needed the first time
 $ python -m pytest -n auto --dist=loadfile -s -v ./examples/
 ```
-In fact, that's how `make test` and `make test-examples` are implemented (sans the `pip install` line)!
+
+In fact, that's how `make test` and `make test-examples` are implemented!

 You can specify a smaller set of tests in order to test only the feature
 you're working on.
@@ -6,7 +6,6 @@ quality:
 	black --check --line-length 119 --target-version py35 examples templates tests src utils
 	isort --check-only --recursive examples templates tests src utils
 	flake8 examples templates tests src utils
-	python utils/check_repo.py

 # Format source code automatically

@@ -24,7 +24,6 @@

 🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides state-of-the-art general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet, T5, CTRL...) for Natural Language Understanding (NLU) and Natural Language Generation (NLG) with over thousands of pretrained models in 100+ languages and deep interoperability between PyTorch & TensorFlow 2.0.

-### Recent contributors
 [![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/0)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/0)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/1)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/1)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/2)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/2)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/3)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/3)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/4)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/4)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/5)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/5)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/6)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/6)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/7)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/7)

 ### Features
@@ -60,7 +59,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
+| Documentation [(master)](https://huggingface.co/transformers/master) [(stable)](https://huggingface.co/transformers/) [(v2.10.0)](https://huggingface.co/transformers/v2.10.0) [(v2.9.0/v2.9.1)](https://huggingface.co/transformers/v2.9.1) [(v2.8.0)](https://huggingface.co/transformers/v2.8.0) [(v2.7.0)](https://huggingface.co/transformers/v2.7.0) [(v2.6.0)](https://huggingface.co/transformers/v2.6.0) [(v2.5.0/v2.5.1)](https://huggingface.co/transformers/v2.5.1) [(v2.4.0/v2.4.1)](https://huggingface.co/transformers/v2.4.0)[(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) | Full API documentation and more |

 ## Installation

@@ -174,19 +173,12 @@ These implementations have been tested on several datasets (see the example scri

 ## Online demo

-You can test our inference API on most model pages from the model hub: https://huggingface.co/models
-
-For example: 
- [Masked word completion with BERT](https://huggingface.co/bert-base-uncased?text=Paris+is+the+%5BMASK%5D+of+France)
- [NER with Electra](https://huggingface.co/dbmdz/electra-large-discriminator-finetuned-conll03-english?text=My+name+is+Sarah+and+I+live+in+London+city)
- [Text generation with GPT-2](https://huggingface.co/gpt2?text=A+long+time+ago%2C+)
- [NLI with RoBERTa](https://huggingface.co/roberta-large-mnli?text=The+dog+was+lost.+Nobody+lost+any+animal)
- [Summarization with BART](https://huggingface.co/facebook/bart-large-cnn?text=The+tower+is+324+metres+%281%2C063+ft%29+tall%2C+about+the+same+height+as+an+81-storey+building%2C+and+the+tallest+structure+in+Paris.+Its+base+is+square%2C+measuring+125+metres+%28410+ft%29+on+each+side.+During+its+construction%2C+the+Eiffel+Tower+surpassed+the+Washington+Monument+to+become+the+tallest+man-made+structure+in+the+world%2C+a+title+it+held+for+41+years+until+the+Chrysler+Building+in+New+York+City+was+finished+in+1930.+It+was+the+first+structure+to+reach+a+height+of+300+metres.+Due+to+the+addition+of+a+broadcasting+aerial+at+the+top+of+the+tower+in+1957%2C+it+is+now+taller+than+the+Chrysler+Building+by+5.2+metres+%2817+ft%29.+Excluding+transmitters%2C+the+Eiffel+Tower+is+the+second+tallest+free-standing+structure+in+France+after+the+Millau+Viaduct)
- [Question answering with DistilBERT](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
- [Translation with T5](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
-
-
 **[Write With Transformer](https://transformer.huggingface.co)**, built by the Hugging Face team at transformer.huggingface.co, is the official demo of this repo’s text generation capabilities.
+You can use it to experiment with completions generated by `GPT2Model`, `TransfoXLModel`, and `XLNetModel`.
+
+> “🦄 Write with transformer is to writing what calculators are to calculus.”
+
+![write_with_transformer](https://transformer.huggingface.co/front/assets/thumbnail-large.png)

 ## Quick tour

@@ -295,8 +287,8 @@ pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf
 sentence_0 = "This research was consistent with his findings."
 sentence_1 = "His findings were compatible with this research."
 sentence_2 = "His findings were not compatible with this research."
-inputs_1 = tokenizer(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
-inputs_2 = tokenizer(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
+inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
+inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')

 pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
 pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()
@@ -0,0 +1,23 @@
+cd docs
+
+function deploy_doc(){
+	echo "Creating doc at commit $1 and pushing to folder $2"
+	git checkout $1
+	if [ ! -z "$2" ] 
+	then
+		echo "Pushing version" $2
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+	else
+		echo "Pushing master"
+		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+	fi
+}
+
+deploy_doc "master" 
+deploy_doc "b33a385" v1.0.0
+deploy_doc "fe02e45" v1.1.0
+deploy_doc "89fd345" v1.2.0
+deploy_doc "fc9faa8" v2.0.0
+deploy_doc "3ddce1d" v2.1.1
+deploy_doc "f2f3294" v2.2.0
+deploy_doc "d0f8b9a" v2.3.0
@@ -1,65 +0,0 @@
-FROM google/cloud-sdk:slim
-
-# Build args.
-ARG GITHUB_REF=refs/heads/master
-
-# TODO: This Dockerfile installs pytorch/xla 3.6 wheels. There are also 3.7
-# wheels available; see below.
-ENV PYTHON_VERSION=3.6
-
-RUN apt-get update && apt-get install -y --no-install-recommends \
-         build-essential \
-         cmake \
-         git \
-         curl \
-         ca-certificates
-
-# Install conda and python.
-# NOTE new Conda does not forward the exit status... https://github.com/conda/conda/issues/8385
-RUN curl -o ~/miniconda.sh https://repo.anaconda.com/miniconda/Miniconda3-4.7.12-Linux-x86_64.sh  && \
-    chmod +x ~/miniconda.sh && \
-    ~/miniconda.sh -b && \
-    rm ~/miniconda.sh
-
-ENV PATH=/root/miniconda3/bin:$PATH
-
-RUN conda create -y --name container python=$PYTHON_VERSION
-
-# Run the rest of commands within the new conda env.
-# Use absolute path to appease Codefactor.
-SHELL ["/root/miniconda3/bin/conda", "run", "-n", "container", "/bin/bash", "-c"]
-RUN conda install -y python=$PYTHON_VERSION mkl
-
-RUN pip uninstall -y torch && \
-    # Python 3.7 wheels are available. Replace cp36-cp36m with cp37-cp37m
-    gsutil cp 'gs://tpu-pytorch/wheels/torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
-    gsutil cp 'gs://tpu-pytorch/wheels/torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
-    gsutil cp 'gs://tpu-pytorch/wheels/torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' . && \
-    pip install 'torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    pip install 'torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    pip install 'torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    rm 'torch-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    rm 'torch_xla-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    rm 'torchvision-nightly-cp${PYTHON_VERSION/./}-cp${PYTHON_VERSION/./}m-linux_x86_64.whl' && \
-    apt-get install -y libomp5
-
-ENV LD_LIBRARY_PATH=root/miniconda3/envs/container/lib
-
-
-# Install huggingface/transformers at the current PR, plus dependencies.
-RUN git clone https://github.com/huggingface/transformers.git && \
-    cd transformers && \
-    git fetch origin $GITHUB_REF:CI && \
-    git checkout CI && \
-    cd .. && \
-    pip install ./transformers && \
-    pip install -r ./transformers/examples/requirements.txt && \
-    pip install pytest
-
-RUN python -c "import torch_xla; print(torch_xla.__version__)"
-RUN python -c "import transformers as trf; print(trf.__version__)"
-RUN conda init bash
-COPY docker-entrypoint.sh /usr/local/bin/
-RUN chmod +x /usr/local/bin/docker-entrypoint.sh
-ENTRYPOINT ["/usr/local/bin/docker-entrypoint.sh"]
-CMD ["bash"]
@@ -1,38 +0,0 @@
-local base = import 'templates/base.libsonnet';
-local tpus = import 'templates/tpus.libsonnet';
-local utils = import "templates/utils.libsonnet";
-local volumes = import "templates/volumes.libsonnet";
-
-local bertBaseCased = base.BaseTest {
-  frameworkPrefix: "hf",
-  modelName: "bert-base-cased",
-  mode: "example",
-  configMaps: [],
-
-  timeout: 3600, # 1 hour, in seconds
-
-  image: std.extVar('image'),
-  imageTag: std.extVar('image-tag'),
-
-  tpuSettings+: {
-    softwareVersion: "pytorch-nightly",
-  },
-  accelerator: tpus.v3_8,
-
-  volumeMap+: {
-    datasets: volumes.PersistentVolumeSpec {
-      name: "huggingface-cluster-disk",
-      mountPath: "/datasets",
-    },
-  },
-  command: utils.scriptCommand(
-    |||
-      python -m pytest -s transformers/examples/test_xla_examples.py -v
-      test_exit_code=$?
-      echo "\nFinished running commands.\n"
-      test $test_exit_code -eq 0
-    |||
-  ),
-};
-
-bertBaseCased.oneshotJob
@@ -1,32 +0,0 @@
-apiVersion: v1
-kind: PersistentVolume
-metadata:
-  name: huggingface-cluster-disk
-spec:
-  storageClassName: ""
-  capacity:
-    storage: 500Gi
-  accessModes:
-    - ReadOnlyMany
-  claimRef:
-    namespace: default
-    name: huggingface-cluster-disk-claim
-  gcePersistentDisk:
-    pdName: huggingface-cluster-disk
-    fsType: ext4
-    readOnly: true
---
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: huggingface-cluster-disk-claim
-spec:
-  # Specify "" as the storageClassName so it matches the PersistentVolume's StorageClass.
-  # A nil storageClassName value uses the default StorageClass. For details, see
-  # https://kubernetes.io/docs/concepts/storage/persistent-volumes/#class-1
-  storageClassName: ""
-  accessModes:
-    - ReadOnlyMany
-  resources:
-    requests:
-      storage: 1Ki
@@ -1,8 +0,0 @@
-#!/bin/bash
-source ~/.bashrc
-echo "running docker-entrypoint.sh"
-conda activate container
-echo $KUBE_GOOGLE_CLOUD_TPU_ENDPOINTS
-echo "printed TPU info"
-export XRT_TPU_CONFIG="tpu_worker;0;${KUBE_GOOGLE_CLOUD_TPU_ENDPOINTS:7}"
-exec "$@"#!/bin/bash
@@ -167,7 +167,7 @@ Here's an example showcasing everything so far:

            Indices can be obtained using :class:`transformers.AlbertTokenizer`.
            See :func:`transformers.PreTrainedTokenizer.encode` and
-            :func:`transformers.PreTrainedTokenizer.__call__` for details.
+            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.

            `What are input IDs? <../glossary.html#input-ids>`__
 ```
@@ -9,8 +9,4 @@

 .highlight .kn, .highlight .nv, .highlight .s2, .highlight .ow {
    color: #6670FF;
-}
-
-.highlight .gp {
-    color: #FB8D68;
 }
@@ -1,81 +1,9 @@
 /* Our DOM objects */

-/* Colab dropdown */
-
-.colab-dropdown {
-    position: relative;
-    display: inline-block;
-}
-  
-.colab-dropdown-content {
-    display: none;
-    position: absolute;
-    background-color: #f9f9f9;
-    min-width: 117px;
-    box-shadow: 0px 8px 16px 0px rgba(0,0,0,0.2);
-    z-index: 1;
-}
-  
-.colab-dropdown-content button {
-    color: #6670FF;
-    background-color: #f9f9f9;
-    font-size: 12px;
-    border: none;
-    min-width: 117px;
-    padding: 5px 5px;
-    text-decoration: none;
-    display: block;
-}
-  
-.colab-dropdown-content button:hover {background-color: #eee;}
-  
-.colab-dropdown:hover .colab-dropdown-content {display: block;}
-
-/* Version control */
-
-.version-button {
-    background-color: #6670FF;
-    color: white;
-    border: none;
-    padding: 5px;
-    font-size: 15px;
-    cursor: pointer;
-}
-
-.version-button:hover, .version-button:focus {
-    background-color: #A6B0FF;
-}
- 
-.version-dropdown {
-    display: none;
-    background-color: #6670FF;
-    min-width: 160px;
-    overflow: auto;
-    font-size: 15px;
-}
-  
-.version-dropdown a {
-    color: white;
-    padding: 3px 4px;
-    text-decoration: none;
-    display: block;
-}
-  
-.version-dropdown a:hover {
-    background-color: #A6B0FF;
-}
-  
-.version-show {
-    display: block;
-}
-
-/* Framework selector */
-
 .framework-selector {
    display: flex;
    flex-direction: row;
    justify-content: flex-end;
-    margin-right: 30px;
 }

 .framework-selector > button {
@@ -92,12 +20,6 @@
    padding: 5px;
 }

-/* Copy button */
-
-a.copybtn {
-    margin: 3px;
-}
-
 /* The literal code blocks */
 .rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
    color: #6670FF;
@@ -116,7 +38,6 @@ a.copybtn {

 /* The research field on top of the toc tree */
 .wy-side-nav-search{
-    padding-top: 0;
    background-color: #6670FF;
 }

@@ -1,38 +1,3 @@
-// These two things need to be updated at each release for the version selector.
-// Last stable version
-const stableVersion = "v3.0.2"
-// Dictionary doc folder to label
-const versionMapping = {
-    "master": "master",
-    "": "v3.0.0/v3.0.1/v3.0.2 (stable)",
-    "v2.11.0": "v2.11.0",
-    "v2.10.0": "v2.10.0",
-    "v2.9.1": "v2.9.0/v2.9.1",
-    "v2.8.0": "v2.8.0",
-    "v2.7.0": "v2.7.0",
-    "v2.6.0": "v2.6.0",
-    "v2.5.1": "v2.5.0/v2.5.1",
-    "v2.4.0": "v2.4.0/v2.4.1",
-    "v2.3.0": "v2.3.0",
-    "v2.2.0": "v2.2.0/v2.2.1/v2.2.2",
-    "v2.1.1": "v2.1.1",
-    "v2.0.0": "v2.0.0",
-    "v1.2.0": "v1.2.0",
-    "v1.1.0": "v1.1.0",
-    "v1.0.0": "v1.0.0"
-}
-// The page that have a notebook and therefore should have the open in colab badge.
-const hasNotebook = [
-    "benchmarks",
-    "multilingual",
-    "perplexity",
-    "preprocessing",
-    "quicktour",
-    "task_summary",
-    "tokenizer_summary",
-    "training"
-];
-
 function addIcon() {
    const huggingFaceLogo = "https://huggingface.co/landing/assets/transformers-docs/huggingface_logo.svg";
    const image = document.createElement("img");
@@ -93,94 +58,11 @@ function addGithubButton() {
    document.querySelector(".wy-side-nav-search .icon-home").insertAdjacentHTML('afterend', div);
 }

-function addColabLink() {
-    const parts = location.toString().split('/');
-    const pageName = parts[parts.length - 1].split(".")[0];
-
-    if (hasNotebook.includes(pageName)) {
-        const baseURL = "https://colab.research.google.com/github/huggingface/notebooks/blob/master/transformers_doc/"
-        const linksColab = `
-        <div class="colab-dropdown">
-            <img alt="Open In Colab" src="https://colab.research.google.com/assets/colab-badge.svg">
-            <div class="colab-dropdown-content">
-                <button onclick=" window.open('${baseURL}${pageName}.ipynb')">Mixed</button>
-                <button onclick=" window.open('${baseURL}pytorch/${pageName}.ipynb')">PyTorch</button>
-                <button onclick=" window.open('${baseURL}tensorflow/${pageName}.ipynb')">TensorFlow</button>
-            </div>
-        </div>`
-        const leftMenu = document.querySelector(".wy-breadcrumbs-aside")
-        leftMenu.innerHTML = linksColab + '\n' + leftMenu.innerHTML
-    }
-}
-
-function addVersionControl() {
-    // To grab the version currently in view, we parse the url
-    const parts = location.toString().split('/');
-    let versionIndex = parts.length - 2;
-    // Index page may not have a last part with filename.html so we need to go up
-    if (parts[parts.length - 1] != "" && ! parts[parts.length - 1].match(/\.html$|^search.html?/)) {
-        versionIndex = parts.length - 1;
-    }
-    // Main classes and models are nested so we need to go deeper
-    else if (parts[versionIndex] == "main_classes" || parts[versionIndex] == "model_doc") {
-        versionIndex = versionIndex - 1;
-    } 
-    const version = parts[versionIndex];
-
-    // Menu with all the links,
-    const versionMenu = document.createElement("div");
-
-    const htmlLines = [];
-    for (const [key, value] of Object.entries(versionMapping)) {
-        let baseUrlIndex = (version == "transformers") ? versionIndex + 1: versionIndex;
-        var urlParts = parts.slice(0, baseUrlIndex);
-        if (key != "") {
-            urlParts = urlParts.concat([key]);
-        }
-        urlParts = urlParts.concat(parts.slice(versionIndex+1));
-        htmlLines.push(`<a href="${urlParts.join('/')}">${value}</a>`);
-    }
-
-    versionMenu.classList.add("version-dropdown");
-    versionMenu.innerHTML = htmlLines.join('\n');
-    
-    // Button for version selection
-    const versionButton = document.createElement("div");
-    versionButton.classList.add("version-button");
-    let label = (version == "transformers") ? stableVersion : version
-    versionButton.innerText = label.concat(" ▼");
-
-    // Toggle the menu when we click on the button
-    versionButton.addEventListener("click", () => {
-        versionMenu.classList.toggle("version-show");
-    });
-
-    // Hide the menu when we click elsewhere
-    window.addEventListener("click", (event) => {
-        if (event.target != versionButton){
-            versionMenu.classList.remove('version-show');
-        }
-    });
-
-    // Container
-    const div = document.createElement("div");
-    div.appendChild(versionButton);
-    div.appendChild(versionMenu);
-    div.style.paddingTop = '25px';
-    div.style.backgroundColor = '#6670FF';
-    div.style.display = 'block';
-    div.style.textAlign = 'center';
-
-    const scrollDiv = document.querySelector(".wy-side-scroll");
-    scrollDiv.insertBefore(div, scrollDiv.children[1]);
-}
-
 function addHfMenu() {
    const div = `
    <div class="menu">
        <a href="/welcome">🔥 Sign in</a>
        <a href="/models">🚀 Models</a>
-        <a href="http://discuss.huggingface.co">💬 Forum</a>
    </div>
    `;
    document.body.insertAdjacentHTML('afterbegin', div);
@@ -190,8 +72,6 @@ function platformToggle() {
    const codeBlocks = Array.from(document.getElementsByClassName("highlight"));
    const pytorchIdentifier = "## PYTORCH CODE";
    const tensorflowIdentifier = "## TENSORFLOW CODE";
-
-    const promptSpanIdentifier = `<span class="gp">&gt;&gt;&gt; </span>`
    const pytorchSpanIdentifier = `<span class="c1">${pytorchIdentifier}</span>`;
    const tensorflowSpanIdentifier = `<span class="c1">${tensorflowIdentifier}</span>`;

@@ -204,22 +84,10 @@ function platformToggle() {
        let tensorflowSpans;

        if(pytorchSpanPosition < tensorflowSpanPosition){
-            const isPrompt = spans.slice(
-                spans.indexOf(tensorflowSpanIdentifier) - promptSpanIdentifier.length,
-                spans.indexOf(tensorflowSpanIdentifier)
-            ) == promptSpanIdentifier;
-            const finalTensorflowSpanPosition = isPrompt ? tensorflowSpanPosition - promptSpanIdentifier.length : tensorflowSpanPosition;
-
-            pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, finalTensorflowSpanPosition);
+            pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, tensorflowSpanPosition);
            tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, spans.length);
        }else{
-            const isPrompt = spans.slice(
-                spans.indexOf(pytorchSpanIdentifier) - promptSpanIdentifier.length,
-                spans.indexOf(pytorchSpanIdentifier)
-            ) == promptSpanIdentifier;
-            const finalPytorchSpanPosition = isPrompt ? pytorchSpanPosition - promptSpanIdentifier.length : pytorchSpanPosition;
-
-            tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, finalPytorchSpanPosition);
+            tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, pytorchSpanPosition);
            pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, spans.length);
        }

@@ -281,12 +149,10 @@ function parseGithubButtons (){"use strict";var e=window.document,t=e.location,o

 function onLoad() {
    addIcon();
-    addVersionControl();
    addCustomFooter();
    addGithubButton();
    parseGithubButtons();
    addHfMenu();
-    addColabLink();
    platformToggle();
 }

@@ -0,0 +1,54 @@
+# Benchmarks
+
+This section is dedicated to the Benchmarks done by the library, both by maintainers, contributors and users. These 
+benchmark will help keep track of the preformance improvements that are brought to our models across versions.
+
+## Benchmarking all models for inference
+
+As of version 2.1 we have benchmarked all models for inference, across many different settings: using PyTorch, with
+and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
+TensorFlow XLA) and GPUs.
+
+The approach is detailed in the [following blogpost](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
+
+The results are available [here](https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing).
+
+## TF2 with mixed precision, XLA, Distribution (@tlkh)
+
+This work was done by [Timothy Liu](https://github.com/tlkh).
+
+There are very positive results to be gained from the various TensorFlow 2.0 features:
+
+- Automatic Mixed Precision (AMP)
+- XLA compiler
+- Distribution strategies (multi-GPU)
+
+The benefits are listed here (tested on CoLA, MRPC, SST-2):
+
+- AMP: Between 1.4x to 1.6x decrease in overall time without change in batch size
+- AMP+XLA: Up to 2.5x decrease in overall time on SST-2 (larger dataset)
+- Distribution: Between 1.4x to 3.4x decrease in overall time on 4xV100
+- Combined: Up to 5.7x decrease in overall training time, or 9.1x training throughput
+
+The model quality (measured by the validation accuracy) fluctuates slightly. Taking an average of 4 training runs 
+on a single GPU gives the following results:
+
+- CoLA: AMP results in slighter lower acc (0.820 vs 0.824)
+- MRPC: AMP results in lower acc (0.823 vs 0.835)
+- SST-2: AMP results in slighter lower acc (0.918 vs 0.922)
+
+However, in a distributed setting with 4xV100 (4x batch size), AMP can yield in better results:
+
+CoLA: AMP results in higher acc (0.828 vs 0.812)
+MRPC: AMP results in lower acc (0.817 vs 0.827)
+SST-2: AMP results in slightly lower acc (0.926 vs 0.929)
+
+The benchmark script is available [here](https://github.com/NVAITC/benchmarking/blob/master/tf2/bert_dist.py).
+
+Note: on some tasks (e.g. MRPC), the dataset is too small. The overhead due to the model compilation with XLA as well
+as the distribution strategy setup does not speed things up. The XLA compile time is also the reason why although throughput 
+can increase a lot (e.g. 2.7x for single GPU), overall (end-to-end) training speed-up is not as fast (as low as 1.4x)
+
+The benefits as seen on SST-2 (larger dataset) is much clear.
+
+All results can be seen on this [Google Sheet](https://docs.google.com/spreadsheets/d/1538MN224EzjbRL239sqSiUy6YY-rAjHyXhTzz_Zptls/edit#gid=960868445).
@@ -1,322 +0,0 @@
-Benchmarks
-==========
-
-Let's take a look at how 🤗 Transformer models can be benchmarked, best practices, and already available benchmarks.
-
-A notebook explaining in more detail how to benchmark 🤗 Transformer models can be found `here <https://github.com/huggingface/transformers/blob/master/notebooks/05-benchmark.ipynb>`__.
-
-How to benchmark 🤗 Transformer models
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-The classes :class:`~transformers.PyTorchBenchmark` and :class:`~transformers.TensorFlowBenchmark` allow to flexibly benchmark 🤗 Transformer models.
-The benchmark classes allow us to measure the `peak memory usage` and `required time` for both 
-`inference` and `training`. 
-
-.. note::
-
-  Hereby, `inference` is defined by a single forward pass, and `training` is defined by a single forward pass and backward pass.
-
-The benchmark classes :class:`~transformers.PyTorchBenchmark` and :class:`~transformers.TensorFlowBenchmark` expect an object of type :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments`, respectively, for instantiation. :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments` are data classes and contain all relevant configurations for their corresponding benchmark class.
-In the following example, it is shown how a BERT model of type `bert-base-cased` can be benchmarked.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import PyTorchBenchmark, PyTorchBenchmarkArguments
-
-    >>> args = PyTorchBenchmarkArguments(models=["bert-base-uncased"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
-    >>> benchmark = PyTorchBenchmark(args)
-
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import TensorFlowBenchmark, TensorFlowBenchmarkArguments
-
-    >>> args = TensorFlowBenchmarkArguments(models=["bert-base-uncased"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
-    >>> benchmark = TensorFlowBenchmark(args)
-
-
-Here, three arguments are given to the benchmark argument data classes, namely ``models``, ``batch_sizes``, and ``sequence_lengths``. The argument ``models`` is required and expects a :obj:`list` of model identifiers from the `model hub <https://huggingface.co/models>`__
-The :obj:`list` arguments ``batch_sizes`` and ``sequence_lengths`` define the size of the ``input_ids`` on which the model is benchmarked. 
-There are many more parameters that can be configured via the benchmark argument data classes. For more detail on these one can either directly consult the files 
-``src/transformers/benchmark/benchmark_args_utils.py``, ``src/transformers/benchmark/benchmark_args.py`` (for PyTorch) and ``src/transformers/benchmark/benchmark_args_tf.py`` (for Tensorflow). 
-Alternatively, running the following shell commands from root will print out a descriptive list of all configurable parameters for PyTorch and Tensorflow respectively.
-
-.. code-block:: bash
-
-    ## PYTORCH CODE
-    python examples/benchmarking/run_benchmark.py --help
-
-    ## TENSORFLOW CODE
-    python examples/benchmarking/run_benchmark_tf.py --help
-
-
-An instantiated benchmark object can then simply be run by calling ``benchmark.run()``.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> results = benchmark.run()
-    >>> print(results)
-    ====================       INFERENCE - SPEED - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length     Time in s                  
-    --------------------------------------------------------------------------------
-    bert-base-uncased          8               8             0.006     
-    bert-base-uncased          8               32            0.006     
-    bert-base-uncased          8              128            0.018     
-    bert-base-uncased          8              512            0.088     
-    --------------------------------------------------------------------------------
-    
-    ====================      INFERENCE - MEMORY - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length    Memory in MB 
-    --------------------------------------------------------------------------------
-    bert-base-uncased          8               8             1227
-    bert-base-uncased          8               32            1281
-    bert-base-uncased          8              128            1307
-    bert-base-uncased          8              512            1539
-    --------------------------------------------------------------------------------
-    
-    ====================        ENVIRONMENT INFORMATION         ====================
-    - transformers_version: 2.11.0
-    - framework: PyTorch
-    - use_torchscript: False
-    - framework_version: 1.4.0
-    - python_version: 3.6.10
-    - system: Linux
-    - cpu: x86_64
-    - architecture: 64bit
-    - date: 2020-06-29
-    - time: 08:58:43.371351
-    - fp16: False
-    - use_multiprocessing: True
-    - only_pretrain_model: False
-    - cpu_ram_mb: 32088
-    - use_gpu: True
-    - num_gpus: 1
-    - gpu: TITAN RTX
-    - gpu_ram_mb: 24217
-    - gpu_power_watts: 280.0
-    - gpu_performance_state: 2
-    - use_tpu: False
-    
-    >>> ## TENSORFLOW CODE
-    >>> results = benchmark.run()
-    >>> print(results)
-    ====================       INFERENCE - SPEED - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length     Time in s                  
-    --------------------------------------------------------------------------------
-    bert-base-uncased          8               8             0.005
-    bert-base-uncased          8               32            0.008
-    bert-base-uncased          8              128            0.022
-    bert-base-uncased          8              512            0.105
-    --------------------------------------------------------------------------------
-    
-    ====================      INFERENCE - MEMORY - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length    Memory in MB 
-    --------------------------------------------------------------------------------
-    bert-base-uncased          8               8             1330
-    bert-base-uncased          8               32            1330
-    bert-base-uncased          8              128            1330
-    bert-base-uncased          8              512            1770
-    --------------------------------------------------------------------------------
-    
-    ====================        ENVIRONMENT INFORMATION         ====================
-    - transformers_version: 2.11.0
-    - framework: Tensorflow
-    - use_xla: False
-    - framework_version: 2.2.0
-    - python_version: 3.6.10
-    - system: Linux
-    - cpu: x86_64
-    - architecture: 64bit
-    - date: 2020-06-29
-    - time: 09:26:35.617317
-    - fp16: False
-    - use_multiprocessing: True
-    - only_pretrain_model: False
-    - cpu_ram_mb: 32088
-    - use_gpu: True
-    - num_gpus: 1
-    - gpu: TITAN RTX
-    - gpu_ram_mb: 24217
-    - gpu_power_watts: 280.0
-    - gpu_performance_state: 2
-    - use_tpu: False
-
-By default, the `time` and the `required memory` for `inference` are benchmarked. 
-In the example output above the first two sections show the result corresponding to `inference time` and `inference memory`. 
-In addition, all relevant information about the computing environment, `e.g.` the GPU type, the system, the library versions, etc... are printed out in the third section under `ENVIRONMENT INFORMATION`.
-This information can optionally be saved in a `.csv` file when adding the argument :obj:`save_to_csv=True` to :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments` respectively.
-In this case, every section is saved in a separate `.csv` file. The path to each `.csv` file can optionally be defined via the argument data classes.
-
-Instead of benchmarking pre-trained models via their model identifier, `e.g.` `bert-base-uncased`, the user can alternatively benchmark an arbitrary configuration of any available model class. 
-In this case, a :obj:`list` of configurations must be inserted with the benchmark args as follows.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import PyTorchBenchmark, PyTorchBenchmarkArguments, BertConfig
-
-    >>> args = PyTorchBenchmarkArguments(models=["bert-base", "bert-384-hid", "bert-6-lay"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
-    >>> config_base = BertConfig()
-    >>> config_384_hid = BertConfig(hidden_size=384)
-    >>> config_6_lay = BertConfig(num_hidden_layers=6)
-
-    >>> benchmark = PyTorchBenchmark(args, configs=[config_base, config_384_hid, config_6_lay])
-    >>> benchmark.run()
-    ====================       INFERENCE - SPEED - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length       Time in s                  
-    --------------------------------------------------------------------------------
-    bert-base                  8              128            0.006
-    bert-base                  8              512            0.006
-    bert-base                  8              128            0.018     
-    bert-base                  8              512            0.088     
-    bert-384-hid              8               8             0.006     
-    bert-384-hid              8               32            0.006     
-    bert-384-hid              8              128            0.011     
-    bert-384-hid              8              512            0.054     
-    bert-6-lay                 8               8             0.003     
-    bert-6-lay                 8               32            0.004     
-    bert-6-lay                 8              128            0.009     
-    bert-6-lay                 8              512            0.044
-    --------------------------------------------------------------------------------
-    
-    ====================      INFERENCE - MEMORY - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length      Memory in MB 
-    --------------------------------------------------------------------------------
-    bert-base                  8               8             1277
-    bert-base                  8               32            1281
-    bert-base                  8              128            1307     
-    bert-base                  8              512            1539     
-    bert-384-hid              8               8             1005     
-    bert-384-hid              8               32            1027     
-    bert-384-hid              8              128            1035     
-    bert-384-hid              8              512            1255     
-    bert-6-lay                 8               8             1097     
-    bert-6-lay                 8               32            1101     
-    bert-6-lay                 8              128            1127     
-    bert-6-lay                 8              512            1359
-    --------------------------------------------------------------------------------
-    
-    ====================        ENVIRONMENT INFORMATION         ====================
-    - transformers_version: 2.11.0
-    - framework: PyTorch
-    - use_torchscript: False
-    - framework_version: 1.4.0
-    - python_version: 3.6.10
-    - system: Linux
-    - cpu: x86_64
-    - architecture: 64bit
-    - date: 2020-06-29
-    - time: 09:35:25.143267
-    - fp16: False
-    - use_multiprocessing: True
-    - only_pretrain_model: False
-    - cpu_ram_mb: 32088
-    - use_gpu: True
-    - num_gpus: 1
-    - gpu: TITAN RTX
-    - gpu_ram_mb: 24217
-    - gpu_power_watts: 280.0
-    - gpu_performance_state: 2
-    - use_tpu: False
-
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import TensorFlowBenchmark, TensorFlowBenchmarkArguments, BertConfig
-
-    >>> args = TensorFlowBenchmarkArguments(models=["bert-base", "bert-384-hid", "bert-6-lay"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
-    >>> config_base = BertConfig()
-    >>> config_384_hid = BertConfig(hidden_size=384)
-    >>> config_6_lay = BertConfig(num_hidden_layers=6)
-
-    >>> benchmark = TensorFlowBenchmark(args, configs=[config_base, config_384_hid, config_6_lay])
-    >>> benchmark.run()
-    ====================       INFERENCE - SPEED - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length       Time in s                  
-    --------------------------------------------------------------------------------
-    bert-base                  8               8             0.005
-    bert-base                  8               32            0.008
-    bert-base                  8              128            0.022
-    bert-base                  8              512            0.106
-    bert-384-hid              8               8             0.005
-    bert-384-hid              8               32            0.007
-    bert-384-hid              8              128            0.018
-    bert-384-hid              8              512            0.064
-    bert-6-lay                 8               8             0.002
-    bert-6-lay                 8               32            0.003
-    bert-6-lay                 8              128            0.0011
-    bert-6-lay                 8              512            0.074
-    --------------------------------------------------------------------------------
-    
-    ====================      INFERENCE - MEMORY - RESULT       ====================
-    --------------------------------------------------------------------------------
-    Model Name             Batch Size     Seq Length      Memory in MB 
-    --------------------------------------------------------------------------------
-    bert-base                  8               8             1330
-    bert-base                  8               32            1330
-    bert-base                  8              128            1330
-    bert-base                  8              512            1770
-    bert-384-hid              8               8             1330
-    bert-384-hid              8               32            1330
-    bert-384-hid              8              128            1330
-    bert-384-hid              8              512            1540
-    bert-6-lay                 8               8             1330
-    bert-6-lay                 8               32            1330
-    bert-6-lay                 8              128            1330
-    bert-6-lay                 8              512            1540
-    --------------------------------------------------------------------------------
-    
-    ====================        ENVIRONMENT INFORMATION         ====================
-    - transformers_version: 2.11.0
-    - framework: Tensorflow
-    - use_xla: False
-    - framework_version: 2.2.0
-    - python_version: 3.6.10
-    - system: Linux
-    - cpu: x86_64
-    - architecture: 64bit
-    - date: 2020-06-29
-    - time: 09:38:15.487125
-    - fp16: False
-    - use_multiprocessing: True
-    - only_pretrain_model: False
-    - cpu_ram_mb: 32088
-    - use_gpu: True
-    - num_gpus: 1
-    - gpu: TITAN RTX
-    - gpu_ram_mb: 24217
-    - gpu_power_watts: 280.0
-    - gpu_performance_state: 2
-    - use_tpu: False
-
-
-Again, `inference time` and `required memory` for `inference` are measured, but this time for customized configurations of the :obj:`BertModel` class. This feature can especially be helpful when 
-deciding for which configuration the model should be trained.
-
-
-Benchmark best practices
-~~~~~~~~~~~~~~~~~~~~~~~~
-
-This section lists a couple of best practices one should be aware of when benchmarking a model.
-
- Currently, only single device benchmarking is supported. When benchmarking on GPU, it is recommended that the user 
-  specifies on which device the code should be run by setting the ``CUDA_VISIBLE_DEVICES`` environment variable in the shell, `e.g.` ``export CUDA_VISIBLE_DEVICES=0`` before running the code.
- The option :obj:`no_multi_processing` should only be set to :obj:`True` for testing and debugging. To ensure accurate memory measurement it is recommended to run each memory benchmark in a separate process by making sure :obj:`no_multi_processing` is set to :obj:`True`.
- One should always state the environment information when sharing the results of a model benchmark. Results can vary heavily between different GPU devices, library versions, etc., so that benchmark results on their own are not very useful for the community.
-
-
-Sharing your benchmark
-~~~~~~~~~~~~~~~~~~~~~~
-
-Previously all available core models (10 at the time) have been benchmarked for `inference time`, across many different settings: using PyTorch, with
-and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
-TensorFlow XLA) and GPUs.
-
-The approach is detailed in the `following blogpost <https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2>`__ and the results are available `here <https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing>`__.
-
-With the new `benchmark` tools, it is easier than ever to share your benchmark results with the community `here <https://github.com/huggingface/transformers/blob/master/examples/benchmarking/README.md>`__.
@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'3.0.2'
+release = u'2.11.0'


 # -- General configuration ---------------------------------------------------
@@ -44,8 +44,7 @@ extensions = [
    'sphinx.ext.napoleon',
    'recommonmark',
    'sphinx.ext.viewcode',
-    'sphinx_markdown_tables',
-    'sphinx_copybutton'
+    'sphinx_markdown_tables'
 ]

 # Add any paths that contain templates here, relative to this directory.
@@ -75,8 +74,6 @@ exclude_patterns = [u'_build', 'Thumbs.db', '.DS_Store']
 # The name of the Pygments (syntax highlighting) style to use.
 pygments_style = None

-# Remove the prompt when copying examples
-copybutton_prompt_text = ">>> "

 # -- Options for HTML output -------------------------------------------------

@@ -190,8 +187,8 @@ epub_title = project
 epub_exclude_files = ['search.html']

 def setup(app):
-    app.add_css_file('css/huggingface.css')
-    app.add_css_file('css/code-snippets.css')
+    app.add_stylesheet('css/huggingface.css')
+    app.add_stylesheet('css/code-snippets.css')
    app.add_js_file('js/custom.js')

 # -- Extension configuration -------------------------------------------------
@@ -11,7 +11,7 @@ General terms
  tokens at a certain timestep.
 - MLM: masked language modeling, a pretraining task where the model sees a corrupted version of the texts, usually done
  by masking some tokens randomly, and has to predict the original text.
- multimodal: a task that combines texts with another kind of inputs (for instance images).
+- multimodal: a task taht combines texts with another kind of inputs (for instance images).
 - NLG: natural language generation, all tasks related to generating text ( for instance talk with transformers,
  translation)
 - NLP: natural language processing, a generic way to say "deal with texts".
@@ -45,16 +45,17 @@ tokenizer, which is a `WordPiece <https://arxiv.org/pdf/1609.08144.pdf>`__ token

 ::

-    >>> from transformers import BertTokenizer
-    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")

-    >>> sequence = "A Titan RTX has 24GB of VRAM"
+    sequence = "A Titan RTX has 24GB of VRAM"

 The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.

 ::

-    >>> tokenized_sequence = tokenizer.tokenize(sequence)
+    tokenized_sequence = tokenizer.tokenize(sequence)
+    print(tokenized_sequence)

 The tokens are either words or subwords. Here for instance, "VRAM" wasn't in the model vocabulary, so it's been split
 in "V", "RA" and "M". To indicate those tokens are not separate words but parts of the same word, a double-dash is
@@ -62,7 +63,6 @@ added for "RA" and "M":

 ::

-    >>> print(tokenized_sequence)
    ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']

 These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding
@@ -71,14 +71,14 @@ the sentence to the tokenizer, which leverages the Rust implementation of

 ::

-    >>> encoded_sequence = tokenizer(sequence)["input_ids"]
+    encoded_sequence = tokenizer(sequence)["input_ids"]
+    print(encoded_sequence)

 The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
 token indices are under the key "input_ids":

 ::

-    >>> print(encoded_sequence)
    [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]

 Note that the tokenizer automatically adds "special tokens" (if the associated model rely on them) which are special
@@ -86,14 +86,13 @@ IDs the model sometimes uses. If we decode the previous sequence of ids,

 ::

-    >>> decoded_sequence = tokenizer.decode(encoded_sequence)
+    tokenizer.decode(encoded_sequence)

 we will see 

 ::

-    >>> print(decoded_sequence)
-    [CLS] A Titan RTX has 24GB of VRAM [SEP]
+    '[CLS] A Titan RTX has 24GB of VRAM [SEP]'

 because this is the way a :class:`~transformers.BertModel` is going to expect its inputs.

@@ -109,20 +108,21 @@ For example, consider these two sequences:

 ::

-    >>> from transformers import BertTokenizer
-    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")

-    >>> sequence_a = "This is a short sequence."
-    >>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
+    sequence_a = "This is a short sequence."
+    sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."

-    >>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
-    >>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
+    encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
+    encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
+    
+    len(encoded_sequence_a), len(encoded_sequence_b)

 The encoded versions have different lengths:

 ::

-    >>> len(encoded_sequence_a), len(encoded_sequence_b)
    (8, 19)

 Therefore, we can't be put then together in a same tensor as-is. The first sequence needs to be padded up to the length
@@ -133,14 +133,15 @@ it to pad like this:

 ::

-    >>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
+    padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
+    padded_sequences["input_ids"]

 We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:

 ::

-    >>> padded_sequences["input_ids"]
-    [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
+    [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]

 This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating
 the position of the padded indices so that the model does not attend to them. For the
@@ -149,8 +150,14 @@ a padded value. This attention mask is in the dictionary returned by the tokeniz

 ::

-    >>> padded_sequences["attention_mask"]
-    [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
+    padded_sequences["attention_mask"]
+
+will give back
+
+::
+
+    [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+     [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]

 .. _token-type-ids:

@@ -163,27 +170,26 @@ tokens. For example, the BERT model builds its two sequence input as such:

 ::

-   >>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
+   # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]

 We can use our tokenizer to automatically generate such a sentence by passing the two sequences as two arguments (and
 not a list like before) like this:

 ::

-    >>> from transformers import BertTokenizer
-    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
-    >>> sequence_a = "HuggingFace is based in NYC"
-    >>> sequence_b = "Where is HuggingFace based?"
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    sequence_a = "HuggingFace is based in NYC"
+    sequence_b = "Where is HuggingFace based?"

-    >>> encoded_dict = tokenizer(sequence_a, sequence_b)
-    >>> decoded = tokenizer.decode(encoded_dict["input_ids"])
+    encoded_dict = tokenizer(sequence_a, sequence_b)
+    tokenizer.decode(encoded_dict["input_ids"])

 which will return:

 ::

-    >>> print(decoded)
-    [CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
+    "[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]"

 This is enough for some models to understand where one sequence ends and where another begins. However, other models
 such as BERT have an additional mechanism, which are the token type IDs (also called segment IDs). They are a binary
@@ -193,7 +199,12 @@ The tokenizer returns in the dictionary under the key "token_type_ids":

 ::

-    >>> encoded_dict['token_type_ids']
+    encoded_dict['token_type_ids']
+
+will return
+
+::
+
    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]

 The first sequence, the "context" used for the question, has all its tokens represented by :obj:`0`, whereas the
@@ -42,7 +42,7 @@ The documentation is organized in five parts:

 - **GET STARTED** contains a quick tour, the installation instructions and some useful information about our philosophy
  and a glossary.
- **USING 🤗 TRANSFORMERS** contains general tutorials on how to use the library.
+- **USING TRANSFORMERS** contains general tutorials on how to use the library.
 - **ADVANCED GUIDES** contains more advanced guides that are more specific to a given script or part of the library.
 - **RESEARCH** focuses on tutorials that have less to do with how to use the library but more about general resarch in
  transformers model
@@ -121,14 +121,7 @@ conversion utilities for the following models:
    trained using `OPUS <http://opus.nlpl.eu/>`_ pretrained_models data by Jörg Tiedemann.
 21. `Longformer <https://github.com/allenai/longformer>`_ (from AllenAI) released with the paper `Longformer: The
    Long-Document Transformer <https://arxiv.org/abs/2004.05150>`_ by Iz Beltagy, Matthew E. Peters, and Arman Cohan.
-22. `DPR <https://github.com/facebookresearch/DPR>`_ (from Facebook) released with the paper `Dense Passage Retrieval
-    for Open-Domain Question Answering <https://arxiv.org/abs/2004.04906>`_ by Vladimir Karpukhin, Barlas Oğuz, Sewon
-    Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
-23. `Pegasus <https://github.com/google-research/pegasus>`_ (from Google) released with the paper `PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization
-    <https://arxiv.org/abs/1912.08777>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
-24. `MBart <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`_ (from Facebook) released with the paper  `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov
-    Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.  
-25. `Other community models <https://huggingface.co/models>`_, contributed by the `community
+22. `Other community models <https://huggingface.co/models>`_, contributed by the `community
    <https://huggingface.co/users>`_.

 .. toctree::
@@ -142,14 +135,12 @@ conversion utilities for the following models:

 .. toctree::
    :maxdepth: 2
-    :caption: Using 🤗 Transformers
+    :caption: Using Transformers

    task_summary
    model_summary
-    preprocessing
-    training
+    serialization
    model_sharing
-    tokenizer_summary
    multilingual

 .. toctree::
@@ -161,15 +152,14 @@ conversion utilities for the following models:
    notebooks
    converting_tensorflow_models
    migration
+    torchscript
    contributing
-    serialization

 .. toctree::
    :maxdepth: 2
    :caption: Research

    bertology
-    perplexity
    benchmarks

 .. toctree::
@@ -177,11 +167,9 @@ conversion utilities for the following models:
    :caption: Package Reference

    main_classes/configuration
-    main_classes/output
    main_classes/model
    main_classes/tokenizer
    main_classes/pipelines
-    main_classes/trainer
    main_classes/optimizer_schedules
    main_classes/processors
    model_doc/auto
@@ -208,9 +196,3 @@ conversion utilities for the following models:
    model_doc/longformer
    model_doc/retribert
    model_doc/mobilebert
-    model_doc/dpr
-    model_doc/pegasus
-    model_doc/mbart
-    internal/modeling_utils
-    internal/tokenization_utils
-    internal/pipelines_utils
@@ -1,88 +0,0 @@
-Custom Layers and Utilities
---------------------------
-
-This page lists all the custom layers used by the library, as well as the utility functions it provides for modeling.
-
-Most of those are only useful if you are studying the code of the models in the library.
-
-
-``Pytorch custom modules``
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_utils.Conv1D
-
-.. autoclass:: transformers.modeling_utils.PoolerStartLogits
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.PoolerEndLogits
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.PoolerAnswerClass
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.SquadHeadOutput
-
-.. autoclass:: transformers.modeling_utils.SQuADHead
-    :members: forward
-
-.. autoclass:: transformers.modeling_utils.SequenceSummary
-    :members: forward
-
-
-``PyTorch Helper Functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autofunction:: transformers.apply_chunking_to_forward
-
-.. autofunction:: transformers.modeling_utils.find_pruneable_heads_and_indices
-
-.. autofunction:: transformers.modeling_utils.prune_layer
-
-.. autofunction:: transformers.modeling_utils.prune_conv1d_layer
-
-.. autofunction:: transformers.modeling_utils.prune_linear_layer
-
-``TensorFlow custom layers``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFConv1D
-
-.. autoclass:: transformers.modeling_tf_utils.TFSharedEmbeddings
-    :members: call
-
-.. autoclass:: transformers.modeling_tf_utils.TFSequenceSummary
-    :members: call
-
-
-``TensorFlow loss functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFCausalLanguageModelingLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFMaskedLanguageModelingLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFMultipleChoiceLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFQuestionAnsweringLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFSequenceClassificationLoss
-    :members:
-
-.. autoclass:: transformers.modeling_tf_utils.TFTokenClassificationLoss
-    :members:
-
-
-``TensorFlow Helper Functions``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autofunction:: transformers.modeling_tf_utils.cast_bool_to_primitive
-
-.. autofunction:: transformers.modeling_tf_utils.get_initializer
-
-.. autofunction:: transformers.modeling_tf_utils.keras_serializable
-
-.. autofunction:: transformers.modeling_tf_utils.shape_list
@@ -1,40 +0,0 @@
-Utilities for pipelines
-----------------------
-
-This page lists all the utility functions the library provides for pipelines.
-
-Most of those are only useful if you are studying the code of the models in the library.
-
-
-Argument handling
-~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.pipelines.ArgumentHandler
-
-.. autoclass:: transformers.pipelines.ZeroShotClassificationArgumentHandler
-
-.. autoclass:: transformers.pipelines.QuestionAnsweringArgumentHandler
-
-
-Data format
-~~~~~~~~~~~
-
-.. autoclass:: transformers.pipelines.PipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.CsvPipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.JsonPipelineDataFormat
-    :members:
-
-.. autoclass:: transformers.pipelines.PipedPipelineDataFormat
-    :members:
-
-
-Utilities
-~~~~~~~~~
-
-.. autofunction:: transformers.pipelines.get_framework
-
-.. autoclass:: transformers.pipelines.PipelineException
@@ -1,38 +0,0 @@
-Utilities for Tokenizers
------------------------
-
-This page lists all the utility functions used by the tokenizers, mainly the class
-:class:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase` that implements the common methods between
-:class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast` and the mixin
-:class:`~transformers.tokenization_utils_base.SpecialTokensMixin`.
-
-Most of those are only useful if you are studying the code of the tokenizers in the library.
-
-``PreTrainedTokenizerBase``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.tokenization_utils_base.PreTrainedTokenizerBase
-    :special-members: __call__
-    :members:
-
-
-``SpecialTokensMixin``
-~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.tokenization_utils_base.SpecialTokensMixin
-    :members:
-
-
-Enums and namedtuples
-~~~~~~~~~~~~~~~~~~~~~
-.. autoclass:: transformers.tokenization_utils_base.ExplicitEnum
-
-.. autoclass:: transformers.tokenization_utils_base.PaddingStrategy
-
-.. autoclass:: transformers.tokenization_utils_base.TensorType
-
-.. autoclass:: transformers.tokenization_utils_base.TruncationStrategy
-
-.. autoclass:: transformers.tokenization_utils_base.CharSpan
-
-.. autoclass:: transformers.tokenization_utils_base.TokenSpan
@@ -1,9 +1,7 @@
 Configuration
 ----------------------------------------------------

-The base class ``PretrainedConfig`` implements the common methods for loading/saving a configuration either from a
-local file or directory, or from a pretrained model configuration provided by the library (downloaded from
-HuggingFace's AWS S3 repository).
+The base class ``PretrainedConfig`` implements the common methods for loading/saving a configuration either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).

 ``PretrainedConfig``
 ~~~~~~~~~~~~~~~~~~~~~
@@ -1,32 +1,23 @@
 Models
 ----------------------------------------------------

-The base classes :class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` implement the
-common methods for loading/saving a model either from a local file or directory, or from a pretrained model
-configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).
+The base class ``PreTrainedModel`` implements the common methods for loading/saving a model either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).

-:class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` also implement a few methods which
-are common among all the models to:
+``PreTrainedModel`` also implements a few methods which are common among all the models to:

 - resize the input token embeddings when new tokens are added to the vocabulary
 - prune the attention heads of the model.

-The other methods that are common to each model are defined in :class:`~transformers.modeling_utils.ModuleUtilsMixin`
-(for the PyTorch models) and :class:`~transformers.modeling_tf_utils.TFModuleUtilsMixin` (for the TensorFlow models).
-
-
 ``PreTrainedModel``
 ~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.PreTrainedModel
    :members:

+``Helper Functions``
+~~~~~~~~~~~~~~~~~~~~~

-``ModuleUtilsMixin``
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_utils.ModuleUtilsMixin
-    :members:
+.. autofunction:: transformers.apply_chunking_to_forward


 ``TFPreTrainedModel``
@@ -34,16 +25,3 @@ The other methods that are common to each model are defined in :class:`~transfor

 .. autoclass:: transformers.TFPreTrainedModel
    :members:
-
-
-``TFModelUtilsMixin``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_tf_utils.TFModelUtilsMixin
-    :members:
-
-
-Generative models
-~~~~~~~~~~~~~~~~~
-
-Coming soon
@@ -1,4 +1,4 @@
-Optimization
+Optimizer
 ----------------------------------------------------

 The ``.optimization`` module provides:
@@ -7,25 +7,24 @@ The ``.optimization`` module provides:
 - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
 - a gradient accumulation class to accumulate the gradients of multiple batches

-``AdamW`` (PyTorch)
-~~~~~~~~~~~~~~~~~~~
+``AdamW``
+~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.AdamW
    :members:

-``AdamWeightDecay`` (TensorFlow)
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+``AdamWeightDecay``
+~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.AdamWeightDecay

 .. autofunction:: transformers.create_optimizer

 Schedules
-~~~~~~~~~~~~~~~~~~~
-
-Learning Rate Schedules (Pytorch)
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+----------------------------------------------------

+Learning Rate Schedules
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 .. autofunction:: transformers.get_constant_schedule


@@ -57,16 +56,16 @@ Learning Rate Schedules (Pytorch)
    :target: /imgs/warmup_linear_schedule.png
    :alt:

-``Warmup`` (TensorFlow)
-^^^^^^^^^^^^^^^^^^^^^^^
+``Warmup``
+~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.WarmUp
    :members:

 Gradient Strategies
-~~~~~~~~~~~~~~~~~~~~
+----------------------------------------------------

-``GradientAccumulator`` (TensorFlow)
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+``GradientAccumulator``
+~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.GradientAccumulator
@@ -1,141 +0,0 @@
-Model outputs
-------------
-
-PyTorch models have outputs that are instances of subclasses of :class:`~transformers.file_utils.ModelOutput`. Those
-are data structures containing all the information returned by the model, but that can also be used as tuples or
-dictionaries.
-
-Let's see of this looks on an example:
-
-.. code-block::
-
-    from transformers import BertTokenizer, BertForSequenceClassification
-    import torch
-
-    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-    model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
-
-    inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
-    labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-    outputs = model(**inputs, labels=labels)
-
-The ``outputs`` object is a :class:`~transformers.modeling_outputs.SequenceClassifierOutput`, as we can see in the
-documentation of that class below, it means it has an optional ``loss``, a ``logits`` an optional ``hidden_states`` and
-an optional ``attentions`` attribute. Here we have the ``loss`` since we passed along ``labels``, but we don't have
-``hidden_states`` and ``attentions`` because we didn't pass ``output_hidden_states=True`` or
-``output_attentions=True``.
-
-You can access each attribute as you would usually do, and if that attribute has not been returned by the model, you
-will get ``None``. Here for instance ``outputs.loss`` is the loss computed by the model, and ``outputs.attentions`` is
-``None``.
-
-When considering our ``outputs`` object as tuple, it only considers the attributes that don't have ``None`` values.
-Here for instance, it has two elements, ``loss`` then ``logits``, so
-
-.. code-block::
-
-    outputs[:2]
-
-will return the tuple ``(outputs.loss, outputs.logits)`` for instance.
-
-When considering our ``outputs`` object as dictionary, it only considers the attributes that don't have ``None``
-values. Here for instance, it has two keys that are ``loss`` and ``logits``.
-
-We document here the generic model outputs that are used by more than one model type. Specific output types are
-documented on their corresponding model page.
-
-``ModelOutput``
-~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.file_utils.ModelOutput
-    :members:
-
-``BaseModelOutput``
-~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.BaseModelOutput
-    :members:
-
-``BaseModelOutputWithPooling``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.BaseModelOutputWithPooling
-    :members:
-
-``BaseModelOutputWithPast``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.BaseModelOutputWithPast
-    :members:
-
-``Seq2SeqModelOutput``
-~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.Seq2SeqModelOutput
-    :members:
-
-``CausalLMOutput``
-~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.CausalLMOutput
-    :members:
-
-``CausalLMOutputWithPast``
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.CausalLMOutputWithPast
-    :members:
-
-``MaskedLMOutput``
-~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.MaskedLMOutput
-    :members:
-
-``Seq2SeqLMOutput``
-~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.Seq2SeqLMOutput
-    :members:
-
-``NextSentencePredictorOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.NextSentencePredictorOutput
-    :members:
-
-``SequenceClassifierOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.SequenceClassifierOutput
-    :members:
-
-``Seq2SeqSequenceClassifierOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.Seq2SeqSequenceClassifierOutput
-    :members:
-
-``MultipleChoiceModelOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.MultipleChoiceModelOutput
-    :members:
-
-``TokenClassifierOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.TokenClassifierOutput
-    :members:
-
-``QuestionAnsweringModelOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.QuestionAnsweringModelOutput
-    :members:
-
-``Seq2SeqQuestionAnsweringModelOutput``
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.Seq2SeqQuestionAnsweringModelOutput
-    :members:
@@ -3,24 +3,13 @@ Pipelines

 The pipelines are a great and easy way to use models for inference. These pipelines are objects that abstract most
 of the complex code from the library, offering a simple API dedicated to several tasks, including Named Entity
-Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering. See the
-:doc:`task summary <../task_summary>` for examples of use.
+Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering.

 There are two categories of pipeline abstractions to be aware about:

- The :func:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines.
- The other task-specific pipelines:
-
-    - :class:`~transformers.ConversationalPipeline`
-    - :class:`~transformers.FeatureExtractionPipeline`
-    - :class:`~transformers.FillMaskPipeline`
-    - :class:`~transformers.QuestionAnsweringPipeline`
-    - :class:`~transformers.SummarizationPipeline`
-    - :class:`~transformers.TextClassificationPipeline`
-    - :class:`~transformers.TextGenerationPipeline`
-    - :class:`~transformers.TokenClassificationPipeline`
-    - :class:`~transformers.TranslationPipeline`
-    - :class:`~transformers.ZeroShotClassificationPipeline`
+- The :func:`~transformers.pipeline` which is the most powerful object encapsulating all other pipelines
+- The other task-specific pipelines, such as :class:`~transformers.TokenClassificationPipeline`
+  or :class:`~transformers.QuestionAnsweringPipeline`

 The pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -32,82 +21,53 @@ other pipeline but requires an additional argument which is the `task`.


 The task specific pipelines
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

-ConversationalPipeline
-==========================================
+Parent class: Pipeline
+=========================================

-.. autoclass:: transformers.Conversation
-
-.. autoclass:: transformers.ConversationalPipeline
-    :special-members: __call__
-    :members:
-
-FeatureExtractionPipeline
-==========================================
-
-.. autoclass:: transformers.FeatureExtractionPipeline
-    :special-members: __call__
-    :members:
-
-FillMaskPipeline
-==========================================
-
-.. autoclass:: transformers.FillMaskPipeline
-    :special-members: __call__
-    :members:
-
-NerPipeline
-==========================================
-
-This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined below. Please refer to that
-pipeline for documentation and usage examples.
-
-QuestionAnsweringPipeline
-==========================================
-
-.. autoclass:: transformers.QuestionAnsweringPipeline
-    :special-members: __call__
-    :members:
-
-SummarizationPipeline
-==========================================
-
-.. autoclass:: transformers.SummarizationPipeline
-    :special-members: __call__
-    :members:
-
-TextClassificationPipeline
-==========================================
-
-.. autoclass:: transformers.TextClassificationPipeline
-    :special-members: __call__
-    :members:
-
-TextGenerationPipeline
-==========================================
-
-.. autoclass:: transformers.TextGenerationPipeline
-    :special-members: __call__
-    :members:
+.. autoclass:: transformers.Pipeline
+    :members: predict, transform, save_pretrained

 TokenClassificationPipeline
 ==========================================

 .. autoclass:: transformers.TokenClassificationPipeline
-    :special-members: __call__
-    :members:

-ZeroShotClassificationPipeline
+NerPipeline
 ==========================================

-.. autoclass:: transformers.ZeroShotClassificationPipeline
-    :special-members: __call__
-    :members:
+This class is an alias of the :class:`~transformers.TokenClassificationPipeline` defined above. Please refer to that pipeline for
+documentation and usage examples.
+
+FillMaskPipeline
+==========================================
+
+.. autoclass:: transformers.FillMaskPipeline
+
+FeatureExtractionPipeline
+==========================================
+
+.. autoclass:: transformers.FeatureExtractionPipeline
+
+TextClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.TextClassificationPipeline
+
+QuestionAnsweringPipeline
+==========================================
+
+.. autoclass:: transformers.QuestionAnsweringPipeline


-Parent class: :obj:`Pipeline`
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+SummarizationPipeline
+==========================================

-.. autoclass:: transformers.Pipeline
-    :members:
+.. autoclass:: transformers.SummarizationPipeline
+
+
+TextGenerationPipeline
+==========================================
+
+.. autoclass:: transformers.TextGenerationPipeline
@@ -1,40 +1,17 @@
 Tokenizer
 ----------------------------------------------------

-A tokenizer is in charge of preparing the inputs for a model. The library contains tokenizers for all the models. Most
-of the tokenizers are available in two flavors: a full python implementation and a "Fast" implementation based on the
-Rust library `tokenizers <https://github.com/huggingface/tokenizers>`__. The "Fast" implementations allows:
+A tokenizer is in charge of preparing the inputs for a model. The library comprise tokenizers for all the models. Most of the tokenizers are available in two flavors: a full python implementation and a "Fast" implementation based on the Rust library `tokenizers`. The "Fast" implementations allows (1) a significant speed-up in particular when doing batched tokenization and (2) additional methods to map between the original string (character and words) and the token space (e.g. getting the index of the token comprising a given character or the span of characters corresponding to a given token). Currently no "Fast" implementation is available for the SentencePiece-based tokenizers (for T5, ALBERT, CamemBERT, XLMRoBERTa and XLNet models).

-1. a significant speed-up in particular when doing batched tokenization and
-2. additional methods to map between the original string (character and words) and the token space (e.g. getting the
-   index of the token comprising a given character or the span of characters corresponding to a given token). Currently
-   no "Fast" implementation is available for the SentencePiece-based tokenizers (for T5, ALBERT, CamemBERT, XLMRoBERTa
-   and XLNet models).
+The base classes ``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` implements the common methods for encoding string inputs in model inputs (see below) and instantiating/saving python and "Fast" tokenizers either from a local file or directory or from a pretrained tokenizer provided by the library (downloaded from HuggingFace's AWS S3 repository).

-The base classes :class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast`
-implement the common methods for encoding string inputs in model inputs (see below) and instantiating/saving python and
-"Fast" tokenizers either from a local file or directory or from a pretrained tokenizer provided by the library
-(downloaded from HuggingFace's AWS S3 repository). They both rely on
-:class:`~transformers.tokenization_utils_base.PreTrainedTokenizerBase` that contains the common methods, and
-:class:`~transformers.tokenization_utils_base.SpecialTokensMixin`.
+``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` thus implements the main methods for using all the tokenizers:

-:class:`~transformers.PreTrainedTokenizer` and :class:`~transformers.PreTrainedTokenizerFast` thus implement the main
-methods for using all the tokenizers:
-
- Tokenizing (splitting strings in sub-word token strings), converting tokens strings to ids and back, and
-  encoding/decoding (i.e., tokenizing and converting to integers).
- Adding new tokens to the vocabulary in a way that is independent of the underlying structure (BPE, SentencePiece...).
- Managing special tokens (like mask, beginning-of-sentence, etc.): adding them, assigning them to attributes in the
-  tokenizer for easy access and making sure they are not split during tokenization.
-
-:class:`~transformers.BatchEncoding` holds the output of the tokenizer's encoding methods (``__call__``,
-``encode_plus`` and ``batch_encode_plus``) and is derived from a Python dictionary. When the tokenizer is a pure python
-tokenizer, this class behaves just like a standard python dictionary and holds the various model inputs computed by these
-methods (``input_ids``, ``attention_mask``...). When the tokenizer is a "Fast" tokenizer (i.e., backed by HuggingFace
-`tokenizers library <https://github.com/huggingface/tokenizers>`__), this class provides in addition several advanced
-alignment methods which can be used to map between the original string (character and words) and the token space (e.g.,
-getting the index of the token comprising a given character or the span of characters corresponding to a given token).
+- tokenizing (spliting strings in sub-word token strings), converting tokens strings to ids and back, and encoding/decoding (i.e. tokenizing + convert to integers),
+- adding new tokens to the vocabulary in a way that is independant of the underlying structure (BPE, SentencePiece...),
+- managing special tokens like mask, beginning-of-sentence, etc tokens (adding them, assigning them to attributes in the tokenizer for easy access and making sure they are not split during tokenization)

+``BatchEncoding`` holds the output of the tokenizer's encoding methods (``encode_plus`` and ``batch_encode_plus``) and is derived from a Python dictionary. When the tokenizer is a pure python tokenizer, this class behave just like a standard python dictionary and hold the various model inputs computed by these methodes (``input_ids``, ``attention_mask``...). When the tokenizer is a "Fast" tokenizer (i.e. backed by HuggingFace tokenizers library), this class provides in addition several advanced alignement methods which can be used to map between the original string (character and words) and the token space (e.g. getting the index of the token comprising a given character or the span of characters corresponding to a given token).

 ``PreTrainedTokenizer``
 ~~~~~~~~~~~~~~~~~~~~~~~~
@@ -43,7 +20,6 @@ getting the index of the token comprising a given character or the span of chara
    :special-members: __call__
    :members:

-
 ``PreTrainedTokenizerFast``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -51,9 +27,14 @@ getting the index of the token comprising a given character or the span of chara
    :special-members: __call__
    :members:

-
 ``BatchEncoding``
 ~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.BatchEncoding
    :members:
+
+``SpecialTokensMixin``
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.SpecialTokensMixin
+    :members:
@@ -1,62 +0,0 @@
-Trainer
----------
-
-The :class:`~transformers.Trainer` and :class:`~transformers.TFTrainer` classes provide an API for feature-complete
-training in most standard use cases. It's used in most of the :doc:`example scripts <../examples>`.
-
-Before instantiating your :class:`~transformers.Trainer`/:class:`~transformers.TFTrainer`, create a 
-:class:`~transformers.TrainingArguments`/:class:`~transformers.TFTrainingArguments` to access all the points of
-customization during training.
-
-The API supports distributed training on multiple GPUs/TPUs, mixed precision through `NVIDIA Apex
-<https://github.com/NVIDIA/apex>`__ for PyTorch and :obj:`tf.keras.mixed_precision` for TensorFlow.
-
-Both :class:`~transformers.Trainer` and :class:`~transformers.TFTrainer` contain the basic training loop supporting the
-previous features. To inject custom behavior you can subclass them and override the following methods:
-
- **get_train_dataloader**/**get_train_tfdataset** -- Creates the training DataLoader (PyTorch) or TF Dataset.
- **get_eval_dataloader**/**get_eval_tfdataset** -- Creates the evaulation DataLoader (PyTorch) or TF Dataset.
- **get_test_dataloader**/**get_test_tfdataset** -- Creates the test DataLoader (PyTorch) or TF Dataset.
- **log** -- Logs information on the various objects watching training.
- **setup_wandb** -- Setups wandb (see `here <https://docs.wandb.com/huggingface>`__ for more information).
- **create_optimizer_and_scheduler** -- Setups the optimizer and learning rate scheduler if they were not passed at
-  init.
- **training_step** -- Performs a training step.
- **prediction_step** -- Performs an evaluation/test step.
- **run_model** (TensorFlow only) -- Basic pass through the model.
- **evaluate** -- Runs an evaluation loop and returns metrics.
- **predict** -- Returns predictions (with metrics if labels are available) on a test set.
-
-
-``Trainer`` 
-~~~~~~~~~~~
-
-.. autoclass:: transformers.Trainer
-    :members:
-
-``TFTrainer`` 
-~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFTrainer
-    :members:
-
-``TrainingArguments``
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TrainingArguments
-    :members:
-
-``TFTrainingArguments``
-~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFTrainingArguments
-    :members:
-
-Utilities
-~~~~~~~~~
-
-.. autoclass:: transformers.EvalPrediction
-
-.. autofunction:: transformers.set_seed
-
-.. autofunction:: transformers.torch_distributed_zero_first
@@ -1,8 +1,8 @@
 # Migrating from previous packages

-## Migrating from pytorch-transformers to 🤗 Transformers
+## Migrating from pytorch-transformers to transformers

-Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to 🤗 Transformers.
+Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.

 ### Positional order of some models' keywords inputs (`attention_mask`, `token_type_ids`...) changed

@@ -14,17 +14,17 @@ If you used to call the models with positional inputs for keyword arguments, e.g

 ## Migrating from pytorch-pretrained-bert

-Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to 🤗 Transformers
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `transformers`

 ### Models always output `tuples`

-The main breaking change when migrating from `pytorch-pretrained-bert` to 🤗 Transformers is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.

 The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/transformers/).

 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.

-Here is a `pytorch-pretrained-bert` to 🤗 Transformers conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to `transformers` conversion example for a `BertForSequenceClassification` classification model:

 ```python
 # Let's load our model
@@ -33,11 +33,11 @@ model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
 # If you used to have this line in pytorch-pretrained-bert:
 loss = model(input_ids, labels=labels)

-# Now just use this line in 🤗 Transformers to extract the loss from the output tuple:
+# Now just use this line in transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]

-# In 🤗 Transformers you can also have access to the logits:
+# In transformers you can also have access to the logits:
 loss, logits = outputs[:2]

 # And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
@@ -109,7 +109,7 @@ for batch in train_data:
    loss.backward()
    optimizer.step()

-### In 🤗 Transformers, optimizer and schedules are splitted and instantiated like this:
+### In Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
 scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:
@@ -47,16 +47,6 @@ AlbertTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


-Albert specific outputs
-~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_albert.AlbertForPreTrainingOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_albert.TFAlbertForPreTrainingOutput
-    :members:
-
-
 AlbertModel
 ~~~~~~~~~~~~~~~~~~~~

@@ -64,13 +54,6 @@ AlbertModel
    :members:


-AlbertForPreTraining
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.AlbertForPreTraining
-    :members:
-
-
 AlbertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -113,13 +96,6 @@ TFAlbertModel
    :members:


-TFAlbertForPreTraining
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFAlbertForPreTraining
-    :members:
-
-
 TFAlbertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -17,23 +17,13 @@ According to the abstract,
 The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/bart>`_


-Implementation Notes
-~~~~~~~~~~~~~~~~~~~~
+Implementation Notes:

 - Bart doesn't use :obj:`token_type_ids` for sequence classification. Use BartTokenizer.encode to get the proper splitting.
 - The forward pass of ``BartModel`` will create decoder inputs (using the helper function ``transformers.modeling_bart._prepare_bart_decoder_inputs``)  if they are not passed. This is different than some other modeling APIs.
 - Model predictions are intended to be identical to the original implementation. This only works, however, if the string you pass to ``fairseq.encode`` starts with a space.
 - ``BartForConditionalGeneration.generate`` should be used for conditional generation tasks like summarization, see the example in that docstrings
 - Models that load the ``"facebook/bart-large-cnn"`` weights will not have a ``mask_token_id``, or be able to perform mask filling tasks.
- for training/forward passes that don't involve beam search, pass ``use_cache=False``
-
-
-BartForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.BartForConditionalGeneration
-    :members: generate, forward
-

 BartConfig
 ~~~~~~~~~~~~~~~~~~~~~
@@ -49,7 +39,6 @@ BartTokenizer
    :members:


-
 BartModel
 ~~~~~~~~~~~~~

@@ -73,3 +62,10 @@ BartForQuestionAnswering
    :members: forward


+BartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartForConditionalGeneration
+    :members: generate, forward
+
+
@@ -27,8 +27,13 @@ Tips:

 - BERT is a model with absolute position embeddings so it's usually advised to pad the inputs on
  the right rather than the left.
- BERT was trained with the masked language modeling (MLM) and next sentence prediction (NSP) objectives. It is efficient at predicting masked
-  tokens and at NLU in general, but is not optimal for text generation.
+- BERT was trained with a 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.
+- Alongside MLM, BERT was trained using a next sentence prediction (NSP) objective using the [CLS] token as a sequence
+  approximate. The user may use this token (the first token in a sequence built with special tokens) to get a sequence
+  prediction rather than a token prediction. However, averaging over the sequence may yield better results than using
+  the [CLS] token.

 The original code can be found `here <https://github.com/google-research/bert>`_.

@@ -54,16 +59,6 @@ BertTokenizerFast
    :members:


-Bert specific outputs
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_bert.BertForPreTrainingOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_bert.TFBertForPreTrainingOutput
-    :members:
-
-
 BertModel
 ~~~~~~~~~~~~~~~~~~~~

@@ -78,13 +73,6 @@ BertForPreTraining
    :members:


-BertModelLMHeadModel
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.BertLMHeadModel
-    :members:
-
-
 BertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -141,13 +129,6 @@ TFBertForPreTraining
    :members:


-TFBertModelLMHeadModel
-~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFBertLMHeadModel
-    :members:
-
-
 TFBertForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -105,13 +105,6 @@ TFCamembertForSequenceClassification
    :members:


-TFCamembertForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFCamembertForMultipleChoice
-    :members:
-
-
 TFCamembertForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -1,102 +0,0 @@
-DPR
----------------------------------------------------
-
-Overview
-~~~~~~~~~~~~~~~~~~~~~
-
-Dense Passage Retrieval (DPR) - is a set of tools and models for state-of-the-art open-domain Q&A research.
-It is based on the following paper:
-
-Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih, Dense Passage Retrieval for Open-Domain Question Answering.
-
-The abstract from the paper is the following:
-
-*Open-domain question answering relies on efficient passage retrieval to select candidate contexts, where traditional
-sparse vector space models, such as TF-IDF or BM25, are the de facto method. In this work, we show that retrieval can
-be practically implemented using dense representations alone, where embeddings are learned from a small number of
-questions and passages by a simple dual-encoder framework. When evaluated on a wide range of open-domain QA datasets,
-our dense retriever outperforms a strong Lucene-BM25 system largely by 9%-19% absolute in terms of top-20 passage
-retrieval accuracy, and helps our end-to-end QA system establish new state-of-the-art on multiple open-domain QA
-benchmarks.*
-
-The original code can be found `here <https://github.com/facebookresearch/DPR>`_.
-
-
-DPRConfig
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRConfig
-    :members:
-
-
-DPRContextEncoderTokenizer
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRContextEncoderTokenizer
-    :members:
-
-
-DPRContextEncoderTokenizerFast
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRContextEncoderTokenizerFast
-    :members:
-
-DPRQuestionEncoderTokenizer
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRQuestionEncoderTokenizer
-    :members:
-
-
-DPRQuestionEncoderTokenizerFast
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRQuestionEncoderTokenizerFast
-    :members:
-
-DPRReaderTokenizer
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRReaderTokenizer
-    :members:
-
-
-DPRReaderTokenizerFast
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRReaderTokenizerFast
-    :members:
-
-
-DPR specific outputs
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_dpr.DPRContextEncoderOutput
-    :members:
-
-.. autoclass:: transformers.modeling_dpr.DPRQuestionEncoderOutput
-    :members:
-
-.. autoclass:: transformers.modeling_dpr.DPRReaderOutput
-    :members:
-
-
-DPRContextEncoder
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRContextEncoder
-    :members:
-
-DPRQuestionEncoder
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRQuestionEncoder
-    :members:
-
-
-DPRReader
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.DPRReader
-    :members:
@@ -71,16 +71,6 @@ ElectraTokenizerFast
    :members:


-Electra specific outputs
-~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_electra.ElectraForPreTrainingOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_electra.TFElectraForPreTrainingOutput
-    :members:
-
-
 ElectraModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -109,13 +99,6 @@ ElectraForSequenceClassification
    :members:


-ElectraForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.ElectraForMultipleChoice
-    :members:
-
-
 ElectraForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -151,20 +134,6 @@ TFElectraForMaskedLM
    :members:


-TFElectraForSequenceClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFElectraForSequenceClassification
-    :members:
-
-
-TFElectraForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFElectraForMultipleChoice
-    :members:
-
-
 TFElectraForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -61,20 +61,6 @@ FlaubertForSequenceClassification
    :members:


-FlaubertForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.FlaubertForMultipleChoice
-    :members:
-
-
-FlaubertForTokenClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.FlaubertForTokenClassification
-    :members:
-
-
 FlaubertForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -128,4 +114,4 @@ TFFlaubertForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.TFFlaubertForQuestionAnsweringSimple
-    :members:
+    :members:
@@ -71,16 +71,6 @@ OpenAIGPTTokenizerFast
    :members:


-OpenAI specific outputs
-~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_openai.OpenAIGPTDoubleHeadsModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_openai.TFOpenAIGPTDoubleHeadsModelOutput
-    :members:
-
-
 OpenAIGPTModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -58,16 +58,6 @@ GPT2TokenizerFast
    :members:


-GPT2 specific outputs
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_gpt2.GPT2DoubleHeadsModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_gpt2.TFGPT2DoubleHeadsModelOutput
-    :members:
-
-
 GPT2Model
 ~~~~~~~~~~~~~~~~~~~~~

@@ -16,7 +16,7 @@ Longformer Self Attention
 ~~~~~~~~~~~~~~~~~~~~~~~~~~
 Longformer self attention employs self attention on both a "local" context and a "global" context.
 Most tokens only attend "locally" to each other meaning that each token attends to its :math:`\frac{1}{2} w` previous tokens and :math:`\frac{1}{2} w` succeding tokens with :math:`w` being the window length as defined in `config.attention_window`. Note that `config.attention_window` can be of type ``list`` to define a different :math:`w` for each layer. 
-A selected few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.
+A selecetd few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.

 Note that "locally" and "globally" attending tokens are projected by different query, key and value matrices.
 Also note that every "locally" attending token not only attends to tokens within its window :math:`w`, but also to all "globally" attending tokens so that global attention is *symmetric*.
@@ -102,25 +102,3 @@ LongformerForQuestionAnswering

 .. autoclass:: transformers.LongformerForQuestionAnswering
    :members:
-
-
-TFLongformerModel
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerModel
-    :members:
-
-
-TFLongformerForMaskedLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerForMaskedLM
-    :members:
-
-
-TFLongformerForQuestionAnswering
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.TFLongformerForQuestionAnswering
-    :members:
-
@@ -48,7 +48,7 @@ Example of translating english to many romance languages, using language codes:
    tokenizer = MarianTokenizer.from_pretrained(model_name)
    print(tokenizer.supported_language_codes)
    model = MarianMTModel.from_pretrained(model_name)
-    translated = model.generate(**tokenizer.prepare_seq2seq_batch(src_text))
+    translated = model.generate(**tokenizer.prepare_translation_batch(src_text))
    tgt_text = [tokenizer.decode(t, skip_special_tokens=True) for t in translated]
    # ["c'est une phrase en anglais que nous voulons traduire en français",
    # 'Isto deve ir para o português.',
@@ -86,14 +86,6 @@ Code to see available pretrained models:
    suffix = [x.split('/')[1] for x in model_ids]
    multi_models = [f'{org}/{s}' for s in suffix if s != s.lower()]

-MarianMTModel
-~~~~~~~~~~~~~
-
-Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
-Model API is identical to BartForConditionalGeneration.
-Available models are listed at `Model List <https://huggingface.co/models?search=Helsinki-NLP>`__
-This class inherits nearly all functionality from ``BartForConditionalGeneration``, see that page for method signatures.
-
 MarianConfig
 ~~~~~~~~~~~~~~~~~~~
 .. autoclass:: transformers.MarianConfig
@@ -104,8 +96,16 @@ MarianTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.MarianTokenizer
-    :members: prepare_seq2seq_batch
+    :members: prepare_translation_batch


+MarianMTModel
+~~~~~~~~~~~~~

+Pytorch version of marian-nmt's transformer.h (c++). Designed for the OPUS-NMT translation checkpoints.
+Model API is identical to BartForConditionalGeneration.
+Available models are listed at `Model List <https://huggingface.co/models?search=Helsinki-NLP>`__
+This class inherits all functionality from ``BartForConditionalGeneration``, see that page for method signatures.

+.. autoclass:: transformers.MarianMTModel
+    :members:
@@ -1,37 +0,0 @@
-MBart
----------------------------------------------------
-**DISCLAIMER:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer
-
-Overview
-~~~~~~~~~~~~~~~~~~~~~
-The MBart model was presented in `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov
-Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer. According to the abstract,
-
-MBART is a sequence-to-sequence denoising auto-encoder pre-trained on large-scale monolingual corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only on the encoder, decoder, or reconstructing parts of the text.
-
-The Authors' code can be found `here <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`__
-
-
-MBartConfig
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MBartConfig
-    :members:
-
-
-MBartTokenizer
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MBartTokenizer
-    :members: build_inputs_with_special_tokens, prepare_seq2seq_batch
-
-
-MBartForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MBartForConditionalGeneration
-    :members: generate, forward
-
-
@@ -56,16 +56,6 @@ MobileBertTokenizerFast
    :members:


-MobileBert specific outputs
-~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_mobilebert.MobileBertForPreTrainingOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_mobilebert.TFMobileBertForPreTrainingOutput
-    :members:
-
-
 MobileBertModel
 ~~~~~~~~~~~~~~~~~~~~

@@ -1,115 +0,0 @@
-Pegasus
----------------------------------------------------
-**DISCLAIMER:** If you see something strange,
-file a `Github Issue <https://github.com/huggingface/transformers/issues/new?assignees=sshleifer&labels=&template=bug-report.md&title>`__ and assign
-@sshleifer.
-
-
-Overview
-~~~~~~~~~~~~~~~~~~~~~
-
-The Pegasus model was `proposed <https://arxiv.org/abs/1910.13461>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu on Dec 18, 2019.
-According to the abstract,
-
- Pegasus' pretraining task is intentionally similar to summarization: important sentences are removed/masked from an input document and are generated together as one output sequence from the remaining sentences, similar to an extractive summary.
- Pegasus achieves SOTA summarization performance on all 12 downstream tasks, as measured by ROUGE and human eval.
-
-The Authors' code can be found `here <https://github.com/google-research/pegasus>`_
-
-
-Checkpoints
-~~~~~~~~~~~
-The `checkpoints <https://huggingface.co/models?search=pegasus>`_ all checkpoints are finetuned for summarization, besides ``pegasus-large``, whence the other checkpoints are finetuned.
- Each checkpoint is 2.2 GB on disk and 568M parameters.
- FP16 is not supported (help/ideas on this appreciated!).
- Summarizing xsum in fp32 takes about 400ms/sample, with default parameters on a v100 GPU.
- For XSUM, The paper reports rouge1,rouge2, rougeL of paper: 47.21/24.56/39.25. As of Aug 9, this port scores 46.91/24.34/39.1.
-The gap is likely because of different alpha/length_penalty implementations in beam search.
-
-
-Implementation Notes
-~~~~~~~~~~~~~~~~~~~~
-
- All models are transformer encoder-decoders with 16 layers in each component.
- The implementation is completely inherited from ``BartForConditionalGeneration``
- Some key configuration differences:
-    - static, sinusoidal position embeddings
-    - no ``layernorm_embedding`` (``PegasusConfig.normalize_embedding=False``)
-    - the model starts generating with pad_token_id (which has 0 token_embedding) as the prefix.
-    - ``num_beams=8``
- All pretrained pegasus checkpoints are the same besides three attributes: ``tokenizer.model_max_length`` (max input size),  ``max_length`` (max num tokens to generate) and ``length_penalty``
- Code to convert checkpoints trained in the author's `repo <https://github.com/google-research/pegasus>`_ can be found in ``convert_pegasus_tf_to_pytorch.py``
-
-
-Usage Example
-~~~~~~~~~~~~~~~~~~~~
-
-.. code-block:: python
-
-    from transformers import PegasusForConditionalGeneration, PegasusTokenizer
-    src_text = [
-        """ PG&E stated it scheduled the blackouts in response to forecasts for high winds amid dry conditions. The aim is to reduce the risk of wildfires. Nearly 800 thousand customers were scheduled to be affected by the shutoffs which were expected to last through at least midday tomorrow."""
-    ]
-
-    model_name = 'google/pegasus-xsum'
-    torch_device = 'cuda' if torch.cuda.is_available() else 'cpu'
-    tokenizer = PegasusTokenizer.from_pretrained(model_name)
-    model = PegasusForConditionalGeneration.from_pretrained(model_name).to(torch_device)
-    batch = tokenizer.prepare_seq2seq_batch(src_text, truncation=True, padding='longest').to(torch_device)
-    translated = model.generate(**batch)
-    tgt_text = tokenizer.batch_decode(translated, skip_special_tokens=True)
-    assert tgt_text[0] == "California's largest electricity provider has turned off power to tens of thousands of customers."
-
-PegasusForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-This class inherits all functionality from ``BartForConditionalGeneration``, see that page for method signatures.
-Available models are listed at `Model List <https://huggingface.co/models?search=pegasus>`__
-
-.. autoclass:: transformers.PegasusForConditionalGeneration
-    :members:
-
-
-PegasusConfig
-~~~~~~~~~~~~~~~~~~~
-This config fully inherits from ``BartConfig``, but pegasus uses different default values:
-Up to date parameter values can be seen in `S3 <https://s3.amazonaws.com/models.huggingface.co/bert/google/pegasus-xsum/config.json>`_.
-As of Aug 10, 2020, they are:
-
-.. code-block:: python
-
-    dict(
-    vocab_size=96103,
-    max_position_embeddings=512,
-    d_model=1024,
-    encoder_ffn_dim=4096,
-    decoder_ffn_dim=4096,
-    encoder_attention_heads=16,
-    decoder_attention_heads=16,
-    encoder_layers=16,
-    decoder_layers=16,
-    dropout=0.1,
-    attention_dropout=0.1,
-    activation_dropout=0.1,
-    pad_token_id=0,
-    eos_token_id=1,
-    is_encoder_decoder=True,
-    normalize_before=True,
-    scale_embedding=True,
-    normalize_embedding=False,
-    add_final_layer_norm=True,
-    static_position_embeddings=True,
-    num_beams=8,
-    activation_function="relu",
-    )
-
-
-PegasusTokenizer
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-warning: ``add_tokens`` does not work at the moment.
-
-.. autoclass:: transformers.PegasusTokenizer
-    :members: __call__, prepare_seq2seq_batch
-
-
-
@@ -112,24 +112,3 @@ ReformerModelWithLMHead

 .. autoclass:: transformers.ReformerModelWithLMHead
    :members:
-
-
-ReformerForMaskedLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.ReformerForMaskedLM
-    :members:
-
-
-ReformerForSequenceClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.ReformerForSequenceClassification
-    :members:
-
-
-ReformerForQuestionAnswering
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.ReformerForQuestionAnswering
-    :members:
@@ -63,13 +63,6 @@ RobertaModel
    :members:


-RobertaForCausalLM
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.RobertaForCausalLM
-    :members:
-
-
 RobertaForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -38,13 +38,13 @@ T5 can be trained / fine-tuned both in a supervised and unsupervised fashion.

  In this setup spans of the input sequence are masked by so-called sentinel tokens (*a.k.a* unique mask tokens) 
  and the output sequence is formed as a concatenation of the same sentinel tokens and the *real* masked tokens. 
-  Each sentinel token represents a unique mask token for this sentence and should start with ``<extra_id_0>``, ``<extra_id_1>``, ... up to ``<extra_id_99>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
+  Each sentinel token represents a unique mask token for this sentence and should start with ``<extra_id_1>``, ``<extra_id_2>``, ... up to ``<extra_id_100>``. As a default 100 sentinel tokens are available in ``T5Tokenizer``.
  *E.g.* the sentence "The cute dog walks in the park" with the masks put on "cute dog" and "the" should be processed as follows: 

 ::

-  input_ids = tokenizer.encode('The <extra_id_0> walks in <extra_id_1> park', return_tensors='pt')
-  labels = tokenizer.encode('<extra_id_0> cute dog <extra_id_1> the <extra_id_2> </s>', return_tensors='pt')
+  input_ids = tokenizer.encode('The <extra_id_1> walks in <extra_id_2> park', return_tensors='pt')
+  labels = tokenizer.encode('<extra_id_1> cute dog <extra_id_2> the <extra_id_3> </s>', return_tensors='pt')
  # the forward function automatically creates the correct decoder_input_ids
  model(input_ids=input_ids, labels=labels)

@@ -54,22 +54,6 @@ TransfoXLTokenizerFast
    :members:


-TransfoXL specific outputs
-~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_transfo_xl.TransfoXLModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_transfo_xl.TransfoXLLMHeadModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_transfo_xl.TFTransfoXLModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_transfo_xl.TFTransfoXLLMHeadModelOutput
-    :members:
-
-
 TransfoXLModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -46,14 +46,6 @@ XLMTokenizer
    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
        create_token_type_ids_from_sequences, save_vocabulary

-
-XLM specific outputs
-~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_xlm.XLMForQuestionAnsweringOutput
-    :members:
-
-
 XLMModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -75,20 +67,6 @@ XLMForSequenceClassification
    :members:


-XLMForMultipleChoice
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.XLMForMultipleChoice
-    :members:
-
-
-XLMForTokenClassification
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.XLMForTokenClassification
-    :members:
-
-
 XLMForQuestionAnsweringSimple
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -50,49 +50,6 @@ XLNetTokenizer
        create_token_type_ids_from_sequences, save_vocabulary


-XLNet specific outputs
-~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_xlnet.XLNetModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetLMHeadModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetForSequenceClassificationOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetForMultipleChoiceOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetForTokenClassificationOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetForQuestionAnsweringSimpleOutput
-    :members:
-
-.. autoclass:: transformers.modeling_xlnet.XLNetForQuestionAnsweringOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetLMHeadModelOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForSequenceClassificationOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForMultipleChoiceOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForTokenClassificationOutput
-    :members:
-
-.. autoclass:: transformers.modeling_tf_xlnet.TFXLNetForQuestionAnsweringSimpleOutput
-    :members:
-
-
 XLNetModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -0,0 +1,55 @@
+# Model upload and sharing
+
+Starting with `v2.2.2`, you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
+
+**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Optionally, join an existing organization or create a new one. Then:
+
+```shell
+transformers-cli login
+# log in using the same credentials as on huggingface.co
+```
+Upload your model:
+```shell
+transformers-cli upload ./path/to/pretrained_model/
+
+# ^^ Upload folder containing weights/tokenizer/config
+# saved via `.save_pretrained()`
+
+transformers-cli upload ./config.json [--filename folder/foobar.json]
+
+# ^^ Upload a single file
+# (you can optionally override its filename, which can be nested inside a folder)
+```
+
+If you want your model to be namespaced by your organization name rather than your username, add the following flag to any command:
+```shell
+--organization organization_name
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username (or organization name) and the folder name above:
+```python
+"username/pretrained_model"
+# or if an org:
+"organization_name/pretrained_model"
+```
+
+**Please add a README.md model card** to the repo under `model_cards/` with: model description, training params (dataset, preprocessing, hardware used, hyperparameters), evaluation results, intended uses & limitations, etc.
+
+Your model now has a page on huggingface.co/models 🔥
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("namespace/pretrained_model")
+model = AutoModel.from_pretrained("namespace/pretrained_model")
+```
+
+List all your files on S3:
+```shell
+transformers-cli s3 ls
+```
+
+You can also delete unneeded files:
+
+```shell
+transformers-cli s3 rm …
+```
@@ -1,217 +0,0 @@
-Model sharing and uploading
-===========================
-
-In this page, we will show you how to share a model you have trained or fine-tuned on new data with the community on
-the `model hub <https://huggingface.co/models>`__.
-
-.. note::
-
-    You will need to create an account on `huggingface.co <https://huggingface.co/join>`__ for this.
-
-    Optionally, you can join an existing organization or create a new one.
-
-Prepare your model for uploading
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-We have seen in the :doc:`training tutorial <training>`: how to fine-tune a model on a given task. You have probably
-done something similar on your task, either using the model directly in your own training loop or using the
-:class:`~.transformers.Trainer`/:class:`~.transformers.TFTrainer` class. Let's see how you can share the result on
-the `model hub <https://huggingface.co/models>`__.
-
-Basic steps
-^^^^^^^^^^^
-
-.. 
-    When #5258 is merged, we can remove the need to create the directory.
-
-First, pick a directory with the name you want your model to have on the model hub (its full name will then be
-`username/awesome-name-you-picked` or `organization/awesome-name-you-picked`) and create it with either
-
-::
-
-    mkdir path/to/awesome-name-you-picked
-
-or in python
-
-::
-
-    import os
-    os.makedirs("path/to/awesome-name-you-picked")
-
-then you can save your model and tokenizer with:
-
-::
-
-    model.save_pretrained("path/to/awesome-name-you-picked")
-    tokenizer.save_pretrained("path/to/awesome-name-you-picked")
-
-Or, if you're using the Trainer API
-
-::
-
-    trainer.save_model("path/to/awesome-name-you-picked")
-    tokenizer.save_pretrained("path/to/awesome-name-you-picked")
-
-Make your model work on all frameworks
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-.. 
-    TODO Sylvain: make this automatic during the upload
-
-You probably have your favorite framework, but so will other users! That's why it's best to upload your model with both
-PyTorch `and` TensorFlow checkpoints to make it easier to use (if you skip this step, users will still be able to load
-your model in another framework, but it will be slower, as it will have to be converted on the fly). Don't worry, it's super easy to do (and in a future version,
-it will all be automatic). You will need to install both PyTorch and TensorFlow for this step, but you don't need to
-worry about the GPU, so it should be very easy. Check the
-`TensorFlow installation page <https://www.tensorflow.org/install/pip#tensorflow-2.0-rc-is-available>`__ 
-and/or the `PyTorch installation page <https://pytorch.org/get-started/locally/#start-locally>`__ to see how.
-
-First check that your model class exists in the other framework, that is try to import the same model by either adding
-or removing TF. For instance, if you trained a :class:`~transformers.DistilBertForSequenceClassification`, try to
-type
-
-::
-
-    from transformers import TFDistilBertForSequenceClassification
-
-and if you trained a :class:`~transformers.TFDistilBertForSequenceClassification`, try to
-type
-
-::
-
-    from transformers import DistilBertForSequenceClassification
-
-This will give back an error if your model does not exist in the other framework (something that should be pretty rare
-since we're aiming for full parity between the two frameworks). In this case, skip this and go to the next step.
-
-Now, if you trained your model in PyTorch and have to create a TensorFlow version, adapt the following code to your
-model class:
-
-::
-
-    tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
-    tf_model.save_pretrained("path/to/awesome-name-you-picked")
-
-and if you trained your model in TensorFlow and have to create a PyTorch version, adapt the following code to your
-model class:
-
-::
-
-    pt_model = DistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_tf=True)
-    pt_model.save_pretrained("path/to/awesome-name-you-picked")
-
-That's all there is to it!
-
-Check the directory before uploading
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Make sure there are no garbage files in the directory you'll upload. It should only have:
-
- a `config.json` file, which saves the :doc:`configuration <main_classes/configuration>` of your model ;
- a `pytorch_model.bin` file, which is the PyTorch checkpoint (unless you can't have it for some reason) ;
- a `tf_model.h5` file, which is the TensorFlow checkpoint (unless you can't have it for some reason) ;
- a `special_tokens_map.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
- a `tokenizer_config.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save;
- a `vocab.txt`, which is the vocabulary of your tokenizer, part of your :doc:`tokenizer <main_classes/tokenizer>`
-  save;
- maybe a `added_tokens.json`, which is part of your :doc:`tokenizer <main_classes/tokenizer>` save.
-
-Other files can safely be deleted.
-
-Upload your model with the CLI
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Now go in a terminal and run the following command. It should be in the virtual enviromnent where you installed 🤗
-Transformers, since that command :obj:`transformers-cli` comes from the library.
-
-::
-
-    transformers-cli login
-
-Then log in using the same credentials as on huggingface.co. To upload your model, just type
-
-::
-
-    transformers-cli upload path/to/awesome-name-you-picked/
-
-This will upload the folder containing the weights, tokenizer and configuration we prepared in the previous section.
-
-If you want to upload a single file (a new version of your model, or the other framework checkpoint you want to add),
-just type:
-
-::
-
-    transformers-cli upload path/to/awesome-name-you-picked/that-file 
-
-or
-
-::
-
-   transformers-cli upload path/to/awesome-name-you-picked/that-file --filename awesome-name-you-picked/new_name
-
-if you want to change its filename.
-
-This uploads the model to your personal account. If you want your model to be namespaced by your organization name
-rather than your username, add the following flag to any command:
-
-::
-
-    --organization organization_name
-
-so for instance:
-
-::
-
-    transformers-cli upload path/to/awesome-name-you-picked/ --organization organization_name
-
-Your model will then be accessible through its identifier, which is, as we saw above,
-`username/awesome-name-you-picked` or `organization/awesome-name-you-picked`.
-
-Add a model card
-^^^^^^^^^^^^^^^^
-
-To make sure everyone knows what your model can do, what its limitations and potential bias or ethetical
-considerations, please add a README.md model card to the 🤗 Transformers repo under `model_cards/`. It should then be
-placed in a subfolder with your username or organization, then another subfolder named like your model
-(`awesome-name-you-picked`). Or just click on the "Create a model card on GitHub" button on the model page, it will
-get you directly to the right location. If you need one, `here <https://github.com/huggingface/model_card>`__ is a
-model card template (meta-suggestions are welcome).
-
-If your model is fine-tuned from another model coming from the model hub (all 🤗 Transformers pretrained models do),
-don't forget to link to its model card so that people can fully trace how your model was built.
-
-If you have never made a pull request to the 🤗 Transformers repo, look at the
-:doc:`contributing guide <contributing>` to see the steps to follow.
-
-.. Note::
-
-    You can also send your model card in the folder you uploaded with the CLI by placing it in a `README.md` file
-    inside `path/to/awesome-name-you-picked/`.
-
-Using your model
-^^^^^^^^^^^^^^^^
-
-Your model now has a page on huggingface.co/models 🔥
-
-Anyone can load it from code:
-
-::
-
-    tokenizer = AutoTokenizer.from_pretrained("namespace/awesome-name-you-picked")
-    model = AutoModel.from_pretrained("namespace/awesome-name-you-picked")
-
-Additional commands
-^^^^^^^^^^^^^^^^^^^
-
-You can list all the files you uploaded on the hub like this:
-
-::
-
-    transformers-cli s3 ls
-
-You can also delete unneeded files with
-
-::
-
-    transformers-cli s3 rm awesome-name-you-picked/filename
-
@@ -1,7 +1,7 @@
 Summary of the models
 ================================================

-This is a summary of the models available in 🤗 Transformers. It assumes you’re familiar with the original 
+This is a summary of the models available in the transformers library. It assumes you’re familiar with the original 
 `transformer model <https://arxiv.org/abs/1706.03762>`_. For a gentle introduction check the `annotated transformer 
 <http://nlp.seas.harvard.edu/2018/04/03/attention.html>`_. Here we focus on the high-level differences between the
 models. You can check them more in detail in their respective documentation. Also checkout the 
@@ -55,7 +55,7 @@ Original GPT
   <a href="https://huggingface.co/models?filter=openai-gpt">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-openai--gpt-blueviolet">
   </a>
-   <a href="model_doc/gpt.html">
+   <a href="/model_doc/gpt">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-openai--gpt-blueviolet">
   </a>

@@ -75,7 +75,7 @@ GPT-2
   <a href="https://huggingface.co/models?filter=gpt2">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-gpt2-blueviolet">
   </a>
-   <a href="model_doc/gpt2.html">
+   <a href="/model_doc/gpt2">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-gpt2-blueviolet">
   </a>

@@ -96,7 +96,7 @@ CTRL
   <a href="https://huggingface.co/models?filter=ctrl">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-ctrl-blueviolet">
   </a>
-   <a href="model_doc/ctrl.html">
+   <a href="/model_doc/ctrl">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-ctrl-blueviolet">
   </a>

@@ -117,7 +117,7 @@ Transformer-XL
   <a href="https://huggingface.co/models?filter=transfo-xl">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-transfo--xl-blueviolet">
   </a>
-   <a href="model_doc/transformerxl.html">
+   <a href="/model_doc/transformerxl">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-transfo--xl-blueviolet">
   </a>

@@ -148,7 +148,7 @@ Reformer
   <a href="https://huggingface.co/models?filter=reformer">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-reformer-blueviolet">
   </a>
-   <a href="model_doc/reformer.html">
+   <a href="/model_doc/reformer">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-reformer-blueviolet">
   </a>

@@ -183,7 +183,7 @@ XLNet
   <a href="https://huggingface.co/models?filter=xlnet">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-xlnet-blueviolet">
   </a>
-   <a href="model_doc/xlnet.html">
+   <a href="/model_doc/xlnet">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-xlnet-blueviolet">
   </a>

@@ -195,7 +195,7 @@ tokens in the sentence, then allows the model to use the last n tokens to predic
 with a mask, the sentence is actually fed in the model in the right order, but instead of masking the first n tokens 
 for n+1, XLNet uses a mask that hides the previous tokens in some given permutation of 1,...,sequence length.

-XLNet also uses the same recurrence mechanism as Transformer-XL to build long-term dependencies. 
+XLNet also uses the same recurrence mechanism as TransformerXL to build long-term dependencies. 

 The library provides a version of the model for language modeling, token classification, sentence classification, 
 multiple choice classification and question answering.
@@ -217,7 +217,7 @@ BERT
   <a href="https://huggingface.co/models?filter=bert">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-bert-blueviolet">
   </a>
-   <a href="model_doc/bert.html">
+   <a href="/model_doc/bert">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-bert-blueviolet">
   </a>

@@ -246,7 +246,7 @@ ALBERT
   <a href="https://huggingface.co/models?filter=albert">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-albert-blueviolet">
   </a>
-   <a href="model_doc/albert.html">
+   <a href="/model_doc/albert">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-albert-blueviolet">
   </a>

@@ -275,7 +275,7 @@ RoBERTa
   <a href="https://huggingface.co/models?filter=roberta">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-roberta-blueviolet">
   </a>
-   <a href="model_doc/roberta.html">
+   <a href="/model_doc/roberta">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-roberta-blueviolet">
   </a>

@@ -301,7 +301,7 @@ DistilBERT
   <a href="https://huggingface.co/models?filter=distilbert">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-distilbert-blueviolet">
   </a>
-   <a href="model_doc/distilbert.html">
+   <a href="/model_doc/distilbert">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-distilbert-blueviolet">
   </a>

@@ -326,7 +326,7 @@ XLM
   <a href="https://huggingface.co/models?filter=xlm">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-xlm-blueviolet">
   </a>
-   <a href="model_doc/xlm.html">
+   <a href="/model_doc/xlm">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-xlm-blueviolet">
   </a>

@@ -361,7 +361,7 @@ XLM-RoBERTa
   <a href="https://huggingface.co/models?filter=xlm-roberta">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-xlm--roberta-blueviolet">
   </a>
-   <a href="model_doc/xlmroberta.html">
+   <a href="/model_doc/xlmroberta">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-xlm--roberta-blueviolet">
   </a>

@@ -383,7 +383,7 @@ FlauBERT
   <a href="https://huggingface.co/models?filter=flaubert">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-flaubert-blueviolet">
   </a>
-   <a href="model_doc/flaubert.html">
+   <a href="/model_doc/flaubert">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-flaubert-blueviolet">
   </a>

@@ -401,7 +401,7 @@ ELECTRA
   <a href="https://huggingface.co/models?filter=electra">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-electra-blueviolet">
   </a>
-   <a href="model_doc/electra.html">
+   <a href="/model_doc/electra">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-electra-blueviolet">
   </a>

@@ -427,7 +427,7 @@ Longformer
   <a href="https://huggingface.co/models?filter=longformer">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-longformer-blueviolet">
   </a>
-   <a href="model_doc/longformer.html">
+   <a href="/model_doc/longformer">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-longformer-blueviolet">
   </a>

@@ -461,7 +461,7 @@ BART
   <a href="https://huggingface.co/models?filter=bart">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-bart-blueviolet">
   </a>
-   <a href="model_doc/bart.html">
+   <a href="/model_doc/bart">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-bart-blueviolet">
   </a>

@@ -488,7 +488,7 @@ MarianMT
   <a href="https://huggingface.co/models?filter=marian">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-marian-blueviolet">
   </a>
-   <a href="model_doc/marian.html">
+   <a href="/model_doc/marian">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-marian-blueviolet">
   </a>

@@ -506,7 +506,7 @@ T5
   <a href="https://huggingface.co/models?filter=t5">
       <img alt="Models" src="https://img.shields.io/badge/All_model_pages-t5-blueviolet">
   </a>
-   <a href="model_doc/t5.html">
+   <a href="/model_doc/t5">
       <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-t5-blueviolet">
   </a>

@@ -555,7 +555,7 @@ model know which part of the input vector corresponds to the text or the image.
 The pretrained model only works for classification.

 ..
-    More information in this :doc:`model documentation </model_doc/mmbt.html>`.
+    More information in this :doc:`model documentation </model_doc/mmbt>`.
    TODO: write this page

 More technical aspects
@@ -36,11 +36,10 @@ Here is an example using the ``xlm-clm-enfr-1024`` checkpoint (Causal language m

 .. code-block::

-    >>> import torch
-    >>> from transformers import XLMTokenizer, XLMWithLMHeadModel
+    import torch
+    from transformers import XLMTokenizer, XLMWithLMHeadModel

-    >>> tokenizer = XLMTokenizer.from_pretrained("xlm-clm-enfr-1024")
-    >>> model = XLMWithLMHeadModel.from_pretrained("xlm-clm-enfr-1024")
+    tokenizer = XLMTokenizer.from_pretrained("xlm-clm-1024-enfr")


 The different languages this model/tokenizer handles, as well as the ids of these languages are visible using the
@@ -48,15 +47,16 @@ The different languages this model/tokenizer handles, as well as the ids of thes

 .. code-block::

-    >>> print(tokenizer.lang2id)
-    {'en': 0, 'fr': 1}
+    # Continuation of the previous script
+    print(tokenizer.lang2id)  # {'en': 0, 'fr': 1}


 These ids should be used when passing a language parameter during a model pass. Let's define our inputs:

 .. code-block::

-    >>> input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1
+    # Continuation of the previous script
+    input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1


 We should now define the language embedding by using the previously defined language id. We want to create a tensor
@@ -64,18 +64,20 @@ filled with the appropriate language ids, of the same size as input_ids. For eng

 .. code-block::

-    >>> language_id = tokenizer.lang2id['en']  # 0
-    >>> langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])
+    # Continuation of the previous script
+    language_id = tokenizer.lang2id['en']  # 0
+    langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])

-    >>> # We reshape it to be of size (batch_size, sequence_length)
-    >>> langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)
+    # We reshape it to be of size (batch_size, sequence_length)
+    langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)


 You can then feed it all as input to your model:

 .. code-block::

-    >>> outputs = model(input_ids, langs=langs)
+    # Continuation of the previous script
+    outputs = model(input_ids, langs=langs)


 The example `run_generation.py <https://github.com/huggingface/transformers/blob/master/examples/text-generation/run_generation.py>`__
@@ -1,151 +0,0 @@
-Perplexity of fixed-length models
-=================================
-
-Perplexity (PPL) is one of the most common metrics for evaluating language
-models. Before diving in, we should note that the metric applies specifically
-to classical language models (sometimes called autoregressive or causal
-language models) and is not well defined for masked language models like BERT
-(see :doc:`summary of the models <model_summary>`).
-
-Perplexity is defined as the exponentiated average log-likelihood of a
-sequence. If we have a tokenized sequence :math:`X = (x_0, x_1, \dots, x_t)`,
-then the perplexity of :math:`X` is,
-
-.. math::
-
-    \text{PPL}(X)
-    = \exp \left\{ {-\frac{1}{t}\sum_i^t \log p_\theta (x_i|x_{<i}) } \right\}
-
-where :math:`\log p_\theta (x_i|x_{<i})` is the log-likelihood of the ith
-token conditioned on the preceding tokens :math:`x_{<i}` according to our
-model. Intuitively, it can be thought of as an evaluation of the model's
-ability to predict uniformly among the set of specified tokens in a corpus.
-Importantly, this means that the tokenization procedure has a direct impact
-on a model's perplexity which should always be taken into consideration when
-comparing different models.
-
-This is also equivalent to the exponentiation of the cross-entropy between
-the data and model predictions. For more intuition about perplexity and its
-relationship to Bits Per Character (BPC) and data compression, check out this
-`fantastic blog post on The Gradient
-<https://thegradient.pub/understanding-evaluation-metrics-for-language-models/>`_.
-
-Calculating PPL with fixed-length models
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-If we weren't limited by a model's context size, we would evaluate the
-model's perplexity by autoregressively factorizing a sequence and
-conditioning on the entire preceding subsequence at each step, as shown
-below.
-
-.. image:: imgs/ppl_full.gif
-    :width: 600
-    :alt: Full decomposition of a sequence with unlimited context length
-
-When working with approximate models, however, we typically have a constraint
-on the number of tokens the model can process. The largest version
-of :doc:`GPT-2 <model_doc/gpt2>`, for example, has a fixed length of 1024
-tokens, so we cannot calculate :math:`p_\theta(x_t|x_{<t})` directly when
-:math:`t` is greater than 1024.
-
-Instead, the sequence is typically broken into subsequences equal to the
-model's maximum input size. If a model's max input size is :math:`k`, we
-then approximate the likelihood of a token :math:`x_t` by conditioning only
-on the :math:`k-1` tokens that precede it rather than the entire context.
-When evaluating the model's perplexity of a sequence, a tempting but
-suboptimal approach is to break the sequence into disjoint chunks and
-add up the decomposed log-likelihoods of each segment independently.
-
-.. image:: imgs/ppl_chunked.gif
-    :width: 600
-    :alt: Suboptimal PPL not taking advantage of full available context
-
-This is quick to compute since the perplexity of each segment can be computed
-in one forward pass, but serves as a poor approximation of the
-fully-factorized perplexity and will typically yield a higher (worse) PPL
-because the model will have less context at most of the prediction steps.
-
-Instead, the PPL of fixed-length models should be evaluated with a
-sliding-window strategy. This involves repeatedly sliding the
-context window so that the model has more context when making each
-prediction.
-
-.. image:: imgs/ppl_sliding.gif
-    :width: 600
-    :alt: Sliding window PPL taking advantage of all available context
-
-This is a closer approximation to the true decomposition of the
-sequence probability and will typically yield a more favorable score.
-The downside is that it requires a separate forward pass for each token in
-the corpus. A good practical compromise is to employ a strided sliding
-window, moving the context by larger strides rather than sliding by 1 token a
-time. This allows computation to procede much faster while still giving the
-model a large context to make predictions at each step.
-
-Example: Calculating perplexity with GPT-2 in 🤗 Transformers
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Let's demonstrate this process with GPT-2.
-
-.. code-block:: python
-
-    from transformers import GPT2LMHeadModel, GPT2TokenizerFast
-    device = 'cuda'
-    model_id = 'gpt2-large'
-    model = GPT2LMHeadModel.from_pretrained(model_id).to(device)
-    tokenizer = GPT2TokenizerFast.from_pretrained(model_id)
-
-We'll load in the WikiText-2 dataset and evaluate the perplexity using a few
-different sliding-window strategies. Since this dataset is small and we're
-just doing one forward pass over the set, we can just load and encode the
-entire dataset in memory.
-
-.. code-block:: python
-
-    from nlp import load_dataset
-    test = load_dataset('wikitext', 'wikitext-2-raw-v1', split='test')
-    encodings = tokenizer('\n\n'.join(test['text']), return_tensors='pt')
-
-With 🤗 Transformers, we can simply pass the ``input_ids`` as the ``labels``
-to our model, and the average log-likelihood for each token is returned as
-the loss. With our sliding window approach, however, there is overlap in the
-tokens we pass to the model at each iteration. We don't want the
-log-likelihood for the tokens we're just treating as context to be included
-in our loss, so we can set these targets to ``-100`` so that they are
-ignored. The following is an example of how we could do this with a stride of
-``512``. This means that the model will have at least 512 tokens for context
-when calculating the conditional likelihood of any one token (provided there
-are 512 preceding tokens available to condition on).
-
-.. code-block:: python
-
-    max_length = model.config.n_positions
-    stride = 512
-
-    lls = []
-    for i in tqdm(range(0, encodings.input_ids.size(1), stride)):
-        begin_loc = max(i + stride - max_length, 0)
-        end_loc = i + stride
-        input_ids = encodings.input_ids[:,begin_loc:end_loc].to(device)
-        target_ids = input_ids.clone()
-        target_ids[:,:-stride] = -100
-
-        with torch.no_grad():
-            outputs = model(input_ids, labels=target_ids)
-            log_likelihood = outputs[0] * stride
-
-        lls.append(log_likelihood)
-    
-    ppl = torch.exp(torch.stack(lls).sum() / i)
-
-Running this with the stride length equal to the max input length is
-equivalent to the suboptimal, non-sliding-window strategy we discussed above.
-The smaller the stride, the more context the model will have in making each
-prediction, and the better the reported perplexity will typically be.
-
-When we run the above with ``stride = 1024``, i.e. no overlap, the resulting
-PPL is ``19.64``, which is about the same as the ``19.93`` reported in the
-GPT-2 paper. By using ``stride = 512`` and thereby employing our striding
-window strategy, this jumps down to ``16.53``. This is not only a more
-favorable score, but is calculated in a way that is closer to the true
-autoregressive decomposition of a sequence likelihood.
@@ -1,7 +1,7 @@
 Philosophy
 ==========

-🤗 Transformers is an opinionated library built for:
+Transformers is an opinionated library built for:

 - NLP researchers and educators seeking to use/study/extend large-scale transformers models
 - hands-on practitioners who want to fine-tune those models and/or serve them in production
@@ -1,338 +0,0 @@
-Preprocessing data
-==================
-
-In this tutorial, we'll explore how to preprocess your data using 🤗 Transformers. The main tool for this is what we
-
-call a :doc:`tokenizer <main_classes/tokenizer>`. You can build one using the tokenizer class associated to the model
-you would like to use, or directly with the :class:`~transformers.AutoTokenizer` class.
-
-As we saw in the :doc:`quicktour </quicktour>`, the tokenizer will first split a given text in words (or part of words,
-punctuation symbols, etc.) usually called `tokens`. Then it will convert those `tokens` into numbers, to be able to
-build a tensor out of them and feed them to the model. It will also add any additional inputs the model might expect to
-work properly.
-
-.. note::
-
-    If you plan on using a pretrained model, it's important to use the associated pretrained tokenizer: it will split
-    the text you give it in tokens the same way for the pretraining corpus, and it will use the same correspondence
-    token to index (that we usually call a `vocab`) as during pretraining.
-
-To automatically download the vocab used during pretraining or fine-tuning a given model, you can use the 
-:func:`~transformers.AutoTokenizer.from_pretrained` method:
-
-.. code-block::
-
-    from transformers import AutoTokenizer
-    tokenizer = AutoTokenizer.from_pretrained('bert-base-cased')
-
-Base use
-~~~~~~~~
-
-A :class:`~transformers.PreTrainedTokenizer` has many methods, but the only one you need to remember for preprocessing
-is its ``__call__``: you just need to feed your sentence to your tokenizer object.
-
-.. code-block::
-
-    >>> encoded_input = tokenizer("Hello, I'm a single sentence!")
-    >>> print(encoded_input)
-    {'input_ids': [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102], 
-     'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
-     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-This returns a dictionary string to list of ints.
-The `input_ids <glossary.html#input-ids>`__ are the indices corresponding to each token in our sentence. We will see
-below what the `attention_mask <glossary.html#attention-mask>`__ is used for and in
-:ref:`the next section <sentence-pairs>` the goal of `token_type_ids <glossary.html#token-type-ids>`__.
-
-The tokenizer can decode a list of token ids in a proper sentence:
-
-.. code-block::
-
-    >>> tokenizer.decode(encoded_input["input_ids"])
-    "[CLS] Hello, I'm a single sentence! [SEP]"
-
-As you can see, the tokenizer automatically added some special tokens that the model expect. Not all model need special
-tokens; for instance, if we had used` gtp2-medium` instead of `bert-base-cased` to create our tokenizer, we would have
-
-seen the same sentence as the original one here. You can disable this behavior (which is only advised if you have added
-those special tokens yourself) by passing ``add_special_tokens=False``.
-
-If you have several sentences you want to process, you can do this efficiently by sending them as a list to the
-tokenizer:
-
-.. code-block::
-
-    >>> batch_sentences = ["Hello I'm a single sentence",
-    ...                    "And another sentence",
-    ...                    "And the very very last one"]
-    >>> encoded_inputs = tokenizer(batch_sentences)
-    >>> print(encoded_inputs)
-    {'input_ids': [[101, 8667, 146, 112, 182, 170, 1423, 5650, 102],
-                   [101, 1262, 1330, 5650, 102],
-                   [101, 1262, 1103, 1304, 1304, 1314, 1141, 102]],
-     'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0],
-                        [0, 0, 0, 0, 0],
-                        [0, 0, 0, 0, 0, 0, 0, 0]],
-     'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1],
-                        [1, 1, 1, 1, 1],
-                        [1, 1, 1, 1, 1, 1, 1, 1]]}
-
-We get back a dictionary once again, this time with values being list of list of ints.
-
-If the purpose of sending several sentences at a time to the tokenizer is to build a batch to feed the model, you will
-probably want:
-
- To pad each sentence to the maximum length there is in your batch.
- To truncate each sentence to the maximum length the model can accept (if applicable).
- To return tensors.
-
-You can do all of this by using the following options when feeding your list of sentences to the tokenizer:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> batch = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="pt")
-    >>> print(batch)
-    {'input_ids': tensor([[ 101, 8667,  146,  112,  182,  170, 1423, 5650,  102],
-                          [ 101, 1262, 1330, 5650,  102,    0,    0,    0,    0],
-                          [ 101, 1262, 1103, 1304, 1304, 1314, 1141,  102,    0]]),
-     'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0],
-                               [0, 0, 0, 0, 0, 0, 0, 0, 0],
-                               [0, 0, 0, 0, 0, 0, 0, 0, 0]]), 
-     'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1],
-                               [1, 1, 1, 1, 1, 0, 0, 0, 0],
-                               [1, 1, 1, 1, 1, 1, 1, 1, 0]])}
-    >>> ## TENSORFLOW CODE
-    >>> batch = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="tf")
-    >>> print(batch)
-    {'input_ids': tf.Tensor([[ 101, 8667,  146,  112,  182,  170, 1423, 5650,  102],
-                          [ 101, 1262, 1330, 5650,  102,    0,    0,    0,    0],
-                          [ 101, 1262, 1103, 1304, 1304, 1314, 1141,  102,    0]]),
-     'token_type_ids': tf.Tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0],
-                               [0, 0, 0, 0, 0, 0, 0, 0, 0],
-                               [0, 0, 0, 0, 0, 0, 0, 0, 0]]), 
-     'attention_mask': tf.Tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1],
-                               [1, 1, 1, 1, 1, 0, 0, 0, 0],
-                               [1, 1, 1, 1, 1, 1, 1, 1, 0]])}
-
-It returns a dictionary string to tensor. We can now see what the `attention_mask <glossary.html#attention-mask>`__ is
-all about: it points out which tokens the model should pay attention to and which ones it should not (because they
-represent padding in this case).
-
-
-Note that if your model does not have a maximum length associated to it, the command above will throw a warning. You
-can safely ignore it. You can also pass ``verbose=False`` to stop the tokenizer to throw those kinds of warnings.
-
-.. _sentence-pairs:
-
-Preprocessing pairs of sentences
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Sometimes you need to feed pair of sentences to your model. For instance, if you want to classify if two sentences in a
-pair are similar, or for question-answering models, which take a context and a question. For BERT models, the input is
-then represented like this: :obj:`[CLS] Sequence A [SEP] Sequence B [SEP]`
-
-You can encode a pair of sentences in the format expected by your model by supplying the two sentences as two arguments
-(not a list since a list of two sentences will be interpreted as a batch of two single sentences, as we saw before).
-This will once again return a dict string to list of ints:
-
-.. code-block::
-
-    >>> encoded_input = tokenizer("How old are you?", "I'm 6 years old")
-    >>> print(encoded_input)
-    {'input_ids': [101, 1731, 1385, 1132, 1128, 136, 102, 146, 112, 182, 127, 1201, 1385, 102], 
-     'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1], 
-     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-This shows us what the `token_type_ids <glossary.html#token-type-ids>`__ are for: they indicate to the model which part
-of the inputs correspond to the first sentence and which part corresponds to the second sentence. Note that
-`token_type_ids` are not required or handled by all models. By default, a tokenizer will only return the inputs that
-its associated model expects. You can force the return (or the non-return) of any of those special arguments by
-using ``return_input_ids`` or ``return_token_type_ids``.
-
-If we decode the token ids we obtained, we will see that the special tokens have been properly added.
-
-.. code-block::
-
-    >>> tokenizer.decode(encoded_input["input_ids"])
-    "[CLS] How old are you? [SEP] I'm 6 years old [SEP]"
-
-If you have a list of pairs of sequences you want to process, you should feed them as two lists to your tokenizer: the
-list of first sentences and the list of second sentences:
-
-.. code-block::
-
-    >>> batch_sentences = ["Hello I'm a single sentence",
-    ...                    "And another sentence",
-    ...                    "And the very very last one"]
-    >>> batch_of_second_sentences = ["I'm a sentence that goes with the first sentence",
-    ...                              "And I should be encoded with the second sentence",
-    ...                              "And I go with the very last one"]
-    >>> encoded_inputs = tokenizer(batch_sentences, batch_of_second_sentences)
-    >>> print(encoded_inputs)
-    {'input_ids': [[101, 8667, 146, 112, 182, 170, 1423, 5650, 102, 146, 112, 182, 170, 5650, 1115, 2947, 1114, 1103, 1148, 5650, 102], 
-                   [101, 1262, 1330, 5650, 102, 1262, 146, 1431, 1129, 12544, 1114, 1103, 1248, 5650, 102], 
-                   [101, 1262, 1103, 1304, 1304, 1314, 1141, 102, 1262, 146, 1301, 1114, 1103, 1304, 1314, 1141, 102]], 
-    'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
-                       [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
-                       [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 
-    'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
-                       [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
-                       [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]}
-
-As we can see, it returns a dictionary with the values being list of lists of ints.
-
-To double-check what is fed to the model, we can decode each list in `input_ids` one by one:
-
-.. code-block::
-
-    >>> for ids in encoded_inputs["input_ids"]:
-    >>>     print(tokenizer.decode(ids))
-    [CLS] Hello I'm a single sentence [SEP] I'm a sentence that goes with the first sentence [SEP]
-    [CLS] And another sentence [SEP] And I should be encoded with the second sentence [SEP]
-    [CLS] And the very very last one [SEP] And I go with the very last one [SEP]
-
-Once again, you can automatically pad your inputs to the maximum sentence length in the batch, truncate to the maximum
-length the model can accept and return tensors directly with the following:
-
-.. code-block::
-
-    ## PYTORCH CODE
-    batch = tokenizer(batch_sentences, batch_of_second_sentences, padding=True, truncation=True, return_tensors="pt")
-    ## TENSORFLOW CODE
-    batch = tokenizer(batch_sentences, batch_of_second_sentences, padding=True, truncation=True, return_tensors="tf")
-
-Everything you always wanted to know about padding and truncation
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-We have seen the commands that will work for most cases (pad your batch to the length of the maximum sentence and
-
-truncate to the maximum length the mode can accept). However, the API supports more strategies if you need them. The
-three arguments you need to know for this are :obj:`padding`, :obj:`truncation` and :obj:`max_length`.
-
- :obj:`padding` controls the padding. It can be a boolean or a string which should be:
-
-    - :obj:`True` or :obj:`'longest'` to pad to the longest sequence in the batch (doing no padding if you only provide
-      a single sequence).
-    - :obj:`'max_length'` to pad to a length specified by the :obj:`max_length` argument or the maximum length accepted
-      by the model if no :obj:`max_length` is provided (``max_length=None``). If you only provide a single sequence,
-      padding will still be applied to it. 
-    - :obj:`False` or :obj:`'do_not_pad'` to not pad the sequences. As we have seen before, this is the default
-      behavior.
-
- :obj:`truncation` controls the truncation. It can be a boolean or a string which should be:
-
-    - :obj:`True` or :obj:`'only_first'` truncate to a maximum length specified by the :obj:`max_length` argument or
-      the maximum length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will
-      only truncate the first sentence of a pair if a pair of sequence (or a batch of pairs of sequences) is provided.
-    - :obj:`'only_second'` truncate to a maximum length specified by the :obj:`max_length` argument or the maximum
-      length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will only truncate
-      the second sentence of a pair if a pair of sequence (or a batch of pairs of sequences) is provided.
-    - :obj:`'longest_first'` truncate to a maximum length specified by the :obj:`max_length` argument or the maximum
-      length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will truncate token
-      by token, removing a token from the longest sequence in the pair until the proper length is reached.
-    - :obj:`False` or :obj:`'do_not_truncate'` to not truncate the sequences. As we have seen before, this is the
-      default behavior.
-
- :obj:`max_length` to control the length of the padding/truncation. It can be an integer or :obj:`None`, in which case
-  it will default to the maximum length the model can accept. If the model has no specific maximum input length,
-  truncation/padding to :obj:`max_length` is deactivated.
-
-Here is a table summarizing the recommend way to setup padding and truncation. If you use pair of inputs sequence in
-any of the following examples, you can replace :obj:`truncation=True` by a :obj:`STRATEGY` selected in 
-:obj:`['only_first', 'only_second', 'longest_first']`, i.e. :obj:`truncation='only_second'` or
-:obj:`truncation= 'longest_first'` to control how both sequence in the pair are truncated as detailed before.
-
-+--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
-| Truncation                           | Padding                           | Instruction                                                                                 |
-+======================================+===================================+=============================================================================================+
-| no truncation                        | no padding                        | :obj:`tokenizer(batch_sentences)`                                                           |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True)` or                                          |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='longest')`                                        |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max model input length | :obj:`tokenizer(batch_sentences, padding='max_length')`                                     |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to specific length        | :obj:`tokenizer(batch_sentences, padding='max_length', max_length=42)`                      |
-+--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
-| truncation to max model input length | no padding                        | :obj:`tokenizer(batch_sentences, truncation=True)` or                                       |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, truncation=STRATEGY)`                                      |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True, truncation=True)` or                         |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, padding=True, truncation=STRATEGY)`                        |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max model input length | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=True)` or                 |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY)`                |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to specific length        | Not possible                                                                                |
-+--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
-| truncation to specific length        | no padding                        | :obj:`tokenizer(batch_sentences, truncation=True, max_length=42)` or                        |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, truncation=STRATEGY, max_length=42)`                       |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True, truncation=True, max_length=42)` or          |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, padding=True, truncation=STRATEGY, max_length=42)`         |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to max model input length | Not possible                                                                                |
-|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
-|                                      | padding to specific length        | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=True, max_length=42)` or  |
-|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY, max_length=42)` |
-+--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
-
-Pre-tokenized inputs
-~~~~~~~~~~~~~~~~~~~~
-
-The tokenizer also accept pre-tokenized inputs. This is particularly useful when you want to compute labels and extract
-predictions in `named entity recognition (NER) <https://en.wikipedia.org/wiki/Named-entity_recognition>`__ or
-`part-of-speech tagging (POS tagging) <https://en.wikipedia.org/wiki/Part-of-speech_tagging>`__.
-
-If you want to use pre-tokenized inputs, just set :obj:`is_pretokenized=True` when passing your inputs to the
-tokenizer. For instance, we have:
-
-.. code-block::
-
-    >>> encoded_input = tokenizer(["Hello", "I'm", "a", "single", "sentence"], is_pretokenized=True)
-    >>> print(encoded_input)
-    {'input_ids': [101, 8667, 146, 112, 182, 170, 1423, 5650, 102],
-     'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0], 
-     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-Note that the tokenizer still adds the ids of special tokens (if applicable) unless you pass
-``add_special_tokens=False``.
-
-This works exactly as before for batch of sentences or batch of pairs of sentences. You can encode a batch of sentences
-like this:
-
-.. code-block::
-
-    batch_sentences = [["Hello", "I'm", "a", "single", "sentence"],
-                       ["And", "another", "sentence"],
-                       ["And", "the", "very", "very", "last", "one"]]
-    encoded_inputs = tokenizer(batch_sentences, is_pretokenized=True)
-
-or a batch of pair sentences like this:
-
-.. code-block::
-
-    batch_of_second_sentences = [["I'm", "a", "sentence", "that", "goes", "with", "the", "first", "sentence"],
-                                 ["And", "I", "should", "be", "encoded", "with", "the", "second", "sentence"],
-                                 ["And", "I", "go", "with", "the", "very", "last", "one"]]
-    encoded_inputs = tokenizer(batch_sentences, batch_of_second_sentences, is_pretokenized=True)
-
-And you can add padding, truncation as well as directly return tensors like before:
-
-.. code-block::
-
-    ## PYTORCH CODE
-    batch = tokenizer(batch_sentences,
-                      batch_of_second_sentences,
-                      is_pretokenized=True,
-                      padding=True,
-                      truncation=True,
-                      return_tensors="pt")
-    ## TENSORFLOW CODE
-    batch = tokenizer(batch_sentences,
-                      batch_of_second_sentences,
-                      is_pretokenized=True,
-                      padding=True,
-                      truncation=True,
-                      return_tensors="tf")
@@ -74,16 +74,14 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   |                                                            | (see `details on dbmdz repository <https://github.com/dbmdz/german-bert>`__).                                                         |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/bert-base-japanese``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
-|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
-|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
-|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
+|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece.                                                               |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``cl-tohoku/bert-base-japanese-whole-word-masking``        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
-|                   |                                                            | | Trained on Japanese text. Text is tokenized with MeCab and WordPiece and this requires some extra dependencies,                     |
-|                   |                                                            | | `fugashi <https://github.com/polm/fugashi>`__ which is a wrapper around `MeCab <https://taku910.github.io/mecab/>`__.               |
-|                   |                                                            | | Use ``pip install transformers["ja"]`` (or ``pip install -e .["ja"]`` if you install from source) to install them.                  |
+|                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                                      |
+|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -353,8 +351,6 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 | MarianMT          | ``Helsinki-NLP/opus-mt-{src}-{tgt}``                       | | 12-layer, 512-hidden, 8-heads, ~74M parameter Machine translation models. Parameter counts vary depending on vocab size.            |
 |                   |                                                            | | (see `model list <https://huggingface.co/Helsinki-NLP>`_)                                                                           |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| Pegasus           | ``google/pegasus-{dataset}``                               | | 16-layer, 1024-hidden, 16-heads, ~568M parameter, 2.2 GB for summary. `model list <https://huggingface.co/models?search=pegasus>`__ |
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | Longformer        | ``allenai/longformer-base-4096``                           | | 12-layer, 768-hidden, 12-heads, ~149M parameters                                                                                    |
 |                   |                                                            | | Starting from RoBERTa-base checkpoint, trained on documents of max length 4,096                                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -1,404 +1,379 @@
-Quick tour
-==========
-
-Let's have a quick look at the 🤗 Transformers library features. The library downloads pretrained models for
-Natural Language Understanding (NLU) tasks, such as analyzing the sentiment of a text, and Natural Language Generation (NLG),
-such as completing a prompt with new text or translating in another language.
-
-First we will see how to easily leverage the pipeline API to quickly use those pretrained models at inference. Then, we
-will dig a little bit more and see how the library gives you access to those models and helps you preprocess your data.
-
-.. note::
-
-    All code examples presented in the documentation have a switch on the top left for Pytorch versus TensorFlow. If
-    not, the code is expected to work for both backends without any change needed.
-
-Getting started on a task with a pipeline
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
-provides the following tasks out of the box:
-
- Sentiment analysis: is a text positive or negative?
- Text generation (in English): provide a prompt and the model will generate what follows.
- Name entity recognition (NER): in an input sentence, label each word with the entity it represents (person, place,
-  etc.)
- Question answering: provide the model with some context and a question, extract the answer from the context.
- Filling masked text: given a text with masked words (e.g., replaced by ``[MASK]``), fill the blanks.
- Summarization: generate a summary of a long text.
- Translation: translate a text in another language.
- Feature extraction: return a tensor representation of the text.
-
-Let's see how this work for sentiment analysis (the other tasks are all covered in the
-:doc:`task summary </task_summary>`):
-
-.. code-block::
-
-    >>> from transformers import pipeline
-    >>> classifier = pipeline('sentiment-analysis')
-
-When typing this command for the first time, a pretrained model and its tokenizer are downloaded and cached. We will
-look at both later on, but as an introduction the tokenizer's job is to preprocess the text for the model, which is
-then responsible for making predictions. The pipeline groups all of that together, and post-process the predictions to
-make them readable. For instance:
-
-
-.. code-block::
-
-    >>> classifier('We are very happy to show you the 🤗 Transformers library.')
-    [{'label': 'POSITIVE', 'score': 0.9997795224189758}]
-
-That's encouraging! You can use it on a list of sentences, which will be preprocessed then fed to the model as a
-`batch`, returning a list of dictionaries like this one:
-
-.. code-block::
-
-    >>> results = classifier(["We are very happy to show you the 🤗 Transformers library.",
-    ...            "We hope you don't hate it."])
-    >>> for result in results:
-    ...     print(f"label: {result['label']}, with score: {round(result['score'], 4)}")
-    label: POSITIVE, with score: 0.9998
-    label: NEGATIVE, with score: 0.5309
-
-You can see the second sentence has been classified as negative (it needs to be positive or negative) but its score is
-fairly neutral.
-
-By default, the model downloaded for this pipeline is called "distilbert-base-uncased-finetuned-sst-2-english". We can
-look at its `model page <https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english>`__ to get more
-information about it. It uses the :doc:`DistilBERT architecture </model_doc/distilbert>` and has been fine-tuned on a
-dataset called SST-2 for the sentiment analysis task.
-
-Let's say we want to use another model; for instance, one that has been trained on French data. We can search through
-the `model hub <https://huggingface.co/models>`__ that gathers models pretrained on a lot of data by research labs, but
-also community models (usually fine-tuned versions of those big models on a specific dataset). Applying the tags
-"French" and "text-classification" gives back a suggestion "nlptown/bert-base-multilingual-uncased-sentiment". Let's
-see how we can use it.
-
-You can directly pass the name of the model to use to :func:`~transformers.pipeline`:
-
-.. code-block::
-
-    >>> classifier = pipeline('sentiment-analysis', model="nlptown/bert-base-multilingual-uncased-sentiment")
-
-This classifier can now deal with texts in English, French, but also Dutch, German, Italian and Spanish! You can also
-replace that name by a local folder where you have saved a pretrained model (see below). You can also pass a model
-object and its associated tokenizer.
-
-We will need two classes for this. The first is :class:`~transformers.AutoTokenizer`, which we will use to download the
-tokenizer associated to the model we picked and instantiate it. The second is
-:class:`~transformers.AutoModelForSequenceClassification` (or
-:class:`~transformers.TFAutoModelForSequenceClassification` if you are using TensorFlow), which we will use to download
-the model itself. Note that if we were using the library on an other task, the class of the model would change. The
-:doc:`task summary </task_summary>` tutorial summarizes which class is used for which task.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-
-Now, to download the models and tokenizer we found previously, we just have to use the
-:func:`~transformers.AutoModelForSequenceClassification.from_pretrained` method (feel free to replace ``model_name`` by
-any other model from the model hub):
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    >>> model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
-    >>> classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-    >>> ## TENSORFLOW CODE
-    >>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    >>> # This model only exists in PyTorch, so we use the `from_pt` flag to import that model in TensorFlow.
-    >>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name, from_pt=True)
-    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
-    >>> classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-
-If you don't find a model that has been pretrained on some data similar to yours, you will need to fine-tune a
-pretrained model on your data. We provide :doc:`example scripts </examples>` to do so. Once you're done, don't forget
-to share your fine-tuned model on the hub with the community, using :doc:`this tutorial </model_sharing>`.
-
-.. _pretrained-model:
-
-Under the hood: pretrained models
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
-using the :obj:`from_pretrained` method:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    >>> pt_model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    >>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
-    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-Using the tokenizer
-^^^^^^^^^^^^^^^^^^^
-
-We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
-words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
-that process (you can learn more about them in the :doc:`tokenizer summary <tokenizer_summary>`, which is why we need
-to instantiate the tokenizer using the name of the model, to make sure we use the same rules as when the model was
-pretrained.
-
-The second step is to convert those `tokens` into numbers, to be able to build a tensor out of them and feed them to
-the model. To do this, the tokenizer has a `vocab`, which is the part we download when we instantiate it with the
-:obj:`from_pretrained` method, since we need to use the same `vocab` as when the model was pretrained.
-
-To apply these steps on a given text, we can just feed it to our tokenizer:
-
-.. code-block::
-
-    >>> inputs = tokenizer("We are very happy to show you the 🤗 Transformers library.")
-
-This returns a dictionary string to list of ints. It contains the `ids of the tokens <glossary.html#input-ids>`__,
-as mentioned before, but also additional arguments that will be useful to the model. Here for instance, we also have an
-`attention mask <glossary.html#attention-mask>`__ that the model will use to have a better understanding of the sequence:
-
-
-.. code-block::
-
-    >>> print(inputs)
-    {'input_ids': [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-You can pass a list of sentences directly to your tokenizer. If your goal is to send them through your model as a
-batch, you probably want to pad them all to the same length, truncate them to the maximum length the model can accept
-and get tensors back. You can specify all of that to the tokenizer:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> pt_batch = tokenizer(
-    ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
-    ...     padding=True,
-    ...     truncation=True,
-    ...     return_tensors="pt"
-    ... )
-    >>> ## TENSORFLOW CODE
-    >>> tf_batch = tokenizer(
-    ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
-    ...     padding=True,
-    ...     truncation=True,
-    ...     return_tensors="tf"
-    ... )
-
-The padding is automatically applied on the side expected by the model (in this case, on the right), with the
-padding token the model was pretrained with. The attention mask is also adapted to take the padding into account:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> for key, value in pt_batch.items():
-    ...     print(f"{key}: {value.numpy().tolist()}")
-    input_ids: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0, 0]]
-    attention_mask: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]]
-    >>> ## TENSORFLOW CODE
-    >>> for key, value in tf_batch.items():
-    ...     print(f"{key}: {value.numpy().tolist()}")
-    input_ids: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0, 0]]
-    attention_mask: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]]
-
-You can learn more about tokenizers :doc:`here <preprocessing>`.
-
-Using the model
-^^^^^^^^^^^^^^^
-
-Once your input has been preprocessed by the tokenizer, you can send it directly to the model. As we mentioned, it will
-contain all the relevant information the model needs. If you're using a TensorFlow model, you can pass the
-dictionary keys directly to tensor, for a PyTorch model, you need to unpack the dictionary by adding :obj:`**`.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> pt_outputs = pt_model(**pt_batch)
-    >>> ## TENSORFLOW CODE
-    >>> tf_outputs = tf_model(tf_batch)
-
-In 🤗 Transformers, all outputs are tuples (with only one element potentially). Here, we get a tuple with just the
-final activations of the model.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> print(pt_outputs)
-    (tensor([[-4.0833,  4.3364],
-            [ 0.0818, -0.0418]], grad_fn=<AddmmBackward>),)
-    >>> ## TENSORFLOW CODE
-    >>> print(tf_outputs)
-    (<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
-    array([[-4.0832963 ,  4.336414  ],
-           [ 0.08181786, -0.04179301]], dtype=float32)>,)
-
-The model can return more than just the final activations, which is why the output is a tuple. Here we only asked for
-the final activations, so we get a tuple with one element.
-.. note::
-
-    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
-    activation function (like SoftMax) since this final activation function is often fused with the loss.
-
-Let's apply the SoftMax activation to get predictions.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> import torch.nn.functional as F
-    >>> pt_predictions = F.softmax(pt_outputs[0], dim=-1)
-    >>> ## TENSORFLOW CODE
-    >>> import tensorflow as tf
-    >>> tf_predictions = tf.nn.softmax(tf_outputs[0], axis=-1)
-
-We can see we get the numbers from before:
-
-.. code-block::
-
-    >>> ## TENSORFLOW CODE
-    >>> print(tf_predictions)
-    tf.Tensor(
-    [[2.2042994e-04 9.9977952e-01]
-     [5.3086340e-01 4.6913657e-01]], shape=(2, 2), dtype=float32)
-    >>> ## PYTORCH CODE
-    >>> print(pt_predictions)
-    tensor([[2.2043e-04, 9.9978e-01],
-            [5.3086e-01, 4.6914e-01]], grad_fn=<SoftmaxBackward>)
-
-If you have labels, you can provide them to the model, it will return a tuple with the loss and the final activations.
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> import torch
-    >>> pt_outputs = pt_model(**pt_batch, labels = torch.tensor([1, 0]))
-    >>> ## TENSORFLOW CODE
-    >>> import tensorflow as tf
-    >>> tf_outputs = tf_model(tf_batch, labels = tf.constant([1, 0]))
-
-Models are standard `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ or
-`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ so you can use them in your usual
-training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer` (or :class:`~transformers.TFTrainer` if
-you are using TensorFlow) class to help with your training (taking care of things such as distributed training, mixed
-precision, etc.). See the :doc:`training tutorial <training>` for more details.
-
-.. note::
-
-    Pytorch model outputs are special dataclasses so that you can get autocompletion for their attributes in an IDE.
-    They also behave like a tuple or a dictionary (e.g., you can index with an integer, a slice or a string) in which
-    case the attributes not set (that have :obj:`None` values) are ignored.
-
-Once your model is fine-tuned, you can save it with its tokenizer in the following way:
-
-.. code-block::
-
-    tokenizer.save_pretrained(save_directory)
-    model.save_pretrained(save_directory)
-
-You can then load this model back using the :func:`~transformers.AutoModel.from_pretrained` method by passing the
-directory name instead of the model name. One cool feature of 🤗 Transformers is that you can easily switch between
-PyTorch and TensorFlow: any model saved as before can be loaded back either in PyTorch or TensorFlow. If you are
-loading a saved PyTorch model in a TensorFlow model, use :func:`~transformers.TFAutoModel.from_pretrained` like this:
-
-.. code-block::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = TFAutoModel.from_pretrained(save_directory, from_pt=True)
-
-and if you are loading a saved TensorFlow model in a PyTorch model, you should use the following code:
-
-.. code-block::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = AutoModel.from_pretrained(save_directory, from_tf=True)
-
-Lastly, you can also ask the model to return all hidden states and all attention weights if you need them:
-
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> pt_outputs = pt_model(**pt_batch, output_hidden_states=True, output_attentions=True)
-    >>> all_hidden_states, all_attentions = pt_outputs[-2:]
-    >>> ## TENSORFLOW CODE
-    >>> tf_outputs = tf_model(tf_batch, output_hidden_states=True, output_attentions=True)
-    >>> all_hidden_states, all_attentions = tf_outputs[-2:]
-
-Accessing the code
-^^^^^^^^^^^^^^^^^^
-
-The :obj:`AutoModel` and :obj:`AutoTokenizer` classes are just shortcuts that will automatically work with any
-pretrained model. Behind the scenes, the library has one model class per combination of architecture plus class, so the
-code is easy to access and tweak if you need to.
-
-In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
-using the :doc:`DistilBERT </model_doc/distilbert>` architecture. As
-:class:`~transformers.AutoModelForSequenceClassification` (or :class:`~transformers.TFAutoModelForSequenceClassification`
-if you are using TensorFlow) was used, the model automatically created is then a
-:class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
-to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
-without the auto magic:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
-    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    >>> model = DistilBertForSequenceClassification.from_pretrained(model_name)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification
-    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    >>> model = TFDistilBertForSequenceClassification.from_pretrained(model_name)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-
-Customizing the model
-^^^^^^^^^^^^^^^^^^^^^
-
-If you want to change how the model itself is built, you can define your custom configuration class. Each architecture
-comes with its own relevant configuration (in the case of DistilBERT, :class:`~transformers.DistilBertConfig`) which
-allows you to specify any of the hidden dimension, dropout rate, etc. If you do core modifications, like changing the
-hidden size, you won't be able to use a pretrained model anymore and will need to train from scratch. You would then
-instantiate the model directly from this configuration.
-
-Here we use the predefined vocabulary of DistilBERT (hence load the tokenizer with the
-:func:`~transformers.DistilBertTokenizer.from_pretrained` method) and initialize the model from scratch (hence
-instantiate the model from the configuration instead of using the
-:func:`~transformers.DistilBertForSequenceClassification.from_pretrained` method).
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    >>> config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    >>> model = DistilBertForSequenceClassification(config)
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    >>> config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    >>> model = TFDistilBertForSequenceClassification(config)
-
-For something that only changes the head of the model (for instance, the number of labels), you can still use a
-pretrained model for the body. For instance, let's define a classifier for 10 different labels using a pretrained body.
-We could create a configuration with all the default values and just change the number of labels, but more easily, you
-can directly pass any argument a configuration would take to the :func:`from_pretrained` method and it will update the
-default configuration with it:
-
-.. code-block::
-
-    >>> ## PYTORCH CODE
-    >>> from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    >>> model_name = "distilbert-base-uncased"
-    >>> model = DistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    >>> ## TENSORFLOW CODE
-    >>> from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    >>> model_name = "distilbert-base-uncased"
-    >>> model = TFDistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+Quick tour
+==========
+
+Let's have a quick look at the 🤗 Transformers library features. The library downloads pretrained models for
+Natural Language Understanding (NLU) tasks, such as analyzing the sentiment of a text, and Natural Language Generation (NLG),
+such as completing a prompt with new text or translating in another language.
+
+First we will see how to easily leverage the pipeline API to quickly use those pretrained models at inference. Then, we
+will dig a little bit more and see how the library gives you access to those models and helps you preprocess your data.
+
+.. note::
+
+    All code examples presented in the documentation have a switch on the top left for Pytorch versus TensorFlow. If
+    not, the code is expected to work for both backends without any change needed.
+
+Getting started on a task with a pipeline
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
+provides the following tasks out of the box:
+
+- Sentiment analysis: is a text positive or negative?
+- Text generation (in English): provide a prompt and the model will generate what follows.
+- Name entity recognition (NER): in an input sentence, label each word with the entity it represents (person, place,
+  etc.)
+- Question answering: provide the model with some context and a question, extract the answer from the context.
+- Filling masked text: given a text with masked words (e.g., replaced by ``[MASK]``), fill the blanks.
+- Summarization: generate a summary of a long text.
+- Translation: translate a text in another language.
+- Feature extraction: return a tensor representation of the text.
+
+Let's see how this work for sentiment analysis (the other tasks are all covered in the
+:doc:`task summary </task_summary>`):
+
+::
+
+    from transformers import pipeline
+    classifier = pipeline('sentiment-analysis')
+
+When typing this command for the first time, a pretrained model and its tokenizer are downloaded and cached. We will
+look at both later on, but as an introduction the tokenizer's job is to preprocess the text for the model, which is
+then responsible for making predictions. The pipeline groups all of that together, and post-process the predictions to
+make them readable. For instance
+
+::
+
+    classifier('We are very happy to show you the Transformers library.')
+
+will return something like this:
+
+::
+
+    [{'label': 'POSITIVE', 'score': 0.999799370765686}]
+
+That's encouraging! You can use it on a list of sentences, which will be preprocessed then fed to the model as a
+`batch`:
+
+::
+
+    classifier(["We are very happy to show you the Transformers library.",
+                "We hope you don't hate it."])
+
+returning a list of dictionaries like this one:
+
+::
+
+    [{'label': 'POSITIVE', 'score': 0.999799370765686},
+     {'label': 'NEGATIVE', 'score': 0.5308589935302734}]
+
+You can see the second sentence has been classified as negative (it needs to be positive or negative) but its score is
+fairly neutral.
+
+By default, the model downloaded for this pipeline is called "distilbert-base-uncased-finetuned-sst-2-english". We can
+look at its `model page <https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english>`__ to get more
+information about it. It uses the :doc:`DistilBERT architecture </model_doc/distilbert>` and has been fine-tuned on a
+dataset called SST-2 for the sentiment analysis task.
+
+Let's say we want to use another model; for instance, one that has been trained on French data. We can search through
+the `model hub <https://huggingface.co/models>`__ that gathers models pretrained on a lot of data by research labs, but
+also community models (usually fine-tuned versions of those big models on a specific dataset). Applying the tags
+"French" and "text-classification" gives back a suggestion "nlptown/bert-base-multilingual-uncased-sentiment". Let's
+see how we can use it. 
+
+You can directly pass the name of the model to use to :func:`~transformers.pipeline`:
+
+::
+
+    classifier = pipeline('sentiment-analysis', model="nlptown/bert-base-multilingual-uncased-sentiment")
+
+This classifier can now deal with texts in English, French, but also Dutch, German, Italian and Spanish! You can also
+replace that name by a local folder where you have saved a pretrained model (see below). You can also pass a model
+object and its associated tokenizer.
+
+We will need two classes for this. The first is :class:`~transformers.AutoTokenizer`, which we will use to download the
+tokenizer associated to the model we picked and instantiate it. The second is
+:class:`~transformers.AutoModelForSequenceClassification` (or
+:class:`~transformers.TFAutoModelForSequenceClassification` if you are using TensorFlow), which we will use to download
+the model itself. Note that if we were using the library on an other task, the class of the model would change. The
+:doc:`task summary </task_summary>` tutorial summarizes which class is used for which task.
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoTokenizer, AutoModelForSequenceClassification
+    ## TENSORFLOW CODE
+    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+
+Now, to download the models and tokenizer we found previously, we just have to use the 
+:func:`~transformers.AutoModelForSequenceClassification.from_pretrained` method (feel free to replace ``model_name`` by
+any other model from the model hub):
+
+::
+
+    ## PYTORCH CODE
+    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
+    model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    pipe = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
+    ## TENSORFLOW CODE
+    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
+    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
+
+If you don't find a model that has been pretrained on some data similar to yours, you will need to fine-tune a
+pretrained model on your data. We provide :doc:`example scripts </examples>` to do so. Once you're done, don't forget
+to share your fine-tuned model on the hub with the community, using :doc:`this tutorial </model_sharing>`.
+
+.. _pretrained-model:
+
+Under the hood: pretrained models
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
+using the :obj:`from_pretrained` method:
+
+::
+
+    ## PYTORCH CODE
+    from transformers import AutoTokenizer, AutoModelForSequenceClassification
+    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    ## TENSORFLOW CODE
+    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+Using the tokenizer
+^^^^^^^^^^^^^^^^^^^
+
+We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
+words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
+that process, which is why we need to instantiate the tokenizer using the name of the model, to make sure we use the
+same rules as when the model was pretrained.
+
+The second step is to convert those `tokens` into numbers, to be able to build a tensor out of them and feed them to
+the model. To do this, the tokenizer has a `vocab`, which is the part we download when we instantiate it with the
+:obj:`from_pretrained` method, since we need to use the same `vocab` as when the model was pretrained.
+
+To apply these steps on a given text, we can just feed it to our tokenizer:
+
+::
+
+    input = tokenizer("We are very happy to show you the Transformers library.")
+    print(input)
+
+This returns a dictionary string to list of ints. It contains the `ids of the tokens <glossary.html#input-ids>`__,
+as mentioned before, but also additional arguments that will be useful to the model. Here for instance, we also have an
+`attention mask <glossary.html#attention-mask>`__ that the model will use to have a better understanding of the sequence:
+
+
+::
+
+    {'input_ids': [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102],
+     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
+
+You can pass a list of sentences directly to your tokenizer. If your goal is to send them through your model as a
+batch, you probably want to pad them all to the same length, truncate them to the maximum length the model can accept
+and get tensors back. You can specify all of that to the tokenizer:
+
+::
+
+    ## PYTORCH CODE
+    batch = tokenizer(
+        ["We are very happy to show you the Transformers library.",
+         "We hope you don't hate it."],
+        padding=True, truncation=True, return_tensors="pt")
+    print(batch)
+    ## TENSORFLOW CODE
+    batch = tokenizer(
+        ["We are very happy to show you the Transformers library.",
+         "We hope you don't hate it."],
+        padding=True, truncation=True, return_tensors="tf")
+    print(batch)
+
+The padding is automatically applied on the side the model expect it (in this case, on the right), with the
+padding token the model was pretrained with. The attention mask is also adapted to take the padding into account:
+
+::
+
+    {'input_ids': tensor([[  101,  2057,  2024,  2200,  3407,  2000,  2265,  2017,  1996, 19081, 3075,  1012,   102],
+                          [  101,  2057,  3246,  2017,  2123,  1005,  1056,  5223,  2009,  1012,  102,     0,     0]]), 
+     'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                               [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])}
+
+You can learn more about tokenizers on their :doc:`doc page <main_classes/tokenizer>` (tutorial coming soon).
+
+Using the model
+^^^^^^^^^^^^^^^
+
+Once your input has been preprocessed by the tokenizer, you can directly send it to the model. As we mentioned, it will
+contain all the relevant information the model needs. If you're using a TensorFlow model, you can directly pass the
+dictionary keys to tensor, for a PyTorch model, you need to unpack the dictionary by adding :obj:`**`.
+
+::
+
+    ## PYTORCH CODE
+    outputs = model(**batch)
+    ## TENSORFLOW CODE
+    outputs = model(batch)
+
+In 🤗 Transformers, all outputs are tuples (with only one element potentially). Here, we get a tuple with just the
+final activations of the model.
+
+::
+
+    (tensor([[-4.1329,  4.3811],
+             [ 0.0818, -0.0418]]),)
+
+.. note::
+
+    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
+    activation function (like SoftMax) since this final activation function is often fused with the loss.
+
+Let's apply the SoftMax activation to get predictions.
+
+::
+
+    ## PYTORCH CODE
+    import torch.nn.functional as F
+    predictions = F.softmax(outputs[0], dim=-1)
+    print(predictions)
+    ## TENSORFLOW CODE
+    predictions = tf.nn.softmax(outputs[0], axis=-1)
+    print(predictions)
+
+We can see we get the numbers from before:
+
+::
+
+    tensor([[2.0060e-04, 9.9980e-01],
+            [5.3086e-01, 4.6914e-01]])
+
+If you have labels, you can provide them to the model, it will return a tuple with the loss and the final activations.
+
+::
+
+    ## PYTORCH CODE
+    import torch
+    outputs = model(**batch, labels = torch.tensor([1, 0])
+    ## TENSORFLOW CODE
+    import tensorflow as tf
+    outputs = model(batch, labels = tf.constant([1, 0])
+
+Models are standard `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ or
+`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ so you can use them in your usual
+training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer` (or :class:`~transformers.TFTrainer` if
+you are using TensorFlow) class to help with your training (taking care of things such as distributed training, mixed
+precision, etc.). See the training tutorial (coming soon) for more details.
+
+Once your model is fine-tuned, you can save it with its tokenizer the following way:
+
+::
+
+    tokenizer.save_pretrained(save_directory)
+    model.save_pretrained(save_directory)
+
+You can then load this model back using the :func:`~transformers.AutoModel.from_pretrained` method by passing the
+directory name instead of the model name. One cool feature of 🤗 Transformers is that you can easily switch between
+PyTorch and TensorFlow: any model saved as before can be loaded back either in PyTorch or TensorFlow. If you are
+loading a saved PyTorch model in a TensorFlow model, use :func:`~transformers.TFAutoModel.from_pretrained` like this:
+
+::
+
+    tokenizer = AutoTokenizer.from_pretrained(save_directory)
+    model = TFAutoModel.from_pretrained(save_directory, from_pt=True)
+
+and if you are loading a saved TensorFlow model in a PyTorch model, you should use the following code:
+
+::
+
+    tokenizer = AutoTokenizer.from_pretrained(save_directory)
+    model = AutoModel.from_pretrained(save_directory, from_tf=True)
+
+Lastly, you can also ask the model to return all hidden states and all attention weights if you need them:
+
+
+::
+
+    ## PYTORCH CODE
+    outputs = model(**batch, output_hidden_states=True, output_attentions=True)
+    all_hidden_states, all_attentions = outputs[-2:]
+    ## TENSORFLOW CODE
+    outputs = model(batch, output_hidden_states=True, output_attentions=True)
+    all_hidden_states, all_attentions = outputs[-2:]
+
+Accessing the code
+^^^^^^^^^^^^^^^^^^
+
+The :obj:`AutoModel` and :obj:`AutoTokenizer` classes are just shortcuts that will automatically work with any
+pretrained model. Behind the scenes, the library has one model class per combination of architecture plus class, so the
+code is easy to access and tweak if you need to.
+
+In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
+using the :doc:`DistilBERT </model_doc/distilbert>` architecture. The model automatically created is then a
+:class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
+to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
+without the auto magic:
+
+::
+
+    ## PYTORCH CODE
+    from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
+    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    model = DistilBertForSequenceClassification.from_pretrained(model_name)
+    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+    ## TENSORFLOW CODE
+    from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification
+    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    model = TFDistilBertForSequenceClassification.from_pretrained(model_name)
+    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+
+Customizing the model
+^^^^^^^^^^^^^^^^^^^^^
+
+If you want to change how the model itself is built, you can define your custom configuration class. Each architecture
+comes with its own relevant configuration (in the case of DistilBERT, :class:`~transformers.DistilBertConfig`) which
+allows you to specify any of the hidden dimension, dropout rate etc. If you do core modifications, like changing the
+hidden size, you won't be able to use a pretrained model anymore and will need to train from scratch. You would then
+instantiate the model directly from this configuration.
+
+Here we use the predefined vocabulary of DistilBERT (hence load the tokenizer with the
+:func:`~transformers.DistilBertTokenizer.from_pretrained` method) and initialize the model from scratch (hence
+instantiate the model from the configuration instead of using the
+:func:`~transformers.DistilBertForSequenceClassification.from_pretrained` method).
+
+::
+
+    ## PYTORCH CODE
+    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
+    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
+    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+    model = DistilBertForSequenceClassification(config)
+    ## TENSORFLOW CODE
+    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
+    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
+    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+    model = TFDistilBertForSequenceClassification(config)
+
+For something that only changes the head of the model (for instance, the number of labels), you can still use a
+pretrained model for the body. For instance, let's define a classifier for 10 different labels using a pretrained body.
+We could create a configuration with all the default values and just change the number of labels, but more easily, you
+can directly pass any argument a configuration would take to the :func:`from_pretrained` method and it will update the
+default configuration with it:
+
+::
+
+    ## PYTORCH CODE
+    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
+    model_name = "distilbert-base-uncased"
+    model = DistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
+    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+    ## TENSORFLOW CODE
+    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
+    model_name = "distilbert-base-uncased"
+    model = TFDistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
+    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
@@ -1,252 +1,89 @@
-**********************************************
-Exporting transformers models
-**********************************************
+Serialization best-practices
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

-ONNX / ONNXRuntime
-==============================================
-
-Projects `ONNX (Open Neural Network eXchange) <http://onnx.ai>`_ and `ONNXRuntime (ORT) <https://microsoft.github.io/onnxruntime/>`_ are part of an effort from leading industries in the AI field
-to provide a unified and community-driven format to store and, by extension, efficiently execute neural network leveraging a variety
-of hardware and dedicated optimizations.
-
-Starting from transformers v2.10.0 we partnered with ONNX Runtime to provide an easy export of transformers models to
-the ONNX format. You can have a look at the effort by looking at our joint blog post `Accelerate your NLP pipelines using
-Hugging Face Transformers and ONNX Runtime <https://medium.com/microsoftazure/accelerate-your-nlp-pipelines-using-hugging-face-transformers-and-onnx-runtime-2443578f4333>`_.
-
-Exporting a model is done through the script `convert_graph_to_onnx.py` at the root of the transformers sources.
-The following command shows how easy it is to export a BERT model from the library, simply run:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased bert-base-cased.onnx
-
-The conversion tool works for both PyTorch and Tensorflow models and ensures:
-
-* The model and its weights are correctly initialized from the Hugging Face model hub or a local checkpoint.
-* The inputs and outputs are correctly generated to their ONNX counterpart.
-* The generated model can be correctly loaded through onnxruntime.
-
-.. note::
-    Currently, inputs and outputs are always exported with dynamic sequence axes preventing some optimizations
-    on the ONNX Runtime. If you would like to see such support for fixed-length inputs/outputs, please
-    open up an issue on transformers.
+This section explain how you can save and re-load a fine-tuned model (BERT, GPT, GPT-2 and Transformer-XL).
+There are three types of files you need to save to be able to reload a fine-tuned model:


-Also, the conversion tool supports different options which let you tune the behavior of the generated model:
+* the model itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
+* the configuration file of the model which is saved as a JSON file, and
+* the vocabulary (and the merges for the BPE-based models GPT and GPT-2).

-* **Change the target opset version of the generated model.**  (More recent opset generally supports more operators and enables faster inference)
-
-* **Export pipeline-specific prediction heads.**  (Allow to export model along with its task-specific prediction head(s))
-
-* **Use the external data format (PyTorch only).**  (Lets you export model which size is above 2Gb (`More info <https://github.com/pytorch/pytorch/pull/33062>`_))
+The *default filenames* of these files are as follow:


-Optimizations
------------------------------------------------
+* the model weights file: ``pytorch_model.bin``\ ,
+* the configuration file: ``config.json``\ ,
+* the vocabulary file: ``vocab.txt`` for BERT and Transformer-XL, ``vocab.json`` for GPT/GPT-2 (BPE vocabulary),
+* for GPT/GPT-2 (BPE vocabulary) the additional merges file: ``merges.txt``.

-ONNXRuntime includes some transformers-specific transformations to leverage optimized operations in the graph.
-Below are some of the operators which can be enabled to speed up inference through ONNXRuntime (*see note below*):
+**If you save a model using these *default filenames*\ , you can then re-load the model and tokenizer using the ``from_pretrained()`` method.**

-* Constant folding
-* Attention Layer fusing
-* Skip connection LayerNormalization fusing
-* FastGeLU approximation
-
-
-Fortunately, you can let ONNXRuntime find all the possible optimized operators for you. Simply add ``--optimize``
-when exporting your model through ``convert_graph_to_onnx.py``.
-
-Example:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased --optimize bert-base-cased.onnx
-
-.. note::
-    For more information about the optimizations enabled by ONNXRuntime, please have a look at the (`ONNXRuntime Github <https://github.com/microsoft/onnxruntime/tree/master/onnxruntime/python/tools/transformers>`_)
-
-Quantization
------------------------------------------------
-
-ONNX exporter supports generating a quantized version of the model to allow efficient inference.
-
-Quantization works by converting the memory representation of the parameters in the neural network
-to a compact integer format. By default, weights of a neural network are stored as single-precision float (`float32`)
-which can express a wide-range of floating-point numbers with decent precision.
-These properties are especially interesting at training where you want fine-grained representation.
-
-On the other hand, after the training phase, it has been shown one can greatly reduce the range and the precision of `float32` numbers
-without changing the performances of the neural network.
-
-More technically, `float32` parameters are converted to a type requiring fewer bits to represent each number, thus reducing
-the overall size of the model. Here, we are enabling `float32` mapping to `int8` values (a non-floating, single byte, number representation)
-according to the following formula:
-
-.. math::
-    y_{float32} = scale * x_{int8} - zero\_point
-
-.. note::
-    The quantization process will infer the parameter `scale` and `zero_point` from the neural network parameters
-
-Leveraging tiny-integers has numerous advantages when it comes to inference:
-
-* Storing fewer bits instead of 32 bits for the `float32` reduces the size of the model and makes it load faster.
-* Integer operations execute a magnitude faster on modern hardware
-* Integer operations require less power to do the computations
-
-In order to convert a transformers model to ONNX IR with quantized weights you just need to specify ``--quantize``
-when using ``convert_graph_to_onnx.py``. Also, you can have a look at the ``quantize()`` utility-method in this
-same script file.
-
-Example of quantized BERT model export:
-
-.. code-block:: bash
-
-    python convert_graph_to_onnx.py --framework <pt, tf> --model bert-base-cased --quantize bert-base-cased.onnx
-
-.. note::
-    Quantization support requires ONNX Runtime >= 1.4.0
-
-.. note::
-    When exporting quantized model you will end up with two different ONNX files. The one specified at the end of the
-    above command will contain the original ONNX model storing `float32` weights.
-    The second one, with ``-quantized`` suffix, will hold the quantized parameters.
-
-.. note::
-    The quantization export gives the best performances when used in combination with ``--optimize``.
-
-TorchScript
-=======================================
-
-.. note::
-    This is the very beginning of our experiments with TorchScript and we are still exploring its capabilities
-    with variable-input-size models. It is a focus of interest to us and we will deepen our analysis in upcoming
-    releases, with more code examples, a more flexible implementation, and benchmarks comparing python-based codes
-    with compiled TorchScript.
-
-
-According to Pytorch's documentation: "TorchScript is a way to create serializable and optimizable models from PyTorch code".
-Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
-their model to be re-used in other programs, such as efficiency-oriented C++ programs.
-
-We have provided an interface that allows the export of 🤗 Transformers models to TorchScript so that they can
-be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
-they can be exported, and what to be mindful of when using these models with TorchScript.
-
-Exporting a model needs two things:
-
-* dummy inputs to execute a model forward pass.
-* the model needs to be instantiated with the ``torchscript`` flag.
-
-These necessities imply several things developers should be careful about. These are detailed below.
-
-
-Implications
------------------------------------------------
-
-TorchScript flag and tied weights
------------------------------------------------
-This flag is necessary because most of the language models in this repository have tied weights between their
-``Embedding`` layer and their ``Decoding`` layer. TorchScript does not allow the export of models that have tied weights,
-it is therefore necessary to untie the weights beforehand.
-
-This implies that models instantiated with the ``torchscript`` flag have their ``Embedding`` layer and ``Decoding`` layer
-separate, which means that they should not be trained down the line. Training would de-synchronize the two layers,
-leading to unexpected results.
-
-This is not the case for models that do not have a Language Model head, as those do not have tied weights. These models
-can be safely exported without the ``torchscript`` flag.
-
-Dummy inputs and standard lengths
------------------------------------------------
-
-The dummy inputs are used to do a model forward pass. While the inputs' values are propagating through the layers,
-Pytorch keeps track of the different operations executed on each tensor. These recorded operations are then used
-to create the "trace" of the model.
-
-The trace is created relatively to the inputs' dimensions. It is therefore constrained by the dimensions of the dummy
-input, and will not work for any other sequence length or batch size. When trying with a different size, an error such
-as:
-
-``The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2``
-
-will be raised. It is therefore recommended to trace the model with a dummy input size at least as large as the largest
-input that will be fed to the model during inference. Padding can be performed to fill the missing values. As the model
-will have been traced with a large input size however, the dimensions of the different matrix will be large as well,
-resulting in more calculations.
-
-It is recommended to be careful of the total number of operations done on each input and to follow performance closely
-when exporting varying sequence-length models.
-
-Using TorchScript in Python
-------------------------------------------------
-
-Below are examples of using the Python to save, load models as well as how to use the trace for inference.
-
-Saving a model
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-This snippet shows how to use TorchScript to export a ``BertModel``. Here the ``BertModel`` is instantiated
-according to a ``BertConfig`` class and then saved to disk under the filename ``traced_bert.pt``
+Here is the recommended way of saving the model, configuration and vocabulary to an ``output_dir`` directory and reloading the model and tokenizer afterwards:

 .. code-block:: python

-    from transformers import BertModel, BertTokenizer, BertConfig
-    import torch
+   from transformers import WEIGHTS_NAME, CONFIG_NAME

-    enc = BertTokenizer.from_pretrained("bert-base-uncased")
+   output_dir = "./models/"

-    # Tokenizing input text
-    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
-    tokenized_text = enc.tokenize(text)
+   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned

-    # Masking one of the input tokens
-    masked_index = 8
-    tokenized_text[masked_index] = '[MASK]'
-    indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
-    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+   # If we have a distributed model, save only the encapsulated model
+   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
+   model_to_save = model.module if hasattr(model, 'module') else model

-    # Creating a dummy input
-    tokens_tensor = torch.tensor([indexed_tokens])
-    segments_tensors = torch.tensor([segments_ids])
-    dummy_input = [tokens_tensor, segments_tensors]
+   # If we save using the predefined names, we can load using `from_pretrained`
+   output_model_file = os.path.join(output_dir, WEIGHTS_NAME)
+   output_config_file = os.path.join(output_dir, CONFIG_NAME)

-    # Initializing the model with the torchscript flag
-    # Flag set to True even though it is not necessary as this model does not have an LM Head.
-    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
-        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
+   torch.save(model_to_save.state_dict(), output_model_file)
+   model_to_save.config.to_json_file(output_config_file)
+   tokenizer.save_pretrained(output_dir)

-    # Instantiating the model
-    model = BertModel(config)
+   # Step 2: Re-load the saved model and vocabulary

-    # The model needs to be in evaluation mode
-    model.eval()
+   # Example for a Bert model
+   model = BertForQuestionAnswering.from_pretrained(output_dir)
+   tokenizer = BertTokenizer.from_pretrained(output_dir)  # Add specific options if needed
+   # Example for a GPT model
+   model = OpenAIGPTDoubleHeadsModel.from_pretrained(output_dir)
+   tokenizer = OpenAIGPTTokenizer.from_pretrained(output_dir)

-    # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
-    model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
-
-    # Creating the trace
-    traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
-    torch.jit.save(traced_model, "traced_bert.pt")
-
-Loading a model
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-This snippet shows how to load the ``BertModel`` that was previously saved to disk under the name ``traced_bert.pt``.
-We are re-using the previously initialised ``dummy_input``.
+Here is another way you can save and reload the model if you want to use specific paths for each type of files:

 .. code-block:: python

-    loaded_model = torch.jit.load("traced_model.pt")
-    loaded_model.eval()
+   output_model_file = "./models/my_own_model_file.bin"
+   output_config_file = "./models/my_own_config_file.bin"
+   output_vocab_file = "./models/my_own_vocab_file.bin"

-    all_encoder_layers, pooled_output = loaded_model(dummy_input)
+   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned

-Using a traced model for inference
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+   # If we have a distributed model, save only the encapsulated model
+   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
+   model_to_save = model.module if hasattr(model, 'module') else model

-Using the traced model for inference is as simple as using its ``__call__`` dunder method:
+   torch.save(model_to_save.state_dict(), output_model_file)
+   model_to_save.config.to_json_file(output_config_file)
+   tokenizer.save_vocabulary(output_vocab_file)

-.. code-block:: python
+   # Step 2: Re-load the saved model and vocabulary
+
+   # We didn't save using the predefined WEIGHTS_NAME, CONFIG_NAME names, we cannot load using `from_pretrained`.
+   # Here is how to do it in this situation:
+
+   # Example for a Bert model
+   config = BertConfig.from_json_file(output_config_file)
+   model = BertForQuestionAnswering(config)
+   state_dict = torch.load(output_model_file)
+   model.load_state_dict(state_dict)
+   tokenizer = BertTokenizer(output_vocab_file, do_lower_case=args.do_lower_case)
+
+   # Example for a GPT model
+   config = OpenAIGPTConfig.from_json_file(output_config_file)
+   model = OpenAIGPTDoubleHeadsModel(config)
+   state_dict = torch.load(output_model_file)
+   model.load_state_dict(state_dict)
+   tokenizer = OpenAIGPTTokenizer(output_vocab_file)

-    traced_model(tokens_tensor, segments_tensors)
@@ -1,243 +0,0 @@
-Tokenizer summary
-----------------
-
-In this page, we will have a closer look at tokenization. As we saw in
-:doc:`the preprocessing tutorial <preprocessing>`, tokenizing a text is splitting it into words or subwords, which then
-are converted to ids. The second part is pretty straightforward, here we will focus on the first part. More
-specifically, we will look at the three main different kinds of tokenizers used in 🤗 Transformers:
-:ref:`Byte-Pair Encoding (BPE) <byte-pair-encoding>`, :ref:`WordPiece <wordpiece>` and
-:ref:`SentencePiece <sentencepiece>`, and provide examples of models using each of those.
-
-Note that on each model page, you can look at the documentation of the associated tokenizer to know which of those
-algorithms the pretrained model used. For instance, if we look at :class:`~transformers.BertTokenizer`, we can see it's
-using :ref:`WordPiece <wordpiece>`.
-
-Introduction to tokenization
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Splitting a text in smaller chunks is a task that's harder than it looks, and there are multiple ways of doing it. For
-instance, let's look at the sentence "Don't you love 🤗 Transformers? We sure do." A first simple way of tokenizing
-this text is just to split it by spaces, which would give:
-
-::
-
-    ["Don't", "you", "love", "🤗", "Transformers?", "We", "sure", "do."]
-
-This is a nice first step, but if we look at the tokens "Transformers?" or "do.", we can see we can do better. Those
-will be different than the tokens "Transformers" and "do" for our model, so we should probably take the punctuation
-into account. This would give:
-
-::
-
-    ["Don", "'", "t", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
-
-which is better already. One thing that is annoying though is how it dealt with "Don't". "Don't" stands for do not, so
-it should probably be better tokenized as ``["Do", "n't"]``. This is where things start getting more complicated, and
-part of the reason each kind of model has its own tokenizer class. Depending on the rules we apply to split our texts
-into tokens, we'll get different tokenized versions of the same text. And of course, a given pretrained model won't
-perform properly if you don't use the exact same rules as the persons who pretrained it.
-
-`spaCy <https://spacy.io/>`__ and `Moses <http://www.statmt.org/moses/?n=Development.GetStarted>`__ are two popular
-rule-based tokenizers. On the text above, they'd output something like:
-
-::
-
-    ["Do", "n't", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
-
-Space/punctuation-tokenization and rule-based tokenization are both examples of word tokenization, which is splitting a
-sentence into words. While it's the most intuitive way to separate texts in smaller chunks, it can have a problem when
-you have a huge corpus: it usually yields a very big vocabulary (the set of all unique tokens used).
-:doc:`Transformer XL <model_doc/transformerxl>` for instance uses space/punctuation-tokenization, and has a vocabulary
-size of 267,735!
-
-A huge vocabulary size means a huge embedding matrix at the start of the model, which will cause memory problems.
-TransformerXL deals with it by using a special kind of embeddings called adaptive embeddings, but in general,
-transformers models rarely have a vocabulary size greater than 50,000, especially if they are trained on a single
-language.
-
-So if tokenizing on words is unsatisfactory, we could go on the opposite direction and simply tokenize on characters.
-While it's very simple and would save a lot of memory, this doesn't allow the model to learn representations of texts
-as meaningful as when using a word tokenization, leading to a loss of performance. So to get the best of both worlds,
-all transformers models use a hybrid between word-level and character-level tokenization called subword tokenization.
-
-Subword tokenization
-^^^^^^^^^^^^^^^^^^^^
-
-Subword tokenization algorithms rely on the principle that most common words should be left as is, but rare words
-should be decomposed in meaningful subword units. For instance "annoyingly" might be considered a rare word and
-decomposed as "annoying" and "ly". This is especially useful in agglutinative languages such as Turkish, where you can
-form (almost) arbitrarily long complex words by stringing together some subwords.
-
-This allows the model to keep a reasonable vocabulary while still learning useful representations for common words or
-subwords. This also enables the model to process words it has never seen before, by decomposing them into
-subwords it knows. For instance, the base :class:`~transformers.BertTokenizer` will tokenize "I have a new GPU!" like
-this:
-
-.. code-block::
-
-    >>> from transformers import BertTokenizer
-    >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-    >>> tokenizer.tokenize("I have a new GPU!")
-    ['i', 'have', 'a', 'new', 'gp', '##u', '!']
-
-Since we are considering the uncased model, the sentence was lowercased first. Then all the words were present in the
-vocabulary of the tokenizer, except for "gpu", so the tokenizer split it in subwords it knows: "gp" and "##u". The "##"
-means that the rest of the token should be attached to the previous one, without space (for when we need to decode
-predictions and reverse the tokenization).
-
-Another example is when we use the base :class:`~transformers.XLNetTokenizer` to tokenize our previous text:
-
-.. code-block::
-
-    >>> from transformers import XLNetTokenizer
-    >>> tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
-    >>> tokenizer.tokenize("Don't you love 🤗 Transformers? We sure do.")
-    ['▁Don', "'", 't', '▁you', '▁love', '▁', '🤗', '▁', 'Transform', 'ers', '?', '▁We', '▁sure', '▁do', '.']
-
-We'll get back to the meaning of those '▁' when we look at :ref:`SentencePiece <sentencepiece>` but you can see
-Transformers has been split into "Transform" and "ers".
-
-Let's now look at how the different subword tokenization algorithms work. Note that they all rely on some form of
-training which is usually done on the corpus the corresponding model will be trained on.
-
-.. _byte-pair-encoding:
-
-Byte-Pair Encoding
-~~~~~~~~~~~~~~~~~~
-
-Byte-Pair Encoding was introduced in `this paper <https://arxiv.org/abs/1508.07909>`__. It relies on a pretokenizer
-splitting the training data into words, which can be a simple space tokenization
-(:doc:`GPT-2 <model_doc/gpt2>` and :doc:`Roberta <model_doc/roberta>` uses this for instance) or a rule-based tokenizer
-(:doc:`XLM <model_doc/xlm>` use Moses for most languages, as does :doc:`FlauBERT <model_doc/flaubert>`),
-
-:doc:`GPT <model_doc/gpt>` uses Spacy and ftfy, and counts the frequency of each word in the training corpus.
-
-It then begins from the list of all characters, and will learn merge rules to form a new token from two symbols in the
-vocabulary until it has learned a vocabulary of the desired size (this is a hyperparameter to pick).
-
-Let's say that after the pre-tokenization we have the following words (the number indicating the frequency of each
-word):
-
-::
-
-    ('hug', 10), ('pug', 5), ('pun', 12), ('bun', 4), ('hugs', 5)
-
-Then the base vocabulary is ['b', 'g', 'h', 'n', 'p', 's', 'u'] and all our words are first split by character:
-
-::
-
-    ('h' 'u' 'g', 10), ('p' 'u' 'g', 5), ('p' 'u' 'n', 12), ('b' 'u' 'n', 4), ('h' 'u' 'g' 's', 5)
-
-We then take each pair of symbols and look at the most frequent. For instance 'hu' is present `10 + 5 = 15` times (10
-times in the 10 occurrences of 'hug', 5 times in the 5 occurrences of 'hugs'). The most frequent here is 'ug', present
-`10 + 5 + 2 + 5 = 22` times in total. So the first merge rule the tokenizer learns is to group all 'u' and 'g' together
-then it adds 'ug' to the vocabulary. Our corpus then becomes
-
-::
-
-    ('h' 'ug', 10), ('p' 'ug', 5), ('p' 'u' 'n', 12), ('b' 'u' 'n', 4), ('h' 'ug' 's', 5)
-
-and we continue by looking at the next most common pair of symbols. It's 'un', present 16 times, so we merge those two
-and add 'un' to the vocabulary. Then it's 'hug' (as 'h' + 'ug'), present 15 times, so we merge those two and add 'hug'
-to the vocabulary.
-
-At this stage, the vocabulary is ``['b', 'g', 'h', 'n', 'p', 's', 'u', 'ug', 'un', 'hug']`` and our corpus is
-represented as
-
-::
-
-    ('hug', 10), ('p' 'ug', 5), ('p' 'un', 12), ('b' 'un', 4), ('hug' 's', 5)
-
-If we stop there, the tokenizer can apply the rules it learned to new words (as long as they don't contain characters that
-were not in the base vocabulary). For instance 'bug' would be tokenized as ``['b', 'ug']`` but mug would be tokenized as
-``['<unk>', 'ug']`` since the 'm' is not in the base vocabulary. This doesn't happen to letters in general (since the
-base corpus uses all of them), but to special characters like emojis.
-
-As we said before, the vocabulary size (which is the base vocabulary size + the number of merges) is a hyperparameter
-to choose. For instance :doc:`GPT <model_doc/gpt>` has a vocabulary size of 40,478 since they have 478 base characters
-and chose to stop the training of the tokenizer at 40,000 merges.
-
-Byte-level BPE
-^^^^^^^^^^^^^^
-
-To deal with the fact the base vocabulary needs to get all base characters, which can be quite big if one allows for
-all unicode characters, the
-`GPT-2 paper <https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf>`__
-introduces a clever trick, which is to use bytes as the base vocabulary (which gives a size of 256). With some
-additional rules to deal with punctuation, this manages to be able to tokenize every text without needing an unknown
-token. For instance, the :doc:`GPT-2 model <model_doc/gpt>` has a vocabulary size of 50,257, which corresponds to the
-256 bytes base tokens, a special end-of-text token and the symbols learned with 50,000 merges.
-
-.. _wordpiece:
-
-WordPiece
-=========
-
-WordPiece is the subword tokenization algorithm used for :doc:`BERT <model_doc/bert>` (as well as
-:doc:`DistilBERT <model_doc/distilbert>` and :doc:`Electra <model_doc/electra>`) and was outlined in
-`this paper <https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf>`__. It relies
-on the same base as BPE, which is to initialize the vocabulary to every character present in the corpus and
-progressively learn a given number of merge rules, the difference is that it doesn't choose the pair that is the most
-frequent but the one that will maximize the likelihood on the corpus once merged.
-
-What does this mean? Well, in the previous example, it means we would only merge 'u' and 'g' if the probability of
-having 'ug' divided by the probability of having 'u' then 'g' is greater than for any other pair of symbols. It's
-subtly different from what BPE does in the sense that it evaluates what it "loses" by merging two symbols and makes
-sure it's `worth it`.
-
-.. _unigram:
-
-Unigram
-=======
-
-Unigram is a subword tokenization algorithm introduced in `this paper <https://arxiv.org/pdf/1804.10959.pdf>`__.
-Instead of starting with a group of base symbols and learning merges with some rule, like BPE or WordPiece, it starts
-from a large vocabulary (for instance, all pretokenized words and the most common substrings) that it will trim down
-progressively. It's not used directly for any of the pretrained models in the library, but it's used in conjunction
-with :ref:`SentencePiece <sentencepiece>`.
-
-More specifically, at a given step, unigram computes a loss from the corpus we have and the current vocabulary, then,
-for each subword, evaluate how much the loss would augment if the subword was removed from the vocabulary. It then
-sorts the subwords by this quantity (that represents how worse the loss becomes if the token is removed) and removes
-all the worst p tokens (for instance p could be 10% or 20%). It then repeats the process until the vocabulary has
-reached the desired size, always keeping the base characters (to be able to tokenize any word written with them, like
-BPE or WordPiece).
-
-Contrary to BPE and WordPiece that work out rules in a certain order that you can then apply in the same order when
-tokenizing new text, Unigram will have several ways of tokenizing a new text. For instance, if it ends up with the
-vocabulary
-
-::
-
-    ['b', 'g', 'h', 'n', 'p', 's', 'u', 'ug', 'un', 'hug']
-
-we had before, it could tokenize "hugs" as ``['hug', 's']``, ``['h', 'ug', 's']`` or ``['h', 'u', 'g', 's']``. So which
-one choose? On top of saving the vocabulary, the trained tokenizer will save the probability of each token in the
-training corpus. You can then give a probability to each tokenization (which is the product of the probabilities of the
-tokens forming it) and pick the most likely one (or if you want to apply some data augmentation, you could sample one
-of the tokenization according to their probabilities).
-
-Those probabilities define the loss that trains the tokenizer: if our corpus consists of the
-words :math:`x_{1}, \dots, x_{N}` and if for the word :math:`x_{i}` we note :math:`S(x_{i})` the set of all possible
-tokenizations of :math:`x_{i}` (with the current vocabulary), then the loss is defined as
-
-.. math::
-    \mathcal{L} = -\sum_{i=1}^{N} \log \left ( \sum_{x \in S(x_{i})} p(x) \right )
-
-.. _sentencepiece:
-
-SentencePiece
-=============
-
-All the methods we have been looking at so far required some form of pretokenization, which has a central problem: not
-all languages use spaces to separate words. This is a problem :doc:`XLM <model_doc/xlm>` solves by using specific
-pretokenizers for each of those languages (in this case, Chinese, Japanese and Thai). To solve this problem,
-SentencePiece (introduced in `this paper <https://arxiv.org/pdf/1808.06226.pdf>`__) treats the input as a raw stream,
-includes the space in the set of characters to use, then uses BPE or unigram to construct the appropriate vocabulary.
-
-That's why in the example we saw before using :class:`~transformers.XLNetTokenizer` (which uses SentencePiece), we had
-the '▁' character, that represents space. Decoding a tokenized text is then super easy: we just have to concatenate
-all of them together and replace '▁' with space.
-
-All transformers models in the library that use SentencePiece use it with unigram. Examples of models using it are
-:doc:`ALBERT <model_doc/albert>`, :doc:`XLNet <model_doc/xlnet>` or the :doc:`Marian framework <model_doc/marian>`.
@@ -0,0 +1,135 @@
+TorchScript
+================================================
+
+.. note::
+    This is the very beginning of our experiments with TorchScript and we are still exploring its capabilities
+    with variable-input-size models. It is a focus of interest to us and we will deepen our analysis in upcoming
+    releases, with more code examples, a more flexible implementation, and benchmarks comparing python-based codes
+    with compiled TorchScript.
+
+
+According to Pytorch's documentation: "TorchScript is a way to create serializable and optimizable models from PyTorch code".
+Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
+their model to be re-used in other programs, such as efficiency-oriented C++ programs.
+
+We have provided an interface that allows the export of `transformers` models to TorchScript so that they can
+be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
+they can be exported, and what to be mindful of when using these models with TorchScript.
+
+Exporting a model needs two things:
+
+* dummy inputs to execute a model forward pass.
+* the model needs to be instantiated with the ``torchscript`` flag.
+
+These necessities imply several things developers should be careful about. These are detailed below.
+
+
+Implications
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+TorchScript flag and tied weights
+------------------------------------------------
+This flag is necessary because most of the language models in this repository have tied weights between their
+``Embedding`` layer and their ``Decoding`` layer. TorchScript does not allow the export of models that have tied weights,
+it is therefore necessary to untie the weights beforehand.
+
+This implies that models instantiated with the ``torchscript`` flag have their ``Embedding`` layer and ``Decoding`` layer
+separate, which means that they should not be trained down the line. Training would de-synchronize the two layers,
+leading to unexpected results.
+
+This is not the case for models that do not have a Language Model head, as those do not have tied weights. These models
+can be safely exported without the ``torchscript`` flag.
+
+Dummy inputs and standard lengths
+------------------------------------------------
+
+The dummy inputs are used to do a model forward pass. While the inputs' values are propagating through the layers,
+Pytorch keeps track of the different operations executed on each tensor. These recorded operations are then used
+to create the "trace" of the model.
+
+The trace is created relatively to the inputs' dimensions. It is therefore constrained by the dimensions of the dummy
+input, and will not work for any other sequence length or batch size. When trying with a different size, an error such
+as:
+
+``The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2``
+
+will be raised. It is therefore recommended to trace the model with a dummy input size at least as large as the largest
+input that will be fed to the model during inference. Padding can be performed to fill the missing values. As the model
+will have been traced with a large input size however, the dimensions of the different matrix will be large as well,
+resulting in more calculations.
+
+It is recommended to be careful of the total number of operations done on each input and to follow performance closely
+when exporting varying sequence-length models.
+
+Using TorchScript in Python
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Below are examples of using the Python to save, load models as well as how to use the trace for inference.
+
+Saving a model
+------------------------------------------------
+
+This snippet shows how to use TorchScript to export a ``BertModel``. Here the ``BertModel`` is instantiated
+according to a ``BertConfig`` class and then saved to disk under the filename ``traced_bert.pt``
+
+.. code-block:: python
+
+    from transformers import BertModel, BertTokenizer, BertConfig
+    import torch
+
+    enc = BertTokenizer.from_pretrained("bert-base-uncased")
+
+    # Tokenizing input text
+    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+    tokenized_text = enc.tokenize(text)
+
+    # Masking one of the input tokens
+    masked_index = 8
+    tokenized_text[masked_index] = '[MASK]'
+    indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
+    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+    # Creating a dummy input
+    tokens_tensor = torch.tensor([indexed_tokens])
+    segments_tensors = torch.tensor([segments_ids])
+    dummy_input = [tokens_tensor, segments_tensors]
+
+    # Initializing the model with the torchscript flag
+    # Flag set to True even though it is not necessary as this model does not have an LM Head.
+    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
+        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
+
+    # Instantiating the model
+    model = BertModel(config)
+
+    # The model needs to be in evaluation mode
+    model.eval()
+
+    # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
+    model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
+
+    # Creating the trace
+    traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
+    torch.jit.save(traced_model, "traced_bert.pt")
+
+Loading a model
+------------------------------------------------
+
+This snippet shows how to load the ``BertModel`` that was previously saved to disk under the name ``traced_bert.pt``.
+We are re-using the previously initialised ``dummy_input``.
+
+.. code-block:: python
+
+    loaded_model = torch.jit.load("traced_model.pt")
+    loaded_model.eval()
+
+    all_encoder_layers, pooled_output = loaded_model(dummy_input)
+
+Using a traced model for inference
+------------------------------------------------
+
+Using the traced model for inference is as simple as using its ``__call__`` dunder method:
+
+.. code-block:: python
+
+    traced_model(tokens_tensor, segments_tensors)
@@ -1,318 +0,0 @@
-Training and fine-tuning
-========================
-
-Model classes in 🤗 Transformers are designed to be compatible with native
-PyTorch and TensorFlow 2 and can be used seemlessly with either. In this
-quickstart, we will show how to fine-tune (or train from scratch) a model
-using the standard training tools available in either framework. We will also
-show how to use our included :func:`~transformers.Trainer` class which
-handles much of the complexity of training for you.
-
-This guide assume that you are already familiar with loading and use our
-models for inference; otherwise, see the :doc:`task summary <task_summary>`. We also assume
-that you are familiar with training deep neural networks in either PyTorch or
-TF2, and focus specifically on the nuances and tools for training models in
-🤗 Transformers.
-
-Sections:
-
-  - :ref:`pytorch`
-  - :ref:`tensorflow`
-  - :ref:`trainer`
-  - :ref:`additional-resources`
-
-.. _pytorch:
-
-Fine-tuning in native PyTorch
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Model classes in 🤗 Transformers that don't begin with ``TF`` are
-`PyTorch Modules <https://pytorch.org/docs/master/generated/torch.nn.Module.html>`_,
-meaning that you can use them just as you would any model in PyTorch for
-both inference and optimization.
-
-Let's consider the common task of fine-tuning a masked language model like
-BERT on a sequence classification dataset. When we instantiate a model with
-:func:`~transformers.PreTrainedModel.from_pretrained`, the model
-configuration and pre-trained weights
-of the specified model are used to initialize the model. The
-library also includes a number of task-specific final layers or 'heads' whose
-weights are instantiated randomly when not present in the specified
-pre-trained model. For example, instantiating a model with
-``BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)``
-will create a BERT model instance with encoder weights copied from the
-``bert-base-uncased`` model and a randomly initialized sequence
-classification head on top of the encoder with an output size of 2. Models
-are initialized in ``eval`` mode by default. We can call ``model.train()`` to
-put it in train mode.
-
-.. code-block:: python
-
-    from transformers import BertForSequenceClassification
-    model = BertForSequenceClassification.from_pretrained('bert-base-uncased', return_dict=True)
-    model.train()
-
-This is useful because it allows us to make use of the pre-trained BERT
-encoder and easily train it on whatever sequence classification dataset we
-choose. We can use any PyTorch optimizer, but our library also provides the
-:func:`~transformers.AdamW` optimizer which implements gradient bias
-correction as well as weight decay.
-
-.. code-block:: python
-
-    from transformers import AdamW
-    optimizer = AdamW(model.parameters(), lr=1e-5)
-
-The optimizer allows us to apply different hyperpameters for specific
-parameter groups. For example, we can apply weight decay to all parameters
-other than bias and layer normalization terms:
-
-.. code-block:: python
-
-    no_decay = ['bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
-        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-    ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=1e-5)
-    
-Now we can set up a simple dummy training batch using
-:func:`~transformers.PreTrainedTokenizer.__call__`. This returns a
-:func:`~transformers.BatchEncoding` instance which
-prepares everything we might need to pass to the model.
-
-.. code-block:: python
-
-    from transformers import BertTokenizer
-    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-    text_batch = ["I love Pixar.", "I don't care for Pixar."]
-    encoding = tokenizer(text_batch, return_tensors='pt', padding=True, truncation=True)
-    input_ids = encoding['input_ids']
-    attention_mask = encoding['attention_mask']
-
-When we call a classification model with the ``labels`` argument, the first
-returned element is the Cross Entropy loss between the predictions and the
-passed labels. Having already set up our optimizer, we can then do a
-backwards pass and update the weights:
-
-.. code-block:: python
-
-    labels = torch.tensor([1,0]).unsqueeze(0)
-    outputs = model(input_ids, attention_mask=attention_mask, labels=labels)
-    loss = outputs.loss
-    loss.backward()
-    optimizer.step()
-
-Alternatively, you can just get the logits and calculate the loss yourself.
-The following is equivalent to the previous example:
-
-.. code-block:: python
-
-    from torch.nn import functional as F
-    labels = torch.tensor([1,0]).unsqueeze(0)
-    outputs = model(input_ids, attention_mask=attention_mask)
-    loss = F.cross_entropy(labels, outputs.logitd)
-    loss.backward()
-    optimizer.step()
-
-Of course, you can train on GPU by calling ``to('cuda')`` on the model and
-inputs as usual.
-
-We also provide a few learning rate scheduling tools. With the following, we
-can set up a scheduler which warms up for ``num_warmup_steps`` and then
-linearly decays to 0 by the end of training.
-
-.. code-block:: python
-
-    from transformers import get_linear_schedule_with_warmup
-    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps, num_train_steps)
-
-Then all we have to do is call ``scheduler.step()`` after ``optimizer.step()``.
-
-.. code-block:: python
-
-    loss.backward()
-    optimizer.step()
-    scheduler.step()
-
-We highly recommend using :func:`~transformers.Trainer`, discussed below,
-which conveniently handles the moving parts of training 🤗 Transformers models
-with features like mixed precision and easy tensorboard logging.
-
-
-Freezing the encoder
--------------------
-
-In some cases, you might be interested in keeping the weights of the
-pre-trained encoder frozen and optimizing only the weights of the head
-layers. To do so, simply set the ``requires_grad`` attribute to ``False`` on
-the encoder parameters, which can be accessed with the ``base_model``
-submodule on any task-specific model in the library:
-
-.. code-block:: python
-
-    for param in model.base_model.parameters():
-        param.requires_grad = False
-
-
-.. _tensorflow:
-
-Fine-tuning in native TensorFlow 2
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-Models can also be trained natively in TensorFlow 2. Just as with PyTorch,
-TensorFlow models can be instantiated with
-:func:`~transformers.PreTrainedModel.from_pretrained` to load the weights of
-the encoder from a pretrained model.
-
-.. code-block:: python
-
-    from transformers import TFBertForSequenceClassification
-    model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
-
-Let's use ``tensorflow_datasets`` to load in the `MRPC dataset
-<https://www.tensorflow.org/datasets/catalog/glue#gluemrpc>`_ from GLUE. We
-can then use our built-in
-:func:`~transformers.data.processors.glue.glue_convert_examples_to_features`
-to tokenize MRPC and convert it to a TensorFlow ``Dataset`` object. Note that
-tokenizers are framework-agnostic, so there is no need to prepend ``TF`` to
-the pretrained tokenizer name.
-
-.. code-block:: python
-
-    from transformers import BertTokenizer, glue_convert_examples_to_features
-    import tensorflow as tf
-    import tensorflow_datasets as tfds
-    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-    data = tfds.load('glue/mrpc')
-    train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, max_length=128, task='mrpc')
-    train_dataset = train_dataset.shuffle(100).batch(32).repeat(2)
-
-The model can then be compiled and trained as any Keras model:
-
-.. code-block:: python
-
-    optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5)
-    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
-    model.compile(optimizer=optimizer, loss=loss)
-    model.fit(train_dataset, epochs=2, steps_per_epoch=115)
-
-With the tight interoperability between TensorFlow and PyTorch models, you
-can even save the model and then reload it as a PyTorch model (or vice-versa):
-
-.. code-block:: python
-
-    from transformers import BertForSequenceClassification
-    model.save_pretrained('./my_mrpc_model/')
-    pytorch_model = BertForSequenceClassification.from_pretrained('./my_mrpc_model/', from_tf=True)
-
-
-.. _trainer:
-
-Trainer
-^^^^^^^
-
-We also provide a simple but feature-complete training and evaluation
-interface through :func:`~transformers.Trainer` and
-:func:`~transformers.TFTrainer`. You can train, fine-tune,
-and evaluate any 🤗 Transformers model with a wide range of training options and
-with built-in features like logging, gradient accumulation, and mixed
-precision.
-
-.. code-block:: python
-
-    ## PYTORCH CODE
-    from transformers import BertForSequenceClassification, Trainer, TrainingArguments
-
-    model = BertForSequenceClassification.from_pretrained("bert-large-uncased")
-
-    training_args = TrainingArguments(
-        output_dir='./results',          # output directory
-        num_train_epochs=3,              # total # of training epochs
-        per_device_train_batch_size=16,  # batch size per device during training
-        per_device_eval_batch_size=64,   # batch size for evaluation
-        warmup_steps=500,                # number of warmup steps for learning rate scheduler
-        weight_decay=0.01,               # strength of weight decay
-        logging_dir='./logs',            # directory for storing logs
-    )
-
-    trainer = Trainer(
-        model=model,                         # the instantiated 🤗 Transformers model to be trained
-        args=training_args,                  # training arguments, defined above
-        train_dataset=train_dataset,         # training dataset
-        eval_dataset=test_dataset            # evaluation dataset
-    )
-    ## TENSORFLOW CODE
-    from transformers import TFBertForSequenceClassification, TFTrainer, TFTrainingArguments
-
-    model = TFBertForSequenceClassification.from_pretrained("bert-large-uncased")
-
-    training_args = TFTrainingArguments(
-        output_dir='./results',          # output directory
-        num_train_epochs=3,              # total # of training epochs
-        per_device_train_batch_size=16,  # batch size per device during training
-        per_device_eval_batch_size=64,   # batch size for evaluation
-        warmup_steps=500,                # number of warmup steps for learning rate scheduler
-        weight_decay=0.01,               # strength of weight decay
-        logging_dir='./logs',            # directory for storing logs
-    )
-
-    trainer = TFTrainer(
-        model=model,                         # the instantiated 🤗 Transformers model to be trained
-        args=training_args,                  # training arguments, defined above
-        train_dataset=tfds_train_dataset,    # tensorflow_datasets training dataset
-        eval_dataset=tfds_test_dataset       # tensorflow_datasets evaluation dataset
-    )
-
-Now simply call ``trainer.train()`` to train and ``trainer.evaluate()`` to
-evaluate. You can use your own module as well, but the first
-argument returned from ``forward`` must be the loss which you wish to
-optimize.
-
-:func:`~transformers.Trainer` uses a built-in default function to collate
-batches and prepare them to be fed into the model. If needed, you can also
-use the ``data_collator`` argument to pass your own collator function which
-takes in the data in the format provided by your dataset and returns a
-batch ready to be fed into the model. Note that
-:func:`~transformers.TFTrainer` expects the passed datasets to be dataset
-objects from ``tensorflow_datasets``.
-
-To calculate additional metrics in addition to the loss, you can also define
-your own ``compute_metrics`` function and pass it to the trainer.
-
-.. code-block:: python
-
-    from sklearn.metrics import precision_recall_fscore_support
-
-    def compute_metrics(pred):
-        labels = pred.label_ids
-        preds = pred.predictions.argmax(-1)
-        precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')
-        acc = accuracy_score(labels, preds)
-        return {
-            'accuracy': acc,
-            'f1': f1,
-            'precision': precision,
-            'recall': recall
-        }
-
-Finally, you can view the results, including any calculated metrics, by
-launching tensorboard in your specified ``logging_dir`` directory.
-
-
-.. _additional-resources:
-
-Additional resources
-^^^^^^^^^^^^^^^^^^^^
-
- `A lightweight colab demo <https://colab.research.google.com/drive/1-JIJlao4dI-Ilww_NnTc0rxtp-ymgDgM?usp=sharing>`_
-  which uses ``Trainer`` for IMDb sentiment classification.
-
- `🤗 Transformers Examples <https://github.com/huggingface/transformers/tree/master/examples>`_
-  including scripts for training and fine-tuning on GLUE, SQuAD, and several other tasks.
-
- `How to train a language model <https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb>`_,
-  a detailed colab notebook which uses ``Trainer`` to train a masked language model from scratch on Esperanto.
-
- `🤗 Transformers Notebooks <notebooks.html>`_ which contain dozens of example notebooks from the community for
-  training and using 🤗 Transformers on a variety of tasks.
@@ -1,7 +1,7 @@
-# Examples
+## Examples

-Version 2.9 of 🤗 Transformers introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.
+Version 2.9 of `transformers` introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
+Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.0+.

 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
@@ -13,7 +13,7 @@ Here is the list of all our examples:
 This is still a work-in-progress – in particular documentation is still sparse – so please **contribute improvements/pull requests.**


-## The Big Table of Tasks
+# The Big Table of Tasks

 | Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab
 |---|---|:---:|:---:|:---:|:---:|
@@ -21,11 +21,11 @@ This is still a work-in-progress – in particular documentation is still sparse
 | [**`text-classification`**](https://github.com/huggingface/transformers/tree/master/examples/text-classification)   | GLUE, XNLI      | ✅ | ✅ | ✅ | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/trainer/01_text_classification.ipynb)
 | [**`token-classification`**](https://github.com/huggingface/transformers/tree/master/examples/token-classification) | CoNLL NER       | ✅ | ✅ | ✅ | -
 | [**`multiple-choice`**](https://github.com/huggingface/transformers/tree/master/examples/multiple-choice)           | SWAG, RACE, ARC | ✅ | ✅ | -  | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ViktorAlm/notebooks/blob/master/MPC_GPU_Demo_for_TF_and_PT.ipynb)
-| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | ✅ | ✅ | -  | -
+| [**`question-answering`**](https://github.com/huggingface/transformers/tree/master/examples/question-answering)     | SQuAD           | -  | ✅ | -  | -
 | [**`text-generation`**](https://github.com/huggingface/transformers/tree/master/examples/text-generation)           | -               | n/a | n/a | n/a | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/02_how_to_generate.ipynb)
 | [**`distillation`**](https://github.com/huggingface/transformers/tree/master/examples/distillation)       | All               | -  | -  | -  | -
-| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)     | CNN/Daily Mail    | -  | -  | ✅  | -
-| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)         | WMT               | -  | -  | ✅  | -
+| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/summarization)     | CNN/Daily Mail    | -  | -  | -  | -
+| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/translation)         | WMT               | -  | -  | -  | -
 | [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)             | -                 | -  | -  | -  | -
 | [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)         | HANS              | ✅ | -  | -  | -

@@ -78,50 +78,3 @@ python examples/xla_spawn.py --num_cores 8 \
 ```

 Feedback and more use cases and benchmarks involving TPUs are welcome, please share with the community.
-
-## Logging & Experiment tracking
-
-You can easily log and monitor your runs code. The following are currently supported:
-
-* [TensorBoard](https://www.tensorflow.org/tensorboard)
-* [Weights & Biases](https://docs.wandb.com/library/integrations/huggingface)
-* [Comet ML](https://www.comet.ml/docs/python-sdk/huggingface/)
-
-### Weights & Biases
-
-To use Weights & Biases, install the wandb package with:
-
-```bash
-pip install wandb
-```
-
-Then log in the command line:
-
-```bash
-wandb login
-```
-
-If you are in Jupyter or Colab, you should login with:
-
-```python
-import wandb
-wandb.login()
-```
-
-Whenever you use `Trainer` or `TFTrainer` classes, your losses, evaluation metrics, model topology and gradients (for `Trainer` only) will automatically be logged.
-
-When using 🤗 Transformers with PyTorch Lightning, runs can be tracked through `WandbLogger`. Refer to related [documentation & examples](https://docs.wandb.com/library/integrations/lightning).
-
-### Comet.ml
-
-To use `comet_ml`, install the Python package with:
-
-```bash
-pip install comet_ml
-```
-
-or if in a Conda environment:
-
-```bash
-conda install -c comet_ml -c anaconda -c conda-forge comet_ml
-```
@@ -33,7 +33,7 @@ from transformers import (
    default_data_collator,
    set_seed,
 )
-from utils_hans import HansDataset, InputFeatures, hans_processors, hans_tasks_num_labels
+from utils_hans import HansDataset, InputFeatures, hans_processors


 logger = logging.getLogger(__name__)
@@ -130,7 +130,9 @@ def main():
    set_seed(training_args.seed)

    try:
-        num_labels = hans_tasks_num_labels[data_args.task_name]
+        processor = hans_processors[data_args.task_name]()
+        label_list = processor.get_labels()
+        num_labels = len(label_list)
    except KeyError:
        raise ValueError("Task not found: %s" % (data_args.task_name))

@@ -212,7 +214,6 @@ def main():

        pair_ids = [ex.pairID for ex in eval_dataset]
        output_eval_file = os.path.join(training_args.output_dir, "hans_predictions.txt")
-        label_list = eval_dataset.get_labels()
        if trainer.is_world_master():
            with open(output_eval_file, "w") as writer:
                writer.write("pairID,gold_label\n")
@@ -22,17 +22,7 @@ from typing import List, Optional, Union
 import tqdm
 from filelock import FileLock

-from transformers import (
-    BartTokenizer,
-    BartTokenizerFast,
-    DataProcessor,
-    PreTrainedTokenizer,
-    RobertaTokenizer,
-    RobertaTokenizerFast,
-    XLMRobertaTokenizer,
-    is_tf_available,
-    is_torch_available,
-)
+from transformers import DataProcessor, PreTrainedTokenizer, is_tf_available, is_torch_available


 logger = logging.getLogger(__name__)
@@ -115,17 +105,6 @@ if is_torch_available():
                    "dev" if evaluate else "train", tokenizer.__class__.__name__, str(max_seq_length), task,
                ),
            )
-            label_list = processor.get_labels()
-            if tokenizer.__class__ in (
-                RobertaTokenizer,
-                RobertaTokenizerFast,
-                XLMRobertaTokenizer,
-                BartTokenizer,
-                BartTokenizerFast,
-            ):
-                # HACK(label indices are swapped in RoBERTa pretrained model)
-                label_list[1], label_list[2] = label_list[2], label_list[1]
-            self.label_list = label_list

            # Make sure only the first process in distributed training processes the dataset,
            # and the others will use the cache.
@@ -137,6 +116,7 @@ if is_torch_available():
                    self.features = torch.load(cached_features_file)
                else:
                    logger.info(f"Creating features from dataset file at {data_dir}")
+                    label_list = processor.get_labels()

                    examples = (
                        processor.get_dev_examples(data_dir) if evaluate else processor.get_train_examples(data_dir)
@@ -153,9 +133,6 @@ if is_torch_available():
        def __getitem__(self, i) -> InputFeatures:
            return self.features[i]

-        def get_labels(self):
-            return self.label_list
-

 if is_tf_available():
    import tensorflow as tf
@@ -179,16 +156,6 @@ if is_tf_available():
        ):
            processor = hans_processors[task]()
            label_list = processor.get_labels()
-            if tokenizer.__class__ in (
-                RobertaTokenizer,
-                RobertaTokenizerFast,
-                XLMRobertaTokenizer,
-                BartTokenizer,
-                BartTokenizerFast,
-            ):
-                # HACK(label indices are swapped in RoBERTa pretrained model)
-                label_list[1], label_list[2] = label_list[2], label_list[1]
-            self.label_list = label_list

            examples = processor.get_dev_examples(data_dir) if evaluate else processor.get_train_examples(data_dir)
            self.features = hans_convert_examples_to_features(examples, label_list, max_seq_length, tokenizer)
@@ -239,9 +206,6 @@ if is_tf_available():
        def __getitem__(self, i) -> InputFeatures:
            return self.features[i]

-        def get_labels(self):
-            return self.label_list
-

 class HansProcessor(DataProcessor):
    """Processor for the HANS data set."""
@@ -255,11 +219,7 @@ class HansProcessor(DataProcessor):
        return self._create_examples(self._read_tsv(os.path.join(data_dir, "heuristics_evaluation_set.txt")), "dev")

    def get_labels(self):
-        """See base class.
-        Note that we follow the standard three labels for MNLI
-        (see :class:`~transformers.data.processors.utils.MnliProcessor`)
-        but the HANS evaluation groups `contradiction` and `neutral` into `non-entailment` (label 0) while
-        `entailment` is label 1."""
+        """See base class."""
        return ["contradiction", "entailment", "neutral"]

    def _create_examples(self, lines, set_type):
@@ -272,7 +232,7 @@ class HansProcessor(DataProcessor):
            text_a = line[5]
            text_b = line[6]
            pairID = line[7][2:] if line[7].startswith("ex") else line[7]
-            label = line[0]
+            label = line[-1]
            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, pairID=pairID))
        return examples

@@ -302,13 +262,12 @@ def hans_convert_examples_to_features(
        if ex_index % 10000 == 0:
            logger.info("Writing example %d" % (ex_index))

-        inputs = tokenizer(
+        inputs = tokenizer.encode_plus(
            example.text_a,
            example.text_b,
            add_special_tokens=True,
            max_length=max_length,
-            padding="max_length",
-            truncation=True,
+            pad_to_max_length=True,
            return_overflowing_tokens=True,
        )

@@ -1,10 +0,0 @@
-# 🤗 Benchmark results
-
-Here, you can find a list of the different benchmark results created by the community.
-
-If you would like to list benchmark results on your favorite models of the [model hub](https://huggingface.co/models) here, please open a Pull Request and add it below.
-
-| Benchmark description | Results | Environment info |      Author      |
-|:----------|:-------------|:-------------|------:|
-| PyTorch Benchmark on inference for `bert-base-cased` |[memory](https://github.com/patrickvonplaten/files_to_link_to/blob/master/bert_benchmark/inference_memory.csv) | [env](https://github.com/patrickvonplaten/files_to_link_to/blob/master/bert_benchmark/env.csv) | [Partick von Platen](https://github.com/patrickvonplaten) | 
-| PyTorch Benchmark on inference for `bert-base-cased` |[time](https://github.com/patrickvonplaten/files_to_link_to/blob/master/bert_benchmark/inference_time.csv) | [env](https://github.com/patrickvonplaten/files_to_link_to/blob/master/bert_benchmark/env.csv) | [Partick von Platen](https://github.com/patrickvonplaten) | 
@@ -1,19 +1,14 @@
 import csv
 from collections import defaultdict
 from dataclasses import dataclass, field
-from typing import List, Optional
+from typing import Optional

 import matplotlib.pyplot as plt
 import numpy as np
-from matplotlib.ticker import ScalarFormatter

 from transformers import HfArgumentParser


-def list_field(default=None, metadata=None):
-    return field(default_factory=lambda: default, metadata=metadata)
-
-
@dataclass
 class PlotArguments:
    """
@@ -29,9 +24,6 @@ class PlotArguments:
        default=False,
        metadata={"help": "Whether the csv file has time results or memory results. Defaults to memory results."},
    )
-    no_log_scale: bool = field(
-        default=False, metadata={"help": "Disable logarithmic scale when plotting"},
-    )
    is_train: bool = field(
        default=False,
        metadata={
@@ -41,25 +33,6 @@ class PlotArguments:
    figure_png_file: Optional[str] = field(
        default=None, metadata={"help": "Filename under which the plot will be saved. If unused no plot is saved."},
    )
-    short_model_names: Optional[List[str]] = list_field(
-        default=None, metadata={"help": "List of model names that are used instead of the ones in the csv file."}
-    )
-
-
-def can_convert_to_int(string):
-    try:
-        int(string)
-        return True
-    except ValueError:
-        return False
-
-
-def can_convert_to_float(string):
-    try:
-        float(string)
-        return True
-    except ValueError:
-        return False


 class Plot:
@@ -73,31 +46,16 @@ class Plot:
                model_name = row["model"]
                self.result_dict[model_name]["bsz"].append(int(row["batch_size"]))
                self.result_dict[model_name]["seq_len"].append(int(row["sequence_length"]))
-                if can_convert_to_int(row["result"]):
-                    # value is not None
-                    self.result_dict[model_name]["result"][
-                        (int(row["batch_size"]), int(row["sequence_length"]))
-                    ] = int(row["result"])
-                elif can_convert_to_float(row["result"]):
-                    # value is not None
-                    self.result_dict[model_name]["result"][
-                        (int(row["batch_size"]), int(row["sequence_length"]))
-                    ] = float(row["result"])
+                self.result_dict[model_name]["result"][(int(row["batch_size"]), int(row["sequence_length"]))] = row[
+                    "result"
+                ]

    def plot(self):
        fig, ax = plt.subplots()
        title_str = "Time usage" if self.args.is_time else "Memory usage"
        title_str = title_str + " for training" if self.args.is_train else title_str + " for inference"

-        if not self.args.no_log_scale:
-            # set logarithm scales
-            ax.set_xscale("log")
-            ax.set_yscale("log")
-
-        for axis in [ax.xaxis, ax.yaxis]:
-            axis.set_major_formatter(ScalarFormatter())
-
-        for model_name_idx, model_name in enumerate(self.result_dict.keys()):
+        for model_name in self.result_dict.keys():
            batch_sizes = sorted(list(set(self.result_dict[model_name]["bsz"])))
            sequence_lengths = sorted(list(set(self.result_dict[model_name]["seq_len"])))
            results = self.result_dict[model_name]["result"]
@@ -106,33 +64,28 @@ class Plot:
                (batch_sizes, sequence_lengths) if self.args.plot_along_batch else (sequence_lengths, batch_sizes)
            )

-            label_model_name = (
-                model_name if self.args.short_model_names is None else self.args.short_model_names[model_name_idx]
-            )
+            plt.xlim(min(x_axis_array), max(x_axis_array))

            for inner_loop_value in inner_loop_array:
                if self.args.plot_along_batch:
-                    y_axis_array = np.asarray(
-                        [results[(x, inner_loop_value)] for x in x_axis_array if (x, inner_loop_value) in results],
-                        dtype=np.int,
-                    )
+                    y_axis_array = np.asarray([results[(x, inner_loop_value)] for x in x_axis_array], dtype=np.int)
                else:
-                    y_axis_array = np.asarray(
-                        [results[(inner_loop_value, x)] for x in x_axis_array if (inner_loop_value, x) in results],
-                        dtype=np.float32,
-                    )
+                    y_axis_array = np.asarray([results[(inner_loop_value, x)] for x in x_axis_array], dtype=np.float32)
+
+                ax.set_xscale("log", basex=2)
+                ax.set_yscale("log", basey=10)

                (x_axis_label, inner_loop_label) = (
-                    ("batch_size", "len") if self.args.plot_along_batch else ("in #tokens", "bsz")
+                    ("batch_size", "sequence_length in #tokens")
+                    if self.args.plot_along_batch
+                    else ("sequence_length in #tokens", "batch_size")
                )

-                x_axis_array = np.asarray(x_axis_array, np.int)[: len(y_axis_array)]
-                plt.scatter(
-                    x_axis_array, y_axis_array, label=f"{label_model_name} - {inner_loop_label}: {inner_loop_value}"
-                )
+                x_axis_array = np.asarray(x_axis_array, np.int)
+                plt.scatter(x_axis_array, y_axis_array, label=f"{model_name} - {inner_loop_label}: {inner_loop_value}")
                plt.plot(x_axis_array, y_axis_array, "--")

-            title_str += f" {label_model_name} vs."
+            title_str += f" {model_name} vs."

        title_str = title_str[:-4]
        y_axis_label = "Time in s" if self.args.is_time else "Memory in MB"
@@ -13,15 +13,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Benchmarking the library on inference and training in TensorFlow"""
+""" Benchmarking the library on inference and training in Tensorflow"""

-from transformers import HfArgumentParser, TensorFlowBenchmark, TensorFlowBenchmarkArguments
+from transformers import HfArgumentParser, TensorflowBenchmark, TensorflowBenchmarkArguments


 def main():
-    parser = HfArgumentParser(TensorFlowBenchmarkArguments)
+    parser = HfArgumentParser(TensorflowBenchmarkArguments)
    benchmark_args = parser.parse_args_into_dataclasses()[0]
-    benchmark = TensorFlowBenchmark(args=benchmark_args)
+    benchmark = TensorflowBenchmark(args=benchmark_args)
    benchmark.run()


@@ -226,6 +226,8 @@ def train(args, train_dataset, model, tokenizer):
                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    # Save model checkpoint
                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
@@ -252,15 +254,12 @@ def train(args, train_dataset, model, tokenizer):
    return global_step, tr_loss / global_step


-def evaluate(args, model, tokenizer, prefix="", patience=0):
+def evaluate(args, model, tokenizer, prefix=""):

+    # PABEE STATS
    if args.model_type == "albert":
-        model.albert.set_regression_threshold(args.regression_threshold)
-        model.albert.set_patience(patience)
        model.albert.reset_stats()
    elif args.model_type == "bert":
-        model.bert.set_regression_threshold(args.regression_threshold)
-        model.bert.set_patience(patience)
        model.bert.reset_stats()
    else:
        raise NotImplementedError()
@@ -332,7 +331,7 @@ def evaluate(args, model, tokenizer, prefix="", patience=0):
                print("  %s = %s" % (key, str(result[key])))
                writer.write("%s = %s\n" % (key, str(result[key])))

-    if args.eval_all_checkpoints and patience != 0:
+    if args.eval_all_checkpoints:
        if args.model_type == "albert":
            model.albert.log_stats()
        elif args.model_type == "bert":
@@ -647,6 +646,10 @@ def main():

    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
@@ -687,7 +690,15 @@ def main():

            print(f"Evaluation for checkpoint {prefix}")
            for patience in patience_list:
-                result = evaluate(args, model, tokenizer, prefix=prefix, patience=patience)
+                if args.model_type == "albert":
+                    model.albert.set_regression_threshold(args.regression_threshold)
+                    model.albert.set_patience(patience)
+                elif args.model_type == "bert":
+                    model.bert.set_regression_threshold(args.regression_threshold)
+                    model.bert.set_patience(patience)
+                else:
+                    raise NotImplementedError()
+                result = evaluate(args, model, tokenizer, prefix=prefix)
                result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
                results.update(result)
    return results
@@ -1,6 +1,5 @@
 import argparse
 import logging
-import shutil
 import sys
 import unittest
 from unittest.mock import patch
@@ -20,10 +19,6 @@ def get_setup_file():
    return args.f


-def clean_test_dir(path):
-    shutil.rmtree(path, ignore_errors=True)
-
-
 class PabeeTests(unittest.TestCase):
    def test_run_glue(self):
        stream_handler = logging.StreamHandler(sys.stdout)
@@ -37,6 +32,7 @@ class PabeeTests(unittest.TestCase):
            --task_name mrpc
            --do_train
            --do_eval
+            --output_dir ./tests/fixtures/tests_samples/temp_dir
            --per_gpu_train_batch_size=2
            --per_gpu_eval_batch_size=1
            --learning_rate=2e-5
@@ -45,13 +41,8 @@ class PabeeTests(unittest.TestCase):
            --overwrite_output_dir
            --seed=42
            --max_seq_length=128
-            """
-        output_dir = "./tests/fixtures/tests_samples/temp_dir_{}".format(hash(testargs))
-        testargs += "--output_dir " + output_dir
-        testargs = testargs.split()
+            """.split()
        with patch.object(sys, "argv", testargs):
            result = run_glue_with_pabee.main()
            for value in result.values():
                self.assertGreaterEqual(value, 0.75)
-
-        clean_test_dir(output_dir)
@@ -1,11 +0,0 @@
-# tests directory-specific settings - this file is run automatically
-# by pytest before any tests are run
-
-import sys
-from os.path import abspath, dirname, join
-
-
-# allow having multiple repository checkouts and not needing to remember to rerun
-# 'pip install -e .[dev]' when switching between checkouts and running tests.
-git_repo_path = abspath(join(dirname(dirname(__file__)), "src"))
-sys.path.insert(1, git_repo_path)
@@ -521,6 +521,10 @@ def main():

    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
@@ -383,6 +383,8 @@ def train(args, train_dataset, model, tokenizer):
                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                    # Save model checkpoint
                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
                    model_to_save = (
                        model.module if hasattr(model, "module") else model
                    )  # Take care of distributed/parallel training
@@ -649,6 +651,10 @@ def main():

    # Save the trained model and the tokenizer
    if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
@@ -1,54 +0,0 @@
-# DeeBERT: Early Exiting for *BERT
-
-This is the code base for the paper [DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference](https://www.aclweb.org/anthology/2020.acl-main.204/), modified from its [original code base](https://github.com/castorini/deebert).
-
-The original code base also has information for downloading sample models that we have trained in advance.
-
-## Usage
-
-There are three scripts in the folder which can be run directly.
-
-In each script, there are several things to modify before running:
-
-* `PATH_TO_DATA`: path to the GLUE dataset.
-* `--output_dir`: path for saving fine-tuned models. Default: `./saved_models`.
-* `--plot_data_dir`: path for saving evaluation results. Default: `./results`. Results are printed to stdout and also saved to `npy` files in this directory to facilitate plotting figures and further analyses.
-* `MODEL_TYPE`: bert or roberta
-* `MODEL_SIZE`: base or large
-* `DATASET`: SST-2, MRPC, RTE, QNLI, QQP, or MNLI
-
-#### train_deebert.sh
-
-This is for fine-tuning DeeBERT models.
-
-#### eval_deebert.sh
-
-This is for evaluating each exit layer for fine-tuned DeeBERT models.
-
-#### entropy_eval.sh
-
-This is for evaluating fine-tuned DeeBERT models, given a number of different early exit entropy thresholds.
-
-
-
-## Citation
-
-Please cite our paper if you find the resource useful:
-```
-@inproceedings{xin-etal-2020-deebert,
-    title = "{D}ee{BERT}: Dynamic Early Exiting for Accelerating {BERT} Inference",
-    author = "Xin, Ji  and
-      Tang, Raphael  and
-      Lee, Jaejun  and
-      Yu, Yaoliang  and
-      Lin, Jimmy",
-    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
-    month = jul,
-    year = "2020",
-    address = "Online",
-    publisher = "Association for Computational Linguistics",
-    url = "https://www.aclweb.org/anthology/2020.acl-main.204",
-    pages = "2246--2251",
-}
-```
-
@@ -1,33 +0,0 @@
-#!/bin/bash
-export CUDA_VISIBLE_DEVICES=0
-
-PATH_TO_DATA=/h/xinji/projects/GLUE
-
-MODEL_TYPE=bert  # bert or roberta
-MODEL_SIZE=base  # base or large
-DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
-
-MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
-if [ $MODEL_TYPE = 'bert' ]
-then
-  MODEL_NAME=${MODEL_NAME}-uncased
-fi
-
-ENTROPIES="0 0.1 0.2 0.3 0.4 0.5 0.6 0.7"
-
-for ENTROPY in $ENTROPIES; do
-  python -u run_glue_deebert.py \
-    --model_type $MODEL_TYPE \
-    --model_name_or_path ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
-    --task_name $DATASET \
-    --do_eval \
-    --do_lower_case \
-    --data_dir $PATH_TO_DATA/$DATASET \
-    --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
-    --plot_data_dir ./results/ \
-    --max_seq_length 128 \
-    --early_exit_entropy $ENTROPY \
-    --eval_highway \
-    --overwrite_cache \
-    --per_gpu_eval_batch_size=1
-done
@@ -1,30 +0,0 @@
-#!/bin/bash
-export CUDA_VISIBLE_DEVICES=0
-
-PATH_TO_DATA=/h/xinji/projects/GLUE
-
-MODEL_TYPE=bert  # bert or roberta
-MODEL_SIZE=base  # base or large
-DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
-
-MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
-if [ $MODEL_TYPE = 'bert' ]
-then
-  MODEL_NAME=${MODEL_NAME}-uncased
-fi
-
-
-python -u run_glue_deebert.py  \
-  --model_type $MODEL_TYPE \
-  --model_name_or_path ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
-  --task_name $DATASET \
-  --do_eval \
-  --do_lower_case \
-  --data_dir $PATH_TO_DATA/$DATASET \
-  --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
-  --plot_data_dir ./results/ \
-  --max_seq_length 128 \
-  --eval_each_highway \
-  --eval_highway \
-  --overwrite_cache \
-  --per_gpu_eval_batch_size=1
@@ -1,720 +0,0 @@
-from __future__ import absolute_import, division, print_function
-
-import argparse
-import glob
-import logging
-import os
-import random
-import time
-
-import numpy as np
-import torch
-from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
-from torch.utils.data.distributed import DistributedSampler
-from tqdm import tqdm, trange
-
-from src.modeling_highway_bert import DeeBertForSequenceClassification
-from src.modeling_highway_roberta import DeeRobertaForSequenceClassification
-from transformers import (
-    WEIGHTS_NAME,
-    AdamW,
-    BertConfig,
-    BertTokenizer,
-    RobertaConfig,
-    RobertaTokenizer,
-    get_linear_schedule_with_warmup,
-)
-from transformers import glue_compute_metrics as compute_metrics
-from transformers import glue_convert_examples_to_features as convert_examples_to_features
-from transformers import glue_output_modes as output_modes
-from transformers import glue_processors as processors
-
-
-try:
-    from torch.utils.tensorboard import SummaryWriter
-except ImportError:
-    from tensorboardX import SummaryWriter
-
-
-logger = logging.getLogger(__name__)
-
-
-MODEL_CLASSES = {
-    "bert": (BertConfig, DeeBertForSequenceClassification, BertTokenizer),
-    "roberta": (RobertaConfig, DeeRobertaForSequenceClassification, RobertaTokenizer),
-}
-
-
-def set_seed(args):
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if args.n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-
-def get_wanted_result(result):
-    if "spearmanr" in result:
-        print_result = result["spearmanr"]
-    elif "f1" in result:
-        print_result = result["f1"]
-    elif "mcc" in result:
-        print_result = result["mcc"]
-    elif "acc" in result:
-        print_result = result["acc"]
-    else:
-        raise ValueError("Primary metric unclear in the results")
-    return print_result
-
-
-def train(args, train_dataset, model, tokenizer, train_highway=False):
-    """ Train the model """
-    if args.local_rank in [-1, 0]:
-        tb_writer = SummaryWriter()
-
-    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
-    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
-    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
-
-    if args.max_steps > 0:
-        t_total = args.max_steps
-        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
-    else:
-        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
-
-    # Prepare optimizer and schedule (linear warmup and decay)
-    no_decay = ["bias", "LayerNorm.weight"]
-    if train_highway:
-        optimizer_grouped_parameters = [
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if ("highway" in n) and (not any(nd in n for nd in no_decay))
-                ],
-                "weight_decay": args.weight_decay,
-            },
-            {
-                "params": [
-                    p for n, p in model.named_parameters() if ("highway" in n) and (any(nd in n for nd in no_decay))
-                ],
-                "weight_decay": 0.0,
-            },
-        ]
-    else:
-        optimizer_grouped_parameters = [
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if ("highway" not in n) and (not any(nd in n for nd in no_decay))
-                ],
-                "weight_decay": args.weight_decay,
-            },
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if ("highway" not in n) and (any(nd in n for nd in no_decay))
-                ],
-                "weight_decay": 0.0,
-            },
-        ]
-    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
-    scheduler = get_linear_schedule_with_warmup(
-        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
-    )
-    if args.fp16:
-        try:
-            from apex import amp
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
-        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
-
-    # multi-gpu training (should be after apex fp16 initialization)
-    if args.n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Distributed training (should be after apex fp16 initialization)
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(
-            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
-        )
-
-    # Train!
-    logger.info("***** Running training *****")
-    logger.info("  Num examples = %d", len(train_dataset))
-    logger.info("  Num Epochs = %d", args.num_train_epochs)
-    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
-    logger.info(
-        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
-        args.train_batch_size
-        * args.gradient_accumulation_steps
-        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
-    )
-    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
-    logger.info("  Total optimization steps = %d", t_total)
-
-    global_step = 0
-    tr_loss, logging_loss = 0.0, 0.0
-    model.zero_grad()
-    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
-    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
-    for _ in train_iterator:
-        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
-        for step, batch in enumerate(epoch_iterator):
-            model.train()
-            batch = tuple(t.to(args.device) for t in batch)
-            inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
-            if args.model_type != "distilbert":
-                inputs["token_type_ids"] = (
-                    batch[2] if args.model_type in ["bert", "xlnet"] else None
-                )  # XLM, DistilBERT and RoBERTa don't use segment_ids
-            inputs["train_highway"] = train_highway
-            outputs = model(**inputs)
-            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
-
-            if args.n_gpu > 1:
-                loss = loss.mean()  # mean() to average on multi-gpu parallel training
-            if args.gradient_accumulation_steps > 1:
-                loss = loss / args.gradient_accumulation_steps
-
-            if args.fp16:
-                with amp.scale_loss(loss, optimizer) as scaled_loss:
-                    scaled_loss.backward()
-            else:
-                loss.backward()
-
-            tr_loss += loss.item()
-            if (step + 1) % args.gradient_accumulation_steps == 0:
-                if args.fp16:
-                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
-                else:
-                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
-
-                optimizer.step()
-                scheduler.step()  # Update learning rate schedule
-                model.zero_grad()
-                global_step += 1
-
-                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
-                    # Log metrics
-                    if (
-                        args.local_rank == -1 and args.evaluate_during_training
-                    ):  # Only evaluate when single GPU otherwise metrics may not average well
-                        results = evaluate(args, model, tokenizer)
-                        for key, value in results.items():
-                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
-                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
-                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
-                    logging_loss = tr_loss
-
-                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
-                    # Save model checkpoint
-                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    model_to_save = (
-                        model.module if hasattr(model, "module") else model
-                    )  # Take care of distributed/parallel training
-                    model_to_save.save_pretrained(output_dir)
-                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
-                    logger.info("Saving model checkpoint to %s", output_dir)
-
-            if args.max_steps > 0 and global_step > args.max_steps:
-                epoch_iterator.close()
-                break
-        if args.max_steps > 0 and global_step > args.max_steps:
-            train_iterator.close()
-            break
-
-    if args.local_rank in [-1, 0]:
-        tb_writer.close()
-
-    return global_step, tr_loss / global_step
-
-
-def evaluate(args, model, tokenizer, prefix="", output_layer=-1, eval_highway=False):
-    # Loop to handle MNLI double evaluation (matched, mis-matched)
-    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
-    eval_outputs_dirs = (args.output_dir, args.output_dir + "-MM") if args.task_name == "mnli" else (args.output_dir,)
-
-    results = {}
-    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
-        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
-
-        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(eval_output_dir)
-
-        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
-        # Note that DistributedSampler samples randomly
-        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
-        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-        # multi-gpu eval
-        if args.n_gpu > 1:
-            model = torch.nn.DataParallel(model)
-
-        # Eval!
-        logger.info("***** Running evaluation {} *****".format(prefix))
-        logger.info("  Num examples = %d", len(eval_dataset))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        eval_loss = 0.0
-        nb_eval_steps = 0
-        preds = None
-        out_label_ids = None
-        exit_layer_counter = {(i + 1): 0 for i in range(model.num_layers)}
-        st = time.time()
-        for batch in tqdm(eval_dataloader, desc="Evaluating"):
-            model.eval()
-            batch = tuple(t.to(args.device) for t in batch)
-
-            with torch.no_grad():
-                inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
-                if args.model_type != "distilbert":
-                    inputs["token_type_ids"] = (
-                        batch[2] if args.model_type in ["bert", "xlnet"] else None
-                    )  # XLM, DistilBERT and RoBERTa don't use segment_ids
-                if output_layer >= 0:
-                    inputs["output_layer"] = output_layer
-                outputs = model(**inputs)
-                if eval_highway:
-                    exit_layer_counter[outputs[-1]] += 1
-                tmp_eval_loss, logits = outputs[:2]
-
-                eval_loss += tmp_eval_loss.mean().item()
-            nb_eval_steps += 1
-            if preds is None:
-                preds = logits.detach().cpu().numpy()
-                out_label_ids = inputs["labels"].detach().cpu().numpy()
-            else:
-                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
-                out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
-        eval_time = time.time() - st
-        logger.info("Eval time: {}".format(eval_time))
-
-        eval_loss = eval_loss / nb_eval_steps
-        if args.output_mode == "classification":
-            preds = np.argmax(preds, axis=1)
-        elif args.output_mode == "regression":
-            preds = np.squeeze(preds)
-        result = compute_metrics(eval_task, preds, out_label_ids)
-        results.update(result)
-
-        if eval_highway:
-            logger.info("Exit layer counter: {}".format(exit_layer_counter))
-            actual_cost = sum([l * c for l, c in exit_layer_counter.items()])
-            full_cost = len(eval_dataloader) * model.num_layers
-            logger.info("Expected saving: {}".format(actual_cost / full_cost))
-            if args.early_exit_entropy >= 0:
-                save_fname = (
-                    args.plot_data_dir
-                    + "/"
-                    + args.model_name_or_path[2:]
-                    + "/entropy_{}.npy".format(args.early_exit_entropy)
-                )
-                if not os.path.exists(os.path.dirname(save_fname)):
-                    os.makedirs(os.path.dirname(save_fname))
-                print_result = get_wanted_result(result)
-                np.save(save_fname, np.array([exit_layer_counter, eval_time, actual_cost / full_cost, print_result]))
-                logger.info("Entropy={}\tResult={:.2f}".format(args.early_exit_entropy, 100 * print_result))
-
-        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            logger.info("***** Eval results {} *****".format(prefix))
-            for key in sorted(result.keys()):
-                logger.info("  %s = %s", key, str(result[key]))
-                writer.write("%s = %s\n" % (key, str(result[key])))
-
-    return results
-
-
-def load_and_cache_examples(args, task, tokenizer, evaluate=False):
-    if args.local_rank not in [-1, 0] and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    processor = processors[task]()
-    output_mode = output_modes[task]
-    # Load data features from cache or dataset file
-    cached_features_file = os.path.join(
-        args.data_dir,
-        "cached_{}_{}_{}_{}".format(
-            "dev" if evaluate else "train",
-            list(filter(None, args.model_name_or_path.split("/"))).pop(),
-            str(args.max_seq_length),
-            str(task),
-        ),
-    )
-    if os.path.exists(cached_features_file) and not args.overwrite_cache:
-        logger.info("Loading features from cached file %s", cached_features_file)
-        features = torch.load(cached_features_file)
-    else:
-        logger.info("Creating features from dataset file at %s", args.data_dir)
-        label_list = processor.get_labels()
-        if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta"]:
-            # HACK(label indices are swapped in RoBERTa pretrained model)
-            label_list[1], label_list[2] = label_list[2], label_list[1]
-        examples = (
-            processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
-        )
-        features = convert_examples_to_features(
-            examples, tokenizer, label_list=label_list, max_length=args.max_seq_length, output_mode=output_mode,
-        )
-        if args.local_rank in [-1, 0]:
-            logger.info("Saving features into cached file %s", cached_features_file)
-            torch.save(features, cached_features_file)
-
-    if args.local_rank == 0 and not evaluate:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
-
-    # Convert to Tensors and build dataset
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
-
-    if features[0].token_type_ids is None:
-        # For RoBERTa (a potential bug!)
-        all_token_type_ids = torch.tensor([[0] * args.max_seq_length for f in features], dtype=torch.long)
-    else:
-        all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
-    if output_mode == "classification":
-        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
-    elif output_mode == "regression":
-        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
-
-    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
-    return dataset
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    # Required parameters
-    parser.add_argument(
-        "--data_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
-    )
-    parser.add_argument(
-        "--model_type",
-        default=None,
-        type=str,
-        required=True,
-        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
-    )
-    parser.add_argument(
-        "--model_name_or_path",
-        default=None,
-        type=str,
-        required=True,
-        help="Path to pre-trained model or shortcut name.",
-    )
-    parser.add_argument(
-        "--task_name",
-        default=None,
-        type=str,
-        required=True,
-        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
-    )
-    parser.add_argument(
-        "--output_dir",
-        default=None,
-        type=str,
-        required=True,
-        help="The output directory where the model predictions and checkpoints will be written.",
-    )
-    parser.add_argument(
-        "--plot_data_dir",
-        default="./plotting/",
-        type=str,
-        required=False,
-        help="The directory to store data for plotting figures.",
-    )
-
-    # Other parameters
-    parser.add_argument(
-        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
-    )
-    parser.add_argument(
-        "--tokenizer_name",
-        default="",
-        type=str,
-        help="Pretrained tokenizer name or path if not the same as model_name",
-    )
-    parser.add_argument(
-        "--cache_dir",
-        default="",
-        type=str,
-        help="Where do you want to store the pre-trained models downloaded from s3",
-    )
-    parser.add_argument(
-        "--max_seq_length",
-        default=128,
-        type=int,
-        help="The maximum total input sequence length after tokenization. Sequences longer "
-        "than this will be truncated, sequences shorter will be padded.",
-    )
-    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
-    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
-    parser.add_argument(
-        "--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
-    )
-    parser.add_argument(
-        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
-    )
-    parser.add_argument("--eval_each_highway", action="store_true", help="Set this flag to evaluate each highway.")
-    parser.add_argument(
-        "--eval_after_first_stage",
-        action="store_true",
-        help="Set this flag to evaluate after training only bert (not highway).",
-    )
-    parser.add_argument("--eval_highway", action="store_true", help="Set this flag if it's evaluating highway models")
-
-    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
-    parser.add_argument(
-        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
-    )
-    parser.add_argument(
-        "--gradient_accumulation_steps",
-        type=int,
-        default=1,
-        help="Number of updates steps to accumulate before performing a backward/update pass.",
-    )
-    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
-    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.")
-    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
-    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
-    parser.add_argument(
-        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
-    )
-    parser.add_argument(
-        "--max_steps",
-        default=-1,
-        type=int,
-        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
-    )
-    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
-    parser.add_argument("--early_exit_entropy", default=-1, type=float, help="Entropy threshold for early exit.")
-
-    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
-    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
-    parser.add_argument(
-        "--eval_all_checkpoints",
-        action="store_true",
-        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
-    )
-    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
-    parser.add_argument(
-        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
-    )
-    parser.add_argument(
-        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
-    )
-    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
-
-    parser.add_argument(
-        "--fp16",
-        action="store_true",
-        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
-    )
-    parser.add_argument(
-        "--fp16_opt_level",
-        type=str,
-        default="O1",
-        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
-        "See details at https://nvidia.github.io/apex/amp.html",
-    )
-    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
-    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
-    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
-    args = parser.parse_args()
-
-    if (
-        os.path.exists(args.output_dir)
-        and os.listdir(args.output_dir)
-        and args.do_train
-        and not args.overwrite_output_dir
-    ):
-        raise ValueError(
-            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
-                args.output_dir
-            )
-        )
-
-    # Setup distant debugging if needed
-    if args.server_ip and args.server_port:
-        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
-        import ptvsd
-
-        print("Waiting for debugger attach")
-        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
-        ptvsd.wait_for_attach()
-
-    # Setup CUDA, GPU & distributed training
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        args.n_gpu = torch.cuda.device_count()
-    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        torch.distributed.init_process_group(backend="nccl")
-        args.n_gpu = 1
-    args.device = device
-
-    # Setup logging
-    logging.basicConfig(
-        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
-        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
-    )
-    logger.warning(
-        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        args.local_rank,
-        device,
-        args.n_gpu,
-        bool(args.local_rank != -1),
-        args.fp16,
-    )
-
-    # Set seed
-    set_seed(args)
-
-    # Prepare GLUE task
-    args.task_name = args.task_name.lower()
-    if args.task_name not in processors:
-        raise ValueError("Task not found: %s" % (args.task_name))
-    processor = processors[args.task_name]()
-    args.output_mode = output_modes[args.task_name]
-    label_list = processor.get_labels()
-    num_labels = len(label_list)
-
-    # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    args.model_type = args.model_type.lower()
-    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
-    config = config_class.from_pretrained(
-        args.config_name if args.config_name else args.model_name_or_path,
-        num_labels=num_labels,
-        finetuning_task=args.task_name,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-    tokenizer = tokenizer_class.from_pretrained(
-        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
-        do_lower_case=args.do_lower_case,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-    model = model_class.from_pretrained(
-        args.model_name_or_path,
-        from_tf=bool(".ckpt" in args.model_name_or_path),
-        config=config,
-        cache_dir=args.cache_dir if args.cache_dir else None,
-    )
-
-    if args.model_type == "bert":
-        model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy)
-        model.bert.init_highway_pooler()
-    elif args.model_type == "roberta":
-        model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy)
-        model.roberta.init_highway_pooler()
-    else:
-        raise NotImplementedError()
-
-    if args.local_rank == 0:
-        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
-
-    model.to(args.device)
-
-    logger.info("Training/evaluation parameters %s", args)
-
-    # Training
-    if args.do_train:
-        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
-        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
-        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
-
-        if args.eval_after_first_stage:
-            result = evaluate(args, model, tokenizer, prefix="")
-            print_result = get_wanted_result(result)
-
-        train(args, train_dataset, model, tokenizer, train_highway=True)
-
-    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        model_to_save = (
-            model.module if hasattr(model, "module") else model
-        )  # Take care of distributed/parallel training
-        model_to_save.save_pretrained(args.output_dir)
-        tokenizer.save_pretrained(args.output_dir)
-
-        # Good practice: save your training arguments together with the trained model
-        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
-
-        # Load a trained model and vocabulary that you have fine-tuned
-        model = model_class.from_pretrained(args.output_dir)
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
-        model.to(args.device)
-
-    # Evaluation
-    results = {}
-    if args.do_eval and args.local_rank in [-1, 0]:
-        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(
-                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
-            )
-            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
-        logger.info("Evaluate the following checkpoints: %s", checkpoints)
-        for checkpoint in checkpoints:
-            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
-            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
-
-            model = model_class.from_pretrained(checkpoint)
-            if args.model_type == "bert":
-                model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy)
-            elif args.model_type == "roberta":
-                model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy)
-            else:
-                raise NotImplementedError()
-
-            model.to(args.device)
-            result = evaluate(args, model, tokenizer, prefix=prefix, eval_highway=args.eval_highway)
-            print_result = get_wanted_result(result)
-            logger.info("Result: {}".format(print_result))
-            if args.eval_each_highway:
-                last_layer_results = print_result
-                each_layer_results = []
-                for i in range(model.num_layers):
-                    logger.info("\n")
-                    _result = evaluate(
-                        args, model, tokenizer, prefix=prefix, output_layer=i, eval_highway=args.eval_highway
-                    )
-                    if i + 1 < model.num_layers:
-                        each_layer_results.append(get_wanted_result(_result))
-                each_layer_results.append(last_layer_results)
-                save_fname = args.plot_data_dir + "/" + args.model_name_or_path[2:] + "/each_layer.npy"
-                if not os.path.exists(os.path.dirname(save_fname)):
-                    os.makedirs(os.path.dirname(save_fname))
-                np.save(save_fname, np.array(each_layer_results))
-                info_str = "Score of each layer:"
-                for i in range(model.num_layers):
-                    info_str += " {:.2f}".format(100 * each_layer_results[i])
-                logger.info(info_str)
-            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
-            results.update(result)
-
-    return results
-
-
-if __name__ == "__main__":
-    main()
@@ -1,396 +0,0 @@
-import torch
-from torch import nn
-from torch.nn import CrossEntropyLoss, MSELoss
-
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from transformers.modeling_bert import (
-    BERT_INPUTS_DOCSTRING,
-    BERT_START_DOCSTRING,
-    BertEmbeddings,
-    BertLayer,
-    BertPooler,
-    BertPreTrainedModel,
-)
-
-
-def entropy(x):
-    """ Calculate entropy of a pre-softmax logit Tensor
-    """
-    exp_x = torch.exp(x)
-    A = torch.sum(exp_x, dim=1)  # sum of exp(x_i)
-    B = torch.sum(x * exp_x, dim=1)  # sum of x_i * exp(x_i)
-    return torch.log(A) - B / A
-
-
-class DeeBertEncoder(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.output_attentions = config.output_attentions
-        self.output_hidden_states = config.output_hidden_states
-        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
-        self.highway = nn.ModuleList([BertHighway(config) for _ in range(config.num_hidden_layers)])
-
-        self.early_exit_entropy = [-1 for _ in range(config.num_hidden_layers)]
-
-    def set_early_exit_entropy(self, x):
-        if (type(x) is float) or (type(x) is int):
-            for i in range(len(self.early_exit_entropy)):
-                self.early_exit_entropy[i] = x
-        else:
-            self.early_exit_entropy = x
-
-    def init_highway_pooler(self, pooler):
-        loaded_model = pooler.state_dict()
-        for highway in self.highway:
-            for name, param in highway.pooler.state_dict().items():
-                param.copy_(loaded_model[name])
-
-    def forward(
-        self,
-        hidden_states,
-        attention_mask=None,
-        head_mask=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-    ):
-        all_hidden_states = ()
-        all_attentions = ()
-        all_highway_exits = ()
-        for i, layer_module in enumerate(self.layer):
-            if self.output_hidden_states:
-                all_hidden_states = all_hidden_states + (hidden_states,)
-
-            layer_outputs = layer_module(
-                hidden_states, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask
-            )
-            hidden_states = layer_outputs[0]
-
-            if self.output_attentions:
-                all_attentions = all_attentions + (layer_outputs[1],)
-
-            current_outputs = (hidden_states,)
-            if self.output_hidden_states:
-                current_outputs = current_outputs + (all_hidden_states,)
-            if self.output_attentions:
-                current_outputs = current_outputs + (all_attentions,)
-
-            highway_exit = self.highway[i](current_outputs)
-            # logits, pooled_output
-
-            if not self.training:
-                highway_logits = highway_exit[0]
-                highway_entropy = entropy(highway_logits)
-                highway_exit = highway_exit + (highway_entropy,)  # logits, hidden_states(?), entropy
-                all_highway_exits = all_highway_exits + (highway_exit,)
-
-                if highway_entropy < self.early_exit_entropy[i]:
-                    new_output = (highway_logits,) + current_outputs[1:] + (all_highway_exits,)
-                    raise HighwayException(new_output, i + 1)
-            else:
-                all_highway_exits = all_highway_exits + (highway_exit,)
-
-        # Add last layer
-        if self.output_hidden_states:
-            all_hidden_states = all_hidden_states + (hidden_states,)
-
-        outputs = (hidden_states,)
-        if self.output_hidden_states:
-            outputs = outputs + (all_hidden_states,)
-        if self.output_attentions:
-            outputs = outputs + (all_attentions,)
-
-        outputs = outputs + (all_highway_exits,)
-        return outputs  # last-layer hidden state, (all hidden states), (all attentions), all highway exits
-
-
-@add_start_docstrings(
-    "The Bert Model transformer with early exiting (DeeBERT). ", BERT_START_DOCSTRING,
-)
-class DeeBertModel(BertPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-        self.config = config
-
-        self.embeddings = BertEmbeddings(config)
-        self.encoder = DeeBertEncoder(config)
-        self.pooler = BertPooler(config)
-
-        self.init_weights()
-
-    def init_highway_pooler(self):
-        self.encoder.init_highway_pooler(self.pooler)
-
-    def get_input_embeddings(self):
-        return self.embeddings.word_embeddings
-
-    def set_input_embeddings(self, value):
-        self.embeddings.word_embeddings = value
-
-    def _prune_heads(self, heads_to_prune):
-        """ Prunes heads of the model.
-            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
-            See base class PreTrainedModel
-        """
-        for layer, heads in heads_to_prune.items():
-            self.encoder.layer[layer].attention.prune_heads(heads)
-
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        encoder_hidden_states=None,
-        encoder_attention_mask=None,
-    ):
-        r"""
-    Return:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
-            Sequence of hidden-states at the output of the last layer of the model.
-        pooler_output (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
-            Last layer hidden-state of the first token of the sequence (classification token)
-            further processed by a Linear layer and a Tanh activation function. The Linear
-            layer weights are trained from the next sentence prediction (classification)
-            objective during pre-training.
-
-            This output is usually *not* a good summary
-            of the semantic content of the input, you're often better with averaging or pooling
-            the sequence of hidden-states for the whole input sequence.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
-            Tuple of each early exit's results (total length: number of layers)
-            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
-        """
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
-        elif input_ids is not None:
-            input_shape = input_ids.size()
-        elif inputs_embeds is not None:
-            input_shape = inputs_embeds.size()[:-1]
-        else:
-            raise ValueError("You have to specify either input_ids or inputs_embeds")
-
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-        if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
-        if encoder_attention_mask is None:
-            encoder_attention_mask = torch.ones(input_shape, device=device)
-        if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
-
-        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
-        # ourselves in which case we just need to make it broadcastable to all heads.
-        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
-
-        # If a 2D ou 3D attention mask is provided for the cross-attention
-        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
-        if encoder_attention_mask.dim() == 3:
-            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
-        if encoder_attention_mask.dim() == 2:
-            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
-
-        encoder_extended_attention_mask = encoder_extended_attention_mask.to(
-            dtype=next(self.parameters()).dtype
-        )  # fp16 compatibility
-        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
-
-        # Prepare head mask if needed
-        # 1.0 in head_mask indicate we keep the head
-        # attention_probs has shape bsz x n_heads x N x N
-        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
-        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
-        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
-
-        embedding_output = self.embeddings(
-            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
-        )
-        encoder_outputs = self.encoder(
-            embedding_output,
-            attention_mask=extended_attention_mask,
-            head_mask=head_mask,
-            encoder_hidden_states=encoder_hidden_states,
-            encoder_attention_mask=encoder_extended_attention_mask,
-        )
-        sequence_output = encoder_outputs[0]
-        pooled_output = self.pooler(sequence_output)
-
-        outputs = (sequence_output, pooled_output,) + encoder_outputs[
-            1:
-        ]  # add hidden_states and attentions if they are here
-        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions), highway exits
-
-
-class HighwayException(Exception):
-    def __init__(self, message, exit_layer):
-        self.message = message
-        self.exit_layer = exit_layer  # start from 1!
-
-
-class BertHighway(nn.Module):
-    """A module to provide a shortcut
-    from (the output of one non-final BertLayer in BertEncoder) to (cross-entropy computation in BertForSequenceClassification)
-    """
-
-    def __init__(self, config):
-        super().__init__()
-        self.pooler = BertPooler(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
-
-    def forward(self, encoder_outputs):
-        # Pooler
-        pooler_input = encoder_outputs[0]
-        pooler_output = self.pooler(pooler_input)
-        # "return" pooler_output
-
-        # BertModel
-        bmodel_output = (pooler_input, pooler_output) + encoder_outputs[1:]
-        # "return" bodel_output
-
-        # Dropout and classification
-        pooled_output = bmodel_output[1]
-
-        pooled_output = self.dropout(pooled_output)
-        logits = self.classifier(pooled_output)
-
-        return logits, pooled_output
-
-
-@add_start_docstrings(
-    """Bert Model (with early exiting - DeeBERT) with a classifier on top,
-    also takes care of multi-layer training. """,
-    BERT_START_DOCSTRING,
-)
-class DeeBertForSequenceClassification(BertPreTrainedModel):
-    def __init__(self, config):
-        super().__init__(config)
-        self.num_labels = config.num_labels
-        self.num_layers = config.num_hidden_layers
-
-        self.bert = DeeBertModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
-
-        self.init_weights()
-
-    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-        output_layer=-1,
-        train_highway=False,
-    ):
-        r"""
-        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-            Labels for computing the sequence classification/regression loss.
-            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
-            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
-            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
-        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
-            Classification (or regression if config.num_labels==1) loss.
-        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
-            Tuple of each early exit's results (total length: number of layers)
-            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
-        """
-
-        exit_layer = self.num_layers
-        try:
-            outputs = self.bert(
-                input_ids,
-                attention_mask=attention_mask,
-                token_type_ids=token_type_ids,
-                position_ids=position_ids,
-                head_mask=head_mask,
-                inputs_embeds=inputs_embeds,
-            )
-            # sequence_output, pooled_output, (hidden_states), (attentions), highway exits
-
-            pooled_output = outputs[1]
-
-            pooled_output = self.dropout(pooled_output)
-            logits = self.classifier(pooled_output)
-            outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-        except HighwayException as e:
-            outputs = e.message
-            exit_layer = e.exit_layer
-            logits = outputs[0]
-
-        if not self.training:
-            original_entropy = entropy(logits)
-            highway_entropy = []
-            highway_logits_all = []
-        if labels is not None:
-            if self.num_labels == 1:
-                #  We are doing regression
-                loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1), labels.view(-1))
-            else:
-                loss_fct = CrossEntropyLoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-
-            # work with highway exits
-            highway_losses = []
-            for highway_exit in outputs[-1]:
-                highway_logits = highway_exit[0]
-                if not self.training:
-                    highway_logits_all.append(highway_logits)
-                    highway_entropy.append(highway_exit[2])
-                if self.num_labels == 1:
-                    #  We are doing regression
-                    loss_fct = MSELoss()
-                    highway_loss = loss_fct(highway_logits.view(-1), labels.view(-1))
-                else:
-                    loss_fct = CrossEntropyLoss()
-                    highway_loss = loss_fct(highway_logits.view(-1, self.num_labels), labels.view(-1))
-                highway_losses.append(highway_loss)
-
-            if train_highway:
-                outputs = (sum(highway_losses[:-1]),) + outputs
-                # exclude the final highway, of course
-            else:
-                outputs = (loss,) + outputs
-        if not self.training:
-            outputs = outputs + ((original_entropy, highway_entropy), exit_layer)
-            if output_layer >= 0:
-                outputs = (
-                    (outputs[0],) + (highway_logits_all[output_layer],) + outputs[2:]
-                )  # use the highway of the last layer
-
-        return outputs  # (loss), logits, (hidden_states), (attentions), (highway_exits)
@@ -1,151 +0,0 @@
-from __future__ import absolute_import, division, print_function, unicode_literals
-
-import torch.nn as nn
-from torch.nn import CrossEntropyLoss, MSELoss
-
-from transformers.configuration_roberta import RobertaConfig
-from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
-from transformers.modeling_roberta import ROBERTA_INPUTS_DOCSTRING, ROBERTA_START_DOCSTRING, RobertaEmbeddings
-
-from .modeling_highway_bert import BertPreTrainedModel, DeeBertModel, HighwayException, entropy
-
-
-@add_start_docstrings(
-    "The RoBERTa Model transformer with early exiting (DeeRoBERTa). ", ROBERTA_START_DOCSTRING,
-)
-class DeeRobertaModel(DeeBertModel):
-
-    config_class = RobertaConfig
-    base_model_prefix = "roberta"
-
-    def __init__(self, config):
-        super().__init__(config)
-
-        self.embeddings = RobertaEmbeddings(config)
-        self.init_weights()
-
-
-@add_start_docstrings(
-    """RoBERTa Model (with early exiting - DeeRoBERTa) with a classifier on top,
-    also takes care of multi-layer training. """,
-    ROBERTA_START_DOCSTRING,
-)
-class DeeRobertaForSequenceClassification(BertPreTrainedModel):
-
-    config_class = RobertaConfig
-    base_model_prefix = "roberta"
-
-    def __init__(self, config):
-        super().__init__(config)
-        self.num_labels = config.num_labels
-        self.num_layers = config.num_hidden_layers
-
-        self.roberta = DeeRobertaModel(config)
-        self.dropout = nn.Dropout(config.hidden_dropout_prob)
-        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
-
-    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        labels=None,
-        output_layer=-1,
-        train_highway=False,
-    ):
-        r"""
-        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
-            Labels for computing the sequence classification/regression loss.
-            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
-            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
-            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
-
-    Returns:
-        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
-        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
-            Classification (or regression if config.num_labels==1) loss.
-        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
-            Classification (or regression if config.num_labels==1) scores (before SoftMax).
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
-            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
-
-            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
-            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
-
-            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
-            heads.
-        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
-            Tuple of each early exit's results (total length: number of layers)
-            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
-        """
-
-        exit_layer = self.num_layers
-        try:
-            outputs = self.roberta(
-                input_ids,
-                attention_mask=attention_mask,
-                token_type_ids=token_type_ids,
-                position_ids=position_ids,
-                head_mask=head_mask,
-                inputs_embeds=inputs_embeds,
-            )
-
-            pooled_output = outputs[1]
-
-            pooled_output = self.dropout(pooled_output)
-            logits = self.classifier(pooled_output)
-            outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
-        except HighwayException as e:
-            outputs = e.message
-            exit_layer = e.exit_layer
-            logits = outputs[0]
-
-        if not self.training:
-            original_entropy = entropy(logits)
-            highway_entropy = []
-            highway_logits_all = []
-        if labels is not None:
-            if self.num_labels == 1:
-                #  We are doing regression
-                loss_fct = MSELoss()
-                loss = loss_fct(logits.view(-1), labels.view(-1))
-            else:
-                loss_fct = CrossEntropyLoss()
-                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
-
-            # work with highway exits
-            highway_losses = []
-            for highway_exit in outputs[-1]:
-                highway_logits = highway_exit[0]
-                if not self.training:
-                    highway_logits_all.append(highway_logits)
-                    highway_entropy.append(highway_exit[2])
-                if self.num_labels == 1:
-                    #  We are doing regression
-                    loss_fct = MSELoss()
-                    highway_loss = loss_fct(highway_logits.view(-1), labels.view(-1))
-                else:
-                    loss_fct = CrossEntropyLoss()
-                    highway_loss = loss_fct(highway_logits.view(-1, self.num_labels), labels.view(-1))
-                highway_losses.append(highway_loss)
-
-            if train_highway:
-                outputs = (sum(highway_losses[:-1]),) + outputs
-                # exclude the final highway, of course
-            else:
-                outputs = (loss,) + outputs
-        if not self.training:
-            outputs = outputs + ((original_entropy, highway_entropy), exit_layer)
-            if output_layer >= 0:
-                outputs = (
-                    (outputs[0],) + (highway_logits_all[output_layer],) + outputs[2:]
-                )  # use the highway of the last layer
-
-        return outputs  # (loss), logits, (hidden_states), (attentions), entropy
@@ -1,98 +0,0 @@
-import argparse
-import logging
-import sys
-import unittest
-from unittest.mock import patch
-
-import run_glue_deebert
-from transformers.testing_utils import slow
-
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-
-
-def get_setup_file():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("-f")
-    args = parser.parse_args()
-    return args.f
-
-
-class DeeBertTests(unittest.TestCase):
-    def setup(self) -> None:
-        stream_handler = logging.StreamHandler(sys.stdout)
-        logger.addHandler(stream_handler)
-
-    @slow
-    def test_glue_deebert_train(self):
-
-        train_args = """
-            run_glue_deebert.py
-            --model_type roberta
-            --model_name_or_path roberta-base
-            --task_name MRPC
-            --do_train
-            --do_eval
-            --do_lower_case
-            --data_dir ./tests/fixtures/tests_samples/MRPC/
-            --max_seq_length 128
-            --per_gpu_eval_batch_size=1
-            --per_gpu_train_batch_size=8
-            --learning_rate 2e-4
-            --num_train_epochs 3
-            --overwrite_output_dir
-            --seed 42
-            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
-            --plot_data_dir ./examples/deebert/results/
-            --save_steps 0
-            --overwrite_cache
-            --eval_after_first_stage
-            """.split()
-        with patch.object(sys, "argv", train_args):
-            result = run_glue_deebert.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)
-
-        eval_args = """
-            run_glue_deebert.py
-            --model_type roberta
-            --model_name_or_path ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
-            --task_name MRPC
-            --do_eval
-            --do_lower_case
-            --data_dir ./tests/fixtures/tests_samples/MRPC/
-            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
-            --plot_data_dir ./examples/deebert/results/
-            --max_seq_length 128
-            --eval_each_highway
-            --eval_highway
-            --overwrite_cache
-            --per_gpu_eval_batch_size=1
-            """.split()
-        with patch.object(sys, "argv", eval_args):
-            result = run_glue_deebert.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)
-
-        entropy_eval_args = """
-            run_glue_deebert.py
-            --model_type roberta
-            --model_name_or_path ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
-            --task_name MRPC
-            --do_eval
-            --do_lower_case
-            --data_dir ./tests/fixtures/tests_samples/MRPC/
-            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
-            --plot_data_dir ./examples/deebert/results/
-            --max_seq_length 128
-            --early_exit_entropy 0.1
-            --eval_highway
-            --overwrite_cache
-            --per_gpu_eval_batch_size=1
-            """.split()
-        with patch.object(sys, "argv", entropy_eval_args):
-            result = run_glue_deebert.main()
-            for value in result.values():
-                self.assertGreaterEqual(value, 0.666)
@@ -1,38 +0,0 @@
-#!/bin/bash
-export CUDA_VISIBLE_DEVICES=0
-
-PATH_TO_DATA=/h/xinji/projects/GLUE
-
-MODEL_TYPE=bert  # bert or roberta
-MODEL_SIZE=base  # base or large
-DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
-
-MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
-EPOCHS=10
-if [ $MODEL_TYPE = 'bert' ]
-then
-  EPOCHS=3
-  MODEL_NAME=${MODEL_NAME}-uncased
-fi
-
-
-python -u run_glue_deebert.py \
-  --model_type $MODEL_TYPE \
-  --model_name_or_path $MODEL_NAME \
-  --task_name $DATASET \
-  --do_train \
-  --do_eval \
-  --do_lower_case \
-  --data_dir $PATH_TO_DATA/$DATASET \
-  --max_seq_length 128 \
-  --per_gpu_eval_batch_size=1 \
-  --per_gpu_train_batch_size=8 \
-  --learning_rate 2e-5 \
-  --num_train_epochs $EPOCHS \
-  --overwrite_output_dir \
-  --seed 42 \
-  --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
-  --plot_data_dir ./results/ \
-  --save_steps 0 \
-  --overwrite_cache \
-  --eval_after_first_stage
@@ -55,7 +55,7 @@ Here are the results on the *test* sets for 6 of the languages available in XNLI

 ## Setup

-This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`.
+This part of the library has only be tested with Python3.6+. There are few specific dependencies to install before launching a distillation, you can install them with the command `pip install -r requirements.txt`. 

 **Important note:** The training scripts have been updated to support PyTorch v1.2.0 (there are breakings changes compared to v1.1.0).

@@ -161,7 +161,7 @@ python -m torch.distributed.launch \
    --master_port $MASTER_PORT \
    train.py \
        --force \
-        --gpus $WORLD_SIZE \
+        --n_gpu $WORLD_SIZE \
        --student_type distilbert \
        --student_config training_configs/distilbert-base-uncased.json \
        --teacher_type bert \
@@ -809,6 +809,10 @@ def main():

    # Save the trained model and the tokenizer
    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
        logger.info("Saving model checkpoint to %s", args.output_dir)
        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
        # They can then be reloaded using `from_pretrained()`
@@ -210,7 +210,7 @@ def main():
        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
        "See details at https://nvidia.github.io/apex/amp.html",
    )
-    parser.add_argument("--gpus", type=int, default=1, help="Number of GPUs in the node.")
+    parser.add_argument("--n_gpu", type=int, default=1, help="Number of GPUs in the node.")
    parser.add_argument("--local_rank", type=int, default=-1, help="Distributed training - Local rank")
    parser.add_argument("--seed", type=int, default=56, help="Random seed")

@@ -60,27 +60,3 @@ python run_language_modeling.py \
    --mlm
 ```

-### XLNet and permutation language modeling
-
-XLNet uses a different training objective, which is permutation language modeling. It is an autoregressive method 
-to learn bidirectional contexts by maximizing the expected likelihood over all permutations of the input 
-sequence factorization order.
-
-We use the `--plm_probability` flag to define the ratio of length of a span of masked tokens to surrounding 
-context length for permutation language modeling.
-
-The `--max_span_length` flag may also be used to limit the length of a span of masked tokens used 
-for permutation language modeling.
-
-```bash
-export TRAIN_FILE=/path/to/dataset/wiki.train.raw
-export TEST_FILE=/path/to/dataset/wiki.test.raw
-
-python run_language_modeling.py \
-    --output_dir=output \
-    --model_name_or_path=xlnet-base-cased \
-    --do_train \
-    --train_data_file=$TRAIN_FILE \
-    --do_eval \
-    --eval_data_file=$TEST_FILE \
-```
@@ -14,9 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, CTRL, BERT, RoBERTa, XLNet).
-GPT, GPT-2 and CTRL are fine-tuned using a causal language modeling (CLM) loss. BERT and RoBERTa are fine-tuned
-using a masked language modeling (MLM) loss. XLNet is fine-tuned using a permutation language modeling (PLM) loss.
+Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
+GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
+using a masked language modeling (MLM) loss.
 """


@@ -33,7 +33,6 @@ from transformers import (
    AutoModelWithLMHead,
    AutoTokenizer,
    DataCollatorForLanguageModeling,
-    DataCollatorForPermutationLanguageModeling,
    HfArgumentParser,
    LineByLineTextDataset,
    PreTrainedTokenizer,
@@ -102,15 +101,6 @@ class DataTrainingArguments:
    mlm_probability: float = field(
        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
    )
-    plm_probability: float = field(
-        default=1 / 6,
-        metadata={
-            "help": "Ratio of length of a span of masked tokens to surrounding context length for permutation language modeling."
-        },
-    )
-    max_span_length: int = field(
-        default=5, metadata={"help": "Maximum length of a span of masked tokens for permutation language modeling."}
-    )

    block_size: int = field(
        default=-1,
@@ -217,8 +207,8 @@ def main():

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
-            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
-            "--mlm flag (masked language modeling)."
+            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
+            "flag (masked language modeling)."
        )

    if data_args.block_size <= 0:
@@ -231,14 +221,9 @@ def main():

    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
-    if config.model_type == "xlnet":
-        data_collator = DataCollatorForPermutationLanguageModeling(
-            tokenizer=tokenizer, plm_probability=data_args.plm_probability, max_span_length=data_args.max_span_length,
-        )
-    else:
-        data_collator = DataCollatorForLanguageModeling(
-            tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
-        )
+    data_collator = DataCollatorForLanguageModeling(
+        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
+    )

    # Initialize our Trainer
    trainer = Trainer(
@@ -1,11 +1,14 @@
 import argparse
 import logging
 import os
+import random
 from pathlib import Path
 from typing import Any, Dict

+import numpy as np
 import pytorch_lightning as pl
-from pytorch_lightning.utilities import rank_zero_info
+import torch
+from pytorch_lightning.utilities import rank_zero_info, rank_zero_only

 from transformers import (
    AdamW,
@@ -20,12 +23,7 @@ from transformers import (
    AutoTokenizer,
    PretrainedConfig,
    PreTrainedTokenizer,
-)
-from transformers.optimization import (
-    get_cosine_schedule_with_warmup,
-    get_cosine_with_hard_restarts_schedule_with_warmup,
    get_linear_schedule_with_warmup,
-    get_polynomial_decay_schedule_with_warmup,
 )


@@ -44,17 +42,12 @@ MODEL_MODES = {
 }


-# update this and the import above to support new schedulers from transformers.optimization
-arg_to_scheduler = {
-    "linear": get_linear_schedule_with_warmup,
-    "cosine": get_cosine_schedule_with_warmup,
-    "cosine_w_restarts": get_cosine_with_hard_restarts_schedule_with_warmup,
-    "polynomial": get_polynomial_decay_schedule_with_warmup,
-    # '': get_constant_schedule,             # not supported for now
-    # '': get_constant_schedule_with_warmup, # not supported for now
-}
-arg_to_scheduler_choices = sorted(arg_to_scheduler.keys())
-arg_to_scheduler_metavar = "{" + ", ".join(arg_to_scheduler_choices) + "}"
+def set_seed(args: argparse.Namespace):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.gpus > 0:
+        torch.cuda.manual_seed_all(args.seed)


 class BaseTransformer(pl.LightningModule):
@@ -70,12 +63,9 @@ class BaseTransformer(pl.LightningModule):
    ):
        """Initialize a model, tokenizer and config."""
        super().__init__()
-        # TODO: move to self.save_hyperparameters()
-        # self.save_hyperparameters()
-        # can also expand arguments into trainer signature for easier reading
-
-        self.save_hyperparameters(hparams)
+        self.hparams = hparams  # TODO: move to self.save_hyperparameters()
        self.step_count = 0
+        self.tfmr_ckpts = {}
        self.output_dir = Path(self.hparams.output_dir)
        cache_dir = self.hparams.cache_dir if self.hparams.cache_dir else None
        if config is None:
@@ -87,13 +77,6 @@ class BaseTransformer(pl.LightningModule):
            )
        else:
            self.config: PretrainedConfig = config
-
-        extra_model_params = ("encoder_layerdrop", "decoder_layerdrop", "dropout", "attention_dropout")
-        for p in extra_model_params:
-            if getattr(self.hparams, p, None):
-                assert hasattr(self.config, p), f"model config doesn't have a `{p}` attribute"
-                setattr(self.config, p, getattr(self.hparams, p))
-
        if tokenizer is None:
            self.tokenizer = AutoTokenizer.from_pretrained(
                self.hparams.tokenizer_name if self.hparams.tokenizer_name else self.hparams.model_name_or_path,
@@ -115,16 +98,8 @@ class BaseTransformer(pl.LightningModule):
    def load_hf_checkpoint(self, *args, **kwargs):
        self.model = self.model_type.from_pretrained(*args, **kwargs)

-    def get_lr_scheduler(self):
-        get_schedule_func = arg_to_scheduler[self.hparams.lr_scheduler]
-        scheduler = get_schedule_func(
-            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=self.total_steps
-        )
-        scheduler = {"scheduler": scheduler, "interval": "step", "frequency": 1}
-        return scheduler
-
    def configure_optimizers(self):
-        """Prepare optimizer and schedule (linear warmup and decay)"""
+        "Prepare optimizer and schedule (linear warmup and decay)"
        model = self.model
        no_decay = ["bias", "LayerNorm.weight"]
        optimizer_grouped_parameters = [
@@ -139,10 +114,20 @@ class BaseTransformer(pl.LightningModule):
        ]
        optimizer = AdamW(optimizer_grouped_parameters, lr=self.hparams.learning_rate, eps=self.hparams.adam_epsilon)
        self.opt = optimizer
+        return [optimizer]

-        scheduler = self.get_lr_scheduler()
+    def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_idx, second_order_closure=None):
+        if self.trainer.use_tpu:
+            xm.optimizer_step(optimizer)
+        else:
+            optimizer.step()
+        optimizer.zero_grad()
+        self.lr_scheduler.step()

-        return [optimizer], [scheduler]
+    def get_tqdm_dict(self):
+        avg_loss = getattr(self.trainer, "avg_loss", 0.0)
+        tqdm_dict = {"loss": "{:.3f}".format(avg_loss), "lr": self.lr_scheduler.get_last_lr()[-1]}
+        return tqdm_dict

    def test_step(self, batch, batch_nb):
        return self.validation_step(batch, batch_nb)
@@ -150,24 +135,26 @@ class BaseTransformer(pl.LightningModule):
    def test_epoch_end(self, outputs):
        return self.validation_end(outputs)

-    def setup(self, step):
-        train_batch_size = self.hparams.train_batch_size
-        dataloader = self.get_dataloader("train", train_batch_size)
-        self.train_loader = dataloader
-        self.total_steps = (
-            (len(dataloader.dataset) // (train_batch_size * max(1, self.hparams.gpus)))
-            // self.hparams.accumulate_grad_batches
-            * float(self.hparams.max_epochs)
-        )
-
    def train_dataloader(self):
-        return self.train_loader
+        train_batch_size = self.hparams.train_batch_size
+        dataloader = self.load_dataset("train", train_batch_size)
+
+        t_total = (
+            (len(dataloader.dataset) // (train_batch_size * max(1, self.hparams.n_gpu)))
+            // self.hparams.gradient_accumulation_steps
+            * float(self.hparams.num_train_epochs)
+        )
+        scheduler = get_linear_schedule_with_warmup(
+            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
+        )
+        self.lr_scheduler = scheduler
+        return dataloader

    def val_dataloader(self):
-        return self.get_dataloader("dev", self.hparams.eval_batch_size)
+        return self.load_dataset("dev", self.hparams.eval_batch_size)

    def test_dataloader(self):
-        return self.get_dataloader("test", self.hparams.eval_batch_size)
+        return self.load_dataset("test", self.hparams.eval_batch_size)

    def _feature_file(self, mode):
        return os.path.join(
@@ -182,9 +169,11 @@ class BaseTransformer(pl.LightningModule):
    @pl.utilities.rank_zero_only
    def on_save_checkpoint(self, checkpoint: Dict[str, Any]) -> None:
        save_path = self.output_dir.joinpath("best_tfmr")
+        save_path.mkdir(exist_ok=True)
        self.model.config.save_step = self.step_count
        self.model.save_pretrained(save_path)
        self.tokenizer.save_pretrained(save_path)
+        self.tfmr_ckpts[self.step_count] = save_path

    @staticmethod
    def add_model_specific_args(parser, root_dir):
@@ -210,46 +199,21 @@ class BaseTransformer(pl.LightningModule):
            type=str,
            help="Where do you want to store the pre-trained models downloaded from s3",
        )
-        parser.add_argument(
-            "--encoder_layerdrop",
-            type=float,
-            help="Encoder layer dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--decoder_layerdrop",
-            type=float,
-            help="Decoder layer dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--dropout", type=float, help="Dropout probability (Optional). Goes into model.config",
-        )
-        parser.add_argument(
-            "--attention_dropout", type=float, help="Attention dropout probability (Optional). Goes into model.config",
-        )
        parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
-        parser.add_argument(
-            "--lr_scheduler",
-            default="linear",
-            choices=arg_to_scheduler_choices,
-            metavar=arg_to_scheduler_metavar,
-            type=str,
-            help="Learning rate scheduler",
-        )
        parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
        parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
-        parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+        parser.add_argument("--warmup_steps", default=500, type=int, help="Linear warmup over warmup_steps.")
        parser.add_argument("--num_workers", default=4, type=int, help="kwarg passed to DataLoader")
-        parser.add_argument("--num_train_epochs", dest="max_epochs", default=3, type=int)
+        parser.add_argument(
+            "--num_train_epochs", default=3, type=int, help="Total number of training epochs to perform."
+        )
+
        parser.add_argument("--train_batch_size", default=32, type=int)
        parser.add_argument("--eval_batch_size", default=32, type=int)


 class LoggingCallback(pl.Callback):
-    def on_batch_end(self, trainer, pl_module):
-        lr_scheduler = trainer.lr_schedulers[0]["scheduler"]
-        lrs = {f"lr_group_{i}": lr for i, lr in enumerate(lr_scheduler.get_lr())}
-        pl_module.logger.log_metrics(lrs)
-
+    @rank_zero_only
    def on_validation_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
        rank_zero_info("***** Validation results *****")
        metrics = trainer.callback_metrics
@@ -258,15 +222,16 @@ class LoggingCallback(pl.Callback):
            if key not in ["log", "progress_bar"]:
                rank_zero_info("{} = {}\n".format(key, str(metrics[key])))

+    @rank_zero_only
    def on_test_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
-        rank_zero_info("***** Test results *****")
+        logger.info("***** Test results *****")
        metrics = trainer.callback_metrics
        # Log and save results to file
        output_test_results_file = os.path.join(pl_module.hparams.output_dir, "test_results.txt")
        with open(output_test_results_file, "w") as writer:
            for key in sorted(metrics):
                if key not in ["log", "progress_bar"]:
-                    rank_zero_info("{} = {}\n".format(key, str(metrics[key])))
+                    logger.info("{} = {}\n".format(key, str(metrics[key])))
                    writer.write("{} = {}\n".format(key, str(metrics[key])))


@@ -279,6 +244,7 @@ def add_generic_args(parser, root_dir) -> None:
        required=True,
        help="The output directory where the model predictions and checkpoints will be written.",
    )
+
    parser.add_argument(
        "--fp16",
        action="store_true",
@@ -288,22 +254,26 @@ def add_generic_args(parser, root_dir) -> None:
    parser.add_argument(
        "--fp16_opt_level",
        type=str,
-        default="O2",
+        default="O1",
        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
        "See details at https://nvidia.github.io/apex/amp.html",
    )
-    parser.add_argument("--n_tpu_cores", dest="tpu_cores", type=int)
-    parser.add_argument("--max_grad_norm", dest="gradient_clip_val", default=1.0, type=float, help="Max gradient norm")
+    parser.add_argument("--fast_dev_run", action="store_true")
+    parser.add_argument("--gpus", type=int, default=1)
+    parser.add_argument("--n_tpu_cores", type=int, default=0)
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
    parser.add_argument("--do_predict", action="store_true", help="Whether to run predictions on the test set.")
    parser.add_argument(
        "--gradient_accumulation_steps",
-        dest="accumulate_grad_batches",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
+
    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--resume_from_checkpoint", type=str, default=None)
+    parser.add_argument("--val_check_interval", default=1.0, type=float)


 def generic_train(
@@ -316,13 +286,10 @@ def generic_train(
    logging_callback=None,
    **extra_train_kwargs
 ):
-    pl.seed_everything(args.seed)
-
    # init model
+    set_seed(args)
    odir = Path(model.hparams.output_dir)
    odir.mkdir(exist_ok=True)
-
-    # add custom checkpoints
    if checkpoint_callback is None:
        checkpoint_callback = pl.callbacks.ModelCheckpoint(
            filepath=args.output_dir, prefix="checkpoint", monitor="val_loss", mode="min", save_top_k=1
@@ -332,25 +299,38 @@ def generic_train(

    train_params = {}

-    # TODO: remove with PyTorch 1.6 since pl uses native amp
    if args.fp16:
-        train_params["precision"] = 16
+        train_params["use_amp"] = args.fp16
        train_params["amp_level"] = args.fp16_opt_level

+    if args.n_tpu_cores > 0:
+        global xm
+        import torch_xla.core.xla_model as xm
+
+        train_params["num_tpu_cores"] = args.n_tpu_cores
+        train_params["gpus"] = 0
+
    if args.gpus > 1:
        train_params["distributed_backend"] = "ddp"

-    trainer = pl.Trainer.from_argparse_args(
-        args,
-        weights_summary=None,
-        callbacks=[logging_callback] + extra_callbacks,
+    trainer = pl.Trainer(
        logger=logger,
-        checkpoint_callback=checkpoint_callback,
+        accumulate_grad_batches=args.gradient_accumulation_steps,
+        gpus=args.gpus,
+        max_epochs=args.num_train_epochs,
        early_stop_callback=early_stopping_callback,
+        gradient_clip_val=args.max_grad_norm,
+        checkpoint_callback=checkpoint_callback,
+        callbacks=[logging_callback] + extra_callbacks,
+        fast_dev_run=args.fast_dev_run,
+        val_check_interval=args.val_check_interval,
+        weights_summary=None,
+        resume_from_checkpoint=args.resume_from_checkpoint,
        **train_params,
    )

    if args.do_train:
        trainer.fit(model)
-
+    trainer.logger.log_hyperparams(args)
+    trainer.logger.save()
    return trainer
@@ -1,10 +1,10 @@
 import faiss
 import nlp
 import numpy as np
-import streamlit as st
 import torch
 from elasticsearch import Elasticsearch

+import streamlit as st
 import transformers
 from eli5_utils import (
    embed_questions_for_retrieval,
@@ -193,12 +193,12 @@ def make_qa_retriever_model(model_name="google/bert_uncased_L-8_H-512_A-8", from
 def make_qa_retriever_batch(qa_list, tokenizer, max_len=64, device="cuda:0"):
    q_ls = [q for q, a in qa_list]
    a_ls = [a for q, a in qa_list]
-    q_toks = tokenizer(q_ls, max_length=max_len, padding="max_length", truncation=True)
+    q_toks = tokenizer.batch_encode_plus(q_ls, max_length=max_len, pad_to_max_length=True)
    q_ids, q_mask = (
        torch.LongTensor(q_toks["input_ids"]).to(device),
        torch.LongTensor(q_toks["attention_mask"]).to(device),
    )
-    a_toks = tokenizer(a_ls, max_length=max_len, padding="max_length", truncation=True)
+    a_toks = tokenizer.batch_encode_plus(a_ls, max_length=max_len, pad_to_max_length=True)
    a_ids, a_mask = (
        torch.LongTensor(a_toks["input_ids"]).to(device),
        torch.LongTensor(a_toks["attention_mask"]).to(device),
@@ -375,12 +375,12 @@ def make_qa_s2s_model(model_name="facebook/bart-large", from_file=None, device="
 def make_qa_s2s_batch(qa_list, tokenizer, max_len=64, max_a_len=360, device="cuda:0"):
    q_ls = [q for q, a in qa_list]
    a_ls = [a for q, a in qa_list]
-    q_toks = tokenizer(q_ls, max_length=max_len, padding="max_length", truncation=True)
+    q_toks = tokenizer.batch_encode_plus(q_ls, max_length=max_len, pad_to_max_length=True)
    q_ids, q_mask = (
        torch.LongTensor(q_toks["input_ids"]).to(device),
        torch.LongTensor(q_toks["attention_mask"]).to(device),
    )
-    a_toks = tokenizer(a_ls, max_length=min(max_len, max_a_len), padding="max_length", truncation=True)
+    a_toks = tokenizer.batch_encode_plus(a_ls, max_length=min(max_len, max_a_len), pad_to_max_length=True)
    a_ids, a_mask = (
        torch.LongTensor(a_toks["input_ids"]).to(device),
        torch.LongTensor(a_toks["attention_mask"]).to(device),
@@ -531,7 +531,7 @@ def qa_s2s_generate(
 # ELI5-trained retrieval model usage
 ###############
 def embed_passages_for_retrieval(passages, tokenizer, qa_embedder, max_length=128, device="cuda:0"):
-    a_toks = tokenizer(passages, max_length=max_length, padding="max_length", truncation=True)
+    a_toks = tokenizer.batch_encode_plus(passages, max_length=max_length, pad_to_max_length=True)
    a_ids, a_mask = (
        torch.LongTensor(a_toks["input_ids"]).to(device),
        torch.LongTensor(a_toks["attention_mask"]).to(device),
@@ -542,7 +542,7 @@ def embed_passages_for_retrieval(passages, tokenizer, qa_embedder, max_length=12


 def embed_questions_for_retrieval(q_ls, tokenizer, qa_embedder, device="cuda:0"):
-    q_toks = tokenizer(q_ls, max_length=128, padding="max_length", truncation=True)
+    q_toks = tokenizer.batch_encode_plus(q_ls, max_length=128, pad_to_max_length=True)
    q_ids, q_mask = (
        torch.LongTensor(q_toks["input_ids"]).to(device),
        torch.LongTensor(q_toks["attention_mask"]).to(device),
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
sshleifer	a07fcff087	boom boom	2020-06-25 02:10:13 +00:00
sshleifer	243f2d98f4	boom boom	2020-06-24 20:44:27 +00:00
sshleifer	7c6802d216	boom boom	2020-06-23 16:10:00 -04:00
sshleifer	6dd3122ad4	tests pass	2020-06-23 14:35:54 -04:00
sshleifer	3a8f32d919	pl 0.8.1 compat	2020-06-19 14:46:51 +00:00
sshleifer	29d1eb57ac	Support top, bottom init	2020-06-19 09:43:20 -04:00
sshleifer	02418a591b	boom boom	2020-06-18 21:38:13 -04:00
sshleifer	be95182ee1	half-merge	2020-06-18 21:37:14 -04:00
sshleifer	970fe3d833	smaller distilbart-clean delta	2020-06-17 16:03:49 -04:00
sshleifer	97e29a9cf3	allstats-option	2020-06-15 21:38:08 -04:00
sshleifer	785a1b6076	RougeTracker	2020-06-15 21:13:36 -04:00
sshleifer	9ef5581d5e	boom boom	2020-06-15 18:05:27 -04:00
sshleifer	657b7fdada	boom boom	2020-06-15 17:57:13 -04:00
sshleifer	3a5999f5d4	boom boom	2020-06-15 17:50:28 -04:00
sshleifer	0efb2830f4	turn off ES callback	2020-06-15 17:43:01 -04:00
sshleifer	b5099747f9	boom boom	2020-06-15 17:36:56 -04:00
sshleifer	34fd62a5ea	init with copy	2020-06-15 17:36:12 -04:00
sshleifer	e167e9fc74	boom boom	2020-06-15 17:19:30 -04:00
sshleifer	1df5e7af4a	odir fix	2020-06-15 17:10:53 -04:00
sshleifer	544f543861	rouge_tracker_csvs	2020-06-15 16:56:58 -04:00
sshleifer	10970ec650	fixed dec_mask, maybe	2020-06-15 12:42:54 -04:00
sshleifer	78b1c850f2	passing, unsolved	2020-06-15 12:24:39 -04:00
sshleifer	ac8d301786	boom boom	2020-06-14 21:56:11 -04:00
sshleifer	9f32be4fbc	boom boom	2020-06-14 21:54:18 -04:00
sshleifer	74c8ac0f84	fixes	2020-06-14 21:51:09 -04:00
sshleifer	1d5e77d87b	test_passing	2020-06-14 21:41:49 -04:00
sshleifer	47b3638593	boom boom	2020-06-14 21:18:15 -04:00
sshleifer	f7c95805cf	boom boom	2020-06-14 21:01:52 -04:00
sshleifer	0169cbbe49	all but t5 passing	2020-06-14 21:01:41 -04:00
sshleifer	8e4a74b1cb	borked	2020-06-14 20:46:20 -04:00
sshleifer	86034ae640	Merge branch 'distilbart-clean' into theseus	2020-06-14 19:51:25 -04:00
sshleifer	5283878c9f	Move stuff to utils	2020-06-14 19:17:31 -04:00
sshleifer	6bdfb14664	boom boom	2020-06-14 19:07:15 -04:00
sshleifer	99de2c3c3c	pass through logger	2020-06-14 19:01:15 -04:00
sshleifer	9e95429843	docs	2020-06-14 18:20:02 -04:00
sshleifer	c9597fef42	docs	2020-06-14 18:15:28 -04:00
sshleifer	c34b886cd1	Wandb logger	2020-06-14 18:11:22 -04:00
sshleifer	f179e7b37c	Allow wandb logger	2020-06-14 17:58:41 -04:00
sshleifer	969c271bd7	more honest docs	2020-06-14 20:18:43 +00:00
sshleifer	9f98a6ac15	docs	2020-06-14 16:07:25 -04:00
sshleifer	4606bda735	Bash cleanup	2020-06-14 19:47:04 +00:00
sshleifer	eb84b9d5ee	boom boom	2020-06-14 15:36:29 -04:00
sshleifer	5526be3499	boom boom	2020-06-14 15:35:47 -04:00
sshleifer	6bf996f58d	boom boom	2020-06-14 15:31:46 -04:00
sshleifer	35a82ee87f	boom boom	2020-06-14 15:01:50 -04:00
sshleifer	379e8c7dba	Cleanup	2020-06-14 15:00:35 -04:00
sshleifer	3e62d96b14	add git-python requirement	2020-06-14 14:28:20 -04:00
sshleifer	350eeb7cd2	style	2020-06-14 13:17:09 -04:00
sshleifer	c06e38add3	boom boom	2020-06-14 13:15:17 -04:00
sshleifer	d94b09d285	started	2020-06-13 20:27:10 -04:00
sshleifer	eab9779732	better mask logic	2020-06-13 20:12:11 -04:00
sshleifer	4a6b23ee40	boom boom	2020-06-13 20:11:41 -04:00
sshleifer	e2d45445c3	style	2020-06-13 19:55:14 -04:00
sshleifer	de7035d9f6	style	2020-06-13 19:49:35 -04:00
sshleifer	abee3e2ecd	Merge branch 'master' into distilbart-clean	2020-06-13 16:40:30 -04:00
sshleifer	0274338aff	original tests pass	2020-06-13 16:40:04 -04:00
sshleifer	e67ad41ae7	boom boom	2020-06-13 15:33:00 -04:00
sshleifer	c5a1de5b75	boom boom	2020-06-13 15:30:03 -04:00
sshleifer	af3f86fd6a	remove some cruft	2020-06-11 23:58:34 -04:00
sshleifer	fb6eb096ec	Failing but minimal	2020-06-11 23:51:55 -04:00
sshleifer	bdf0b1ab88	boom boom	2020-06-11 23:45:32 -04:00
sshleifer	9d99761b1b	passing	2020-06-11 22:55:51 -04:00
sshleifer	3f3040ae12	boom boom	2020-06-11 10:51:05 -04:00
sshleifer	4224ac999d	boom boom	2020-06-11 09:29:05 -04:00
sshleifer	c7d05f70db	boom boom	2020-06-11 09:26:30 -04:00
sshleifer	ac47dbdfdf	boom boom	2020-06-11 09:24:46 -04:00
sshleifer	10b092ea27	boom boom	2020-06-11 09:22:15 -04:00
sshleifer	e5728e4ddb	boom boom	2020-06-11 02:31:05 +00:00
sshleifer	5a281449df	fp16_ever=False	2020-06-11 02:29:48 +00:00
sshleifer	585bb576ba	passing cpu	2020-06-10 19:47:02 -04:00
sshleifer	6aa28bd1f5	ignore student	2020-06-10 23:27:33 +00:00
sshleifer	a2f2d5da31	undo some chg	2020-06-10 23:23:20 +00:00
sshleifer	b2b222d0db	boom boom	2020-06-10 17:20:26 -04:00
sshleifer	574ffbb1eb	boom boom	2020-06-10 17:17:59 -04:00
sshleifer	a549aaac3e	boom boom	2020-06-10 17:17:03 -04:00
sshleifer	bad1b75a91	boom boom	2020-06-10 17:04:13 -04:00
sshleifer	6624b63c24	boom boom	2020-06-10 16:56:42 -04:00
sshleifer	7f3f9fae22	boom boom	2020-06-10 16:36:06 -04:00
sshleifer	c553e1214c	t5 works	2020-06-09 23:54:39 -04:00
sshleifer	d26f5dac45	stuck on t5 mask	2020-06-09 23:35:59 -04:00
sshleifer	913bb8f770	boom boom	2020-06-09 22:53:11 -04:00
sshleifer	af05c0e004	boom boom	2020-06-09 22:52:35 -04:00
sshleifer	dc785f0d10	boom boom	2020-06-09 22:51:59 -04:00
sshleifer	d3b8694cd8	boom boom	2020-06-09 22:50:44 -04:00
sshleifer	b5adf48404	boom boom	2020-06-09 22:49:23 -04:00
sshleifer	315ab6be60	boom boom	2020-06-09 22:48:38 -04:00
sshleifer	76377c9854	boom boom	2020-06-09 22:45:35 -04:00
sshleifer	a387c057db	boom boom	2020-06-09 22:45:03 -04:00
sshleifer	0098872c04	Deleted SummarizationDistiller	2020-06-09 22:41:36 -04:00
sshleifer	a79f5fb77e	boom boom	2020-06-09 22:09:57 -04:00
sshleifer	0dc32e6f27	boom boom	2020-06-09 21:11:07 -04:00
sshleifer	fa7f8185e8	t5 failing	2020-06-09 17:37:38 -04:00
sshleifer	31b4c8cc08	boom boom	2020-06-09 15:31:31 -04:00
sshleifer	318a585f2b	boom boom	2020-06-09 17:44:45 +00:00
sshleifer	630af0c3fa	boom boom	2020-06-09 13:27:31 -04:00
sshleifer	3c384445a7	boom boom	2020-06-08 21:58:24 +00:00
sshleifer	9f6596ebc9	boom boom	2020-06-08 14:30:54 -04:00
sshleifer	582efc0319	boom boom	2020-06-08 14:00:07 -04:00
sshleifer	3467f1c7f6	boom boom	2020-06-08 17:58:33 +00:00
sshleifer	bcb4624d42	boom boom	2020-06-08 13:57:57 -04:00
sshleifer	e46e10150d	boom boom	2020-06-08 13:29:51 -04:00
sshleifer	ecff6f92fa	V8 and generate_summaries fixes	2020-06-08 13:14:46 -04:00
sshleifer	c59be9b2e0	boom boom	2020-06-08 13:07:01 -04:00
sshleifer	b32d283d26	boom boom	2020-06-08 12:17:51 -04:00
sshleifer	7ccd2366de	Evaluate_checkpoint	2020-06-08 11:32:29 -04:00
sshleifer	c294e5e331	boom boom	2020-06-07 14:23:46 -04:00
sshleifer	3b304afdf7	boom boom	2020-06-07 14:21:49 -04:00
sshleifer	09f199b790	Merge branch 'distilbart' of github.com:sshleifer/transformers_fork into distilbart	2020-06-07 14:15:28 -04:00
sshleifer	bb1bffc89e	Save top1	2020-06-07 14:15:23 -04:00
sshleifer	47c226ea3f	Merge branch 'distilbart' of github.com:sshleifer/transformers_fork into distilbart	2020-06-07 17:48:10 +00:00
sshleifer	3abbf78a48	boom boom	2020-06-07 17:48:06 +00:00
sshleifer	9fc692d9f3	boom boom	2020-06-06 13:40:07 -04:00
sshleifer	e146713686	passing	2020-06-06 13:19:46 -04:00
sshleifer	014c061f3d	boom boom	2020-06-06 11:19:45 -04:00
sshleifer	e446fd24ef	boom boom	2020-06-06 13:13:55 +00:00
sshleifer	055167ba12	boom boom	2020-06-06 01:00:50 -04:00
sshleifer	8f68b91e81	boom boom	2020-06-05 20:10:55 -04:00
sshleifer	8b98c4cfa8	boom boom	2020-06-05 20:02:23 -04:00
sshleifer	e3a00a402b	boom boom	2020-06-05 19:33:35 -04:00
sshleifer	24f483cf06	boom boom	2020-06-05 19:33:12 -04:00
sshleifer	8d4f075245	boom boom	2020-06-05 19:26:14 -04:00
sshleifer	0b98750ecd	boom boom	2020-06-05 17:26:11 -04:00
sshleifer	cc83d34a7d	boom boom	2020-06-05 15:13:24 -04:00
sshleifer	a65ab52d74	boom boom	2020-06-05 19:10:02 +00:00
sshleifer	3e47e61a1a	Disable cos loss	2020-06-05 12:51:13 -04:00
sshleifer	20aedfbcf7	boom boom	2020-06-04 09:45:47 -04:00
sshleifer	440744a111	Distiller brewer does not break	2020-06-04 09:38:02 -04:00
sshleifer	a2ba031b92	boom boom	2020-06-04 07:52:34 -04:00
sshleifer	8f614af5b7	broken	2020-06-04 07:15:34 -04:00
sshleifer	df0b28a581	boom boom	2020-06-04 02:09:27 -04:00
sshleifer	8c6769ee41	boom boom	2020-06-03 13:04:02 +00:00
sshleifer	b11fc6977e	boom boom	2020-06-03 13:02:37 +00:00
sshleifer	fb6adb2a47	backup rouge dfs	2020-06-03 12:53:05 +00:00
sshleifer	789f06e9d8	less annoying requirements	2020-06-02 18:59:11 -04:00
sshleifer	e17df7499c	merged master	2020-06-02 13:45:10 -04:00
sshleifer	82a6b8345f	Merge branch 'master' into distilbart	2020-06-02 13:44:40 -04:00
sshleifer	60a90143ee	rouge tracker fix	2020-06-02 11:17:09 -04:00
sshleifer	7b9557fbf5	boom boom	2020-05-31 09:16:20 -04:00
sshleifer	105d83d595	undo rename	2020-05-31 08:52:04 -04:00
sshleifer	bd361eae11	rename generic_train -> build_trainer	2020-05-31 08:38:36 -04:00
sshleifer	8fb24ce3ce	boom boom	2020-05-27 15:49:36 -04:00
sshleifer	6689b594fc	boom boom	2020-05-27 15:46:50 -04:00
sshleifer	25edee17ed	boom boom	2020-05-27 14:22:36 -04:00
sshleifer	7c52e73dfe	boom boom	2020-05-27 14:19:08 -04:00
sshleifer	d53236c840	boom boom	2020-05-27 14:17:34 -04:00
sshleifer	eb4b3cce7a	boom boom	2020-05-27 13:50:07 -04:00
sshleifer	324d4dcfce	No wandb if ddp	2020-05-27 12:49:11 -04:00
sshleifer	3f6463a895	attempt logger.name fix	2020-05-27 11:51:49 -04:00
sshleifer	3c34cae08a	distrib hacks	2020-05-27 15:47:45 +00:00
sshleifer	175efd864a	getstate	2020-05-27 14:33:15 +00:00
sshleifer	516bbcabd0	boom boom	2020-05-27 10:28:34 -04:00
sshleifer	0c3d6d8cbf	boom boom	2020-05-27 10:19:14 -04:00
sshleifer	84ea9fac23	Fixed tests	2020-05-27 09:37:52 -04:00
sshleifer	44743e8dff	fix ckpt bug maybe	2020-05-27 08:49:18 -04:00
sshleifer	7f1af24a05	freeze decoder	2020-05-27 01:10:20 -04:00
sshleifer	99e7161eb2	boom boom	2020-05-26 21:23:47 -04:00
sshleifer	d92a69f223	boom boom	2020-05-26 21:22:08 -04:00
sshleifer	4f19747960	boom boom	2020-05-26 21:20:53 -04:00
sshleifer	9b54356b0b	Logging adjustments	2020-05-26 20:08:41 -04:00
sshleifer	f750d801ff	boom boom	2020-05-26 15:25:28 -04:00
sshleifer	298ec5361c	--freeze_encoder	2020-05-26 15:21:00 -04:00
sshleifer	ee18f81852	Merge branch 'distilbart' of github.com:sshleifer/transformers_fork into distilbart	2020-05-26 15:05:36 -04:00
sshleifer	e7d9dc3d22	dont default teacher	2020-05-26 19:05:30 +00:00
sshleifer	d2cd7b4d9d	Somehow passing with better freezing logic	2020-05-26 15:01:56 -04:00
sshleifer	d3611c9844	Run distiller: no data defaults	2020-05-26 12:00:12 -04:00
sshleifer	c10bb9fb23	boom boom	2020-05-25 23:15:20 -04:00
sshleifer	81e939fb54	boom boom	2020-05-25 21:23:39 -04:00
sshleifer	0c3d22ea86	boom boom	2020-05-25 21:20:08 -04:00
sshleifer	63521f19a9	boom boom	2020-05-25 21:18:18 -04:00
sshleifer	56aedacc27	boom boom	2020-05-25 21:15:00 -04:00
sshleifer	0933e118e0	boom boom	2020-05-25 21:07:02 -04:00
sshleifer	12f2d9e6bc	early quitting for encoder only loss	2020-05-25 21:01:47 -04:00
sshleifer	20a59685a9	passing with enc mse	2020-05-25 20:01:04 -04:00
sshleifer	f5e697bbe6	Dont require model_name_or_path	2020-05-25 19:31:17 -04:00
sshleifer	db52f3708f	save fewer checkpoints	2020-05-25 17:18:57 -04:00
sshleifer	800dcdcce2	boom boom	2020-05-25 16:51:21 -04:00
sshleifer	88f9f917a6	Sortish Sampler	2020-05-25 16:44:48 -04:00
sshleifer	a42f09eec6	warmup_steps=500	2020-05-25 16:24:49 -04:00
sshleifer	c154fb3f8b	style	2020-05-25 10:12:23 -04:00
sshleifer	c7f6e62c3e	boom boom	2020-05-25 09:26:59 -04:00
sshleifer	af8b962edf	boom boom	2020-05-25 09:26:11 -04:00
sshleifer	6f1757d2b8	boom boom	2020-05-25 09:23:48 -04:00
sshleifer	4d3617acc6	boom boom	2020-05-24 22:45:23 -04:00
sshleifer	2443a7eefc	boom boom	2020-05-24 22:42:50 -04:00
sshleifer	6f03e39b30	boom boom	2020-05-24 22:39:17 -04:00
sshleifer	9f214624f4	boom boom	2020-05-24 22:37:28 -04:00
sshleifer	e48a8299b9	boom boom	2020-05-24 22:36:48 -04:00
sshleifer	20abeec1f5	passing	2020-05-24 14:22:45 -04:00
sshleifer	584c329ffe	encoder different	2020-05-24 14:20:29 -04:00
sshleifer	d32887160a	add RougeTracker	2020-05-22 13:02:59 -04:00
sshleifer	ef45ee9a7e	boom boom	2020-05-22 15:58:06 +00:00
sshleifer	8fbdd7d205	boom boom	2020-05-22 11:54:34 -04:00
sshleifer	1ca011ff75	test_mtl=140 default	2020-05-22 11:49:34 -04:00
sshleifer	a78603cee7	boom boom	2020-05-21 23:14:56 -04:00
sshleifer	d4830f9d0a	rename	2020-05-21 23:13:57 -04:00
sshleifer	c779c38297	boom boom	2020-05-21 17:04:19 -04:00
sshleifer	021d2b9f80	only train	2020-05-21 15:04:20 -04:00
sshleifer	fa0eda43b6	boom boom	2020-05-21 14:19:55 -04:00
sshleifer	c6f7e14d0b	progbar	2020-05-21 14:07:16 -04:00
sshleifer	db3d85ad2c	boom boom	2020-05-21 13:56:35 -04:00
sshleifer	705ed2ffb4	Better desc	2020-05-21 13:43:47 -04:00
sshleifer	29499c9ca5	boom boom	2020-05-21 13:19:07 -04:00
sshleifer	bcb9996ae5	overcautious freezing	2020-05-21 13:17:17 -04:00
sshleifer	aa995f98a3	freeze after init	2020-05-21 13:02:00 -04:00
sshleifer	5d751729ec	Freeze encoder fix	2020-05-21 12:55:59 -04:00
sshleifer	f5606df47b	boom boom	2020-05-21 12:26:07 -04:00
sshleifer	ca82946559	boom boom	2020-05-21 12:24:06 -04:00
sshleifer	3461e24b8d	boom boom	2020-05-21 12:21:42 -04:00
sshleifer	15a2d4ee79	boom boom	2020-05-21 11:45:07 -04:00
sshleifer	3f73cb76a7	boom boom	2020-05-21 11:41:13 -04:00
sshleifer	88b4b970d3	switch copy logic	2020-05-21 11:39:29 -04:00
sshleifer	ac72c7aadc	convenience method	2020-05-21 09:55:38 -04:00
sshleifer	dd418e8d1d	boom boom	2020-05-21 09:41:17 -04:00
sshleifer	07da0623ad	avoid losing bart grouped batch sampler work	2020-05-21 08:49:24 -04:00
sshleifer	21a2cc0eb1	delete run_distiller -> main	2020-05-21 01:20:11 -04:00
sshleifer	2c2a10e5d7	orig tests pass	2020-05-21 01:08:43 -04:00
sshleifer	910da0fb96	boom boom	2020-05-21 01:00:57 -04:00
sshleifer	7a7dcd397f	l2copy[1] = 0	2020-05-20 22:25:18 -04:00
sshleifer	b39c39fe87	boom boom	2020-05-20 22:09:49 -04:00
sshleifer	ce0de9073a	boom boom	2020-05-20 22:07:35 -04:00
sshleifer	cf47a6ff9d	boom boom	2020-05-20 22:05:05 -04:00
sshleifer	a6420767e9	boom boom	2020-05-20 22:04:48 -04:00
sshleifer	fee71b4fef	npars	2020-05-20 22:01:26 -04:00
sshleifer	a2671a6eee	boom boom	2020-05-20 21:30:49 -04:00
sshleifer	18524d8396	boom boom	2020-05-20 21:30:36 -04:00
sshleifer	a250d1c6aa	boom boom	2020-05-20 21:30:09 -04:00
sshleifer	b62bfe666d	spelling	2020-05-20 18:36:14 -04:00
sshleifer	1fc653f626	log_metrics	2020-05-20 18:18:11 -04:00
sshleifer	f281724e92	wandb	2020-05-20 18:14:27 -04:00
sshleifer	67f9553d4b	Run encoder once	2020-05-20 17:58:43 -04:00
sshleifer	cd461e13ff	fix	2020-05-20 16:40:27 -04:00
sshleifer	05bfb1eb1d	Support fewer layers	2020-05-20 16:38:20 -04:00
sshleifer	c79351a64a	fix flatten	2020-05-20 15:59:22 -04:00
sshleifer	c774160bd2	imports	2020-05-20 15:30:32 -04:00
sshleifer	beda65dd17	bash	2020-05-20 19:29:21 +00:00
sshleifer	59c658de2f	fixes	2020-05-20 15:28:05 -04:00
sshleifer	5947812374	add files	2020-05-20 15:08:41 -04:00
sshleifer	09cc1e64cf	save preds	2020-05-20 15:04:38 -04:00
sshleifer	5fbf42258c	summaries for file	2020-05-20 12:56:26 -04:00
sshleifer	af315c157e	boom boom	2020-05-20 09:55:53 -04:00
sshleifer	2c7294830e	val check test passing with smaller batch size	2020-05-20 09:31:11 -04:00
sshleifer	2b7132c25e	metrics saving, but no val_check_interval honored	2020-05-20 09:20:55 -04:00
sshleifer	bef77fc211	boom boom	2020-05-19 15:59:22 -04:00
sshleifer	abb81df15f	boom boom	2020-05-19 15:58:04 -04:00
sshleifer	51221fbbf8	boom boom	2020-05-19 15:57:04 -04:00
sshleifer	2ee9388492	boom boom	2020-05-19 15:54:12 -04:00
sshleifer	c4530f91ab	batch	2020-05-19 15:22:13 -04:00
sshleifer	bbc4e52d3b	assert student small	2020-05-19 15:18:50 -04:00
sshleifer	74704cedab	boom boom	2020-05-19 14:46:35 -04:00
sshleifer	bf5782e080	boom boom	2020-05-19 14:38:41 -04:00
sshleifer	3cebd56848	boom boom	2020-05-19 14:35:05 -04:00
sshleifer	696c8c28f2	boom boom	2020-05-19 14:34:20 -04:00
sshleifer	04a8ace0df	boom boom	2020-05-19 14:32:41 -04:00
sshleifer	0ed41562eb	boom boom	2020-05-19 14:30:36 -04:00
sshleifer	7081d0f3fd	add rouge	2020-05-19 14:29:14 -04:00
sshleifer	ca9c685453	rouge	2020-05-19 14:05:46 -04:00
sshleifer	70cf536ae2	Merge branch 'distilbart' of github.com:sshleifer/transformers_fork into distilbart	2020-05-19 11:35:22 -04:00
sshleifer	5a3ed998f0	Cache tokenized	2020-05-19 11:35:17 -04:00
sshleifer	6302fb02ef	bs=8	2020-05-19 14:44:21 +00:00
sshleifer	5a35811635	boom boom	2020-05-19 09:55:26 -04:00
sshleifer	174aaf374d	Fast dev run	2020-05-19 09:54:43 -04:00
sshleifer	1edc50f670	bash	2020-05-19 13:50:01 +00:00
sshleifer	d2cc12b182	boom boom	2020-05-19 09:16:02 -04:00
sshleifer	8afb88ce24	relatif importschlossen	2020-05-19 09:09:15 -04:00
sshleifer	4be1287671	real data test passes	2020-05-19 09:00:33 -04:00
sshleifer	4f5790f652	passing	2020-05-18 15:16:41 -04:00
sshleifer	fcc49a0c5c	can import	2020-05-18 14:31:36 -04:00
sshleifer	937d3d6ae8	Merge branch 'master' into distilbart	2020-05-18 14:30:34 -04:00
sshleifer	b01f1d594c	Failing test	2020-05-18 14:19:46 -04:00
sshleifer	71850ad26b	copy decoder layers	2020-05-18 13:33:30 -04:00
sshleifer	1368969930	passing	2020-05-18 11:55:30 -04:00
sshleifer	f2de528d04	passing	2020-05-18 11:49:12 -04:00
sshleifer	eabb3c5f9a	test passing, loss doesnt change	2020-05-18 11:13:54 -04:00
sshleifer	762dd832f1	boom boom	2020-05-17 15:43:35 -04:00
sshleifer	8280dad80e	roberta extract test works	2020-05-16 11:09:26 -04:00