Add examples/bert-loses-patience who can help

Fix TPU Convergence bug introduced by PR#6151 (#6488 )
Currently with the bug introduced we're taking two optimizer steps per batch: one global one, where `xm.optimizer_step` injects a CRS between all cores in training, and one without. This has been affecting training accuracy (for example, XLNet GLUE on MNLI is not converging, etc.).
2020-08-15 19:54:24 +08:00 · 2020-08-14 12:47:37 -04:00 · 2020-08-14 09:46:39 -04:00 · 2020-08-14 08:16:16 -04:00 · 2020-08-14 10:36:58 +02:00 · 2020-08-14 09:43:29 +02:00
594 changed files with 34917 additions and 11813 deletions
@@ -1,4 +1,66 @@
-version: 2
+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}')
+
+
+
+
 jobs:
    run_tests_torch_and_tf:
        working_directory: ~/transformers
@@ -10,8 +72,17 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install .[sklearn,tf-cpu,torch,testing]
-            - run: sudo pip install codecov pytest-cov
+            - 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: codecov
            - store_artifacts:
@@ -27,12 +98,20 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install .[sklearn,torch,testing]
+            - 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_tests_tf:
        working_directory: ~/transformers
        docker:
@@ -43,7 +122,16 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install .[sklearn,tf-cpu,testing]
+            - 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
@@ -56,7 +144,16 @@ jobs:
            RUN_CUSTOM_TOKENIZERS: yes
        steps:
            - checkout
-            - run: sudo pip install .[mecab,testing]
+            - 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
@@ -71,9 +168,18 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install .[sklearn,torch,testing]
-            - run: sudo pip install -r examples/requirements.txt
-            - run: python -m pytest -n 8 --dist=loadfile -s ./examples/ | tee output.txt
+            - 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
@@ -83,7 +189,16 @@ jobs:
            - image: circleci/python:3.6
        steps:
            - checkout
-            - run: sudo pip install .[tf,torch,docs]
+            - 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: cd docs && make html SPHINXOPTS="-W"
            - store_artifacts:
                path: ./docs/_build
@@ -96,7 +211,15 @@ jobs:
                fingerprints:
                    - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
            - checkout
-            - run: sudo pip install .[tf,torch,docs]
+            - 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: ./.circleci/deploy.sh
    check_code_quality:
        working_directory: ~/transformers
@@ -106,12 +229,22 @@ 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: sudo pip install git+git://github.com/timothycrosley/isort.git@e63ae06ec7d70b06df9e528357650281a3d3ec22#egg=isort
-            - run: sudo pip install .[tf,torch,quality]
+            - 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: 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:
@@ -120,8 +253,37 @@ jobs:
        parallelism: 1
        steps:
            - checkout
-            - run: sudo pip install requests
+            - run: 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:
@@ -140,3 +302,15 @@ 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
@@ -47,4 +47,4 @@ 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 "b0892fa" #v3.0.2 Latest stable release
+deploy_doc "7fb8bdf" #v3.0.2 Latest stable release
@@ -7,14 +7,52 @@ assignees: ''

 ---

-# 🐛 Bug
+
+## 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
+ examples/bert-loses-patience: @JetRunner
+ tensorflow: @jplu 
+ documentation: @sgugger
+ -->

 ## 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)
@@ -38,15 +76,3 @@ 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 Stack Overflow tagged huggingface-transformers
+about: Post your general questions on the Hugging Face forum or Stack Overflow tagged huggingface-transformers
 title: ''
 labels: ''
 assignees: ''
@@ -11,19 +11,17 @@ 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 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: 
+     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: 
     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 SO, and only if
+<!-- You should first ask your question on the forum or SO, and only if
     you didn't get an answer ask it here on GitHub. -->
-**A link to original question on Stack Overflow**:
+**A link to original question on the forum/Stack Overflow**:
@@ -1,19 +0,0 @@
-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]
-
-
-
@@ -18,8 +18,17 @@ 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

@@ -25,6 +25,14 @@ 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
@@ -35,7 +43,8 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install torch
+        pip install --upgrade pip
+        pip install torch!=1.6.0
        pip install .[sklearn,testing]

    - name: Are GPUs recognized by our DL frameworks
@@ -51,11 +60,4 @@ jobs:
        USE_CUDA: yes
      run: |
        source .env/bin/activate
-        python -m pytest -n 2 --dist=loadfile -s ./tests/ | tee output.txt
-    - name: cat output.txt
-      run: cat output.txt
-    - name: Upload output.txt
-      uses: actions/upload-artifact@v1
-      with:
-        name: pytest_output
-        path: output.txt
+        python -m pytest -n 2 --dist=loadfile -s ./tests/
@@ -13,6 +13,14 @@ 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
@@ -22,6 +30,7 @@ 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
@@ -31,7 +40,9 @@ jobs:
    - name: Install dependencies
      run: |
        source .env/bin/activate
-        pip install .[sklearn,torch,testing]
+        pip install --upgrade pip
+        pip install torch!=1.6.0
+        pip install .[sklearn,testing]

    - name: Are GPUs recognized by our DL frameworks
      run: |
@@ -46,11 +57,14 @@ jobs:
        USE_CUDA: yes
      run: |
        source .env/bin/activate
-        python -m pytest -n 1 --dist=loadfile -s ./tests/  | tee output.txt
-    - name: cat output.txt
-      run: cat output.txt
-    - name: Upload output.txt
-      uses: actions/upload-artifact@v1
-      with:
-        name: pytest_output
-        path: output.txt
+        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
@@ -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/templates) folder.
+in the [`templates`](https://github.com/huggingface/transformers/tree/master/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/templates) 
+models in the library. You can find them in the [`templates`](https://github.com/huggingface/transformers/tree/master/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. 
+   CircleCI does not run the slow tests, but github actions does every night!
 6. All public methods must have informative docstrings that work nicely with sphinx. See `modeling_ctrl.py` for an 
   example.

@@ -238,8 +238,7 @@ 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!
+In fact, that's how `make test` and `make test-examples` are implemented (sans the `pip install` line)!

 You can specify a smaller set of tests in order to test only the feature
 you're working on.
@@ -6,6 +6,7 @@ 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

@@ -174,12 +174,19 @@ 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

@@ -1,23 +0,0 @@
-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
@@ -0,0 +1,65 @@
+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 $GITHUB_REF:CI && \
+    git checkout CI && \
+    cd .. && \
+    pip install ./transformers && \
+    pip install -r ./transformers/examples/requirements.txt && \
+    pip install pytest
+
+RUN python -c "import torch_xla; print(torch_xla.__version__)"
+RUN python -c "import transformers as trf; print(trf.__version__)"
+RUN conda init bash
+COPY docker-entrypoint.sh /usr/local/bin/
+RUN chmod +x /usr/local/bin/docker-entrypoint.sh
+ENTRYPOINT ["/usr/local/bin/docker-entrypoint.sh"]
+CMD ["bash"]
@@ -0,0 +1,38 @@
+local base = import 'templates/base.libsonnet';
+local tpus = import 'templates/tpus.libsonnet';
+local utils = import "templates/utils.libsonnet";
+local volumes = import "templates/volumes.libsonnet";
+
+local bertBaseCased = base.BaseTest {
+  frameworkPrefix: "hf",
+  modelName: "bert-base-cased",
+  mode: "example",
+  configMaps: [],
+
+  timeout: 3600, # 1 hour, in seconds
+
+  image: std.extVar('image'),
+  imageTag: std.extVar('image-tag'),
+
+  tpuSettings+: {
+    softwareVersion: "pytorch-nightly",
+  },
+  accelerator: tpus.v3_8,
+
+  volumeMap+: {
+    datasets: volumes.PersistentVolumeSpec {
+      name: "huggingface-cluster-disk",
+      mountPath: "/datasets",
+    },
+  },
+  command: utils.scriptCommand(
+    |||
+      python -m pytest -s transformers/examples/test_xla_examples.py -v
+      test_exit_code=$?
+      echo "\nFinished running commands.\n"
+      test $test_exit_code -eq 0
+    |||
+  ),
+};
+
+bertBaseCased.oneshotJob
@@ -0,0 +1,32 @@
+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
@@ -0,0 +1,8 @@
+#!/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
@@ -1,5 +1,36 @@
 /* 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 {
@@ -21,6 +21,17 @@ const versionMapping = {
    "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";
@@ -82,6 +93,26 @@ 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('/');
@@ -149,6 +180,7 @@ function addHfMenu() {
    <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);
@@ -254,6 +286,7 @@ function onLoad() {
    addGithubButton();
    parseGithubButtons();
    addHfMenu();
+    addColabLink();
    platformToggle();
 }

@@ -40,12 +40,12 @@ There are many more parameters that can be configured via the benchmark argument
 ``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::
+.. code-block:: bash

-    >>> ## PYTORCH CODE
+    ## PYTORCH CODE
    python examples/benchmarking/run_benchmark.py --help

-    >>> ## TENSORFLOW CODE
+    ## TENSORFLOW CODE
    python examples/benchmarking/run_benchmark_tf.py --help


@@ -124,7 +124,11 @@ conversion utilities for the following models:
 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. `Other community models <https://huggingface.co/models>`_, contributed by the `community
+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
    <https://huggingface.co/users>`_.

 .. toctree::
@@ -157,14 +161,15 @@ conversion utilities for the following models:
    notebooks
    converting_tensorflow_models
    migration
-    torchscript
    contributing
+    serialization

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

    bertology
+    perplexity
    benchmarks

 .. toctree::
@@ -172,12 +177,13 @@ 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
-    main_classes/trainer
    model_doc/auto
    model_doc/encoderdecoder
    model_doc/bert
@@ -203,3 +209,8 @@ conversion utilities for the following models:
    model_doc/retribert
    model_doc/mobilebert
    model_doc/dpr
+    model_doc/pegasus
+    model_doc/mbart
+    internal/modeling_utils
+    internal/tokenization_utils
+    internal/pipelines_utils
@@ -0,0 +1,88 @@
+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
@@ -0,0 +1,40 @@
+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
@@ -0,0 +1,38 @@
+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
@@ -1,7 +1,9 @@
 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``
 ~~~~~~~~~~~~~~~~~~~~~
@@ -1,23 +1,34 @@
 Models
 ----------------------------------------------------

-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).
+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).

-``PreTrainedModel`` also implements a few methods which are common among all the models to:
+:class:`~transformers.PreTrainedModel` and :class:`~transformers.TFPreTrainedModel` also implement 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) or
+for text generation, :class:`~transformers.generation_utils.GenerationMixin` (for the PyTorch models) and
+:class:`~transformers.generation_tf_utils.TFGenerationMixin` (for the TensorFlow models)
+
+
 ``PreTrainedModel``
 ~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.PreTrainedModel
    :members:

-``Helper Functions``
-~~~~~~~~~~~~~~~~~~~~~

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


 ``TFPreTrainedModel``
@@ -25,3 +36,20 @@ The base class ``PreTrainedModel`` implements the common methods for loading/sav

 .. autoclass:: transformers.TFPreTrainedModel
    :members:
+
+
+``TFModelUtilsMixin``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_tf_utils.TFModelUtilsMixin
+    :members:
+
+
+Generative models
+~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.generation_utils.GenerationMixin
+    :members:
+
+.. autoclass:: transformers.generation_tf_utils.TFGenerationMixin
+    :members:
@@ -0,0 +1,141 @@
+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:
@@ -3,13 +3,24 @@ 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.
+Recognition, Masked Language Modeling, Sentiment Analysis, Feature Extraction and Question Answering. See the
+:doc:`task summary <../task_summary>` for examples of use.

 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, such as :class:`~transformers.TokenClassificationPipeline`
-  or :class:`~transformers.QuestionAnsweringPipeline`
+- 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 pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -21,53 +32,82 @@ other pipeline but requires an additional argument which is the `task`.


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

-Parent class: Pipeline
-=========================================
-
-.. autoclass:: transformers.Pipeline
-    :members: predict, transform, save_pretrained
-
-TokenClassificationPipeline
+ConversationalPipeline
 ==========================================

-.. autoclass:: transformers.TokenClassificationPipeline
+.. autoclass:: transformers.Conversation

-NerPipeline
-==========================================
-
-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
+.. autoclass:: transformers.ConversationalPipeline
+    :special-members: __call__
+    :members:

 FeatureExtractionPipeline
 ==========================================

 .. autoclass:: transformers.FeatureExtractionPipeline
+    :special-members: __call__
+    :members:

-TextClassificationPipeline
+FillMaskPipeline
 ==========================================

-.. autoclass:: transformers.TextClassificationPipeline
+.. 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:
+
+TokenClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.TokenClassificationPipeline
+    :special-members: __call__
+    :members:
+
+ZeroShotClassificationPipeline
+==========================================
+
+.. autoclass:: transformers.ZeroShotClassificationPipeline
+    :special-members: __call__
+    :members:
+
+
+Parent class: :obj:`Pipeline`
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.Pipeline
+    :members:
@@ -1,17 +1,40 @@
 Tokenizer
 ----------------------------------------------------

-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).
+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:

-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).
+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).

-``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` thus implements the main methods for using all the tokenizers:
+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`.

- 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)
+: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).

-``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 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``
 ~~~~~~~~~~~~~~~~~~~~~~~~
@@ -20,6 +43,7 @@ The base classes ``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` impleme
    :special-members: __call__
    :members:

+
 ``PreTrainedTokenizerFast``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -27,14 +51,9 @@ The base classes ``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` impleme
    :special-members: __call__
    :members:

+
 ``BatchEncoding``
 ~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.BatchEncoding
    :members:
-
-``SpecialTokensMixin``
-~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.SpecialTokensMixin
-    :members:
@@ -11,6 +11,23 @@ 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`` 
 ~~~~~~~~~~~

@@ -47,6 +47,16 @@ 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
 ~~~~~~~~~~~~~~~~~~~~

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


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

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


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

@@ -17,13 +17,23 @@ 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
 ~~~~~~~~~~~~~~~~~~~~~
@@ -39,17 +49,6 @@ BartTokenizer
    :members:


-MBartTokenizer
-~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.MBartTokenizer
-    :members: build_inputs_with_special_tokens, prepare_translation_batch
-
-BartForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.BartForConditionalGeneration
-    :members: generate, forward

 BartModel
 ~~~~~~~~~~~~~
@@ -27,13 +27,8 @@ 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 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.
+- 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.

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

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


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

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


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

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


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

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


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

@@ -69,6 +69,19 @@ DPRReaderTokenizerFast
    :members:


+DPR specific outputs
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_dpr.DPRContextEncoderOutput
+    :members:
+
+.. autoclass:: transformers.modeling_dpr.DPRQuestionEncoderOutput
+    :members:
+
+.. autoclass:: transformers.modeling_dpr.DPRReaderOutput
+    :members:
+
+
 DPRContextEncoder
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -71,6 +71,16 @@ ElectraTokenizerFast
    :members:


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

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


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

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


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

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


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

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

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


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

@@ -58,6 +58,16 @@ 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 selecetd few tokens attend "globally" to all other tokens, as it is conventionally done for all tokens in *e.g.* `BertSelfAttention`.
+A selected 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*.
@@ -102,3 +102,25 @@ 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_translation_batch(src_text))
+    translated = model.generate(**tokenizer.prepare_seq2seq_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.',
@@ -86,6 +86,14 @@ 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
@@ -96,16 +104,8 @@ MarianTokenizer
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

 .. autoclass:: transformers.MarianTokenizer
-    :members: prepare_translation_batch
+    :members: prepare_seq2seq_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:
@@ -0,0 +1,37 @@
+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
+
+
@@ -56,6 +56,16 @@ MobileBertTokenizerFast
    :members:


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

@@ -0,0 +1,115 @@
+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
+
+
+
@@ -121,6 +121,13 @@ ReformerForMaskedLM
    :members:


+ReformerForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ReformerForSequenceClassification
+    :members:
+
+
 ReformerForQuestionAnswering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -63,6 +63,13 @@ 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_1>``, ``<extra_id_2>``, ... up to ``<extra_id_100>``. 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_0>``, ``<extra_id_1>``, ... up to ``<extra_id_99>``. 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_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')
+  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')
  # the forward function automatically creates the correct decoder_input_ids
  model(input_ids=input_ids, labels=labels)

@@ -54,6 +54,22 @@ 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
 ~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -46,6 +46,14 @@ 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
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

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


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

@@ -50,6 +50,49 @@ 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
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@@ -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">
+   <a href="model_doc/gpt.html">
       <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">
+   <a href="model_doc/gpt2.html">
       <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">
+   <a href="model_doc/ctrl.html">
       <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">
+   <a href="model_doc/transformerxl.html">
       <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">
+   <a href="model_doc/reformer.html">
       <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">
+   <a href="model_doc/xlnet.html">
       <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 TransformerXL to build long-term dependencies. 
+XLNet also uses the same recurrence mechanism as Transformer-XL 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">
+   <a href="model_doc/bert.html">
       <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">
+   <a href="model_doc/albert.html">
       <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">
+   <a href="model_doc/roberta.html">
       <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">
+   <a href="model_doc/distilbert.html">
       <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">
+   <a href="model_doc/xlm.html">
       <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">
+   <a href="model_doc/xlmroberta.html">
       <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">
+   <a href="model_doc/flaubert.html">
       <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">
+   <a href="model_doc/electra.html">
       <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">
+   <a href="model_doc/longformer.html">
       <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">
+   <a href="model_doc/bart.html">
       <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">
+   <a href="model_doc/marian.html">
       <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">
+   <a href="model_doc/t5.html">
       <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>`.
+    More information in this :doc:`model documentation </model_doc/mmbt.html>`.
    TODO: write this page

 More technical aspects
@@ -0,0 +1,151 @@
+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.
@@ -20,7 +20,7 @@ work properly.
 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')
@@ -31,33 +31,24 @@ 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.

-::
-
-    encoded_input = tokenizer("Hello, I'm a single sentence!")
-    print(encoded_input)
-
-This will return a dictionary string to list of ints like this one:
-
-::
+.. 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:

-::
-
-    tokenizer.decode(encoded_input["input_ids"])
-
-which should return
-
-::
+.. 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
@@ -69,18 +60,13 @@ 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:

-::
-
-    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)
-
-We get back a dictionary once again, this time with values being list of list of ints:
-
-::
+.. 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]],
@@ -91,6 +77,8 @@ We get back a dictionary once again, this time with values being list of list of
                        [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:

@@ -100,19 +88,11 @@ probably want:

 You can do all of this by using the following options when feeding your list of sentences to the tokenizer:

-::
-
-    ## PYTORCH CODE
-    batch = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="pt")
-    print(batch)
-    ## TENSORFLOW CODE
-    batch = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="tf")
-    print(batch)
-
-which should now return a dictionary string to tensor like this:
-
-::
+.. 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]]),
@@ -122,9 +102,22 @@ which should now return a dictionary string to tensor like this:
     '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]])}

-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).
+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
@@ -137,26 +130,16 @@ 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:
-
-::
-
-    [CLS] Sequence A [SEP] Sequence B [SEP]
+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).
-
-
-::
-
-    encoded_input = tokenizer("How old are you?", "I'm 6 years old")
-    print(encoded_input)
-
 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]}
@@ -169,34 +152,24 @@ 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.

-::
-
-    tokenizer.decode(encoded_input["input_ids"])
-
-will return:
-
-::
+.. 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:

-::
-
-    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)
-
-will return a dict with the values being list of lists of ints:
-
-::
+.. 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]], 
@@ -207,17 +180,14 @@ will return a dict with the values being list of lists of ints:
                       [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:

-::
-
-    for ids in encoded_inputs["input_ids"]:
-        print(tokenizer.decode(ids))
-
-which will return:
-
-::
+.. 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]
@@ -225,7 +195,7 @@ which will return:
 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")
@@ -316,17 +286,12 @@ predictions in `named entity recognition (NER) <https://en.wikipedia.org/wiki/Na
 `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:
+tokenizer. For instance, we have:

-::
-
-    encoded_input = tokenizer(["Hello", "I'm", "a", "single", "sentence"], is_pretokenized=True)
-    print(encoded_input)
-
-will return:
-
-::
+.. 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]}
@@ -337,7 +302,7 @@ Note that the tokenizer still adds the ids of special tokens (if applicable) unl
 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"],
@@ -346,7 +311,7 @@ like this:

 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"],
@@ -355,7 +320,7 @@ or a batch of pair sentences like this:

 And you can add padding, truncation as well as directly return tensors like before:

-::
+.. code-block::

    ## PYTORCH CODE
    batch = tokenizer(batch_sentences,
@@ -74,14 +74,16 @@ 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.                                                               |
-|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
+|                   |                                                            | | 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.                  |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (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 using Whole-Word-Masking. Text is tokenized with MeCab and WordPiece.                                      |
-|                   |                                                            | | `MeCab <https://taku910.github.io/mecab/>`__ is required for tokenization.                                                          |
+|                   |                                                            | | 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.                  |
 |                   |                                                            |                                                                                                                                       |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -351,6 +353,8 @@ 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                                                     |
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
@@ -108,11 +108,11 @@ any other model from the model hub):
    >>> 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)
+    >>> 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) 
+    >>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name, from_pt=True)
    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
    >>> classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)

@@ -128,7 +128,7 @@ 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
@@ -146,7 +146,7 @@ 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
+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.

@@ -191,7 +191,7 @@ and get tensors back. You can specify all of that to the tokenizer:
    ...     return_tensors="tf"
    ... )

-The padding is automatically applied on the side the model expect it (in this case, on the right), with the
+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::
@@ -212,9 +212,9 @@ You can learn more about tokenizers :doc:`here <preprocessing>`.
 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:`**`.
+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::

@@ -235,9 +235,11 @@ final activations of the model.
    >>> ## TENSORFLOW CODE
    >>> print(tf_outputs)
    (<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
-    array([[-4.0832963 ,  4.3364134 ],
-           [ 0.08181238, -0.04178794]], dtype=float32)>,)
+    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
@@ -262,7 +264,7 @@ We can see we get the numbers from before:
    >>> print(tf_predictions)
    tf.Tensor(
    [[2.2042994e-04 9.9977952e-01]
-     [5.3086078e-01 4.6913919e-01]], shape=(2, 2), dtype=float32)
+     [5.3086340e-01 4.6913657e-01]], shape=(2, 2), dtype=float32)
    >>> ## PYTORCH CODE
    >>> print(pt_predictions)
    tensor([[2.2043e-04, 9.9978e-01],
@@ -285,9 +287,15 @@ training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer`
 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.

-Once your model is fine-tuned, you can save it with its tokenizer the following way:
+.. 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)
@@ -297,14 +305,14 @@ directory name instead of the model name. One cool feature of 🤗 Transformers
 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)
@@ -312,7 +320,7 @@ and if you are loading a saved TensorFlow model in a PyTorch model, you should u
 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)
@@ -329,7 +337,9 @@ pretrained model. Behind the scenes, the library has one model class per combina
 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
+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:
@@ -352,7 +362,7 @@ 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
+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.

@@ -0,0 +1,252 @@
+**********************************************
+Exporting transformers models
+**********************************************
+
+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.
+
+
+Also, the conversion tool supports different options which let you tune the behavior of the generated model:
+
+* **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>`_))
+
+
+Optimizations
+------------------------------------------------
+
+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*):
+
+* 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``
+
+.. 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)
@@ -98,8 +98,8 @@ of each other. The process is the following:
    >>> paraphrase = tokenizer(sequence_0, sequence_2, return_tensors="pt")
    >>> not_paraphrase = tokenizer(sequence_0, sequence_1, return_tensors="pt")

-    >>> paraphrase_classification_logits = model(**paraphrase)[0]
-    >>> not_paraphrase_classification_logits = model(**not_paraphrase)[0]
+    >>> paraphrase_classification_logits = model(**paraphrase).logits
+    >>> not_paraphrase_classification_logits = model(**not_paraphrase).logits

    >>> paraphrase_results = torch.softmax(paraphrase_classification_logits, dim=1).tolist()[0]
    >>> not_paraphrase_results = torch.softmax(not_paraphrase_classification_logits, dim=1).tolist()[0]
@@ -375,7 +375,7 @@ Here is an example doing masked language modeling using a model and a tokenizer.
    >>> input = tokenizer.encode(sequence, return_tensors="pt")
    >>> mask_token_index = torch.where(input == tokenizer.mask_token_id)[1]

-    >>> token_logits = model(input)[0]
+    >>> token_logits = model(input).logits
    >>> mask_token_logits = token_logits[0, mask_token_index, :]

    >>> top_5_tokens = torch.topk(mask_token_logits, 5, dim=1).indices[0].tolist()
@@ -436,7 +436,7 @@ Here is an example using the tokenizer and model and leveraging the :func:`~tran
    >>> input_ids = tokenizer.encode(sequence, return_tensors="pt")

    >>> # get logits of last hidden state
-    >>> next_token_logits = model(input_ids)[0][:, -1, :]
+    >>> next_token_logits = model(input_ids).logits[:, -1, :]

    >>> # filter
    >>> filtered_next_token_logits = top_k_top_p_filtering(next_token_logits, top_k=50, top_p=1.0)
@@ -477,7 +477,7 @@ This outputs a (hopefully) coherent next token following the original sequence,

 .. code-block::

-    print(resulting_string)
+    >>> print(resulting_string)
    Hugging Face is based in DUMBO, New York City, and has

 In the next section, we show how this functionality is leveraged in :func:`~transformers.PreTrainedModel.generate` to generate multiple tokens up to a user-defined length.
@@ -604,8 +604,7 @@ expected results:

 .. code-block::

-    print(nlp(sequence))
-
+    >>> print(nlp(sequence))
    [
        {'word': 'Hu', 'score': 0.9995632767677307, 'entity': 'I-ORG'},
        {'word': '##gging', 'score': 0.9915938973426819, 'entity': 'I-ORG'},
@@ -666,7 +665,7 @@ Here is an example doing named entity recognition using a model and a tokenizer.
    >>> tokens = tokenizer.tokenize(tokenizer.decode(tokenizer.encode(sequence)))
    >>> inputs = tokenizer.encode(sequence, return_tensors="pt")

-    >>> outputs = model(inputs)[0]
+    >>> outputs = model(inputs).logits
    >>> predictions = torch.argmax(outputs, dim=2)
    >>> ## TENSORFLOW CODE
    >>> from transformers import TFAutoModelForTokenClassification, AutoTokenizer
@@ -803,11 +802,6 @@ translation results nevertheless.

 Because the translation pipeline depends on the ``PretrainedModel.generate()`` method, we can override the default arguments 
 of ``PretrainedModel.generate()`` directly in the pipeline as is shown for ``max_length`` above.
-This outputs the following translation into German:
-
-::
-
-  Hugging Face ist ein Technologieunternehmen mit Sitz in New York und Paris.
  
 Here is an example doing translation using a model and a tokenizer. The process is the following:

@@ -52,7 +52,7 @@ 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 model rarely have a vocabulary size greater than 50,000, especially if they are trained on a single
+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.
@@ -69,11 +69,11 @@ decomposed as "annoying" and "ly". This is especially useful in agglutinative la
 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 gives the ability to the model to process words it has never seen before, by decomposing them into
+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')
@@ -87,7 +87,7 @@ 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')
@@ -110,7 +110,7 @@ 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.
+: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).
@@ -178,7 +178,7 @@ WordPiece is the subword tokenization algorithm used for :doc:`BERT <model_doc/b
 `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. 
+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
@@ -217,7 +217,7 @@ training corpus. You can then give a probability to each tokenization (which is
 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 are what are used to define the loss that trains the tokenizer: if our corpus consists of the
+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

@@ -229,15 +229,15 @@ tokenizations of :math:`x_{i}` (with the current vocabulary), then the loss is d
 SentencePiece
 =============

-All the methods we have been looking at so far required some from of pretrokenization, which has a central problem: not
+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
-some '▁' characters, that represent spaces. Decoding a tokenized text is then super easy: we just have to concatenate
-all of them together and replace those '▁' by spaces.
+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>`.
@@ -1,135 +0,0 @@
-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)
@@ -16,10 +16,10 @@ TF2, and focus specifically on the nuances and tools for training models in

 Sections:

-  * :ref:`pytorch`
-  * :ref:`tensorflow`
-  * :ref:`trainer`
-  * :ref:`additional-resources`
+  - :ref:`pytorch`
+  - :ref:`tensorflow`
+  - :ref:`trainer`
+  - :ref:`additional-resources`

 .. _pytorch:

@@ -49,7 +49,7 @@ put it in train mode.
 .. code-block:: python

    from transformers import BertForSequenceClassification
-    model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+    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
@@ -99,7 +99,7 @@ backwards pass and update the weights:

    labels = torch.tensor([1,0]).unsqueeze(0)
    outputs = model(input_ids, attention_mask=attention_mask, labels=labels)
-    loss = outputs[0]
+    loss = outputs.loss
    loss.backward()
    optimizer.step()

@@ -111,7 +111,7 @@ The following is equivalent to the previous example:
    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[0])
+    loss = F.cross_entropy(labels, outputs.logitd)
    loss.backward()
    optimizer.step()

@@ -131,7 +131,6 @@ Then all we have to do is call ``scheduler.step()`` after ``optimizer.step()``.

 .. code-block:: python

-    ...
    loss.backward()
    optimizer.step()
    scheduler.step()
@@ -151,7 +150,7 @@ 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

@@ -182,6 +181,7 @@ 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')
@@ -191,7 +191,7 @@ the pretrained tokenizer name.
 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)
@@ -282,7 +282,7 @@ your own ``compute_metrics`` function and pass it to the trainer.

 .. code-block:: python

-    from sklearn.metrics import precision_recall_fscore_support
+    from sklearn.metrics import accuracy_score, precision_recall_fscore_support

    def compute_metrics(pred):
        labels = pred.label_ids
@@ -305,19 +305,14 @@ launching tensorboard in your specified ``logging_dir`` directory.
 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.
+- `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.
+- `🤗 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.
+- `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.
+- `🤗 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

 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 2.1+.
+Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.2+.

 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
@@ -21,7 +21,7 @@ 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    | -  | -  | ✅  | -
@@ -78,3 +78,50 @@ 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
+```
@@ -255,7 +255,11 @@ 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."""
+        """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."""
        return ["contradiction", "entailment", "neutral"]

    def _create_examples(self, lines, set_type):
@@ -268,7 +272,7 @@ 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[-1]
+            label = line[0]
            examples.append(InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label, pairID=pairID))
        return examples

@@ -0,0 +1,10 @@
+# 🤗 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) | 
@@ -1,5 +1,6 @@
 import argparse
 import logging
+import shutil
 import sys
 import unittest
 from unittest.mock import patch
@@ -19,6 +20,10 @@ 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)
@@ -32,7 +37,6 @@ 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
@@ -41,8 +45,13 @@ class PabeeTests(unittest.TestCase):
            --overwrite_output_dir
            --seed=42
            --max_seq_length=128
-            """.split()
+            """
+        output_dir = "./tests/fixtures/tests_samples/temp_dir_{}".format(hash(testargs))
+        testargs += "--output_dir " + output_dir
+        testargs = testargs.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)
@@ -0,0 +1,11 @@
+# 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)
@@ -0,0 +1,54 @@
+# 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",
+}
+```
+
@@ -0,0 +1,33 @@
+#!/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 $ENTROPY \
+    --eval_highway \
+    --overwrite_cache \
+    --per_gpu_eval_batch_size=1
+done
@@ -0,0 +1,30 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+PATH_TO_DATA=/h/xinji/projects/GLUE
+
+MODEL_TYPE=bert  # bert or roberta
+MODEL_SIZE=base  # base or large
+DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
+
+MODEL_NAME=${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
@@ -0,0 +1,720 @@
+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()
@@ -0,0 +1,396 @@
+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)
@@ -0,0 +1,151 @@
+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
@@ -0,0 +1,98 @@
+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)
@@ -0,0 +1,38 @@
+#!/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 \
-        --n_gpu $WORLD_SIZE \
+        --gpus $WORLD_SIZE \
        --student_type distilbert \
        --student_config training_configs/distilbert-base-uncased.json \
        --teacher_type bert \
@@ -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("--n_gpu", type=int, default=1, help="Number of GPUs in the node.")
+    parser.add_argument("--gpus", 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")

@@ -60,3 +60,27 @@ 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 \
+```
@@ -1,14 +1,11 @@
 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
-import torch
-from pytorch_lightning.utilities import rank_zero_info, rank_zero_only
+from pytorch_lightning.utilities import rank_zero_info

 from transformers import (
    AdamW,
@@ -23,7 +20,12 @@ 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,
 )


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


-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)
+# 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) + "}"


 class BaseTransformer(pl.LightningModule):
@@ -63,9 +70,12 @@ class BaseTransformer(pl.LightningModule):
    ):
        """Initialize a model, tokenizer and config."""
        super().__init__()
-        self.hparams = hparams  # TODO: move to self.save_hyperparameters()
+        # TODO: move to self.save_hyperparameters()
+        # self.save_hyperparameters()
+        # can also expand arguments into trainer signature for easier reading
+
+        self.save_hyperparameters(hparams)
        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:
@@ -77,6 +87,13 @@ 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,
@@ -98,8 +115,16 @@ 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 = [
@@ -114,17 +139,10 @@ class BaseTransformer(pl.LightningModule):
        ]
        optimizer = AdamW(optimizer_grouped_parameters, lr=self.hparams.learning_rate, eps=self.hparams.adam_epsilon)
        self.opt = optimizer
-        return [optimizer]

-    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()  # By default, PL will only step every epoch.
-        lrs = {f"lr_group_{i}": lr for i, lr in enumerate(self.lr_scheduler.get_lr())}
-        self.logger.log_metrics(lrs)
+        scheduler = self.get_lr_scheduler()
+
+        return [optimizer], [scheduler]

    def test_step(self, batch, batch_nb):
        return self.validation_step(batch, batch_nb)
@@ -132,26 +150,24 @@ class BaseTransformer(pl.LightningModule):
    def test_epoch_end(self, outputs):
        return self.validation_end(outputs)

-    def train_dataloader(self):
+    def setup(self, step):
        train_batch_size = self.hparams.train_batch_size
-        dataloader = self.load_dataset("train", 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)
+        )

-        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 train_dataloader(self):
+        return self.train_loader

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

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

    def _feature_file(self, mode):
        return os.path.join(
@@ -166,11 +182,9 @@ 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):
@@ -196,21 +210,46 @@ 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("--num_workers", default=4, type=int, help="kwarg passed to DataLoader")
-        parser.add_argument(
-            "--num_train_epochs", default=3, type=int, help="Total number of training epochs to perform."
-        )
-
+        parser.add_argument("--num_train_epochs", dest="max_epochs", default=3, type=int)
        parser.add_argument("--train_batch_size", default=32, type=int)
        parser.add_argument("--eval_batch_size", default=32, type=int)


 class LoggingCallback(pl.Callback):
-    @rank_zero_only
+    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)
+
    def on_validation_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
        rank_zero_info("***** Validation results *****")
        metrics = trainer.callback_metrics
@@ -219,16 +258,15 @@ 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):
-        logger.info("***** Test results *****")
+        rank_zero_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"]:
-                    logger.info("{} = {}\n".format(key, str(metrics[key])))
+                    rank_zero_info("{} = {}\n".format(key, str(metrics[key])))
                    writer.write("{} = {}\n".format(key, str(metrics[key])))


@@ -241,7 +279,6 @@ 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",
@@ -251,26 +288,22 @@ def add_generic_args(parser, root_dir) -> None:
    parser.add_argument(
        "--fp16_opt_level",
        type=str,
-        default="O1",
+        default="O2",
        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("--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("--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("--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(
@@ -283,10 +316,13 @@ 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
@@ -296,38 +332,25 @@ def generic_train(

    train_params = {}

+    # TODO: remove with PyTorch 1.6 since pl uses native amp
    if args.fp16:
-        train_params["use_amp"] = args.fp16
+        train_params["precision"] = 16
        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(
-        logger=logger,
-        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,
+    trainer = pl.Trainer.from_argparse_args(
+        args,
        weights_summary=None,
-        resume_from_checkpoint=args.resume_from_checkpoint,
+        callbacks=[logging_callback] + extra_callbacks,
+        logger=logger,
+        checkpoint_callback=checkpoint_callback,
+        early_stop_callback=early_stopping_callback,
        **train_params,
    )

    if args.do_train:
        trainer.fit(model)
-    trainer.logger.log_hyperparams(args)
-    trainer.logger.save()
+
    return trainer
@@ -204,6 +204,8 @@ if is_tf_available():
            )

        def get_dataset(self):
+            self.dataset = self.dataset.apply(tf.data.experimental.assert_cardinality(len(self.features)))
+
            return self.dataset

        def __len__(self):
@@ -13,147 +13,807 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Fine-tuning the library models for question-answering."""
+""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""


+import argparse
+import glob
 import logging
 import os
-import sys
-from dataclasses import dataclass, field
-from typing import Optional
+import random
+import timeit

-from transformers import AutoConfig, AutoModelForQuestionAnswering, AutoTokenizer, HfArgumentParser, SquadDataset
-from transformers import SquadDataTrainingArguments as DataTrainingArguments
-from transformers import Trainer, TrainingArguments
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from transformers import (
+    MODEL_FOR_QUESTION_ANSWERING_MAPPING,
+    WEIGHTS_NAME,
+    AdamW,
+    AutoConfig,
+    AutoModelForQuestionAnswering,
+    AutoTokenizer,
+    get_linear_schedule_with_warmup,
+    squad_convert_examples_to_features,
+)
+from transformers.data.metrics.squad_metrics import (
+    compute_predictions_log_probs,
+    compute_predictions_logits,
+    squad_evaluate,
+)
+from transformers.data.processors.squad import SquadResult, SquadV1Processor, SquadV2Processor
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter


 logger = logging.getLogger(__name__)

+MODEL_CONFIG_CLASSES = list(MODEL_FOR_QUESTION_ANSWERING_MAPPING.keys())
+MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)

-@dataclass
-class ModelArguments:
-    """
-    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
-    """

-    model_name_or_path: str = field(
-        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+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 to_list(tensor):
+    return tensor.detach().cpu().tolist()
+
+
+def train(args, train_dataset, model, tokenizer):
+    """ 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"]
+    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": args.weight_decay,
+        },
+        {"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=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
    )
-    config_name: Optional[str] = field(
-        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
+        os.path.join(args.model_name_or_path, "scheduler.pt")
+    ):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
+
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
+        )
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size
+        * args.gradient_accumulation_steps
+        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
    )
-    tokenizer_name: Optional[str] = field(
-        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 1
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        try:
+            # set global_step to gobal_step of last saved checkpoint from model path
+            checkpoint_suffix = args.model_name_or_path.split("-")[-1].split("/")[0]
+            global_step = int(checkpoint_suffix)
+            epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+            steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+            logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+            logger.info("  Continuing training from epoch %d", epochs_trained)
+            logger.info("  Continuing training from global step %d", global_step)
+            logger.info("  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch)
+        except ValueError:
+            logger.info("  Starting fine-tuning.")
+
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(
+        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]
    )
-    use_fast: bool = field(default=False, metadata={"help": "Set this flag to use fast tokenization."})
-    # If you want to tweak more attributes on your tokenizer, you should do it in a distinct script,
-    # or just modify its tokenizer_config.json.
-    cache_dir: Optional[str] = field(
-        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    # Added here for reproductibility
+    set_seed(args)
+
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+
+            inputs = {
+                "input_ids": batch[0],
+                "attention_mask": batch[1],
+                "token_type_ids": batch[2],
+                "start_positions": batch[3],
+                "end_positions": batch[4],
+            }
+
+            if args.model_type in ["xlm", "roberta", "distilbert", "camembert"]:
+                del inputs["token_type_ids"]
+
+            if args.model_type in ["xlnet", "xlm"]:
+                inputs.update({"cls_index": batch[5], "p_mask": batch[6]})
+                if args.version_2_with_negative:
+                    inputs.update({"is_impossible": batch[7]})
+                if hasattr(model, "config") and hasattr(model.config, "lang2id"):
+                    inputs.update(
+                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
+                    )
+
+            outputs = model(**inputs)
+            # model outputs are always tuple in transformers (see doc)
+            loss = outputs[0]
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) 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
+
+                # Log metrics
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Only evaluate when single GPU otherwise metrics may not average well
+                    if args.local_rank == -1 and args.evaluate_during_training:
+                        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
+
+                # Save model checkpoint
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    # Take care of distributed/parallel training
+                    model_to_save = model.module if hasattr(model, "module") else model
+                    model_to_save.save_pretrained(output_dir)
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                    logger.info("Saving optimizer and scheduler states to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    dataset, examples, features = load_and_cache_examples(args, tokenizer, evaluate=True, output_examples=True)
+
+    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(args.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(dataset)
+    eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # multi-gpu evaluate
+    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
+        model = torch.nn.DataParallel(model)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+
+    all_results = []
+    start_time = timeit.default_timer()
+
+    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],
+                "token_type_ids": batch[2],
+            }
+
+            if args.model_type in ["xlm", "roberta", "distilbert", "camembert"]:
+                del inputs["token_type_ids"]
+
+            feature_indices = batch[3]
+
+            # XLNet and XLM use more arguments for their predictions
+            if args.model_type in ["xlnet", "xlm"]:
+                inputs.update({"cls_index": batch[4], "p_mask": batch[5]})
+                # for lang_id-sensitive xlm models
+                if hasattr(model, "config") and hasattr(model.config, "lang2id"):
+                    inputs.update(
+                        {"langs": (torch.ones(batch[0].shape, dtype=torch.int64) * args.lang_id).to(args.device)}
+                    )
+            outputs = model(**inputs)
+
+        for i, feature_index in enumerate(feature_indices):
+            eval_feature = features[feature_index.item()]
+            unique_id = int(eval_feature.unique_id)
+
+            output = [to_list(output[i]) for output in outputs]
+
+            # Some models (XLNet, XLM) use 5 arguments for their predictions, while the other "simpler"
+            # models only use two.
+            if len(output) >= 5:
+                start_logits = output[0]
+                start_top_index = output[1]
+                end_logits = output[2]
+                end_top_index = output[3]
+                cls_logits = output[4]
+
+                result = SquadResult(
+                    unique_id,
+                    start_logits,
+                    end_logits,
+                    start_top_index=start_top_index,
+                    end_top_index=end_top_index,
+                    cls_logits=cls_logits,
+                )
+
+            else:
+                start_logits, end_logits = output
+                result = SquadResult(unique_id, start_logits, end_logits)
+
+            all_results.append(result)
+
+    evalTime = timeit.default_timer() - start_time
+    logger.info("  Evaluation done in total %f secs (%f sec per example)", evalTime, evalTime / len(dataset))
+
+    # Compute predictions
+    output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix))
+    output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix))
+
+    if args.version_2_with_negative:
+        output_null_log_odds_file = os.path.join(args.output_dir, "null_odds_{}.json".format(prefix))
+    else:
+        output_null_log_odds_file = None
+
+    # XLNet and XLM use a more complex post-processing procedure
+    if args.model_type in ["xlnet", "xlm"]:
+        start_n_top = model.config.start_n_top if hasattr(model, "config") else model.module.config.start_n_top
+        end_n_top = model.config.end_n_top if hasattr(model, "config") else model.module.config.end_n_top
+
+        predictions = compute_predictions_log_probs(
+            examples,
+            features,
+            all_results,
+            args.n_best_size,
+            args.max_answer_length,
+            output_prediction_file,
+            output_nbest_file,
+            output_null_log_odds_file,
+            start_n_top,
+            end_n_top,
+            args.version_2_with_negative,
+            tokenizer,
+            args.verbose_logging,
+        )
+    else:
+        predictions = compute_predictions_logits(
+            examples,
+            features,
+            all_results,
+            args.n_best_size,
+            args.max_answer_length,
+            args.do_lower_case,
+            output_prediction_file,
+            output_nbest_file,
+            output_null_log_odds_file,
+            args.verbose_logging,
+            args.version_2_with_negative,
+            args.null_score_diff_threshold,
+            tokenizer,
+        )
+
+    # Compute the F1 and exact scores.
+    results = squad_evaluate(examples, predictions)
+    return results
+
+
+def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
+
+    # Load data features from cache or dataset file
+    input_dir = args.data_dir if args.data_dir else "."
+    cached_features_file = os.path.join(
+        input_dir,
+        "cached_{}_{}_{}".format(
+            "dev" if evaluate else "train",
+            list(filter(None, args.model_name_or_path.split("/"))).pop(),
+            str(args.max_seq_length),
+        ),
    )

+    # Init features and dataset from cache if it exists
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features_and_dataset = torch.load(cached_features_file)
+        features, dataset, examples = (
+            features_and_dataset["features"],
+            features_and_dataset["dataset"],
+            features_and_dataset["examples"],
+        )
+    else:
+        logger.info("Creating features from dataset file at %s", input_dir)
+
+        if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
+            try:
+                import tensorflow_datasets as tfds
+            except ImportError:
+                raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")
+
+            if args.version_2_with_negative:
+                logger.warn("tensorflow_datasets does not handle version 2 of SQuAD.")
+
+            tfds_examples = tfds.load("squad")
+            examples = SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=evaluate)
+        else:
+            processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
+            if evaluate:
+                examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
+            else:
+                examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
+
+        features, dataset = squad_convert_examples_to_features(
+            examples=examples,
+            tokenizer=tokenizer,
+            max_seq_length=args.max_seq_length,
+            doc_stride=args.doc_stride,
+            max_query_length=args.max_query_length,
+            is_training=not evaluate,
+            return_dataset="pt",
+            threads=args.threads,
+        )
+
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save({"features": features, "dataset": dataset, "examples": examples}, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+        torch.distributed.barrier()
+
+    if output_examples:
+        return dataset, examples, features
+    return dataset
+

 def main():
-    # See all possible arguments in src/transformers/training_args.py
-    # or by passing the --help flag to this script.
-    # We now keep distinct sets of args, for a cleaner separation of concerns.
+    parser = argparse.ArgumentParser()

-    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    # Required parameters
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_TYPES),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models",
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model checkpoints and predictions will be written.",
+    )

-    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
-        # If we pass only one argument to the script and it's the path to a json file,
-        # let's parse it to get our arguments.
-        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
-    else:
-        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+    # Other parameters
+    parser.add_argument(
+        "--data_dir",
+        default=None,
+        type=str,
+        help="The input data dir. Should contain the .json files for the task."
+        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
+    )
+    parser.add_argument(
+        "--train_file",
+        default=None,
+        type=str,
+        help="The input training file. If a data dir is specified, will look for the file there"
+        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
+    )
+    parser.add_argument(
+        "--predict_file",
+        default=None,
+        type=str,
+        help="The input evaluation file. If a data dir is specified, will look for the file there"
+        + "If no data dir or train/predict files are specified, will run with tensorflow_datasets.",
+    )
+    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(
+        "--version_2_with_negative",
+        action="store_true",
+        help="If true, the SQuAD examples contain some that do not have an answer.",
+    )
+    parser.add_argument(
+        "--null_score_diff_threshold",
+        type=float,
+        default=0.0,
+        help="If null_score - best_non_null is greater than the threshold predict null.",
+    )
+
+    parser.add_argument(
+        "--max_seq_length",
+        default=384,
+        type=int,
+        help="The maximum total input sequence length after WordPiece tokenization. Sequences "
+        "longer than this will be truncated, and sequences shorter than this will be padded.",
+    )
+    parser.add_argument(
+        "--doc_stride",
+        default=128,
+        type=int,
+        help="When splitting up a long document into chunks, how much stride to take between chunks.",
+    )
+    parser.add_argument(
+        "--max_query_length",
+        default=64,
+        type=int,
+        help="The maximum number of tokens for the question. Questions longer than this will "
+        "be truncated to this length.",
+    )
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--evaluate_during_training", action="store_true", help="Run evaluation during training at each logging step."
+    )
+    parser.add_argument(
+        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
+    )
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
+    parser.add_argument(
+        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
+    )
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument(
+        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+    parser.add_argument(
+        "--n_best_size",
+        default=20,
+        type=int,
+        help="The total number of n-best predictions to generate in the nbest_predictions.json output file.",
+    )
+    parser.add_argument(
+        "--max_answer_length",
+        default=30,
+        type=int,
+        help="The maximum length of an answer that can be generated. This is needed because the start "
+        "and end predictions are not conditioned on one another.",
+    )
+    parser.add_argument(
+        "--verbose_logging",
+        action="store_true",
+        help="If true, all of the warnings related to data processing will be printed. "
+        "A number of warnings are expected for a normal SQuAD evaluation.",
+    )
+    parser.add_argument(
+        "--lang_id",
+        default=0,
+        type=int,
+        help="language id of input for language-specific xlm models (see tokenization_xlm.PRETRAINED_INIT_CONFIGURATION)",
+    )
+
+    parser.add_argument("--logging_steps", type=int, default=500, help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=500, 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="Whether not to use 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("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
+    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("--server_ip", type=str, default="", help="Can be used for distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="Can be used for distant debugging.")
+
+    parser.add_argument("--threads", type=int, default=1, help="multiple threads for converting example to features")
+    args = parser.parse_args()
+
+    if args.doc_stride >= args.max_seq_length - args.max_query_length:
+        logger.warning(
+            "WARNING - You've set a doc stride which may be superior to the document length in some "
+            "examples. This could result in errors when building features from the examples. Please reduce the doc "
+            "stride or increase the maximum length to ensure the features are correctly built."
+        )

    if (
-        os.path.exists(training_args.output_dir)
-        and os.listdir(training_args.output_dir)
-        and training_args.do_train
-        and not training_args.overwrite_output_dir
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.overwrite_output_dir
    ):
        raise ValueError(
-            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+            "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 = 0 if args.no_cuda else torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
-        level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
+        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",
-        training_args.local_rank,
-        training_args.device,
-        training_args.n_gpu,
-        bool(training_args.local_rank != -1),
-        training_args.fp16,
+        args.local_rank,
+        device,
+        args.n_gpu,
+        bool(args.local_rank != -1),
+        args.fp16,
    )
-    logger.info("Training/evaluation parameters %s", training_args)

-    # Prepare Question-Answering task
+    # Set seed
+    set_seed(args)
+
    # Load pretrained model and tokenizer
-    #
-    # Distributed training:
-    # The .from_pretrained methods guarantee that only one local process can concurrently
-    # download model & vocab.
+    if args.local_rank not in [-1, 0]:
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()

+    args.model_type = args.model_type.lower()
    config = AutoConfig.from_pretrained(
-        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
-        cache_dir=model_args.cache_dir,
+        args.config_name if args.config_name else args.model_name_or_path,
+        cache_dir=args.cache_dir if args.cache_dir else None,
    )
    tokenizer = AutoTokenizer.from_pretrained(
-        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
-        cache_dir=model_args.cache_dir,
+        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 = AutoModelForQuestionAnswering.from_pretrained(
-        model_args.model_name_or_path,
-        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        args.model_name_or_path,
+        from_tf=bool(".ckpt" in args.model_name_or_path),
        config=config,
-        cache_dir=model_args.cache_dir,
+        cache_dir=args.cache_dir if args.cache_dir else None,
    )

-    # Get datasets
-    is_language_sensitive = hasattr(model.config, "lang2id")
-    train_dataset = (
-        SquadDataset(
-            data_args, tokenizer=tokenizer, is_language_sensitive=is_language_sensitive, cache_dir=model_args.cache_dir
-        )
-        if training_args.do_train
-        else None
-    )
-    eval_dataset = (
-        SquadDataset(
-            data_args,
-            tokenizer=tokenizer,
-            mode="dev",
-            is_language_sensitive=is_language_sensitive,
-            cache_dir=model_args.cache_dir,
-        )
-        if training_args.do_eval
-        else None
-    )
+    if args.local_rank == 0:
+        # Make sure only the first process in distributed training will download model & vocab
+        torch.distributed.barrier()

-    # Initialize our Trainer
-    trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is set.
+    # Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running `--fp16_opt_level="O2"` will
+    # remove the need for this code, but it is still valid.
+    if args.fp16:
+        try:
+            import apex
+
+            apex.amp.register_half_function(torch, "einsum")
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")

    # Training
-    if training_args.do_train:
-        trainer.train(
-            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
-        )
-        trainer.save_model()
-        # For convenience, we also re-save the tokenizer to the same directory,
-        # so that you can share your model easily on huggingface.co/models =)
-        if trainer.is_world_master():
-            tokenizer.save_pretrained(training_args.output_dir)
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

+    # Save the trained model and the tokenizer
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        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()`
+        # Take care of distributed/parallel training
+        model_to_save = model.module if hasattr(model, "module") else model
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)

-def _mp_fn(index):
-    # For xla_spawn (TPUs)
-    main()
+        # 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 = AutoModelForQuestionAnswering.from_pretrained(args.output_dir)  # , force_download=True)
+        tokenizer = AutoTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+    # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        if args.do_train:
+            logger.info("Loading checkpoints saved during training for evaluation")
+            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 model loading logs
+        else:
+            logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
+            checkpoints = [args.model_name_or_path]
+
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+
+        for checkpoint in checkpoints:
+            # Reload the model
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            model = AutoModelForQuestionAnswering.from_pretrained(checkpoint)  # , force_download=True)
+            model.to(args.device)
+
+            # Evaluate
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+
+            result = dict((k + ("_{}".format(global_step) if global_step else ""), v) for k, v in result.items())
+            results.update(result)
+
+    logger.info("Results: {}".format(results))
+
+    return results


 if __name__ == "__main__":
@@ -0,0 +1,160 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Fine-tuning the library models for question-answering."""
+
+
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
+
+from transformers import AutoConfig, AutoModelForQuestionAnswering, AutoTokenizer, HfArgumentParser, SquadDataset
+from transformers import SquadDataTrainingArguments as DataTrainingArguments
+from transformers import Trainer, TrainingArguments
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
+    )
+    use_fast: bool = field(default=False, metadata={"help": "Set this flag to use fast tokenization."})
+    # If you want to tweak more attributes on your tokenizer, you should do it in a distinct script,
+    # or just modify its tokenizer_config.json.
+    cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
+    )
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    if (
+        os.path.exists(training_args.output_dir)
+        and os.listdir(training_args.output_dir)
+        and training_args.do_train
+        and not training_args.overwrite_output_dir
+    ):
+        raise ValueError(
+            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if training_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",
+        training_args.local_rank,
+        training_args.device,
+        training_args.n_gpu,
+        bool(training_args.local_rank != -1),
+        training_args.fp16,
+    )
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Prepare Question-Answering task
+    # Load pretrained model and tokenizer
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
+
+    config = AutoConfig.from_pretrained(
+        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    tokenizer = AutoTokenizer.from_pretrained(
+        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+    )
+    model = AutoModelForQuestionAnswering.from_pretrained(
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
+        config=config,
+        cache_dir=model_args.cache_dir,
+    )
+
+    # Get datasets
+    is_language_sensitive = hasattr(model.config, "lang2id")
+    train_dataset = (
+        SquadDataset(
+            data_args, tokenizer=tokenizer, is_language_sensitive=is_language_sensitive, cache_dir=model_args.cache_dir
+        )
+        if training_args.do_train
+        else None
+    )
+    eval_dataset = (
+        SquadDataset(
+            data_args,
+            tokenizer=tokenizer,
+            mode="dev",
+            is_language_sensitive=is_language_sensitive,
+            cache_dir=model_args.cache_dir,
+        )
+        if training_args.do_eval
+        else None
+    )
+
+    # Initialize our Trainer
+    trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)
+
+    # Training
+    if training_args.do_train:
+        trainer.train(
+            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
+        )
+        trainer.save_model()
+        # For convenience, we also re-save the tokenizer to the same directory,
+        # so that you can share your model easily on huggingface.co/models =)
+        if trainer.is_world_master():
+            tokenizer.save_pretrained(training_args.output_dir)
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()
@@ -21,6 +21,8 @@ import os
 from dataclasses import dataclass, field
 from typing import Optional

+import tensorflow as tf
+
 from transformers import (
    AutoConfig,
    AutoTokenizer,
@@ -68,6 +70,7 @@ class DataTrainingArguments:
    data_dir: Optional[str] = field(
        default=None, metadata={"help": "The input data dir. Should contain the .json files for the SQuAD task."}
    )
+    use_tfds: Optional[bool] = field(default=True, metadata={"help": "If TFDS should be used or not."})
    max_seq_length: int = field(
        default=128,
        metadata={
@@ -170,7 +173,7 @@ def main():
        )

    # Get datasets
-    if not data_args.data_dir:
+    if data_args.use_tfds:
        if data_args.version_2_with_negative:
            logger.warn("tensorflow_datasets does not handle version 2 of SQuAD. Switch to version 1 automatically")

@@ -179,7 +182,7 @@ def main():
        except ImportError:
            raise ImportError("If not data_dir is specified, tensorflow_datasets needs to be installed.")

-        tfds_examples = tfds.load("squad")
+        tfds_examples = tfds.load("squad", data_dir=data_args.data_dir)
        train_examples = (
            SquadV1Processor().get_examples_from_dataset(tfds_examples, evaluate=False)
            if training_args.do_train
@@ -209,6 +212,8 @@ def main():
        else None
    )

+    train_dataset = train_dataset.apply(tf.data.experimental.assert_cardinality(len(train_examples)))
+
    eval_dataset = (
        squad_convert_examples_to_features(
            examples=eval_examples,
@@ -223,6 +228,8 @@ def main():
        else None
    )

+    eval_dataset = eval_dataset.apply(tf.data.experimental.assert_cardinality(len(eval_examples)))
+
    # Initialize our Trainer
    trainer = TFTrainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)

@@ -5,7 +5,7 @@ psutil
 sacrebleu
 rouge-score
 tensorflow_datasets
-pytorch-lightning==0.8.1
+pytorch-lightning==0.8.5
 matplotlib
 git-python==1.0.3
 faiss
@@ -13,3 +13,6 @@ streamlit
 elasticsearch
 pandas
 nlp
+fire
+pytest
+conllu
@@ -3,23 +3,10 @@
 This directory contains examples for finetuning and evaluating transformers on summarization and translation tasks.
 Summarization support is more mature than translation support.
 Please tag @sshleifer with any issues/unexpected behaviors, or send a PR!
-For `bertabs` instructions, see `bertabs/README.md`. 
+For `bertabs` instructions, see [`bertabs/README.md`](bertabs/README.md).


 ### Data
-
-CNN/DailyMail data
-```bash
-cd examples/seq2seq
-wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/cnn_dm.tgz
-tar -xzvf cnn_dm.tgz
-
-export CNN_DIR=${PWD}/cnn_dm
-```
-
-this should make a directory called cnn_dm/ with files like `test.source`.
-To use your own data, copy that files format. Each article to be summarized is on its own line.
-
 XSUM Data:
 ```bash
 cd examples/seq2seq
@@ -27,42 +14,79 @@ wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/xsum.tar.gz
 tar -xzvf xsum.tar.gz
 export XSUM_DIR=${PWD}/xsum
 ```
+this should make a directory called `xsum/` with files like `test.source`.
+To use your own data, copy that files format. Each article to be summarized is on its own line.

-
-WMT16 English-Romanian Translation Data:
+CNN/DailyMail data
 ```bash
 cd examples/seq2seq
+wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/cnn_dm.tgz
+tar -xzvf cnn_dm.tgz
+export CNN_DIR=${PWD}/cnn_dm
+this should make a directory called `cnn_dm/` with files like `test.source`.
+```
+
+WMT16 English-Romanian Translation Data:
+download with this command:
+```bash
 wget https://s3.amazonaws.com/datasets.huggingface.co/translation/wmt_en_ro.tar.gz
 tar -xzvf wmt_en_ro.tar.gz
 export ENRO_DIR=${PWD}/wmt_en_ro
+this should make a directory called `wmt_en_ro/` with files like `test.source`.
 ```

-If you are using your own data, it must be formatted as one directory with 6 files: train.source, train.target, val.source, val.target, test.source, test.target.  
+If you are using your own data, it must be formatted as one directory with 6 files: train.source, train.target, val.source, val.target, test.source, test.target.
 The `.source` files are the input, the `.target` files are the desired output.


 ### Tips and Tricks

 General Tips:
- since you need to run from `examples/seq2seq`, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run `pip install -e .` before you get started.   
+- since you need to run from `examples/seq2seq`, and likely need to modify code, the easiest workflow is fork transformers, clone your fork, and run `pip install -e .` before you get started.
 - try `--freeze_encoder` or `--freeze_embeds` for faster training/larger batch size.  (3hr per epoch with bs=8, see the "xsum_shared_task" command below)
 - `fp16_opt_level=O1` (the default works best).
 - In addition to the pytorch-lightning .ckpt checkpoint, a transformers checkpoint will be saved.
 Load it with `BartForConditionalGeneration.from_pretrained(f'{output_dir}/best_tfmr)`.
 - At the moment, `--do_predict` does not work in a multi-gpu setting. You need to use `evaluate_checkpoint` or the `run_eval.py` code.
- This warning can be safely ignored: 
+- This warning can be safely ignored:
    > "Some weights of BartForConditionalGeneration were not initialized from the model checkpoint at facebook/bart-large-xsum and are newly initialized: ['final_logits_bias']"
 - Both finetuning and eval are 30% faster with `--fp16`. For that you need to [install apex](https://github.com/NVIDIA/apex#quick-start).
- Read scripts before you run them! 
+- Read scripts before you run them!

 Summarization Tips:
 - (summ) 1 epoch at batch size 1 for bart-large takes 24 hours and requires 13GB GPU RAM with fp16 on an NVIDIA-V100.
 - If you want to run experiments on improving the summarization finetuning process, try the XSUM Shared Task (below). It's faster to train than CNNDM because the summaries are shorter.
 - For CNN/DailyMail, the default `val_max_target_length` and `test_max_target_length` will truncate the ground truth labels, resulting in slightly higher rouge scores. To get accurate rouge scores, you should rerun calculate_rouge on the `{output_dir}/test_generations.txt` file saved by `trainer.test()`
 - `--max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 ` is a reasonable setting for XSUM.
- `wandb` can be used by specifying `--logger wandb`. It is useful for reproducibility. Specify the environment variable `WANDB_PROJECT='hf_xsum'` to do the XSUM shared task.
+- `wandb` can be used by specifying `--logger_name wandb`. It is useful for reproducibility. Specify the environment variable `WANDB_PROJECT='hf_xsum'` to do the XSUM shared task.
 - If you are finetuning on your own dataset, start from `distilbart-cnn-12-6` if you want long summaries and `distilbart-xsum-12-6` if you want short summaries.
-(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods). 
+(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
+
+**Update 2018-07-18**
+Datasets: `Seq2SeqDataset` should be used for all tokenizers without a `prepare_seq2seq_batch` method. For those who do (like Marian, MBart), `TranslationDataset` should be used.**
+A new dataset is needed to support multilingual tasks.
+
+
+### Command Line Options
+
+To see all the possible command line options, run:
+
+```bash
+./finetune.sh --help  # this calls python finetune.py --help
+```
+
+### Finetuning Training Params
+
+To override the pretrained model's training params, you can pass them to `./finetune.sh`:
+
+```bash
+./finetune.sh \
+    [...]
+    --encoder_layerdrop 0.1 \
+    --decoder_layerdrop 0.1 \
+    --dropout 0.1 \
+    --attention_dropout 0.1 \
+```

 ### Summarization Finetuning
 Run/modify `finetune.sh`
@@ -78,23 +102,35 @@ The following command should work on a 16GB GPU:
    --model_name_or_path facebook/bart-large
 ```

-*Note*: The following tips mostly apply to summarization finetuning.
-
 ### Translation Finetuning

 First, follow the wmt_en_ro download instructions.
 Then you can finetune mbart_cc25 on english-romanian with the following command.
-**Recommendation:** Read and potentially modify the fairly opinionated defaults in `train_mbart_cc25_enro.sh` script before running it. 
+**Recommendation:** Read and potentially modify the fairly opinionated defaults in `train_mbart_cc25_enro.sh` script before running it.
+
+Best performing command:
 ```bash
-export ENRO_DIR=${PWD}/wmt_en_ro   # may need to be fixed depending on where you downloaded
+# optionally
+export ENRO_DIR='wmt_en_ro' # Download instructions above
+# export WANDB_PROJECT="MT" # optional
+export MAX_LEN=128
 export BS=4
-export GAS=8
-./train_mbart_cc25_enro.sh --output_dir cc25_v1_frozen/
+./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --label_smoothing 0.1 --fp16_opt_level=O1 --logger_name wandb --sortish_sampler
 ```
+This should take < 6h/epoch on a 16GB v100 and achieve test BLEU above 26
+To get results in line with fairseq, you need to do some postprocessing. (see `romanian_postprocessing.md`)

-
-### Finetuning Outputs 
-As you train, `output_dir` will be filled with files, that look kind of like this (comments are mine). 
+MultiGPU command
+(using 8 GPUS as an example)
+```bash
+export ENRO_DIR='wmt_en_ro' # Download instructions above
+ # export WANDB_PROJECT="MT" # optional
+export MAX_LEN=128
+export BS=4
+./train_mbart_cc25_enro.sh --output_dir enro_finetune_baseline --gpus 8 --logger_name wandb
+```
+### Finetuning Outputs
+As you train, `output_dir` will be filled with files, that look kind of like this (comments are mine).
 Some of them are metrics, some of them are checkpoints, some of them are metadata. Here is a quick tour:

 ```bash
@@ -107,13 +143,13 @@ output_dir
 │   ├── tokenizer_config.json
 │   └── vocab.json
 ├── git_log.json   # repo, branch, and commit hash
-├── val_avg_rouge2=0.1984-step_count=11.ckpt  # this is a pytorch lightning checkpoint associated with the best val score.
+├── val_avg_rouge2=0.1984-step_count=11.ckpt  # this is a pytorch lightning checkpoint associated with the best val score. (it will be called BLEU for MT)
 ├── metrics.json  # new validation metrics will continually be appended to this
 ├── student  # this is a huggingface checkpoint generated by SummarizationDistiller. It is the student before it gets finetuned.
 │   ├── config.json
 │   └── pytorch_model.bin
-├── test_generations.txt 
-# ^^ are the summaries or translations produced by your best checkpoint on the test data. Populated when training is done 
+├── test_generations.txt
+# ^^ are the summaries or translations produced by your best checkpoint on the test data. Populated when training is done
 ├── test_results.txt  # a convenience file with the test set metrics. This data is also in metrics.json['test']
 ├── hparams.pkl  # the command line args passed after some light preprocessing. Should be saved fairly quickly.
 ```
@@ -123,23 +159,6 @@ from transformers import AutoModelForSeq2SeqLM
 model = AutoModelForSeq2SeqLM.from_pretrained(f'{output_dir}/best_tfmr')
 ```

-#### XSUM Shared Task
-Compare XSUM results with others by using `--logger wandb_shared`. This requires `wandb` registration.
-
-Here is an example command, but you can do whatever you want. Hopefully this will make debugging and collaboration easier!
-```bash
-WANDB_PROJECT='hf_xsum' ./finetune.sh \
-    --data_dir $XSUM_DIR \
-    --output_dir xsum_frozen_embs \
-    --model_name_or_path facebook/bart-large \
-    --train_batch_size 16 --eval_batch_size 16 --freeze_embeds --freeze_encoder \
-    --num_train_epochs 6 \
-    --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
-    --logger wandb
-```
-
-You can see your wandb logs [here](https://app.wandb.ai/sshleifer/hf_xsum?workspace=user-)
-
 ### Evaluation Commands

 To create summaries for each article in dataset, we use `run_eval.py`, here are a few commands that run eval for different tasks and models.
@@ -148,7 +167,7 @@ If 'translation' is in your task name, the computed metric will be BLEU. Otherwi
 For t5, you need to specify --task translation_{src}_to_{tgt} as follows:
 ```bash
 export DATA_DIR=wmt_en_ro
-python run_eval.py t5_base \
+python run_eval.py t5-base \
    $DATA_DIR/val.source t5_val_generations.txt \
    --reference_path $DATA_DIR/val.target \
    --score_path enro_bleu.json \
@@ -181,6 +200,8 @@ python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_
    --task summarization \
    --n_obs 100 \
    --device cuda \
+    --max_source_length 1024 \
+    --max_target_length 56 \
    --fp16 \
    --bs 32
 ```
@@ -190,7 +211,7 @@ python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_
 ![DBART](https://huggingface.co/front/thumbnails/distilbart_large.png)

 For the CNN/DailyMail dataset, (relatively longer, more extractive summaries), we found a simple technique that works:
-you just copy alternating layers from `bart-large-cnn` and finetune more on the same data. 
+you just copy alternating layers from `bart-large-cnn` and finetune more on the same data.

 For the XSUM dataset, that didn’t work as well so we used that same initialization strategy followed by a combination of Distillbert’s ce_loss and the hidden states MSE loss used in the tinybert paper.

@@ -206,7 +227,7 @@ They are initialized by copying layers from the associated `bart-large-{cnn|xsum
 The command that produced `sshleifer/distilbart-cnn-12-6` is
 ```bash
 ./train_distilbart_cnn.sh
-```  
+```
 runtime: 6H on NVIDIA RTX 24GB GPU

 *Note*: You can get the same simple distillation logic by using `./run_distiller.sh --no_teacher` followed by identical arguments as the ones in `train_distilbart_cnn.sh`.
@@ -222,15 +243,15 @@ This is how `sshleifer/distilbart-xsum*` checkpoints were produced.
 The command that produced `sshleifer/distilbart-xsum-12-6` is:

 ```bash
-./train_distilbart_xsum.sh  
+./train_distilbart_xsum.sh
 ```

-runtime: 13H on V-100 16GB GPU. 
+runtime: 13H on V-100 16GB GPU.

 ### Contributing
 - follow the standard contributing guidelines and code of conduct.
 - add tests to `test_seq2seq_examples.py`
 - To run only the seq2seq tests, you must be in the root of the repository and run:
 ```bash
-pytest examples/seq2seq/  
+pytest examples/seq2seq/
 ```
@@ -5,7 +5,7 @@ from pathlib import Path
 import numpy as np
 import pytorch_lightning as pl
 import torch
-from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import EarlyStopping, ModelCheckpoint
 from pytorch_lightning.utilities import rank_zero_only


@@ -19,6 +19,10 @@ logger = logging.getLogger(__name__)


 class Seq2SeqLoggingCallback(pl.Callback):
+    def on_batch_end(self, trainer, pl_module):
+        lrs = {f"lr_group_{i}": param["lr"] for i, param in enumerate(pl_module.trainer.optimizers[0].param_groups)}
+        pl_module.logger.log_metrics(lrs)
+
    @rank_zero_only
    def _write_logs(
        self, trainer: pl.Trainer, pl_module: pl.LightningModule, type_path: str, save_generations=True
@@ -90,3 +94,7 @@ def get_checkpoint_callback(output_dir, metric):
        period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.
    )
    return checkpoint_callback
+
+
+def get_early_stopping_callback(metric, patience):
+    return EarlyStopping(monitor=f"val_{metric}", mode="max", patience=patience, verbose=True,)
@@ -0,0 +1,21 @@
+from typing import Union
+
+import fire
+import torch
+from tqdm import tqdm
+
+
+def convert(src_path: str, map_location: str = "cpu", save_path: Union[str, None] = None) -> None:
+    """Convert a pytorch_model.bin or model.pt file to torch.float16 for faster downloads, less disk space."""
+    state_dict = torch.load(src_path, map_location=map_location)
+    for k, v in tqdm(state_dict.items()):
+        if not isinstance(v, torch.Tensor):
+            raise TypeError("FP16 conversion only works on paths that are saved state dics, like pytorch_model.bin")
+        state_dict[k] = v.half()
+    if save_path is None:  # overwrite src_path
+        save_path = src_path
+    torch.save(state_dict, save_path)
+
+
+if __name__ == "__main__":
+    fire.Fire(convert)
@@ -10,32 +10,39 @@ from torch import nn
 from torch.nn import functional as F

 from lightning_base import generic_train
-from transformers import AdamW, BartConfig, BartForConditionalGeneration, T5Config, T5ForConditionalGeneration
+from transformers import (
+    AdamW,
+    BartConfig,
+    BartForConditionalGeneration,
+    MBartTokenizer,
+    T5Config,
+    T5ForConditionalGeneration,
+)


 try:
-    from .finetune import SummarizationModule
+    from .finetune import SummarizationModule, TranslationModule
    from .initialization_utils import init_student, copy_layers
    from .utils import (
        use_task_specific_params,
-        SummarizationDataset,
        pickle_load,
        freeze_params,
        assert_all_frozen,
        any_requires_grad,
+        calculate_bleu_score,
    )
    from .finetune import main as ft_main
 except ImportError:
-    from finetune import SummarizationModule
+    from finetune import SummarizationModule, TranslationModule
    from finetune import main as ft_main
    from initialization_utils import init_student, copy_layers
    from utils import (
        use_task_specific_params,
-        SummarizationDataset,
        pickle_load,
        freeze_params,
        assert_all_frozen,
        any_requires_grad,
+        calculate_bleu_score,
    )


@@ -115,11 +122,6 @@ class BartSummarizationDistiller(SummarizationModule):
        if self.different_encoder:
            copy_layers(teacher.encoder.block, student.encoder.block, e_layers_to_copy)

-    def get_dataset(self, type_path) -> SummarizationDataset:
-        n_obs = self.n_obs[type_path]
-        dataset = SummarizationDataset(self.tokenizer, type_path=type_path, n_obs=n_obs, **self.dataset_kwargs)
-        return dataset
-
    def calc_mse_loss(self, teacher_outputs: torch.Tensor, student_outputs: torch.Tensor, mask) -> torch.FloatTensor:
        if mask is not None:
            # mask has False at padding_idx
@@ -177,17 +179,7 @@ class BartSummarizationDistiller(SummarizationModule):
    @staticmethod
    def add_model_specific_args(parser, root_dir):
        SummarizationModule.add_model_specific_args(parser, root_dir)
-        parser.add_argument("--teacher", default="facebook/bart-large-cnn", type=str)
-        parser.add_argument("--alpha_ce", default=0.8, type=float)
-        parser.add_argument("--alpha_mlm", default=0.2, type=float)
-        # parser.add_argument("--alpha_cos", default=0.0, type=float)
-        parser.add_argument("--alpha_encoder_loss", default=0.0, type=float)
-        parser.add_argument("--alpha_hid", default=0.0, type=float, required=False)
-        parser.add_argument("--student_decoder_layers", default=12, type=int, required=False)
-        parser.add_argument("--student_encoder_layers", default=12, type=int, required=False)
-        parser.add_argument("--no_teacher", action="store_true", default=False)
-        parser.add_argument("--length_penalty", type=float, default=-1)
-
+        add_distill_args(parser)
        return parser

    def _step(self, batch):
@@ -265,6 +257,44 @@ class BartSummarizationDistiller(SummarizationModule):
        return sum(hidden_losses)


+def add_distill_args(parser):
+    parser.add_argument("--teacher", default="facebook/bart-large-cnn", type=str)
+    parser.add_argument("--alpha_ce", default=0.8, type=float)
+    parser.add_argument("--alpha_mlm", default=0.2, type=float)
+    parser.add_argument("--alpha_encoder_loss", default=0.0, type=float)
+    parser.add_argument("--alpha_hid", default=0.0, type=float, required=False)
+    parser.add_argument("--student_decoder_layers", default=12, type=int, required=False)
+    parser.add_argument("--student_encoder_layers", default=12, type=int, required=False)
+    parser.add_argument("--no_teacher", action="store_true", default=False)
+    parser.add_argument("--length_penalty", type=float, default=-1)
+
+
+class BartTranslationDistiller(BartSummarizationDistiller):
+    mode = "translation"
+    loss_names = ["loss"]
+    metric_names = ["bleu"]
+    val_metric = "bleu"
+
+    def __init__(self, hparams, **kwargs):
+        super().__init__(hparams, **kwargs)
+        assert isinstance(self.tokenizer, MBartTokenizer)
+        assert hparams.src_lang is not None
+        assert hparams.tgt_lang is not None
+        self.dataset_kwargs["src_lang"] = hparams.src_lang
+        self.dataset_kwargs["tgt_lang"] = hparams.tgt_lang
+        if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
+            self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]
+
+    def calc_generative_metrics(self, preds, target) -> dict:
+        return calculate_bleu_score(preds, target)
+
+    @staticmethod
+    def add_model_specific_args(parser, root_dir):
+        TranslationModule.add_model_specific_args(parser, root_dir)
+        add_distill_args(parser)
+        return parser
+
+
 class T5SummarizationDistiller(BartSummarizationDistiller):
    def pre_init(self, hparams):
        raise NotImplementedError("T5 Distillation does not work yet")
@@ -382,15 +412,14 @@ class T5SummarizationDistiller(BartSummarizationDistiller):
 def create_module(args):
    t5 = "t5" in args.model_name_or_path
    if args.no_teacher:
-        assert not args.enc_only
-        module_cls = SummarizationModule
-    elif t5:
+        module_cls = TranslationModule if "translation" in args.task else SummarizationModule
+    elif t5:  # DISTILL T5 WITH TEACHER FOR SUMMARIZATION
+        assert "translation" not in args.task, "t5 translation distillation not supported"
        module_cls = T5SummarizationDistiller
-    elif args.enc_only:
-        raise ValueError("Deleted that")
-    else:
-        module_cls = BartSummarizationDistiller
+    else:  # DISTILL WITH TEACHER
+        module_cls = BartTranslationDistiller if "translation" in args.task else BartSummarizationDistiller
    args.setup_cls: str = module_cls.__name__
+    print(f"using module {args.setup_cls}")
    model = module_cls(args)
    return model

@@ -431,6 +460,18 @@ def get_layers_to_copy(n_to_get, tot):
            12: all_layers,
        }
        return layers_to_copy[n_to_get]
+    elif tot == 16:
+        layers_to_copy = {  # maps  num layers in student -> which teacher layers to copy
+            1: [0],
+            2: [0, 8],
+            3: [0, 8, 15],
+            4: [0, 5, 10, 15],
+            6: [0, 3, 6, 9, 12, 15],
+            8: [0, 2, 4, 6, 8, 10, 12, 15],
+            9: [0, 1, 3, 5, 7, 9, 11, 13, 15],
+            16: all_layers,
+        }
+        return layers_to_copy[n_to_get]
    else:
        return all_layers[:n_to_get]  # TODO: better version on theseus-bart branch

@@ -446,6 +487,7 @@ def distill_main(args):

 if __name__ == "__main__":
    parser = argparse.ArgumentParser()
+    parser = pl.Trainer.add_argparse_args(parser)
    parser = BartSummarizationDistiller.add_model_specific_args(parser, os.getcwd())
    args = parser.parse_args()

@@ -0,0 +1,52 @@
+from pathlib import Path
+
+import fire
+from tqdm import tqdm
+
+
+def download_wmt_dataset(src_lang="ro", tgt_lang="en", dataset="wmt16", save_dir=None) -> None:
+    """Download a dataset using the nlp package and save it to the format expected by finetune.py
+    Format of save_dir: train.source, train.target, val.source, val.target, test.source, test.target.
+
+    Args:
+        src_lang: <str> source language
+        tgt_lang: <str> target language
+        dataset: <str> wmt16, wmt17, etc. wmt16 is a good start as it's small. To get the full list run `import nlp; print([d.id for d in nlp.list_datasets() if "wmt" in d.id])`
+        save_dir: <str>, where to save the datasets, defaults to f'{dataset}-{src_lang}-{tgt_lang}'
+
+    Usage:
+        >>> download_wmt_dataset('ro', 'en', dataset='wmt16') # saves to wmt16-ro-en
+    """
+    try:
+        import nlp
+    except (ModuleNotFoundError, ImportError):
+        raise ImportError("run pip install nlp")
+    pair = f"{src_lang}-{tgt_lang}"
+    print(f"Converting {dataset}-{pair}")
+    ds = nlp.load_dataset(dataset, pair)
+    if save_dir is None:
+        save_dir = f"{dataset}-{pair}"
+    save_dir = Path(save_dir)
+    save_dir.mkdir(exist_ok=True)
+
+    for split in ds.keys():
+        print(f"Splitting {split} with {ds[split].num_rows} records")
+
+        # to save to val.source, val.target like summary datasets
+        fn = "val" if split == "validation" else split
+        src_path = save_dir.joinpath(f"{fn}.source")
+        tgt_path = save_dir.joinpath(f"{fn}.target")
+        src_fp = src_path.open("w+")
+        tgt_fp = tgt_path.open("w+")
+
+        # reader is the bottleneck so writing one record at a time doesn't slow things down
+        for x in tqdm(ds[split]):
+            ex = x["translation"]
+            src_fp.write(ex[src_lang] + "\n")
+            tgt_fp.write(ex[tgt_lang] + "\n")
+
+    print(f"Saved {dataset} dataset to {save_dir}")
+
+
+if __name__ == "__main__":
+    fire.Fire(download_wmt_dataset)
@@ -14,14 +14,13 @@ import torch
 from torch.utils.data import DataLoader

 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from transformers import MBartTokenizer, get_linear_schedule_with_warmup
+from transformers import MarianTokenizer, MBartTokenizer, T5ForConditionalGeneration, get_linear_schedule_with_warmup


 try:
    from .utils import (
        assert_all_frozen,
        use_task_specific_params,
-        SummarizationDataset,
        lmap,
        flatten_list,
        pickle_save,
@@ -32,12 +31,18 @@ try:
        get_git_info,
        ROUGE_KEYS,
        calculate_bleu_score,
+        Seq2SeqDataset,
+        TranslationDataset,
+        label_smoothed_nll_loss,
    )
-    from .callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback
+
+    from .callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback, get_early_stopping_callback
 except ImportError:
    from utils import (
+        Seq2SeqDataset,
+        TranslationDataset,
+        assert_all_frozen,
        use_task_specific_params,
-        SummarizationDataset,
        lmap,
        flatten_list,
        pickle_save,
@@ -48,9 +53,9 @@ except ImportError:
        get_git_info,
        ROUGE_KEYS,
        calculate_bleu_score,
-        assert_all_frozen,
+        label_smoothed_nll_loss,
    )
-    from callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback
+    from callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback, get_early_stopping_callback

 logger = logging.getLogger(__name__)

@@ -100,6 +105,13 @@ class SummarizationModule(BaseTransformer):
        self.hparams.git_sha = get_git_info()["repo_sha"]
        self.num_workers = hparams.num_workers
        self.decoder_start_token_id = None
+        if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
+            self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]
+            self.model.config.decoder_start_token_id = self.decoder_start_token_id
+        if isinstance(self.tokenizer, MBartTokenizer) or isinstance(self.tokenizer, MarianTokenizer):
+            self.dataset_class = TranslationDataset
+        else:
+            self.dataset_class = Seq2SeqDataset

    def freeze_embeds(self):
        """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
@@ -124,17 +136,40 @@ class SummarizationModule(BaseTransformer):

    def _step(self, batch: dict) -> Tuple:
        pad_token_id = self.tokenizer.pad_token_id
-        source_ids, source_mask, y = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
-        y_ids = y[:, :-1].contiguous()
-        lm_labels = y[:, 1:].clone()
-        lm_labels[y[:, 1:] == pad_token_id] = -100
-        outputs = self(source_ids, attention_mask=source_mask, decoder_input_ids=y_ids, labels=lm_labels,)
-        loss = outputs[0]
+        source_ids, source_mask, target_ids = batch["input_ids"], batch["attention_mask"], batch["decoder_input_ids"]
+
+        if isinstance(self.model, T5ForConditionalGeneration):
+            decoder_input_ids = self.model._shift_right(target_ids)
+            lm_labels = target_ids
+        else:
+            decoder_input_ids = target_ids[:, :-1].contiguous()  # Why this line?
+            lm_labels = target_ids[:, 1:].clone()  # why clone?
+
+        outputs = self(source_ids, attention_mask=source_mask, decoder_input_ids=decoder_input_ids, use_cache=False)
+
+        if self.hparams.label_smoothing == 0:
+            # Same behavior as modeling_bart.py
+            loss_fct = torch.nn.CrossEntropyLoss(ignore_index=pad_token_id)
+            lm_logits = outputs[0]
+            assert lm_logits.shape[-1] == self.model.config.vocab_size
+            loss = loss_fct(lm_logits.view(-1, lm_logits.shape[-1]), lm_labels.view(-1))
+        else:
+            lprobs = torch.nn.functional.log_softmax(outputs[0], dim=-1)
+            loss, nll_loss = label_smoothed_nll_loss(
+                lprobs, lm_labels, self.hparams.label_smoothing, ignore_index=pad_token_id
+            )
        return (loss,)

+    @property
+    def pad(self) -> int:
+        return self.tokenizer.pad_token_id
+
    def training_step(self, batch, batch_idx) -> Dict:
        loss_tensors = self._step(batch)
+
        logs = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
+        # tokens per batch
+        logs["tpb"] = batch["input_ids"].ne(self.pad).sum() + batch["decoder_input_ids"].ne(self.pad).sum()
        return {"loss": loss_tensors[0], "log": logs}

    def validation_step(self, batch, batch_idx) -> Dict:
@@ -144,7 +179,7 @@ class SummarizationModule(BaseTransformer):
        self.step_count += 1
        losses = {k: torch.stack([x[k] for x in outputs]).mean() for k in self.loss_names}
        loss = losses["loss"]
-        rouges = {k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "summ_len"]}
+        rouges = {k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "gen_len"]}
        rouge_tensor: torch.FloatTensor = torch.tensor(rouges[self.val_metric]).type_as(loss)
        rouges.update({k: v.item() for k, v in losses.items()})
        losses.update(rouges)
@@ -162,23 +197,21 @@ class SummarizationModule(BaseTransformer):
        return calculate_rouge(preds, target)

    def _generative_step(self, batch: dict) -> dict:
-        pad_token_id = self.tokenizer.pad_token_id
-        source_ids, source_mask, y = SummarizationDataset.trim_seq2seq_batch(batch, pad_token_id)
        t0 = time.time()
        generated_ids = self.model.generate(
-            input_ids=source_ids,
-            attention_mask=source_mask,
+            batch["input_ids"],
+            attention_mask=batch["attention_mask"],
            use_cache=True,
            decoder_start_token_id=self.decoder_start_token_id,
        )
-        gen_time = (time.time() - t0) / source_ids.shape[0]
-        preds = self.ids_to_clean_text(generated_ids)
-        target = self.ids_to_clean_text(y)
+        gen_time = (time.time() - t0) / batch["input_ids"].shape[0]
+        preds: List[str] = self.ids_to_clean_text(generated_ids)
+        target: List[str] = self.ids_to_clean_text(batch["decoder_input_ids"])
        loss_tensors = self._step(batch)
        base_metrics = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
        rouge: Dict = self.calc_generative_metrics(preds, target)
        summ_len = np.mean(lmap(len, generated_ids))
-        base_metrics.update(gen_time=gen_time, summ_len=summ_len, preds=preds, target=target, **rouge)
+        base_metrics.update(gen_time=gen_time, gen_len=summ_len, preds=preds, target=target, **rouge)
        return base_metrics

    def test_step(self, batch, batch_idx):
@@ -187,10 +220,10 @@ class SummarizationModule(BaseTransformer):
    def test_epoch_end(self, outputs):
        return self.validation_epoch_end(outputs, prefix="test")

-    def get_dataset(self, type_path) -> SummarizationDataset:
+    def get_dataset(self, type_path) -> Seq2SeqDataset:
        n_obs = self.n_obs[type_path]
        max_target_length = self.target_lens[type_path]
-        dataset = SummarizationDataset(
+        dataset = self.dataset_class(
            self.tokenizer,
            type_path=type_path,
            n_obs=n_obs,
@@ -221,8 +254,8 @@ class SummarizationModule(BaseTransformer):
        dataloader = self.get_dataloader("train", batch_size=self.hparams.train_batch_size, shuffle=True)
        t_total = (
            (len(dataloader.dataset) // (self.hparams.train_batch_size * max(1, self.hparams.gpus)))
-            // self.hparams.gradient_accumulation_steps
-            * float(self.hparams.num_train_epochs)
+            // self.hparams.accumulate_grad_batches
+            * float(self.hparams.max_epochs)
        )
        scheduler = get_linear_schedule_with_warmup(
            self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
@@ -279,16 +312,23 @@ class SummarizationModule(BaseTransformer):
        parser.add_argument("--freeze_encoder", action="store_true")
        parser.add_argument("--freeze_embeds", action="store_true")
        parser.add_argument("--sortish_sampler", action="store_true", default=False)
-        parser.add_argument("--logger", type=str, choices=["default", "wandb", "wandb_shared"], default="default")
+        parser.add_argument("--logger_name", type=str, choices=["default", "wandb", "wandb_shared"], default="default")
        parser.add_argument("--n_train", type=int, default=-1, required=False, help="# examples. -1 means use all.")
        parser.add_argument("--n_val", type=int, default=500, required=False, help="# examples. -1 means use all.")
        parser.add_argument("--n_test", type=int, default=-1, required=False, help="# examples. -1 means use all.")
        parser.add_argument(
            "--task", type=str, default="summarization", required=False, help="# examples. -1 means use all."
        )
+        parser.add_argument("--label_smoothing", type=float, default=0.0, required=False)
        parser.add_argument("--src_lang", type=str, default="", required=False)
        parser.add_argument("--tgt_lang", type=str, default="", required=False)
-
+        parser.add_argument(
+            "--early_stopping_patience",
+            type=int,
+            default=-1,
+            required=False,
+            help="-1 means never early stop. early_stopping_patience is measured in validation checks, not epochs. So val_check_interval will effect it.",
+        )
        return parser


@@ -302,8 +342,6 @@ class TranslationModule(SummarizationModule):
        super().__init__(hparams, **kwargs)
        self.dataset_kwargs["src_lang"] = hparams.src_lang
        self.dataset_kwargs["tgt_lang"] = hparams.tgt_lang
-        if self.model.config.decoder_start_token_id is None and isinstance(self.tokenizer, MBartTokenizer):
-            self.decoder_start_token_id = self.tokenizer.lang_code_to_id[hparams.tgt_lang]

    def calc_generative_metrics(self, preds, target) -> dict:
        return calculate_bleu_score(preds, target)
@@ -318,27 +356,36 @@ def main(args, model=None) -> SummarizationModule:
            model: SummarizationModule = SummarizationModule(args)
        else:
            model: SummarizationModule = TranslationModule(args)
+
+    dataset = Path(args.data_dir).name
    if (
-        args.logger == "default"
+        args.logger_name == "default"
        or args.fast_dev_run
        or str(args.output_dir).startswith("/tmp")
        or str(args.output_dir).startswith("/var")
    ):
        logger = True  # don't pollute wandb logs unnecessarily
-    elif args.logger == "wandb":
+    elif args.logger_name == "wandb":
        from pytorch_lightning.loggers import WandbLogger

-        logger = WandbLogger(name=model.output_dir.name)
+        project = os.environ.get("WANDB_PROJECT", dataset)
+        logger = WandbLogger(name=model.output_dir.name, project=project)

-    elif args.logger == "wandb_shared":
+    elif args.logger_name == "wandb_shared":
        from pytorch_lightning.loggers import WandbLogger

-        logger = WandbLogger(name=model.output_dir.name)
+        logger = WandbLogger(name=model.output_dir.name, project=f"hf_{dataset}")
+
+    if args.early_stopping_patience >= 0:
+        es_callback = get_early_stopping_callback(model.val_metric, args.early_stopping_patience)
+    else:
+        es_callback = False
    trainer: pl.Trainer = generic_train(
        model,
        args,
        logging_callback=Seq2SeqLoggingCallback(),
        checkpoint_callback=get_checkpoint_callback(args.output_dir, model.val_metric),
+        early_stopping_callback=es_callback,
        logger=logger,
        # TODO: early stopping callback seems messed up
    )
@@ -352,13 +399,17 @@ def main(args, model=None) -> SummarizationModule:
        model.hparams.test_checkpoint = checkpoints[-1]
        trainer.resume_from_checkpoint = checkpoints[-1]
    trainer.logger.log_hyperparams(model.hparams)
-    trainer.test(model)  # this breaks in DDP, known lightning issue. See evaluate_checkpoint to recover metrics.
+
+    # test() without a model tests using the best checkpoint automatically
+    trainer.test()
    return model


 if __name__ == "__main__":
    parser = argparse.ArgumentParser()
+    parser = pl.Trainer.add_argparse_args(parser)
    parser = SummarizationModule.add_model_specific_args(parser, os.getcwd())
+
    args = parser.parse_args()

    main(args)
@@ -2,6 +2,7 @@
 export PYTHONPATH="../":"${PYTHONPATH}"

 # the proper usage is documented in the README, you need to specify data_dir, output_dir and model_name_or_path
+# run ./finetune.sh --help to see all the possible options
 python finetune.py \
    --learning_rate=3e-5 \
    --fp16 \
@@ -10,5 +11,4 @@ python finetune.py \
    --do_predict \
    --n_val 1000 \
    --val_check_interval 0.1 \
-    --sortish_sampler \
    $@
@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+export PYTHONPATH="../":"${PYTHONPATH}"
+
+# From appendix C of paper https://arxiv.org/abs/1912.08777
+# Set --gradient_accumulation_steps  so that effective batch size is 256 (2*128, 4*64, 8*32, 16*16)
+python finetune.py \
+    --learning_rate=1e-4 \
+    --do_train \
+    --do_predict \
+    --n_val 1000 \
+    --val_check_interval 0.25 \
+    --max_source_length 512 --max_target_length 56 \
+    --freeze_embeds --max_target_length 56 --label_smoothing 0.1 \
+    $@
@@ -8,6 +8,7 @@ python finetune.py \
 --eval_batch_size=$BS \
 --output_dir=$OUTPUT_DIR \
 --max_source_length=512 \
+--max_target_length=56 \
 --val_check_interval=0.1 --n_val=200 \
 --do_train --do_predict \
 $@
@@ -0,0 +1,19 @@
+from pathlib import Path
+
+import fire
+
+
+def minify(src_dir: str, dest_dir: str, n: int):
+    """Write first n lines of each file f in src_dir to dest_dir/f """
+    src_dir = Path(src_dir)
+    dest_dir = Path(dest_dir)
+    dest_dir.mkdir(exist_ok=True)
+    for path in src_dir.iterdir():
+        new = [x.rstrip() for x in list(path.open().readlines())][:n]
+        dest_path = dest_dir.joinpath(path.name)
+        print(dest_path)
+        dest_path.open("w").write("\n".join(new))
+
+
+if __name__ == "__main__":
+    fire.Fire(minify)
@@ -0,0 +1,74 @@
+"""Fill examples with bitext up to max_tokens without breaking up examples.
+[['I went', 'yo fui'],
+['to the store', 'a la tienda']
+]
+=> ['I went to the store', 'yo fui a la tienda']
+"""
+
+import argparse
+import shutil
+from pathlib import Path
+
+from tqdm import tqdm
+
+from transformers import AutoTokenizer
+
+
+def pack_examples(tok, src_examples, tgt_examples, max_tokens=1024):
+
+    finished_src, finished_tgt = [], []
+
+    sorted_examples = list(zip(src_examples, tgt_examples))
+    new_src, new_tgt = sorted_examples[0]
+
+    def is_too_big(strang):
+        return tok(strang, return_tensors="pt").input_ids.shape[1] > max_tokens
+
+    for src, tgt in tqdm(sorted_examples[1:]):
+        cand_src = new_src + " " + src
+        cand_tgt = new_tgt + " " + tgt
+        if is_too_big(cand_src) or is_too_big(cand_tgt):  # cant fit, finalize example
+            finished_src.append(new_src)
+            finished_tgt.append(new_tgt)
+            new_src, new_tgt = src, tgt
+        else:  # can fit, keep adding
+            new_src, new_tgt = cand_src, cand_tgt
+
+    # cleanup
+    if new_src:
+        assert new_tgt
+        finished_src.append(new_src)
+        finished_tgt.append(new_tgt)
+    return finished_src, finished_tgt
+
+
+def pack_data_dir(tok, data_dir: Path, max_tokens, save_path):
+    save_path = Path(save_path)
+    save_path.mkdir(exist_ok=True)
+    for split in ["train"]:
+        src_path, tgt_path = data_dir / f"{split}.source", data_dir / f"{split}.target"
+        src_docs = [x.rstrip() for x in Path(src_path).open().readlines()]
+        tgt_docs = [x.rstrip() for x in Path(tgt_path).open().readlines()]
+        packed_src, packed_tgt = pack_examples(tok, src_docs, tgt_docs, max_tokens)
+        print(f"packed {split} split from {len(src_docs)} examples -> {len(packed_src)}.")
+        Path(save_path / f"{split}.source").open("w").write("\n".join(packed_src))
+        Path(save_path / f"{split}.target").open("w").write("\n".join(packed_tgt))
+    for split in ["val", "test"]:
+        src_path, tgt_path = data_dir / f"{split}.source", data_dir / f"{split}.target"
+        shutil.copyfile(src_path, save_path / f"{split}.source")
+        shutil.copyfile(tgt_path, save_path / f"{split}.target")
+
+
+def packer_cli():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--tok_name", type=str, help="like facebook/bart-large-cnn,t5-base, etc.")
+    parser.add_argument("--max_seq_len", type=int, default=128)
+    parser.add_argument("--data_dir", type=str)
+    parser.add_argument("--save_path", type=str)
+    args = parser.parse_args()
+    tokenizer = AutoTokenizer.from_pretrained(args.tok_name)
+    return pack_data_dir(tokenizer, Path(args.data_dir), args.max_seq_len, args.save_path)
+
+
+if __name__ == "__main__":
+    packer_cli()
@@ -0,0 +1,65 @@
+### Motivation
+Without processing, english-> romanian mbart-large-en-ro gets BLEU score 26.8 on the WMT data.
+With post processing, it can score 37..
+Here is the postprocessing code, stolen from @mjpost in this [issue](https://github.com/pytorch/fairseq/issues/1758)
+
+
+
+### Instructions
+Note: You need to have your test_generations.txt before you start this process.
+(1) Setup `mosesdecoder` and `wmt16-scripts`
+```bash
+cd $HOME
+git clone git@github.com:moses-smt/mosesdecoder.git
+cd mosesdecoder  
+git@github.com:rsennrich/wmt16-scripts.git
+```
+
+(2) define a function for post processing.
+ It removes diacritics and does other things I don't understand 
+```bash
+ro_post_process () {
+  sys=$1
+  ref=$2
+  export MOSES_PATH=$HOME/mosesdecoder
+  REPLACE_UNICODE_PUNCT=$MOSES_PATH/scripts/tokenizer/replace-unicode-punctuation.perl
+  NORM_PUNC=$MOSES_PATH/scripts/tokenizer/normalize-punctuation.perl
+  REM_NON_PRINT_CHAR=$MOSES_PATH/scripts/tokenizer/remove-non-printing-char.perl
+  REMOVE_DIACRITICS=$MOSES_PATH/wmt16-scripts/preprocess/remove-diacritics.py
+  NORMALIZE_ROMANIAN=$MOSES_PATH/wmt16-scripts/preprocess/normalise-romanian.py
+  TOKENIZER=$MOSES_PATH/scripts/tokenizer/tokenizer.perl
+
+
+
+  lang=ro
+  for file in $sys $ref; do
+    cat $file \
+    | $REPLACE_UNICODE_PUNCT \
+    | $NORM_PUNC -l $lang \
+    | $REM_NON_PRINT_CHAR \
+    | $NORMALIZE_ROMANIAN \
+    | $REMOVE_DIACRITICS \
+    | $TOKENIZER -no-escape -l $lang \
+    > $(basename $file).tok
+  done
+  # compute BLEU
+  cat $(basename $sys).tok | sacrebleu -tok none -s none -b $(basename $ref).tok
+}
+```
+
+(3) Call the function on test_generations.txt and test.target
+For example,
+```bash
+ro_post_process enro_finetune/test_generations.txt wmt_en_ro/test.target
+```
+This will split out a new blue score and write a new fine called `test_generations.tok` with post-processed outputs.
+
+
+
+
+
+
+
+
+
+```
@@ -30,6 +30,7 @@ def generate_summaries_or_translations(
    device: str = DEFAULT_DEVICE,
    fp16=False,
    task="summarization",
+    decoder_start_token_id=None,
    **gen_kwargs,
 ) -> None:
    fout = Path(out_file).open("w", encoding="utf-8")
@@ -37,6 +38,8 @@ def generate_summaries_or_translations(
    model = AutoModelForSeq2SeqLM.from_pretrained(model_name).to(device)
    if fp16:
        model = model.half()
+    if decoder_start_token_id is None:
+        decoder_start_token_id = gen_kwargs.pop("decoder_start_token_id", None)

    tokenizer = AutoTokenizer.from_pretrained(model_name)

@@ -46,11 +49,14 @@ def generate_summaries_or_translations(
    for batch in tqdm(list(chunks(examples, batch_size))):
        if "t5" in model_name:
            batch = [model.config.prefix + text for text in batch]
-        batch = tokenizer(batch, max_length=1024, return_tensors="pt", truncation=True, padding="max_length").to(
-            device
-        )
+        batch = tokenizer(batch, return_tensors="pt", truncation=True, padding="max_length").to(device)
        input_ids, attention_mask = trim_batch(**batch, pad_token_id=tokenizer.pad_token_id)
-        summaries = model.generate(input_ids=input_ids, attention_mask=attention_mask, **gen_kwargs)
+        summaries = model.generate(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            decoder_start_token_id=decoder_start_token_id,
+            **gen_kwargs,
+        )
        dec = tokenizer.batch_decode(summaries, skip_special_tokens=True, clean_up_tokenization_spaces=False)
        for hypothesis in dec:
            fout.write(hypothesis + "\n")
@@ -68,6 +74,13 @@ def run_generate():
    parser.add_argument("--device", type=str, required=False, default=DEFAULT_DEVICE, help="cuda, cuda:1, cpu etc.")
    parser.add_argument("--task", type=str, default="summarization", help="typically translation or summarization")
    parser.add_argument("--bs", type=int, default=8, required=False, help="batch size")
+    parser.add_argument(
+        "--decoder_start_token_id",
+        type=int,
+        default=None,
+        required=False,
+        help="decoder_start_token_id (otherwise will look at config)",
+    )
    parser.add_argument(
        "--n_obs", type=int, default=-1, required=False, help="How many observations. Defaults to all."
    )
@@ -76,7 +89,7 @@ def run_generate():
    examples = [" " + x.rstrip() if "t5" in args.model_name else x.rstrip() for x in open(args.input_path).readlines()]
    if args.n_obs > 0:
        examples = examples[: args.n_obs]
-
+    Path(args.save_path).parent.mkdir(exist_ok=True)
    generate_summaries_or_translations(
        examples,
        args.save_path,
@@ -85,6 +98,7 @@ def run_generate():
        device=args.device,
        fp16=args.fp16,
        task=args.task,
+        decoder_start_token_id=args.decoder_start_token_id,
    )
    if args.reference_path is None:
        return
@@ -93,6 +107,7 @@ def run_generate():
    output_lns = [x.rstrip() for x in open(args.save_path).readlines()]
    reference_lns = [x.rstrip() for x in open(args.reference_path).readlines()][: len(output_lns)]
    scores: dict = score_fn(output_lns, reference_lns)
+    print(scores)
    if args.score_path is not None:
        json.dump(scores, open(args.score_path, "w+"))
    return scores
@@ -0,0 +1,108 @@
+import argparse
+import os
+import sys
+import tempfile
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+import pytorch_lightning as pl
+import timeout_decorator
+import torch
+
+from transformers import BartForConditionalGeneration
+from transformers.testing_utils import slow
+
+from .finetune import SummarizationModule, main
+from .test_seq2seq_examples import CUDA_AVAILABLE, MBART_TINY
+from .utils import load_json
+
+
+MODEL_NAME = MBART_TINY
+# TODO(SS): MODEL_NAME = "sshleifer/student_mbart_en_ro_1_1"
+
+
+@slow
+@pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
+def test_model_download():
+    """This warms up the cache so that we can time the next test without including download time, which varies between machines."""
+    BartForConditionalGeneration.from_pretrained(MODEL_NAME)
+
+
+@timeout_decorator.timeout(120)
+@slow
+@pytest.mark.skipif(not CUDA_AVAILABLE, reason="too slow to run on CPU")
+def test_train_mbart_cc25_enro_script():
+    data_dir = "examples/seq2seq/test_data/wmt_en_ro"
+    env_vars_to_replace = {
+        "$MAX_LEN": 200,
+        "$BS": 4,
+        "$GAS": 1,
+        "$ENRO_DIR": data_dir,
+        "facebook/mbart-large-cc25": MODEL_NAME,
+        # 1 encoder and 1 decoder layer from finetuned mbart en-ro. Should be able to start >0 and improve quickly.
+        # Download is 600MB in previous test.
+        "val_check_interval=0.25": "val_check_interval=1.0",
+    }
+
+    # Clean up bash script
+    bash_script = Path("examples/seq2seq/train_mbart_cc25_enro.sh").open().read().split("finetune.py")[1].strip()
+    bash_script = bash_script.replace("\\\n", "").strip().replace("$@", "")
+    for k, v in env_vars_to_replace.items():
+        bash_script = bash_script.replace(k, str(v))
+    output_dir = tempfile.mkdtemp(prefix="output")
+
+    if CUDA_AVAILABLE:
+        gpus = 1  # torch.cuda.device_count()
+    else:
+        gpus = 0
+    bash_script = bash_script.replace("--fp16", "")
+    testargs = (
+        ["finetune.py"]
+        + bash_script.split()
+        + [
+            f"--output_dir={output_dir}",
+            f"--gpus={gpus}",
+            "--learning_rate=3e-1",
+            "--warmup_steps=0",
+            "--val_check_interval=1.0",
+            "--tokenizer_name=facebook/mbart-large-en-ro",
+        ]
+    )
+    with patch.object(sys, "argv", testargs):
+        parser = argparse.ArgumentParser()
+        parser = pl.Trainer.add_argparse_args(parser)
+        parser = SummarizationModule.add_model_specific_args(parser, os.getcwd())
+        args = parser.parse_args()
+        args.do_predict = False
+        # assert args.gpus == gpus THIS BREAKS for multigpu
+        model = main(args)
+
+    # Check metrics
+    metrics = load_json(model.metrics_save_path)
+    first_step_stats = metrics["val"][0]
+    last_step_stats = metrics["val"][-1]
+    assert len(metrics["val"]) == (args.max_epochs / args.val_check_interval)  # +1 accounts for val_sanity_check
+
+    assert last_step_stats["val_avg_gen_time"] >= 0.01
+
+    assert first_step_stats["val_avg_bleu"] < last_step_stats["val_avg_bleu"]  # model learned nothing
+    assert 1.0 >= last_step_stats["val_avg_gen_time"]  # model hanging on generate. Maybe bad config was saved.
+    assert isinstance(last_step_stats[f"val_avg_{model.val_metric}"], float)
+
+    # check lightning ckpt can be loaded and has a reasonable statedict
+    contents = os.listdir(output_dir)
+    ckpt_path = [x for x in contents if x.endswith(".ckpt")][0]
+    full_path = os.path.join(args.output_dir, ckpt_path)
+    ckpt = torch.load(full_path, map_location="cpu")
+    expected_key = "model.model.decoder.layers.0.encoder_attn_layer_norm.weight"
+    assert expected_key in ckpt["state_dict"]
+    assert ckpt["state_dict"]["model.model.decoder.layers.0.encoder_attn_layer_norm.weight"].dtype == torch.float32
+
+    # TODO(SS): turn on args.do_predict when PL bug fixed.
+    if args.do_predict:
+        contents = {os.path.basename(p) for p in contents}
+        assert "test_generations.txt" in contents
+        assert "test_results.txt" in contents
+        # assert len(metrics["val"]) ==  desired_n_evals
+        assert len(metrics["test"]) == 1
@@ -0,0 +1,20 @@
+UN Chief Says There Is No Military Solution in Syria Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people. The U.N. chief again urged all parties, including the divided U.N. Security Council, to unite and support inclusive negotiations to find a political solution. Ban told a news conference Wednesday that he plans to meet with foreign ministers of the five permanent council nations - the U.S., Russia, China, Britain and France - on the sidelines of the General Assembly's ministerial session later this month to discuss Syria.
+He expressed regret that divisions in the council and among the Syrian people and regional powers "made this situation unsolvable." Ban urged the five permanent members to show the solidarity and unity they did in achieving an Iran nuclear deal in addressing the Syria crisis. 8 Poll Numbers That Show Donald Trump Is For Real Some have tried to label him a flip-flopper. Others have dismissed him as a joke. And some are holding out for an implosion. But no matter how some Republicans are trying to drag Donald Trump down from atop the polls, it hasn't worked (yet).
+Ten of the last 11 national polls have shown Donald Trump's lead at double digits, and some are starting to ask seriously what it means for the real estate mogul's nomination chances. Of course, it's still early in the election cycle. None of this is to say that Trump is likely to win the Republican nomination. Pundits point out that at this time in 2011, Rick Perry's lead was giving way to a rising Herman Cain, neither of whom won even one state in the nomination process. And there are many reasons he would struggle in a general election. But outside groups like Jeb Bush's Super PAC and the economic conservative group Club for Growth are recognizing Trump's staying power and beginning to unload their dollars to topple him.
+Here are some recent poll numbers that suggest that the real estate mogul isn't just a passing phase: Trump's favorability ratings have turned 180 degrees. Right before Donald Trump announced his candidacy in mid-June, a Monmouth University poll showed only two in 10 Republicans had a positive view of the real estate mogul. By mid-July, it was 40 percent. In early August, it was 52 percent. Now, six in 10 Republicans have a favorable view of Donald Trump. Roughly three in 10 say they have a negative view. And these numbers hold up in early states. A Quinnipiac poll in Iowa last week found that 60 percent of Republicans there had a favorable view of Trump.
+Two-thirds of GOP voters would be happy with Trump as the nominee. In a CNN/ORC poll last week, 67 percent of Republicans said they would be either "enthusiastic" or "satisfied" if Trump were the nominee. Only two in 10 say they would be "upset" if he were the nominee. Only Ben Carson generates roughly the same level of enthusiasm as Trump (43 percent say they would be "enthusiastic" vs. 40 percent who say the same of Trump). The next closest in enthusiasm? Marco Rubio with only 21 percent.
+On the flip side, 47 percent of Republican voters say they would be "dissatisfied" or "upset" if establishment favorite Jeb Bush becomes the nominee. A majority of Republicans don't see Trump's temperament as a problem. While Donald Trump has been widely criticized for his bombast and insults, 52 percent of leaned Republican voters nationwide think that the real estate mogul has the right temperament to be president, according to Monday's ABC News/Washington Post poll. The same number holds in the first-in-the-nation caucus state of Iowa, where the same 52 percent of Republicans think he has the personality to be commander in chief, according to Quinnipiac last week.
+Still, 44 percent think he doesn't have the personality to serve effectively, and almost six in 10 independents say his temperament does not belong in the White House, according to ABC/Post. Republican voters are getting used to the idea. When they put on their pundit hats, Republican voters think Trump is for real. When asked who is most likely to win the GOP nomination, four in 10 said Trump was the best bet, according to a CNN/ORC poll out last week. That's a change from when four in 10 placed their money on Jeb Bush in late July. Full disclosure: GOP voters haven't had the clearest crystal ball in the past.
+At this time last cycle, four in 10 Republicans picked Rick Perry to win the nomination, vs. only 28 percent for eventual nominee Mitt Romney. Still, it shows that a plurality of GOP voters see Trump's campaign as plausible. Even if Republicans rallied around another candidate, Trump still beats almost everyone. Some pundits point out that the splintered field is likely contributing to Trump's lead, while anti-Trump support is be spread diffusely among more than a dozen other candidates. But a Monmouth University poll in early September shows that, in a hypothetical head-to-head matchup between Trump and most other Republican candidates, Trump almost always garners majority support.
+He leads Carly Fiorina by 13 points, Marco Rubio by 14 points, Walker by 15 points, Jeb Bush by 19 points, and, finally, Rand Paul, John Kasich and Chris Christie by 33 points each. He's in a dead heat with Ted Cruz. The only candidate who beats him? Ben Carson would lead the businessman by a wide 19 points in a hypothetical head-to-head. A bare majority of Donald Trump's supporters say they've made up their minds. A new CBS/NYT poll out on Tuesday shows that just more than half of voters who support Trump say they have locked in their votes. Obviously, a lot can happen to change that, and no one can really say they would never change their mind.
+46 percent said they are leaving the door open to switching candidates. Still, Trump's strongest competition at the moment is from fellow outsider neurosurgeon Ben Carson, but voters who say they have made up their minds are twice as likely to go for Trump. Six in 10 Republicans say they agree with Trump on immigration. Even since Donald Trump called immigrants from Mexico "rapists" in his campaign announcement speech two months ago, immigration has been front and center in the 2016 conversation. Some are worried that Trump's bombast will drive crucial Hispanic voters away from the Republican Party and damage rebranding efforts.
+But according to Monday's new ABC/Post poll, six in 10 Republicans say they agree with Trump on immigration issues. So as long as immigration remains in the spotlight, it seems Donald Trump will remain too. Frustration with government is climbing to new highs. Donald Trump and Ben Carson now account for roughly half of the support from Republican voters, largely due to their outsider status. Six in 10 Republicans in Monday's new ABC/Post poll say they want a political outsider over someone with government experience. And they are angry at Washington, too.
+A Des Moines Register/Bloomberg poll in Iowa from two weeks ago shows that three in four Iowa Republicans are frustrated with Republicans in Congress, with 54 percent "unsatisfied" and 21 percent "mad as hell." Jeremy Corbyn to make debut at Prime Minister's Questions Since his election, Mr Corbyn's debut at PMQs has been keenly awaited New Labour leader Jeremy Corbyn is to make his debut at Prime Minister's Questions later, taking on David Cameron for the first time.
+Mr Corbyn will rise to ask the first of his six allotted questions shortly after midday, with his performance likely to be closely scrutinised by the media and Labour MPs. He has called for "less theatre and more facts" at the weekly showpiece. He has also said he could skip some sessions, leaving them to colleagues. The encounter will be the first parliamentary test of Mr Corbyn's leadership, coming after his appointment of a shadow cabinet and his speech to the TUC annual congress on Tuesday.
+Meanwhile, the Labour leader's decision to stand in silence during the singing of the national anthem at a service on Tuesday to mark the 75th anniversary of the Battle of Britain has attracted criticism from a number of Tory MPs and is the focus of several front page stories in the newspapers. Mr Corbyn's decision not to sing the national anthem has attracted attention A spokesman for Mr Corbyn said he had "stood in respectful silence" and did recognise the "heroism of the Royal Air Force in the Battle of Britain."
+But a member of Mr Corbyn's shadow cabinet, Owen Smith, told BBC Two's Newsnight programme he would have advised the Labour leader to sing the national anthem "irrespective" of his belief that the monarchy should be abolished. Nearly a dozen shadow ministers have refused to serve in Mr Corbyn's top team, citing differences over the economy, defence and foreign affairs, while less than a sixth of the parliamentary party originally backed him as leader. BBC political correspondent Robin Brant says policy differences are also "stacking up" within Labour following Mr Corbyn's appointment over its position on the European Union and the government's cap on benefits.
+Mr Corbyn told the TUC conference Labour was putting forward amendments to remove the whole idea of a cap altogether. Hours later Mr Smith, the shadow work and pensions secretary, said the party was "very clear" that it was only opposing government plans to reduce the level of cap from £26,000 to £23,000. Mr Corbyn will be the fifth Labour leader that David Cameron has faced across the despatch box over the past decade since he became Tory leader. The Labour leader, who has promised a different approach to politics, says he has "crowd sourced" ideas for questions to ask Mr Cameron and has been given more than 30,000 suggestions.
+The Islington North MP has said PMQs is too confrontational and that he will refrain from both "repartee" and trading barbs, instead vowing to focus on serious issues such as poverty, inequality and the challenges facing young people. Mr Corbyn has said that Angela Eagle, the shadow business secretary, will deputise for him at PMQs when he does not attend - for instance when Mr Cameron is travelling abroad. He has also floated the idea of allowing other colleagues to take the floor on occasion, saying he had approached the Commons Speaker John Bercow to discuss the issue.
+When he became leader in 2005, Mr Cameron said he wanted to move away from the "Punch and Judy" style of politics often associated with PMQs but admitted some years later that he had failed. Since it was first televised in 1990, PMQs has been seen as a key barometer of a leader's judgement, their command of the Commons and their standing among their fellow MPs although critics have argued it has become a caricature and is in need of far-reaching reforms. 'Shot in Joburg': Homeless youth trained as photographers Downtown Johannesburg is a tough place to be homeless.
+But one group of former street children have found a way to learn a skill and make a living. "I was shot in Joburg" is a non-profit studio that teaches homeless youngsters how to take photographs of their neighbourhood and make a profit from it. BBC News went to meet one of the project's first graduates. JD Sports boss says higher wages could hurt expansion JD Sports Executive Chairman Peter Cowgill says a higher minimum wage for UK workers could mean "more spending power in the pockets of potential consumers." But that spending power is unlikely to outweigh the higher labour costs at his firm, he says.
+The costs could hit JD Sports' expansion plans, he added, which could mean fewer extra jobs. Thanasi Kokkinakis backed by Tennis Australia president Steve Healy Thanasi Kokkinakis deserves kudos rather than criticism for his behaviour. Thanasi Kokkinakis has been the collateral damage in the recent storm around his friend Nick Kyrgios and deserves kudos rather than criticism for his own behaviour, according to Tennis Australia president Steve Healy.
--- a/Show More
+++ b/Show More