Rename files

improve unittests for finetuning, especially w.r.t testing frozen parameters
fix freeze_embeds for T5
add streamlit setup.cfg

.circleci/config.yml

@@ -12,9 +12,11 @@ jobs:
             - checkout
             - run: sudo pip install .[sklearn,tf-cpu,torch,testing]
             - run: sudo pip install codecov pytest-cov
-            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
+            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ --cov  | tee output.txt
             - run: codecov
-
+            - store_artifacts:
+                  path: ~/transformers/output.txt
+                  destination: test_output.txt
     run_tests_torch:
         working_directory: ~/transformers
         docker:
@@ -26,9 +28,11 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install .[sklearn,torch,testing]
-            - run: sudo pip install codecov pytest-cov
-            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
-            - run: codecov
+            - run: 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:
@@ -40,9 +44,10 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install .[sklearn,tf-cpu,testing]
-            - run: sudo pip install codecov pytest-cov
-            - run: python -m pytest -n 8 --dist=loadfile -s -v ./tests/ --cov
-            - run: codecov
+            - run: python -m pytest -n 8 --dist=loadfile -s ./tests/ | tee output.txt
+            - store_artifacts:
+               path: ~/transformers/output.txt
+               destination: test_output.txt
     run_tests_custom_tokenizers:
         working_directory: ~/transformers
         docker:
@@ -52,7 +57,10 @@ jobs:
         steps:
             - checkout
             - run: sudo pip install .[mecab,testing]
-            - run: python -m pytest -sv ./tests/test_tokenization_bert_japanese.py
+            - run: python -m pytest -s ./tests/test_tokenization_bert_japanese.py | tee output.txt
+            - store_artifacts:
+                path: ~/transformers/output.txt
+                destination: test_output.txt
     run_examples_torch:
         working_directory: ~/transformers
         docker:
@@ -65,7 +73,10 @@ jobs:
             - 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 -v ./examples/
+            - run: python -m pytest -n 8 --dist=loadfile -s ./examples/ | tee output.txt
+            - store_artifacts:
+                  path: ~/transformers/output.txt
+                  destination: test_output.txt
     build_doc:
         working_directory: ~/transformers
         docker:

.circleci/deploy.sh

@@ -5,15 +5,26 @@ function deploy_doc(){
 	git checkout $1
 	if [ ! -z "$2" ]
 	then
-		if [ -d "$dir/$2" ]; then
+		if [ "$2" == "master" ]; then
+		    echo "Pushing master"
+			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir/$2/
+			cp -r _build/html/_static .
+		elif ssh -oStrictHostKeyChecking=no $doc "[ -d $dir/$2 ]"; then
 			echo "Directory" $2 "already exists"
+			scp -r -oStrictHostKeyChecking=no _static/* $doc:$dir/$2/_static/
 		else
 			echo "Pushing version" $2
-			make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
+			make clean && make html
+			rm -rf _build/html/_static
+			cp -r _static _build/html
+			scp -r -oStrictHostKeyChecking=no _build/html $doc:$dir/$2
 		fi
 	else
-		echo "Pushing master"
-		make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+		echo "Pushing stable"
+		make clean && make html
+		rm -rf _build/html/_static
+		cp -r _static _build/html
+		scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
 	fi
 }
 
@@ -35,4 +46,5 @@ deploy_doc "11c3257" v2.8.0
 deploy_doc "e7cfc1a" v2.9.0
 deploy_doc "7cb203f" v2.9.1
 deploy_doc "10d7239" v2.10.0 
-deploy_doc "b42586e" #v2.11.0 Latest stable release
+deploy_doc "b42586e" v2.11.0
+deploy_doc "b0892fa" #v3.0.2 Latest stable release

.github/workflows/self-push.yml

@@ -51,4 +51,11 @@ jobs:
         USE_CUDA: yes
       run: |
         source .env/bin/activate
-        python -m pytest -n 2 --dist=loadfile -s -v ./tests/
+        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

.github/workflows/self-scheduled.yml

@@ -46,5 +46,11 @@ jobs:
         USE_CUDA: yes
       run: |
         source .env/bin/activate
-        python -m pytest -n 1 --dist=loadfile -s -v ./tests/
-        
+        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

README.md

@@ -24,6 +24,7 @@
 
 🤗 Transformers (formerly known as `pytorch-transformers` and `pytorch-pretrained-bert`) provides state-of-the-art general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet, T5, CTRL...) for Natural Language Understanding (NLU) and Natural Language Generation (NLG) with over thousands of pretrained models in 100+ languages and deep interoperability between PyTorch & TensorFlow 2.0.
 
+### Recent contributors
 [![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/0)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/0)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/1)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/1)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/2)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/2)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/3)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/3)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/4)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/4)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/5)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/5)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/6)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/6)[![](https://sourcerer.io/fame/clmnt/huggingface/transformers/images/7)](https://sourcerer.io/fame/clmnt/huggingface/transformers/links/7)
 
 ### Features
@@ -59,7 +60,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
-| Documentation [(master)](https://huggingface.co/transformers/master) [(stable)](https://huggingface.co/transformers/) [(v2.10.0)](https://huggingface.co/transformers/v2.10.0) [(v2.9.0/v2.9.1)](https://huggingface.co/transformers/v2.9.1) [(v2.8.0)](https://huggingface.co/transformers/v2.8.0) [(v2.7.0)](https://huggingface.co/transformers/v2.7.0) [(v2.6.0)](https://huggingface.co/transformers/v2.6.0) [(v2.5.0/v2.5.1)](https://huggingface.co/transformers/v2.5.1) [(v2.4.0/v2.4.1)](https://huggingface.co/transformers/v2.4.0)[(v2.3.0)](https://huggingface.co/transformers/v2.3.0)[(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) | Full API documentation and more |
+| [Documentation](https://huggingface.co/transformers/) | Full API documentation and more |
 
 ## Installation
 
@@ -287,8 +288,8 @@ pytorch_model = BertForSequenceClassification.from_pretrained('./save/', from_tf
 sentence_0 = "This research was consistent with his findings."
 sentence_1 = "His findings were compatible with this research."
 sentence_2 = "His findings were not compatible with this research."
-inputs_1 = tokenizer.encode_plus(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
-inputs_2 = tokenizer.encode_plus(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
+inputs_1 = tokenizer(sentence_0, sentence_1, add_special_tokens=True, return_tensors='pt')
+inputs_2 = tokenizer(sentence_0, sentence_2, add_special_tokens=True, return_tensors='pt')
 
 pred_1 = pytorch_model(inputs_1['input_ids'], token_type_ids=inputs_1['token_type_ids'])[0].argmax().item()
 pred_2 = pytorch_model(inputs_2['input_ids'], token_type_ids=inputs_2['token_type_ids'])[0].argmax().item()

docs/README.md

@@ -167,7 +167,7 @@ Here's an example showcasing everything so far:
 
             Indices can be obtained using :class:`transformers.AlbertTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
-            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
+            :func:`transformers.PreTrainedTokenizer.__call__` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
 ```

docs/source/_static/css/code-snippets.css

@@ -9,4 +9,8 @@
 
 .highlight .kn, .highlight .nv, .highlight .s2, .highlight .ow {
     color: #6670FF;
+}
+
+.highlight .gp {
+    color: #FB8D68;
 }

docs/source/_static/css/huggingface.css

@@ -1,9 +1,50 @@
 /* Our DOM objects */
 
+/* Version control */
+
+.version-button {
+    background-color: #6670FF;
+    color: white;
+    border: none;
+    padding: 5px;
+    font-size: 15px;
+    cursor: pointer;
+}
+
+.version-button:hover, .version-button:focus {
+    background-color: #A6B0FF;
+}
+ 
+.version-dropdown {
+    display: none;
+    background-color: #6670FF;
+    min-width: 160px;
+    overflow: auto;
+    font-size: 15px;
+}
+  
+.version-dropdown a {
+    color: white;
+    padding: 3px 4px;
+    text-decoration: none;
+    display: block;
+}
+  
+.version-dropdown a:hover {
+    background-color: #A6B0FF;
+}
+  
+.version-show {
+    display: block;
+}
+
+/* Framework selector */
+
 .framework-selector {
     display: flex;
     flex-direction: row;
     justify-content: flex-end;
+    margin-right: 30px;
 }
 
 .framework-selector > button {
@@ -20,6 +61,12 @@
     padding: 5px;
 }
 
+/* Copy button */
+
+a.copybtn {
+    margin: 3px;
+}
+
 /* The literal code blocks */
 .rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
     color: #6670FF;
@@ -38,6 +85,7 @@
 
 /* The research field on top of the toc tree */
 .wy-side-nav-search{
+    padding-top: 0;
     background-color: #6670FF;
 }
 

docs/source/_static/js/custom.js

@@ -1,3 +1,27 @@
+// These two things need to be updated at each release for the version selector.
+// Last stable version
+const stableVersion = "v3.0.2"
+// Dictionary doc folder to label
+const versionMapping = {
+    "master": "master",
+    "": "v3.0.0/v3.0.1/v3.0.2 (stable)",
+    "v2.11.0": "v2.11.0",
+    "v2.10.0": "v2.10.0",
+    "v2.9.1": "v2.9.0/v2.9.1",
+    "v2.8.0": "v2.8.0",
+    "v2.7.0": "v2.7.0",
+    "v2.6.0": "v2.6.0",
+    "v2.5.1": "v2.5.0/v2.5.1",
+    "v2.4.0": "v2.4.0/v2.4.1",
+    "v2.3.0": "v2.3.0",
+    "v2.2.0": "v2.2.0/v2.2.1/v2.2.2",
+    "v2.1.1": "v2.1.1",
+    "v2.0.0": "v2.0.0",
+    "v1.2.0": "v1.2.0",
+    "v1.1.0": "v1.1.0",
+    "v1.0.0": "v1.0.0"
+}
+
 function addIcon() {
     const huggingFaceLogo = "https://huggingface.co/landing/assets/transformers-docs/huggingface_logo.svg";
     const image = document.createElement("img");
@@ -58,6 +82,68 @@ function addGithubButton() {
     document.querySelector(".wy-side-nav-search .icon-home").insertAdjacentHTML('afterend', div);
 }
 
+function addVersionControl() {
+    // To grab the version currently in view, we parse the url
+    const parts = location.toString().split('/');
+    let versionIndex = parts.length - 2;
+    // Index page may not have a last part with filename.html so we need to go up
+    if (parts[parts.length - 1] != "" && ! parts[parts.length - 1].match(/\.html$|^search.html?/)) {
+        versionIndex = parts.length - 1;
+    }
+    // Main classes and models are nested so we need to go deeper
+    else if (parts[versionIndex] == "main_classes" || parts[versionIndex] == "model_doc") {
+        versionIndex = versionIndex - 1;
+    } 
+    const version = parts[versionIndex];
+
+    // Menu with all the links,
+    const versionMenu = document.createElement("div");
+
+    const htmlLines = [];
+    for (const [key, value] of Object.entries(versionMapping)) {
+        let baseUrlIndex = (version == "transformers") ? versionIndex + 1: versionIndex;
+        var urlParts = parts.slice(0, baseUrlIndex);
+        if (key != "") {
+            urlParts = urlParts.concat([key]);
+        }
+        urlParts = urlParts.concat(parts.slice(versionIndex+1));
+        htmlLines.push(`<a href="${urlParts.join('/')}">${value}</a>`);
+    }
+
+    versionMenu.classList.add("version-dropdown");
+    versionMenu.innerHTML = htmlLines.join('\n');
+    
+    // Button for version selection
+    const versionButton = document.createElement("div");
+    versionButton.classList.add("version-button");
+    let label = (version == "transformers") ? stableVersion : version
+    versionButton.innerText = label.concat(" ▼");
+
+    // Toggle the menu when we click on the button
+    versionButton.addEventListener("click", () => {
+        versionMenu.classList.toggle("version-show");
+    });
+
+    // Hide the menu when we click elsewhere
+    window.addEventListener("click", (event) => {
+        if (event.target != versionButton){
+            versionMenu.classList.remove('version-show');
+        }
+    });
+
+    // Container
+    const div = document.createElement("div");
+    div.appendChild(versionButton);
+    div.appendChild(versionMenu);
+    div.style.paddingTop = '25px';
+    div.style.backgroundColor = '#6670FF';
+    div.style.display = 'block';
+    div.style.textAlign = 'center';
+
+    const scrollDiv = document.querySelector(".wy-side-scroll");
+    scrollDiv.insertBefore(div, scrollDiv.children[1]);
+}
+
 function addHfMenu() {
     const div = `
     <div class="menu">
@@ -72,6 +158,8 @@ function platformToggle() {
     const codeBlocks = Array.from(document.getElementsByClassName("highlight"));
     const pytorchIdentifier = "## PYTORCH CODE";
     const tensorflowIdentifier = "## TENSORFLOW CODE";
+
+    const promptSpanIdentifier = `<span class="gp">&gt;&gt;&gt; </span>`
     const pytorchSpanIdentifier = `<span class="c1">${pytorchIdentifier}</span>`;
     const tensorflowSpanIdentifier = `<span class="c1">${tensorflowIdentifier}</span>`;
 
@@ -84,10 +172,22 @@ function platformToggle() {
         let tensorflowSpans;
 
         if(pytorchSpanPosition < tensorflowSpanPosition){
-            pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, tensorflowSpanPosition);
+            const isPrompt = spans.slice(
+                spans.indexOf(tensorflowSpanIdentifier) - promptSpanIdentifier.length,
+                spans.indexOf(tensorflowSpanIdentifier)
+            ) == promptSpanIdentifier;
+            const finalTensorflowSpanPosition = isPrompt ? tensorflowSpanPosition - promptSpanIdentifier.length : tensorflowSpanPosition;
+
+            pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, finalTensorflowSpanPosition);
             tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, spans.length);
         }else{
-            tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, pytorchSpanPosition);
+            const isPrompt = spans.slice(
+                spans.indexOf(pytorchSpanIdentifier) - promptSpanIdentifier.length,
+                spans.indexOf(pytorchSpanIdentifier)
+            ) == promptSpanIdentifier;
+            const finalPytorchSpanPosition = isPrompt ? pytorchSpanPosition - promptSpanIdentifier.length : pytorchSpanPosition;
+
+            tensorflowSpans = spans.slice(tensorflowSpanPosition + tensorflowSpanIdentifier.length + 1, finalPytorchSpanPosition);
             pytorchSpans = spans.slice(pytorchSpanPosition + pytorchSpanIdentifier.length + 1, spans.length);
         }
 
@@ -149,6 +249,7 @@ function parseGithubButtons (){"use strict";var e=window.document,t=e.location,o
 
 function onLoad() {
     addIcon();
+    addVersionControl();
     addCustomFooter();
     addGithubButton();
     parseGithubButtons();

docs/source/benchmarks.md

@@ -1,54 +0,0 @@
-# Benchmarks
-
-This section is dedicated to the Benchmarks done by the library, both by maintainers, contributors and users. These 
-benchmark will help keep track of the preformance improvements that are brought to our models across versions.
-
-## Benchmarking all models for inference
-
-As of version 2.1 we have benchmarked all models for inference, across many different settings: using PyTorch, with
-and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
-TensorFlow XLA) and GPUs.
-
-The approach is detailed in the [following blogpost](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
-
-The results are available [here](https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing).
-
-## TF2 with mixed precision, XLA, Distribution (@tlkh)
-
-This work was done by [Timothy Liu](https://github.com/tlkh).
-
-There are very positive results to be gained from the various TensorFlow 2.0 features:
-
-- Automatic Mixed Precision (AMP)
-- XLA compiler
-- Distribution strategies (multi-GPU)
-
-The benefits are listed here (tested on CoLA, MRPC, SST-2):
-
-- AMP: Between 1.4x to 1.6x decrease in overall time without change in batch size
-- AMP+XLA: Up to 2.5x decrease in overall time on SST-2 (larger dataset)
-- Distribution: Between 1.4x to 3.4x decrease in overall time on 4xV100
-- Combined: Up to 5.7x decrease in overall training time, or 9.1x training throughput
-
-The model quality (measured by the validation accuracy) fluctuates slightly. Taking an average of 4 training runs 
-on a single GPU gives the following results:
-
-- CoLA: AMP results in slighter lower acc (0.820 vs 0.824)
-- MRPC: AMP results in lower acc (0.823 vs 0.835)
-- SST-2: AMP results in slighter lower acc (0.918 vs 0.922)
-
-However, in a distributed setting with 4xV100 (4x batch size), AMP can yield in better results:
-
-CoLA: AMP results in higher acc (0.828 vs 0.812)
-MRPC: AMP results in lower acc (0.817 vs 0.827)
-SST-2: AMP results in slightly lower acc (0.926 vs 0.929)
-
-The benchmark script is available [here](https://github.com/NVAITC/benchmarking/blob/master/tf2/bert_dist.py).
-
-Note: on some tasks (e.g. MRPC), the dataset is too small. The overhead due to the model compilation with XLA as well
-as the distribution strategy setup does not speed things up. The XLA compile time is also the reason why although throughput 
-can increase a lot (e.g. 2.7x for single GPU), overall (end-to-end) training speed-up is not as fast (as low as 1.4x)
-
-The benefits as seen on SST-2 (larger dataset) is much clear.
-
-All results can be seen on this [Google Sheet](https://docs.google.com/spreadsheets/d/1538MN224EzjbRL239sqSiUy6YY-rAjHyXhTzz_Zptls/edit#gid=960868445).

docs/source/benchmarks.rst

@@ -0,0 +1,322 @@
+Benchmarks
+==========
+
+Let's take a look at how 🤗 Transformer models can be benchmarked, best practices, and already available benchmarks.
+
+A notebook explaining in more detail how to benchmark 🤗 Transformer models can be found `here <https://github.com/huggingface/transformers/blob/master/notebooks/05-benchmark.ipynb>`__.
+
+How to benchmark 🤗 Transformer models
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The classes :class:`~transformers.PyTorchBenchmark` and :class:`~transformers.TensorFlowBenchmark` allow to flexibly benchmark 🤗 Transformer models.
+The benchmark classes allow us to measure the `peak memory usage` and `required time` for both 
+`inference` and `training`. 
+
+.. note::
+
+  Hereby, `inference` is defined by a single forward pass, and `training` is defined by a single forward pass and backward pass.
+
+The benchmark classes :class:`~transformers.PyTorchBenchmark` and :class:`~transformers.TensorFlowBenchmark` expect an object of type :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments`, respectively, for instantiation. :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments` are data classes and contain all relevant configurations for their corresponding benchmark class.
+In the following example, it is shown how a BERT model of type `bert-base-cased` can be benchmarked.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import PyTorchBenchmark, PyTorchBenchmarkArguments
+
+    >>> args = PyTorchBenchmarkArguments(models=["bert-base-uncased"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
+    >>> benchmark = PyTorchBenchmark(args)
+
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import TensorFlowBenchmark, TensorFlowBenchmarkArguments
+
+    >>> args = TensorFlowBenchmarkArguments(models=["bert-base-uncased"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
+    >>> benchmark = TensorFlowBenchmark(args)
+
+
+Here, three arguments are given to the benchmark argument data classes, namely ``models``, ``batch_sizes``, and ``sequence_lengths``. The argument ``models`` is required and expects a :obj:`list` of model identifiers from the `model hub <https://huggingface.co/models>`__
+The :obj:`list` arguments ``batch_sizes`` and ``sequence_lengths`` define the size of the ``input_ids`` on which the model is benchmarked. 
+There are many more parameters that can be configured via the benchmark argument data classes. For more detail on these one can either directly consult the files 
+``src/transformers/benchmark/benchmark_args_utils.py``, ``src/transformers/benchmark/benchmark_args.py`` (for PyTorch) and ``src/transformers/benchmark/benchmark_args_tf.py`` (for Tensorflow). 
+Alternatively, running the following shell commands from root will print out a descriptive list of all configurable parameters for PyTorch and Tensorflow respectively.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    python examples/benchmarking/run_benchmark.py --help
+
+    >>> ## TENSORFLOW CODE
+    python examples/benchmarking/run_benchmark_tf.py --help
+
+
+An instantiated benchmark object can then simply be run by calling ``benchmark.run()``.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> results = benchmark.run()
+    >>> print(results)
+    ====================       INFERENCE - SPEED - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length     Time in s                  
+    --------------------------------------------------------------------------------
+    bert-base-uncased          8               8             0.006     
+    bert-base-uncased          8               32            0.006     
+    bert-base-uncased          8              128            0.018     
+    bert-base-uncased          8              512            0.088     
+    --------------------------------------------------------------------------------
+    
+    ====================      INFERENCE - MEMORY - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length    Memory in MB 
+    --------------------------------------------------------------------------------
+    bert-base-uncased          8               8             1227
+    bert-base-uncased          8               32            1281
+    bert-base-uncased          8              128            1307
+    bert-base-uncased          8              512            1539
+    --------------------------------------------------------------------------------
+    
+    ====================        ENVIRONMENT INFORMATION         ====================
+    - transformers_version: 2.11.0
+    - framework: PyTorch
+    - use_torchscript: False
+    - framework_version: 1.4.0
+    - python_version: 3.6.10
+    - system: Linux
+    - cpu: x86_64
+    - architecture: 64bit
+    - date: 2020-06-29
+    - time: 08:58:43.371351
+    - fp16: False
+    - use_multiprocessing: True
+    - only_pretrain_model: False
+    - cpu_ram_mb: 32088
+    - use_gpu: True
+    - num_gpus: 1
+    - gpu: TITAN RTX
+    - gpu_ram_mb: 24217
+    - gpu_power_watts: 280.0
+    - gpu_performance_state: 2
+    - use_tpu: False
+    
+    >>> ## TENSORFLOW CODE
+    >>> results = benchmark.run()
+    >>> print(results)
+    ====================       INFERENCE - SPEED - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length     Time in s                  
+    --------------------------------------------------------------------------------
+    bert-base-uncased          8               8             0.005
+    bert-base-uncased          8               32            0.008
+    bert-base-uncased          8              128            0.022
+    bert-base-uncased          8              512            0.105
+    --------------------------------------------------------------------------------
+    
+    ====================      INFERENCE - MEMORY - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length    Memory in MB 
+    --------------------------------------------------------------------------------
+    bert-base-uncased          8               8             1330
+    bert-base-uncased          8               32            1330
+    bert-base-uncased          8              128            1330
+    bert-base-uncased          8              512            1770
+    --------------------------------------------------------------------------------
+    
+    ====================        ENVIRONMENT INFORMATION         ====================
+    - transformers_version: 2.11.0
+    - framework: Tensorflow
+    - use_xla: False
+    - framework_version: 2.2.0
+    - python_version: 3.6.10
+    - system: Linux
+    - cpu: x86_64
+    - architecture: 64bit
+    - date: 2020-06-29
+    - time: 09:26:35.617317
+    - fp16: False
+    - use_multiprocessing: True
+    - only_pretrain_model: False
+    - cpu_ram_mb: 32088
+    - use_gpu: True
+    - num_gpus: 1
+    - gpu: TITAN RTX
+    - gpu_ram_mb: 24217
+    - gpu_power_watts: 280.0
+    - gpu_performance_state: 2
+    - use_tpu: False
+
+By default, the `time` and the `required memory` for `inference` are benchmarked. 
+In the example output above the first two sections show the result corresponding to `inference time` and `inference memory`. 
+In addition, all relevant information about the computing environment, `e.g.` the GPU type, the system, the library versions, etc... are printed out in the third section under `ENVIRONMENT INFORMATION`.
+This information can optionally be saved in a `.csv` file when adding the argument :obj:`save_to_csv=True` to :class:`~transformers.PyTorchBenchmarkArguments` and :class:`~transformers.TensorFlowBenchmarkArguments` respectively.
+In this case, every section is saved in a separate `.csv` file. The path to each `.csv` file can optionally be defined via the argument data classes.
+
+Instead of benchmarking pre-trained models via their model identifier, `e.g.` `bert-base-uncased`, the user can alternatively benchmark an arbitrary configuration of any available model class. 
+In this case, a :obj:`list` of configurations must be inserted with the benchmark args as follows.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import PyTorchBenchmark, PyTorchBenchmarkArguments, BertConfig
+
+    >>> args = PyTorchBenchmarkArguments(models=["bert-base", "bert-384-hid", "bert-6-lay"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
+    >>> config_base = BertConfig()
+    >>> config_384_hid = BertConfig(hidden_size=384)
+    >>> config_6_lay = BertConfig(num_hidden_layers=6)
+
+    >>> benchmark = PyTorchBenchmark(args, configs=[config_base, config_384_hid, config_6_lay])
+    >>> benchmark.run()
+    ====================       INFERENCE - SPEED - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length       Time in s                  
+    --------------------------------------------------------------------------------
+    bert-base                  8              128            0.006
+    bert-base                  8              512            0.006
+    bert-base                  8              128            0.018     
+    bert-base                  8              512            0.088     
+    bert-384-hid              8               8             0.006     
+    bert-384-hid              8               32            0.006     
+    bert-384-hid              8              128            0.011     
+    bert-384-hid              8              512            0.054     
+    bert-6-lay                 8               8             0.003     
+    bert-6-lay                 8               32            0.004     
+    bert-6-lay                 8              128            0.009     
+    bert-6-lay                 8              512            0.044
+    --------------------------------------------------------------------------------
+    
+    ====================      INFERENCE - MEMORY - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length      Memory in MB 
+    --------------------------------------------------------------------------------
+    bert-base                  8               8             1277
+    bert-base                  8               32            1281
+    bert-base                  8              128            1307     
+    bert-base                  8              512            1539     
+    bert-384-hid              8               8             1005     
+    bert-384-hid              8               32            1027     
+    bert-384-hid              8              128            1035     
+    bert-384-hid              8              512            1255     
+    bert-6-lay                 8               8             1097     
+    bert-6-lay                 8               32            1101     
+    bert-6-lay                 8              128            1127     
+    bert-6-lay                 8              512            1359
+    --------------------------------------------------------------------------------
+    
+    ====================        ENVIRONMENT INFORMATION         ====================
+    - transformers_version: 2.11.0
+    - framework: PyTorch
+    - use_torchscript: False
+    - framework_version: 1.4.0
+    - python_version: 3.6.10
+    - system: Linux
+    - cpu: x86_64
+    - architecture: 64bit
+    - date: 2020-06-29
+    - time: 09:35:25.143267
+    - fp16: False
+    - use_multiprocessing: True
+    - only_pretrain_model: False
+    - cpu_ram_mb: 32088
+    - use_gpu: True
+    - num_gpus: 1
+    - gpu: TITAN RTX
+    - gpu_ram_mb: 24217
+    - gpu_power_watts: 280.0
+    - gpu_performance_state: 2
+    - use_tpu: False
+
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import TensorFlowBenchmark, TensorFlowBenchmarkArguments, BertConfig
+
+    >>> args = TensorFlowBenchmarkArguments(models=["bert-base", "bert-384-hid", "bert-6-lay"], batch_sizes=[8], sequence_lengths=[8, 32, 128, 512])
+    >>> config_base = BertConfig()
+    >>> config_384_hid = BertConfig(hidden_size=384)
+    >>> config_6_lay = BertConfig(num_hidden_layers=6)
+
+    >>> benchmark = TensorFlowBenchmark(args, configs=[config_base, config_384_hid, config_6_lay])
+    >>> benchmark.run()
+    ====================       INFERENCE - SPEED - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length       Time in s                  
+    --------------------------------------------------------------------------------
+    bert-base                  8               8             0.005
+    bert-base                  8               32            0.008
+    bert-base                  8              128            0.022
+    bert-base                  8              512            0.106
+    bert-384-hid              8               8             0.005
+    bert-384-hid              8               32            0.007
+    bert-384-hid              8              128            0.018
+    bert-384-hid              8              512            0.064
+    bert-6-lay                 8               8             0.002
+    bert-6-lay                 8               32            0.003
+    bert-6-lay                 8              128            0.0011
+    bert-6-lay                 8              512            0.074
+    --------------------------------------------------------------------------------
+    
+    ====================      INFERENCE - MEMORY - RESULT       ====================
+    --------------------------------------------------------------------------------
+    Model Name             Batch Size     Seq Length      Memory in MB 
+    --------------------------------------------------------------------------------
+    bert-base                  8               8             1330
+    bert-base                  8               32            1330
+    bert-base                  8              128            1330
+    bert-base                  8              512            1770
+    bert-384-hid              8               8             1330
+    bert-384-hid              8               32            1330
+    bert-384-hid              8              128            1330
+    bert-384-hid              8              512            1540
+    bert-6-lay                 8               8             1330
+    bert-6-lay                 8               32            1330
+    bert-6-lay                 8              128            1330
+    bert-6-lay                 8              512            1540
+    --------------------------------------------------------------------------------
+    
+    ====================        ENVIRONMENT INFORMATION         ====================
+    - transformers_version: 2.11.0
+    - framework: Tensorflow
+    - use_xla: False
+    - framework_version: 2.2.0
+    - python_version: 3.6.10
+    - system: Linux
+    - cpu: x86_64
+    - architecture: 64bit
+    - date: 2020-06-29
+    - time: 09:38:15.487125
+    - fp16: False
+    - use_multiprocessing: True
+    - only_pretrain_model: False
+    - cpu_ram_mb: 32088
+    - use_gpu: True
+    - num_gpus: 1
+    - gpu: TITAN RTX
+    - gpu_ram_mb: 24217
+    - gpu_power_watts: 280.0
+    - gpu_performance_state: 2
+    - use_tpu: False
+
+
+Again, `inference time` and `required memory` for `inference` are measured, but this time for customized configurations of the :obj:`BertModel` class. This feature can especially be helpful when 
+deciding for which configuration the model should be trained.
+
+
+Benchmark best practices
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+This section lists a couple of best practices one should be aware of when benchmarking a model.
+
+- Currently, only single device benchmarking is supported. When benchmarking on GPU, it is recommended that the user 
+  specifies on which device the code should be run by setting the ``CUDA_VISIBLE_DEVICES`` environment variable in the shell, `e.g.` ``export CUDA_VISIBLE_DEVICES=0`` before running the code.
+- The option :obj:`no_multi_processing` should only be set to :obj:`True` for testing and debugging. To ensure accurate memory measurement it is recommended to run each memory benchmark in a separate process by making sure :obj:`no_multi_processing` is set to :obj:`True`.
+- One should always state the environment information when sharing the results of a model benchmark. Results can vary heavily between different GPU devices, library versions, etc., so that benchmark results on their own are not very useful for the community.
+
+
+Sharing your benchmark
+~~~~~~~~~~~~~~~~~~~~~~
+
+Previously all available core models (10 at the time) have been benchmarked for `inference time`, across many different settings: using PyTorch, with
+and without TorchScript, using TensorFlow, with and without XLA. All of those tests were done across CPUs (except for
+TensorFlow XLA) and GPUs.
+
+The approach is detailed in the `following blogpost <https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2>`__ and the results are available `here <https://docs.google.com/spreadsheets/d/1sryqufw2D0XlUH4sq3e9Wnxu5EAQkaohzrJbd5HdQ_w/edit?usp=sharing>`__.
+
+With the new `benchmark` tools, it is easier than ever to share your benchmark results with the community `here <https://github.com/huggingface/transformers/blob/master/examples/benchmarking/README.md>`__.

docs/source/conf.py

@@ -26,7 +26,7 @@ author = u'huggingface'
 # The short X.Y version
 version = u''
 # The full version, including alpha/beta/rc tags
-release = u'2.11.0'
+release = u'3.0.2'
 
 
 # -- General configuration ---------------------------------------------------
@@ -44,7 +44,8 @@ extensions = [
     'sphinx.ext.napoleon',
     'recommonmark',
     'sphinx.ext.viewcode',
-    'sphinx_markdown_tables'
+    'sphinx_markdown_tables',
+    'sphinx_copybutton'
 ]
 
 # Add any paths that contain templates here, relative to this directory.
@@ -74,6 +75,8 @@ exclude_patterns = [u'_build', 'Thumbs.db', '.DS_Store']
 # The name of the Pygments (syntax highlighting) style to use.
 pygments_style = None
 
+# Remove the prompt when copying examples
+copybutton_prompt_text = ">>> "
 
 # -- Options for HTML output -------------------------------------------------
 
@@ -187,8 +190,8 @@ epub_title = project
 epub_exclude_files = ['search.html']
 
 def setup(app):
-    app.add_stylesheet('css/huggingface.css')
-    app.add_stylesheet('css/code-snippets.css')
+    app.add_css_file('css/huggingface.css')
+    app.add_css_file('css/code-snippets.css')
     app.add_js_file('js/custom.js')
 
 # -- Extension configuration -------------------------------------------------

docs/source/glossary.rst

@@ -11,7 +11,7 @@ General terms
   tokens at a certain timestep.
 - MLM: masked language modeling, a pretraining task where the model sees a corrupted version of the texts, usually done
   by masking some tokens randomly, and has to predict the original text.
-- multimodal: a task taht combines texts with another kind of inputs (for instance images).
+- multimodal: a task that combines texts with another kind of inputs (for instance images).
 - NLG: natural language generation, all tasks related to generating text ( for instance talk with transformers,
   translation)
 - NLP: natural language processing, a generic way to say "deal with texts".
@@ -45,17 +45,16 @@ tokenizer, which is a `WordPiece <https://arxiv.org/pdf/1609.08144.pdf>`__ token
 
 ::
 
-    from transformers import BertTokenizer
-    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    >>> from transformers import BertTokenizer
+    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
 
-    sequence = "A Titan RTX has 24GB of VRAM"
+    >>> sequence = "A Titan RTX has 24GB of VRAM"
 
 The tokenizer takes care of splitting the sequence into tokens available in the tokenizer vocabulary.
 
 ::
 
-    tokenized_sequence = tokenizer.tokenize(sequence)
-    print(tokenized_sequence)
+    >>> tokenized_sequence = tokenizer.tokenize(sequence)
 
 The tokens are either words or subwords. Here for instance, "VRAM" wasn't in the model vocabulary, so it's been split
 in "V", "RA" and "M". To indicate those tokens are not separate words but parts of the same word, a double-dash is
@@ -63,6 +62,7 @@ added for "RA" and "M":
 
 ::
 
+    >>> print(tokenized_sequence)
     ['A', 'Titan', 'R', '##T', '##X', 'has', '24', '##GB', 'of', 'V', '##RA', '##M']
 
 These tokens can then be converted into IDs which are understandable by the model. This can be done by directly feeding
@@ -71,14 +71,14 @@ the sentence to the tokenizer, which leverages the Rust implementation of
 
 ::
 
-    encoded_sequence = tokenizer(sequence)["input_ids"]
-    print(encoded_sequence)
+    >>> encoded_sequence = tokenizer(sequence)["input_ids"]
 
 The tokenizer returns a dictionary with all the arguments necessary for its corresponding model to work properly. The
 token indices are under the key "input_ids":
 
 ::
 
+    >>> print(encoded_sequence)
     [101, 138, 18696, 155, 1942, 3190, 1144, 1572, 13745, 1104, 159, 9664, 2107, 102]
 
 Note that the tokenizer automatically adds "special tokens" (if the associated model rely on them) which are special
@@ -86,13 +86,14 @@ IDs the model sometimes uses. If we decode the previous sequence of ids,
 
 ::
 
-    tokenizer.decode(encoded_sequence)
+    >>> decoded_sequence = tokenizer.decode(encoded_sequence)
 
 we will see 
 
 ::
 
-    '[CLS] A Titan RTX has 24GB of VRAM [SEP]'
+    >>> print(decoded_sequence)
+    [CLS] A Titan RTX has 24GB of VRAM [SEP]
 
 because this is the way a :class:`~transformers.BertModel` is going to expect its inputs.
 
@@ -108,21 +109,20 @@ For example, consider these two sequences:
 
 ::
 
-    from transformers import BertTokenizer
-    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    >>> from transformers import BertTokenizer
+    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
 
-    sequence_a = "This is a short sequence."
-    sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
+    >>> sequence_a = "This is a short sequence."
+    >>> sequence_b = "This is a rather long sequence. It is at least longer than the sequence A."
 
-    encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
-    encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
-    
-    len(encoded_sequence_a), len(encoded_sequence_b)
+    >>> encoded_sequence_a = tokenizer(sequence_a)["input_ids"]
+    >>> encoded_sequence_b = tokenizer(sequence_b)["input_ids"]
 
 The encoded versions have different lengths:
 
 ::
 
+    >>> len(encoded_sequence_a), len(encoded_sequence_b)
     (8, 19)
 
 Therefore, we can't be put then together in a same tensor as-is. The first sequence needs to be padded up to the length
@@ -133,15 +133,14 @@ it to pad like this:
 
 ::
 
-    padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
-    padded_sequences["input_ids"]
+    >>> padded_sequences = tokenizer([sequence_a, sequence_b], padding=True)
 
 We can see that 0s have been added on the right of the first sentence to make it the same length as the second one:
 
 ::
 
-    [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
-     [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
+    >>> padded_sequences["input_ids"]
+    [[101, 1188, 1110, 170, 1603, 4954, 119, 102, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [101, 1188, 1110, 170, 1897, 1263, 4954, 119, 1135, 1110, 1120, 1655, 2039, 1190, 1103, 4954, 138, 119, 102]]
 
 This can then be converted into a tensor in PyTorch or TensorFlow. The attention mask is a binary tensor indicating
 the position of the padded indices so that the model does not attend to them. For the
@@ -150,14 +149,8 @@ a padded value. This attention mask is in the dictionary returned by the tokeniz
 
 ::
 
-    padded_sequences["attention_mask"]
-
-will give back
-
-::
-
-    [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
-     [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
+    >>> padded_sequences["attention_mask"]
+    [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]
 
 .. _token-type-ids:
 
@@ -170,26 +163,27 @@ tokens. For example, the BERT model builds its two sequence input as such:
 
 ::
 
-   # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
+   >>> # [CLS] SEQUENCE_A [SEP] SEQUENCE_B [SEP]
 
 We can use our tokenizer to automatically generate such a sentence by passing the two sequences as two arguments (and
 not a list like before) like this:
 
 ::
 
-    from transformers import BertTokenizer
-    tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
-    sequence_a = "HuggingFace is based in NYC"
-    sequence_b = "Where is HuggingFace based?"
+    >>> from transformers import BertTokenizer
+    >>> tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
+    >>> sequence_a = "HuggingFace is based in NYC"
+    >>> sequence_b = "Where is HuggingFace based?"
 
-    encoded_dict = tokenizer(sequence_a, sequence_b)
-    tokenizer.decode(encoded_dict["input_ids"])
+    >>> encoded_dict = tokenizer(sequence_a, sequence_b)
+    >>> decoded = tokenizer.decode(encoded_dict["input_ids"])
 
 which will return:
 
 ::
 
-    "[CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]"
+    >>> print(decoded)
+    [CLS] HuggingFace is based in NYC [SEP] Where is HuggingFace based? [SEP]
 
 This is enough for some models to understand where one sequence ends and where another begins. However, other models
 such as BERT have an additional mechanism, which are the token type IDs (also called segment IDs). They are a binary
@@ -199,12 +193,7 @@ The tokenizer returns in the dictionary under the key "token_type_ids":
 
 ::
 
-    encoded_dict['token_type_ids']
-
-will return
-
-::
-
+    >>> encoded_dict['token_type_ids']
     [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]
 
 The first sequence, the "context" used for the question, has all its tokens represented by :obj:`0`, whereas the

docs/source/index.rst

@@ -42,7 +42,7 @@ The documentation is organized in five parts:
 
 - **GET STARTED** contains a quick tour, the installation instructions and some useful information about our philosophy
   and a glossary.
-- **USING TRANSFORMERS** contains general tutorials on how to use the library.
+- **USING 🤗 TRANSFORMERS** contains general tutorials on how to use the library.
 - **ADVANCED GUIDES** contains more advanced guides that are more specific to a given script or part of the library.
 - **RESEARCH** focuses on tutorials that have less to do with how to use the library but more about general resarch in
   transformers model
@@ -121,7 +121,10 @@ conversion utilities for the following models:
     trained using `OPUS <http://opus.nlpl.eu/>`_ pretrained_models data by Jörg Tiedemann.
 21. `Longformer <https://github.com/allenai/longformer>`_ (from AllenAI) released with the paper `Longformer: The
     Long-Document Transformer <https://arxiv.org/abs/2004.05150>`_ by Iz Beltagy, Matthew E. Peters, and Arman Cohan.
-22. `Other community models <https://huggingface.co/models>`_, contributed by the `community
+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
     <https://huggingface.co/users>`_.
 
 .. toctree::
@@ -135,12 +138,14 @@ conversion utilities for the following models:
 
 .. toctree::
     :maxdepth: 2
-    :caption: Using Transformers
+    :caption: Using 🤗 Transformers
 
     task_summary
     model_summary
-    serialization
+    preprocessing
+    training
     model_sharing
+    tokenizer_summary
     multilingual
 
 .. toctree::
@@ -172,6 +177,7 @@ conversion utilities for the following models:
     main_classes/pipelines
     main_classes/optimizer_schedules
     main_classes/processors
+    main_classes/trainer
     model_doc/auto
     model_doc/encoderdecoder
     model_doc/bert
@@ -196,3 +202,4 @@ conversion utilities for the following models:
     model_doc/longformer
     model_doc/retribert
     model_doc/mobilebert
+    model_doc/dpr

docs/source/main_classes/optimizer_schedules.rst

@@ -1,4 +1,4 @@
-Optimizer
+Optimization
 ----------------------------------------------------
 
 The ``.optimization`` module provides:
@@ -7,24 +7,25 @@ The ``.optimization`` module provides:
 - several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
 - a gradient accumulation class to accumulate the gradients of multiple batches
 
-``AdamW``
-~~~~~~~~~~~~~~~~
+``AdamW`` (PyTorch)
+~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.AdamW
     :members:
 
-``AdamWeightDecay``
-~~~~~~~~~~~~~~~~~~~
+``AdamWeightDecay`` (TensorFlow)
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.AdamWeightDecay
 
 .. autofunction:: transformers.create_optimizer
 
 Schedules
-----------------------------------------------------
+~~~~~~~~~~~~~~~~~~~
+
+Learning Rate Schedules (Pytorch)
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
-Learning Rate Schedules
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 .. autofunction:: transformers.get_constant_schedule
 
 
@@ -56,16 +57,16 @@ Learning Rate Schedules
     :target: /imgs/warmup_linear_schedule.png
     :alt:
 
-``Warmup``
-~~~~~~~~~~~~~~~~
+``Warmup`` (TensorFlow)
+^^^^^^^^^^^^^^^^^^^^^^^
 
 .. autoclass:: transformers.WarmUp
     :members:
 
 Gradient Strategies
-----------------------------------------------------
+~~~~~~~~~~~~~~~~~~~~
 
-``GradientAccumulator``
-~~~~~~~~~~~~~~~~~~~~~~~
+``GradientAccumulator`` (TensorFlow)
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
 .. autoclass:: transformers.GradientAccumulator

docs/source/main_classes/tokenizer.rst

@@ -11,7 +11,7 @@ The base classes ``PreTrainedTokenizer`` and ``PreTrainedTokenizerFast`` impleme
 - adding new tokens to the vocabulary in a way that is independant of the underlying structure (BPE, SentencePiece...),
 - managing special tokens like mask, beginning-of-sentence, etc tokens (adding them, assigning them to attributes in the tokenizer for easy access and making sure they are not split during tokenization)
 
-``BatchEncoding`` holds the output of the tokenizer's encoding methods (``encode_plus`` and ``batch_encode_plus``) and is derived from a Python dictionary. When the tokenizer is a pure python tokenizer, this class behave just like a standard python dictionary and hold the various model inputs computed by these methodes (``input_ids``, ``attention_mask``...). When the tokenizer is a "Fast" tokenizer (i.e. backed by HuggingFace tokenizers library), this class provides in addition several advanced alignement methods which can be used to map between the original string (character and words) and the token space (e.g. getting the index of the token comprising a given character or the span of characters corresponding to a given token).
+``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``
 ~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/main_classes/trainer.rst

@@ -0,0 +1,45 @@
+Trainer
+----------
+
+The :class:`~transformers.Trainer` and :class:`~transformers.TFTrainer` classes provide an API for feature-complete
+training in most standard use cases. It's used in most of the :doc:`example scripts <../examples>`.
+
+Before instantiating your :class:`~transformers.Trainer`/:class:`~transformers.TFTrainer`, create a 
+:class:`~transformers.TrainingArguments`/:class:`~transformers.TFTrainingArguments` to access all the points of
+customization during training.
+
+The API supports distributed training on multiple GPUs/TPUs, mixed precision through `NVIDIA Apex
+<https://github.com/NVIDIA/apex>`__ for PyTorch and :obj:`tf.keras.mixed_precision` for TensorFlow.
+
+``Trainer`` 
+~~~~~~~~~~~
+
+.. autoclass:: transformers.Trainer
+    :members:
+
+``TFTrainer`` 
+~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFTrainer
+    :members:
+
+``TrainingArguments``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TrainingArguments
+    :members:
+
+``TFTrainingArguments``
+~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFTrainingArguments
+    :members:
+
+Utilities
+~~~~~~~~~
+
+.. autoclass:: transformers.EvalPrediction
+
+.. autofunction:: transformers.set_seed
+
+.. autofunction:: transformers.torch_distributed_zero_first

docs/source/migration.md

@@ -1,8 +1,8 @@
 # Migrating from previous packages
 
-## Migrating from pytorch-transformers to transformers
+## Migrating from pytorch-transformers to 🤗 Transformers
 
-Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.
+Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to 🤗 Transformers.
 
 ### Positional order of some models' keywords inputs (`attention_mask`, `token_type_ids`...) changed
 
@@ -14,17 +14,17 @@ If you used to call the models with positional inputs for keyword arguments, e.g
 
 ## Migrating from pytorch-pretrained-bert
 
-Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `transformers`
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to 🤗 Transformers
 
 ### Models always output `tuples`
 
-The main breaking change when migrating from `pytorch-pretrained-bert` to `transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+The main breaking change when migrating from `pytorch-pretrained-bert` to 🤗 Transformers is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
 
 The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/transformers/).
 
 In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
 
-Here is a `pytorch-pretrained-bert` to `transformers` conversion example for a `BertForSequenceClassification` classification model:
+Here is a `pytorch-pretrained-bert` to 🤗 Transformers conversion example for a `BertForSequenceClassification` classification model:
 
 ```python
 # Let's load our model
@@ -33,11 +33,11 @@ model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
 # If you used to have this line in pytorch-pretrained-bert:
 loss = model(input_ids, labels=labels)
 
-# Now just use this line in transformers to extract the loss from the output tuple:
+# Now just use this line in 🤗 Transformers to extract the loss from the output tuple:
 outputs = model(input_ids, labels=labels)
 loss = outputs[0]
 
-# In transformers you can also have access to the logits:
+# In 🤗 Transformers you can also have access to the logits:
 loss, logits = outputs[:2]
 
 # And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
@@ -109,7 +109,7 @@ for batch in train_data:
     loss.backward()
     optimizer.step()
 
-### In Transformers, optimizer and schedules are splitted and instantiated like this:
+### In 🤗 Transformers, optimizer and schedules are splitted and instantiated like this:
 optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
 scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps)  # PyTorch scheduler
 ### and used like this:

docs/source/model_doc/bart.rst

@@ -39,6 +39,18 @@ BartTokenizer
     :members:
 
 
+MBartTokenizer
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MBartTokenizer
+    :members: build_inputs_with_special_tokens, prepare_translation_batch
+
+BartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.BartForConditionalGeneration
+    :members: generate, forward
+
 BartModel
 ~~~~~~~~~~~~~
 
@@ -62,10 +74,3 @@ BartForQuestionAnswering
     :members: forward
 
 
-BartForConditionalGeneration
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.BartForConditionalGeneration
-    :members: generate, forward
-
-

docs/source/model_doc/dpr.rst

@@ -0,0 +1,89 @@
+DPR
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+Dense Passage Retrieval (DPR) - is a set of tools and models for state-of-the-art open-domain Q&A research.
+It is based on the following paper:
+
+Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih, Dense Passage Retrieval for Open-Domain Question Answering.
+
+The abstract from the paper is the following:
+
+*Open-domain question answering relies on efficient passage retrieval to select candidate contexts, where traditional
+sparse vector space models, such as TF-IDF or BM25, are the de facto method. In this work, we show that retrieval can
+be practically implemented using dense representations alone, where embeddings are learned from a small number of
+questions and passages by a simple dual-encoder framework. When evaluated on a wide range of open-domain QA datasets,
+our dense retriever outperforms a strong Lucene-BM25 system largely by 9%-19% absolute in terms of top-20 passage
+retrieval accuracy, and helps our end-to-end QA system establish new state-of-the-art on multiple open-domain QA
+benchmarks.*
+
+The original code can be found `here <https://github.com/facebookresearch/DPR>`_.
+
+
+DPRConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRConfig
+    :members:
+
+
+DPRContextEncoderTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRContextEncoderTokenizer
+    :members:
+
+
+DPRContextEncoderTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRContextEncoderTokenizerFast
+    :members:
+
+DPRQuestionEncoderTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRQuestionEncoderTokenizer
+    :members:
+
+
+DPRQuestionEncoderTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRQuestionEncoderTokenizerFast
+    :members:
+
+DPRReaderTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRReaderTokenizer
+    :members:
+
+
+DPRReaderTokenizerFast
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRReaderTokenizerFast
+    :members:
+
+
+DPRContextEncoder
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRContextEncoder
+    :members:
+
+DPRQuestionEncoder
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRQuestionEncoder
+    :members:
+
+
+DPRReader
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DPRReader
+    :members:

docs/source/model_doc/reformer.rst

@@ -112,3 +112,17 @@ ReformerModelWithLMHead
 
 .. autoclass:: transformers.ReformerModelWithLMHead
     :members:
+
+
+ReformerForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ReformerForMaskedLM
+    :members:
+
+
+ReformerForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.ReformerForQuestionAnswering
+    :members:

docs/source/model_sharing.md

@@ -1,55 +0,0 @@
-# Model upload and sharing
-
-Starting with `v2.2.2`, you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
-
-**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Optionally, join an existing organization or create a new one. Then:
-
-```shell
-transformers-cli login
-# log in using the same credentials as on huggingface.co
-```
-Upload your model:
-```shell
-transformers-cli upload ./path/to/pretrained_model/
-
-# ^^ Upload folder containing weights/tokenizer/config
-# saved via `.save_pretrained()`
-
-transformers-cli upload ./config.json [--filename folder/foobar.json]
-
-# ^^ Upload a single file
-# (you can optionally override its filename, which can be nested inside a folder)
-```
-
-If you want your model to be namespaced by your organization name rather than your username, add the following flag to any command:
-```shell
---organization organization_name
-```
-
-Your model will then be accessible through its identifier, a concatenation of your username (or organization name) and the folder name above:
-```python
-"username/pretrained_model"
-# or if an org:
-"organization_name/pretrained_model"
-```
-
-**Please add a README.md model card** to the repo under `model_cards/` with: model description, training params (dataset, preprocessing, hardware used, hyperparameters), evaluation results, intended uses & limitations, etc.
-
-Your model now has a page on huggingface.co/models 🔥
-
-Anyone can load it from code:
-```python
-tokenizer = AutoTokenizer.from_pretrained("namespace/pretrained_model")
-model = AutoModel.from_pretrained("namespace/pretrained_model")
-```
-
-List all your files on S3:
-```shell
-transformers-cli s3 ls
-```
-
-You can also delete unneeded files:
-
-```shell
-transformers-cli s3 rm …
-```

docs/source/model_sharing.rst

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

docs/source/model_summary.rst

@@ -1,7 +1,7 @@
 Summary of the models
 ================================================
 
-This is a summary of the models available in the transformers library. It assumes you’re familiar with the original 
+This is a summary of the models available in 🤗 Transformers. It assumes you’re familiar with the original 
 `transformer model <https://arxiv.org/abs/1706.03762>`_. For a gentle introduction check the `annotated transformer 
 <http://nlp.seas.harvard.edu/2018/04/03/attention.html>`_. Here we focus on the high-level differences between the
 models. You can check them more in detail in their respective documentation. Also checkout the 
@@ -55,7 +55,7 @@ Original GPT
    <a href="https://huggingface.co/models?filter=openai-gpt">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-openai--gpt-blueviolet">
    </a>
-   <a href="/model_doc/gpt">
+   <a href="model_doc/gpt">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-openai--gpt-blueviolet">
    </a>
 
@@ -75,7 +75,7 @@ GPT-2
    <a href="https://huggingface.co/models?filter=gpt2">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-gpt2-blueviolet">
    </a>
-   <a href="/model_doc/gpt2">
+   <a href="model_doc/gpt2">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-gpt2-blueviolet">
    </a>
 
@@ -96,7 +96,7 @@ CTRL
    <a href="https://huggingface.co/models?filter=ctrl">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-ctrl-blueviolet">
    </a>
-   <a href="/model_doc/ctrl">
+   <a href="model_doc/ctrl">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-ctrl-blueviolet">
    </a>
 
@@ -117,7 +117,7 @@ Transformer-XL
    <a href="https://huggingface.co/models?filter=transfo-xl">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-transfo--xl-blueviolet">
    </a>
-   <a href="/model_doc/transformerxl">
+   <a href="model_doc/transformerxl">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-transfo--xl-blueviolet">
    </a>
 
@@ -148,7 +148,7 @@ Reformer
    <a href="https://huggingface.co/models?filter=reformer">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-reformer-blueviolet">
    </a>
-   <a href="/model_doc/reformer">
+   <a href="model_doc/reformer">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-reformer-blueviolet">
    </a>
 
@@ -183,7 +183,7 @@ XLNet
    <a href="https://huggingface.co/models?filter=xlnet">
        <img alt="Models" src="https://img.shields.io/badge/All_model_pages-xlnet-blueviolet">
    </a>
-   <a href="/model_doc/xlnet">
+   <a href="model_doc/xlnet">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-xlnet-blueviolet">
    </a>
 
@@ -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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <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">
        <img alt="Doc" src="https://img.shields.io/badge/Model_documentation-t5-blueviolet">
    </a>
 

docs/source/multilingual.rst

@@ -36,10 +36,11 @@ Here is an example using the ``xlm-clm-enfr-1024`` checkpoint (Causal language m
 
 .. code-block::
 
-    import torch
-    from transformers import XLMTokenizer, XLMWithLMHeadModel
+    >>> import torch
+    >>> from transformers import XLMTokenizer, XLMWithLMHeadModel
 
-    tokenizer = XLMTokenizer.from_pretrained("xlm-clm-1024-enfr")
+    >>> tokenizer = XLMTokenizer.from_pretrained("xlm-clm-enfr-1024")
+    >>> model = XLMWithLMHeadModel.from_pretrained("xlm-clm-enfr-1024")
 
 
 The different languages this model/tokenizer handles, as well as the ids of these languages are visible using the
@@ -47,16 +48,15 @@ The different languages this model/tokenizer handles, as well as the ids of thes
 
 .. code-block::
 
-    # Continuation of the previous script
-    print(tokenizer.lang2id)  # {'en': 0, 'fr': 1}
+    >>> print(tokenizer.lang2id)
+    {'en': 0, 'fr': 1}
 
 
 These ids should be used when passing a language parameter during a model pass. Let's define our inputs:
 
 .. code-block::
 
-    # Continuation of the previous script
-    input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1
+    >>> input_ids = torch.tensor([tokenizer.encode("Wikipedia was used to")]) # batch size of 1
 
 
 We should now define the language embedding by using the previously defined language id. We want to create a tensor
@@ -64,20 +64,18 @@ filled with the appropriate language ids, of the same size as input_ids. For eng
 
 .. code-block::
 
-    # Continuation of the previous script
-    language_id = tokenizer.lang2id['en']  # 0
-    langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])
+    >>> language_id = tokenizer.lang2id['en']  # 0
+    >>> langs = torch.tensor([language_id] * input_ids.shape[1])  # torch.tensor([0, 0, 0, ..., 0])
 
-    # We reshape it to be of size (batch_size, sequence_length)
-    langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)
+    >>> # We reshape it to be of size (batch_size, sequence_length)
+    >>> langs = langs.view(1, -1) # is now of shape [1, sequence_length] (we have a batch size of 1)
 
 
 You can then feed it all as input to your model:
 
 .. code-block::
 
-    # Continuation of the previous script
-    outputs = model(input_ids, langs=langs)
+    >>> outputs = model(input_ids, langs=langs)
 
 
 The example `run_generation.py <https://github.com/huggingface/transformers/blob/master/examples/text-generation/run_generation.py>`__

docs/source/philosophy.rst

@@ -1,7 +1,7 @@
 Philosophy
 ==========
 
-Transformers is an opinionated library built for:
+🤗 Transformers is an opinionated library built for:
 
 - NLP researchers and educators seeking to use/study/extend large-scale transformers models
 - hands-on practitioners who want to fine-tune those models and/or serve them in production

docs/source/preprocessing.rst

@@ -0,0 +1,373 @@
+Preprocessing data
+==================
+
+In this tutorial, we'll explore how to preprocess your data using 🤗 Transformers. The main tool for this is what we
+
+call a :doc:`tokenizer <main_classes/tokenizer>`. You can build one using the tokenizer class associated to the model
+you would like to use, or directly with the :class:`~transformers.AutoTokenizer` class.
+
+As we saw in the :doc:`quicktour </quicktour>`, the tokenizer will first split a given text in words (or part of words,
+punctuation symbols, etc.) usually called `tokens`. Then it will convert those `tokens` into numbers, to be able to
+build a tensor out of them and feed them to the model. It will also add any additional inputs the model might expect to
+work properly.
+
+.. note::
+
+    If you plan on using a pretrained model, it's important to use the associated pretrained tokenizer: it will split
+    the text you give it in tokens the same way for the pretraining corpus, and it will use the same correspondence
+    token to index (that we usually call a `vocab`) as during pretraining.
+
+To automatically download the vocab used during pretraining or fine-tuning a given model, you can use the 
+:func:`~transformers.AutoTokenizer.from_pretrained` method:
+
+::
+
+    from transformers import AutoTokenizer
+    tokenizer = AutoTokenizer.from_pretrained('bert-base-cased')
+
+Base use
+~~~~~~~~
+
+A :class:`~transformers.PreTrainedTokenizer` has many methods, but the only one you need to remember for preprocessing
+is its ``__call__``: you just need to feed your sentence to your tokenizer object.
+
+::
+
+    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:
+
+::
+
+    {'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]}
+
+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
+
+::
+
+    "[CLS] Hello, I'm a single sentence! [SEP]"
+
+As you can see, the tokenizer automatically added some special tokens that the model expect. Not all model need special
+tokens; for instance, if we had used` gtp2-medium` instead of `bert-base-cased` to create our tokenizer, we would have
+
+seen the same sentence as the original one here. You can disable this behavior (which is only advised if you have added
+those special tokens yourself) by passing ``add_special_tokens=False``.
+
+If you have several sentences you want to process, you can do this efficiently by sending them as a list to the
+tokenizer:
+
+::
+
+    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:
+
+::
+
+    {'input_ids': [[101, 8667, 146, 112, 182, 170, 1423, 5650, 102],
+                   [101, 1262, 1330, 5650, 102],
+                   [101, 1262, 1103, 1304, 1304, 1314, 1141, 102]],
+     'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0],
+                        [0, 0, 0, 0, 0, 0, 0, 0]],
+     'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1],
+                        [1, 1, 1, 1, 1, 1, 1, 1]]}
+
+If the purpose of sending several sentences at a time to the tokenizer is to build a batch to feed the model, you will
+probably want:
+
+- To pad each sentence to the maximum length there is in your batch.
+- To truncate each sentence to the maximum length the model can accept (if applicable).
+- To return tensors.
+
+You can do all of this by using the following options when feeding your list of sentences to the tokenizer:
+
+::
+
+    ## 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:
+
+::
+
+    {'input_ids': tensor([[ 101, 8667,  146,  112,  182,  170, 1423, 5650,  102],
+                          [ 101, 1262, 1330, 5650,  102,    0,    0,    0,    0],
+                          [ 101, 1262, 1103, 1304, 1304, 1314, 1141,  102,    0]]),
+     'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0],
+                               [0, 0, 0, 0, 0, 0, 0, 0, 0],
+                               [0, 0, 0, 0, 0, 0, 0, 0, 0]]), 
+     'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1],
+                               [1, 1, 1, 1, 1, 0, 0, 0, 0],
+                               [1, 1, 1, 1, 1, 1, 1, 1, 0]])}
+
+We can now see what the `attention_mask <glossary.html#attention-mask>`__ is all about: it points out which tokens the
+model should pay attention to and which ones it should not (because they represent padding in this case).
+
+
+Note that if your model does not have a maximum length associated to it, the command above will throw a warning. You
+can safely ignore it. You can also pass ``verbose=False`` to stop the tokenizer to throw those kinds of warnings.
+
+.. _sentence-pairs:
+
+Preprocessing pairs of sentences
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Sometimes you need to feed pair of sentences to your model. For instance, if you want to classify if two sentences in a
+pair are similar, or for question-answering models, which take a context and a question. For BERT models, the input is
+then represented like this:
+
+::
+
+    [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:
+
+::
+
+    {'input_ids': [101, 1731, 1385, 1132, 1128, 136, 102, 146, 112, 182, 127, 1201, 1385, 102], 
+     'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1], 
+     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
+
+This shows us what the `token_type_ids <glossary.html#token-type-ids>`__ are for: they indicate to the model which part
+of the inputs correspond to the first sentence and which part corresponds to the second sentence. Note that
+`token_type_ids` are not required or handled by all models. By default, a tokenizer will only return the inputs that
+its associated model expects. You can force the return (or the non-return) of any of those special arguments by
+using ``return_input_ids`` or ``return_token_type_ids``.
+
+If we decode the token ids we obtained, we will see that the special tokens have been properly added.
+
+::
+
+    tokenizer.decode(encoded_input["input_ids"])
+
+will return:
+
+::
+
+    "[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:
+
+::
+
+    {'input_ids': [[101, 8667, 146, 112, 182, 170, 1423, 5650, 102, 146, 112, 182, 170, 5650, 1115, 2947, 1114, 1103, 1148, 5650, 102], 
+                   [101, 1262, 1330, 5650, 102, 1262, 146, 1431, 1129, 12544, 1114, 1103, 1248, 5650, 102], 
+                   [101, 1262, 1103, 1304, 1304, 1314, 1141, 102, 1262, 146, 1301, 1114, 1103, 1304, 1314, 1141, 102]], 
+    'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                       [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                       [0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1]], 
+    'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                       [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], 
+                       [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]}
+
+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:
+
+::
+
+    [CLS] Hello I'm a single sentence [SEP] I'm a sentence that goes with the first sentence [SEP]
+    [CLS] And another sentence [SEP] And I should be encoded with the second sentence [SEP]
+    [CLS] And the very very last one [SEP] And I go with the very last one [SEP]
+
+Once again, you can automatically pad your inputs to the maximum sentence length in the batch, truncate to the maximum
+length the model can accept and return tensors directly with the following:
+
+::
+
+    ## PYTORCH CODE
+    batch = tokenizer(batch_sentences, batch_of_second_sentences, padding=True, truncation=True, return_tensors="pt")
+    ## TENSORFLOW CODE
+    batch = tokenizer(batch_sentences, batch_of_second_sentences, padding=True, truncation=True, return_tensors="tf")
+
+Everything you always wanted to know about padding and truncation
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+We have seen the commands that will work for most cases (pad your batch to the length of the maximum sentence and
+
+truncate to the maximum length the mode can accept). However, the API supports more strategies if you need them. The
+three arguments you need to know for this are :obj:`padding`, :obj:`truncation` and :obj:`max_length`.
+
+- :obj:`padding` controls the padding. It can be a boolean or a string which should be:
+
+    - :obj:`True` or :obj:`'longest'` to pad to the longest sequence in the batch (doing no padding if you only provide
+      a single sequence).
+    - :obj:`'max_length'` to pad to a length specified by the :obj:`max_length` argument or the maximum length accepted
+      by the model if no :obj:`max_length` is provided (``max_length=None``). If you only provide a single sequence,
+      padding will still be applied to it. 
+    - :obj:`False` or :obj:`'do_not_pad'` to not pad the sequences. As we have seen before, this is the default
+      behavior.
+
+- :obj:`truncation` controls the truncation. It can be a boolean or a string which should be:
+
+    - :obj:`True` or :obj:`'only_first'` truncate to a maximum length specified by the :obj:`max_length` argument or
+      the maximum length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will
+      only truncate the first sentence of a pair if a pair of sequence (or a batch of pairs of sequences) is provided.
+    - :obj:`'only_second'` truncate to a maximum length specified by the :obj:`max_length` argument or the maximum
+      length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will only truncate
+      the second sentence of a pair if a pair of sequence (or a batch of pairs of sequences) is provided.
+    - :obj:`'longest_first'` truncate to a maximum length specified by the :obj:`max_length` argument or the maximum
+      length accepted by the model if no :obj:`max_length` is provided (``max_length=None``). This will truncate token
+      by token, removing a token from the longest sequence in the pair until the proper length is reached.
+    - :obj:`False` or :obj:`'do_not_truncate'` to not truncate the sequences. As we have seen before, this is the
+      default behavior.
+
+- :obj:`max_length` to control the length of the padding/truncation. It can be an integer or :obj:`None`, in which case
+  it will default to the maximum length the model can accept. If the model has no specific maximum input length,
+  truncation/padding to :obj:`max_length` is deactivated.
+
+Here is a table summarizing the recommend way to setup padding and truncation. If you use pair of inputs sequence in
+any of the following examples, you can replace :obj:`truncation=True` by a :obj:`STRATEGY` selected in 
+:obj:`['only_first', 'only_second', 'longest_first']`, i.e. :obj:`truncation='only_second'` or
+:obj:`truncation= 'longest_first'` to control how both sequence in the pair are truncated as detailed before.
+
++--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
+| Truncation                           | Padding                           | Instruction                                                                                 |
++======================================+===================================+=============================================================================================+
+| no truncation                        | no padding                        | :obj:`tokenizer(batch_sentences)`                                                           |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True)` or                                          |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='longest')`                                        |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max model input length | :obj:`tokenizer(batch_sentences, padding='max_length')`                                     |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to specific length        | :obj:`tokenizer(batch_sentences, padding='max_length', max_length=42)`                      |
++--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
+| truncation to max model input length | no padding                        | :obj:`tokenizer(batch_sentences, truncation=True)` or                                       |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, truncation=STRATEGY)`                                      |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True, truncation=True)` or                         |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, padding=True, truncation=STRATEGY)`                        |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max model input length | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=True)` or                 |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY)`                |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to specific length        | Not possible                                                                                |
++--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
+| truncation to specific length        | no padding                        | :obj:`tokenizer(batch_sentences, truncation=True, max_length=42)` or                        |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, truncation=STRATEGY, max_length=42)`                       |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max sequence in batch  | :obj:`tokenizer(batch_sentences, padding=True, truncation=True, max_length=42)` or          |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, padding=True, truncation=STRATEGY, max_length=42)`         |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to max model input length | Not possible                                                                                |
+|                                      +-----------------------------------+---------------------------------------------------------------------------------------------+
+|                                      | padding to specific length        | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=True, max_length=42)` or  |
+|                                      |                                   | :obj:`tokenizer(batch_sentences, padding='max_length', truncation=STRATEGY, max_length=42)` |
++--------------------------------------+-----------------------------------+---------------------------------------------------------------------------------------------+
+
+Pre-tokenized inputs
+~~~~~~~~~~~~~~~~~~~~
+
+The tokenizer also accept pre-tokenized inputs. This is particularly useful when you want to compute labels and extract
+predictions in `named entity recognition (NER) <https://en.wikipedia.org/wiki/Named-entity_recognition>`__ or
+`part-of-speech tagging (POS tagging) <https://en.wikipedia.org/wiki/Part-of-speech_tagging>`__.
+
+If you want to use pre-tokenized inputs, just set :obj:`is_pretokenized=True` when passing your inputs to the
+tokenizer. For instance:
+
+::
+
+    encoded_input = tokenizer(["Hello", "I'm", "a", "single", "sentence"], is_pretokenized=True)
+    print(encoded_input)
+
+will return:
+
+::
+
+    {'input_ids': [101, 8667, 146, 112, 182, 170, 1423, 5650, 102],
+     'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0], 
+     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1]}
+
+Note that the tokenizer still adds the ids of special tokens (if applicable) unless you pass
+``add_special_tokens=False``.
+
+This works exactly as before for batch of sentences or batch of pairs of sentences. You can encode a batch of sentences
+like this:
+
+::
+
+    batch_sentences = [["Hello", "I'm", "a", "single", "sentence"],
+                       ["And", "another", "sentence"],
+                       ["And", "the", "very", "very", "last", "one"]]
+    encoded_inputs = tokenizer(batch_sentences, is_pretokenized=True)
+
+or a batch of pair sentences like this:
+
+::
+
+    batch_of_second_sentences = [["I'm", "a", "sentence", "that", "goes", "with", "the", "first", "sentence"],
+                                 ["And", "I", "should", "be", "encoded", "with", "the", "second", "sentence"],
+                                 ["And", "I", "go", "with", "the", "very", "last", "one"]]
+    encoded_inputs = tokenizer(batch_sentences, batch_of_second_sentences, is_pretokenized=True)
+
+And you can add padding, truncation as well as directly return tensors like before:
+
+::
+
+    ## PYTORCH CODE
+    batch = tokenizer(batch_sentences,
+                      batch_of_second_sentences,
+                      is_pretokenized=True,
+                      padding=True,
+                      truncation=True,
+                      return_tensors="pt")
+    ## TENSORFLOW CODE
+    batch = tokenizer(batch_sentences,
+                      batch_of_second_sentences,
+                      is_pretokenized=True,
+                      padding=True,
+                      truncation=True,
+                      return_tensors="tf")

docs/source/quicktour.rst

@@ -1,379 +1,394 @@
-Quick tour
-==========
-
-Let's have a quick look at the 🤗 Transformers library features. The library downloads pretrained models for
-Natural Language Understanding (NLU) tasks, such as analyzing the sentiment of a text, and Natural Language Generation (NLG),
-such as completing a prompt with new text or translating in another language.
-
-First we will see how to easily leverage the pipeline API to quickly use those pretrained models at inference. Then, we
-will dig a little bit more and see how the library gives you access to those models and helps you preprocess your data.
-
-.. note::
-
-    All code examples presented in the documentation have a switch on the top left for Pytorch versus TensorFlow. If
-    not, the code is expected to work for both backends without any change needed.
-
-Getting started on a task with a pipeline
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
-provides the following tasks out of the box:
-
-- Sentiment analysis: is a text positive or negative?
-- Text generation (in English): provide a prompt and the model will generate what follows.
-- Name entity recognition (NER): in an input sentence, label each word with the entity it represents (person, place,
-  etc.)
-- Question answering: provide the model with some context and a question, extract the answer from the context.
-- Filling masked text: given a text with masked words (e.g., replaced by ``[MASK]``), fill the blanks.
-- Summarization: generate a summary of a long text.
-- Translation: translate a text in another language.
-- Feature extraction: return a tensor representation of the text.
-
-Let's see how this work for sentiment analysis (the other tasks are all covered in the
-:doc:`task summary </task_summary>`):
-
-::
-
-    from transformers import pipeline
-    classifier = pipeline('sentiment-analysis')
-
-When typing this command for the first time, a pretrained model and its tokenizer are downloaded and cached. We will
-look at both later on, but as an introduction the tokenizer's job is to preprocess the text for the model, which is
-then responsible for making predictions. The pipeline groups all of that together, and post-process the predictions to
-make them readable. For instance
-
-::
-
-    classifier('We are very happy to show you the Transformers library.')
-
-will return something like this:
-
-::
-
-    [{'label': 'POSITIVE', 'score': 0.999799370765686}]
-
-That's encouraging! You can use it on a list of sentences, which will be preprocessed then fed to the model as a
-`batch`:
-
-::
-
-    classifier(["We are very happy to show you the Transformers library.",
-                "We hope you don't hate it."])
-
-returning a list of dictionaries like this one:
-
-::
-
-    [{'label': 'POSITIVE', 'score': 0.999799370765686},
-     {'label': 'NEGATIVE', 'score': 0.5308589935302734}]
-
-You can see the second sentence has been classified as negative (it needs to be positive or negative) but its score is
-fairly neutral.
-
-By default, the model downloaded for this pipeline is called "distilbert-base-uncased-finetuned-sst-2-english". We can
-look at its `model page <https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english>`__ to get more
-information about it. It uses the :doc:`DistilBERT architecture </model_doc/distilbert>` and has been fine-tuned on a
-dataset called SST-2 for the sentiment analysis task.
-
-Let's say we want to use another model; for instance, one that has been trained on French data. We can search through
-the `model hub <https://huggingface.co/models>`__ that gathers models pretrained on a lot of data by research labs, but
-also community models (usually fine-tuned versions of those big models on a specific dataset). Applying the tags
-"French" and "text-classification" gives back a suggestion "nlptown/bert-base-multilingual-uncased-sentiment". Let's
-see how we can use it. 
-
-You can directly pass the name of the model to use to :func:`~transformers.pipeline`:
-
-::
-
-    classifier = pipeline('sentiment-analysis', model="nlptown/bert-base-multilingual-uncased-sentiment")
-
-This classifier can now deal with texts in English, French, but also Dutch, German, Italian and Spanish! You can also
-replace that name by a local folder where you have saved a pretrained model (see below). You can also pass a model
-object and its associated tokenizer.
-
-We will need two classes for this. The first is :class:`~transformers.AutoTokenizer`, which we will use to download the
-tokenizer associated to the model we picked and instantiate it. The second is
-:class:`~transformers.AutoModelForSequenceClassification` (or
-:class:`~transformers.TFAutoModelForSequenceClassification` if you are using TensorFlow), which we will use to download
-the model itself. Note that if we were using the library on an other task, the class of the model would change. The
-:doc:`task summary </task_summary>` tutorial summarizes which class is used for which task.
-
-::
-
-    ## PYTORCH CODE
-    from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    ## TENSORFLOW CODE
-    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-
-Now, to download the models and tokenizer we found previously, we just have to use the 
-:func:`~transformers.AutoModelForSequenceClassification.from_pretrained` method (feel free to replace ``model_name`` by
-any other model from the model hub):
-
-::
-
-    ## PYTORCH CODE
-    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    pipe = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-    ## TENSORFLOW CODE
-    model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
-
-If you don't find a model that has been pretrained on some data similar to yours, you will need to fine-tune a
-pretrained model on your data. We provide :doc:`example scripts </examples>` to do so. Once you're done, don't forget
-to share your fine-tuned model on the hub with the community, using :doc:`this tutorial </model_sharing>`.
-
-.. _pretrained-model:
-
-Under the hood: pretrained models
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
-using the :obj:`from_pretrained` method:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import AutoTokenizer, AutoModelForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = AutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
-    tokenizer = AutoTokenizer.from_pretrained(model_name)
-
-Using the tokenizer
-^^^^^^^^^^^^^^^^^^^
-
-We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
-words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
-that process, which is why we need to instantiate the tokenizer using the name of the model, to make sure we use the
-same rules as when the model was pretrained.
-
-The second step is to convert those `tokens` into numbers, to be able to build a tensor out of them and feed them to
-the model. To do this, the tokenizer has a `vocab`, which is the part we download when we instantiate it with the
-:obj:`from_pretrained` method, since we need to use the same `vocab` as when the model was pretrained.
-
-To apply these steps on a given text, we can just feed it to our tokenizer:
-
-::
-
-    input = tokenizer("We are very happy to show you the Transformers library.")
-    print(input)
-
-This returns a dictionary string to list of ints. It contains the `ids of the tokens <glossary.html#input-ids>`__,
-as mentioned before, but also additional arguments that will be useful to the model. Here for instance, we also have an
-`attention mask <glossary.html#attention-mask>`__ that the model will use to have a better understanding of the sequence:
-
-
-::
-
-    {'input_ids': [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 19081, 3075, 1012, 102],
-     'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
-
-You can pass a list of sentences directly to your tokenizer. If your goal is to send them through your model as a
-batch, you probably want to pad them all to the same length, truncate them to the maximum length the model can accept
-and get tensors back. You can specify all of that to the tokenizer:
-
-::
-
-    ## PYTORCH CODE
-    batch = tokenizer(
-        ["We are very happy to show you the Transformers library.",
-         "We hope you don't hate it."],
-        padding=True, truncation=True, return_tensors="pt")
-    print(batch)
-    ## TENSORFLOW CODE
-    batch = tokenizer(
-        ["We are very happy to show you the Transformers library.",
-         "We hope you don't hate it."],
-        padding=True, truncation=True, return_tensors="tf")
-    print(batch)
-
-The padding is automatically applied on the side the model expect it (in this case, on the right), with the
-padding token the model was pretrained with. The attention mask is also adapted to take the padding into account:
-
-::
-
-    {'input_ids': tensor([[  101,  2057,  2024,  2200,  3407,  2000,  2265,  2017,  1996, 19081, 3075,  1012,   102],
-                          [  101,  2057,  3246,  2017,  2123,  1005,  1056,  5223,  2009,  1012,  102,     0,     0]]), 
-     'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-                               [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])}
-
-You can learn more about tokenizers on their :doc:`doc page <main_classes/tokenizer>` (tutorial coming soon).
-
-Using the model
-^^^^^^^^^^^^^^^
-
-Once your input has been preprocessed by the tokenizer, you can directly send it to the model. As we mentioned, it will
-contain all the relevant information the model needs. If you're using a TensorFlow model, you can directly pass the
-dictionary keys to tensor, for a PyTorch model, you need to unpack the dictionary by adding :obj:`**`.
-
-::
-
-    ## PYTORCH CODE
-    outputs = model(**batch)
-    ## TENSORFLOW CODE
-    outputs = model(batch)
-
-In 🤗 Transformers, all outputs are tuples (with only one element potentially). Here, we get a tuple with just the
-final activations of the model.
-
-::
-
-    (tensor([[-4.1329,  4.3811],
-             [ 0.0818, -0.0418]]),)
-
-.. note::
-
-    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
-    activation function (like SoftMax) since this final activation function is often fused with the loss.
-
-Let's apply the SoftMax activation to get predictions.
-
-::
-
-    ## PYTORCH CODE
-    import torch.nn.functional as F
-    predictions = F.softmax(outputs[0], dim=-1)
-    print(predictions)
-    ## TENSORFLOW CODE
-    predictions = tf.nn.softmax(outputs[0], axis=-1)
-    print(predictions)
-
-We can see we get the numbers from before:
-
-::
-
-    tensor([[2.0060e-04, 9.9980e-01],
-            [5.3086e-01, 4.6914e-01]])
-
-If you have labels, you can provide them to the model, it will return a tuple with the loss and the final activations.
-
-::
-
-    ## PYTORCH CODE
-    import torch
-    outputs = model(**batch, labels = torch.tensor([1, 0])
-    ## TENSORFLOW CODE
-    import tensorflow as tf
-    outputs = model(batch, labels = tf.constant([1, 0])
-
-Models are standard `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ or
-`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ so you can use them in your usual
-training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer` (or :class:`~transformers.TFTrainer` if
-you are using TensorFlow) class to help with your training (taking care of things such as distributed training, mixed
-precision, etc.). See the training tutorial (coming soon) for more details.
-
-Once your model is fine-tuned, you can save it with its tokenizer the following way:
-
-::
-
-    tokenizer.save_pretrained(save_directory)
-    model.save_pretrained(save_directory)
-
-You can then load this model back using the :func:`~transformers.AutoModel.from_pretrained` method by passing the
-directory name instead of the model name. One cool feature of 🤗 Transformers is that you can easily switch between
-PyTorch and TensorFlow: any model saved as before can be loaded back either in PyTorch or TensorFlow. If you are
-loading a saved PyTorch model in a TensorFlow model, use :func:`~transformers.TFAutoModel.from_pretrained` like this:
-
-::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = TFAutoModel.from_pretrained(save_directory, from_pt=True)
-
-and if you are loading a saved TensorFlow model in a PyTorch model, you should use the following code:
-
-::
-
-    tokenizer = AutoTokenizer.from_pretrained(save_directory)
-    model = AutoModel.from_pretrained(save_directory, from_tf=True)
-
-Lastly, you can also ask the model to return all hidden states and all attention weights if you need them:
-
-
-::
-
-    ## PYTORCH CODE
-    outputs = model(**batch, output_hidden_states=True, output_attentions=True)
-    all_hidden_states, all_attentions = outputs[-2:]
-    ## TENSORFLOW CODE
-    outputs = model(batch, output_hidden_states=True, output_attentions=True)
-    all_hidden_states, all_attentions = outputs[-2:]
-
-Accessing the code
-^^^^^^^^^^^^^^^^^^
-
-The :obj:`AutoModel` and :obj:`AutoTokenizer` classes are just shortcuts that will automatically work with any
-pretrained model. Behind the scenes, the library has one model class per combination of architecture plus class, so the
-code is easy to access and tweak if you need to.
-
-In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
-using the :doc:`DistilBERT </model_doc/distilbert>` architecture. The model automatically created is then a
-:class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
-to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
-without the auto magic:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = DistilBertForSequenceClassification.from_pretrained(model_name)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased-finetuned-sst-2-english"
-    model = TFDistilBertForSequenceClassification.from_pretrained(model_name)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-
-Customizing the model
-^^^^^^^^^^^^^^^^^^^^^
-
-If you want to change how the model itself is built, you can define your custom configuration class. Each architecture
-comes with its own relevant configuration (in the case of DistilBERT, :class:`~transformers.DistilBertConfig`) which
-allows you to specify any of the hidden dimension, dropout rate etc. If you do core modifications, like changing the
-hidden size, you won't be able to use a pretrained model anymore and will need to train from scratch. You would then
-instantiate the model directly from this configuration.
-
-Here we use the predefined vocabulary of DistilBERT (hence load the tokenizer with the
-:func:`~transformers.DistilBertTokenizer.from_pretrained` method) and initialize the model from scratch (hence
-instantiate the model from the configuration instead of using the
-:func:`~transformers.DistilBertForSequenceClassification.from_pretrained` method).
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    model = DistilBertForSequenceClassification(config)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
-    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
-    model = TFDistilBertForSequenceClassification(config)
-
-For something that only changes the head of the model (for instance, the number of labels), you can still use a
-pretrained model for the body. For instance, let's define a classifier for 10 different labels using a pretrained body.
-We could create a configuration with all the default values and just change the number of labels, but more easily, you
-can directly pass any argument a configuration would take to the :func:`from_pretrained` method and it will update the
-default configuration with it:
-
-::
-
-    ## PYTORCH CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased"
-    model = DistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
-    ## TENSORFLOW CODE
-    from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
-    model_name = "distilbert-base-uncased"
-    model = TFDistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
-    tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+Quick tour
+==========
+
+Let's have a quick look at the 🤗 Transformers library features. The library downloads pretrained models for
+Natural Language Understanding (NLU) tasks, such as analyzing the sentiment of a text, and Natural Language Generation (NLG),
+such as completing a prompt with new text or translating in another language.
+
+First we will see how to easily leverage the pipeline API to quickly use those pretrained models at inference. Then, we
+will dig a little bit more and see how the library gives you access to those models and helps you preprocess your data.
+
+.. note::
+
+    All code examples presented in the documentation have a switch on the top left for Pytorch versus TensorFlow. If
+    not, the code is expected to work for both backends without any change needed.
+
+Getting started on a task with a pipeline
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The easiest way to use a pretrained model on a given task is to use :func:`~transformers.pipeline`. 🤗 Transformers
+provides the following tasks out of the box:
+
+- Sentiment analysis: is a text positive or negative?
+- Text generation (in English): provide a prompt and the model will generate what follows.
+- Name entity recognition (NER): in an input sentence, label each word with the entity it represents (person, place,
+  etc.)
+- Question answering: provide the model with some context and a question, extract the answer from the context.
+- Filling masked text: given a text with masked words (e.g., replaced by ``[MASK]``), fill the blanks.
+- Summarization: generate a summary of a long text.
+- Translation: translate a text in another language.
+- Feature extraction: return a tensor representation of the text.
+
+Let's see how this work for sentiment analysis (the other tasks are all covered in the
+:doc:`task summary </task_summary>`):
+
+.. code-block::
+
+    >>> from transformers import pipeline
+    >>> classifier = pipeline('sentiment-analysis')
+
+When typing this command for the first time, a pretrained model and its tokenizer are downloaded and cached. We will
+look at both later on, but as an introduction the tokenizer's job is to preprocess the text for the model, which is
+then responsible for making predictions. The pipeline groups all of that together, and post-process the predictions to
+make them readable. For instance:
+
+
+.. code-block::
+
+    >>> classifier('We are very happy to show you the 🤗 Transformers library.')
+    [{'label': 'POSITIVE', 'score': 0.9997795224189758}]
+
+That's encouraging! You can use it on a list of sentences, which will be preprocessed then fed to the model as a
+`batch`, returning a list of dictionaries like this one:
+
+.. code-block::
+
+    >>> results = classifier(["We are very happy to show you the 🤗 Transformers library.",
+    ...            "We hope you don't hate it."])
+    >>> for result in results:
+    ...     print(f"label: {result['label']}, with score: {round(result['score'], 4)}")
+    label: POSITIVE, with score: 0.9998
+    label: NEGATIVE, with score: 0.5309
+
+You can see the second sentence has been classified as negative (it needs to be positive or negative) but its score is
+fairly neutral.
+
+By default, the model downloaded for this pipeline is called "distilbert-base-uncased-finetuned-sst-2-english". We can
+look at its `model page <https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english>`__ to get more
+information about it. It uses the :doc:`DistilBERT architecture </model_doc/distilbert>` and has been fine-tuned on a
+dataset called SST-2 for the sentiment analysis task.
+
+Let's say we want to use another model; for instance, one that has been trained on French data. We can search through
+the `model hub <https://huggingface.co/models>`__ that gathers models pretrained on a lot of data by research labs, but
+also community models (usually fine-tuned versions of those big models on a specific dataset). Applying the tags
+"French" and "text-classification" gives back a suggestion "nlptown/bert-base-multilingual-uncased-sentiment". Let's
+see how we can use it.
+
+You can directly pass the name of the model to use to :func:`~transformers.pipeline`:
+
+.. code-block::
+
+    >>> classifier = pipeline('sentiment-analysis', model="nlptown/bert-base-multilingual-uncased-sentiment")
+
+This classifier can now deal with texts in English, French, but also Dutch, German, Italian and Spanish! You can also
+replace that name by a local folder where you have saved a pretrained model (see below). You can also pass a model
+object and its associated tokenizer.
+
+We will need two classes for this. The first is :class:`~transformers.AutoTokenizer`, which we will use to download the
+tokenizer associated to the model we picked and instantiate it. The second is
+:class:`~transformers.AutoModelForSequenceClassification` (or
+:class:`~transformers.TFAutoModelForSequenceClassification` if you are using TensorFlow), which we will use to download
+the model itself. Note that if we were using the library on an other task, the class of the model would change. The
+:doc:`task summary </task_summary>` tutorial summarizes which class is used for which task.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+
+Now, to download the models and tokenizer we found previously, we just have to use the
+:func:`~transformers.AutoModelForSequenceClassification.from_pretrained` method (feel free to replace ``model_name`` by
+any other model from the model hub):
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
+    >>> model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
+    >>> pipe = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
+    >>> ## TENSORFLOW CODE
+    >>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
+    >>> # This model only exists in PyTorch, so we use the `from_pt` flag to import that model in TensorFlow.
+    >>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name, from_pt=True) 
+    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
+    >>> classifier = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
+
+If you don't find a model that has been pretrained on some data similar to yours, you will need to fine-tune a
+pretrained model on your data. We provide :doc:`example scripts </examples>` to do so. Once you're done, don't forget
+to share your fine-tuned model on the hub with the community, using :doc:`this tutorial </model_sharing>`.
+
+.. _pretrained-model:
+
+Under the hood: pretrained models
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Let's now see what happens beneath the hood when using those pipelines. As we saw, the model and tokenizer are created
+using the :obj:`from_pretrained` method:
+
+::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import AutoTokenizer, AutoModelForSequenceClassification
+    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    >>> pt_model = AutoModelForSequenceClassification.from_pretrained(model_name)
+    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
+    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    >>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
+    >>> tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+Using the tokenizer
+^^^^^^^^^^^^^^^^^^^
+
+We mentioned the tokenizer is responsible for the preprocessing of your texts. First, it will split a given text in
+words (or part of words, punctuation symbols, etc.) usually called `tokens`. There are multiple rules that can govern
+that process (you can learn more about them in the :doc:`tokenizer_summary <tokenizer_summary>`, which is why we need
+to instantiate the tokenizer using the name of the model, to make sure we use the same rules as when the model was
+pretrained.
+
+The second step is to convert those `tokens` into numbers, to be able to build a tensor out of them and feed them to
+the model. To do this, the tokenizer has a `vocab`, which is the part we download when we instantiate it with the
+:obj:`from_pretrained` method, since we need to use the same `vocab` as when the model was pretrained.
+
+To apply these steps on a given text, we can just feed it to our tokenizer:
+
+.. code-block::
+
+    >>> inputs = tokenizer("We are very happy to show you the 🤗 Transformers library.")
+
+This returns a dictionary string to list of ints. It contains the `ids of the tokens <glossary.html#input-ids>`__,
+as mentioned before, but also additional arguments that will be useful to the model. Here for instance, we also have an
+`attention mask <glossary.html#attention-mask>`__ that the model will use to have a better understanding of the sequence:
+
+
+.. code-block::
+
+    >>> print(inputs)
+    {'input_ids': [101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
+
+You can pass a list of sentences directly to your tokenizer. If your goal is to send them through your model as a
+batch, you probably want to pad them all to the same length, truncate them to the maximum length the model can accept
+and get tensors back. You can specify all of that to the tokenizer:
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> pt_batch = tokenizer(
+    ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
+    ...     padding=True,
+    ...     truncation=True,
+    ...     return_tensors="pt"
+    ... )
+    >>> ## TENSORFLOW CODE
+    >>> tf_batch = tokenizer(
+    ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
+    ...     padding=True,
+    ...     truncation=True,
+    ...     return_tensors="tf"
+    ... )
+
+The padding is automatically applied on the side the model expect it (in this case, on the right), with the
+padding token the model was pretrained with. The attention mask is also adapted to take the padding into account:
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> for key, value in pt_batch.items():
+    ...     print(f"{key}: {value.numpy().tolist()}")
+    input_ids: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0, 0]]
+    attention_mask: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]]
+    >>> ## TENSORFLOW CODE
+    >>> for key, value in tf_batch.items():
+    ...     print(f"{key}: {value.numpy().tolist()}")
+    input_ids: [[101, 2057, 2024, 2200, 3407, 2000, 2265, 2017, 1996, 100, 19081, 3075, 1012, 102], [101, 2057, 3246, 2017, 2123, 1005, 1056, 5223, 2009, 1012, 102, 0, 0, 0]]
+    attention_mask: [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0]]
+
+You can learn more about tokenizers :doc:`here <preprocessing>`.
+
+Using the model
+^^^^^^^^^^^^^^^
+
+Once your input has been preprocessed by the tokenizer, you can 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:`**`.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> pt_outputs = pt_model(**pt_batch)
+    >>> ## TENSORFLOW CODE
+    >>> tf_outputs = tf_model(tf_batch)
+
+In 🤗 Transformers, all outputs are tuples (with only one element potentially). Here, we get a tuple with just the
+final activations of the model.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> print(pt_outputs)
+    (tensor([[-4.0833,  4.3364],
+            [ 0.0818, -0.0418]], grad_fn=<AddmmBackward>),)
+    >>> ## TENSORFLOW CODE
+    >>> print(tf_outputs)
+    (<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
+    array([[-4.0832963 ,  4.3364134 ],
+           [ 0.08181238, -0.04178794]], dtype=float32)>,)
+
+.. note::
+
+    All 🤗 Transformers models (PyTorch or TensorFlow) return the activations of the model *before* the final
+    activation function (like SoftMax) since this final activation function is often fused with the loss.
+
+Let's apply the SoftMax activation to get predictions.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> import torch.nn.functional as F
+    >>> pt_predictions = F.softmax(pt_outputs[0], dim=-1)
+    >>> ## TENSORFLOW CODE
+    >>> import tensorflow as tf
+    >>> tf_predictions = tf.nn.softmax(tf_outputs[0], axis=-1)
+
+We can see we get the numbers from before:
+
+.. code-block::
+
+    >>> ## TENSORFLOW CODE
+    >>> print(tf_predictions)
+    tf.Tensor(
+    [[2.2042994e-04 9.9977952e-01]
+     [5.3086078e-01 4.6913919e-01]], shape=(2, 2), dtype=float32)
+    >>> ## PYTORCH CODE
+    >>> print(pt_predictions)
+    tensor([[2.2043e-04, 9.9978e-01],
+            [5.3086e-01, 4.6914e-01]], grad_fn=<SoftmaxBackward>)
+
+If you have labels, you can provide them to the model, it will return a tuple with the loss and the final activations.
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> import torch
+    >>> pt_outputs = pt_model(**pt_batch, labels = torch.tensor([1, 0]))
+    >>> ## TENSORFLOW CODE
+    >>> import tensorflow as tf
+    >>> tf_outputs = tf_model(tf_batch, labels = tf.constant([1, 0]))
+
+Models are standard `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__ or
+`tf.keras.Model <https://www.tensorflow.org/api_docs/python/tf/keras/Model>`__ so you can use them in your usual
+training loop. 🤗 Transformers also provides a :class:`~transformers.Trainer` (or :class:`~transformers.TFTrainer` if
+you are using TensorFlow) class to help with your training (taking care of things such as distributed training, mixed
+precision, etc.). See the :doc:`training tutorial <training>` for more details.
+
+Once your model is fine-tuned, you can save it with its tokenizer the following way:
+
+::
+
+    tokenizer.save_pretrained(save_directory)
+    model.save_pretrained(save_directory)
+
+You can then load this model back using the :func:`~transformers.AutoModel.from_pretrained` method by passing the
+directory name instead of the model name. One cool feature of 🤗 Transformers is that you can easily switch between
+PyTorch and TensorFlow: any model saved as before can be loaded back either in PyTorch or TensorFlow. If you are
+loading a saved PyTorch model in a TensorFlow model, use :func:`~transformers.TFAutoModel.from_pretrained` like this:
+
+::
+
+    tokenizer = AutoTokenizer.from_pretrained(save_directory)
+    model = TFAutoModel.from_pretrained(save_directory, from_pt=True)
+
+and if you are loading a saved TensorFlow model in a PyTorch model, you should use the following code:
+
+::
+
+    tokenizer = AutoTokenizer.from_pretrained(save_directory)
+    model = AutoModel.from_pretrained(save_directory, from_tf=True)
+
+Lastly, you can also ask the model to return all hidden states and all attention weights if you need them:
+
+
+::
+
+    >>> ## PYTORCH CODE
+    >>> pt_outputs = pt_model(**pt_batch, output_hidden_states=True, output_attentions=True)
+    >>> all_hidden_states, all_attentions = pt_outputs[-2:]
+    >>> ## TENSORFLOW CODE
+    >>> tf_outputs = tf_model(tf_batch, output_hidden_states=True, output_attentions=True)
+    >>> all_hidden_states, all_attentions = tf_outputs[-2:]
+
+Accessing the code
+^^^^^^^^^^^^^^^^^^
+
+The :obj:`AutoModel` and :obj:`AutoTokenizer` classes are just shortcuts that will automatically work with any
+pretrained model. Behind the scenes, the library has one model class per combination of architecture plus class, so the
+code is easy to access and tweak if you need to.
+
+In our previous example, the model was called "distilbert-base-uncased-finetuned-sst-2-english", which means it's
+using the :doc:`DistilBERT </model_doc/distilbert>` architecture. The model automatically created is then a
+:class:`~transformers.DistilBertForSequenceClassification`. You can look at its documentation for all details relevant
+to that specific model, or browse the source code. This is how you would directly instantiate model and tokenizer
+without the auto magic:
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import DistilBertTokenizer, DistilBertForSequenceClassification
+    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    >>> model = DistilBertForSequenceClassification.from_pretrained(model_name)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import DistilBertTokenizer, TFDistilBertForSequenceClassification
+    >>> model_name = "distilbert-base-uncased-finetuned-sst-2-english"
+    >>> model = TFDistilBertForSequenceClassification.from_pretrained(model_name)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+
+Customizing the model
+^^^^^^^^^^^^^^^^^^^^^
+
+If you want to change how the model itself is built, you can define your custom configuration class. Each architecture
+comes with its own relevant configuration (in the case of DistilBERT, :class:`~transformers.DistilBertConfig`) which
+allows you to specify any of the hidden dimension, dropout rate etc. If you do core modifications, like changing the
+hidden size, you won't be able to use a pretrained model anymore and will need to train from scratch. You would then
+instantiate the model directly from this configuration.
+
+Here we use the predefined vocabulary of DistilBERT (hence load the tokenizer with the
+:func:`~transformers.DistilBertTokenizer.from_pretrained` method) and initialize the model from scratch (hence
+instantiate the model from the configuration instead of using the
+:func:`~transformers.DistilBertForSequenceClassification.from_pretrained` method).
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
+    >>> config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+    >>> model = DistilBertForSequenceClassification(config)
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
+    >>> config = DistilBertConfig(n_heads=8, dim=512, hidden_dim=4*512)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+    >>> model = TFDistilBertForSequenceClassification(config)
+
+For something that only changes the head of the model (for instance, the number of labels), you can still use a
+pretrained model for the body. For instance, let's define a classifier for 10 different labels using a pretrained body.
+We could create a configuration with all the default values and just change the number of labels, but more easily, you
+can directly pass any argument a configuration would take to the :func:`from_pretrained` method and it will update the
+default configuration with it:
+
+.. code-block::
+
+    >>> ## PYTORCH CODE
+    >>> from transformers import DistilBertConfig, DistilBertTokenizer, DistilBertForSequenceClassification
+    >>> model_name = "distilbert-base-uncased"
+    >>> model = DistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)
+    >>> ## TENSORFLOW CODE
+    >>> from transformers import DistilBertConfig, DistilBertTokenizer, TFDistilBertForSequenceClassification
+    >>> model_name = "distilbert-base-uncased"
+    >>> model = TFDistilBertForSequenceClassification.from_pretrained(model_name, num_labels=10)
+    >>> tokenizer = DistilBertTokenizer.from_pretrained(model_name)

docs/source/serialization.rst

@@ -1,89 +0,0 @@
-Serialization best-practices
-^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-This section explain how you can save and re-load a fine-tuned model (BERT, GPT, GPT-2 and Transformer-XL).
-There are three types of files you need to save to be able to reload a fine-tuned model:
-
-
-* the model itself which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
-* the configuration file of the model which is saved as a JSON file, and
-* the vocabulary (and the merges for the BPE-based models GPT and GPT-2).
-
-The *default filenames* of these files are as follow:
-
-
-* the model weights file: ``pytorch_model.bin``\ ,
-* the configuration file: ``config.json``\ ,
-* the vocabulary file: ``vocab.txt`` for BERT and Transformer-XL, ``vocab.json`` for GPT/GPT-2 (BPE vocabulary),
-* for GPT/GPT-2 (BPE vocabulary) the additional merges file: ``merges.txt``.
-
-**If you save a model using these *default filenames*\ , you can then re-load the model and tokenizer using the ``from_pretrained()`` method.**
-
-Here is the recommended way of saving the model, configuration and vocabulary to an ``output_dir`` directory and reloading the model and tokenizer afterwards:
-
-.. code-block:: python
-
-   from transformers import WEIGHTS_NAME, CONFIG_NAME
-
-   output_dir = "./models/"
-
-   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned
-
-   # If we have a distributed model, save only the encapsulated model
-   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
-   model_to_save = model.module if hasattr(model, 'module') else model
-
-   # If we save using the predefined names, we can load using `from_pretrained`
-   output_model_file = os.path.join(output_dir, WEIGHTS_NAME)
-   output_config_file = os.path.join(output_dir, CONFIG_NAME)
-
-   torch.save(model_to_save.state_dict(), output_model_file)
-   model_to_save.config.to_json_file(output_config_file)
-   tokenizer.save_pretrained(output_dir)
-
-   # Step 2: Re-load the saved model and vocabulary
-
-   # Example for a Bert model
-   model = BertForQuestionAnswering.from_pretrained(output_dir)
-   tokenizer = BertTokenizer.from_pretrained(output_dir)  # Add specific options if needed
-   # Example for a GPT model
-   model = OpenAIGPTDoubleHeadsModel.from_pretrained(output_dir)
-   tokenizer = OpenAIGPTTokenizer.from_pretrained(output_dir)
-
-Here is another way you can save and reload the model if you want to use specific paths for each type of files:
-
-.. code-block:: python
-
-   output_model_file = "./models/my_own_model_file.bin"
-   output_config_file = "./models/my_own_config_file.bin"
-   output_vocab_file = "./models/my_own_vocab_file.bin"
-
-   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned
-
-   # If we have a distributed model, save only the encapsulated model
-   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
-   model_to_save = model.module if hasattr(model, 'module') else model
-
-   torch.save(model_to_save.state_dict(), output_model_file)
-   model_to_save.config.to_json_file(output_config_file)
-   tokenizer.save_vocabulary(output_vocab_file)
-
-   # Step 2: Re-load the saved model and vocabulary
-
-   # We didn't save using the predefined WEIGHTS_NAME, CONFIG_NAME names, we cannot load using `from_pretrained`.
-   # Here is how to do it in this situation:
-
-   # Example for a Bert model
-   config = BertConfig.from_json_file(output_config_file)
-   model = BertForQuestionAnswering(config)
-   state_dict = torch.load(output_model_file)
-   model.load_state_dict(state_dict)
-   tokenizer = BertTokenizer(output_vocab_file, do_lower_case=args.do_lower_case)
-
-   # Example for a GPT model
-   config = OpenAIGPTConfig.from_json_file(output_config_file)
-   model = OpenAIGPTDoubleHeadsModel(config)
-   state_dict = torch.load(output_model_file)
-   model.load_state_dict(state_dict)
-   tokenizer = OpenAIGPTTokenizer(output_vocab_file)
-

docs/source/tokenizer_summary.rst

@@ -0,0 +1,243 @@
+Tokenizer summary
+-----------------
+
+In this page, we will have a closer look at tokenization. As we saw in
+:doc:`the preprocessing tutorial <preprocessing>`, tokenizing a text is splitting it into words or subwords, which then
+are converted to ids. The second part is pretty straightforward, here we will focus on the first part. More
+specifically, we will look at the three main different kinds of tokenizers used in 🤗 Transformers:
+:ref:`Byte-Pair Encoding (BPE) <byte-pair-encoding>`, :ref:`WordPiece <wordpiece>` and
+:ref:`SentencePiece <sentencepiece>`, and provide examples of models using each of those.
+
+Note that on each model page, you can look at the documentation of the associated tokenizer to know which of those
+algorithms the pretrained model used. For instance, if we look at :class:`~transformers.BertTokenizer`, we can see it's
+using :ref:`WordPiece <wordpiece>`.
+
+Introduction to tokenization
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Splitting a text in smaller chunks is a task that's harder than it looks, and there are multiple ways of doing it. For
+instance, let's look at the sentence "Don't you love 🤗 Transformers? We sure do." A first simple way of tokenizing
+this text is just to split it by spaces, which would give:
+
+::
+
+    ["Don't", "you", "love", "🤗", "Transformers?", "We", "sure", "do."]
+
+This is a nice first step, but if we look at the tokens "Transformers?" or "do.", we can see we can do better. Those
+will be different than the tokens "Transformers" and "do" for our model, so we should probably take the punctuation
+into account. This would give:
+
+::
+
+    ["Don", "'", "t", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
+
+which is better already. One thing that is annoying though is how it dealt with "Don't". "Don't" stands for do not, so
+it should probably be better tokenized as ``["Do", "n't"]``. This is where things start getting more complicated, and
+part of the reason each kind of model has its own tokenizer class. Depending on the rules we apply to split our texts
+into tokens, we'll get different tokenized versions of the same text. And of course, a given pretrained model won't
+perform properly if you don't use the exact same rules as the persons who pretrained it.
+
+`spaCy <https://spacy.io/>`__ and `Moses <http://www.statmt.org/moses/?n=Development.GetStarted>`__ are two popular
+rule-based tokenizers. On the text above, they'd output something like:
+
+::
+
+    ["Do", "n't", "you", "love", "🤗", "Transformers", "?", "We", "sure", "do", "."]
+
+Space/punctuation-tokenization and rule-based tokenization are both examples of word tokenization, which is splitting a
+sentence into words. While it's the most intuitive way to separate texts in smaller chunks, it can have a problem when
+you have a huge corpus: it usually yields a very big vocabulary (the set of all unique tokens used).
+:doc:`Transformer XL <model_doc/transformerxl>` for instance uses space/punctuation-tokenization, and has a vocabulary
+size of 267,735!
+
+A huge vocabulary size means a huge embedding matrix at the start of the model, which will cause memory problems.
+TransformerXL deals with it by using a special kind of embeddings called adaptive embeddings, but in general,
+transformers model rarely have a vocabulary size greater than 50,000, especially if they are trained on a single
+language.
+
+So if tokenizing on words is unsatisfactory, we could go on the opposite direction and simply tokenize on characters.
+While it's very simple and would save a lot of memory, this doesn't allow the model to learn representations of texts
+as meaningful as when using a word tokenization, leading to a loss of performance. So to get the best of both worlds,
+all transformers models use a hybrid between word-level and character-level tokenization called subword tokenization.
+
+Subword tokenization
+^^^^^^^^^^^^^^^^^^^^
+
+Subword tokenization algorithms rely on the principle that most common words should be left as is, but rare words
+should be decomposed in meaningful subword units. For instance "annoyingly" might be considered a rare word and
+decomposed as "annoying" and "ly". This is especially useful in agglutinative languages such as Turkish, where you can
+form (almost) arbitrarily long complex words by stringing together some subwords.
+
+This allows the model to keep a reasonable vocabulary while still learning useful representations for common words or
+subwords. This also gives the ability to 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:
+
+::
+
+    >>> from transformers import BertTokenizer
+    >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+    >>> tokenizer.tokenize("I have a new GPU!")
+    ['i', 'have', 'a', 'new', 'gp', '##u', '!']
+
+Since we are considering the uncased model, the sentence was lowercased first. Then all the words were present in the
+vocabulary of the tokenizer, except for "gpu", so the tokenizer split it in subwords it knows: "gp" and "##u". The "##"
+means that the rest of the token should be attached to the previous one, without space (for when we need to decode
+predictions and reverse the tokenization).
+
+Another example is when we use the base :class:`~transformers.XLNetTokenizer` to tokenize our previous text:
+
+::
+
+    >>> from transformers import XLNetTokenizer
+    >>> tokenizer = XLNetTokenizer.from_pretrained('xlnet-base-cased')
+    >>> tokenizer.tokenize("Don't you love 🤗 Transformers? We sure do.")
+    ['▁Don', "'", 't', '▁you', '▁love', '▁', '🤗', '▁', 'Transform', 'ers', '?', '▁We', '▁sure', '▁do', '.']
+
+We'll get back to the meaning of those '▁' when we look at :ref:`SentencePiece <sentencepiece>` but you can see
+Transformers has been split into "Transform" and "ers".
+
+Let's now look at how the different subword tokenization algorithms work. Note that they all rely on some form of
+training which is usually done on the corpus the corresponding model will be trained on.
+
+.. _byte-pair-encoding:
+
+Byte-Pair Encoding
+~~~~~~~~~~~~~~~~~~
+
+Byte-Pair Encoding was introduced in `this paper <https://arxiv.org/abs/1508.07909>`__. It relies on a pretokenizer
+splitting the training data into words, which can be a simple space tokenization
+(:doc:`GPT-2 <model_doc/gpt2>` and :doc:`Roberta <model_doc/roberta>` uses this for instance) or a rule-based tokenizer
+(:doc:`XLM <model_doc/xlm>` use Moses for most languages, as does :doc:`FlauBERT <model_doc/flaubert>`),
+
+:doc:`GPT <model_doc/gpt>` uses Spacy and ftfy) and, counts the frequency of each word in the training corpus.
+
+It then begins from the list of all characters, and will learn merge rules to form a new token from two symbols in the
+vocabulary until it has learned a vocabulary of the desired size (this is a hyperparameter to pick).
+
+Let's say that after the pre-tokenization we have the following words (the number indicating the frequency of each
+word):
+
+::
+
+    ('hug', 10), ('pug', 5), ('pun', 12), ('bun', 4), ('hugs', 5)
+
+Then the base vocabulary is ['b', 'g', 'h', 'n', 'p', 's', 'u'] and all our words are first split by character:
+
+::
+
+    ('h' 'u' 'g', 10), ('p' 'u' 'g', 5), ('p' 'u' 'n', 12), ('b' 'u' 'n', 4), ('h' 'u' 'g' 's', 5)
+
+We then take each pair of symbols and look at the most frequent. For instance 'hu' is present `10 + 5 = 15` times (10
+times in the 10 occurrences of 'hug', 5 times in the 5 occurrences of 'hugs'). The most frequent here is 'ug', present
+`10 + 5 + 2 + 5 = 22` times in total. So the first merge rule the tokenizer learns is to group all 'u' and 'g' together
+then it adds 'ug' to the vocabulary. Our corpus then becomes
+
+::
+
+    ('h' 'ug', 10), ('p' 'ug', 5), ('p' 'u' 'n', 12), ('b' 'u' 'n', 4), ('h' 'ug' 's', 5)
+
+and we continue by looking at the next most common pair of symbols. It's 'un', present 16 times, so we merge those two
+and add 'un' to the vocabulary. Then it's 'hug' (as 'h' + 'ug'), present 15 times, so we merge those two and add 'hug'
+to the vocabulary.
+
+At this stage, the vocabulary is ``['b', 'g', 'h', 'n', 'p', 's', 'u', 'ug', 'un', 'hug']`` and our corpus is
+represented as
+
+::
+
+    ('hug', 10), ('p' 'ug', 5), ('p' 'un', 12), ('b' 'un', 4), ('hug' 's', 5)
+
+If we stop there, the tokenizer can apply the rules it learned to new words (as long as they don't contain characters that
+were not in the base vocabulary). For instance 'bug' would be tokenized as ``['b', 'ug']`` but mug would be tokenized as
+``['<unk>', 'ug']`` since the 'm' is not in the base vocabulary. This doesn't happen to letters in general (since the
+base corpus uses all of them), but to special characters like emojis.
+
+As we said before, the vocabulary size (which is the base vocabulary size + the number of merges) is a hyperparameter
+to choose. For instance :doc:`GPT <model_doc/gpt>` has a vocabulary size of 40,478 since they have 478 base characters
+and chose to stop the training of the tokenizer at 40,000 merges.
+
+Byte-level BPE
+^^^^^^^^^^^^^^
+
+To deal with the fact the base vocabulary needs to get all base characters, which can be quite big if one allows for
+all unicode characters, the
+`GPT-2 paper <https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf>`__
+introduces a clever trick, which is to use bytes as the base vocabulary (which gives a size of 256). With some
+additional rules to deal with punctuation, this manages to be able to tokenize every text without needing an unknown
+token. For instance, the :doc:`GPT-2 model <model_doc/gpt>` has a vocabulary size of 50,257, which corresponds to the
+256 bytes base tokens, a special end-of-text token and the symbols learned with 50,000 merges.
+
+.. _wordpiece:
+
+WordPiece
+=========
+
+WordPiece is the subword tokenization algorithm used for :doc:`BERT <model_doc/bert>` (as well as
+:doc:`DistilBERT <model_doc/distilbert>` and :doc:`Electra <model_doc/electra>`) and was outlined in
+`this paper <https://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/37842.pdf>`__. It relies
+on the same base as BPE, which is to initialize the vocabulary to every character present in the corpus and
+progressively learn a given number of merge rules, the difference is that it doesn't choose the pair that is the most
+frequent but the one that will maximize the likelihood on the corpus once merged. 
+
+What does this mean? Well, in the previous example, it means we would only merge 'u' and 'g' if the probability of
+having 'ug' divided by the probability of having 'u' then 'g' is greater than for any other pair of symbols. It's
+subtly different from what BPE does in the sense that it evaluates what it "loses" by merging two symbols and makes
+sure it's `worth it`.
+
+.. _unigram:
+
+Unigram
+=======
+
+Unigram is a subword tokenization algorithm introduced in `this paper <https://arxiv.org/pdf/1804.10959.pdf>`__.
+Instead of starting with a group of base symbols and learning merges with some rule, like BPE or WordPiece, it starts
+from a large vocabulary (for instance, all pretokenized words and the most common substrings) that it will trim down
+progressively. It's not used directly for any of the pretrained models in the library, but it's used in conjunction
+with :ref:`SentencePiece <sentencepiece>`.
+
+More specifically, at a given step, unigram computes a loss from the corpus we have and the current vocabulary, then,
+for each subword, evaluate how much the loss would augment if the subword was removed from the vocabulary. It then
+sorts the subwords by this quantity (that represents how worse the loss becomes if the token is removed) and removes
+all the worst p tokens (for instance p could be 10% or 20%). It then repeats the process until the vocabulary has
+reached the desired size, always keeping the base characters (to be able to tokenize any word written with them, like
+BPE or WordPiece).
+
+Contrary to BPE and WordPiece that work out rules in a certain order that you can then apply in the same order when
+tokenizing new text, Unigram will have several ways of tokenizing a new text. For instance, if it ends up with the
+vocabulary
+
+::
+
+    ['b', 'g', 'h', 'n', 'p', 's', 'u', 'ug', 'un', 'hug']
+
+we had before, it could tokenize "hugs" as ``['hug', 's']``, ``['h', 'ug', 's']`` or ``['h', 'u', 'g', 's']``. So which
+one choose? On top of saving the vocabulary, the trained tokenizer will save the probability of each token in the
+training corpus. You can then give a probability to each tokenization (which is the product of the probabilities of the
+tokens forming it) and pick the most likely one (or if you want to apply some data augmentation, you could sample one
+of the tokenization according to their probabilities).
+
+Those probabilities are what are used to define the loss that trains the tokenizer: if our corpus consists of the
+words :math:`x_{1}, \dots, x_{N}` and if for the word :math:`x_{i}` we note :math:`S(x_{i})` the set of all possible
+tokenizations of :math:`x_{i}` (with the current vocabulary), then the loss is defined as
+
+.. math::
+    \mathcal{L} = -\sum_{i=1}^{N} \log \left ( \sum_{x \in S(x_{i})} p(x) \right )
+
+.. _sentencepiece:
+
+SentencePiece
+=============
+
+All the methods we have been looking at so far required some from of pretrokenization, 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.
+
+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>`.

docs/source/torchscript.rst

@@ -12,7 +12,7 @@ According to Pytorch's documentation: "TorchScript is a way to create serializab
 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
+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.
 

docs/source/training.rst

@@ -0,0 +1,323 @@
+Training and fine-tuning
+========================
+
+Model classes in 🤗 Transformers are designed to be compatible with native
+PyTorch and TensorFlow 2 and can be used seemlessly with either. In this
+quickstart, we will show how to fine-tune (or train from scratch) a model
+using the standard training tools available in either framework. We will also
+show how to use our included :func:`~transformers.Trainer` class which
+handles much of the complexity of training for you.
+
+This guide assume that you are already familiar with loading and use our
+models for inference; otherwise, see the :doc:`task summary <task_summary>`. We also assume
+that you are familiar with training deep neural networks in either PyTorch or
+TF2, and focus specifically on the nuances and tools for training models in
+🤗 Transformers.
+
+Sections:
+
+  * :ref:`pytorch`
+  * :ref:`tensorflow`
+  * :ref:`trainer`
+  * :ref:`additional-resources`
+
+.. _pytorch:
+
+Fine-tuning in native PyTorch
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Model classes in 🤗 Transformers that don't begin with ``TF`` are
+`PyTorch Modules <https://pytorch.org/docs/master/generated/torch.nn.Module.html>`_,
+meaning that you can use them just as you would any model in PyTorch for
+both inference and optimization.
+
+Let's consider the common task of fine-tuning a masked language model like
+BERT on a sequence classification dataset. When we instantiate a model with
+:func:`~transformers.PreTrainedModel.from_pretrained`, the model
+configuration and pre-trained weights
+of the specified model are used to initialize the model. The
+library also includes a number of task-specific final layers or 'heads' whose
+weights are instantiated randomly when not present in the specified
+pre-trained model. For example, instantiating a model with
+``BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)``
+will create a BERT model instance with encoder weights copied from the
+``bert-base-uncased`` model and a randomly initialized sequence
+classification head on top of the encoder with an output size of 2. Models
+are initialized in ``eval`` mode by default. We can call ``model.train()`` to
+put it in train mode.
+
+.. code-block:: python
+
+    from transformers import BertForSequenceClassification
+    model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+    model.train()
+
+This is useful because it allows us to make use of the pre-trained BERT
+encoder and easily train it on whatever sequence classification dataset we
+choose. We can use any PyTorch optimizer, but our library also provides the
+:func:`~transformers.AdamW` optimizer which implements gradient bias
+correction as well as weight decay.
+
+.. code-block:: python
+
+    from transformers import AdamW
+    optimizer = AdamW(model.parameters(), lr=1e-5)
+
+The optimizer allows us to apply different hyperpameters for specific
+parameter groups. For example, we can apply weight decay to all parameters
+other than bias and layer normalization terms:
+
+.. code-block:: python
+
+    no_decay = ['bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
+        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+    ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=1e-5)
+    
+Now we can set up a simple dummy training batch using
+:func:`~transformers.PreTrainedTokenizer.__call__`. This returns a
+:func:`~transformers.BatchEncoding` instance which
+prepares everything we might need to pass to the model.
+
+.. code-block:: python
+
+    from transformers import BertTokenizer
+    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+    text_batch = ["I love Pixar.", "I don't care for Pixar."]
+    encoding = tokenizer(text_batch, return_tensors='pt', padding=True, truncation=True)
+    input_ids = encoding['input_ids']
+    attention_mask = encoding['attention_mask']
+
+When we call a classification model with the ``labels`` argument, the first
+returned element is the Cross Entropy loss between the predictions and the
+passed labels. Having already set up our optimizer, we can then do a
+backwards pass and update the weights:
+
+.. code-block:: python
+
+    labels = torch.tensor([1,0]).unsqueeze(0)
+    outputs = model(input_ids, attention_mask=attention_mask, labels=labels)
+    loss = outputs[0]
+    loss.backward()
+    optimizer.step()
+
+Alternatively, you can just get the logits and calculate the loss yourself.
+The following is equivalent to the previous example:
+
+.. code-block:: python
+
+    from torch.nn import functional as F
+    labels = torch.tensor([1,0]).unsqueeze(0)
+    outputs = model(input_ids, attention_mask=attention_mask)
+    loss = F.cross_entropy(labels, outputs[0])
+    loss.backward()
+    optimizer.step()
+
+Of course, you can train on GPU by calling ``to('cuda')`` on the model and
+inputs as usual.
+
+We also provide a few learning rate scheduling tools. With the following, we
+can set up a scheduler which warms up for ``num_warmup_steps`` and then
+linearly decays to 0 by the end of training.
+
+.. code-block:: python
+
+    from transformers import get_linear_schedule_with_warmup
+    scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps, num_train_steps)
+
+Then all we have to do is call ``scheduler.step()`` after ``optimizer.step()``.
+
+.. code-block:: python
+
+    ...
+    loss.backward()
+    optimizer.step()
+    scheduler.step()
+
+We highly recommend using :func:`~transformers.Trainer`, discussed below,
+which conveniently handles the moving parts of training 🤗 Transformers models
+with features like mixed precision and easy tensorboard logging.
+
+
+Freezing the encoder
+--------------------
+
+In some cases, you might be interested in keeping the weights of the
+pre-trained encoder frozen and optimizing only the weights of the head
+layers. To do so, simply set the ``requires_grad`` attribute to ``False`` on
+the encoder parameters, which can be accessed with the ``base_model``
+submodule on any task-specific model in the library:
+
+.. code-block:: python
+   
+    for param in model.base_model.parameters():
+        param.requires_grad = False
+
+
+.. _tensorflow:
+
+Fine-tuning in native TensorFlow 2
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Models can also be trained natively in TensorFlow 2. Just as with PyTorch,
+TensorFlow models can be instantiated with
+:func:`~transformers.PreTrainedModel.from_pretrained` to load the weights of
+the encoder from a pretrained model.
+
+.. code-block:: python
+
+    from transformers import TFBertForSequenceClassification
+    model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
+
+Let's use ``tensorflow_datasets`` to load in the `MRPC dataset
+<https://www.tensorflow.org/datasets/catalog/glue#gluemrpc>`_ from GLUE. We
+can then use our built-in
+:func:`~transformers.data.processors.glue.glue_convert_examples_to_features`
+to tokenize MRPC and convert it to a TensorFlow ``Dataset`` object. Note that
+tokenizers are framework-agnostic, so there is no need to prepend ``TF`` to
+the pretrained tokenizer name.
+
+.. code-block:: python
+
+    from transformers import BertTokenizer, glue_convert_examples_to_features
+    import tensorflow_datasets as tfds
+    tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+    data = tfds.load('glue/mrpc')
+    train_dataset = glue_convert_examples_to_features(data['train'], tokenizer, max_length=128, task='mrpc')
+    train_dataset = train_dataset.shuffle(100).batch(32).repeat(2)
+
+The model can then be compiled and trained as any Keras model:
+
+.. code-block:: python
+    
+    optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5)
+    loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+    model.compile(optimizer=optimizer, loss=loss)
+    model.fit(train_dataset, epochs=2, steps_per_epoch=115)
+
+With the tight interoperability between TensorFlow and PyTorch models, you
+can even save the model and then reload it as a PyTorch model (or vice-versa):
+
+.. code-block:: python
+
+    from transformers import BertForSequenceClassification
+    model.save_pretrained('./my_mrpc_model/')
+    pytorch_model = BertForSequenceClassification.from_pretrained('./my_mrpc_model/', from_tf=True)
+
+
+.. _trainer:
+
+Trainer
+^^^^^^^
+
+We also provide a simple but feature-complete training and evaluation
+interface through :func:`~transformers.Trainer` and
+:func:`~transformers.TFTrainer`. You can train, fine-tune,
+and evaluate any 🤗 Transformers model with a wide range of training options and
+with built-in features like logging, gradient accumulation, and mixed
+precision.
+
+.. code-block:: python
+
+    ## PYTORCH CODE
+    from transformers import BertForSequenceClassification, Trainer, TrainingArguments
+
+    model = BertForSequenceClassification.from_pretrained("bert-large-uncased")
+
+    training_args = TrainingArguments(
+        output_dir='./results',          # output directory
+        num_train_epochs=3,              # total # of training epochs
+        per_device_train_batch_size=16,  # batch size per device during training
+        per_device_eval_batch_size=64,   # batch size for evaluation
+        warmup_steps=500,                # number of warmup steps for learning rate scheduler
+        weight_decay=0.01,               # strength of weight decay
+        logging_dir='./logs',            # directory for storing logs
+    )
+
+    trainer = Trainer(
+        model=model,                         # the instantiated 🤗 Transformers model to be trained
+        args=training_args,                  # training arguments, defined above
+        train_dataset=train_dataset,         # training dataset
+        eval_dataset=test_dataset            # evaluation dataset
+    )
+    ## TENSORFLOW CODE
+    from transformers import TFBertForSequenceClassification, TFTrainer, TFTrainingArguments
+
+    model = TFBertForSequenceClassification.from_pretrained("bert-large-uncased")
+
+    training_args = TFTrainingArguments(
+        output_dir='./results',          # output directory
+        num_train_epochs=3,              # total # of training epochs
+        per_device_train_batch_size=16,  # batch size per device during training
+        per_device_eval_batch_size=64,   # batch size for evaluation
+        warmup_steps=500,                # number of warmup steps for learning rate scheduler
+        weight_decay=0.01,               # strength of weight decay
+        logging_dir='./logs',            # directory for storing logs
+    )
+
+    trainer = TFTrainer(
+        model=model,                         # the instantiated 🤗 Transformers model to be trained
+        args=training_args,                  # training arguments, defined above
+        train_dataset=tfds_train_dataset,    # tensorflow_datasets training dataset
+        eval_dataset=tfds_test_dataset       # tensorflow_datasets evaluation dataset
+    )
+
+Now simply call ``trainer.train()`` to train and ``trainer.evaluate()`` to
+evaluate. You can use your own module as well, but the first
+argument returned from ``forward`` must be the loss which you wish to
+optimize.
+
+:func:`~transformers.Trainer` uses a built-in default function to collate
+batches and prepare them to be fed into the model. If needed, you can also
+use the ``data_collator`` argument to pass your own collator function which
+takes in the data in the format provided by your dataset and returns a
+batch ready to be fed into the model. Note that
+:func:`~transformers.TFTrainer` expects the passed datasets to be dataset
+objects from ``tensorflow_datasets``.
+
+To calculate additional metrics in addition to the loss, you can also define
+your own ``compute_metrics`` function and pass it to the trainer.
+
+.. code-block:: python
+
+    from sklearn.metrics import precision_recall_fscore_support
+
+    def compute_metrics(pred):
+        labels = pred.label_ids
+        preds = pred.predictions.argmax(-1)
+        precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')
+        acc = accuracy_score(labels, preds)
+        return {
+            'accuracy': acc,
+            'f1': f1,
+            'precision': precision,
+            'recall': recall
+        }
+
+Finally, you can view the results, including any calculated metrics, by
+launching tensorboard in your specified ``logging_dir`` directory.
+
+
+.. _additional-resources:
+
+Additional resources
+^^^^^^^^^^^^^^^^^^^^
+
+    * `A lightweight colab demo
+      <https://colab.research.google.com/drive/1-JIJlao4dI-Ilww_NnTc0rxtp-ymgDgM?usp=sharing>`_
+      which uses ``Trainer`` for IMDb sentiment classification.
+
+    * `🤗 Transformers Examples <https://github.com/huggingface/transformers/tree/master/examples>`_
+      including scripts for training and fine-tuning on GLUE, SQuAD, and
+      several other tasks.
+
+    * `How to train a language model
+      <https://colab.research.google.com/github/huggingface/blog/blob/master/notebooks/01_how_to_train.ipynb>`_,
+      a detailed colab notebook which uses ``Trainer`` to train a masked
+      language model from scratch on Esperanto.
+
+    * `🤗 Transformers Notebooks <./notebooks.html>`_ which contain dozens
+      of example notebooks from the community for training and using
+      🤗 Transformers on a variety of tasks.

examples/README.md

@@ -1,7 +1,7 @@
-## Examples
+# Examples
 
-Version 2.9 of `transformers` introduces a new [`Trainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer.py) class for PyTorch, and its equivalent [`TFTrainer`](https://github.com/huggingface/transformers/blob/master/src/transformers/trainer_tf.py) for TF 2.
-Running the examples requires PyTorch 1.3.1+ or TensorFlow 2.0+.
+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+.
 
 Here is the list of all our examples:
 - **grouped by task** (all official examples work for multiple models)
@@ -13,7 +13,7 @@ Here is the list of all our examples:
 This is still a work-in-progress – in particular documentation is still sparse – so please **contribute improvements/pull requests.**
 
 
-# The Big Table of Tasks
+## The Big Table of Tasks
 
 | Task | Example datasets | Trainer support | TFTrainer support | pytorch-lightning | Colab
 |---|---|:---:|:---:|:---:|:---:|
@@ -24,8 +24,8 @@ This is still a work-in-progress – in particular documentation is still sparse
 | [**`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/summarization)     | CNN/Daily Mail    | -  | -  | -  | -
-| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/translation)         | WMT               | -  | -  | -  | -
+| [**`summarization`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)     | CNN/Daily Mail    | -  | -  | ✅  | -
+| [**`translation`**](https://github.com/huggingface/transformers/tree/master/examples/seq2seq)         | WMT               | -  | -  | ✅  | -
 | [**`bertology`**](https://github.com/huggingface/transformers/tree/master/examples/bertology)             | -                 | -  | -  | -  | -
 | [**`adversarial`**](https://github.com/huggingface/transformers/tree/master/examples/adversarial)         | HANS              | ✅ | -  | -  | -
 

examples/adversarial/run_hans.py

@@ -33,7 +33,7 @@ from transformers import (
     default_data_collator,
     set_seed,
 )
-from utils_hans import HansDataset, InputFeatures, hans_processors
+from utils_hans import HansDataset, InputFeatures, hans_processors, hans_tasks_num_labels
 
 
 logger = logging.getLogger(__name__)
@@ -130,9 +130,7 @@ def main():
     set_seed(training_args.seed)
 
     try:
-        processor = hans_processors[data_args.task_name]()
-        label_list = processor.get_labels()
-        num_labels = len(label_list)
+        num_labels = hans_tasks_num_labels[data_args.task_name]
     except KeyError:
         raise ValueError("Task not found: %s" % (data_args.task_name))
 
@@ -214,6 +212,7 @@ def main():
 
         pair_ids = [ex.pairID for ex in eval_dataset]
         output_eval_file = os.path.join(training_args.output_dir, "hans_predictions.txt")
+        label_list = eval_dataset.get_labels()
         if trainer.is_world_master():
             with open(output_eval_file, "w") as writer:
                 writer.write("pairID,gold_label\n")

examples/adversarial/utils_hans.py

@@ -22,7 +22,17 @@ from typing import List, Optional, Union
 import tqdm
 from filelock import FileLock
 
-from transformers import DataProcessor, PreTrainedTokenizer, is_tf_available, is_torch_available
+from transformers import (
+    BartTokenizer,
+    BartTokenizerFast,
+    DataProcessor,
+    PreTrainedTokenizer,
+    RobertaTokenizer,
+    RobertaTokenizerFast,
+    XLMRobertaTokenizer,
+    is_tf_available,
+    is_torch_available,
+)
 
 
 logger = logging.getLogger(__name__)
@@ -105,6 +115,17 @@ if is_torch_available():
                     "dev" if evaluate else "train", tokenizer.__class__.__name__, str(max_seq_length), task,
                 ),
             )
+            label_list = processor.get_labels()
+            if tokenizer.__class__ in (
+                RobertaTokenizer,
+                RobertaTokenizerFast,
+                XLMRobertaTokenizer,
+                BartTokenizer,
+                BartTokenizerFast,
+            ):
+                # HACK(label indices are swapped in RoBERTa pretrained model)
+                label_list[1], label_list[2] = label_list[2], label_list[1]
+            self.label_list = label_list
 
             # Make sure only the first process in distributed training processes the dataset,
             # and the others will use the cache.
@@ -116,7 +137,6 @@ if is_torch_available():
                     self.features = torch.load(cached_features_file)
                 else:
                     logger.info(f"Creating features from dataset file at {data_dir}")
-                    label_list = processor.get_labels()
 
                     examples = (
                         processor.get_dev_examples(data_dir) if evaluate else processor.get_train_examples(data_dir)
@@ -133,6 +153,9 @@ if is_torch_available():
         def __getitem__(self, i) -> InputFeatures:
             return self.features[i]
 
+        def get_labels(self):
+            return self.label_list
+
 
 if is_tf_available():
     import tensorflow as tf
@@ -156,6 +179,16 @@ if is_tf_available():
         ):
             processor = hans_processors[task]()
             label_list = processor.get_labels()
+            if tokenizer.__class__ in (
+                RobertaTokenizer,
+                RobertaTokenizerFast,
+                XLMRobertaTokenizer,
+                BartTokenizer,
+                BartTokenizerFast,
+            ):
+                # HACK(label indices are swapped in RoBERTa pretrained model)
+                label_list[1], label_list[2] = label_list[2], label_list[1]
+            self.label_list = label_list
 
             examples = processor.get_dev_examples(data_dir) if evaluate else processor.get_train_examples(data_dir)
             self.features = hans_convert_examples_to_features(examples, label_list, max_seq_length, tokenizer)
@@ -206,6 +239,9 @@ if is_tf_available():
         def __getitem__(self, i) -> InputFeatures:
             return self.features[i]
 
+        def get_labels(self):
+            return self.label_list
+
 
 class HansProcessor(DataProcessor):
     """Processor for the HANS data set."""
@@ -262,12 +298,13 @@ def hans_convert_examples_to_features(
         if ex_index % 10000 == 0:
             logger.info("Writing example %d" % (ex_index))
 
-        inputs = tokenizer.encode_plus(
+        inputs = tokenizer(
             example.text_a,
             example.text_b,
             add_special_tokens=True,
             max_length=max_length,
-            pad_to_max_length=True,
+            padding="max_length",
+            truncation=True,
             return_overflowing_tokens=True,
         )
 

examples/benchmarking/plot_csv_file.py

@@ -1,14 +1,19 @@
 import csv
 from collections import defaultdict
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import List, Optional
 
 import matplotlib.pyplot as plt
 import numpy as np
+from matplotlib.ticker import ScalarFormatter
 
 from transformers import HfArgumentParser
 
 
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
 @dataclass
 class PlotArguments:
     """
@@ -24,6 +29,9 @@ class PlotArguments:
         default=False,
         metadata={"help": "Whether the csv file has time results or memory results. Defaults to memory results."},
     )
+    no_log_scale: bool = field(
+        default=False, metadata={"help": "Disable logarithmic scale when plotting"},
+    )
     is_train: bool = field(
         default=False,
         metadata={
@@ -33,6 +41,25 @@ class PlotArguments:
     figure_png_file: Optional[str] = field(
         default=None, metadata={"help": "Filename under which the plot will be saved. If unused no plot is saved."},
     )
+    short_model_names: Optional[List[str]] = list_field(
+        default=None, metadata={"help": "List of model names that are used instead of the ones in the csv file."}
+    )
+
+
+def can_convert_to_int(string):
+    try:
+        int(string)
+        return True
+    except ValueError:
+        return False
+
+
+def can_convert_to_float(string):
+    try:
+        float(string)
+        return True
+    except ValueError:
+        return False
 
 
 class Plot:
@@ -46,16 +73,31 @@ class Plot:
                 model_name = row["model"]
                 self.result_dict[model_name]["bsz"].append(int(row["batch_size"]))
                 self.result_dict[model_name]["seq_len"].append(int(row["sequence_length"]))
-                self.result_dict[model_name]["result"][(int(row["batch_size"]), int(row["sequence_length"]))] = row[
-                    "result"
-                ]
+                if can_convert_to_int(row["result"]):
+                    # value is not None
+                    self.result_dict[model_name]["result"][
+                        (int(row["batch_size"]), int(row["sequence_length"]))
+                    ] = int(row["result"])
+                elif can_convert_to_float(row["result"]):
+                    # value is not None
+                    self.result_dict[model_name]["result"][
+                        (int(row["batch_size"]), int(row["sequence_length"]))
+                    ] = float(row["result"])
 
     def plot(self):
         fig, ax = plt.subplots()
         title_str = "Time usage" if self.args.is_time else "Memory usage"
         title_str = title_str + " for training" if self.args.is_train else title_str + " for inference"
 
-        for model_name in self.result_dict.keys():
+        if not self.args.no_log_scale:
+            # set logarithm scales
+            ax.set_xscale("log")
+            ax.set_yscale("log")
+
+        for axis in [ax.xaxis, ax.yaxis]:
+            axis.set_major_formatter(ScalarFormatter())
+
+        for model_name_idx, model_name in enumerate(self.result_dict.keys()):
             batch_sizes = sorted(list(set(self.result_dict[model_name]["bsz"])))
             sequence_lengths = sorted(list(set(self.result_dict[model_name]["seq_len"])))
             results = self.result_dict[model_name]["result"]
@@ -64,28 +106,33 @@ class Plot:
                 (batch_sizes, sequence_lengths) if self.args.plot_along_batch else (sequence_lengths, batch_sizes)
             )
 
-            plt.xlim(min(x_axis_array), max(x_axis_array))
+            label_model_name = (
+                model_name if self.args.short_model_names is None else self.args.short_model_names[model_name_idx]
+            )
 
             for inner_loop_value in inner_loop_array:
                 if self.args.plot_along_batch:
-                    y_axis_array = np.asarray([results[(x, inner_loop_value)] for x in x_axis_array], dtype=np.int)
+                    y_axis_array = np.asarray(
+                        [results[(x, inner_loop_value)] for x in x_axis_array if (x, inner_loop_value) in results],
+                        dtype=np.int,
+                    )
                 else:
-                    y_axis_array = np.asarray([results[(inner_loop_value, x)] for x in x_axis_array], dtype=np.float32)
-
-                ax.set_xscale("log", basex=2)
-                ax.set_yscale("log", basey=10)
+                    y_axis_array = np.asarray(
+                        [results[(inner_loop_value, x)] for x in x_axis_array if (inner_loop_value, x) in results],
+                        dtype=np.float32,
+                    )
 
                 (x_axis_label, inner_loop_label) = (
-                    ("batch_size", "sequence_length in #tokens")
-                    if self.args.plot_along_batch
-                    else ("sequence_length in #tokens", "batch_size")
+                    ("batch_size", "len") if self.args.plot_along_batch else ("in #tokens", "bsz")
                 )
 
-                x_axis_array = np.asarray(x_axis_array, np.int)
-                plt.scatter(x_axis_array, y_axis_array, label=f"{model_name} - {inner_loop_label}: {inner_loop_value}")
+                x_axis_array = np.asarray(x_axis_array, np.int)[: len(y_axis_array)]
+                plt.scatter(
+                    x_axis_array, y_axis_array, label=f"{label_model_name} - {inner_loop_label}: {inner_loop_value}"
+                )
                 plt.plot(x_axis_array, y_axis_array, "--")
 
-            title_str += f" {model_name} vs."
+            title_str += f" {label_model_name} vs."
 
         title_str = title_str[:-4]
         y_axis_label = "Time in s" if self.args.is_time else "Memory in MB"

examples/benchmarking/run_benchmark_tf.py

@@ -13,15 +13,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-""" Benchmarking the library on inference and training in Tensorflow"""
+""" Benchmarking the library on inference and training in TensorFlow"""
 
-from transformers import HfArgumentParser, TensorflowBenchmark, TensorflowBenchmarkArguments
+from transformers import HfArgumentParser, TensorFlowBenchmark, TensorFlowBenchmarkArguments
 
 
 def main():
-    parser = HfArgumentParser(TensorflowBenchmarkArguments)
+    parser = HfArgumentParser(TensorFlowBenchmarkArguments)
     benchmark_args = parser.parse_args_into_dataclasses()[0]
-    benchmark = TensorflowBenchmark(args=benchmark_args)
+    benchmark = TensorFlowBenchmark(args=benchmark_args)
     benchmark.run()
 
 

examples/bert-loses-patience/run_glue_with_pabee.py

@@ -226,8 +226,6 @@ def train(args, train_dataset, model, tokenizer):
                 if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
                     # Save model checkpoint
                     output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
                     model_to_save = (
                         model.module if hasattr(model, "module") else model
                     )  # Take care of distributed/parallel training
@@ -254,12 +252,15 @@ def train(args, train_dataset, model, tokenizer):
     return global_step, tr_loss / global_step
 
 
-def evaluate(args, model, tokenizer, prefix=""):
+def evaluate(args, model, tokenizer, prefix="", patience=0):
 
-    # PABEE STATS
     if args.model_type == "albert":
+        model.albert.set_regression_threshold(args.regression_threshold)
+        model.albert.set_patience(patience)
         model.albert.reset_stats()
     elif args.model_type == "bert":
+        model.bert.set_regression_threshold(args.regression_threshold)
+        model.bert.set_patience(patience)
         model.bert.reset_stats()
     else:
         raise NotImplementedError()
@@ -331,7 +332,7 @@ def evaluate(args, model, tokenizer, prefix=""):
                 print("  %s = %s" % (key, str(result[key])))
                 writer.write("%s = %s\n" % (key, str(result[key])))
 
-    if args.eval_all_checkpoints:
+    if args.eval_all_checkpoints and patience != 0:
         if args.model_type == "albert":
             model.albert.log_stats()
         elif args.model_type == "bert":
@@ -646,10 +647,6 @@ def main():
 
     # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`
@@ -690,15 +687,7 @@ def main():
 
             print(f"Evaluation for checkpoint {prefix}")
             for patience in patience_list:
-                if args.model_type == "albert":
-                    model.albert.set_regression_threshold(args.regression_threshold)
-                    model.albert.set_patience(patience)
-                elif args.model_type == "bert":
-                    model.bert.set_regression_threshold(args.regression_threshold)
-                    model.bert.set_patience(patience)
-                else:
-                    raise NotImplementedError()
-                result = evaluate(args, model, tokenizer, prefix=prefix)
+                result = evaluate(args, model, tokenizer, prefix=prefix, patience=patience)
                 result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
                 results.update(result)
     return results

examples/contrib/mm-imdb/run_mmimdb.py

@@ -521,10 +521,6 @@ def main():
 
     # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`

examples/contrib/run_swag.py

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

examples/distillation/run_squad_w_distillation.py

@@ -809,10 +809,6 @@ def main():
 
     # Save the trained model and the tokenizer
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`

examples/language-modeling/run_language_modeling.py

@@ -14,9 +14,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, BERT, RoBERTa).
-GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
-using a masked language modeling (MLM) loss.
+Fine-tuning the library models for language modeling on a text file (GPT, GPT-2, CTRL, BERT, RoBERTa, XLNet).
+GPT, GPT-2 and CTRL are fine-tuned using a causal language modeling (CLM) loss. BERT and RoBERTa are fine-tuned
+using a masked language modeling (MLM) loss. XLNet is fine-tuned using a permutation language modeling (PLM) loss.
 """
 
 
@@ -33,6 +33,7 @@ from transformers import (
     AutoModelWithLMHead,
     AutoTokenizer,
     DataCollatorForLanguageModeling,
+    DataCollatorForPermutationLanguageModeling,
     HfArgumentParser,
     LineByLineTextDataset,
     PreTrainedTokenizer,
@@ -101,6 +102,15 @@ class DataTrainingArguments:
     mlm_probability: float = field(
         default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
     )
+    plm_probability: float = field(
+        default=1 / 6,
+        metadata={
+            "help": "Ratio of length of a span of masked tokens to surrounding context length for permutation language modeling."
+        },
+    )
+    max_span_length: int = field(
+        default=5, metadata={"help": "Maximum length of a span of masked tokens for permutation language modeling."}
+    )
 
     block_size: int = field(
         default=-1,
@@ -207,8 +217,8 @@ def main():
 
     if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
         raise ValueError(
-            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
-            "flag (masked language modeling)."
+            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
+            "--mlm flag (masked language modeling)."
         )
 
     if data_args.block_size <= 0:
@@ -221,9 +231,14 @@ def main():
 
     train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
     eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
-    data_collator = DataCollatorForLanguageModeling(
-        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
-    )
+    if config.model_type == "xlnet":
+        data_collator = DataCollatorForPermutationLanguageModeling(
+            tokenizer=tokenizer, plm_probability=data_args.plm_probability, max_span_length=data_args.max_span_length,
+        )
+    else:
+        data_collator = DataCollatorForLanguageModeling(
+            tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
+        )
 
     # Initialize our Trainer
     trainer = Trainer(

examples/lightning_base.py

@@ -116,6 +116,16 @@ class BaseTransformer(pl.LightningModule):
         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)
+
     def test_step(self, batch, batch_nb):
         return self.validation_step(batch, batch_nb)
 
@@ -189,7 +199,7 @@ class BaseTransformer(pl.LightningModule):
         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 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=500, type=int, help="Linear warmup over warmup_steps.")
+        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."

examples/longform-qa/eli5_app.py

@@ -1,10 +1,10 @@
 import faiss
 import nlp
 import numpy as np
+import streamlit as st
 import torch
 from elasticsearch import Elasticsearch
 
-import streamlit as st
 import transformers
 from eli5_utils import (
     embed_questions_for_retrieval,

examples/longform-qa/eli5_utils.py

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

examples/movement-pruning/emmental/modeling_bert_masked.py

@@ -424,7 +424,7 @@ MASKED_BERT_INPUTS_DOCSTRING = r"""
 
             Indices can be obtained using :class:`transformers.BertTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
-            :func:`transformers.PreTrainedTokenizer.encode_plus` for details.
+            :func:`transformers.PreTrainedTokenizer.__call__` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):

examples/movement-pruning/masked_run_glue.py

@@ -875,10 +875,6 @@ def main():
 
     # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`

examples/movement-pruning/masked_run_squad.py

@@ -1059,10 +1059,6 @@ def main():
 
     # Save the trained model and the tokenizer
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`

examples/multiple-choice/run_tf_multiple_choice.py

@@ -108,7 +108,10 @@ def main():
         level=logging.INFO,
     )
     logger.warning(
-        "device: %s, n_gpu: %s, 16-bits training: %s", training_args.device, training_args.n_gpu, training_args.fp16,
+        "device: %s, n_replicas: %s, 16-bits training: %s",
+        training_args.device,
+        training_args.n_replicas,
+        training_args.fp16,
     )
     logger.info("Training/evaluation parameters %s", training_args)
 

examples/multiple-choice/utils_multiple_choice.py

@@ -510,12 +510,13 @@ def convert_examples_to_features(
             else:
                 text_b = example.question + " " + ending
 
-            inputs = tokenizer.encode_plus(
+            inputs = tokenizer(
                 text_a,
                 text_b,
                 add_special_tokens=True,
                 max_length=max_length,
-                pad_to_max_length=True,
+                padding="max_length",
+                truncation=True,
                 return_overflowing_tokens=True,
             )
             if "num_truncated_tokens" in inputs and inputs["num_truncated_tokens"] > 0:

examples/question-answering/README.md

@@ -77,7 +77,7 @@ exact_match = 86.91
 ```
 
 This fine-tuned model is available as a checkpoint under the reference
-`bert-large-uncased-whole-word-masking-finetuned-squad`.
+[`bert-large-uncased-whole-word-masking-finetuned-squad`](https://huggingface.co/bert-large-uncased-whole-word-masking-finetuned-squad).
 
 #### Fine-tuning XLNet on SQuAD
 
@@ -176,4 +176,5 @@ python run_tf_squad.py \
     --doc_stride 128    
 ```
 
-For the moment the evaluation is not available in the Tensorflow Trainer only the training.
+
+For the moment evaluation is not available in the Tensorflow Trainer only the training.

examples/question-answering/run_squad.py

@@ -13,814 +13,147 @@
 # 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.
-""" Finetuning the library models for question-answering on SQuAD (DistilBERT, Bert, XLM, XLNet)."""
+""" Fine-tuning the library models for question-answering."""
 
 
-import argparse
-import glob
 import logging
 import os
-import random
-import timeit
+import sys
+from dataclasses import dataclass, field
+from typing import Optional
 
-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
+from transformers import AutoConfig, AutoModelForQuestionAnswering, AutoTokenizer, HfArgumentParser, SquadDataset
+from transformers import SquadDataTrainingArguments as DataTrainingArguments
+from transformers import Trainer, TrainingArguments
 
 
 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.
+    """
 
-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
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
     )
-
-    # 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),
+    config_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained config 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]
+    tokenizer_name: Optional[str] = field(
+        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
     )
-    # 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))
-                    if not os.path.exists(output_dir):
-                        os.makedirs(output_dir)
-                    # 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),
-        ),
+    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"}
     )
 
-    # 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():
-    parser = argparse.ArgumentParser()
+    # 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.
 
-    # 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.",
-    )
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
 
-    # 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 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(args.output_dir)
-        and os.listdir(args.output_dir)
-        and args.do_train
-        and not args.overwrite_output_dir
+        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(
-            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
-                args.output_dir
-            )
+            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
         )
 
-    # 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 args.local_rank in [-1, 0] else logging.WARN,
+        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",
-        args.local_rank,
-        device,
-        args.n_gpu,
-        bool(args.local_rank != -1),
-        args.fp16,
+        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)
 
-    # Set seed
-    set_seed(args)
-
+    # Prepare Question-Answering task
     # Load pretrained model and tokenizer
-    if args.local_rank not in [-1, 0]:
-        # Make sure only the first process in distributed training will download model & vocab
-        torch.distributed.barrier()
+    #
+    # Distributed training:
+    # The .from_pretrained methods guarantee that only one local process can concurrently
+    # download model & vocab.
 
-    args.model_type = args.model_type.lower()
     config = AutoConfig.from_pretrained(
-        args.config_name if args.config_name else args.model_name_or_path,
-        cache_dir=args.cache_dir if args.cache_dir else None,
+        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(
-        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_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(
-        args.model_name_or_path,
-        from_tf=bool(".ckpt" in args.model_name_or_path),
+        model_args.model_name_or_path,
+        from_tf=bool(".ckpt" in model_args.model_name_or_path),
         config=config,
-        cache_dir=args.cache_dir if args.cache_dir else None,
+        cache_dir=model_args.cache_dir,
     )
 
-    if args.local_rank == 0:
-        # Make sure only the first process in distributed training will download model & vocab
-        torch.distributed.barrier()
+    # 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
+    )
 
-    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.")
+    # Initialize our Trainer
+    trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset, eval_dataset=eval_dataset,)
 
     # Training
-    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)
+    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)
 
-    # Save the trained model and the tokenizer
-    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
 
-        logger.info("Saving model checkpoint to %s", args.output_dir)
-        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
-        # They can then be reloaded using `from_pretrained()`
-        # 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)
-
-        # 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
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
 
 
 if __name__ == "__main__":

examples/question-answering/run_tf_squad.py

@@ -137,9 +137,9 @@ def main():
         level=logging.INFO,
     )
     logger.info(
-        "n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        training_args.n_gpu,
-        bool(training_args.n_gpu > 1),
+        "n_replicas: %s, distributed training: %s, 16-bits training: %s",
+        training_args.n_replicas,
+        bool(training_args.n_replicas > 1),
         training_args.fp16,
     )
     logger.info("Training/evaluation parameters %s", training_args)

examples/seq2seq/README.md

@@ -0,0 +1,236 @@
+## Sequence to Sequence
+
+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`. 
+
+
+### 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=${PWD}/cnn_dm
+```
+
+this should make a directory called cnn_dm/ with files like `test.source`.
+To use your own data, copy that files format. Each article to be summarized is on its own line.
+
+XSUM Data:
+```bash
+cd examples/seq2seq
+wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/xsum.tar.gz
+tar -xzvf xsum.tar.gz
+export XSUM_DIR=${PWD}/xsum
+```
+
+
+WMT16 English-Romanian Translation Data:
+```bash
+cd examples/seq2seq
+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
+```
+
+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.   
+- 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: 
+    > "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! 
+
+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.
+- 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). 
+
+### Summarization Finetuning
+Run/modify `finetune.sh`
+
+The following command should work on a 16GB GPU:
+```bash
+./finetune.sh \
+    --data_dir $XSUM_DIR \
+    --train_batch_size=1 \
+    --eval_batch_size=1 \
+    --output_dir=xsum_results \
+    --num_train_epochs 1 \
+    --model_name_or_path facebook/bart-large
+```
+
+*Note*: The following tips mostly apply to summarization finetuning.
+
+### Translation Finetuning
+
+First, follow the wmt_en_ro download instructions.
+Then you can finetune mbart_cc25 on english-romanian with the following command.
+**Recommendation:** Read and potentially modify the fairly opinionated defaults in `train_mbart_cc25_enro.sh` script before running it. 
+```bash
+export ENRO_DIR=${PWD}/wmt_en_ro   # may need to be fixed depending on where you downloaded
+export BS=4
+export GAS=8
+./train_mbart_cc25_enro.sh --output_dir cc25_v1_frozen/
+```
+
+
+### 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
+output_dir
+├── best_tfmr  # this is a huggingface checkpoint generated by save_pretrained. It is the same model as the PL .ckpt file below
+│   ├── config.json
+│   ├── merges.txt
+│   ├── pytorch_model.bin
+│   ├── special_tokens_map.json
+│   ├── 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.
+├── 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_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.
+```
+After training, you can recover the best checkpoint by running
+```python
+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.
+If 'translation' is in your task name, the computed metric will be BLEU. Otherwise, ROUGE will be used.
+
+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 \
+    $DATA_DIR/val.source t5_val_generations.txt \
+    --reference_path $DATA_DIR/val.target \
+    --score_path enro_bleu.json \
+    --task translation_en_to_ro \
+    --n_obs 100 \
+    --device cuda \
+    --fp16 \
+    --bs 32
+```
+
+This command works for MBART, although the BLEU score is suspiciously low.
+```bash
+export DATA_DIR=wmt_en_ro
+python run_eval.py facebook/mbart-large-en-ro $DATA_DIR/val.source mbart_val_generations.txt \
+    --reference_path $DATA_DIR/val.target \
+    --score_path enro_bleu.json \
+    --task translation \
+    --n_obs 100 \
+    --device cuda \
+    --fp16 \
+    --bs 32
+```
+
+Summarization (xsum will be very similar):
+```bash
+export DATA_DIR=cnn_dm
+python run_eval.py sshleifer/distilbart-cnn-12-6 $DATA_DIR/val.source dbart_val_generations.txt \
+    --reference_path $DATA_DIR/val.target \
+    --score_path cnn_rouge.json \
+    --task summarization \
+    --n_obs 100 \
+    --device cuda \
+    --fp16 \
+    --bs 32
+```
+
+
+### DistilBART
+![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. 
+
+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.
+
+You can see the performance tradeoffs of model sizes [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=0).
+and more granular timing results [here](https://docs.google.com/spreadsheets/d/1EkhDMwVO02m8jCD1cG3RoFPLicpcL1GQHTQjfvDYgIM/edit#gid=1753259047&range=B2:I23).
+
+#### No Teacher Distillation
+To run the simpler distilbart-cnn style distillation all you need is data, a GPU, and a properly initialized student.
+You don't even need `distillation.py`.
+
+Some [un-finetuned students](https://huggingface.co/models?search=sshleifer%2Fstudent) are available for replication purposes.
+They are initialized by copying layers from the associated `bart-large-{cnn|xsum}` teacher using `--init_strategy alternate`. (You can read about that in `initialization_utils.py`)
+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`.
+If you are using `wandb` and comparing the two distillation methods, using this entry point will make your logs consistent,
+because you will have the same hyperparameters logged in every run.
+
+#### With a teacher
+*Note* only BART variants are supported
+
+In this method, we use try to enforce that the student and teacher produce similar encoder_outputs, logits, and hidden_states using `BartSummarizationDistiller`.
+This is how `sshleifer/distilbart-xsum*` checkpoints were produced.
+
+The command that produced `sshleifer/distilbart-xsum-12-6` is:
+
+```bash
+./train_distilbart_xsum.sh  
+```
+
+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/  
+```

examples/seq2seq/bertabs/README.md

@@ -12,7 +12,7 @@ The model is loaded with the pre-trained weights for the abstractive summarizati
 git clone https://github.com/huggingface/transformers && cd transformers
 pip install .
 pip install nltk py-rouge
-cd examples/summarization
+cd examples/seq2seq/bertabs
 ```
 
 ## Reproduce the authors'  ROUGE score

examples/seq2seq/bertabs/run_summarization.py

@@ -30,7 +30,7 @@ Batch = namedtuple("Batch", ["document_names", "batch_size", "src", "segs", "mas
 
 def evaluate(args):
     tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", do_lower_case=True)
-    model = BertAbs.from_pretrained("bertabs-finetuned-cnndm")
+    model = BertAbs.from_pretrained("remi/bertabs-finetuned-extractive-abstractive-summarization")
     model.to(args.device)
     model.eval()
 

examples/seq2seq/callbacks.py

@@ -32,9 +32,12 @@ class Seq2SeqLoggingCallback(pl.Callback):
             results_file = od / "test_results.txt"
             generations_file = od / "test_generations.txt"
         else:
-            results_file = od / f"{type_path}_results_{trainer.global_step:05d}.txt"
-            generations_file = od / f"{type_path}_generations_{trainer.global_step:05d}.txt"
-
+            # this never gets hit. I prefer not to save intermediate generations, and results are in metrics.json
+            # If people want this it will be easy enough to add back.
+            results_file = od / f"{type_path}_results/{trainer.global_step:05d}.txt"
+            generations_file = od / f"{type_path}_generations/{trainer.global_step:05d}.txt"
+            results_file.parent.mkdir(exist_ok=True)
+            generations_file.parent.mkdir(exist_ok=True)
         with open(results_file, "a+") as writer:
             for key in sorted(metrics):
                 if key in ["log", "progress_bar", "preds"]:
@@ -63,20 +66,25 @@ class Seq2SeqLoggingCallback(pl.Callback):
         # mp stands for million parameters
         trainer.logger.log_metrics({"n_params": npars, "mp": npars / 1e6, "grad_mp": n_trainable_pars / 1e6})
 
-    @rank_zero_only
-    def on_validation_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
-        return self._write_logs(trainer, pl_module, "val")
-
     @rank_zero_only
     def on_test_end(self, trainer: pl.Trainer, pl_module: pl.LightningModule):
         return self._write_logs(trainer, pl_module, "test")
 
 
-def get_rouge2_checkpoint_callback(output_dir):
+def get_checkpoint_callback(output_dir, metric):
     """Saves the best model by validation ROUGE2 score."""
+    if metric == "rouge2":
+        exp = "{val_avg_rouge2:.4f}-{step_count}"
+    elif metric == "bleu":
+        exp = "{val_avg_bleu:.4f}-{step_count}"
+    else:
+        raise NotImplementedError(
+            f"seq2seq callbacks only support rouge2 and bleu, got {metric}, You can make your own by adding to this function."
+        )
+
     checkpoint_callback = ModelCheckpoint(
-        filepath=os.path.join(output_dir, "{val_avg_rouge2:.4f}-{step_count}"),
-        monitor="val_rouge",
+        filepath=os.path.join(output_dir, exp),
+        monitor=f"val_{metric}",
         mode="max",
         save_top_k=1,
         period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.

examples/seq2seq/distillation.py

@@ -39,13 +39,12 @@ except ImportError:
     )
 
 
-class SummarizationDistiller(SummarizationModule):
+class BartSummarizationDistiller(SummarizationModule):
     loss_names = ["loss", "ce_loss", "mlm_loss", "enc_mse_loss", "hid_loss_enc", "hid_loss_dec"]
 
     def __init__(self, hparams):
         assert Path(hparams.data_dir).exists()
-
-        d_layers_to_copy, student, student_cfg, teacher = self.pre_init(hparams)
+        student, student_cfg, teacher = self.pre_init(hparams)
 
         super().__init__(hparams, model=student, config=student_cfg)
         self.teacher = teacher
@@ -73,12 +72,15 @@ class SummarizationDistiller(SummarizationModule):
             del self.teacher.model.encoder
 
     def pre_init(self, hparams):
-        # Dump empty student model at a path, then call from_pretrained on it
+        self.output_dir = Path(hparams.output_dir)
+        self.output_dir.mkdir(exist_ok=True)
         teacher = BartForConditionalGeneration.from_pretrained(hparams.teacher).eval()
         student_updates = {
             "decoder_layers": hparams.student_decoder_layers,
             "encoder_layers": hparams.student_encoder_layers,
         }
+        if hparams.length_penalty != -1:
+            student_updates["length_penalty"] = hparams.length_penalty
         d_layers_to_copy = get_layers_to_copy(student_updates["decoder_layers"], teacher.config.decoder_layers)
         e_layers_to_copy: List = get_layers_to_copy(student_updates["encoder_layers"], teacher.config.encoder_layers)
         hparams.d_layer_to_copy = d_layers_to_copy
@@ -89,9 +91,11 @@ class SummarizationDistiller(SummarizationModule):
         student_cfg = BartConfig(**kw)
         student = BartForConditionalGeneration(student_cfg)
         student, _ = init_student(student, teacher)
+        save_dir = self.output_dir.joinpath("student")
         self.copy_to_student(d_layers_to_copy, e_layers_to_copy, hparams, student, teacher)
-        Path(hparams.output_dir).mkdir(exist_ok=True)
-        return d_layers_to_copy, student, student_cfg, teacher
+        student.save_pretrained(save_dir)
+        hparams.model_name_or_path = str(save_dir)
+        return student, student_cfg, teacher
 
     def copy_to_student(self, d_layers_to_copy, e_layers_to_copy, hparams, student, teacher):
         if teacher.config.model_type == "t5":
@@ -154,7 +158,6 @@ class SummarizationDistiller(SummarizationModule):
 
     def configure_optimizers(self):
         "Prepare optimizer and schedule (linear warmup and decay)"
-
         model = self.model
         no_decay = ["bias", "LayerNorm.weight"]
         optimizer_grouped_parameters = [
@@ -180,18 +183,11 @@ class SummarizationDistiller(SummarizationModule):
         # 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(  # TODO: remove
-            "--enc_only", action="store_true", default=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)
+
         return parser
 
     def _step(self, batch):
@@ -269,12 +265,14 @@ class SummarizationDistiller(SummarizationModule):
         return sum(hidden_losses)
 
 
-class T5SummarizationDistiller(SummarizationDistiller):
+class T5SummarizationDistiller(BartSummarizationDistiller):
     def pre_init(self, hparams):
         raise NotImplementedError("T5 Distillation does not work yet")
+        self.output_dir = Path(hparams.output_dir)
+        self.output_dir.mkdir(exist_ok=True)
         teacher = T5ForConditionalGeneration.from_pretrained(hparams.teacher)
         n_layer = hparams.student_decoder_layers
-        assert n_layer == hparams.student_encoder_layers  # TODO(SS): relax this
+        assert n_layer == hparams.student_encoder_layers  # TODO(SS): relax this constraint so that we can do 12-6.
         d_layers_to_copy = get_layers_to_copy(n_layer, len(teacher.decoder.block))
         e_layers_to_copy: List = get_layers_to_copy(n_layer, len(teacher.encoder.block))
         student_updates = {"num_layers": n_layer}
@@ -291,8 +289,13 @@ class T5SummarizationDistiller(SummarizationDistiller):
         Path(hparams.output_dir).mkdir(exist_ok=True)
         task_specific_params = student.config.task_specific_params
         if task_specific_params is not None:
-            student.config.update(task_specific_params.get("summarization", {}))
-        return d_layers_to_copy, student, student_cfg, teacher
+            student.config.update(task_specific_params.get("summarization", {}))  # TODO: dont hardcode
+        save_dir = self.output_dir.joinpath("student")
+        save_dir.mkdir(exist_ok=True)
+
+        student.save_pretrained(save_dir)
+        hparams.model_name_or_path = str(save_dir)
+        return student, student_cfg, teacher
 
     def freeze_embeds(self):
         freeze_params(self.model.shared)
@@ -386,7 +389,7 @@ def create_module(args):
     elif args.enc_only:
         raise ValueError("Deleted that")
     else:
-        module_cls = SummarizationDistiller
+        module_cls = BartSummarizationDistiller
     args.setup_cls: str = module_cls.__name__
     model = module_cls(args)
     return model
@@ -418,18 +421,18 @@ def evaluate_checkpoint(ckpt_path: Path, dest_dir=None):
 def get_layers_to_copy(n_to_get, tot):
     all_layers = list(range(tot))
     if tot == 12:  # Alternating for special cases
-        layers_to_copy = {  # maps # layers in student -> which teacher layers to copy
-            6: [0, 2, 4, 7, 9, 11],
-            1: [11],
+        layers_to_copy = {  # maps  num layers in student -> which teacher layers to copy
+            1: [0],
+            2: [0, 6],
             3: [0, 6, 11],
-            2: [0, 11],
             4: [0, 4, 8, 11],
+            6: [0, 2, 4, 7, 9, 11],
             9: [0, 1, 2, 4, 5, 7, 9, 10, 11],
             12: all_layers,
         }
         return layers_to_copy[n_to_get]
     else:
-        return all_layers[:n_to_get]
+        return all_layers[:n_to_get]  # TODO: better version on theseus-bart branch
 
 
 def distill_main(args):
@@ -443,7 +446,7 @@ def distill_main(args):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser = SummarizationDistiller.add_model_specific_args(parser, os.getcwd())
+    parser = BartSummarizationDistiller.add_model_specific_args(parser, os.getcwd())
     args = parser.parse_args()
 
     distill_main(args)

examples/seq2seq/finetune.py

@@ -3,6 +3,8 @@ import glob
 import logging
 import os
 import time
+import warnings
+from collections import defaultdict
 from pathlib import Path
 from typing import Dict, List, Tuple
 
@@ -12,23 +14,26 @@ import torch
 from torch.utils.data import DataLoader
 
 from lightning_base import BaseTransformer, add_generic_args, generic_train
-from transformers import get_linear_schedule_with_warmup
+from transformers import MBartTokenizer, get_linear_schedule_with_warmup
 
 
 try:
     from .utils import (
+        assert_all_frozen,
         use_task_specific_params,
         SummarizationDataset,
         lmap,
         flatten_list,
         pickle_save,
         save_git_info,
+        save_json,
         freeze_params,
         calculate_rouge,
         get_git_info,
         ROUGE_KEYS,
+        calculate_bleu_score,
     )
-    from .callbacks import Seq2SeqLoggingCallback, get_rouge2_checkpoint_callback
+    from .callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback
 except ImportError:
     from utils import (
         use_task_specific_params,
@@ -37,12 +42,15 @@ except ImportError:
         flatten_list,
         pickle_save,
         save_git_info,
+        save_json,
         freeze_params,
         calculate_rouge,
         get_git_info,
         ROUGE_KEYS,
+        calculate_bleu_score,
+        assert_all_frozen,
     )
-    from callbacks import Seq2SeqLoggingCallback, get_rouge2_checkpoint_callback
+    from callbacks import Seq2SeqLoggingCallback, get_checkpoint_callback
 
 logger = logging.getLogger(__name__)
 
@@ -50,15 +58,18 @@ logger = logging.getLogger(__name__)
 class SummarizationModule(BaseTransformer):
     mode = "summarization"
     loss_names = ["loss"]
+    metric_names = ROUGE_KEYS
+    val_metric = "rouge2"
 
     def __init__(self, hparams, **kwargs):
         super().__init__(hparams, num_labels=None, mode=self.mode, **kwargs)
         use_task_specific_params(self.model, "summarization")
         save_git_info(self.hparams.output_dir)
-        self.metrics_save_path = Path(self.output_dir) / "metrics.pkl"
+        self.metrics_save_path = Path(self.output_dir) / "metrics.json"
         self.hparams_save_path = Path(self.output_dir) / "hparams.pkl"
+        pickle_save(self.hparams, self.hparams_save_path)
         self.step_count = 0
-        self.metrics = {"train": [], "val": [], "test": []}
+        self.metrics = defaultdict(list)
 
         self.dataset_kwargs: dict = dict(
             data_dir=self.hparams.data_dir,
@@ -83,18 +94,21 @@ class SummarizationModule(BaseTransformer):
         if self.hparams.freeze_embeds:
             self.freeze_embeds()
         if self.hparams.freeze_encoder:
-            freeze_params(self.model.model.encoder)  # TODO: this will break for t5
+            freeze_params(self.model.get_encoder())
+            assert_all_frozen(self.model.get_encoder())
+
         self.hparams.git_sha = get_git_info()["repo_sha"]
         self.num_workers = hparams.num_workers
+        self.decoder_start_token_id = None
 
     def freeze_embeds(self):
         """Freeze token embeddings and positional embeddings for bart, just token embeddings for t5."""
-        if self.model.config.model_type == "bart":
+        try:
             freeze_params(self.model.model.shared)
             for d in [self.model.model.encoder, self.model.model.decoder]:
                 freeze_params(d.embed_positions)
                 freeze_params(d.embed_tokens)
-        else:
+        except AttributeError:
             freeze_params(self.model.shared)
             for d in [self.model.encoder, self.model.decoder]:
                 freeze_params(d.embed_tokens)
@@ -130,31 +144,39 @@ 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 ROUGE_KEYS + ["gen_time", "summ_len"]}
-        rouge_tensor: torch.FloatTensor = torch.tensor(rouges["rouge2"]).type_as(loss)
+        rouges = {k: np.array([x[k] for x in outputs]).mean() for k in self.metric_names + ["gen_time", "summ_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)
         metrics = {f"{prefix}_avg_{k}": x for k, x in losses.items()}
         metrics["step_count"] = self.step_count
         self.save_metrics(metrics, prefix)  # writes to self.metrics_save_path
         preds = flatten_list([x["preds"] for x in outputs])
-        return {"log": metrics, "preds": preds, f"{prefix}_loss": loss, f"{prefix}_rouge": rouge_tensor}
+        return {"log": metrics, "preds": preds, f"{prefix}_loss": loss, f"{prefix}_{self.val_metric}": rouge_tensor}
 
-    def save_metrics(self, metrics, prefix) -> None:
-        self.metrics[prefix].append(metrics)
-        pickle_save(self.metrics, self.metrics_save_path)
+    def save_metrics(self, latest_metrics, type_path) -> None:
+        self.metrics[type_path].append(latest_metrics)
+        save_json(self.metrics, self.metrics_save_path)
+
+    def calc_generative_metrics(self, preds, target) -> Dict:
+        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, use_cache=True,)
-        gen_time = time.time() - t0 / source_ids.shape[0]
+        generated_ids = self.model.generate(
+            input_ids=source_ids,
+            attention_mask=source_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)
         loss_tensors = self._step(batch)
         base_metrics = {name: loss for name, loss in zip(self.loss_names, loss_tensors)}
-        rouge: Dict = calculate_rouge(preds, target)
+        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)
         return base_metrics
@@ -205,6 +227,8 @@ class SummarizationModule(BaseTransformer):
         scheduler = get_linear_schedule_with_warmup(
             self.opt, num_warmup_steps=self.hparams.warmup_steps, num_training_steps=t_total
         )
+        if max(scheduler.get_last_lr()) > 0:
+            warnings.warn("All learning rates are 0")
         self.lr_scheduler = scheduler
         return dataloader
 
@@ -259,15 +283,41 @@ class SummarizationModule(BaseTransformer):
         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("--src_lang", type=str, default="", required=False)
+        parser.add_argument("--tgt_lang", type=str, default="", required=False)
+
         return parser
 
 
+class TranslationModule(SummarizationModule):
+    mode = "translation"
+    loss_names = ["loss"]
+    metric_names = ["bleu"]
+    val_metric = "bleu"
+
+    def __init__(self, hparams, **kwargs):
+        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)
+
+
 def main(args, model=None) -> SummarizationModule:
     Path(args.output_dir).mkdir(exist_ok=True)
     if len(os.listdir(args.output_dir)) > 3 and args.do_train:
         raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
     if model is None:
-        model: BaseTransformer = SummarizationModule(args)
+        if args.task == "summarization":
+            model: SummarizationModule = SummarizationModule(args)
+        else:
+            model: SummarizationModule = TranslationModule(args)
     if (
         args.logger == "default"
         or args.fast_dev_run
@@ -279,16 +329,16 @@ def main(args, model=None) -> SummarizationModule:
         from pytorch_lightning.loggers import WandbLogger
 
         logger = WandbLogger(name=model.output_dir.name)
+
     elif args.logger == "wandb_shared":
         from pytorch_lightning.loggers import WandbLogger
 
-        # TODO: separate LB for CNN, we should use Path(args.data_dir).name to determine the correct LB.
-        logger = WandbLogger(name=model.output_dir.name, project="hf_summarization")
+        logger = WandbLogger(name=model.output_dir.name)
     trainer: pl.Trainer = generic_train(
         model,
         args,
         logging_callback=Seq2SeqLoggingCallback(),
-        checkpoint_callback=get_rouge2_checkpoint_callback(args.output_dir),
+        checkpoint_callback=get_checkpoint_callback(args.output_dir, model.val_metric),
         logger=logger,
         # TODO: early stopping callback seems messed up
     )

examples/seq2seq/finetune.sh

@@ -1,13 +1,8 @@
-
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"
 
-
-# --model_name_or_path=t5-base for t5
-
-# the proper usage is documented in the README
+# the proper usage is documented in the README, you need to specify data_dir, output_dir and model_name_or_path
 python finetune.py \
-    --model_name_or_path=facebook/bart-large \
     --learning_rate=3e-5 \
     --fp16 \
     --gpus 1 \
@@ -16,5 +11,4 @@ python finetune.py \
     --n_val 1000 \
     --val_check_interval 0.1 \
     --sortish_sampler \
-    --max_target_length=56 \
     $@

examples/seq2seq/finetune_bart_tiny.sh

@@ -12,7 +12,7 @@ export OUTPUT_DIR=${CURRENT_DIR}/${OUTPUT_DIR_NAME}
 # Make output directory if it doesn't exist
 mkdir -p $OUTPUT_DIR
 
-# Add parent directory to python path to access lightning_base.py and utils.py
+# Add parent directory to python path to access lightning_base.py and testing_utils.py
 export PYTHONPATH="../":"${PYTHONPATH}"
 python finetune.py \
 --data_dir=cnn_tiny/ \

examples/seq2seq/finetune_t5.sh

@@ -0,0 +1,13 @@
+# Add parent directory to python path to access lightning_base.py
+export PYTHONPATH="../":"${PYTHONPATH}"
+
+python finetune.py \
+--data_dir=$CNN_DIR \
+--learning_rate=3e-5 \
+--train_batch_size=$BS \
+--eval_batch_size=$BS \
+--output_dir=$OUTPUT_DIR \
+--max_source_length=512 \
+--val_check_interval=0.1 --n_val=200 \
+--do_train --do_predict \
+ $@

examples/seq2seq/run_distiller.sh

@@ -1,5 +1,3 @@
-#CNN_DIR = /home/shleifer/transformers_fork/examples/summarization/bart/cnn_dm
-
 # Add parent directory to python path to access lightning_base.py
 export PYTHONPATH="../":"${PYTHONPATH}"
 
@@ -7,5 +5,6 @@ python distillation.py \
 --learning_rate=3e-4 \
 --do_train \
 --do_predict \
+--fp16 \
 --val_check_interval 0.1 \
 $@

examples/seq2seq/run_eval.py

@@ -9,9 +9,9 @@ from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
 
 
 try:
-    from .finetune import calculate_rouge, use_task_specific_params
+    from .utils import calculate_rouge, use_task_specific_params, calculate_bleu_score, trim_batch
 except ImportError:
-    from finetune import calculate_rouge, use_task_specific_params
+    from utils import calculate_rouge, use_task_specific_params, calculate_bleu_score, trim_batch
 
 DEFAULT_DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
@@ -22,8 +22,15 @@ def chunks(lst, n):
         yield lst[i : i + n]
 
 
-def generate_summaries(
-    examples: list, out_file: str, model_name: str, batch_size: int = 8, device: str = DEFAULT_DEVICE, fp16=False,
+def generate_summaries_or_translations(
+    examples: list,
+    out_file: str,
+    model_name: str,
+    batch_size: int = 8,
+    device: str = DEFAULT_DEVICE,
+    fp16=False,
+    task="summarization",
+    **gen_kwargs,
 ) -> None:
     fout = Path(out_file).open("w", encoding="utf-8")
     model_name = str(model_name)
@@ -34,16 +41,16 @@ def generate_summaries(
     tokenizer = AutoTokenizer.from_pretrained(model_name)
 
     # update config with summarization specific params
-    use_task_specific_params(model, "summarization")
+    use_task_specific_params(model, task)
 
     for batch in tqdm(list(chunks(examples, batch_size))):
         if "t5" in model_name:
             batch = [model.config.prefix + text for text in batch]
-        dct = tokenizer.batch_encode_plus(batch, max_length=1024, return_tensors="pt", pad_to_max_length=True).to(
+        batch = tokenizer(batch, max_length=1024, return_tensors="pt", truncation=True, padding="max_length").to(
             device
         )
-        summaries = model.generate(**dct)
-
+        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)
         dec = tokenizer.batch_decode(summaries, skip_special_tokens=True, clean_up_tokenization_spaces=False)
         for hypothesis in dec:
             fout.write(hypothesis + "\n")
@@ -52,27 +59,43 @@ def generate_summaries(
 
 def run_generate():
     parser = argparse.ArgumentParser()
-    parser.add_argument("input_path", type=str, help="like cnn_dm/test.source")
-    parser.add_argument("output_path", type=str, help="where to save summaries")
     parser.add_argument("model_name", type=str, help="like facebook/bart-large-cnn,t5-base, etc.")
+    parser.add_argument("input_path", type=str, help="like cnn_dm/test.source")
+    parser.add_argument("save_path", type=str, help="where to save summaries")
+
     parser.add_argument("--reference_path", type=str, required=False, help="like cnn_dm/test_reference_summaries.txt")
     parser.add_argument("--score_path", type=str, required=False, help="where to save the rouge score in json format")
     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(
+        "--n_obs", type=int, default=-1, required=False, help="How many observations. Defaults to all."
+    )
     parser.add_argument("--fp16", action="store_true")
     args = parser.parse_args()
     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]
 
-    generate_summaries(
-        examples, args.output_path, args.model_name, batch_size=args.bs, device=args.device, fp16=args.fp16
+    generate_summaries_or_translations(
+        examples,
+        args.save_path,
+        args.model_name,
+        batch_size=args.bs,
+        device=args.device,
+        fp16=args.fp16,
+        task=args.task,
     )
+    if args.reference_path is None:
+        return
+    # Compute scores
+    score_fn = calculate_bleu_score if "translation" in args.task else calculate_rouge
+    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)
     if args.score_path is not None:
-        output_lns = [x.rstrip() for x in open(args.output_path).readlines()]
-        reference_lns = [x.rstrip() for x in open(args.reference_path).readlines()]
-
-        rouge: dict = calculate_rouge(output_lns, reference_lns)
-
-        json.dump(rouge, open("score_path", "w+"))
+        json.dump(scores, open(args.score_path, "w+"))
+    return scores
 
 
 if __name__ == "__main__":

examples/seq2seq/test_seq2seq_examples.py

@@ -0,0 +1,277 @@
+import argparse
+import logging
+import os
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+import torch
+from torch.utils.data import DataLoader
+
+from transformers import AutoTokenizer
+from transformers.testing_utils import require_multigpu
+
+from .distillation import distill_main, evaluate_checkpoint
+from .finetune import main
+from .run_eval import generate_summaries_or_translations, run_generate
+from .utils import SummarizationDataset, lmap, load_json
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+CUDA_AVAILABLE = torch.cuda.is_available()
+CHEAP_ARGS = {
+    "logger": "default",
+    "length_penalty": 0.5,
+    "cache_dir": "",
+    "task": "summarization",
+    "num_workers": 2,
+    "alpha_hid": 0,
+    "freeze_embeds": True,
+    "enc_only": False,
+    "tgt_suffix": "",
+    "resume_from_checkpoint": None,
+    "sortish_sampler": True,
+    "student_decoder_layers": 1,
+    "val_check_interval": 1.0,
+    "output_dir": "",
+    "fp16": CUDA_AVAILABLE,
+    "no_teacher": False,
+    "fp16_opt_level": "O1",
+    "gpus": 1 if CUDA_AVAILABLE else 0,
+    "n_tpu_cores": 0,
+    "max_grad_norm": 1.0,
+    "do_train": True,
+    "do_predict": True,
+    "gradient_accumulation_steps": 1,
+    "server_ip": "",
+    "server_port": "",
+    "seed": 42,
+    "model_name_or_path": "sshleifer/bart-tiny-random",
+    "config_name": "",
+    "tokenizer_name": "facebook/bart-large",
+    "do_lower_case": False,
+    "learning_rate": 0.3,
+    "weight_decay": 0.0,
+    "adam_epsilon": 1e-08,
+    "warmup_steps": 0,
+    "num_train_epochs": 1,
+    "train_batch_size": 2,
+    "eval_batch_size": 2,
+    "max_source_length": 12,
+    "max_target_length": 12,
+    "val_max_target_length": 12,
+    "test_max_target_length": 12,
+    "fast_dev_run": False,
+    "no_cache": False,
+    "n_train": -1,
+    "n_val": -1,
+    "n_test": -1,
+    "student_encoder_layers": 1,
+    "alpha_loss_encoder": 0.0,
+    "freeze_encoder": False,
+    "auto_scale_batch_size": False,
+}
+
+
+def _dump_articles(path: Path, articles: list):
+    with path.open("w") as f:
+        f.write("\n".join(articles))
+
+
+ARTICLES = [" Sam ate lunch today", "Sams lunch ingredients"]
+SUMMARIES = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
+T5_TINY = "patrickvonplaten/t5-tiny-random"
+BART_TINY = "sshleifer/bart-tiny-random"
+MBART_TINY = "sshleifer/tiny-mbart"
+MARIAN_TINY = "sshleifer/tiny-marian-en-de"
+stream_handler = logging.StreamHandler(sys.stdout)
+logger.addHandler(stream_handler)
+logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
+
+
+def make_test_data_dir(**kwargs):
+    tmp_dir = Path(tempfile.mkdtemp(**kwargs))
+    for split in ["train", "val", "test"]:
+        _dump_articles((tmp_dir / f"{split}.source"), ARTICLES)
+        _dump_articles((tmp_dir / f"{split}.target"), SUMMARIES)
+    return tmp_dir
+
+
+class TestSummarizationDistiller(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
+        return cls
+
+    @require_multigpu
+    def test_multigpu(self):
+        updates = dict(no_teacher=True, freeze_encoder=True, gpus=2, sortish_sampler=False,)
+        self._test_distiller_cli(updates)
+
+    def test_distill_no_teacher(self):
+        updates = dict(student_encoder_layers=2, student_decoder_layers=1, no_teacher=True)
+        self._test_distiller_cli(updates)
+
+    def test_distill_checkpointing_with_teacher(self):
+        updates = dict(
+            student_encoder_layers=2,
+            student_decoder_layers=1,
+            num_train_epochs=4,
+            val_check_interval=0.25,
+            alpha_hid=2.0,
+            model_name_or_path="IGNORE_THIS_IT_DOESNT_GET_USED",
+        )
+        model = self._test_distiller_cli(updates, check_contents=False)
+
+        ckpts = list(Path(model.output_dir).glob("*.ckpt"))
+        self.assertEqual(1, len(ckpts))
+        transformer_ckpts = list(Path(model.output_dir).glob("**/*.bin"))
+        self.assertEqual(len(transformer_ckpts), 2)
+        examples = lmap(str.strip, model.hparams.data_dir.joinpath("test.source").open().readlines())
+        out_path = tempfile.mktemp()
+        generate_summaries_or_translations(examples, out_path, str(model.output_dir / "best_tfmr"))
+        self.assertTrue(Path(out_path).exists())
+
+        evaluate_checkpoint(ckpts[0], dest_dir=Path(tempfile.mkdtemp()))
+
+    @unittest.skip("T5 distillation is broken at the moment")
+    def test_distill_t5(self):
+        updates = dict(
+            student_encoder_layers=1,
+            student_decoder_layers=1,
+            alpha_hid=2.0,
+            teacher=T5_TINY,
+            model_name_or_path=T5_TINY,
+            tokenizer_name=T5_TINY,
+        )
+        self._test_distiller_cli(updates)
+
+    def _test_distiller_cli(self, updates, check_contents=True):
+        default_updates = dict(
+            train_batch_size=1,
+            eval_batch_size=2,
+            num_train_epochs=2,
+            alpha_mlm=0.2,
+            alpha_ce=0.8,
+            do_predict=True,
+            model_name_or_path="sshleifer/tinier_bart",
+            teacher=CHEAP_ARGS["model_name_or_path"],
+            val_check_interval=0.5,
+            alpha_encoder_loss=0.4,
+        )
+        default_updates.update(updates)
+        args_d: dict = CHEAP_ARGS.copy()
+        tmp_dir = make_test_data_dir()
+        output_dir = tempfile.mkdtemp(prefix="output_")
+
+        args_d.update(data_dir=tmp_dir, output_dir=output_dir, **default_updates)
+        model = distill_main(argparse.Namespace(**args_d))
+        if not check_contents:
+            return model
+        contents = os.listdir(output_dir)
+        ckpt_name = "val_avg_rouge2=0.0000-step_count=2.ckpt"  # "val_avg_rouge2=0.0000-epoch=1.ckpt"  # "epoch=1-val_avg_rouge2=0.0000.ckpt"
+        contents = {os.path.basename(p) for p in contents}
+        self.assertIn(ckpt_name, contents)
+
+        self.assertIn("test_generations.txt", contents)
+        self.assertIn("test_results.txt", contents)
+
+        metrics = load_json(model.metrics_save_path)
+        last_step_stats = metrics["val"][-1]
+        self.assertGreaterEqual(last_step_stats["val_avg_gen_time"], 0.01)
+        self.assertGreaterEqual(1.0, last_step_stats["val_avg_gen_time"])
+        self.assertIsInstance(last_step_stats[f"val_avg_{model.val_metric}"], float)
+        desired_n_evals = int(args_d["num_train_epochs"] * (1 / args_d["val_check_interval"]) + 1)
+        self.assertEqual(len(metrics["val"]), desired_n_evals)
+        self.assertEqual(len(metrics["test"]), 1)
+        return model
+
+
+@pytest.mark.parametrize(["model"], [pytest.param(T5_TINY), pytest.param(BART_TINY), pytest.param(MBART_TINY)])
+def test_run_eval_bart(model):
+    input_file_name = Path(tempfile.mkdtemp()) / "utest_input.source"
+    output_file_name = input_file_name.parent / "utest_output.txt"
+    assert not output_file_name.exists()
+    articles = [" New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]
+    _dump_articles(input_file_name, articles)
+    testargs = ["run_eval.py", model, str(input_file_name), str(output_file_name)]  # TODO: test score_path
+    with patch.object(sys, "argv", testargs):
+        run_generate()
+        assert Path(output_file_name).exists()
+        os.remove(Path(output_file_name))
+
+
+@pytest.mark.parametrize(
+    ["model"], [pytest.param(T5_TINY), pytest.param(BART_TINY), pytest.param(MBART_TINY), pytest.param(MARIAN_TINY)]
+)
+def test_finetune(model):
+    args_d: dict = CHEAP_ARGS.copy()
+    task = "translation" if model in [MBART_TINY, MARIAN_TINY] else "summarization"
+    tmp_dir = make_test_data_dir()
+    output_dir = tempfile.mkdtemp(prefix="output_")
+    args_d.update(
+        data_dir=tmp_dir,
+        model_name_or_path=model,
+        tokenizer_name=None,
+        train_batch_size=2,
+        eval_batch_size=2,
+        output_dir=output_dir,
+        do_predict=True,
+        task=task,
+        src_lang="en_XX",
+        tgt_lang="ro_RO",
+        freeze_encoder=True,
+        freeze_embeds=True,
+    )
+    assert "n_train" in args_d
+    args = argparse.Namespace(**args_d)
+    module = main(args)
+
+    input_embeds = module.model.get_input_embeddings()
+    assert not input_embeds.weight.requires_grad
+    if model == T5_TINY:
+        lm_head = module.model.lm_head
+        assert not lm_head.weight.requires_grad
+        assert (lm_head.weight == input_embeds.weight).all().item()
+
+    else:
+        bart = module.model.model
+        embed_pos = bart.decoder.embed_positions
+        assert not embed_pos.weight.requires_grad
+        assert not bart.shared.weight.requires_grad
+        # check that embeds are the same
+        assert bart.decoder.embed_tokens == bart.encoder.embed_tokens
+        assert bart.decoder.embed_tokens == bart.shared
+
+
+@pytest.mark.parametrize(
+    ["tok"], [pytest.param(T5_TINY), pytest.param(BART_TINY), pytest.param(MBART_TINY), pytest.param(MARIAN_TINY)]
+)
+def test_dataset(tok):
+    tokenizer = AutoTokenizer.from_pretrained(tok)
+    tmp_dir = make_test_data_dir()
+    max_len_source = max(len(tokenizer.encode(a)) for a in ARTICLES)
+    max_len_target = max(len(tokenizer.encode(a)) for a in SUMMARIES)
+    trunc_target = 4
+    train_dataset = SummarizationDataset(
+        tokenizer,
+        data_dir=tmp_dir,
+        type_path="train",
+        max_source_length=20,
+        max_target_length=trunc_target,
+        tgt_lang="ro_RO",
+    )
+    dataloader = DataLoader(train_dataset, batch_size=2, collate_fn=train_dataset.collate_fn)
+    for batch in dataloader:
+        assert batch["attention_mask"].shape == batch["input_ids"].shape
+        # show that articles were trimmed.
+        assert batch["input_ids"].shape[1] == max_len_source
+        assert 20 >= batch["input_ids"].shape[1]  # trimmed significantly
+        # show that targets were truncated
+        assert batch["decoder_input_ids"].shape[1] == trunc_target  # Truncated
+        assert max_len_target > trunc_target  # Truncated

examples/seq2seq/train_distilbart_cnn.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+export PYTHONPATH="../":"${PYTHONPATH}"
+
+export BS=32
+export GAS=1
+
+python finetune.py \
+    --learning_rate=3e-5 \
+    --fp16 \
+    --gpus 1 \
+    --do_train \
+    --do_predict \
+    --val_check_interval 0.25 \
+    --n_val 500 \
+    --num_train_epochs 2 \
+    --freeze_encoder --freeze_embeds --data_dir $CNN_DIR \
+    --max_target_length 142 --val_max_target_length=142 \
+    --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS \
+    --model_name_or_path sshleifer/student_cnn_12_6 \
+    --tokenizer_name facebook/bart-large \
+    --warmup_steps 500 \
+    --output_dir distilbart-cnn-12-6 \
+    $@
+

examples/seq2seq/train_distilbart_xsum.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+export PYTHONPATH="../":"${PYTHONPATH}"
+export BS=16
+export GAS=2
+python distillation.py \
+  --learning_rate=3e-4 \
+  --do_train \
+  --do_predict \
+  --fp16 \
+  --val_check_interval 0.1 --n_val 1000 \
+  --teacher facebook/bart-large-xsum --data_dir $XSUM_DIR \
+  --max_target_length=60 --val_max_target_length=60 --test_max_target_length=100 \
+  --student_decoder_layers 6 --student_encoder_layers 12 \
+  --freeze_encoder --freeze_embeds \
+  --model_name_or_path IGNORED \
+  --alpha_hid=3. --length_penalty=0.5 \
+  --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS --num_train_epochs=6 \
+  --tokenizer_name facebook/bart-large \
+  --warmup_steps 500 \
+  --output_dir distilbart_xsum_12_6 \
+  $@

examples/seq2seq/train_mbart_cc25_enro.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+export PYTHONPATH="../":"${PYTHONPATH}"
+
+python finetune.py \
+    --learning_rate=3e-5 \
+    --fp16 \
+    --gpus 1 \
+    --do_train \
+    --do_predict \
+    --val_check_interval 0.1 \
+    --n_val 500 \
+    --adam_eps 1e-06 \
+    --num_train_epochs 3 --src_lang en_XX --tgt_lang ro_RO \
+    --freeze_encoder --freeze_embeds --data_dir $ENRO_DIR \
+    --max_source_length=300 --max_target_length 300 --val_max_target_length=300 --test_max_target_length 300 \
+    --train_batch_size=$BS --eval_batch_size=$BS --gradient_accumulation_steps=$GAS \
+    --model_name_or_path facebook/mbart-large-cc25 \
+    --task translation \
+    --warmup_steps 500 \
+    --logger wandb --sortish_sampler \
+    $@

examples/seq2seq/utils.py

@@ -3,16 +3,19 @@ import json
 import os
 import pickle
 from pathlib import Path
-from typing import Dict, Iterable, List
+from typing import Callable, Dict, Iterable, List
 
 import git
 import numpy as np
 import torch
 from rouge_score import rouge_scorer, scoring
+from sacrebleu import corpus_bleu
 from torch import nn
 from torch.utils.data import Dataset, Sampler
 from tqdm import tqdm
 
+from transformers import BartTokenizer
+
 
 def encode_file(
     tokenizer,
@@ -24,6 +27,7 @@ def encode_file(
     prefix="",
     tok_name="",
 ):
+    extra_kw = {"add_prefix_space": True} if isinstance(tokenizer, BartTokenizer) else {}
     cache_path = Path(f"{data_path}_{tok_name}{max_length}.pt")
     if not overwrite_cache and cache_path.exists():
         try:
@@ -40,13 +44,13 @@ def encode_file(
     assert lns, f"found empty file at {data_path}"
     examples = []
     for text in tqdm(lns, desc=f"Tokenizing {data_path.name}"):
-        tokenized = tokenizer.batch_encode_plus(
-            [text],  # DONT ADD SPACES
+        tokenized = tokenizer(
+            [text],
             max_length=max_length,
-            pad_to_max_length=pad_to_max_length,
-            add_prefix_space=True,
+            padding="max_length" if pad_to_max_length else None,
             truncation=True,
             return_tensors=return_tensors,
+            **extra_kw,
         )
         assert tokenized.input_ids.shape[1] == max_length
         examples.append(tokenized)
@@ -54,11 +58,14 @@ def encode_file(
     return examples
 
 
-def lmap(f, x):
+def lmap(f: Callable, x: Iterable) -> List:
+    """list(map(f, x))"""
     return list(map(f, x))
 
 
-T5_PREFIX = "summarize: "  # HACK, fixme
+def calculate_bleu_score(output_lns, refs_lns, **kwargs) -> dict:
+    """Uses sacrebleu's corpus_bleu implementation."""
+    return {"bleu": corpus_bleu(output_lns, [refs_lns], **kwargs).score}
 
 
 def trim_batch(
@@ -83,9 +90,14 @@ class SummarizationDataset(Dataset):
         n_obs=None,
         overwrite_cache=False,
         prefix="",
+        src_lang=None,
+        tgt_lang=None,
     ):
         super().__init__()
+        # FIXME: the rstrip logic strips all the chars, it seems.
         tok_name = tokenizer.__class__.__name__.lower().rstrip("tokenizer")
+        if hasattr(tokenizer, "set_lang") and src_lang is not None:
+            tokenizer.set_lang(src_lang)  # HACK: only applies to mbart
         self.source = encode_file(
             tokenizer,
             os.path.join(data_dir, type_path + ".source"),
@@ -95,6 +107,9 @@ class SummarizationDataset(Dataset):
             tok_name=tok_name,
         )
         tgt_path = os.path.join(data_dir, type_path + ".target")
+        if hasattr(tokenizer, "set_lang"):
+            assert tgt_lang is not None, "--tgt_lang must be passed to build a translation"
+            tokenizer.set_lang(tgt_lang)  # HACK: only applies to mbart
         self.target = encode_file(
             tokenizer, tgt_path, max_target_length, overwrite_cache=overwrite_cache, tok_name=tok_name
         )
@@ -189,14 +204,20 @@ def flatten_list(summary_ids: List[List]):
     return [x for x in itertools.chain.from_iterable(summary_ids)]
 
 
-def save_git_info(folder_path: str):
-    """
-    Log commit info.
-    """
+def save_git_info(folder_path: str) -> None:
+    """Save git information to output_dir/git_log.json"""
     repo_infos = get_git_info()
+    save_json(repo_infos, os.path.join(folder_path, "git_log.json"))
 
-    with open(os.path.join(folder_path, "git_log.json"), "w") as f:
-        json.dump(repo_infos, f, indent=4)
+
+def save_json(content, path):
+    with open(path, "w") as f:
+        json.dump(content, f, indent=4)
+
+
+def load_json(path):
+    with open(path) as f:
+        return json.load(f)
 
 
 def get_git_info():
@@ -212,8 +233,8 @@ def get_git_info():
 ROUGE_KEYS = ["rouge1", "rouge2", "rougeL"]
 
 
-def calculate_rouge(output_lns: List[str], reference_lns: List[str]) -> Dict:
-    scorer = rouge_scorer.RougeScorer(ROUGE_KEYS, use_stemmer=True)
+def calculate_rouge(output_lns: List[str], reference_lns: List[str], use_stemmer=True) -> Dict:
+    scorer = rouge_scorer.RougeScorer(ROUGE_KEYS, use_stemmer=use_stemmer)
     aggregator = scoring.BootstrapAggregator()
 
     for reference_ln, output_ln in zip(reference_lns, output_lns):

examples/summarization/README.md

@@ -1,70 +0,0 @@
-### Data
-
-CNN/DailyMail data
-```bash
-cd examples/summarization
-wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/cnn_dm.tgz
-tar -xzvf cnn_dm.tgz
-export CNN_DIR=${PWD}/cnn_dm
-```
-
-this should make a directory called cnn_dm/ with files like `test.source`.
-To use your own data, copy that files format. Each article to be summarized is on its own line.
-
-XSUM Data:
-```bash
-cd examples/summarization
-wget https://s3.amazonaws.com/datasets.huggingface.co/summarization/xsum.tar.gz
-tar -xzvf xsum.tar.gz
-export XSUM_DIR=${PWD}/xsum
-```
-
-
-### Evaluation
-
-To create summaries for each article in dataset, run:
-```bash
-python run_eval.py <path_to_test.source> test_generations.txt <model-name>  --score_path rouge_scores.txt
-```
-The default batch size, 4, fits in 16GB GPU memory, but may need to be adjusted to fit your system.
-
-
-### Training
-Run/modify `finetune.sh`
-
-The following command should work on a 16GB GPU:
-```bash
-export me=`git config user.name`
-./finetune.sh \
-    --data_dir $XSUM_DIR \
-    --train_batch_size=1 \
-    --eval_batch_size=1 \
-    --output_dir="$me"_xsum_results \
-    --num_train_epochs 1
-```
-
-Tips:
-- 1 epoch at batch size 1 for bart-large takes 24 hours, requires 13GB GPU RAM with fp16 on an NVIDIA-V100. 
-- try `bart-base`, `--freeze_encoder` or `--freeze_embeds` for faster training/larger batch size.  (3hr/epoch with bs=8, see below)
-- `fp16_opt_level=O1` (the default works best).
-- If you are finetuning on your own dataset, start from `bart-large-cnn` if you want long summaries and `bart-large-xsum` if you want short summaries.
-(It rarely makes sense to start from `bart-large` unless you are a researching finetuning methods).
-- 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.
-- 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.    
-
-### XSUM Shared Task
-Compare XSUM results with others by using `--logger wandb_shared`. This requires `wandb` registration.
-Here is an example command
-```bash
-export me=`git config user.name`
-./finetune.sh \
-    --data_dir $XSUM_DIR \
-    --output_dir "$me"_xsum_frozen_embs \
-    --logger wandb_shared \
-    --train_batch_size 16 --eval_batch_size 16 --freeze_embeds --freeze_encoder \
-    --num_train_epochs 6
-```
-
-Results can be viewed [here](https://app.wandb.ai/sshleifer/hf_summarization/table?workspace=user-)

examples/summarization/finetune_t5.sh

@@ -1,18 +0,0 @@
-export OUTPUT_DIR_NAME=t5
-export CURRENT_DIR=${PWD}
-export OUTPUT_DIR=${CURRENT_DIR}/${OUTPUT_DIR_NAME}
-
-# Make output directory if it doesn't exist
-mkdir -p $OUTPUT_DIR
-
-# Add parent directory to python path to access lightning_base.py
-export PYTHONPATH="../":"${PYTHONPATH}"
-
-python finetune.py \
---data_dir=./cnn-dailymail/cnn_dm \
---model_name_or_path=t5-large \
---learning_rate=3e-5 \
---train_batch_size=4 \
---eval_batch_size=4 \
---output_dir=$OUTPUT_DIR \
---do_train  $@

examples/summarization/test_summarization_examples.py

@@ -1,267 +0,0 @@
-import argparse
-import logging
-import os
-import sys
-import tempfile
-import unittest
-from pathlib import Path
-from unittest.mock import patch
-
-import torch
-from torch.utils.data import DataLoader
-
-from transformers import BartTokenizer
-
-from .distillation import distill_main, evaluate_checkpoint
-from .finetune import main
-from .run_eval import generate_summaries, run_generate
-from .utils import SummarizationDataset, lmap, pickle_load
-
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-FP16_EVER = False
-CHEAP_ARGS = {
-    "logger": "default",
-    "num_workers": 2,
-    "alpha_hid": 0,
-    "freeze_embeds": True,
-    "enc_only": False,
-    "tgt_suffix": "",
-    "resume_from_checkpoint": None,
-    "sortish_sampler": True,
-    "student_decoder_layers": 1,
-    "val_check_interval": 1.0,
-    "output_dir": "",
-    "fp16": False,
-    "no_teacher": False,
-    "fp16_opt_level": "O1",
-    "gpus": 1 if torch.cuda.is_available() else 0,
-    "n_tpu_cores": 0,
-    "max_grad_norm": 1.0,
-    "do_train": True,
-    "do_predict": True,
-    "gradient_accumulation_steps": 1,
-    "server_ip": "",
-    "server_port": "",
-    "seed": 42,
-    "model_type": "bart",
-    "model_name_or_path": "sshleifer/bart-tiny-random",
-    "config_name": "",
-    "tokenizer_name": "facebook/bart-large",
-    "cache_dir": "",
-    "do_lower_case": False,
-    "learning_rate": 3e-05,
-    "weight_decay": 0.0,
-    "adam_epsilon": 1e-08,
-    "warmup_steps": 0,
-    "num_train_epochs": 1,
-    "train_batch_size": 2,
-    "eval_batch_size": 2,
-    "max_source_length": 12,
-    "max_target_length": 12,
-    "val_max_target_length": 12,
-    "test_max_target_length": 12,
-    "fast_dev_run": False,
-    "no_cache": False,
-    "n_train": -1,
-    "n_val": -1,
-    "n_test": -1,
-    "student_encoder_layers": 1,
-    "alpha_loss_encoder": 0.0,
-    "freeze_encoder": False,
-    "auto_scale_batch_size": False,
-}
-
-
-def _dump_articles(path: Path, articles: list):
-    with path.open("w") as f:
-        f.write("\n".join(articles))
-
-
-MSG = "T5 is broken at the moment"
-T5_TINY = "patrickvonplaten/t5-tiny-random"
-
-
-def make_test_data_dir():
-    tmp_dir = Path(tempfile.gettempdir())
-    articles = [" Sam ate lunch today", "Sams lunch ingredients"]
-    summaries = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
-    for split in ["train", "val", "test"]:
-        _dump_articles((tmp_dir / f"{split}.source"), articles)
-        _dump_articles((tmp_dir / f"{split}.target"), summaries)
-    return tmp_dir
-
-
-class TestSummarizationDistiller(unittest.TestCase):
-    @classmethod
-    def setUpClass(cls):
-        logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
-        return cls
-
-    @unittest.skipUnless(torch.cuda.device_count() > 1, "skipping multiGPU test")
-    def test_bdc_multigpu(self):
-        updates = dict(
-            student_encoder_layers=2,
-            student_decoder_layers=1,
-            no_teacher=True,
-            freeze_encoder=True,
-            gpus=2,
-            sortish_sampler=False,
-            fp16_opt_level="O1",
-            fp16=FP16_EVER,
-        )
-        self._bart_distiller_cli(updates)
-
-    def test_bdc_t5_train(self):
-        updates = dict(
-            fp16=FP16_EVER,
-            gpus=1 if torch.cuda.is_available() else 0,
-            model_type="t5",
-            model_name_or_path=T5_TINY,
-            do_train=True,
-            do_predict=True,
-            tokenizer_name=T5_TINY,
-            no_teacher=True,
-            alpha_hid=2.0,
-        )
-        self._bart_distiller_cli(updates)
-
-    def test_bdc_no_teacher(self):
-        updates = dict(student_encoder_layers=2, student_decoder_layers=1, no_teacher=True,)
-        self._bart_distiller_cli(updates)
-
-    def test_bdc_yes_teacher(self):
-        updates = dict(student_encoder_layers=2, student_decoder_layers=1,)
-        self._bart_distiller_cli(updates)
-
-    def test_bdc_checkpointing(self):
-        updates = dict(
-            student_encoder_layers=2,
-            student_decoder_layers=1,
-            num_train_epochs=4,
-            val_check_interval=0.25,
-            alpha_hid=2.0,
-        )
-        model = self._bart_distiller_cli(updates, check_contents=False)
-
-        ckpts = list(Path(model.output_dir).glob("*.ckpt"))
-        self.assertEqual(1, len(ckpts))
-        transformer_ckpts = list(Path(model.output_dir).glob("**/*.bin"))
-        self.assertEqual(len(transformer_ckpts), len(ckpts))
-        new_transformer_ckpts = list(Path(model.output_dir).glob("**/*.bin"))
-        self.assertEqual(len(new_transformer_ckpts), 1)
-        examples = lmap(str.strip, model.hparams.data_dir.joinpath("test.source").open().readlines())
-        out_path = tempfile.mktemp()
-        generate_summaries(examples, out_path, new_transformer_ckpts[0].parent)
-        self.assertTrue(Path(out_path).exists())
-
-        evaluate_checkpoint(ckpts[0], dest_dir=Path(tempfile.mkdtemp()))
-
-    def _bart_distiller_cli(self, updates, check_contents=True):
-        default_updates = dict(
-            train_batch_size=1,
-            eval_batch_size=2,
-            num_train_epochs=2,
-            alpha_mlm=0.2,
-            alpha_ce=0.8,
-            do_predict=True,
-            gpus=1 if torch.cuda.is_available() else 0,
-            model_name_or_path="sshleifer/tinier_bart",
-            teacher=CHEAP_ARGS["model_name_or_path"],
-            val_check_interval=0.5,
-            alpha_encoder_loss=0.4,
-        )
-        default_updates.update(updates)
-        args_d: dict = CHEAP_ARGS.copy()
-        tmp_dir = make_test_data_dir()
-        output_dir = tempfile.mkdtemp(prefix="output_")
-
-        args_d.update(data_dir=tmp_dir, output_dir=output_dir, **default_updates)
-        model = distill_main(argparse.Namespace(**args_d))
-        if not check_contents:
-            return model
-        contents = os.listdir(output_dir)
-        ckpt_name = "val_avg_rouge2=0.0000-step_count=2.ckpt"  # "val_avg_rouge2=0.0000-epoch=1.ckpt"  # "epoch=1-val_avg_rouge2=0.0000.ckpt"
-        contents = {os.path.basename(p) for p in contents}
-        self.assertIn(ckpt_name, contents)
-        self.assertIn("metrics.pkl", contents)
-        self.assertIn("test_generations.txt", contents)
-        self.assertIn("val_generations_00001.txt", contents)
-        self.assertIn("val_results_00001.txt", contents)
-        self.assertIn("test_results.txt", contents)
-
-        metrics = pickle_load(Path(output_dir) / "metrics.pkl")
-        desired_n_evals = int(args_d["num_train_epochs"] * (1 / args_d["val_check_interval"]) + 1)
-        self.assertEqual(len(metrics["val"]), desired_n_evals)
-        self.assertEqual(len(metrics["train"]), 0)  # doesn't get logged here
-        return model
-
-
-class TestBartExamples(unittest.TestCase):
-    @classmethod
-    def setUpClass(cls):
-        stream_handler = logging.StreamHandler(sys.stdout)
-        logger.addHandler(stream_handler)
-        logging.disable(logging.CRITICAL)  # remove noisy download output from tracebacks
-        return cls
-
-    def test_bart_cnn_cli(self):
-        tmp = Path(tempfile.gettempdir()) / "utest_generations_bart_sum.hypo"
-        output_file_name = Path(tempfile.gettempdir()) / "utest_output_bart_sum.hypo"
-        articles = [" New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County."]
-        _dump_articles(tmp, articles)
-        testargs = ["run_eval.py", str(tmp), str(output_file_name), "sshleifer/bart-tiny-random"]
-        with patch.object(sys, "argv", testargs):
-            run_generate()
-            self.assertTrue(Path(output_file_name).exists())
-            os.remove(Path(output_file_name))
-
-    def test_t5_run_sum_cli(self):
-        args_d: dict = CHEAP_ARGS.copy()
-
-        tmp_dir = make_test_data_dir()
-        output_dir = tempfile.mkdtemp(prefix="output_")
-        args_d.update(
-            data_dir=tmp_dir,
-            model_name_or_path=T5_TINY,
-            tokenizer_name=None,  # T5_TINY,
-            train_batch_size=2,
-            eval_batch_size=2,
-            gpus=0,
-            output_dir=output_dir,
-            do_predict=True,
-        )
-        assert "n_train" in args_d
-        args = argparse.Namespace(**args_d)
-        main(args)
-
-    def test_bart_summarization_dataset(self):
-        tmp_dir = Path(tempfile.gettempdir())
-        articles = [" Sam ate lunch today", "Sams lunch ingredients"]
-        summaries = ["A very interesting story about what I ate for lunch.", "Avocado, celery, turkey, coffee"]
-        _dump_articles((tmp_dir / "train.source"), articles)
-        _dump_articles((tmp_dir / "train.target"), summaries)
-        tokenizer = BartTokenizer.from_pretrained("facebook/bart-large")
-        max_len_source = max(len(tokenizer.encode(a)) for a in articles)
-        max_len_target = max(len(tokenizer.encode(a)) for a in summaries)
-        trunc_target = 4
-        train_dataset = SummarizationDataset(
-            tokenizer, data_dir=tmp_dir, type_path="train", max_source_length=20, max_target_length=trunc_target,
-        )
-        dataloader = DataLoader(train_dataset, batch_size=2, collate_fn=train_dataset.collate_fn)
-        for batch in dataloader:
-            self.assertEqual(batch["attention_mask"].shape, batch["input_ids"].shape)
-            # show that articles were trimmed.
-            self.assertEqual(batch["input_ids"].shape[1], max_len_source)
-            self.assertGreater(20, batch["input_ids"].shape[1])  # trimmed significantly
-
-            # show that targets were truncated
-            self.assertEqual(batch["decoder_input_ids"].shape[1], trunc_target)  # Truncated
-            self.assertGreater(max_len_target, trunc_target)  # Truncated
-
-
-def list_to_text_file(lst, path):
-    dest = Path(path)
-    dest.open("w+").writelines(lst)

examples/text-classification/README.md

@@ -52,9 +52,12 @@ Some of these results are significantly different from the ones reported on the
 of GLUE benchmark on the website. For QQP and WNLI, please refer to [FAQ #12](https://gluebenchmark.com/faq) on the webite.
 
 Before running any one of these GLUE tasks you should download the
-[GLUE data](https://gluebenchmark.com/tasks) by running
-[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
-and unpack it to some directory `$GLUE_DIR`.
+[GLUE data](https://gluebenchmark.com/tasks) by running the following lines at the root of the repo
+```
+python utils/download_glue_data.py --data_dir /path/to/glue --tasks all
+```
+
+after replacing *path/to/glue* with a value that you like. Then you can run
 
 ```bash
 export GLUE_DIR=/path/to/glue

examples/text-classification/run_tf_glue.py

@@ -131,9 +131,9 @@ def main():
         level=logging.INFO,
     )
     logger.info(
-        "n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        training_args.n_gpu,
-        bool(training_args.n_gpu > 1),
+        "n_replicas: %s, distributed training: %s, 16-bits training: %s",
+        training_args.n_replicas,
+        bool(training_args.n_replicas > 1),
         training_args.fp16,
     )
     logger.info("Training/evaluation parameters %s", training_args)

examples/text-classification/run_xnli.py

@@ -573,10 +573,6 @@ def main():
 
     # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
     if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        # Create output directory if needed
-        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
-            os.makedirs(args.output_dir)
-
         logger.info("Saving model checkpoint to %s", args.output_dir)
         # Save a trained model, configuration and tokenizer using `save_pretrained()`.
         # They can then be reloaded using `from_pretrained()`

examples/text-generation/run_generation.py

@@ -214,8 +214,14 @@ def main():
     if requires_preprocessing:
         prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
         preprocessed_prompt_text = prepare_input(args, model, tokenizer, prompt_text)
+
+        if model.__class__.__name__ in ["TransfoXLLMHeadModel"]:
+            tokenizer_kwargs = {"add_space_before_punct_symbol": True}
+        else:
+            tokenizer_kwargs = {}
+
         encoded_prompt = tokenizer.encode(
-            preprocessed_prompt_text, add_special_tokens=False, return_tensors="pt", add_space_before_punct_symbol=True
+            preprocessed_prompt_text, add_special_tokens=False, return_tensors="pt", **tokenizer_kwargs
         )
     else:
         encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors="pt")

examples/token-classification/run_ner.py

@@ -75,7 +75,8 @@ class DataTrainingArguments:
         metadata={"help": "The input data dir. Should contain the .txt files for a CoNLL-2003-formatted task."}
     )
     labels: Optional[str] = field(
-        metadata={"help": "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used."}
+        default=None,
+        metadata={"help": "Path to a file containing all labels. If not specified, CoNLL-2003 labels are used."},
     )
     max_seq_length: int = field(
         default=128,

examples/token-classification/run_tf_ner.py

@@ -17,6 +17,7 @@
 
 import logging
 import os
+import warnings
 from dataclasses import dataclass, field
 from typing import Dict, List, Optional, Tuple
 
@@ -109,9 +110,9 @@ def main():
         level=logging.INFO,
     )
     logger.info(
-        "n_gpu: %s, distributed training: %s, 16-bits training: %s",
-        training_args.n_gpu,
-        bool(training_args.n_gpu > 1),
+        "n_replicas: %s, distributed training: %s, 16-bits training: %s",
+        training_args.n_replicas,
+        bool(training_args.n_replicas > 1),
         training_args.fp16,
     )
     logger.info("Training/evaluation parameters %s", training_args)
@@ -184,7 +185,12 @@ def main():
 
         for i in range(batch_size):
             for j in range(seq_len):
-                if label_ids[i, j] != -1:
+                if label_ids[i, j] == -1:
+                    label_ids[i, j] = -100
+                    warnings.warn(
+                        "Using `-1` to mask the loss for the token is depreciated. Please use `-100` instead."
+                    )
+                if label_ids[i, j] != -100:
                     out_label_list[i].append(label_map[label_ids[i][j]])
                     preds_list[i].append(label_map[preds[i][j]])
 

examples/translation/t5/README.md

@@ -1,51 +0,0 @@
-***This script evaluates the multitask pre-trained checkpoint for ``t5-base`` (see paper [here](https://arxiv.org/pdf/1910.10683.pdf)) on the English to German WMT dataset. Please note that the results in the paper were attained using a model fine-tuned on translation, so that results will be worse here by approx. 1.5 BLEU points***
-
-### Intro
-
-This example shows how T5 (here the official [paper](https://arxiv.org/abs/1910.10683)) can be
-evaluated on the WMT English-German dataset.
-
-### Get the WMT Data
-
-To be able to reproduce the authors' results on WMT English to German, you first need to download 
-the WMT14 en-de news datasets.
-Go on Stanford's official NLP [website](https://nlp.stanford.edu/projects/nmt/) and find "newstest2014.en" and "newstest2014.de" under WMT'14 English-German data or download the dataset directly via:
-
-```bash
-curl https://nlp.stanford.edu/projects/nmt/data/wmt14.en-de/newstest2014.en > newstest2014.en
-curl https://nlp.stanford.edu/projects/nmt/data/wmt14.en-de/newstest2014.de > newstest2014.de
-```
-
-You should have 2737 sentences in each file. You can verify this by running:
-
-```bash
-wc -l newstest2014.en  # should give 2737
-```
-
-### Usage
-
-Let's check the longest and shortest sentence in our file to find reasonable decoding hyperparameters: 
-
-Get the longest and shortest sentence:
-
-```bash 
-awk '{print NF}' newstest2014.en | sort -n | head -1 # shortest sentence has 2 word
-awk '{print NF}' newstest2014.en | sort -n | tail -1 # longest sentence has 91 words
-```
-
-We will set our `max_length` to ~3 times the longest sentence and leave `min_length` to its default value of 0.
-We decode with beam search `num_beams=4` as proposed in the paper. Also as is common in beam search we set `early_stopping=True` and `length_penalty=2.0`.
-
-To create translation for each in dataset and get a final BLEU score, run:
-```bash
-python evaluate_wmt.py <path_to_newstest2014.en> newstest2014_de_translations.txt <path_to_newstest2014.de> newsstest2014_en_de_bleu.txt
-```
-the default batch size, 16, fits in 16GB GPU memory, but may need to be adjusted to fit your system.
-
-### Where is the code?
-The core model is in `src/transformers/modeling_t5.py`. This directory only contains examples.
-
-### BLEU Scores
-
-The BLEU score is calculated using [sacrebleu](https://github.com/mjpost/sacreBLEU) by mjpost.
-To get the BLEU score we used 

examples/translation/t5/evaluate_wmt.py

@@ -1,103 +0,0 @@
-import argparse
-from pathlib import Path
-
-import torch
-from sacrebleu import corpus_bleu
-from tqdm import tqdm
-
-from transformers import T5ForConditionalGeneration, T5Tokenizer
-
-
-def chunks(lst, n):
-    """Yield successive n-sized chunks from lst."""
-    for i in range(0, len(lst), n):
-        yield lst[i : i + n]
-
-
-def generate_translations(lns, output_file_path, model_size, batch_size, device):
-    model = T5ForConditionalGeneration.from_pretrained(model_size)
-    model.to(device)
-
-    tokenizer = T5Tokenizer.from_pretrained(model_size)
-
-    # update config with summarization specific params
-    task_specific_params = model.config.task_specific_params
-    if task_specific_params is not None:
-        model.config.update(task_specific_params.get("translation_en_to_de", {}))
-
-    with Path(output_file_path).open("w") as output_file:
-        for batch in tqdm(list(chunks(lns, batch_size))):
-            batch = [model.config.prefix + text for text in batch]
-
-            dct = tokenizer.batch_encode_plus(batch, max_length=512, return_tensors="pt", pad_to_max_length=True)
-
-            input_ids = dct["input_ids"].to(device)
-            attention_mask = dct["attention_mask"].to(device)
-
-            translations = model.generate(input_ids=input_ids, attention_mask=attention_mask)
-            dec = [
-                tokenizer.decode(g, skip_special_tokens=True, clean_up_tokenization_spaces=False) for g in translations
-            ]
-
-            for hypothesis in dec:
-                output_file.write(hypothesis + "\n")
-
-
-def calculate_bleu_score(output_lns, refs_lns, score_path):
-    bleu = corpus_bleu(output_lns, [refs_lns])
-    result = "BLEU score: {}".format(bleu.score)
-    with Path(score_path).open("w") as score_file:
-        score_file.write(result)
-
-
-def run_generate():
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "model_size",
-        type=str,
-        help="T5 model size, either 't5-small', 't5-base', 't5-large', 't5-3b', 't5-11b'. Defaults to 't5-base'.",
-        default="t5-base",
-    )
-    parser.add_argument(
-        "input_path", type=str, help="like wmt/newstest2014.en",
-    )
-    parser.add_argument(
-        "output_path", type=str, help="where to save translation",
-    )
-    parser.add_argument(
-        "reference_path", type=str, help="like wmt/newstest2014.de",
-    )
-    parser.add_argument(
-        "score_path", type=str, help="where to save the bleu score",
-    )
-    parser.add_argument(
-        "--batch_size", type=int, default=16, required=False, help="batch size: how many to summarize at a time",
-    )
-    parser.add_argument(
-        "--no_cuda", default=False, type=bool, help="Whether to force the execution on CPU.",
-    )
-
-    args = parser.parse_args()
-    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-
-    dash_pattern = (" ##AT##-##AT## ", "-")
-
-    # Read input lines into python
-    with open(args.input_path, "r") as input_file:
-        input_lns = [x.strip().replace(dash_pattern[0], dash_pattern[1]) for x in input_file.readlines()]
-
-    generate_translations(input_lns, args.output_path, args.model_size, args.batch_size, args.device)
-
-    # Read generated lines into python
-    with open(args.output_path, "r") as output_file:
-        output_lns = [x.strip() for x in output_file.readlines()]
-
-    # Read reference lines into python
-    with open(args.reference_path, "r") as reference_file:
-        refs_lns = [x.strip().replace(dash_pattern[0], dash_pattern[1]) for x in reference_file.readlines()]
-
-    calculate_bleu_score(output_lns, refs_lns, args.score_path)
-
-
-if __name__ == "__main__":
-    run_generate()

examples/translation/t5/test_t5_examples.py

@@ -1,50 +0,0 @@
-import logging
-import sys
-import tempfile
-import unittest
-from pathlib import Path
-from unittest.mock import patch
-
-from .evaluate_wmt import run_generate
-
-
-text = ["When Liana Barrientos was 23 years old, she got married in Westchester County."]
-translation = ["Als Liana Barrientos 23 Jahre alt war, heiratete sie in Westchester County."]
-
-output_file_name = "output_t5_trans.txt"
-score_file_name = "score_t5_trans.txt"
-
-logging.basicConfig(level=logging.DEBUG)
-
-logger = logging.getLogger()
-
-
-class TestT5Examples(unittest.TestCase):
-    def test_t5_cli(self):
-        stream_handler = logging.StreamHandler(sys.stdout)
-        logger.addHandler(stream_handler)
-
-        tmp_source = Path(tempfile.gettempdir()) / "utest_generations_t5_trans.hypo"
-        with tmp_source.open("w") as f:
-            f.write("\n".join(text))
-
-        tmp_target = Path(tempfile.gettempdir()) / "utest_generations_t5_trans.target"
-        with tmp_target.open("w") as f:
-            f.write("\n".join(translation))
-
-        output_file_name = Path(tempfile.gettempdir()) / "utest_output_trans.hypo"
-        score_file_name = Path(tempfile.gettempdir()) / "utest_score.hypo"
-
-        testargs = [
-            "evaluate_wmt.py",
-            "patrickvonplaten/t5-tiny-random",
-            str(tmp_source),
-            str(output_file_name),
-            str(tmp_target),
-            str(score_file_name),
-        ]
-
-        with patch.object(sys, "argv", testargs):
-            run_generate()
-            self.assertTrue(Path(output_file_name).exists())
-            self.assertTrue(Path(score_file_name).exists())

model_cards/Hate-speech-CNERG/dehatebert-mono-arabic/README.md

@@ -1,2 +1,2 @@
 This model is used detecting **hatespeech** in **Arabic language**. The mono in the name refers to the monolingual setting, where the model is trained using only Arabic language data. It is finetuned on multilingual bert model.
-The model is trained with different learning rates and the best validation score achieved is 0.8674776 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)
+The model is trained with different learning rates and the best validation score achieved is 0.877609 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-english/README.md

@@ -1,2 +1,2 @@
 This model is used detecting **hatespeech** in **English language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
-The model is trained with different learning rates and the best validation score achieved is 0.7069374 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)
+The model is trained with different learning rates and the best validation score achieved is 0.726030 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-french/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **French language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.692094 for a learning rate of 3e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-german/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **German language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.649794 for a learning rate of 3e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-indonesian/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **Indonesian language**. The mono in the name refers to the monolingual setting, where the model is trained using only Arabic language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.844494 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-italian/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **Italian language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.837288 for a learning rate of 3e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-polish/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **Polish language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.723254 for a learning rate of 2e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)

model_cards/Hate-speech-CNERG/dehatebert-mono-portugese/README.md

@@ -0,0 +1,2 @@
+This model is used detecting **hatespeech** in **Portuguese language**. The mono in the name refers to the monolingual setting, where the model is trained using only English language data. It is finetuned on multilingual bert model.
+The model is trained with different learning rates and the best validation score achieved is 0.716119 for a learning rate of 3e-5. Training code can be found at this [url](https://github.com/punyajoy/DE-LIMIT)