init commit

CODE_OF_CONDUCT.md

@@ -0,0 +1,4 @@
+## Code of Conduct
+This project has adopted the [Amazon Open Source Code of Conduct](https://aws.github.io/code-of-conduct).
+For more information see the [Code of Conduct FAQ](https://aws.github.io/code-of-conduct-faq) or contact
+opensource-codeofconduct@amazon.com with any additional questions or comments.

CONTRIBUTING.md

@@ -0,0 +1,59 @@
+# Contributing Guidelines
+
+Thank you for your interest in contributing to our project. Whether it's a bug report, new feature, correction, or additional
+documentation, we greatly value feedback and contributions from our community.
+
+Please read through this document before submitting any issues or pull requests to ensure we have all the necessary
+information to effectively respond to your bug report or contribution.
+
+
+## Reporting Bugs/Feature Requests
+
+We welcome you to use the GitHub issue tracker to report bugs or suggest features.
+
+When filing an issue, please check existing open, or recently closed, issues to make sure somebody else hasn't already
+reported the issue. Please try to include as much information as you can. Details like these are incredibly useful:
+
+* A reproducible test case or series of steps
+* The version of our code being used
+* Any modifications you've made relevant to the bug
+* Anything unusual about your environment or deployment
+
+
+## Contributing via Pull Requests
+Contributions via pull requests are much appreciated. Before sending us a pull request, please ensure that:
+
+1. You are working against the latest source on the *main* branch.
+2. You check existing open, and recently merged, pull requests to make sure someone else hasn't addressed the problem already.
+3. You open an issue to discuss any significant work - we would hate for your time to be wasted.
+
+To send us a pull request, please:
+
+1. Fork the repository.
+2. Modify the source; please focus on the specific change you are contributing. If you also reformat all the code, it will be hard for us to focus on your change.
+3. Ensure local tests pass.
+4. Commit to your fork using clear commit messages.
+5. Send us a pull request, answering any default questions in the pull request interface.
+6. Pay attention to any automated CI failures reported in the pull request, and stay involved in the conversation.
+
+GitHub provides additional document on [forking a repository](https://help.github.com/articles/fork-a-repo/) and
+[creating a pull request](https://help.github.com/articles/creating-a-pull-request/).
+
+
+## Finding contributions to work on
+Looking at the existing issues is a great way to find something to contribute on. As our projects, by default, use the default GitHub issue labels (enhancement/bug/duplicate/help wanted/invalid/question/wontfix), looking at any 'help wanted' issues is a great place to start.
+
+
+## Code of Conduct
+This project has adopted the [Amazon Open Source Code of Conduct](https://aws.github.io/code-of-conduct).
+For more information see the [Code of Conduct FAQ](https://aws.github.io/code-of-conduct-faq) or contact
+opensource-codeofconduct@amazon.com with any additional questions or comments.
+
+
+## Security issue notifications
+If you discover a potential security issue in this project we ask that you notify AWS/Amazon Security via our [vulnerability reporting page](http://aws.amazon.com/security/vulnerability-reporting/). Please do **not** create a public github issue.
+
+
+## Licensing
+
+See the [LICENSE](LICENSE) file for our project's licensing. We will ask you to confirm the licensing of your contribution.

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 1999-2022 Alibaba Group Holding Ltd.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1 @@
+Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.

README.md

@@ -1,13 +1,185 @@
----
-title: PolyFormer
-emoji: 🔥
-colorFrom: blue
-colorTo: blue
-sdk: gradio
-sdk_version: 3.29.0
-app_file: app.py
-pinned: false
-license: apache-2.0
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# PolyFormer: Referring Image Segmentation as Sequential Polygon Generation (CVPR 2023)
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/polyformer-referring-image-segmentation-as/referring-expression-segmentation-on-refcocog)](https://paperswithcode.com/sota/referring-expression-segmentation-on-refcocog?p=polyformer-referring-image-segmentation-as)
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/polyformer-referring-image-segmentation-as/referring-expression-segmentation-on-refcoco)](https://paperswithcode.com/sota/referring-expression-segmentation-on-refcoco?p=polyformer-referring-image-segmentation-as)
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/polyformer-referring-image-segmentation-as/referring-expression-segmentation-on-refcoco-1)](https://paperswithcode.com/sota/referring-expression-segmentation-on-refcoco-1?p=polyformer-referring-image-segmentation-as)
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/polyformer-referring-image-segmentation-as/referring-expression-comprehension-on-refcoco)](https://paperswithcode.com/sota/referring-expression-comprehension-on-refcoco?p=polyformer-referring-image-segmentation-as)
+[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/polyformer-referring-image-segmentation-as/referring-expression-comprehension-on-refcoco-1)](https://paperswithcode.com/sota/referring-expression-comprehension-on-refcoco-1?p=polyformer-referring-image-segmentation-as)
+
+
+\[[Project Page](https://polyformer.github.io/)\]   \[[Paper](https://arxiv.org/abs/2302.07387)\]   
+
+by [Jiang Liu*](https://joellliu.github.io/), [Hui Ding*](http://www.huiding.org/), [Zhaowei Cai](https://zhaoweicai.github.io/),  [Yuting Zhang](https://scholar.google.com/citations?user=9UfZJskAAAAJ&hl=en), [Ravi Kumar Satzoda](https://scholar.google.com.sg/citations?user=4ngycwIAAAAJ&hl=en), [Vijay Mahadevan](https://scholar.google.com/citations?user=n9fRgvkAAAAJ&hl=en), [R. Manmatha](https://ciir.cs.umass.edu/~manmatha/).
+
+
+## :notes: Introduction
+![github_figure](pipeline.gif)
+PolyFormer is a unified model for referring image segmentation (polygon vertex sequence) and referring expression comprehension (bounding box corner points). The polygons are converted to segmentation masks in the end.
+
+**Contributions:**
+
+* State-of-the-art results on referring image segmentation and referring expression comprehension on 6 datasets; 
+* A unified framework for referring image segmentation (RIS) and referring expression comprehension (REC) by formulating them as a sequence-to-sequence (seq2seq) prediction problem; 
+* A regression-based decoder for accurate coordinate prediction, which outputs continuous 2D coordinates directly without quantization error..
+
+
+
+## Getting Started
+### Installation
+```bash
+conda create -n polyformer python=3.7.4
+conda activate polyformer
+python -m pip install -r requirements.txt
+```
+Note: if you are getting import errors from `fairseq`, try the following:
+```bash
+python -m pip install pip==21.2.4
+pip uninstall fairseq
+pip install -r requirements.txt
+```
+
+## Datasets 
+### Prepare Pretraining Data
+1. Create the dataset folders
+```bash
+mkdir datasets
+mkdir datasets/images
+mkdir datasets/annotations
+```
+2. Download the *2014 Train images [83K/13GB]* from [COCO](https://cocodataset.org/#download), 
+original [Flickr30K images](http://shannon.cs.illinois.edu/DenotationGraph/),
+[ReferItGame images](https://drive.google.com/file/d/1R6Tm7tQTHCil6A_eOhjudK3rgaBxkD2t/view?usp=sharing), 
+and [Visual Genome images](http://visualgenome.org/api/v0/api_home.html), and extract them to `datasets/images`. 
+3. Download the annotation file for pretraining datasets [instances.json](https://drive.google.com/drive/folders/1O4hzL8_s3aUsnj_JZnM3CwANd7TejcJO) 
+provided by [SeqTR](https://github.com/sean-zhuh/SeqTR) and store it in `datasets/annotations`. 
+The workspace directory should be organized like this:
+```
+PolyFormer/
+├── datasets/
+│   ├── images
+│   │   ├── flickr30k/*.jpg
+│   │   ├── mscoco/
+│   │   │   └── train2014/*.jpg
+│   │   ├── saiaprtc12/*.jpg
+│   │   └── visual-genome/*.jpg
+│   └── annotations
+│       └── instances.json
+└── ...
+```
+4. Generate the tsv files for pretraining
+```bash
+python data/create_pretraining_data.py
+```
+### Prepare Finetuning Data
+1. Follow the instructions in the `./refer` directory to set up subdirectories
+and download annotations.
+This directory is based on the [refer](https://github.com/lichengunc/refer) API.
+
+2. Generate the tsv files for finetuning
+```bash
+python data/create_finetuning_data.py
+```
+
+
+
+
+## Pretraining
+1. Create the checkpoints folder
+```bash
+mkdir weights
+```
+2. Download pretrain weights of [Swin-base](https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window12_384_22k.pth),
+[Swin-large](https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window12_384_22k.pth),
+[BERT-base](https://cdn.huggingface.co/bert-base-uncased-pytorch_model.bin)
+and put the weight files in `./pretrained_weights`.
+These weights are needed for training to initialize the model.
+
+
+3. Run the pretraining scripts for model pretraining on the referring expression comprehension task:
+```bash
+cd run_scripts/pretrain
+bash pretrain_polyformer_b.sh  # for pretraining PolyFormer-B model
+bash pretrain_polyformer_l.sh  # for pretraining PolyFormer-L model
+```
+
+## Finetuning
+Run the finetuning scripts for model pretraining on the referring image segmentation and referring expression comprehension tasks:
+```bash
+cd run_scripts/finetune
+bash train_polyformer_b.sh  # for finetuning PolyFormer-B model
+bash train_polyformer_l.sh  # for finetuning PolyFormer-L model
+```
+Please make sure to link the pretrain weight paths (Line 20) in the finetuning scripts to the best pretraining checkpoints. 
+
+## Evaluation
+Run the evaluation scripts for evaluating on the referring image segmentation and referring expression comprehension tasks:
+```bash
+cd run_scripts/evaluation
+
+# for evaluating PolyFormer-B model
+bash evaluate_polyformer_b_refcoco.sh 
+bash evaluate_polyformer_b_refcoco+.sh 
+bash evaluate_polyformer_b_refcocog.sh 
+
+# for evaluating PolyFormer-L model
+bash evaluate_polyformer_l_refcoco.sh 
+bash evaluate_polyformer_l_refcoco+.sh 
+bash evaluate_polyformer_l_refcocog.sh 
+```
+
+## Model Zoo
+Download the model weights to `./weights` if you want to use our trained models for finetuning and evaluation.
+
+|                                                                                                       | Refcoco val|           | | Refcoco testA|       |       | Refcoco testB| ||
+|-------------------------------------------------------------------------------------------------------|------|------|---------|------|-------|------|-----|------|------|
+| Model                                                                                                 | oIoU | mIoU | Prec@0.5 | oIoU | mIoU  |Prec@0.5 | oIoU | mIoU  |Prec@0.5  | 
+| [PolyFormer-B](https://drive.google.com/file/d/1K0y-WBO6cL7gBzNnJaHAeNu3pgq4DbJ9/view?usp=share_link) | 74.82| 75.96 | 89.73   |76.64| 77.09 | 91.73| 71.06| 73.22 | 86.03 | 
+| [PolyFormer-L](https://drive.google.com/file/d/15P6m5RI6HAQE2QXQXMAjw_oBsaPii7b3/view?usp=share_link) | 75.96| 76.94 | 90.38   |78.29| 78.49 | 92.89| 73.25| 74.83 |  87.16| 
+
+
+|                                  [test_demo.py](..%2F..%2FDownloads%2Ftest_demo.py)                                                                     | Refcoco val|           | | Refcoco testA|       |       | Refcoco testB| ||
+|--------------------------------------------------------------------------------------------------------|------|------|------|------|------|------|------|------|------|
+| Model                                                                                                  | oIoU | mIoU |Prec@0.5| oIoU | mIoU  |Prec@0.5 | oIoU | mIoU  |Prec@0.5  | 
+| [PolyFormer-B ](https://drive.google.com/file/d/12_ylFhsbqGySxDqgeEByn8nKoJtT2n2w/view?usp=share_link) |  67.64| 70.65 | 83.73 | 72.89| 74.51 | 88.60 | 59.33| 64.64 | 76.38 | 67.76| 69.36  | 
+| [PolyFormer-L](https://drive.google.com/file/d/1lUCv7dUPctEz4vEpPr7aI8A8ZmfYCB8y/view?usp=share_link)  |  69.33| 72.15 | 84.98 | 74.56| 75.71 | 89.77 | 61.87| 66.73 | 77.97 | 69.20| 71.15 | 
+
+
+|                                                                                                       |  Refcocog val| || | Refcocog test| | 
+|-------------------------------------------------------------------------------------------------------|------|------|------|------|------|------|
+| Model                                                                                                 |  oIoU | mIoU   |Prec@0.5 | oIoU | mIoU   |Prec@0.5 |
+| [PolyFormer-B](https://drive.google.com/file/d/12_ylFhsbqGySxDqgeEByn8nKoJtT2n2w/view?usp=share_link) |  67.76| 69.36  | 84.46| 69.05| 69.88 | 84.96 | 
+| [PolyFormer-L](https://drive.google.com/file/d/1lUCv7dUPctEz4vEpPr7aI8A8ZmfYCB8y/view?usp=share_link) |  69.20| 71.15 | 85.83 | 70.19| 71.17  | 85.91| 
+
+* Pretrained weights:
+  * [PolyFormer-B](https://drive.google.com/file/d/1sAzfChYDdHdaeatB2K14lrJjG4uiXAol/view?usp=share_link)
+  * [PolyFormer-L](https://drive.google.com/file/d/1knRxgM1lmEkuZZ-cOm_fmwKP1H0bJGU9/view?usp=share_link)
+
+# Acknowlegement
+This codebase is developed based on [OFA](https://github.com/OFA-Sys/OFA). 
+Other related codebases include:
+* [Fairseq](https://github.com/pytorch/fairseq)
+* [refer](https://github.com/lichengunc/refer)
+* [LAVT-RIS](https://github.com/yz93/LAVT-RIS/)
+* [SeqTR](https://github.com/sean-zhuh/SeqTR)
+
+
+
+# Citation
+Please cite our paper if you find this codebase helpful :)
+
+```
+@inproceedings{liu2023polyformer,
+  title={PolyFormer: Referring Image Segmentation as Sequential Polygon Generation},
+  author={Liu, Jiang and Ding, Hui and Cai, Zhaowei and Zhang, Yuting and Satzoda, Ravi Kumar and Mahadevan, Vijay and Manmatha, R},
+  booktitle={CVPR},
+  year={2023}
+}
+```
+
+## Security
+
+See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for more information.
+
+## License
+
+This project is licensed under the Apache-2.0 License.
+

app.py

@@ -0,0 +1,47 @@
+import os
+import torch
+import numpy as np
+from fairseq import utils,tasks
+from utils.checkpoint_utils import load_model_ensemble_and_task
+from models.polyformer import PolyFormerModel
+import cv2
+
+import torch
+import numpy as np
+from fairseq import utils, tasks
+from fairseq import checkpoint_utils
+from utils.eval_utils import eval_step
+from tasks.refcoco import RefcocoTask
+from models.polyformer import PolyFormerModel
+from PIL import Image
+from torchvision import transforms
+import cv2
+import gradio as gr
+import math
+from io import BytesIO
+import base64
+import re
+from demo import visual_grounding
+
+title = "PolyFormer-Visual_Grounding"
+description = "Gradio Demo for PolyFormer-Visual_Grounding. Upload your own image or click any one of the examples, " \
+              "and write a description about a certain object.  " \
+              "Then click \"Submit\" and wait for the result of grounding. For help or to provide feedback, please contact: Hui Ding (@huidin)"
+article = "<p style='text-align: center'><a href='https://arxiv.org/pdf/2302.07387.pdf' target='_blank'>PolyFormer CVPR2023" \
+          "</a></p> "
+# examples = [['A bear astronaut in the space.jpeg', 'a bear astronaut in the space'],
+#             ['A unicorn doing computer vision research.jpeg', 'a unicorn doing computer vision research'],
+#             ['pig.jpeg', 'a pig robot preparing a delicious meal'],
+#             ['otta.png', 'a gentleman otter in a 19th century portrait'],
+#             ['pikachu.jpeg', 'a pikachu fine-dining  with  a view  to  the  Eiffel Tower'],
+#             ['A small cabin on top of a snowy mountain in the style of Disney artstation.jpeg', 'a small cabin on top of a snowy mountain in the style of Disney artstation'],
+#
+#             ]
+examples = []
+io = gr.Interface(fn=visual_grounding, inputs=[gr.inputs.Image(type='pil'), "textbox"],
+                  outputs=[gr.outputs.Image(label="output", type='numpy'), gr.outputs.Image(label="predicted mask", type='numpy')],
+                  title=title, description=description, article=article, examples=examples,
+                  allow_flagging=False, allow_screenshot=False)
+# io.launch(cache_examples=True)
+io.launch(share=True)
+

bert/activations.py

@@ -0,0 +1,56 @@
+import logging
+import math
+
+import torch
+import torch.nn.functional as F
+
+
+logger = logging.getLogger(__name__)
+
+
+def swish(x):
+    return x * torch.sigmoid(x)
+
+
+def _gelu_python(x):
+    """ Original Implementation of the gelu activation function in Google Bert repo when initially created.
+        For information: OpenAI GPT's gelu is slightly different (and gives slightly different results):
+        0.5 * x * (1 + torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3))))
+        This is now written in C in torch.nn.functional
+        Also see https://arxiv.org/abs/1606.08415
+    """
+    return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+
+def gelu_new(x):
+    """ Implementation of the gelu activation function currently in Google Bert repo (identical to OpenAI GPT).
+        Also see https://arxiv.org/abs/1606.08415
+    """
+    return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+
+if torch.__version__ < "1.4.0":
+    gelu = _gelu_python
+else:
+    gelu = F.gelu
+
+
+def gelu_fast(x):
+    return 0.5 * x * (1.0 + torch.tanh(x * 0.7978845608 * (1.0 + 0.044715 * x * x)))
+
+
+ACT2FN = {
+    "relu": F.relu,
+    "swish": swish,
+    "gelu": gelu,
+    "tanh": torch.tanh,
+    "gelu_new": gelu_new,
+    "gelu_fast": gelu_fast,
+}
+
+
+def get_activation(activation_string):
+    if activation_string in ACT2FN:
+        return ACT2FN[activation_string]
+    else:
+        raise KeyError("function {} not found in ACT2FN mapping {}".format(activation_string, list(ACT2FN.keys())))

bert/configuration_bert.py

@@ -0,0 +1,143 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BERT model configuration """
+
+
+import logging
+
+from .configuration_utils import PretrainedConfig
+
+
+logger = logging.getLogger(__name__)
+
+BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "bert-base-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-config.json",
+    "bert-large-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-config.json",
+    "bert-base-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-config.json",
+    "bert-large-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-config.json",
+    "bert-base-multilingual-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-config.json",
+    "bert-base-multilingual-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-config.json",
+    "bert-base-chinese": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-config.json",
+    "bert-base-german-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-cased-config.json",
+    "bert-large-uncased-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-config.json",
+    "bert-large-cased-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-config.json",
+    "bert-large-uncased-whole-word-masking-finetuned-squad": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-config.json",
+    "bert-large-cased-whole-word-masking-finetuned-squad": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-config.json",
+    "bert-base-cased-finetuned-mrpc": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-config.json",
+    "bert-base-german-dbmdz-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-config.json",
+    "bert-base-german-dbmdz-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-config.json",
+    "cl-tohoku/bert-base-japanese": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese/config.json",
+    "cl-tohoku/bert-base-japanese-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking/config.json",
+    "cl-tohoku/bert-base-japanese-char": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char/config.json",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking/config.json",
+    "TurkuNLP/bert-base-finnish-cased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
+    "TurkuNLP/bert-base-finnish-uncased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
+    "wietsedv/bert-base-dutch-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/wietsedv/bert-base-dutch-cased/config.json",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+}
+
+
+class BertConfig(PretrainedConfig):
+    r"""
+        This is the configuration class to store the configuration of a :class:`~transformers.BertModel`.
+        It is used to instantiate an BERT model according to the specified arguments, defining the model
+        architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+        the BERT `bert-base-uncased <https://huggingface.co/bert-base-uncased>`__ architecture.
+
+        Configuration objects inherit from  :class:`~transformers.PretrainedConfig` and can be used
+        to control the model outputs. Read the documentation from  :class:`~transformers.PretrainedConfig`
+        for more information.
+
+
+        Args:
+            vocab_size (:obj:`int`, optional, defaults to 30522):
+                Vocabulary size of the BERT model. Defines the different tokens that
+                can be represented by the `inputs_ids` passed to the forward method of :class:`~transformers.BertModel`.
+            hidden_size (:obj:`int`, optional, defaults to 768):
+                Dimensionality of the encoder layers and the pooler layer.
+            num_hidden_layers (:obj:`int`, optional, defaults to 12):
+                Number of hidden layers in the Transformer encoder.
+            num_attention_heads (:obj:`int`, optional, defaults to 12):
+                Number of attention heads for each attention layer in the Transformer encoder.
+            intermediate_size (:obj:`int`, optional, defaults to 3072):
+                Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+            hidden_act (:obj:`str` or :obj:`function`, optional, defaults to "gelu"):
+                The non-linear activation function (function or string) in the encoder and pooler.
+                If string, "gelu", "relu", "swish" and "gelu_new" are supported.
+            hidden_dropout_prob (:obj:`float`, optional, defaults to 0.1):
+                The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
+                The dropout ratio for the attention probabilities.
+            max_position_embeddings (:obj:`int`, optional, defaults to 512):
+                The maximum sequence length that this model might ever be used with.
+                Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+            type_vocab_size (:obj:`int`, optional, defaults to 2):
+                The vocabulary size of the `token_type_ids` passed into :class:`~transformers.BertModel`.
+            initializer_range (:obj:`float`, optional, defaults to 0.02):
+                The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+            layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
+                The epsilon used by the layer normalization layers.
+            gradient_checkpointing (:obj:`bool`, optional, defaults to False):
+                If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+
+        Example::
+
+            >>> from transformers import BertModel, BertConfig
+
+            >>> # Initializing a BERT bert-base-uncased style configuration
+            >>> configuration = BertConfig()
+
+            >>> # Initializing a model from the bert-base-uncased style configuration
+            >>> model = BertModel(configuration)
+
+            >>> # Accessing the model configuration
+            >>> configuration = model.config
+    """
+    model_type = "bert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        gradient_checkpointing=False,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.gradient_checkpointing = gradient_checkpointing

bert/configuration_utils.py

@@ -0,0 +1,408 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+
+import copy
+import json
+import logging
+import os
+from typing import Dict, Tuple
+
+from .file_utils import CONFIG_NAME, cached_path, hf_bucket_url, is_remote_url
+
+
+logger = logging.getLogger(__name__)
+
+
+class PretrainedConfig(object):
+    r""" Base class for all configuration classes.
+        Handles a few parameters common to all models' configurations as well as methods for loading/downloading/saving configurations.
+
+        Note:
+            A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to initialize a model does **not** load the model weights.
+            It only affects the model's configuration.
+
+        Class attributes (overridden by derived classes):
+            - ``model_type``: a string that identifies the model type, that we serialize into the JSON file, and that we use to recreate the correct object in :class:`~transformers.AutoConfig`.
+
+        Args:
+            finetuning_task (:obj:`string` or :obj:`None`, `optional`, defaults to :obj:`None`):
+                Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow or PyTorch) checkpoint.
+            num_labels (:obj:`int`, `optional`, defaults to `2`):
+                Number of classes to use when the model is a classification model (sequences/tokens)
+            output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Should the model returns all hidden-states.
+            output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Should the model returns all attentions.
+            torchscript (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Is the model used with Torchscript (for PyTorch models).
+    """
+    model_type: str = ""
+
+    def __init__(self, **kwargs):
+        # Attributes with defaults
+        self.output_hidden_states = kwargs.pop("output_hidden_states", False)
+        self.output_attentions = kwargs.pop("output_attentions", False)
+        self.use_cache = kwargs.pop("use_cache", True)  # Not used by all models
+        self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
+        self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
+        self.pruned_heads = kwargs.pop("pruned_heads", {})
+
+        # Is decoder is used in encoder-decoder models to differentiate encoder from decoder
+        self.is_encoder_decoder = kwargs.pop("is_encoder_decoder", False)
+        self.is_decoder = kwargs.pop("is_decoder", False)
+
+        # Parameters for sequence generation
+        self.max_length = kwargs.pop("max_length", 20)
+        self.min_length = kwargs.pop("min_length", 0)
+        self.do_sample = kwargs.pop("do_sample", False)
+        self.early_stopping = kwargs.pop("early_stopping", False)
+        self.num_beams = kwargs.pop("num_beams", 1)
+        self.temperature = kwargs.pop("temperature", 1.0)
+        self.top_k = kwargs.pop("top_k", 50)
+        self.top_p = kwargs.pop("top_p", 1.0)
+        self.repetition_penalty = kwargs.pop("repetition_penalty", 1.0)
+        self.length_penalty = kwargs.pop("length_penalty", 1.0)
+        self.no_repeat_ngram_size = kwargs.pop("no_repeat_ngram_size", 0)
+        self.bad_words_ids = kwargs.pop("bad_words_ids", None)
+        self.num_return_sequences = kwargs.pop("num_return_sequences", 1)
+
+        # Fine-tuning task arguments
+        self.architectures = kwargs.pop("architectures", None)
+        self.finetuning_task = kwargs.pop("finetuning_task", None)
+        self.id2label = kwargs.pop("id2label", None)
+        self.label2id = kwargs.pop("label2id", None)
+        if self.id2label is not None:
+            kwargs.pop("num_labels", None)
+            self.id2label = dict((int(key), value) for key, value in self.id2label.items())
+            # Keys are always strings in JSON so convert ids to int here.
+        else:
+            self.num_labels = kwargs.pop("num_labels", 2)
+
+        # Tokenizer arguments TODO: eventually tokenizer and models should share the same config
+        self.prefix = kwargs.pop("prefix", None)
+        self.bos_token_id = kwargs.pop("bos_token_id", None)
+        self.pad_token_id = kwargs.pop("pad_token_id", None)
+        self.eos_token_id = kwargs.pop("eos_token_id", None)
+        self.decoder_start_token_id = kwargs.pop("decoder_start_token_id", None)
+
+        # task specific arguments
+        self.task_specific_params = kwargs.pop("task_specific_params", None)
+
+        # TPU arguments
+        self.xla_device = kwargs.pop("xla_device", None)
+
+        # Additional attributes without default values
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
+    @property
+    def num_labels(self):
+        return len(self.id2label)
+
+    @num_labels.setter
+    def num_labels(self, num_labels):
+        self.id2label = {i: "LABEL_{}".format(i) for i in range(num_labels)}
+        self.label2id = dict(zip(self.id2label.values(), self.id2label.keys()))
+
+    def save_pretrained(self, save_directory):
+        """
+        Save a configuration object to the directory `save_directory`, so that it
+        can be re-loaded using the :func:`~transformers.PretrainedConfig.from_pretrained` class method.
+
+        Args:
+            save_directory (:obj:`string`):
+                Directory where the configuration JSON file will be saved.
+        """
+        if os.path.isfile(save_directory):
+            raise AssertionError("Provided path ({}) should be a directory, not a file".format(save_directory))
+        os.makedirs(save_directory, exist_ok=True)
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_config_file = os.path.join(save_directory, CONFIG_NAME)
+
+        self.to_json_file(output_config_file, use_diff=True)
+        logger.info("Configuration saved in {}".format(output_config_file))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs) -> "PretrainedConfig":
+        r"""
+
+        Instantiate a :class:`~transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.
+
+        Args:
+            pretrained_model_name_or_path (:obj:`string`):
+                either:
+                  - a string with the `shortcut name` of a pre-trained model configuration to load from cache or
+                    download, e.g.: ``bert-base-uncased``.
+                  - a string with the `identifier name` of a pre-trained model configuration that was user-uploaded to
+                    our S3, e.g.: ``dbmdz/bert-base-german-cased``.
+                  - a path to a `directory` containing a configuration file saved using the
+                    :func:`~transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                  - a path or url to a saved configuration JSON `file`, e.g.:
+                    ``./my_model_directory/configuration.json``.
+            cache_dir (:obj:`string`, `optional`):
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+            kwargs (:obj:`Dict[str, any]`, `optional`):
+                The values in kwargs of any keys which are configuration attributes will be used to override the loaded
+                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is
+                controlled by the `return_unused_kwargs` keyword parameter.
+            force_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exist.
+            resume_download (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+            proxies (:obj:`Dict`, `optional`):
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.:
+                :obj:`{'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.`
+                The proxies are used on each request.
+            return_unused_kwargs: (`optional`) bool:
+                If False, then this function returns just the final configuration object.
+                If True, then this functions returns a :obj:`Tuple(config, unused_kwargs)` where `unused_kwargs` is a
+                dictionary consisting of the key/value pairs whose keys are not configuration attributes: ie the part
+                of kwargs which has not been used to update `config` and is otherwise ignored.
+
+        Returns:
+            :class:`PretrainedConfig`: An instance of a configuration object
+
+        Examples::
+
+            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
+            # derived class: BertConfig
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
+            config = BertConfig.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
+            assert config.output_attention == True
+            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attention=True,
+                                                               foo=False, return_unused_kwargs=True)
+            assert config.output_attention == True
+            assert unused_kwargs == {'foo': False}
+
+        """
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+        return cls.from_dict(config_dict, **kwargs)
+
+    @classmethod
+    def get_config_dict(cls, pretrained_model_name_or_path: str, **kwargs) -> Tuple[Dict, Dict]:
+        """
+        From a `pretrained_model_name_or_path`, resolve to a dictionary of parameters, to be used
+        for instantiating a Config using `from_dict`.
+
+        Parameters:
+            pretrained_model_name_or_path (:obj:`string`):
+                The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.
+
+        Returns:
+            :obj:`Tuple[Dict, Dict]`: The dictionary that will be used to instantiate the configuration object.
+
+        """
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", False)
+
+        if os.path.isdir(pretrained_model_name_or_path):
+            config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME)
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+            config_file = pretrained_model_name_or_path
+        else:
+            config_file = hf_bucket_url(pretrained_model_name_or_path, filename=CONFIG_NAME, use_cdn=False)
+
+        try:
+            # Load from URL or cache if already cached
+            resolved_config_file = cached_path(
+                config_file,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                local_files_only=local_files_only,
+            )
+            # Load config dict
+            if resolved_config_file is None:
+                raise EnvironmentError
+            config_dict = cls._dict_from_json_file(resolved_config_file)
+
+        except EnvironmentError:
+            msg = (
+                f"Can't load config for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
+                f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a {CONFIG_NAME} file\n\n"
+            )
+            raise EnvironmentError(msg)
+
+        except json.JSONDecodeError:
+            msg = (
+                "Couldn't reach server at '{}' to download configuration file or "
+                "configuration file is not a valid JSON file. "
+                "Please check network or file content here: {}.".format(config_file, resolved_config_file)
+            )
+            raise EnvironmentError(msg)
+
+        if resolved_config_file == config_file:
+            logger.info("loading configuration file {}".format(config_file))
+        else:
+            logger.info("loading configuration file {} from cache at {}".format(config_file, resolved_config_file))
+
+        return config_dict, kwargs
+
+    @classmethod
+    def from_dict(cls, config_dict: Dict, **kwargs) -> "PretrainedConfig":
+        """
+        Constructs a `Config` from a Python dictionary of parameters.
+
+        Args:
+            config_dict (:obj:`Dict[str, any]`):
+                Dictionary that will be used to instantiate the configuration object. Such a dictionary can be retrieved
+                from a pre-trained checkpoint by leveraging the :func:`~transformers.PretrainedConfig.get_config_dict`
+                method.
+            kwargs (:obj:`Dict[str, any]`):
+                Additional parameters from which to initialize the configuration object.
+
+        Returns:
+            :class:`PretrainedConfig`: An instance of a configuration object
+        """
+        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
+
+        config = cls(**config_dict)
+
+        if hasattr(config, "pruned_heads"):
+            config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
+
+        # Update config with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(config, key):
+                setattr(config, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info("Model config %s", str(config))
+        if return_unused_kwargs:
+            return config, kwargs
+        else:
+            return config
+
+    @classmethod
+    def from_json_file(cls, json_file: str) -> "PretrainedConfig":
+        """
+        Constructs a `Config` from the path to a json file of parameters.
+
+        Args:
+            json_file (:obj:`string`):
+                Path to the JSON file containing the parameters.
+
+        Returns:
+            :class:`PretrainedConfig`: An instance of a configuration object
+
+        """
+        config_dict = cls._dict_from_json_file(json_file)
+        return cls(**config_dict)
+
+    @classmethod
+    def _dict_from_json_file(cls, json_file: str):
+        with open(json_file, "r", encoding="utf-8") as reader:
+            text = reader.read()
+        return json.loads(text)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return "{} {}".format(self.__class__.__name__, self.to_json_string())
+
+    def to_diff_dict(self):
+        """
+        Removes all attributes from config which correspond to the default
+        config attributes for better readability and serializes to a Python
+        dictionary.
+
+        Returns:
+            :obj:`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        config_dict = self.to_dict()
+
+        # get the default config dict
+        default_config_dict = PretrainedConfig().to_dict()
+
+        serializable_config_dict = {}
+
+        # only serialize values that differ from the default config
+        for key, value in config_dict.items():
+            if key not in default_config_dict or value != default_config_dict[key]:
+                serializable_config_dict[key] = value
+
+        return serializable_config_dict
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary.
+
+        Returns:
+            :obj:`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        if hasattr(self.__class__, "model_type"):
+            output["model_type"] = self.__class__.model_type
+        return output
+
+    def to_json_string(self, use_diff=True):
+        """
+        Serializes this instance to a JSON string.
+
+        Args:
+            use_diff (:obj:`bool`):
+                If set to True, only the difference between the config instance and the default PretrainedConfig() is serialized to JSON string.
+
+        Returns:
+            :obj:`string`: String containing all the attributes that make up this configuration instance in JSON format.
+        """
+        if use_diff is True:
+            config_dict = self.to_diff_dict()
+        else:
+            config_dict = self.to_dict()
+        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path, use_diff=True):
+        """
+        Save this instance to a json file.
+
+        Args:
+            json_file_path (:obj:`string`):
+                Path to the JSON file in which this configuration instance's parameters will be saved.
+            use_diff (:obj:`bool`):
+                If set to True, only the difference between the config instance and the default PretrainedConfig() is serialized to JSON file.
+        """
+        with open(json_file_path, "w", encoding="utf-8") as writer:
+            writer.write(self.to_json_string(use_diff=use_diff))
+
+    def update(self, config_dict: Dict):
+        """
+        Updates attributes of this class
+        with attributes from `config_dict`.
+
+        Args:
+            :obj:`Dict[str, any]`: Dictionary of attributes that shall be updated for this class.
+        """
+        for key, value in config_dict.items():
+            setattr(self, key, value)

bert/file_utils.py

@@ -0,0 +1,808 @@
+"""
+Utilities for working with the local dataset cache.
+This file is adapted from the AllenNLP library at https://github.com/allenai/allennlp
+Copyright by the AllenNLP authors.
+"""
+
+import fnmatch
+import json
+import logging
+import os
+import shutil
+import sys
+import tarfile
+import tempfile
+from contextlib import contextmanager
+from functools import partial, wraps
+from hashlib import sha256
+from pathlib import Path
+from typing import Dict, Optional, Union
+from urllib.parse import urlparse
+from zipfile import ZipFile, is_zipfile
+
+import requests
+from filelock import FileLock
+from tqdm.auto import tqdm
+
+#from . import __version__
+__version__ = "3.0.2"
+
+logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
+
+try:
+    USE_TF = os.environ.get("USE_TF", "AUTO").upper()
+    USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()
+    if USE_TORCH in ("1", "ON", "YES", "AUTO") and USE_TF not in ("1", "ON", "YES"):
+        import torch
+
+        _torch_available = True  # pylint: disable=invalid-name
+        logger.info("PyTorch version {} available.".format(torch.__version__))
+    else:
+        logger.info("Disabling PyTorch because USE_TF is set")
+        _torch_available = False
+except ImportError:
+    _torch_available = False  # pylint: disable=invalid-name
+
+try:
+    USE_TF = os.environ.get("USE_TF", "AUTO").upper()
+    USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()
+
+    if USE_TF in ("1", "ON", "YES", "AUTO") and USE_TORCH not in ("1", "ON", "YES"):
+        import tensorflow as tf
+
+        assert hasattr(tf, "__version__") and int(tf.__version__[0]) >= 2
+        _tf_available = True  # pylint: disable=invalid-name
+        logger.info("TensorFlow version {} available.".format(tf.__version__))
+    else:
+        logger.info("Disabling Tensorflow because USE_TORCH is set")
+        _tf_available = False
+except (ImportError, AssertionError):
+    _tf_available = False  # pylint: disable=invalid-name
+
+
+try:
+    from torch.hub import _get_torch_home
+
+    torch_cache_home = _get_torch_home()
+except ImportError:
+    torch_cache_home = os.path.expanduser(
+        os.getenv("TORCH_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "torch"))
+    )
+
+
+try:
+    import torch_xla.core.xla_model as xm  # noqa: F401
+
+    if _torch_available:
+        _torch_tpu_available = True  # pylint: disable=
+    else:
+        _torch_tpu_available = False
+except ImportError:
+    _torch_tpu_available = False
+
+
+try:
+    import psutil  # noqa: F401
+
+    _psutil_available = True
+
+except ImportError:
+    _psutil_available = False
+
+
+try:
+    import py3nvml  # noqa: F401
+
+    _py3nvml_available = True
+
+except ImportError:
+    _py3nvml_available = False
+
+
+try:
+    from apex import amp  # noqa: F401
+
+    _has_apex = True
+except ImportError:
+    _has_apex = False
+
+default_cache_path = os.path.join(torch_cache_home, "transformers")
+
+
+PYTORCH_PRETRAINED_BERT_CACHE = os.getenv("PYTORCH_PRETRAINED_BERT_CACHE", default_cache_path)
+PYTORCH_TRANSFORMERS_CACHE = os.getenv("PYTORCH_TRANSFORMERS_CACHE", PYTORCH_PRETRAINED_BERT_CACHE)
+TRANSFORMERS_CACHE = os.getenv("TRANSFORMERS_CACHE", PYTORCH_TRANSFORMERS_CACHE)
+
+WEIGHTS_NAME = "pytorch_model.bin"
+TF2_WEIGHTS_NAME = "tf_model.h5"
+TF_WEIGHTS_NAME = "model.ckpt"
+CONFIG_NAME = "config.json"
+MODEL_CARD_NAME = "modelcard.json"
+
+
+MULTIPLE_CHOICE_DUMMY_INPUTS = [[[0], [1]], [[0], [1]]]
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
+
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+CLOUDFRONT_DISTRIB_PREFIX = "https://cdn.huggingface.co"
+
+
+def is_torch_available():
+    return _torch_available
+
+
+def is_tf_available():
+    return _tf_available
+
+
+def is_torch_tpu_available():
+    return _torch_tpu_available
+
+
+def is_psutil_available():
+    return _psutil_available
+
+
+def is_py3nvml_available():
+    return _py3nvml_available
+
+
+def is_apex_available():
+    return _has_apex
+
+
+def add_start_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_start_docstrings_to_callable(*docstr):
+    def docstring_decorator(fn):
+        class_name = ":class:`~transformers.{}`".format(fn.__qualname__.split(".")[0])
+        intro = "   The {} forward method, overrides the :func:`__call__` special method.".format(class_name)
+        note = r"""
+
+    .. note::
+        Although the recipe for forward pass needs to be defined within
+        this function, one should call the :class:`Module` instance afterwards
+        instead of this since the former takes care of running the
+        pre and post processing steps while the latter silently ignores them.
+        """
+        fn.__doc__ = intro + note + "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")
+        return fn
+
+    return docstring_decorator
+
+
+def add_end_docstrings(*docstr):
+    def docstring_decorator(fn):
+        fn.__doc__ = fn.__doc__ + "".join(docstr)
+        return fn
+
+    return docstring_decorator
+
+
+PT_TOKEN_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1] * inputs["input_ids"].size(1)).unsqueeze(0)  # Batch size 1
+
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss, scores = outputs[:2]
+"""
+
+PT_QUESTION_ANSWERING_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> start_positions = torch.tensor([1])
+        >>> end_positions = torch.tensor([3])
+
+        >>> outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)
+        >>> loss, start_scores, end_scores = outputs[:3]
+"""
+
+PT_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        >>> outputs = model(**inputs, labels=labels)
+        >>> loss, logits = outputs[:2]
+"""
+
+PT_MASKED_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> input_ids = tokenizer("Hello, my dog is cute", return_tensors="pt")["input_ids"]
+
+        >>> outputs = model(input_ids, labels=input_ids)
+        >>> loss, prediction_scores = outputs[:2]
+"""
+
+PT_BASE_MODEL_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+"""
+
+PT_MULTIPLE_CHOICE_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import torch
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> choice0 = "It is eaten with a fork and a knife."
+        >>> choice1 = "It is eaten while held in the hand."
+        >>> labels = torch.tensor(0).unsqueeze(0)  # choice0 is correct (according to Wikipedia ;)), batch size 1
+
+        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='pt', padding=True)
+        >>> outputs = model(**{{k: v.unsqueeze(0) for k,v in encoding.items()}}, labels=labels)  # batch size is 1
+
+        >>> # the linear classifier still needs to be trained
+        >>> loss, logits = outputs[:2]
+"""
+
+PT_CAUSAL_LM_SAMPLE = r"""
+    Example::
+
+        >>> import torch
+        >>> from transformers import {tokenizer_class}, {model_class}
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs, labels=inputs["input_ids"])
+        >>> loss, logits = outputs[:2]
+"""
+
+TF_TOKEN_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> input_ids = inputs["input_ids"]
+        >>> inputs["labels"] = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids))) # Batch size 1
+
+        >>> outputs = model(inputs)
+        >>> loss, scores = outputs[:2]
+"""
+
+TF_QUESTION_ANSWERING_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
+        >>> input_dict = tokenizer(question, text, return_tensors='tf')
+        >>> start_scores, end_scores = model(input_dict)
+
+        >>> all_tokens = tokenizer.convert_ids_to_tokens(input_dict["input_ids"].numpy()[0])
+        >>> answer = ' '.join(all_tokens[tf.math.argmax(start_scores, 1)[0] : tf.math.argmax(end_scores, 1)[0]+1])
+"""
+
+TF_SEQUENCE_CLASSIFICATION_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> inputs["labels"] = tf.reshape(tf.constant(1), (-1, 1)) # Batch size 1
+
+        >>> outputs = model(inputs)
+        >>> loss, logits = outputs[:2]
+"""
+
+TF_MASKED_LM_SAMPLE = r"""
+    Example::
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
+
+        >>> outputs = model(input_ids)
+        >>> prediction_scores = outputs[0]
+"""
+
+TF_BASE_MODEL_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> outputs = model(inputs)
+
+        >>> last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+"""
+
+TF_MULTIPLE_CHOICE_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> choice0 = "It is eaten with a fork and a knife."
+        >>> choice1 = "It is eaten while held in the hand."
+
+        >>> encoding = tokenizer([[prompt, prompt], [choice0, choice1]], return_tensors='tf', padding=True)
+        >>> inputs = {{k: tf.expand_dims(v, 0) for k, v in encoding.items()}}
+        >>> outputs = model(inputs)  # batch size is 1
+
+        >>> # the linear classifier still needs to be trained
+        >>> logits = outputs[0]
+"""
+
+TF_CAUSAL_LM_SAMPLE = r"""
+    Example::
+
+        >>> from transformers import {tokenizer_class}, {model_class}
+        >>> import tensorflow as tf
+
+        >>> tokenizer = {tokenizer_class}.from_pretrained('{checkpoint}')
+        >>> model = {model_class}.from_pretrained('{checkpoint}')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
+        >>> outputs = model(inputs)
+        >>> logits = outputs[0]
+"""
+
+
+def add_code_sample_docstrings(*docstr, tokenizer_class=None, checkpoint=None):
+    def docstring_decorator(fn):
+        model_class = fn.__qualname__.split(".")[0]
+        is_tf_class = model_class[:2] == "TF"
+
+        if "SequenceClassification" in model_class:
+            code_sample = TF_SEQUENCE_CLASSIFICATION_SAMPLE if is_tf_class else PT_SEQUENCE_CLASSIFICATION_SAMPLE
+        elif "QuestionAnswering" in model_class:
+            code_sample = TF_QUESTION_ANSWERING_SAMPLE if is_tf_class else PT_QUESTION_ANSWERING_SAMPLE
+        elif "TokenClassification" in model_class:
+            code_sample = TF_TOKEN_CLASSIFICATION_SAMPLE if is_tf_class else PT_TOKEN_CLASSIFICATION_SAMPLE
+        elif "MultipleChoice" in model_class:
+            code_sample = TF_MULTIPLE_CHOICE_SAMPLE if is_tf_class else PT_MULTIPLE_CHOICE_SAMPLE
+        elif "MaskedLM" in model_class:
+            code_sample = TF_MASKED_LM_SAMPLE if is_tf_class else PT_MASKED_LM_SAMPLE
+        elif "LMHead" in model_class:
+            code_sample = TF_CAUSAL_LM_SAMPLE if is_tf_class else PT_CAUSAL_LM_SAMPLE
+        elif "Model" in model_class:
+            code_sample = TF_BASE_MODEL_SAMPLE if is_tf_class else PT_BASE_MODEL_SAMPLE
+        else:
+            raise ValueError(f"Docstring can't be built for model {model_class}")
+
+        built_doc = code_sample.format(model_class=model_class, tokenizer_class=tokenizer_class, checkpoint=checkpoint)
+        fn.__doc__ = (fn.__doc__ or "") + "".join(docstr) + built_doc
+        return fn
+
+    return docstring_decorator
+
+
+def is_remote_url(url_or_filename):
+    parsed = urlparse(url_or_filename)
+    return parsed.scheme in ("http", "https")
+
+
+def hf_bucket_url(model_id: str, filename: str, use_cdn=True) -> str:
+    """
+    Resolve a model identifier, and a file name, to a HF-hosted url
+    on either S3 or Cloudfront (a Content Delivery Network, or CDN).
+
+    Cloudfront is replicated over the globe so downloads are way faster
+    for the end user (and it also lowers our bandwidth costs). However, it
+    is more aggressively cached by default, so may not always reflect the
+    latest changes to the underlying file (default TTL is 24 hours).
+
+    In terms of client-side caching from this library, even though
+    Cloudfront relays the ETags from S3, using one or the other
+    (or switching from one to the other) will affect caching: cached files
+    are not shared between the two because the cached file's name contains
+    a hash of the url.
+    """
+    endpoint = CLOUDFRONT_DISTRIB_PREFIX if use_cdn else S3_BUCKET_PREFIX
+    legacy_format = "/" not in model_id
+    if legacy_format:
+        return f"{endpoint}/{model_id}-{filename}"
+    else:
+        return f"{endpoint}/{model_id}/{filename}"
+
+
+def url_to_filename(url, etag=None):
+    """
+    Convert `url` into a hashed filename in a repeatable way.
+    If `etag` is specified, append its hash to the url's, delimited
+    by a period.
+    If the url ends with .h5 (Keras HDF5 weights) adds '.h5' to the name
+    so that TF 2.0 can identify it as a HDF5 file
+    (see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1380)
+    """
+    url_bytes = url.encode("utf-8")
+    url_hash = sha256(url_bytes)
+    filename = url_hash.hexdigest()
+
+    if etag:
+        etag_bytes = etag.encode("utf-8")
+        etag_hash = sha256(etag_bytes)
+        filename += "." + etag_hash.hexdigest()
+
+    if url.endswith(".h5"):
+        filename += ".h5"
+
+    return filename
+
+
+def filename_to_url(filename, cache_dir=None):
+    """
+    Return the url and etag (which may be ``None``) stored for `filename`.
+    Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist.
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    cache_path = os.path.join(cache_dir, filename)
+    if not os.path.exists(cache_path):
+        raise EnvironmentError("file {} not found".format(cache_path))
+
+    meta_path = cache_path + ".json"
+    if not os.path.exists(meta_path):
+        raise EnvironmentError("file {} not found".format(meta_path))
+
+    with open(meta_path, encoding="utf-8") as meta_file:
+        metadata = json.load(meta_file)
+    url = metadata["url"]
+    etag = metadata["etag"]
+
+    return url, etag
+
+
+def cached_path(
+    url_or_filename,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    extract_compressed_file=False,
+    force_extract=False,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given something that might be a URL (or might be a local path),
+    determine which. If it's a URL, download the file and cache it, and
+    return the path to the cached file. If it's already a local path,
+    make sure the file exists and then return the path.
+    Args:
+        cache_dir: specify a cache directory to save the file to (overwrite the default cache dir).
+        force_download: if True, re-dowload the file even if it's already cached in the cache dir.
+        resume_download: if True, resume the download if incompletly recieved file is found.
+        user_agent: Optional string or dict that will be appended to the user-agent on remote requests.
+        extract_compressed_file: if True and the path point to a zip or tar file, extract the compressed
+            file in a folder along the archive.
+        force_extract: if True when extract_compressed_file is True and the archive was already extracted,
+            re-extract the archive and overide the folder where it was extracted.
+
+    Return:
+        None in case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+        Local path (string) otherwise
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(url_or_filename, Path):
+        url_or_filename = str(url_or_filename)
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    if is_remote_url(url_or_filename):
+        # URL, so get it from the cache (downloading if necessary)
+        output_path = get_from_cache(
+            url_or_filename,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            resume_download=resume_download,
+            user_agent=user_agent,
+            local_files_only=local_files_only,
+        )
+    elif os.path.exists(url_or_filename):
+        # File, and it exists.
+        output_path = url_or_filename
+    elif urlparse(url_or_filename).scheme == "":
+        # File, but it doesn't exist.
+        raise EnvironmentError("file {} not found".format(url_or_filename))
+    else:
+        # Something unknown
+        raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename))
+
+    if extract_compressed_file:
+        if not is_zipfile(output_path) and not tarfile.is_tarfile(output_path):
+            return output_path
+
+        # Path where we extract compressed archives
+        # We avoid '.' in dir name and add "-extracted" at the end: "./model.zip" => "./model-zip-extracted/"
+        output_dir, output_file = os.path.split(output_path)
+        output_extract_dir_name = output_file.replace(".", "-") + "-extracted"
+        output_path_extracted = os.path.join(output_dir, output_extract_dir_name)
+
+        if os.path.isdir(output_path_extracted) and os.listdir(output_path_extracted) and not force_extract:
+            return output_path_extracted
+
+        # Prevent parallel extractions
+        lock_path = output_path + ".lock"
+        with FileLock(lock_path):
+            shutil.rmtree(output_path_extracted, ignore_errors=True)
+            os.makedirs(output_path_extracted)
+            if is_zipfile(output_path):
+                with ZipFile(output_path, "r") as zip_file:
+                    zip_file.extractall(output_path_extracted)
+                    zip_file.close()
+            elif tarfile.is_tarfile(output_path):
+                tar_file = tarfile.open(output_path)
+                tar_file.extractall(output_path_extracted)
+                tar_file.close()
+            else:
+                raise EnvironmentError("Archive format of {} could not be identified".format(output_path))
+
+        return output_path_extracted
+
+    return output_path
+
+
+def http_get(url, temp_file, proxies=None, resume_size=0, user_agent: Union[Dict, str, None] = None):
+    ua = "transformers/{}; python/{}".format(__version__, sys.version.split()[0])
+    if is_torch_available():
+        ua += "; torch/{}".format(torch.__version__)
+    if is_tf_available():
+        ua += "; tensorflow/{}".format(tf.__version__)
+    if isinstance(user_agent, dict):
+        ua += "; " + "; ".join("{}/{}".format(k, v) for k, v in user_agent.items())
+    elif isinstance(user_agent, str):
+        ua += "; " + user_agent
+    headers = {"user-agent": ua}
+    if resume_size > 0:
+        headers["Range"] = "bytes=%d-" % (resume_size,)
+    response = requests.get(url, stream=True, proxies=proxies, headers=headers)
+    if response.status_code == 416:  # Range not satisfiable
+        return
+    content_length = response.headers.get("Content-Length")
+    total = resume_size + int(content_length) if content_length is not None else None
+    progress = tqdm(
+        unit="B",
+        unit_scale=True,
+        total=total,
+        initial=resume_size,
+        desc="Downloading",
+        disable=bool(logger.getEffectiveLevel() == logging.NOTSET),
+    )
+    for chunk in response.iter_content(chunk_size=1024):
+        if chunk:  # filter out keep-alive new chunks
+            progress.update(len(chunk))
+            temp_file.write(chunk)
+    progress.close()
+
+
+def get_from_cache(
+    url,
+    cache_dir=None,
+    force_download=False,
+    proxies=None,
+    etag_timeout=10,
+    resume_download=False,
+    user_agent: Union[Dict, str, None] = None,
+    local_files_only=False,
+) -> Optional[str]:
+    """
+    Given a URL, look for the corresponding file in the local cache.
+    If it's not there, download it. Then return the path to the cached file.
+
+    Return:
+        None in case of non-recoverable file (non-existent or inaccessible url + no cache on disk).
+        Local path (string) otherwise
+    """
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    if isinstance(cache_dir, Path):
+        cache_dir = str(cache_dir)
+
+    os.makedirs(cache_dir, exist_ok=True)
+
+    etag = None
+    if not local_files_only:
+        try:
+            response = requests.head(url, allow_redirects=True, proxies=proxies, timeout=etag_timeout)
+            if response.status_code == 200:
+                etag = response.headers.get("ETag")
+        except (EnvironmentError, requests.exceptions.Timeout):
+            # etag is already None
+            pass
+
+    filename = url_to_filename(url, etag)
+
+    # get cache path to put the file
+    cache_path = os.path.join(cache_dir, filename)
+
+    # etag is None = we don't have a connection, or url doesn't exist, or is otherwise inaccessible.
+    # try to get the last downloaded one
+    if etag is None:
+        if os.path.exists(cache_path):
+            return cache_path
+        else:
+            matching_files = [
+                file
+                for file in fnmatch.filter(os.listdir(cache_dir), filename + ".*")
+                if not file.endswith(".json") and not file.endswith(".lock")
+            ]
+            if len(matching_files) > 0:
+                return os.path.join(cache_dir, matching_files[-1])
+            else:
+                # If files cannot be found and local_files_only=True,
+                # the models might've been found if local_files_only=False
+                # Notify the user about that
+                if local_files_only:
+                    raise ValueError(
+                        "Cannot find the requested files in the cached path and outgoing traffic has been"
+                        " disabled. To enable model look-ups and downloads online, set 'local_files_only'"
+                        " to False."
+                    )
+                return None
+
+    # From now on, etag is not None.
+    if os.path.exists(cache_path) and not force_download:
+        return cache_path
+
+    # Prevent parallel downloads of the same file with a lock.
+    lock_path = cache_path + ".lock"
+    with FileLock(lock_path):
+
+        # If the download just completed while the lock was activated.
+        if os.path.exists(cache_path) and not force_download:
+            # Even if returning early like here, the lock will be released.
+            return cache_path
+
+        if resume_download:
+            incomplete_path = cache_path + ".incomplete"
+
+            @contextmanager
+            def _resumable_file_manager():
+                with open(incomplete_path, "a+b") as f:
+                    yield f
+
+            temp_file_manager = _resumable_file_manager
+            if os.path.exists(incomplete_path):
+                resume_size = os.stat(incomplete_path).st_size
+            else:
+                resume_size = 0
+        else:
+            temp_file_manager = partial(tempfile.NamedTemporaryFile, dir=cache_dir, delete=False)
+            resume_size = 0
+
+        # Download to temporary file, then copy to cache dir once finished.
+        # Otherwise you get corrupt cache entries if the download gets interrupted.
+        with temp_file_manager() as temp_file:
+            logger.info("%s not found in cache or force_download set to True, downloading to %s", url, temp_file.name)
+
+            http_get(url, temp_file, proxies=proxies, resume_size=resume_size, user_agent=user_agent)
+
+        logger.info("storing %s in cache at %s", url, cache_path)
+        os.replace(temp_file.name, cache_path)
+
+        logger.info("creating metadata file for %s", cache_path)
+        meta = {"url": url, "etag": etag}
+        meta_path = cache_path + ".json"
+        with open(meta_path, "w") as meta_file:
+            json.dump(meta, meta_file)
+
+    return cache_path
+
+
+class cached_property(property):
+    """
+    Descriptor that mimics @property but caches output in member variable.
+
+    From tensorflow_datasets
+
+    Built-in in functools from Python 3.8.
+    """
+
+    def __get__(self, obj, objtype=None):
+        # See docs.python.org/3/howto/descriptor.html#properties
+        if obj is None:
+            return self
+        if self.fget is None:
+            raise AttributeError("unreadable attribute")
+        attr = "__cached_" + self.fget.__name__
+        cached = getattr(obj, attr, None)
+        if cached is None:
+            cached = self.fget(obj)
+            setattr(obj, attr, cached)
+        return cached
+
+
+def torch_required(func):
+    # Chose a different decorator name than in tests so it's clear they are not the same.
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        if is_torch_available():
+            return func(*args, **kwargs)
+        else:
+            raise ImportError(f"Method `{func.__name__}` requires PyTorch.")
+
+    return wrapper
+
+
+def tf_required(func):
+    # Chose a different decorator name than in tests so it's clear they are not the same.
+    @wraps(func)
+    def wrapper(*args, **kwargs):
+        if is_tf_available():
+            return func(*args, **kwargs)
+        else:
+            raise ImportError(f"Method `{func.__name__}` requires TF.")
+
+    return wrapper

bert/generation_utils.py

@@ -0,0 +1,993 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from typing import Iterable, Optional, Tuple
+
+import torch
+from torch import Tensor
+from torch.nn import functional as F
+
+
+logger = logging.getLogger(__name__)
+
+
+class GenerationMixin:
+    """
+    A class contraining all of the functions supporting generation, to be used as a mixin in PreTrainedModel.
+    """
+
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        return {"input_ids": input_ids}
+
+    def adjust_logits_during_generation(self, logits, **kwargs):
+        return logits
+
+    def _use_cache(self, outputs, use_cache):
+        """During generation, decide whether to pass the `past` variable to the next forward pass."""
+        if len(outputs) <= 1 or use_cache is False:
+            return False
+        if hasattr(self.config, "mem_len") and self.config.mem_len == 0:
+            return False
+        return True
+
+    def enforce_repetition_penalty_(self, lprobs, batch_size, num_beams, prev_output_tokens, repetition_penalty):
+        """repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858). """
+        for i in range(batch_size * num_beams):
+            for previous_token in set(prev_output_tokens[i].tolist()):
+                # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if lprobs[i, previous_token] < 0:
+                    lprobs[i, previous_token] *= repetition_penalty
+                else:
+                    lprobs[i, previous_token] /= repetition_penalty
+
+    def postprocess_next_token_scores(
+        self,
+        scores,
+        input_ids,
+        no_repeat_ngram_size,
+        bad_words_ids,
+        cur_len,
+        min_length,
+        max_length,
+        eos_token_id,
+        repetition_penalty,
+        batch_size,
+        num_beams,
+    ):
+        # repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858)
+        if repetition_penalty != 1.0:
+            self.enforce_repetition_penalty_(
+                scores, batch_size, num_beams, input_ids, repetition_penalty,
+            )
+
+        # set eos token prob to zero if min_length is not reached
+        if eos_token_id is not None and cur_len < min_length:
+            scores[:, eos_token_id] = -float("inf")
+
+        if no_repeat_ngram_size > 0:
+            # calculate a list of banned tokens to prevent repetitively generating the same ngrams
+            num_batch_hypotheses = batch_size * num_beams
+            # from fairseq: https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345
+            banned_batch_tokens = calc_banned_ngram_tokens(
+                input_ids, num_batch_hypotheses, no_repeat_ngram_size, cur_len
+            )
+            for i, banned_tokens in enumerate(banned_batch_tokens):
+                scores[i, banned_tokens] = -float("inf")
+
+        if bad_words_ids is not None:
+            # calculate a list of banned tokens according to bad words
+            banned_tokens = calc_banned_bad_words_ids(input_ids, bad_words_ids)
+
+            for i, banned_tokens in enumerate(banned_tokens):
+                scores[i, banned_tokens] = -float("inf")
+
+        return scores
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        max_length: Optional[int] = None,
+        min_length: Optional[int] = None,
+        do_sample: Optional[bool] = None,
+        early_stopping: Optional[bool] = None,
+        num_beams: Optional[int] = None,
+        temperature: Optional[float] = None,
+        top_k: Optional[int] = None,
+        top_p: Optional[float] = None,
+        repetition_penalty: Optional[float] = None,
+        bad_words_ids: Optional[Iterable[int]] = None,
+        bos_token_id: Optional[int] = None,
+        pad_token_id: Optional[int] = None,
+        eos_token_id: Optional[int] = None,
+        length_penalty: Optional[float] = None,
+        no_repeat_ngram_size: Optional[int] = None,
+        num_return_sequences: Optional[int] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_start_token_id: Optional[int] = None,
+        use_cache: Optional[bool] = None,
+        **model_specific_kwargs
+    ) -> torch.LongTensor:
+        r""" Generates sequences for models with a LM head. The method currently supports greedy decoding, beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.
+
+        Adapted in part from `Facebook's XLM beam search code`_.
+
+        .. _`Facebook's XLM beam search code`:
+           https://github.com/facebookresearch/XLM/blob/9e6f6814d17be4fe5b15f2e6c43eb2b2d76daeb4/src/model/transformer.py#L529
+
+
+        Parameters:
+
+            input_ids: (`optional`) `torch.LongTensor` of shape `(batch_size, sequence_length)`
+                The sequence used as a prompt for the generation. If `None` the method initializes
+                it as an empty `torch.LongTensor` of shape `(1,)`.
+
+            max_length: (`optional`) int
+                The max length of the sequence to be generated.  Between `min_length` and infinity. Default to 20.
+
+            min_length: (`optional`) int
+                The min length of the sequence to be generated.  Between 0 and infinity. Default to 0.
+
+            do_sample: (`optional`) bool
+                If set to `False` greedy decoding is used. Otherwise sampling is used. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.
+
+            early_stopping: (`optional`) bool
+                if set to `True` beam search is stopped when at least `num_beams` sentences finished per batch. Defaults to `False` as defined in `configuration_utils.PretrainedConfig`.
+
+            num_beams: (`optional`) int
+                Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.
+
+            temperature: (`optional`) float
+                The value used to module the next token probabilities. Must be strictly positive. Default to 1.0.
+
+            top_k: (`optional`) int
+                The number of highest probability vocabulary tokens to keep for top-k-filtering. Between 1 and infinity. Default to 50.
+
+            top_p: (`optional`) float
+                The cumulative probability of parameter highest probability vocabulary tokens to keep for nucleus sampling. Must be between 0 and 1. Default to 1.
+
+            repetition_penalty: (`optional`) float
+                The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.
+
+            pad_token_id: (`optional`) int
+                Padding token. Default to specicic model pad_token_id or None if it does not exist.
+
+            bos_token_id: (`optional`) int
+                BOS token. Defaults to `bos_token_id` as defined in the models config.
+
+            eos_token_id: (`optional`) int
+                EOS token. Defaults to `eos_token_id` as defined in the models config.
+
+            length_penalty: (`optional`) float
+                Exponential penalty to the length. Default to 1.
+
+            no_repeat_ngram_size: (`optional`) int
+                If set to int > 0, all ngrams of size `no_repeat_ngram_size` can only occur once.
+            bad_words_ids: (`optional`) list of lists of int
+                `bad_words_ids` contains tokens that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use `tokenizer.encode(bad_word, add_prefix_space=True)`.
+
+            num_return_sequences: (`optional`) int
+                The number of independently computed returned sequences for each element in the batch. Default to 1.
+
+            attention_mask (`optional`) obj: `torch.LongTensor` of same shape as `input_ids`
+                Mask to avoid performing attention on padding token indices.
+                Mask values selected in ``[0, 1]``:
+                ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+                Defaults to `None`.
+
+                `What are attention masks? <../glossary.html#attention-mask>`__
+
+            decoder_start_token_id=None: (`optional`) int
+                If an encoder-decoder model starts decoding with a different token than BOS.
+                Defaults to `None` and is changed to `BOS` later.
+
+            use_cache: (`optional`) bool
+                If `use_cache` is True, past key values are used to speed up decoding if applicable to model. Defaults to `True`.
+
+            model_specific_kwargs: (`optional`) dict
+                Additional model specific kwargs will be forwarded to the `forward` function of the model.
+
+        Return:
+
+            output: `torch.LongTensor` of shape `(batch_size * num_return_sequences, sequence_length)`
+                sequence_length is either equal to max_length or shorter if all batches finished early due to the `eos_token_id`
+
+        Examples::
+
+            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
+            outputs = model.generate(max_length=40)  # do greedy decoding
+            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('openai-gpt')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('openai-gpt')    # Download model and configuration from S3 and cache.
+            input_context = 'The dog'
+            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
+            outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5)  # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
+            for i in range(3): #  3 output sequences were generated
+                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from S3 and cache.
+            input_context = 'The dog'
+            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3)  # 3 generate sequences using by sampling
+            for i in range(3): #  3 output sequences were generated
+                print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('ctrl')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('ctrl')    # Download model and configuration from S3 and cache.
+            input_context = 'Legal My neighbor is'  # "Legal" is one of the control codes for ctrl
+            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
+            print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))
+
+            tokenizer = AutoTokenizer.from_pretrained('gpt2')   # Initialize tokenizer
+            model = AutoModelWithLMHead.from_pretrained('gpt2')    # Download model and configuration from S3 and cache.
+            input_context = 'My cute dog'  # "Legal" is one of the control codes for ctrl
+            bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']]
+            input_ids = tokenizer.encode(input_context, return_tensors='pt')  # encode input context
+            outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids)  # generate sequences without allowing bad_words to be generated
+        """
+
+        # We cannot generate if the model does not have a LM head
+        if self.get_output_embeddings() is None:
+            raise AttributeError(
+                "You tried to generate sequences with a model that does not have a LM Head."
+                "Please use another model class (e.g. `OpenAIGPTLMHeadModel`, `XLNetLMHeadModel`, `GPT2LMHeadModel`, `CTRLLMHeadModel`, `T5WithLMHeadModel`, `TransfoXLLMHeadModel`, `XLMWithLMHeadModel`, `BartForConditionalGeneration` )"
+            )
+
+        max_length = max_length if max_length is not None else self.config.max_length
+        min_length = min_length if min_length is not None else self.config.min_length
+        do_sample = do_sample if do_sample is not None else self.config.do_sample
+        early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        num_beams = num_beams if num_beams is not None else self.config.num_beams
+        temperature = temperature if temperature is not None else self.config.temperature
+        top_k = top_k if top_k is not None else self.config.top_k
+        top_p = top_p if top_p is not None else self.config.top_p
+        repetition_penalty = repetition_penalty if repetition_penalty is not None else self.config.repetition_penalty
+        bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
+        pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
+        eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
+        length_penalty = length_penalty if length_penalty is not None else self.config.length_penalty
+        no_repeat_ngram_size = (
+            no_repeat_ngram_size if no_repeat_ngram_size is not None else self.config.no_repeat_ngram_size
+        )
+        bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
+        num_return_sequences = (
+            num_return_sequences if num_return_sequences is not None else self.config.num_return_sequences
+        )
+        decoder_start_token_id = (
+            decoder_start_token_id if decoder_start_token_id is not None else self.config.decoder_start_token_id
+        )
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]  # overriden by the input batch_size
+        else:
+            batch_size = 1
+
+        assert isinstance(max_length, int) and max_length > 0, "`max_length` should be a strictly positive integer."
+        assert isinstance(min_length, int) and min_length >= 0, "`min_length` should be a positive integer."
+        assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
+        assert isinstance(early_stopping, bool), "`early_stopping` should be a boolean."
+        assert isinstance(use_cache, bool), "`use_cache` should be a boolean."
+        assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
+        assert temperature > 0, "`temperature` should be strictly positive."
+        assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."
+        assert 0 <= top_p <= 1, "`top_p` should be between 0 and 1."
+        assert repetition_penalty >= 1.0, "`repetition_penalty` should be >= 1."
+        assert input_ids is not None or (
+            isinstance(bos_token_id, int) and bos_token_id >= 0
+        ), "If input_ids is not defined, `bos_token_id` should be a positive integer."
+        assert pad_token_id is None or (
+            isinstance(pad_token_id, int) and (pad_token_id >= 0)
+        ), "`pad_token_id` should be a positive integer."
+        assert (eos_token_id is None) or (
+            isinstance(eos_token_id, int) and (eos_token_id >= 0)
+        ), "`eos_token_id` should be a positive integer."
+        assert length_penalty > 0, "`length_penalty` should be strictly positive."
+        assert (
+            isinstance(no_repeat_ngram_size, int) and no_repeat_ngram_size >= 0
+        ), "`no_repeat_ngram_size` should be a positive integer."
+        assert (
+            isinstance(num_return_sequences, int) and num_return_sequences > 0
+        ), "`num_return_sequences` should be a strictly positive integer."
+        assert (
+            bad_words_ids is None or isinstance(bad_words_ids, list) and isinstance(bad_words_ids[0], list)
+        ), "`bad_words_ids` is either `None` or a list of lists of tokens that should not be generated"
+
+        if input_ids is None:
+            assert isinstance(bos_token_id, int) and bos_token_id >= 0, (
+                "you should either supply a context to complete as `input_ids` input "
+                "or a `bos_token_id` (integer >= 0) as a first token to start the generation."
+            )
+            input_ids = torch.full(
+                (batch_size, 1), bos_token_id, dtype=torch.long, device=next(self.parameters()).device,
+            )
+        else:
+            assert input_ids.dim() == 2, "Input prompt should be of shape (batch_size, sequence length)."
+
+        # not allow to duplicate outputs when greedy decoding
+        if do_sample is False:
+            if num_beams == 1:
+                # no_beam_search greedy generation conditions
+                assert (
+                    num_return_sequences == 1
+                ), "Greedy decoding will always produce the same output for num_beams == 1 and num_return_sequences > 1. Please set num_return_sequences = 1"
+
+            else:
+                # beam_search greedy generation conditions
+                assert (
+                    num_beams >= num_return_sequences
+                ), "Greedy beam search decoding cannot return more sequences than it has beams. Please set num_beams >= num_return_sequences"
+
+        # create attention mask if necessary
+        # TODO (PVP): this should later be handled by the forward fn() in each model in the future see PR 3140
+        if (attention_mask is None) and (pad_token_id is not None) and (pad_token_id in input_ids):
+            attention_mask = input_ids.ne(pad_token_id).long()
+        elif attention_mask is None:
+            attention_mask = input_ids.new_ones(input_ids.shape)
+
+        # set pad_token_id to eos_token_id if not set. Important that this is done after
+        # attention_mask is created
+        if pad_token_id is None and eos_token_id is not None:
+            logger.warning(
+                "Setting `pad_token_id` to {} (first `eos_token_id`) to generate sequence".format(eos_token_id)
+            )
+            pad_token_id = eos_token_id
+
+        # current position and vocab size
+        if hasattr(self.config, "vocab_size"):
+            vocab_size = self.config.vocab_size
+        elif (
+            self.config.is_encoder_decoder
+            and hasattr(self.config, "decoder")
+            and hasattr(self.config.decoder, "vocab_size")
+        ):
+            vocab_size = self.config.decoder.vocab_size
+
+        # set effective batch size and effective batch multiplier according to do_sample
+        if do_sample:
+            effective_batch_size = batch_size * num_return_sequences
+            effective_batch_mult = num_return_sequences
+        else:
+            effective_batch_size = batch_size
+            effective_batch_mult = 1
+
+        if self.config.is_encoder_decoder:
+            if decoder_start_token_id is None:
+                decoder_start_token_id = bos_token_id
+
+            assert (
+                decoder_start_token_id is not None
+            ), "decoder_start_token_id or bos_token_id has to be defined for encoder-decoder generation"
+            assert hasattr(self, "get_encoder"), "{} should have a 'get_encoder' function defined".format(self)
+            assert callable(self.get_encoder), "{} should be a method".format(self.get_encoder)
+
+            # get encoder and store encoder outputs
+            encoder = self.get_encoder()
+
+            encoder_outputs: tuple = encoder(input_ids, attention_mask=attention_mask)
+
+        # Expand input ids if num_beams > 1 or num_return_sequences > 1
+        if num_return_sequences > 1 or num_beams > 1:
+            input_ids_len = input_ids.shape[-1]
+            input_ids = input_ids.unsqueeze(1).expand(batch_size, effective_batch_mult * num_beams, input_ids_len)
+            attention_mask = attention_mask.unsqueeze(1).expand(
+                batch_size, effective_batch_mult * num_beams, input_ids_len
+            )
+
+            input_ids = input_ids.contiguous().view(
+                effective_batch_size * num_beams, input_ids_len
+            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
+            attention_mask = attention_mask.contiguous().view(
+                effective_batch_size * num_beams, input_ids_len
+            )  # shape: (batch_size * num_return_sequences * num_beams, cur_len)
+
+        if self.config.is_encoder_decoder:
+            # create empty decoder_input_ids
+            input_ids = torch.full(
+                (effective_batch_size * num_beams, 1),
+                decoder_start_token_id,
+                dtype=torch.long,
+                device=next(self.parameters()).device,
+            )
+            cur_len = 1
+
+            assert (
+                batch_size == encoder_outputs[0].shape[0]
+            ), f"expected encoder_outputs[0] to have 1st dimension bs={batch_size}, got {encoder_outputs[0].shape[0]} "
+
+            # expand batch_idx to assign correct encoder output for expanded input_ids (due to num_beams > 1 and num_return_sequences > 1)
+            expanded_batch_idxs = (
+                torch.arange(batch_size)
+                .view(-1, 1)
+                .repeat(1, num_beams * effective_batch_mult)
+                .view(-1)
+                .to(input_ids.device)
+            )
+            # expand encoder_outputs
+            encoder_outputs = (encoder_outputs[0].index_select(0, expanded_batch_idxs), *encoder_outputs[1:])
+
+        else:
+            encoder_outputs = None
+            cur_len = input_ids.shape[-1]
+
+        assert (
+            cur_len < max_length
+        ), f"The context has {cur_len} number of tokens, but `max_length` is only {max_length}. Please make sure that `max_length` is bigger than the number of tokens, by setting either `generate(max_length=...,...)` or `config.max_length = ...`"
+
+        if num_beams > 1:
+            output = self._generate_beam_search(
+                input_ids,
+                cur_len=cur_len,
+                max_length=max_length,
+                min_length=min_length,
+                do_sample=do_sample,
+                early_stopping=early_stopping,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
+                pad_token_id=pad_token_id,
+                eos_token_id=eos_token_id,
+                batch_size=effective_batch_size,
+                num_return_sequences=num_return_sequences,
+                length_penalty=length_penalty,
+                num_beams=num_beams,
+                vocab_size=vocab_size,
+                encoder_outputs=encoder_outputs,
+                attention_mask=attention_mask,
+                use_cache=use_cache,
+                model_specific_kwargs=model_specific_kwargs,
+            )
+        else:
+            output = self._generate_no_beam_search(
+                input_ids,
+                cur_len=cur_len,
+                max_length=max_length,
+                min_length=min_length,
+                do_sample=do_sample,
+                temperature=temperature,
+                top_k=top_k,
+                top_p=top_p,
+                repetition_penalty=repetition_penalty,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
+                pad_token_id=pad_token_id,
+                eos_token_id=eos_token_id,
+                batch_size=effective_batch_size,
+                encoder_outputs=encoder_outputs,
+                attention_mask=attention_mask,
+                use_cache=use_cache,
+                model_specific_kwargs=model_specific_kwargs,
+            )
+
+        return output
+
+    def _generate_no_beam_search(
+        self,
+        input_ids,
+        cur_len,
+        max_length,
+        min_length,
+        do_sample,
+        temperature,
+        top_k,
+        top_p,
+        repetition_penalty,
+        no_repeat_ngram_size,
+        bad_words_ids,
+        pad_token_id,
+        eos_token_id,
+        batch_size,
+        encoder_outputs,
+        attention_mask,
+        use_cache,
+        model_specific_kwargs,
+    ):
+        """ Generate sequences for each example without beam search (num_beams == 1).
+            All returned sequence are generated independantly.
+        """
+        # length of generated sentences / unfinished sentences
+        unfinished_sents = input_ids.new(batch_size).fill_(1)
+        sent_lengths = input_ids.new(batch_size).fill_(max_length)
+
+        past = (encoder_outputs, None) if encoder_outputs is not None else None
+
+        while cur_len < max_length:
+            model_inputs = self.prepare_inputs_for_generation(
+                input_ids, past=past, attention_mask=attention_mask, use_cache=use_cache, **model_specific_kwargs
+            )
+
+            outputs = self(**model_inputs)
+            next_token_logits = outputs[0][:, -1, :]
+
+            scores = self.postprocess_next_token_scores(
+                scores=next_token_logits,
+                input_ids=input_ids,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
+                cur_len=cur_len,
+                min_length=min_length,
+                max_length=max_length,
+                eos_token_id=eos_token_id,
+                repetition_penalty=repetition_penalty,
+                batch_size=batch_size,
+                num_beams=1,
+            )
+
+            # if model has past, then set the past variable to speed up decoding
+            if self._use_cache(outputs, use_cache):
+                past = outputs[1]
+
+            if do_sample:
+                # Temperature (higher temperature => more likely to sample low probability tokens)
+                if temperature != 1.0:
+                    scores = scores / temperature
+                # Top-p/top-k filtering
+                next_token_logscores = top_k_top_p_filtering(scores, top_k=top_k, top_p=top_p)
+                # Sample
+                probs = F.softmax(next_token_logscores, dim=-1)
+                next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
+            else:
+                # Greedy decoding
+                next_token = torch.argmax(next_token_logits, dim=-1)
+
+            # update generations and finished sentences
+            if eos_token_id is not None:
+                # pad finished sentences if eos_token_id exist
+                tokens_to_add = next_token * unfinished_sents + (pad_token_id) * (1 - unfinished_sents)
+            else:
+                tokens_to_add = next_token
+
+            # add token and increase length by one
+            input_ids = torch.cat([input_ids, tokens_to_add.unsqueeze(-1)], dim=-1)
+            cur_len = cur_len + 1
+
+            if eos_token_id is not None:
+                eos_in_sents = tokens_to_add == eos_token_id
+                # if sentence is unfinished and the token to add is eos, sent_lengths is filled with current length
+                is_sents_unfinished_and_token_to_add_is_eos = unfinished_sents.mul(eos_in_sents.long()).bool()
+                sent_lengths.masked_fill_(is_sents_unfinished_and_token_to_add_is_eos, cur_len)
+                # unfinished_sents is set to zero if eos in sentence
+                unfinished_sents.mul_((~eos_in_sents).long())
+
+            # stop when there is a </s> in each sentence, or if we exceed the maximul length
+            if unfinished_sents.max() == 0:
+                break
+
+            # extend attention_mask for new generated input if only decoder
+            if self.config.is_encoder_decoder is False:
+                attention_mask = torch.cat(
+                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+                )
+
+        return input_ids
+
+    def _generate_beam_search(
+        self,
+        input_ids,
+        cur_len,
+        max_length,
+        min_length,
+        do_sample,
+        early_stopping,
+        temperature,
+        top_k,
+        top_p,
+        repetition_penalty,
+        no_repeat_ngram_size,
+        bad_words_ids,
+        pad_token_id,
+        eos_token_id,
+        batch_size,
+        num_return_sequences,
+        length_penalty,
+        num_beams,
+        vocab_size,
+        encoder_outputs,
+        attention_mask,
+        use_cache,
+        model_specific_kwargs,
+    ):
+        """ Generate sequences for each example with beam search.
+        """
+
+        # generated hypotheses
+        generated_hyps = [
+            BeamHypotheses(num_beams, max_length, length_penalty, early_stopping=early_stopping)
+            for _ in range(batch_size)
+        ]
+
+        # scores for each sentence in the beam
+        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
+
+        # for greedy decoding it is made sure that only tokens of the first beam are considered to avoid sampling the exact same tokens three times
+        if do_sample is False:
+            beam_scores[:, 1:] = -1e9
+        beam_scores = beam_scores.view(-1)  # shape (batch_size * num_beams,)
+
+        # cache compute states
+        past = (encoder_outputs, None) if encoder_outputs is not None else None
+
+        # done sentences
+        done = [False for _ in range(batch_size)]
+
+        while cur_len < max_length:
+            model_inputs = self.prepare_inputs_for_generation(
+                input_ids, past=past, attention_mask=attention_mask, use_cache=use_cache, **model_specific_kwargs
+            )
+            outputs = self(**model_inputs)  # (batch_size * num_beams, cur_len, vocab_size)
+            next_token_logits = outputs[0][:, -1, :]  # (batch_size * num_beams, vocab_size)
+
+            # if model has past, then set the past variable to speed up decoding
+            if self._use_cache(outputs, use_cache):
+                past = outputs[1]
+            if self.config.is_encoder_decoder and do_sample is False:
+                # TODO (PVP) still a bit hacky here - there might be a better solution
+                next_token_logits = self.adjust_logits_during_generation(
+                    next_token_logits, cur_len=cur_len, max_length=max_length
+                )
+
+            scores = F.log_softmax(next_token_logits, dim=-1)  # (batch_size * num_beams, vocab_size)
+
+            scores = self.postprocess_next_token_scores(
+                scores=scores,
+                input_ids=input_ids,
+                no_repeat_ngram_size=no_repeat_ngram_size,
+                bad_words_ids=bad_words_ids,
+                cur_len=cur_len,
+                min_length=min_length,
+                max_length=max_length,
+                eos_token_id=eos_token_id,
+                repetition_penalty=repetition_penalty,
+                batch_size=batch_size,
+                num_beams=num_beams,
+            )
+
+            assert scores.shape == (batch_size * num_beams, vocab_size), "Shapes of scores: {} != {}".format(
+                scores.shape, (batch_size * num_beams, vocab_size)
+            )
+
+            if do_sample:
+                _scores = scores + beam_scores[:, None].expand_as(scores)  # (batch_size * num_beams, vocab_size)
+                # Temperature
+                if temperature != 1.0:
+                    _scores = _scores / temperature
+                # Top-p/top-k filtering
+                _scores = top_k_top_p_filtering(
+                    _scores, top_k=top_k, top_p=top_p, min_tokens_to_keep=2
+                )  # (batch_size * num_beams, vocab_size)
+                # re-organize to group the beam together to sample from all beam_idxs
+                _scores = _scores.contiguous().view(
+                    batch_size, num_beams * vocab_size
+                )  # (batch_size, num_beams * vocab_size)
+
+                # Sample 2 next tokens for each beam (so we have some spare tokens and match output of greedy beam search)
+                probs = F.softmax(_scores, dim=-1)
+                next_tokens = torch.multinomial(probs, num_samples=2 * num_beams)  # (batch_size, num_beams * 2)
+                # Compute next scores
+                next_scores = torch.gather(_scores, -1, next_tokens)  # (batch_size, num_beams * 2)
+                # sort the sampled vector to make sure that the first num_beams samples are the best
+                next_scores, next_scores_indices = torch.sort(next_scores, descending=True, dim=1)
+                next_tokens = torch.gather(next_tokens, -1, next_scores_indices)  # (batch_size, num_beams * 2)
+
+            else:
+                next_scores = scores + beam_scores[:, None].expand_as(scores)  # (batch_size * num_beams, vocab_size)
+
+                # re-organize to group the beam together (we are keeping top hypothesis accross beams)
+                next_scores = next_scores.view(
+                    batch_size, num_beams * vocab_size
+                )  # (batch_size, num_beams * vocab_size)
+
+                next_scores, next_tokens = torch.topk(next_scores, 2 * num_beams, dim=1, largest=True, sorted=True)
+
+            assert next_scores.size() == next_tokens.size() == (batch_size, 2 * num_beams)
+
+            # next batch beam content
+            next_batch_beam = []
+
+            # for each sentence
+            for batch_idx in range(batch_size):
+
+                # if we are done with this sentence, add a pad token
+                if done[batch_idx]:
+                    assert (
+                        len(generated_hyps[batch_idx]) >= num_beams
+                    ), "Batch can only be done if at least {} beams have been generated".format(num_beams)
+                    assert (
+                        eos_token_id is not None and pad_token_id is not None
+                    ), "generated beams >= num_beams -> eos_token_id and pad_token have to be defined"
+                    next_batch_beam.extend([(0, pad_token_id, 0)] * num_beams)  # pad the batch
+                    continue
+
+                # next sentence beam content, this will get added to next_batch_beam
+                next_sent_beam = []
+
+                # next tokens for this sentence
+                for beam_token_rank, (beam_token_id, beam_token_score) in enumerate(
+                    zip(next_tokens[batch_idx], next_scores[batch_idx])
+                ):
+                    # get beam and token IDs
+                    beam_id = beam_token_id // vocab_size
+                    token_id = beam_token_id % vocab_size
+
+                    effective_beam_id = batch_idx * num_beams + beam_id
+                    # add to generated hypotheses if end of sentence
+                    if (eos_token_id is not None) and (token_id.item() == eos_token_id):
+                        # if beam_token does not belong to top num_beams tokens, it should not be added
+                        is_beam_token_worse_than_top_num_beams = beam_token_rank >= num_beams
+                        if is_beam_token_worse_than_top_num_beams:
+                            continue
+                        generated_hyps[batch_idx].add(
+                            input_ids[effective_beam_id].clone(), beam_token_score.item(),
+                        )
+                    else:
+                        # add next predicted token since it is not eos_token
+                        next_sent_beam.append((beam_token_score, token_id, effective_beam_id))
+
+                    # once the beam for next step is full, don't add more tokens to it.
+                    if len(next_sent_beam) == num_beams:
+                        break
+
+                # Check if we are done so that we can save a pad step if all(done)
+                done[batch_idx] = done[batch_idx] or generated_hyps[batch_idx].is_done(
+                    next_scores[batch_idx].max().item(), cur_len
+                )
+
+                # update next beam content
+                assert len(next_sent_beam) == num_beams, "Beam should always be full"
+                next_batch_beam.extend(next_sent_beam)
+                assert len(next_batch_beam) == num_beams * (batch_idx + 1), "We should have added num_beams each step"
+
+            # stop when we are done with each sentence
+            if all(done):
+                break
+
+            # sanity check / prepare next batch
+            assert len(next_batch_beam) == batch_size * num_beams
+            beam_scores = beam_scores.new([x[0] for x in next_batch_beam])
+            beam_tokens = input_ids.new([x[1] for x in next_batch_beam])
+            beam_idx = input_ids.new([x[2] for x in next_batch_beam])
+
+            # re-order batch and update current length
+            input_ids = input_ids[beam_idx, :]
+            input_ids = torch.cat([input_ids, beam_tokens.unsqueeze(1)], dim=-1)
+            cur_len = cur_len + 1
+
+            # re-order internal states
+            if past is not None:
+                past = self._reorder_cache(past, beam_idx)
+
+            # extend attention_mask for new generated input if only decoder
+            if self.config.is_encoder_decoder is False:
+                attention_mask = torch.cat(
+                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
+                )
+
+        # finalize all open beam hypotheses and add to generated hypotheses
+        for batch_idx in range(batch_size):
+            if done[batch_idx]:
+                continue
+
+            # test that beam scores match previously calculated scores if not eos and batch_idx not done
+            if eos_token_id is not None and all(
+                (token_id % vocab_size).item() != eos_token_id for token_id in next_tokens[batch_idx]
+            ):
+                assert torch.all(
+                    next_scores[batch_idx, :num_beams] == beam_scores.view(batch_size, num_beams)[batch_idx]
+                ), "If batch_idx is not done, final next scores: {} have to equal to accumulated beam_scores: {}".format(
+                    next_scores[:, :num_beams][batch_idx], beam_scores.view(batch_size, num_beams)[batch_idx],
+                )
+
+            # need to add best num_beams hypotheses to generated hyps
+            for beam_id in range(num_beams):
+                effective_beam_id = batch_idx * num_beams + beam_id
+                final_score = beam_scores[effective_beam_id].item()
+                final_tokens = input_ids[effective_beam_id]
+                generated_hyps[batch_idx].add(final_tokens, final_score)
+
+        # depending on whether greedy generation is wanted or not define different output_batch_size and output_num_return_sequences_per_batch
+        output_batch_size = batch_size if do_sample else batch_size * num_return_sequences
+        output_num_return_sequences_per_batch = 1 if do_sample else num_return_sequences
+
+        # select the best hypotheses
+        sent_lengths = input_ids.new(output_batch_size)
+        best = []
+
+        # retrieve best hypotheses
+        for i, hypotheses in enumerate(generated_hyps):
+            sorted_hyps = sorted(hypotheses.beams, key=lambda x: x[0])
+            for j in range(output_num_return_sequences_per_batch):
+                effective_batch_idx = output_num_return_sequences_per_batch * i + j
+                best_hyp = sorted_hyps.pop()[1]
+                sent_lengths[effective_batch_idx] = len(best_hyp)
+                best.append(best_hyp)
+
+        # shorter batches are padded
+        if sent_lengths.min().item() != sent_lengths.max().item():
+            assert pad_token_id is not None, "`Pad_token_id` has to be defined"
+            sent_max_len = min(sent_lengths.max().item() + 1, max_length)
+            decoded = input_ids.new(output_batch_size, sent_max_len).fill_(pad_token_id)
+
+            # fill with hypothesis and eos_token_id if necessary
+            for i, hypo in enumerate(best):
+                decoded[i, : sent_lengths[i]] = hypo
+                if sent_lengths[i] < max_length:
+                    decoded[i, sent_lengths[i]] = eos_token_id
+        else:
+            # none of the hypotheses have an eos_token
+            assert (len(hypo) == max_length for hypo in best)
+            decoded = torch.stack(best).type(torch.long).to(next(self.parameters()).device)
+
+        return decoded
+
+    @staticmethod
+    def _reorder_cache(past: Tuple, beam_idx: Tensor) -> Tuple[Tensor]:
+        return tuple(layer_past.index_select(1, beam_idx) for layer_past in past)
+
+
+def calc_banned_ngram_tokens(prev_input_ids: Tensor, num_hypos: int, no_repeat_ngram_size: int, cur_len: int) -> None:
+    """Copied from fairseq for no_repeat_ngram in beam_search"""
+    if cur_len + 1 < no_repeat_ngram_size:
+        # return no banned tokens if we haven't generated no_repeat_ngram_size tokens yet
+        return [[] for _ in range(num_hypos)]
+    generated_ngrams = [{} for _ in range(num_hypos)]
+    for idx in range(num_hypos):
+        gen_tokens = prev_input_ids[idx].tolist()
+        generated_ngram = generated_ngrams[idx]
+        for ngram in zip(*[gen_tokens[i:] for i in range(no_repeat_ngram_size)]):
+            prev_ngram_tuple = tuple(ngram[:-1])
+            generated_ngram[prev_ngram_tuple] = generated_ngram.get(prev_ngram_tuple, []) + [ngram[-1]]
+
+    def _get_generated_ngrams(hypo_idx):
+        # Before decoding the next token, prevent decoding of ngrams that have already appeared
+        start_idx = cur_len + 1 - no_repeat_ngram_size
+        ngram_idx = tuple(prev_input_ids[hypo_idx, start_idx:cur_len].tolist())
+        return generated_ngrams[hypo_idx].get(ngram_idx, [])
+
+    banned_tokens = [_get_generated_ngrams(hypo_idx) for hypo_idx in range(num_hypos)]
+    return banned_tokens
+
+
+def calc_banned_bad_words_ids(prev_input_ids: Iterable[int], bad_words_ids: Iterable[int]) -> Iterable[int]:
+    banned_tokens = []
+
+    def _tokens_match(prev_tokens, tokens):
+        if len(tokens) == 0:
+            # if bad word tokens is just one token always ban it
+            return True
+        if len(tokens) > len(prev_input_ids):
+            # if bad word tokens are longer then prev input_ids they can't be equal
+            return False
+
+        if prev_tokens[-len(tokens) :] == tokens:
+            # if tokens match
+            return True
+        else:
+            return False
+
+    for prev_input_ids_slice in prev_input_ids:
+        banned_tokens_slice = []
+
+        for banned_token_seq in bad_words_ids:
+            assert len(banned_token_seq) > 0, "Banned words token sequences {} cannot have an empty list".format(
+                bad_words_ids
+            )
+
+            if _tokens_match(prev_input_ids_slice.tolist(), banned_token_seq[:-1]) is False:
+                # if tokens do not match continue
+                continue
+
+            banned_tokens_slice.append(banned_token_seq[-1])
+
+        banned_tokens.append(banned_tokens_slice)
+
+    return banned_tokens
+
+
+def top_k_top_p_filtering(
+    logits: Tensor,
+    top_k: int = 0,
+    top_p: float = 1.0,
+    filter_value: float = -float("Inf"),
+    min_tokens_to_keep: int = 1,
+) -> Tensor:
+    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+        Args:
+            logits: logits distribution shape (batch size, vocabulary size)
+            if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+            if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+            Make sure we keep at least min_tokens_to_keep per batch example in the output
+        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    if top_k > 0:
+        top_k = min(max(top_k, min_tokens_to_keep), logits.size(-1))  # Safety check
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p < 1.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold (token with 0 are kept)
+        sorted_indices_to_remove = cumulative_probs > top_p
+        if min_tokens_to_keep > 1:
+            # Keep at least min_tokens_to_keep (set to min_tokens_to_keep-1 because we add the first one below)
+            sorted_indices_to_remove[..., :min_tokens_to_keep] = 0
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        # scatter sorted tensors to original indexing
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        logits[indices_to_remove] = filter_value
+    return logits
+
+
+class BeamHypotheses(object):
+    def __init__(self, num_beams, max_length, length_penalty, early_stopping):
+        """
+        Initialize n-best list of hypotheses.
+        """
+        self.max_length = max_length - 1  # ignoring bos_token
+        self.length_penalty = length_penalty
+        self.early_stopping = early_stopping
+        self.num_beams = num_beams
+        self.beams = []
+        self.worst_score = 1e9
+
+    def __len__(self):
+        """
+        Number of hypotheses in the list.
+        """
+        return len(self.beams)
+
+    def add(self, hyp, sum_logprobs):
+        """
+        Add a new hypothesis to the list.
+        """
+        score = sum_logprobs / len(hyp) ** self.length_penalty
+        if len(self) < self.num_beams or score > self.worst_score:
+            self.beams.append((score, hyp))
+            if len(self) > self.num_beams:
+                sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.beams)])
+                del self.beams[sorted_scores[0][1]]
+                self.worst_score = sorted_scores[1][0]
+            else:
+                self.worst_score = min(score, self.worst_score)
+
+    def is_done(self, best_sum_logprobs, cur_len):
+        """
+        If there are enough hypotheses and that none of the hypotheses being generated
+        can become better than the worst one in the heap, then we are done with this sentence.
+        """
+
+        if len(self) < self.num_beams:
+            return False
+        elif self.early_stopping:
+            return True
+        else:
+            cur_score = best_sum_logprobs / cur_len ** self.length_penalty
+            ret = self.worst_score >= cur_score
+            return ret

bert/modeling_bert.py

@@ -0,0 +1,1569 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch BERT model. """
+
+
+import logging
+import math
+import os
+import warnings
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .activations import gelu, gelu_new, swish
+from .configuration_bert import BertConfig
+from .file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_callable
+from .modeling_utils import PreTrainedModel, find_pruneable_heads_and_indices, prune_linear_layer
+
+
+logger = logging.getLogger(__name__)
+
+_TOKENIZER_FOR_DOC = "BertTokenizer"
+
+BERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "bert-base-uncased",
+    "bert-large-uncased",
+    "bert-base-cased",
+    "bert-large-cased",
+    "bert-base-multilingual-uncased",
+    "bert-base-multilingual-cased",
+    "bert-base-chinese",
+    "bert-base-german-cased",
+    "bert-large-uncased-whole-word-masking",
+    "bert-large-cased-whole-word-masking",
+    "bert-large-uncased-whole-word-masking-finetuned-squad",
+    "bert-large-cased-whole-word-masking-finetuned-squad",
+    "bert-base-cased-finetuned-mrpc",
+    "bert-base-german-dbmdz-cased",
+    "bert-base-german-dbmdz-uncased",
+    "cl-tohoku/bert-base-japanese",
+    "cl-tohoku/bert-base-japanese-whole-word-masking",
+    "cl-tohoku/bert-base-japanese-char",
+    "cl-tohoku/bert-base-japanese-char-whole-word-masking",
+    "TurkuNLP/bert-base-finnish-cased-v1",
+    "TurkuNLP/bert-base-finnish-uncased-v1",
+    "wietsedv/bert-base-dutch-cased",
+    # See all BERT models at https://huggingface.co/models?filter=bert
+]
+
+
+def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+            n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+            for n in name
+        ):
+            logger.info("Skipping {}".format("/".join(name)))
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+def mish(x):
+    return x * torch.tanh(nn.functional.softplus(x))
+
+
+ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_new": gelu_new, "mish": mish}
+
+
+BertLayerNorm = torch.nn.LayerNorm
+
+
+class BertEmbeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        if position_ids is None:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
+            position_ids = position_ids.unsqueeze(0).expand(input_shape)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=False,
+    ):
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        if encoder_hidden_states is not None:
+            mixed_key_layer = self.key(encoder_hidden_states)
+            mixed_value_layer = self.value(encoder_hidden_states)
+            attention_mask = encoder_attention_mask
+        else:
+            mixed_key_layer = self.key(hidden_states)
+            mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+        return outputs
+
+
+class BertSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = BertSelfAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=False,
+    ):
+        self_outputs = self.self(
+            hidden_states, attention_mask, head_mask, encoder_hidden_states, encoder_attention_mask, output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class BertOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BertLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = BertAttention(config)
+        self.is_decoder = config.is_decoder
+        if self.is_decoder:
+            self.crossattention = BertAttention(config)
+        self.intermediate = BertIntermediate(config)
+        self.output = BertOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=False,
+    ):
+        self_attention_outputs = self.attention(
+            hidden_states, attention_mask, head_mask, output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        if self.is_decoder and encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        outputs = (layer_output,) + outputs
+        return outputs
+
+
+class BertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=False,
+        output_hidden_states=False,
+    ):
+        all_hidden_states = ()
+        all_attentions = ()
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions,
+                )
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        # Add last layer
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+class BertPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class BertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class BertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class BertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+class BertOnlyNSPHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, pooled_output):
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return seq_relationship_score
+
+
+class BertPreTrainingHeads(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BertLMPredictionHead(config)
+        self.seq_relationship = nn.Linear(config.hidden_size, 2)
+
+    def forward(self, sequence_output, pooled_output):
+        prediction_scores = self.predictions(sequence_output)
+        seq_relationship_score = self.seq_relationship(pooled_output)
+        return prediction_scores, seq_relationship_score
+
+
+class BertPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for downloading and loading pretrained models.
+    """
+
+    config_class = BertConfig
+    load_tf_weights = load_tf_weights_in_bert
+    base_model_prefix = "bert"
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, BertLayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+BERT_START_DOCSTRING = r"""
+    This model is a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`_ sub-class.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
+    usage and behavior.
+
+    Parameters:
+        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+BERT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`transformers.BertTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.__call__` for details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            Segment token indices to indicate first and second portions of the inputs.
+            Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
+            corresponds to a `sentence B` token
+
+            `What are token type IDs? <../glossary.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
+
+            `What are position IDs? <../glossary.html#position-ids>`_
+        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        encoder_hidden_states  (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+            if the model is configured as a decoder.
+        encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask
+            is used in the cross-attention if the model is configured as a decoder.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        output_attentions (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
+"""
+
+
+@add_start_docstrings(
+    "The bare Bert Model transformer outputting raw hidden-states without any specific head on top.",
+    BERT_START_DOCSTRING,
+)
+class BertModel(BertPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well
+    as a decoder, in which case a layer of cross-attention is added between
+    the self-attention layers, following the architecture described in `Attention is all you need`_ by Ashish Vaswani,
+    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the
+    :obj:`is_decoder` argument of the configuration set to :obj:`True`; an
+    :obj:`encoder_hidden_states` is expected as an input to the forward pass.
+
+    .. _`Attention is all you need`:
+        https://arxiv.org/abs/1706.03762
+
+    """
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = BertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during pre-training.
+
+            This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[
+            1:
+        ]  # add hidden_states and attentions if they are here
+        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model with two heads on top as done during the pre-training: a `masked language modeling` head and
+    a `next sentence prediction (classification)` head. """,
+    BERT_START_DOCSTRING,
+)
+class BertForPreTraining(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertPreTrainingHeads(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        next_sentence_label=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        **kwargs
+    ):
+        r"""
+        labels (``torch.LongTensor`` of shape ``(batch_size, sequence_length)``, `optional`, defaults to :obj:`None`):
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+        next_sentence_label (``torch.LongTensor`` of shape ``(batch_size,)``, `optional`, defaults to :obj:`None`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see :obj:`input_ids` docstring)
+            Indices should be in ``[0, 1]``.
+            ``0`` indicates sequence B is a continuation of sequence A,
+            ``1`` indicates sequence B is a random sequence.
+        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
+            Used to hide legacy arguments that have been deprecated.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
+        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False
+            continuation before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+
+    Examples::
+
+        >>> from transformers import BertTokenizer, BertForPreTraining
+        >>> import torch
+
+        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        >>> model = BertForPreTraining.from_pretrained('bert-base-uncased')
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> prediction_scores, seq_relationship_scores = outputs[:2]
+
+        """
+        if "masked_lm_labels" in kwargs:
+            warnings.warn(
+                "The `masked_lm_labels` argument is deprecated and will be removed in a future version, use `labels` instead.",
+                DeprecationWarning,
+            )
+            labels = kwargs.pop("masked_lm_labels")
+        assert kwargs == {}, f"Unexpected keyword arguments: {list(kwargs.keys())}."
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output, pooled_output = outputs[:2]
+        prediction_scores, seq_relationship_score = self.cls(sequence_output, pooled_output)
+
+        outputs = (prediction_scores, seq_relationship_score,) + outputs[
+            2:
+        ]  # add hidden states and attention if they are here
+
+        if labels is not None and next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            total_loss = masked_lm_loss + next_sentence_loss
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), prediction_scores, seq_relationship_score, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model with a `language modeling` head on top for CLM fine-tuning. """, BERT_START_DOCSTRING
+)
+class BertLMHeadModel(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        assert config.is_decoder, "If you want to use `BertLMHeadModel` as a standalone, add `is_decoder=True`."
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        **kwargs
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the left-to-right language modeling loss (next word prediction).
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
+            Used to hide legacy arguments that have been deprecated.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        ltr_lm_loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Next token prediction loss.
+        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Example::
+
+        >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig
+        >>> import torch
+
+        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
+        >>> config = BertConfig.from_pretrained("bert-base-cased")
+        >>> config.is_decoder = True
+        >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+        """
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            ltr_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            outputs = (ltr_lm_loss,) + outputs
+
+        return outputs  # (ltr_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask}
+
+
+@add_start_docstrings("""Bert Model with a `language modeling` head on top. """, BERT_START_DOCSTRING)
+class BertForMaskedLM(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        assert (
+            not config.is_decoder
+        ), "If you want to use `BertForMaskedLM` make sure `config.is_decoder=False` for bi-directional self-attention."
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        **kwargs
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-100`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+        kwargs (:obj:`Dict[str, any]`, optional, defaults to `{}`):
+            Used to hide legacy arguments that have been deprecated.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        masked_lm_loss (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        prediction_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`)
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+        if "masked_lm_labels" in kwargs:
+            warnings.warn(
+                "The `masked_lm_labels` argument is deprecated and will be removed in a future version, use `labels` instead.",
+                DeprecationWarning,
+            )
+            labels = kwargs.pop("masked_lm_labels")
+        assert "lm_labels" not in kwargs, "Use `BertWithLMHead` for autoregressive language modeling task."
+        assert kwargs == {}, f"Unexpected keyword arguments: {list(kwargs.keys())}."
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        effective_batch_size = input_shape[0]
+
+        #  add a dummy token
+        assert self.config.pad_token_id is not None, "The PAD token should be defined for generation"
+        attention_mask = torch.cat([attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))], dim=-1)
+        dummy_token = torch.full(
+            (effective_batch_size, 1), self.config.pad_token_id, dtype=torch.long, device=input_ids.device
+        )
+        input_ids = torch.cat([input_ids, dummy_token], dim=1)
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask}
+
+
+@add_start_docstrings(
+    """Bert Model with a `next sentence prediction (classification)` head on top. """, BERT_START_DOCSTRING,
+)
+class BertForNextSentencePrediction(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.cls = BertOnlyNSPHead(config)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        next_sentence_label=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+        next_sentence_label (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the next sequence prediction (classification) loss. Input should be a sequence pair (see ``input_ids`` docstring)
+            Indices should be in ``[0, 1]``.
+            ``0`` indicates sequence B is a continuation of sequence A,
+            ``1`` indicates sequence B is a random sequence.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`next_sentence_label` is provided):
+            Next sequence prediction (classification) loss.
+        seq_relationship_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, 2)`):
+            Prediction scores of the next sequence prediction (classification) head (scores of True/False continuation before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+
+    Examples::
+
+        >>> from transformers import BertTokenizer, BertForNextSentencePrediction
+        >>> import torch
+
+        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        >>> model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+
+        >>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
+        >>> next_sentence = "The sky is blue due to the shorter wavelength of blue light."
+        >>> encoding = tokenizer(prompt, next_sentence, return_tensors='pt')
+
+        >>> loss, logits = model(**encoding, next_sentence_label=torch.LongTensor([1]))
+        >>> assert logits[0, 0] < logits[0, 1] # next sentence was random
+        """
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = outputs[1]
+
+        seq_relationship_score = self.cls(pooled_output)
+
+        outputs = (seq_relationship_score,) + outputs[2:]  # add hidden states and attention if they are here
+        if next_sentence_label is not None:
+            loss_fct = CrossEntropyLoss()
+            next_sentence_loss = loss_fct(seq_relationship_score.view(-1, 2), next_sentence_label.view(-1))
+            outputs = (next_sentence_loss,) + outputs
+
+        return outputs  # (next_sentence_loss), seq_relationship_score, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    BERT_START_DOCSTRING,
+)
+class BertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    BERT_START_DOCSTRING,
+)
+class BertForMultipleChoice(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, num_choices, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the multiple choice classification loss.
+            Indices should be in ``[0, ..., num_choices-1]`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above)
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape `(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Classification loss.
+        classification_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
+            `num_choices` is the second dimension of the input tensors. (see `input_ids` above).
+
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        outputs = (reshaped_logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), reshaped_logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    BERT_START_DOCSTRING,
+)
+class BertForTokenClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the token classification loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when ``labels`` is provided) :
+            Classification loss.
+        scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.num_labels)`)
+            Classification scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings(
+    """Bert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`). """,
+    BERT_START_DOCSTRING,
+)
+class BertForQuestionAnswering(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
+    @add_code_sample_docstrings(tokenizer_class=_TOKENIZER_FOR_DOC, checkpoint="bert-base-uncased")
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        start_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-start scores (before SoftMax).
+        end_scores (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length,)`):
+            Span-end scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        """
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)
+        end_logits = end_logits.squeeze(-1)
+
+        outputs = (start_logits, end_logits,) + outputs[2:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

bert/modeling_utils.py

@@ -0,0 +1,1269 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import logging
+import os
+from typing import Callable, Dict, List, Optional, Tuple
+
+import torch
+from torch import Tensor, device, dtype, nn
+from torch.nn import CrossEntropyLoss
+from torch.nn import functional as F
+
+from .activations import get_activation
+from .configuration_utils import PretrainedConfig
+from .file_utils import (
+    DUMMY_INPUTS,
+    TF2_WEIGHTS_NAME,
+    TF_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    cached_path,
+    hf_bucket_url,
+    is_remote_url,
+)
+from .generation_utils import GenerationMixin
+
+
+logger = logging.getLogger(__name__)
+
+
+try:
+    from torch.nn import Identity
+except ImportError:
+    # Older PyTorch compatibility
+    class Identity(nn.Module):
+        r"""A placeholder identity operator that is argument-insensitive.
+        """
+
+        def __init__(self, *args, **kwargs):
+            super().__init__()
+
+        def forward(self, input):
+            return input
+
+
+def find_pruneable_heads_and_indices(
+    heads: List, n_heads: int, head_size: int, already_pruned_heads: set
+) -> Tuple[set, "torch.LongTensor"]:
+    mask = torch.ones(n_heads, head_size)
+    heads = set(heads) - already_pruned_heads  # Convert to set and remove already pruned heads
+    for head in heads:
+        # Compute how many pruned heads are before the head and move the index accordingly
+        head = head - sum(1 if h < head else 0 for h in already_pruned_heads)
+        mask[head] = 0
+    mask = mask.view(-1).contiguous().eq(1)
+    index: torch.LongTensor = torch.arange(len(mask))[mask].long()
+    return heads, index
+
+
+class ModuleUtilsMixin:
+    """
+    A few utilities for torch.nn.Modules, to be used as a mixin.
+    """
+
+    def num_parameters(self, only_trainable: bool = False) -> int:
+        """
+        Get number of (optionally, trainable) parameters in the module.
+        """
+        params = filter(lambda x: x.requires_grad, self.parameters()) if only_trainable else self.parameters()
+        return sum(p.numel() for p in params)
+
+    @staticmethod
+    def _hook_rss_memory_pre_forward(module, *args, **kwargs):
+        try:
+            import psutil
+        except (ImportError):
+            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")
+
+        process = psutil.Process(os.getpid())
+        mem = process.memory_info()
+        module.mem_rss_pre_forward = mem.rss
+        return None
+
+    @staticmethod
+    def _hook_rss_memory_post_forward(module, *args, **kwargs):
+        try:
+            import psutil
+        except (ImportError):
+            raise ImportError("You need to install psutil (pip install psutil) to use memory tracing.")
+
+        process = psutil.Process(os.getpid())
+        mem = process.memory_info()
+        module.mem_rss_post_forward = mem.rss
+        mem_rss_diff = module.mem_rss_post_forward - module.mem_rss_pre_forward
+        module.mem_rss_diff = mem_rss_diff + (module.mem_rss_diff if hasattr(module, "mem_rss_diff") else 0)
+        return None
+
+    def add_memory_hooks(self):
+        """ Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.
+            Increase in memory consumption is stored in a `mem_rss_diff` attribute for each module and can be reset to zero with `model.reset_memory_hooks_state()`
+        """
+        for module in self.modules():
+            module.register_forward_pre_hook(self._hook_rss_memory_pre_forward)
+            module.register_forward_hook(self._hook_rss_memory_post_forward)
+        self.reset_memory_hooks_state()
+
+    def reset_memory_hooks_state(self):
+        for module in self.modules():
+            module.mem_rss_diff = 0
+            module.mem_rss_post_forward = 0
+            module.mem_rss_pre_forward = 0
+
+    @property
+    def device(self) -> device:
+        """
+        Get torch.device from module, assuming that the whole module has one device.
+        """
+        try:
+            return next(self.parameters()).device
+        except StopIteration:
+            # For nn.DataParallel compatibility in PyTorch 1.5
+
+            def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
+                tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+                return tuples
+
+            gen = self._named_members(get_members_fn=find_tensor_attributes)
+            first_tuple = next(gen)
+            return first_tuple[1].device
+
+    @property
+    def dtype(self) -> dtype:
+        """
+        Get torch.dtype from module, assuming that the whole module has one dtype.
+        """
+        try:
+            return next(self.parameters()).dtype
+        except StopIteration:
+            # For nn.DataParallel compatibility in PyTorch 1.5
+
+            def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
+                tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+                return tuples
+
+            gen = self._named_members(get_members_fn=find_tensor_attributes)
+            first_tuple = next(gen)
+            return first_tuple[1].dtype
+
+    def invert_attention_mask(self, encoder_attention_mask: Tensor) -> Tensor:
+        """type: torch.Tensor -> torch.Tensor"""
+        if encoder_attention_mask.dim() == 3:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+        if encoder_attention_mask.dim() == 2:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow
+        # /transformer/transformer_layers.py#L270
+        # encoder_extended_attention_mask = (encoder_extended_attention_mask ==
+        # encoder_extended_attention_mask.transpose(-1, -2))
+        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+
+        if self.dtype == torch.float16:
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e4
+        elif self.dtype == torch.float32:
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            raise ValueError(
+                "{} not recognized. `dtype` should be set to either `torch.float32` or `torch.float16`".format(
+                    self.dtype
+                )
+            )
+
+        return encoder_extended_attention_mask
+
+    def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple, device: device) -> Tensor:
+        """Makes broadcastable attention mask and causal mask so that future and maked tokens are ignored.
+
+        Arguments:
+            attention_mask: torch.Tensor with 1 indicating tokens to ATTEND to
+            input_shape: tuple, shape of input_ids
+            device: torch.Device, usually self.device
+
+        Returns:
+            torch.Tensor with dtype of attention_mask.dtype
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                batch_size, seq_length = input_shape
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def get_head_mask(self, head_mask: Tensor, num_hidden_layers: int, is_attention_chunked: bool = False) -> Tensor:
+        """
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        attention_probs has shape bsz x n_heads x N x N
+        Arguments:
+            head_mask: torch.Tensor or None: has shape [num_heads] or [num_hidden_layers x num_heads]
+            num_hidden_layers: int
+        Returns:
+             Tensor of shape shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+             or list with [None] for each layer
+        """
+        if head_mask is not None:
+            head_mask = self._convert_head_mask_to_5d(head_mask, num_hidden_layers)
+            if is_attention_chunked is True:
+                head_mask = head_mask.unsqueeze(-1)
+        else:
+            head_mask = [None] * num_hidden_layers
+
+        return head_mask
+
+    def _convert_head_mask_to_5d(self, head_mask, num_hidden_layers):
+        """-> [num_hidden_layers x batch x num_heads x seq_length x seq_length]"""
+        if head_mask.dim() == 1:
+            head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+            head_mask = head_mask.expand(num_hidden_layers, -1, -1, -1, -1)
+        elif head_mask.dim() == 2:
+            head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+        assert head_mask.dim() == 5, f"head_mask.dim != 5, instead {head_mask.dim()}"
+        head_mask = head_mask.to(dtype=self.dtype)  # switch to fload if need + fp16 compatibility
+        return head_mask
+
+
+class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin):
+    r""" Base class for all models.
+
+        :class:`~transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
+        as well as a few methods common to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
+
+        Class attributes (overridden by derived classes):
+            - ``config_class``: a class derived from :class:`~transformers.PretrainedConfig` to use as configuration class for this model architecture.
+            - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:
+
+                - ``model``: an instance of the relevant subclass of :class:`~transformers.PreTrainedModel`,
+                - ``config``: an instance of the relevant subclass of :class:`~transformers.PretrainedConfig`,
+                - ``path``: a path (string) to the TensorFlow checkpoint.
+
+            - ``base_model_prefix``: a string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.
+    """
+    config_class = None
+    base_model_prefix = ""
+
+    @property
+    def dummy_inputs(self):
+        """ Dummy inputs to do a forward pass in the network.
+
+        Returns:
+            torch.Tensor with dummy inputs
+        """
+        return {"input_ids": torch.tensor(DUMMY_INPUTS)}
+
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__()
+        if not isinstance(config, PretrainedConfig):
+            raise ValueError(
+                "Parameter config in `{}(config)` should be an instance of class `PretrainedConfig`. "
+                "To create a model from a pretrained model use "
+                "`model = {}.from_pretrained(PRETRAINED_MODEL_NAME)`".format(
+                    self.__class__.__name__, self.__class__.__name__
+                )
+            )
+        # Save config in model
+        self.config = config
+
+    @property
+    def base_model(self):
+        return getattr(self, self.base_model_prefix, self)
+
+    def get_input_embeddings(self):
+        """
+        Returns the model's input embeddings.
+
+        Returns:
+            :obj:`nn.Module`:
+                A torch module mapping vocabulary to hidden states.
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            return base_model.get_input_embeddings()
+        else:
+            raise NotImplementedError
+
+    def set_input_embeddings(self, value: nn.Module):
+        """
+        Set model's input embeddings
+
+        Args:
+            value (:obj:`nn.Module`):
+                A module mapping vocabulary to hidden states.
+        """
+        base_model = getattr(self, self.base_model_prefix, self)
+        if base_model is not self:
+            base_model.set_input_embeddings(value)
+        else:
+            raise NotImplementedError
+
+    def get_output_embeddings(self):
+        """
+        Returns the model's output embeddings.
+
+        Returns:
+            :obj:`nn.Module`:
+                A torch module mapping hidden states to vocabulary.
+        """
+        return None  # Overwrite for models with output embeddings
+
+    def tie_weights(self):
+        """
+        Tie the weights between the input embeddings and the output embeddings.
+        If the `torchscript` flag is set in the configuration, can't handle parameter sharing so we are cloning
+        the weights instead.
+        """
+        output_embeddings = self.get_output_embeddings()
+        if output_embeddings is not None:
+            self._tie_or_clone_weights(output_embeddings, self.get_input_embeddings())
+
+    def _tie_or_clone_weights(self, output_embeddings, input_embeddings):
+        """ Tie or clone module weights depending of whether we are using TorchScript or not
+        """
+        if self.config.torchscript:
+            output_embeddings.weight = nn.Parameter(input_embeddings.weight.clone())
+        else:
+            output_embeddings.weight = input_embeddings.weight
+
+        if getattr(output_embeddings, "bias", None) is not None:
+            output_embeddings.bias.data = torch.nn.functional.pad(
+                output_embeddings.bias.data,
+                (0, output_embeddings.weight.shape[0] - output_embeddings.bias.shape[0],),
+                "constant",
+                0,
+            )
+        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
+            output_embeddings.out_features = input_embeddings.num_embeddings
+
+    def resize_token_embeddings(self, new_num_tokens: Optional[int] = None):
+        """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
+        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
+
+        Arguments:
+
+            new_num_tokens: (`optional`) int:
+                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end.
+                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
+
+        Return: ``torch.nn.Embeddings``
+            Pointer to the input tokens Embeddings Module of the model
+        """
+        base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+        model_embeds = base_model._resize_token_embeddings(new_num_tokens)
+        if new_num_tokens is None:
+            return model_embeds
+
+        # Update base model and current model config
+        self.config.vocab_size = new_num_tokens
+        base_model.vocab_size = new_num_tokens
+
+        # Tie weights again if needed
+        self.tie_weights()
+
+        return model_embeds
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.get_input_embeddings()
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.set_input_embeddings(new_embeddings)
+        return self.get_input_embeddings()
+
+    def _get_resized_embeddings(
+        self, old_embeddings: torch.nn.Embedding, new_num_tokens: Optional[int] = None
+    ) -> torch.nn.Embedding:
+        """ Build a resized Embedding Module from a provided token Embedding Module.
+            Increasing the size will add newly initialized vectors at the end
+            Reducing the size will remove vectors from the end
+
+        Args:
+            old_embeddings: ``torch.nn.Embedding``
+                Old embeddings to be resized.
+            new_num_tokens: (`optional`) int
+                New number of tokens in the embedding matrix.
+                Increasing the size will add newly initialized vectors at the end
+                Reducing the size will remove vectors from the end
+                If not provided or None: return the provided token Embedding Module.
+        Return: ``torch.nn.Embedding``
+            Pointer to the resized Embedding Module or the old Embedding Module if new_num_tokens is None
+        """
+        if new_num_tokens is None:
+            return old_embeddings
+
+        old_num_tokens, old_embedding_dim = old_embeddings.weight.size()
+        if old_num_tokens == new_num_tokens:
+            return old_embeddings
+
+        # Build new embeddings
+        new_embeddings = nn.Embedding(new_num_tokens, old_embedding_dim)
+        new_embeddings.to(old_embeddings.weight.device)
+
+        # initialize all new embeddings (in particular added tokens)
+        self._init_weights(new_embeddings)
+
+        # Copy token embeddings from the previous weights
+        num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
+        new_embeddings.weight.data[:num_tokens_to_copy, :] = old_embeddings.weight.data[:num_tokens_to_copy, :]
+
+        return new_embeddings
+
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
+        # Tie weights if needed
+        self.tie_weights()
+
+    def prune_heads(self, heads_to_prune: Dict):
+        """ Prunes heads of the base model.
+
+            Arguments:
+
+                heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
+                E.g. {1: [0, 2], 2: [2, 3]} will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
+        """
+        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
+        for layer, heads in heads_to_prune.items():
+            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
+            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON
+
+        self.base_model._prune_heads(heads_to_prune)
+
+    def save_pretrained(self, save_directory):
+        """ Save a model and its configuration file to a directory, so that it
+            can be re-loaded using the `:func:`~transformers.PreTrainedModel.from_pretrained`` class method.
+
+            Arguments:
+                save_directory: directory to which to save.
+        """
+        if os.path.isfile(save_directory):
+            logger.error("Provided path ({}) should be a directory, not a file".format(save_directory))
+            return
+        os.makedirs(save_directory, exist_ok=True)
+
+        # Only save the model itself if we are using distributed training
+        model_to_save = self.module if hasattr(self, "module") else self
+
+        # Attach architecture to the config
+        model_to_save.config.architectures = [model_to_save.__class__.__name__]
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(save_directory, WEIGHTS_NAME)
+
+        if getattr(self.config, "xla_device", False):
+            import torch_xla.core.xla_model as xm
+
+            if xm.is_master_ordinal():
+                # Save configuration file
+                model_to_save.config.save_pretrained(save_directory)
+            # xm.save takes care of saving only from master
+            xm.save(model_to_save.state_dict(), output_model_file)
+        else:
+            model_to_save.config.save_pretrained(save_directory)
+            torch.save(model_to_save.state_dict(), output_model_file)
+
+        logger.info("Model weights saved in {}".format(output_model_file))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r"""Instantiate a pretrained pytorch model from a pre-trained model configuration.
+
+        The model is set in evaluation mode by default using ``model.eval()`` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with ``model.train()``
+
+        The warning ``Weights from XXX not initialized from pretrained model`` means that the weights of XXX do not come pre-trained with the rest of the model.
+        It is up to you to train those weights with a downstream fine-tuning task.
+
+        The warning ``Weights from XXX not used in YYY`` means that the layer XXX is not used by YYY, therefore those weights are discarded.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+              - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+              - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
+              - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+              - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
+              - None if you are both providing the configuration and state dictionary (resp. with keyword arguments ``config`` and ``state_dict``)
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) one of:
+                - an instance of a class derived from :class:`~transformers.PretrainedConfig`, or
+                - a string valid as input to :func:`~transformers.PretrainedConfig.from_pretrained()`
+
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+                    - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                    - the model was saved using :func:`~transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                    - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~transformers.PreTrainedModel.save_pretrained` and :func:`~transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                configuration should be cached if the standard cache should not be used.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
+                Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
+
+        Examples::
+
+            # For example purposes. Not runnable.
+            model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
+            model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+        config = kwargs.pop("config", None)
+        state_dict = kwargs.pop("state_dict", None)
+        cache_dir = kwargs.pop("cache_dir", None)
+        from_tf = kwargs.pop("from_tf", False)
+        force_download = kwargs.pop("force_download", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        output_loading_info = kwargs.pop("output_loading_info", False)
+        local_files_only = kwargs.pop("local_files_only", False)
+        use_cdn = kwargs.pop("use_cdn", True)
+
+        # Load config if we don't provide a configuration
+        if not isinstance(config, PretrainedConfig):
+            config_path = config if config is not None else pretrained_model_name_or_path
+            config, model_kwargs = cls.config_class.from_pretrained(
+                config_path,
+                *model_args,
+                cache_dir=cache_dir,
+                return_unused_kwargs=True,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                **kwargs,
+            )
+        else:
+            model_kwargs = kwargs
+
+        # Load model
+        if pretrained_model_name_or_path is not None:
+            if os.path.isdir(pretrained_model_name_or_path):
+                if from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")):
+                    # Load from a TF 1.0 checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, TF_WEIGHTS_NAME + ".index")
+                elif from_tf and os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
+                    # Load from a TF 2.0 checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
+                elif os.path.isfile(os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)):
+                    # Load from a PyTorch checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME)
+                else:
+                    raise EnvironmentError(
+                        "Error no file named {} found in directory {} or `from_tf` set to False".format(
+                            [WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME + ".index"],
+                            pretrained_model_name_or_path,
+                        )
+                    )
+            elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+                archive_file = pretrained_model_name_or_path
+            elif os.path.isfile(pretrained_model_name_or_path + ".index"):
+                assert (
+                    from_tf
+                ), "We found a TensorFlow checkpoint at {}, please set from_tf to True to load from this checkpoint".format(
+                    pretrained_model_name_or_path + ".index"
+                )
+                archive_file = pretrained_model_name_or_path + ".index"
+            else:
+                archive_file = hf_bucket_url(
+                    pretrained_model_name_or_path,
+                    filename=(TF2_WEIGHTS_NAME if from_tf else WEIGHTS_NAME),
+                    use_cdn=use_cdn,
+                )
+                # pytorch_model.bin
+                # https://cdn.huggingface.co/bert-base-uncased-pytorch_model.bin
+            try:
+                # Load from URL or cache if already cached
+                resolved_archive_file = cached_path(
+                    archive_file,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    proxies=proxies,
+                    resume_download=resume_download,
+                    local_files_only=local_files_only,
+                )
+                if resolved_archive_file is None:
+                    raise EnvironmentError
+            except EnvironmentError:
+                msg = (
+                    f"Can't load weights for '{pretrained_model_name_or_path}'. Make sure that:\n\n"
+                    f"- '{pretrained_model_name_or_path}' is a correct model identifier listed on 'https://huggingface.co/models'\n\n"
+                    f"- or '{pretrained_model_name_or_path}' is the correct path to a directory containing a file named one of {WEIGHTS_NAME}, {TF2_WEIGHTS_NAME}, {TF_WEIGHTS_NAME}.\n\n"
+                )
+                raise EnvironmentError(msg)
+
+            if resolved_archive_file == archive_file:
+                logger.info("loading weights file {}".format(archive_file))
+            else:
+                logger.info("loading weights file {} from cache at {}".format(archive_file, resolved_archive_file))
+        else:
+            resolved_archive_file = None
+
+        # Instantiate model.
+        model = cls(config, *model_args, **model_kwargs)
+
+        if state_dict is None and not from_tf:
+            try:
+                state_dict = torch.load(resolved_archive_file, map_location="cpu")
+            except Exception:
+                raise OSError(
+                    "Unable to load weights from pytorch checkpoint file. "
+                    "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True. "
+                )
+
+        missing_keys = []
+        unexpected_keys = []
+        error_msgs = []
+
+        if from_tf:
+            if resolved_archive_file.endswith(".index"):
+                # Load from a TensorFlow 1.X checkpoint - provided by original authors
+                model = cls.load_tf_weights(model, config, resolved_archive_file[:-6])  # Remove the '.index'
+            else:
+                # Load from our TensorFlow 2.0 checkpoints
+                try:
+                    from transformers import load_tf2_checkpoint_in_pytorch_model
+
+                    model = load_tf2_checkpoint_in_pytorch_model(model, resolved_archive_file, allow_missing_keys=True)
+                except ImportError:
+                    logger.error(
+                        "Loading a TensorFlow model in PyTorch, requires both PyTorch and TensorFlow to be installed. Please see "
+                        "https://pytorch.org/ and https://www.tensorflow.org/install/ for installation instructions."
+                    )
+                    raise
+        else:
+            # Convert old format to new format if needed from a PyTorch state_dict
+            old_keys = []
+            new_keys = []
+            for key in state_dict.keys():
+                new_key = None
+                if "gamma" in key:
+                    new_key = key.replace("gamma", "weight")
+                if "beta" in key:
+                    new_key = key.replace("beta", "bias")
+                if new_key:
+                    old_keys.append(key)
+                    new_keys.append(new_key)
+            for old_key, new_key in zip(old_keys, new_keys):
+                state_dict[new_key] = state_dict.pop(old_key)
+
+            # copy state_dict so _load_from_state_dict can modify it
+            metadata = getattr(state_dict, "_metadata", None)
+            state_dict = state_dict.copy()
+            if metadata is not None:
+                state_dict._metadata = metadata
+
+            ##############################################################################################
+            # Print out state_dict's contents: keys
+            '''
+            for key, _ in state_dict.items():
+                print(key)
+            '''
+            ##############################################################################################
+
+
+            # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+            # so we need to apply the function recursively.
+            def load(module: nn.Module, prefix=""):
+                local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
+                module._load_from_state_dict(
+                    state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs,
+                )
+                for name, child in module._modules.items():
+                    if child is not None:
+                        load(child, prefix + name + ".")
+
+            # Make sure we are able to load base models as well as derived models (with heads)
+            start_prefix = ""
+            model_to_load = model
+            has_prefix_module = any(s.startswith(cls.base_model_prefix) for s in state_dict.keys())
+            if not hasattr(model, cls.base_model_prefix) and has_prefix_module:
+                start_prefix = cls.base_model_prefix + "."
+            if hasattr(model, cls.base_model_prefix) and not has_prefix_module:
+                model_to_load = getattr(model, cls.base_model_prefix)
+
+            load(model_to_load, prefix=start_prefix)
+
+            if model.__class__.__name__ != model_to_load.__class__.__name__:
+                base_model_state_dict = model_to_load.state_dict().keys()
+                head_model_state_dict_without_base_prefix = [
+                    key.split(cls.base_model_prefix + ".")[-1] for key in model.state_dict().keys()
+                ]
+
+                missing_keys.extend(head_model_state_dict_without_base_prefix - base_model_state_dict)
+
+            if len(unexpected_keys) > 0:
+                logger.warning(
+                    f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when "
+                    f"initializing {model.__class__.__name__}: {unexpected_keys}\n"
+                    f"- This IS expected if you are initializing {model.__class__.__name__} from the checkpoint of a model trained on another task "
+                    f"or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPretraining model).\n"
+                    f"- This IS NOT expected if you are initializing {model.__class__.__name__} from the checkpoint of a model that you expect "
+                    f"to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model)."
+                )
+            else:
+                logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+            if len(missing_keys) > 0:
+                logger.warning(
+                    f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at {pretrained_model_name_or_path} "
+                    f"and are newly initialized: {missing_keys}\n"
+                    f"You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+                )
+            else:
+                logger.info(
+                    f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at {pretrained_model_name_or_path}.\n"
+                    f"If your task is similar to the task the model of the ckeckpoint was trained on, "
+                    f"you can already use {model.__class__.__name__} for predictions without further training."
+                )
+            if len(error_msgs) > 0:
+                raise RuntimeError(
+                    "Error(s) in loading state_dict for {}:\n\t{}".format(
+                        model.__class__.__name__, "\n\t".join(error_msgs)
+                    )
+                )
+        model.tie_weights()  # make sure token embedding weights are still tied if needed
+
+        # Set model in evaluation mode to deactivate DropOut modules by default
+        model.eval()
+
+        if output_loading_info:
+            loading_info = {
+                "missing_keys": missing_keys,
+                "unexpected_keys": unexpected_keys,
+                "error_msgs": error_msgs,
+            }
+            return model, loading_info
+
+        if hasattr(config, "xla_device") and config.xla_device:
+            import torch_xla.core.xla_model as xm
+
+            model = xm.send_cpu_data_to_device(model, xm.xla_device())
+            model.to(xm.xla_device())
+
+        return model
+
+
+class Conv1D(nn.Module):
+    def __init__(self, nf, nx):
+        """ Conv1D layer as defined by Radford et al. for OpenAI GPT (and also used in GPT-2)
+            Basically works like a Linear layer but the weights are transposed
+        """
+        super().__init__()
+        self.nf = nf
+        w = torch.empty(nx, nf)
+        nn.init.normal_(w, std=0.02)
+        self.weight = nn.Parameter(w)
+        self.bias = nn.Parameter(torch.zeros(nf))
+
+    def forward(self, x):
+        size_out = x.size()[:-1] + (self.nf,)
+        x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
+        x = x.view(*size_out)
+        return x
+
+
+class PoolerStartLogits(nn.Module):
+    """ Compute SQuAD start_logits from sequence hidden states. """
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, 1)
+
+    def forward(self, hidden_states, p_mask=None):
+        """ Args:
+            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape `(batch_size, seq_len)`
+                invalid position mask such as query and special symbols (PAD, SEP, CLS)
+                1.0 means token should be masked.
+        """
+        x = self.dense(hidden_states).squeeze(-1)
+
+        if p_mask is not None:
+            if next(self.parameters()).dtype == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
+                x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerEndLogits(nn.Module):
+    """ Compute SQuAD end_logits from sequence hidden states and start token hidden state.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dense_1 = nn.Linear(config.hidden_size, 1)
+
+    def forward(self, hidden_states, start_states=None, start_positions=None, p_mask=None):
+        """ Args:
+            One of ``start_states``, ``start_positions`` should be not None.
+            If both are set, ``start_positions`` overrides ``start_states``.
+
+            **start_states**: ``torch.LongTensor`` of shape identical to hidden_states
+                hidden states of the first tokens for the labeled span.
+            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+                position of the first token for the labeled span:
+            **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
+                Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
+                1.0 means token should be masked.
+        """
+        assert (
+            start_states is not None or start_positions is not None
+        ), "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            slen, hsz = hidden_states.shape[-2:]
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions)  # shape (bsz, 1, hsz)
+            start_states = start_states.expand(-1, slen, -1)  # shape (bsz, slen, hsz)
+
+        x = self.dense_0(torch.cat([hidden_states, start_states], dim=-1))
+        x = self.activation(x)
+        x = self.LayerNorm(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        if p_mask is not None:
+            if next(self.parameters()).dtype == torch.float16:
+                x = x * (1 - p_mask) - 65500 * p_mask
+            else:
+                x = x * (1 - p_mask) - 1e30 * p_mask
+
+        return x
+
+
+class PoolerAnswerClass(nn.Module):
+    """ Compute SQuAD 2.0 answer class from classification and start tokens hidden states. """
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense_0 = nn.Linear(config.hidden_size * 2, config.hidden_size)
+        self.activation = nn.Tanh()
+        self.dense_1 = nn.Linear(config.hidden_size, 1, bias=False)
+
+    def forward(self, hidden_states, start_states=None, start_positions=None, cls_index=None):
+        """
+        Args:
+            One of ``start_states``, ``start_positions`` should be not None.
+            If both are set, ``start_positions`` overrides ``start_states``.
+
+            **start_states**: ``torch.LongTensor`` of shape identical to ``hidden_states``.
+                hidden states of the first tokens for the labeled span.
+            **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+                position of the first token for the labeled span.
+            **cls_index**: torch.LongTensor of shape ``(batch_size,)``
+                position of the CLS token. If None, take the last token.
+
+            note(Original repo):
+                no dependency on end_feature so that we can obtain one single `cls_logits`
+                for each sample
+        """
+        hsz = hidden_states.shape[-1]
+        assert (
+            start_states is not None or start_positions is not None
+        ), "One of start_states, start_positions should be not None"
+        if start_positions is not None:
+            start_positions = start_positions[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            start_states = hidden_states.gather(-2, start_positions).squeeze(-2)  # shape (bsz, hsz)
+
+        if cls_index is not None:
+            cls_index = cls_index[:, None, None].expand(-1, -1, hsz)  # shape (bsz, 1, hsz)
+            cls_token_state = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, hsz)
+        else:
+            cls_token_state = hidden_states[:, -1, :]  # shape (bsz, hsz)
+
+        x = self.dense_0(torch.cat([start_states, cls_token_state], dim=-1))
+        x = self.activation(x)
+        x = self.dense_1(x).squeeze(-1)
+
+        return x
+
+
+class SQuADHead(nn.Module):
+    r""" A SQuAD head inspired by XLNet.
+
+    Parameters:
+        config (:class:`~transformers.XLNetConfig`): Model configuration class with all the parameters of the model.
+
+    Inputs:
+        **hidden_states**: ``torch.FloatTensor`` of shape ``(batch_size, seq_len, hidden_size)``
+            hidden states of sequence tokens
+        **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            position of the first token for the labeled span.
+        **end_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            position of the last token for the labeled span.
+        **cls_index**: torch.LongTensor of shape ``(batch_size,)``
+            position of the CLS token. If None, take the last token.
+        **is_impossible**: ``torch.LongTensor`` of shape ``(batch_size,)``
+            Whether the question has a possible answer in the paragraph or not.
+        **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
+            Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
+            1.0 means token should be masked.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned if both ``start_positions`` and ``end_positions`` are provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification loss as the sum of start token, end token (and is_impossible if provided) classification losses.
+        **start_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top)``
+            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
+        **start_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top)``
+            Indices for the top config.start_n_top start token possibilities (beam-search).
+        **end_top_log_probs**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
+            Log probabilities for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        **end_top_index**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.LongTensor`` of shape ``(batch_size, config.start_n_top * config.end_n_top)``
+            Indices for the top ``config.start_n_top * config.end_n_top`` end token possibilities (beam-search).
+        **cls_logits**: (`optional`, returned if ``start_positions`` or ``end_positions`` is not provided)
+            ``torch.FloatTensor`` of shape ``(batch_size,)``
+            Log probabilities for the ``is_impossible`` label of the answers.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.start_n_top = config.start_n_top
+        self.end_n_top = config.end_n_top
+
+        self.start_logits = PoolerStartLogits(config)
+        self.end_logits = PoolerEndLogits(config)
+        self.answer_class = PoolerAnswerClass(config)
+
+    def forward(
+        self, hidden_states, start_positions=None, end_positions=None, cls_index=None, is_impossible=None, p_mask=None,
+    ):
+        outputs = ()
+
+        start_logits = self.start_logits(hidden_states, p_mask=p_mask)
+
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, let's remove the dimension added by batch splitting
+            for x in (start_positions, end_positions, cls_index, is_impossible):
+                if x is not None and x.dim() > 1:
+                    x.squeeze_(-1)
+
+            # during training, compute the end logits based on the ground truth of the start position
+            end_logits = self.end_logits(hidden_states, start_positions=start_positions, p_mask=p_mask)
+
+            loss_fct = CrossEntropyLoss()
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+
+            if cls_index is not None and is_impossible is not None:
+                # Predict answerability from the representation of CLS and START
+                cls_logits = self.answer_class(hidden_states, start_positions=start_positions, cls_index=cls_index)
+                loss_fct_cls = nn.BCEWithLogitsLoss()
+                cls_loss = loss_fct_cls(cls_logits, is_impossible)
+
+                # note(zhiliny): by default multiply the loss by 0.5 so that the scale is comparable to start_loss and end_loss
+                total_loss += cls_loss * 0.5
+
+            outputs = (total_loss,) + outputs
+
+        else:
+            # during inference, compute the end logits based on beam search
+            bsz, slen, hsz = hidden_states.size()
+            start_log_probs = F.softmax(start_logits, dim=-1)  # shape (bsz, slen)
+
+            start_top_log_probs, start_top_index = torch.topk(
+                start_log_probs, self.start_n_top, dim=-1
+            )  # shape (bsz, start_n_top)
+            start_top_index_exp = start_top_index.unsqueeze(-1).expand(-1, -1, hsz)  # shape (bsz, start_n_top, hsz)
+            start_states = torch.gather(hidden_states, -2, start_top_index_exp)  # shape (bsz, start_n_top, hsz)
+            start_states = start_states.unsqueeze(1).expand(-1, slen, -1, -1)  # shape (bsz, slen, start_n_top, hsz)
+
+            hidden_states_expanded = hidden_states.unsqueeze(2).expand_as(
+                start_states
+            )  # shape (bsz, slen, start_n_top, hsz)
+            p_mask = p_mask.unsqueeze(-1) if p_mask is not None else None
+            end_logits = self.end_logits(hidden_states_expanded, start_states=start_states, p_mask=p_mask)
+            end_log_probs = F.softmax(end_logits, dim=1)  # shape (bsz, slen, start_n_top)
+
+            end_top_log_probs, end_top_index = torch.topk(
+                end_log_probs, self.end_n_top, dim=1
+            )  # shape (bsz, end_n_top, start_n_top)
+            end_top_log_probs = end_top_log_probs.view(-1, self.start_n_top * self.end_n_top)
+            end_top_index = end_top_index.view(-1, self.start_n_top * self.end_n_top)
+
+            start_states = torch.einsum("blh,bl->bh", hidden_states, start_log_probs)
+            cls_logits = self.answer_class(hidden_states, start_states=start_states, cls_index=cls_index)
+
+            outputs = (start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits,) + outputs
+
+        # return start_top_log_probs, start_top_index, end_top_log_probs, end_top_index, cls_logits
+        # or (if labels are provided) (total_loss,)
+        return outputs
+
+
+class SequenceSummary(nn.Module):
+    r""" Compute a single vector summary of a sequence hidden states according to various possibilities:
+        Args of the config class:
+            summary_type:
+                - 'last' => [default] take the last token hidden state (like XLNet)
+                - 'first' => take the first token hidden state (like Bert)
+                - 'mean' => take the mean of all tokens hidden states
+                - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
+                - 'attn' => Not implemented now, use multi-head attention
+            summary_use_proj: Add a projection after the vector extraction
+            summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False.
+            summary_activation: 'tanh' or another string => add an activation to the output, Other => no activation. Default
+            summary_first_dropout: Add a dropout before the projection and activation
+            summary_last_dropout: Add a dropout after the projection and activation
+    """
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+
+        self.summary_type = getattr(config, "summary_type", "last")
+        if self.summary_type == "attn":
+            # We should use a standard multi-head attention module with absolute positional embedding for that.
+            # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
+            # We can probably just use the multi-head attention module of PyTorch >=1.1.0
+            raise NotImplementedError
+
+        self.summary = Identity()
+        if hasattr(config, "summary_use_proj") and config.summary_use_proj:
+            if hasattr(config, "summary_proj_to_labels") and config.summary_proj_to_labels and config.num_labels > 0:
+                num_classes = config.num_labels
+            else:
+                num_classes = config.hidden_size
+            self.summary = nn.Linear(config.hidden_size, num_classes)
+
+        activation_string = getattr(config, "summary_activation", None)
+        self.activation: Callable = (get_activation(activation_string) if activation_string else Identity())
+
+        self.first_dropout = Identity()
+        if hasattr(config, "summary_first_dropout") and config.summary_first_dropout > 0:
+            self.first_dropout = nn.Dropout(config.summary_first_dropout)
+
+        self.last_dropout = Identity()
+        if hasattr(config, "summary_last_dropout") and config.summary_last_dropout > 0:
+            self.last_dropout = nn.Dropout(config.summary_last_dropout)
+
+    def forward(self, hidden_states, cls_index=None):
+        """ hidden_states: float Tensor in shape [bsz, ..., seq_len, hidden_size], the hidden-states of the last layer.
+            cls_index: [optional] position of the classification token if summary_type == 'cls_index',
+                shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states.
+                if summary_type == 'cls_index' and cls_index is None:
+                    we take the last token of the sequence as classification token
+        """
+        if self.summary_type == "last":
+            output = hidden_states[:, -1]
+        elif self.summary_type == "first":
+            output = hidden_states[:, 0]
+        elif self.summary_type == "mean":
+            output = hidden_states.mean(dim=1)
+        elif self.summary_type == "cls_index":
+            if cls_index is None:
+                cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2] - 1, dtype=torch.long,)
+            else:
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand((-1,) * (cls_index.dim() - 1) + (hidden_states.size(-1),))
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(-2)  # shape (bsz, XX, hidden_size)
+        elif self.summary_type == "attn":
+            raise NotImplementedError
+
+        output = self.first_dropout(output)
+        output = self.summary(output)
+        output = self.activation(output)
+        output = self.last_dropout(output)
+
+        return output
+
+
+def prune_linear_layer(layer, index, dim=0):
+    """ Prune a linear layer (a model parameters) to keep only entries in index.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if layer.bias is not None:
+        if dim == 1:
+            b = layer.bias.clone().detach()
+        else:
+            b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = nn.Linear(new_size[1], new_size[0], bias=layer.bias is not None).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    if layer.bias is not None:
+        new_layer.bias.requires_grad = False
+        new_layer.bias.copy_(b.contiguous())
+        new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_conv1d_layer(layer, index, dim=1):
+    """ Prune a Conv1D layer (a model parameters) to keep only entries in index.
+        A Conv1D work as a Linear layer (see e.g. BERT) but the weights are transposed.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    index = index.to(layer.weight.device)
+    W = layer.weight.index_select(dim, index).clone().detach()
+    if dim == 0:
+        b = layer.bias.clone().detach()
+    else:
+        b = layer.bias[index].clone().detach()
+    new_size = list(layer.weight.size())
+    new_size[dim] = len(index)
+    new_layer = Conv1D(new_size[1], new_size[0]).to(layer.weight.device)
+    new_layer.weight.requires_grad = False
+    new_layer.weight.copy_(W.contiguous())
+    new_layer.weight.requires_grad = True
+    new_layer.bias.requires_grad = False
+    new_layer.bias.copy_(b.contiguous())
+    new_layer.bias.requires_grad = True
+    return new_layer
+
+
+def prune_layer(layer, index, dim=None):
+    """ Prune a Conv1D or nn.Linear layer (a model parameters) to keep only entries in index.
+        Return the pruned layer as a new layer with requires_grad=True.
+        Used to remove heads.
+    """
+    if isinstance(layer, nn.Linear):
+        return prune_linear_layer(layer, index, dim=0 if dim is None else dim)
+    elif isinstance(layer, Conv1D):
+        return prune_conv1d_layer(layer, index, dim=1 if dim is None else dim)
+    else:
+        raise ValueError("Can't prune layer of class {}".format(layer.__class__))
+
+
+def apply_chunking_to_forward(
+    chunk_size: int, chunk_dim: int, forward_fn: Callable[..., torch.Tensor], *input_tensors
+) -> torch.Tensor:
+    """
+    This function chunks the `input_tensors` into smaller input tensor parts of size `chunk_size` over the dimension `chunk_dim`.
+    It then applies a layer `forward_fn` to each chunk independently to save memory.
+    If the `forward_fn` is independent across the `chunk_dim` this function will yield the
+    same result as not applying it.
+
+    Args:
+        chunk_size: int - the chunk size of a chunked tensor. `num_chunks` = `len(input_tensors[0]) / chunk_size`
+        chunk_dim: int - the dimension over which the input_tensors should be chunked
+        forward_fn: fn - the forward fn of the model
+        input_tensors: tuple(torch.Tensor) - the input tensors of `forward_fn` which are chunked
+    Returns:
+        a Tensor with the same shape the foward_fn would have given if applied
+
+
+    Examples::
+
+        # rename the usual forward() fn to forward_chunk()
+        def forward_chunk(self, hidden_states):
+            hidden_states = self.decoder(hidden_states)
+            return hidden_states
+
+        # implement a chunked forward function
+        def forward(self, hidden_states):
+            return apply_chunking_to_forward(self.chunk_size_lm_head, self.seq_len_dim, self.forward_chunk, hidden_states)
+    """
+
+    assert len(input_tensors) > 0, "{} has to be a tuple/list of tensors".format(input_tensors)
+    tensor_shape = input_tensors[0].shape
+    assert all(
+        input_tensor.shape == tensor_shape for input_tensor in input_tensors
+    ), "All input tenors have to be of the same shape"
+
+    # inspect.signature exist since python 3.5 and is a python method -> no problem with backward compability
+    num_args_in_forward_chunk_fn = len(inspect.signature(forward_fn).parameters)
+    assert num_args_in_forward_chunk_fn == len(
+        input_tensors
+    ), "forward_chunk_fn expects {} arguments, but only {} input tensors are given".format(
+        num_args_in_forward_chunk_fn, len(input_tensors)
+    )
+
+    if chunk_size > 0:
+        assert (
+            input_tensors[0].shape[chunk_dim] % chunk_size == 0
+        ), "The dimension to be chunked {} has to be a multiple of the chunk size {}".format(
+            input_tensors[0].shape[chunk_dim], chunk_size
+        )
+
+        num_chunks = input_tensors[0].shape[chunk_dim] // chunk_size
+
+        # chunk input tensor into tuples
+        input_tensors_chunks = tuple(input_tensor.chunk(num_chunks, dim=chunk_dim) for input_tensor in input_tensors)
+        # apply forward fn to every tuple
+        output_chunks = tuple(forward_fn(*input_tensors_chunk) for input_tensors_chunk in zip(*input_tensors_chunks))
+        # concatenate output at same dimension
+        return torch.cat(output_chunks, dim=chunk_dim)
+
+    return forward_fn(*input_tensors)

bert/tokenization_bert.py

@@ -0,0 +1,545 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes."""
+
+
+import collections
+import logging
+import os
+import unicodedata
+from typing import List, Optional
+
+from .tokenization_utils import PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
+
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "bert-base-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
+        "bert-large-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
+        "bert-base-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-vocab.txt",
+        "bert-large-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-vocab.txt",
+        "bert-base-multilingual-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-vocab.txt",
+        "bert-base-multilingual-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-vocab.txt",
+        "bert-base-chinese": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-vocab.txt",
+        "bert-base-german-cased": "https://int-deepset-models-bert.s3.eu-central-1.amazonaws.com/pytorch/bert-base-german-cased-vocab.txt",
+        "bert-large-uncased-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-vocab.txt",
+        "bert-large-cased-whole-word-masking": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-vocab.txt",
+        "bert-large-uncased-whole-word-masking-finetuned-squad": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-vocab.txt",
+        "bert-large-cased-whole-word-masking-finetuned-squad": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-vocab.txt",
+        "bert-base-cased-finetuned-mrpc": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-vocab.txt",
+        "bert-base-german-dbmdz-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-vocab.txt",
+        "bert-base-german-dbmdz-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-vocab.txt",
+        "TurkuNLP/bert-base-finnish-cased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/vocab.txt",
+        "TurkuNLP/bert-base-finnish-uncased-v1": "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/vocab.txt",
+        "wietsedv/bert-base-dutch-cased": "https://s3.amazonaws.com/models.huggingface.co/bert/wietsedv/bert-base-dutch-cased/vocab.txt",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "bert-base-uncased": 512,
+    "bert-large-uncased": 512,
+    "bert-base-cased": 512,
+    "bert-large-cased": 512,
+    "bert-base-multilingual-uncased": 512,
+    "bert-base-multilingual-cased": 512,
+    "bert-base-chinese": 512,
+    "bert-base-german-cased": 512,
+    "bert-large-uncased-whole-word-masking": 512,
+    "bert-large-cased-whole-word-masking": 512,
+    "bert-large-uncased-whole-word-masking-finetuned-squad": 512,
+    "bert-large-cased-whole-word-masking-finetuned-squad": 512,
+    "bert-base-cased-finetuned-mrpc": 512,
+    "bert-base-german-dbmdz-cased": 512,
+    "bert-base-german-dbmdz-uncased": 512,
+    "TurkuNLP/bert-base-finnish-cased-v1": 512,
+    "TurkuNLP/bert-base-finnish-uncased-v1": 512,
+    "wietsedv/bert-base-dutch-cased": 512,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "bert-base-uncased": {"do_lower_case": True},
+    "bert-large-uncased": {"do_lower_case": True},
+    "bert-base-cased": {"do_lower_case": False},
+    "bert-large-cased": {"do_lower_case": False},
+    "bert-base-multilingual-uncased": {"do_lower_case": True},
+    "bert-base-multilingual-cased": {"do_lower_case": False},
+    "bert-base-chinese": {"do_lower_case": False},
+    "bert-base-german-cased": {"do_lower_case": False},
+    "bert-large-uncased-whole-word-masking": {"do_lower_case": True},
+    "bert-large-cased-whole-word-masking": {"do_lower_case": False},
+    "bert-large-uncased-whole-word-masking-finetuned-squad": {"do_lower_case": True},
+    "bert-large-cased-whole-word-masking-finetuned-squad": {"do_lower_case": False},
+    "bert-base-cased-finetuned-mrpc": {"do_lower_case": False},
+    "bert-base-german-dbmdz-cased": {"do_lower_case": False},
+    "bert-base-german-dbmdz-uncased": {"do_lower_case": True},
+    "TurkuNLP/bert-base-finnish-cased-v1": {"do_lower_case": False},
+    "TurkuNLP/bert-base-finnish-uncased-v1": {"do_lower_case": True},
+    "wietsedv/bert-base-dutch-cased": {"do_lower_case": False},
+}
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    with open(vocab_file, "r", encoding="utf-8") as reader:
+        tokens = reader.readlines()
+    for index, token in enumerate(tokens):
+        token = token.rstrip("\n")
+        vocab[token] = index
+    return vocab
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a piece of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class BertTokenizer(PreTrainedTokenizer):
+    r"""
+    Constructs a BERT tokenizer. Based on WordPiece.
+
+    This tokenizer inherits from :class:`~transformers.PreTrainedTokenizer` which contains most of the methods. Users
+    should refer to the superclass for more information regarding methods.
+
+    Args:
+        vocab_file (:obj:`string`):
+            File containing the vocabulary.
+        do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to lowercase the input when tokenizing.
+        do_basic_tokenize (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to do basic tokenization before WordPiece.
+        never_split (:obj:`Iterable`, `optional`, defaults to :obj:`None`):
+            Collection of tokens which will never be split during tokenization. Only has an effect when
+            :obj:`do_basic_tokenize=True`
+        unk_token (:obj:`string`, `optional`, defaults to "[UNK]"):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (:obj:`string`, `optional`, defaults to "[SEP]"):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences
+            for sequence classification or for a text and a question for question answering.
+            It is also used as the last token of a sequence built with special tokens.
+        pad_token (:obj:`string`, `optional`, defaults to "[PAD]"):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (:obj:`string`, `optional`, defaults to "[CLS]"):
+            The classifier token which is used when doing sequence classification (classification of the whole
+            sequence instead of per-token classification). It is the first token of the sequence when built with
+            special tokens.
+        mask_token (:obj:`string`, `optional`, defaults to "[MASK]"):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        tokenize_chinese_chars (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to tokenize Chinese characters.
+            This should likely be deactivated for Japanese:
+            see: https://github.com/huggingface/transformers/issues/328
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(
+        self,
+        vocab_file,
+        do_lower_case=True,
+        do_basic_tokenize=True,
+        never_split=None,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        tokenize_chinese_chars=True,
+        **kwargs
+    ):
+        super().__init__(
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            **kwargs,
+        )
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file)
+            )
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict([(ids, tok) for tok, ids in self.vocab.items()])
+        self.do_basic_tokenize = do_basic_tokenize
+        if do_basic_tokenize:
+            self.basic_tokenizer = BasicTokenizer(
+                do_lower_case=do_lower_case, never_split=never_split, tokenize_chinese_chars=tokenize_chinese_chars
+            )
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab, unk_token=self.unk_token)
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    def get_vocab(self):
+        return dict(self.vocab, **self.added_tokens_encoder)
+
+    def _tokenize(self, text):
+        split_tokens = []
+        if self.do_basic_tokenize:
+            for token in self.basic_tokenizer.tokenize(text, never_split=self.all_special_tokens):
+
+                # If the token is part of the never_split set
+                if token in self.basic_tokenizer.never_split:
+                    split_tokens.append(token)
+                else:
+                    split_tokens += self.wordpiece_tokenizer.tokenize(token)
+        else:
+            split_tokens = self.wordpiece_tokenizer.tokenize(text)
+        return split_tokens
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str) in an id using the vocab. """
+        return self.vocab.get(token, self.vocab.get(self.unk_token))
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.ids_to_tokens.get(index, self.unk_token)
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        out_string = " ".join(tokens).replace(" ##", "").strip()
+        return out_string
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks
+        by concatenating and adding special tokens.
+        A BERT sequence has the following format:
+
+        - single sequence: ``[CLS] X [SEP]``
+        - pair of sequences: ``[CLS] A [SEP] B [SEP]``
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: list of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` method.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of ids.
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Set to True if the token list is already formatted with special tokens for the model
+
+        Returns:
+            :obj:`List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError(
+                    "You should not supply a second sequence if the provided sequence of "
+                    "ids is already formated with special tokens for the model."
+                )
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A BERT sequence pair mask has the following format:
+
+        ::
+
+            0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+            | first sequence    | second sequence |
+
+        if token_ids_1 is None, only returns the first portion of the mask (0's).
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of ids.
+            token_ids_1 (:obj:`List[int]`, `optional`, defaults to :obj:`None`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
+            sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, vocab_path):
+        """
+        Save the sentencepiece vocabulary (copy original file) and special tokens file to a directory.
+
+        Args:
+            vocab_path (:obj:`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            :obj:`Tuple(str)`: Paths to the files saved.
+        """
+        index = 0
+        if os.path.isdir(vocab_path):
+            vocab_file = os.path.join(vocab_path, VOCAB_FILES_NAMES["vocab_file"])
+        else:
+            vocab_file = vocab_path
+        with open(vocab_file, "w", encoding="utf-8") as writer:
+            for token, token_index in sorted(self.vocab.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        "Saving vocabulary to {}: vocabulary indices are not consecutive."
+                        " Please check that the vocabulary is not corrupted!".format(vocab_file)
+                    )
+                    index = token_index
+                writer.write(token + "\n")
+                index += 1
+        return (vocab_file,)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True, never_split=None, tokenize_chinese_chars=True):
+        """ Constructs a BasicTokenizer.
+
+        Args:
+            **do_lower_case**: Whether to lower case the input.
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+            **tokenize_chinese_chars**: (`optional`) boolean (default True)
+                Whether to tokenize Chinese characters.
+                This should likely be deactivated for Japanese:
+                see: https://github.com/huggingface/pytorch-pretrained-BERT/issues/328
+        """
+        if never_split is None:
+            never_split = []
+        self.do_lower_case = do_lower_case
+        self.never_split = set(never_split)
+        self.tokenize_chinese_chars = tokenize_chinese_chars
+
+    def tokenize(self, text, never_split=None):
+        """ Basic Tokenization of a piece of text.
+            Split on "white spaces" only, for sub-word tokenization, see WordPieceTokenizer.
+
+        Args:
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+        """
+        # union() returns a new set by concatenating the two sets.
+        never_split = self.never_split.union(set(never_split)) if never_split else self.never_split
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        if self.tokenize_chinese_chars:
+            text = self._tokenize_chinese_chars(text)
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token, never_split))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text, never_split=None):
+        """Splits punctuation on a piece of text."""
+        if never_split is not None and text in never_split:
+            return [text]
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if (
+            (cp >= 0x4E00 and cp <= 0x9FFF)
+            or (cp >= 0x3400 and cp <= 0x4DBF)  #
+            or (cp >= 0x20000 and cp <= 0x2A6DF)  #
+            or (cp >= 0x2A700 and cp <= 0x2B73F)  #
+            or (cp >= 0x2B740 and cp <= 0x2B81F)  #
+            or (cp >= 0x2B820 and cp <= 0x2CEAF)  #
+            or (cp >= 0xF900 and cp <= 0xFAFF)
+            or (cp >= 0x2F800 and cp <= 0x2FA1F)  #
+        ):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xFFFD or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenization."""
+
+    def __init__(self, vocab, unk_token, max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+          input = "unaffable"
+          output = ["un", "##aff", "##able"]
+
+        Args:
+          text: A single token or whitespace separated tokens. This should have
+            already been passed through `BasicTokenizer`.
+
+        Returns:
+          A list of wordpiece tokens.
+        """
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+

bert/tokenization_utils.py

@@ -0,0 +1,723 @@
+# coding=utf-8
+# Copyright 2020 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Tokenization classes for python tokenizers.
+    For fast tokenizers (provided by HuggingFace's tokenizers library) see tokenization_utils_fast.py
+"""
+
+import itertools
+import logging
+import re
+import unicodedata
+from typing import Dict, List, Optional, Tuple, Union
+
+from .file_utils import add_end_docstrings
+from .tokenization_utils_base import (
+    ENCODE_KWARGS_DOCSTRING,
+    ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING,
+    AddedToken,
+    BatchEncoding,
+    EncodedInput,
+    EncodedInputPair,
+    PaddingStrategy,
+    PreTokenizedInput,
+    PreTokenizedInputPair,
+    PreTrainedTokenizerBase,
+    TensorType,
+    TextInput,
+    TextInputPair,
+    TruncationStrategy,
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if (cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False
+
+
+def _is_end_of_word(text):
+    """Checks whether the last character in text is one of a punctuation, control or whitespace character."""
+    last_char = text[-1]
+    return bool(_is_control(last_char) | _is_punctuation(last_char) | _is_whitespace(last_char))
+
+
+def _is_start_of_word(text):
+    """Checks whether the first character in text is one of a punctuation, control or whitespace character."""
+    first_char = text[0]
+    return bool(_is_control(first_char) | _is_punctuation(first_char) | _is_whitespace(first_char))
+
+
+class PreTrainedTokenizer(PreTrainedTokenizerBase):
+    """ Base class for all slow tokenizers.
+
+    Handle all the shared methods for tokenization and special tokens as well as methods
+    downloading/caching/loading pretrained tokenizers as well as adding tokens to the vocabulary.
+
+    This class also contain the added tokens in a unified way on top of all tokenizers so we don't
+    have to handle the specific vocabulary augmentation methods of the various underlying
+    dictionary structures (BPE, sentencepiece...).
+
+    Class attributes (overridden by derived classes):
+
+    - ``vocab_files_names``: a python ``dict`` with, as keys, the ``__init__`` keyword name of each vocabulary file
+      required by the model, and as associated values, the filename for saving the associated file (string).
+    - ``pretrained_vocab_files_map``: a python ``dict of dict`` the high-level keys
+      being the ``__init__`` keyword name of each vocabulary file required by the model, the low-level being the
+      `short-cut-names` (string) of the pretrained models with, as associated values, the `url` (string) to the
+      associated pretrained vocabulary file.
+    - ``max_model_input_sizes``: a python ``dict`` with, as keys, the `short-cut-names` (string) of the pretrained
+      models, and as associated values, the maximum length of the sequence inputs of this model, or None if the
+      model has no maximum input size.
+    - ``pretrained_init_configuration``: a python ``dict`` with, as keys, the `short-cut-names` (string) of the
+      pretrained models, and as associated values, a dictionnary of specific arguments to pass to the
+      ``__init__``method of the tokenizer class for this pretrained model when loading the tokenizer with the
+      ``from_pretrained()`` method.
+
+    Args:
+        - ``model_max_length``: (`Optional`) int: the maximum length in number of tokens for the inputs to the transformer model.
+            When the tokenizer is loaded with `from_pretrained`, this will be set to the value stored for the associated
+            model in ``max_model_input_sizes`` (see above). If no value is provided, will default to VERY_LARGE_INTEGER (`int(1e30)`).
+            no associated max_length can be found in ``max_model_input_sizes``.
+        - ``padding_side``: (`Optional`) string: the side on which the model should have padding applied.
+            Should be selected between ['right', 'left']
+        - ``model_input_names``: (`Optional`) List[string]: the list of the forward pass inputs accepted by the
+            model ("token_type_ids", "attention_mask"...).
+        - ``bos_token``: (`Optional`) string: a beginning of sentence token.
+            Will be associated to ``self.bos_token`` and ``self.bos_token_id``
+        - ``eos_token``: (`Optional`) string: an end of sentence token.
+            Will be associated to ``self.eos_token`` and ``self.eos_token_id``
+        - ``unk_token``: (`Optional`) string: an unknown token.
+            Will be associated to ``self.unk_token`` and ``self.unk_token_id``
+        - ``sep_token``: (`Optional`) string: a separation token (e.g. to separate context and query in an input sequence).
+            Will be associated to ``self.sep_token`` and ``self.sep_token_id``
+        - ``pad_token``: (`Optional`) string: a padding token.
+            Will be associated to ``self.pad_token`` and ``self.pad_token_id``
+        - ``cls_token``: (`Optional`) string: a classification token (e.g. to extract a summary of an input sequence
+            leveraging self-attention along the full depth of the model).
+            Will be associated to ``self.cls_token`` and ``self.cls_token_id``
+        - ``mask_token``: (`Optional`) string: a masking token (e.g. when training a model with masked-language
+            modeling). Will be associated to ``self.mask_token`` and ``self.mask_token_id``
+        - ``additional_special_tokens``: (`Optional`) list: a list of additional special tokens.
+            Adding all special tokens here ensure they won't be split by the tokenization process.
+            Will be associated to ``self.additional_special_tokens`` and ``self.additional_special_tokens_ids``
+
+
+    .. automethod:: __call__
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        # Added tokens - We store this for both slow and fast tokenizers
+        # until the serialization of Fast tokenizers is updated
+        self.added_tokens_encoder: Dict[str, int] = {}
+        self.added_tokens_decoder: Dict[int, str] = {}
+        self.unique_no_split_tokens: List[str] = []
+
+    @property
+    def is_fast(self) -> bool:
+        return False
+
+    @property
+    def vocab_size(self) -> int:
+        """ Size of the base vocabulary (without the added tokens) """
+        raise NotImplementedError
+
+    def get_vocab(self):
+        """ Returns the vocabulary as a dict of {token: index} pairs. `tokenizer.get_vocab()[token]` is equivalent to `tokenizer.convert_tokens_to_ids(token)` when `token` is in the vocab. """
+        raise NotImplementedError()
+
+    def get_added_vocab(self) -> Dict[str, int]:
+        return self.added_tokens_encoder
+
+    def __len__(self):
+        """ Size of the full vocabulary with the added tokens """
+        return self.vocab_size + len(self.added_tokens_encoder)
+
+    def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens=False) -> int:
+        """
+        Add a list of new tokens to the tokenizer class. If the new tokens are not in the
+        vocabulary, they are added to it with indices starting from length of the current vocabulary.
+
+        Args:
+            new_tokens: string or list of string. Each string is a token to add. Tokens are only added if they are not
+                already in the vocabulary (tested by checking if the tokenizer assign the index of the ``unk_token`` to them).
+
+        Returns:
+            Number of tokens added to the vocabulary.
+
+        Examples::
+
+            # Let's see how to increase the vocabulary of Bert model and tokenizer
+            tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+            model = BertModel.from_pretrained('bert-base-uncased')
+
+            num_added_toks = tokenizer.add_tokens(['new_tok1', 'my_new-tok2'])
+            print('We have added', num_added_toks, 'tokens')
+            model.resize_token_embeddings(len(tokenizer))  # Notice: resize_token_embeddings expect to receive the full size of the new vocabulary, i.e. the length of the tokenizer.
+        """
+        new_tokens = [str(tok) for tok in new_tokens]
+
+        tokens_to_add = []
+        for token in new_tokens:
+            assert isinstance(token, str)
+            if not special_tokens and self.init_kwargs.get("do_lower_case", False):
+                token = token.lower()
+            if (
+                token != self.unk_token
+                and self.convert_tokens_to_ids(token) == self.convert_tokens_to_ids(self.unk_token)
+                and token not in tokens_to_add
+            ):
+                tokens_to_add.append(token)
+                if self.verbose:
+                    logger.info("Adding %s to the vocabulary", token)
+
+        added_tok_encoder = dict((tok, len(self) + i) for i, tok in enumerate(tokens_to_add))
+        added_tok_decoder = {v: k for k, v in added_tok_encoder.items()}
+        self.added_tokens_encoder.update(added_tok_encoder)
+        self.added_tokens_decoder.update(added_tok_decoder)
+
+        # Make sure we don't split on any special tokens (even they were already in the vocab before e.g. for Albert)
+        if special_tokens:
+            self.unique_no_split_tokens = list(set(self.unique_no_split_tokens).union(set(new_tokens)))
+        else:
+            # Or on the newly added tokens
+            self.unique_no_split_tokens = list(set(self.unique_no_split_tokens).union(set(tokens_to_add)))
+
+        return len(tokens_to_add)
+
+    def num_special_tokens_to_add(self, pair=False):
+        """
+        Returns the number of added tokens when encoding a sequence with special tokens.
+
+        Note:
+            This encodes inputs and checks the number of added tokens, and is therefore not efficient. Do not put this
+            inside your training loop.
+
+        Args:
+            pair: Returns the number of added tokens in the case of a sequence pair if set to True, returns the
+                number of added tokens in the case of a single sequence if set to False.
+
+        Returns:
+            Number of tokens added to sequences
+        """
+        token_ids_0 = []
+        token_ids_1 = []
+        return len(self.build_inputs_with_special_tokens(token_ids_0, token_ids_1 if pair else None))
+
+    def tokenize(self, text: TextInput, **kwargs):
+        """ Converts a string in a sequence of tokens (string), using the tokenizer.
+            Split in words for word-based vocabulary or sub-words for sub-word-based
+            vocabularies (BPE/SentencePieces/WordPieces).
+
+            Take care of added tokens.
+
+            Args:
+                text (:obj:`string`): The sequence to be encoded.
+                **kwargs (:obj: `dict`): Arguments passed to the model-specific `prepare_for_tokenization` preprocessing method.
+        """
+        # Simple mapping string => AddedToken for special tokens with specific tokenization behaviors
+        all_special_tokens_extended = dict(
+            (str(t), t) for t in self.all_special_tokens_extended if isinstance(t, AddedToken)
+        )
+
+        text, kwargs = self.prepare_for_tokenization(text, **kwargs)
+
+        if kwargs:
+            logger.warning(f"Keyword arguments {kwargs} not recognized.")
+
+        # TODO: should this be in the base class?
+        if self.init_kwargs.get("do_lower_case", False):
+            # convert non-special tokens to lowercase
+            escaped_special_toks = [re.escape(s_tok) for s_tok in self.all_special_tokens]
+            pattern = r"(" + r"|".join(escaped_special_toks) + r")|" + r"(.+?)"
+            text = re.sub(pattern, lambda m: m.groups()[0] or m.groups()[1].lower(), text)
+
+        def split_on_token(tok, text):
+            result = []
+            tok_extended = all_special_tokens_extended.get(tok, None)
+            split_text = text.split(tok)
+            full_word = ""
+            for i, sub_text in enumerate(split_text):
+                # AddedToken can control whitespace stripping around them.
+                # We use them for GPT2 and Roberta to have different behavior depending on the special token
+                # Cf. https://github.com/huggingface/transformers/pull/2778
+                # and https://github.com/huggingface/transformers/issues/3788
+                if isinstance(tok_extended, AddedToken):
+                    if tok_extended.single_word:
+                        # Try to avoid splitting on token
+                        if (
+                            i < len(split_text) - 1
+                            and not _is_end_of_word(sub_text)
+                            and not _is_start_of_word(split_text[i + 1])
+                        ):
+                            # Don't extract the special token
+                            full_word += sub_text + tok
+                        elif full_word:
+                            full_word += sub_text
+                            result += [full_word]
+                            full_word = ""
+                            continue
+                    # Strip white spaces on the right
+                    if tok_extended.rstrip and i > 0:
+                        # A bit counter-intuitive but we strip the left of the string
+                        # since tok_extended.rstrip means the special token is eating all white spaces on its right
+                        sub_text = sub_text.lstrip()
+                    # Strip white spaces on the left
+                    if tok_extended.lstrip and i < len(split_text) - 1:
+                        sub_text = sub_text.rstrip()  # Opposite here
+                else:
+                    # We strip left and right by default
+                    if i < len(split_text) - 1:
+                        sub_text = sub_text.rstrip()
+                    if i > 0:
+                        sub_text = sub_text.lstrip()
+
+                if i == 0 and not sub_text:
+                    result += [tok]
+                elif i == len(split_text) - 1:
+                    if sub_text:
+                        result += [sub_text]
+                    else:
+                        pass
+                else:
+                    if sub_text:
+                        result += [sub_text]
+                    result += [tok]
+            return result
+
+        def split_on_tokens(tok_list, text):
+            if not text.strip():
+                return []
+            if not tok_list:
+                return self._tokenize(text)
+
+            tokenized_text = []
+            text_list = [text]
+            for tok in tok_list:
+                tokenized_text = []
+                for sub_text in text_list:
+                    if sub_text not in self.unique_no_split_tokens:
+                        tokenized_text += split_on_token(tok, sub_text)
+                    else:
+                        tokenized_text += [sub_text]
+                text_list = tokenized_text
+
+            return list(
+                itertools.chain.from_iterable(
+                    (
+                        self._tokenize(token) if token not in self.unique_no_split_tokens else [token]
+                        for token in tokenized_text
+                    )
+                )
+            )
+
+        no_split_token = self.unique_no_split_tokens
+        tokenized_text = split_on_tokens(no_split_token, text)
+        return tokenized_text
+
+    def _tokenize(self, text, **kwargs):
+        """ Converts a string in a sequence of tokens (string), using the tokenizer.
+            Split in words for word-based vocabulary or sub-words for sub-word-based
+            vocabularies (BPE/SentencePieces/WordPieces).
+
+            Do NOT take care of added tokens.
+        """
+        raise NotImplementedError
+
+    def convert_tokens_to_ids(self, tokens):
+        """ Converts a token string (or a sequence of tokens) in a single integer id
+            (or a sequence of ids), using the vocabulary.
+        """
+        if tokens is None:
+            return None
+
+        if isinstance(tokens, str):
+            return self._convert_token_to_id_with_added_voc(tokens)
+
+        ids = []
+        for token in tokens:
+            ids.append(self._convert_token_to_id_with_added_voc(token))
+        return ids
+
+    def _convert_token_to_id_with_added_voc(self, token):
+        if token is None:
+            return None
+
+        if token in self.added_tokens_encoder:
+            return self.added_tokens_encoder[token]
+        return self._convert_token_to_id(token)
+
+    def _convert_token_to_id(self, token):
+        raise NotImplementedError
+
+    def _encode_plus(
+        self,
+        text: Union[TextInput, PreTokenizedInput, EncodedInput],
+        text_pair: Optional[Union[TextInput, PreTokenizedInput, EncodedInput]] = None,
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        is_pretokenized: bool = False,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        **kwargs
+    ) -> BatchEncoding:
+        def get_input_ids(text):
+            if isinstance(text, str):
+                tokens = self.tokenize(text, **kwargs)
+                return self.convert_tokens_to_ids(tokens)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], str):
+                if is_pretokenized:
+                    tokens = list(itertools.chain(*(self.tokenize(t, is_pretokenized=True, **kwargs) for t in text)))
+                    return self.convert_tokens_to_ids(tokens)
+                else:
+                    return self.convert_tokens_to_ids(text)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
+                return text
+            else:
+                if is_pretokenized:
+                    raise ValueError(
+                        f"Input {text} is not valid. Should be a string or a list/tuple of strings when `is_pretokenized=True`."
+                    )
+                else:
+                    raise ValueError(
+                        f"Input {text} is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
+                    )
+
+        if return_offsets_mapping:
+            raise NotImplementedError(
+                "return_offset_mapping is not available when using Python tokenizers."
+                "To use this feature, change your tokenizer to one deriving from "
+                "transformers.PreTrainedTokenizerFast."
+                "More information on available tokenizers at "
+                "https://github.com/huggingface/transformers/pull/2674"
+            )
+
+        first_ids = get_input_ids(text)
+        second_ids = get_input_ids(text_pair) if text_pair is not None else None
+
+        return self.prepare_for_model(
+            first_ids,
+            pair_ids=second_ids,
+            add_special_tokens=add_special_tokens,
+            padding=padding_strategy.value,
+            truncation=truncation_strategy.value,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_tensors=return_tensors,
+            prepend_batch_axis=True,
+            return_attention_mask=return_attention_mask,
+            return_token_type_ids=return_token_type_ids,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_length=return_length,
+            verbose=verbose,
+        )
+
+    def _batch_encode_plus(
+        self,
+        batch_text_or_text_pairs: Union[
+            List[TextInput],
+            List[TextInputPair],
+            List[PreTokenizedInput],
+            List[PreTokenizedInputPair],
+            List[EncodedInput],
+            List[EncodedInputPair],
+        ],
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        is_pretokenized: bool = False,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        **kwargs
+    ) -> BatchEncoding:
+        def get_input_ids(text):
+            if isinstance(text, str):
+                tokens = self.tokenize(text, **kwargs)
+                return self.convert_tokens_to_ids(tokens)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], str):
+                if is_pretokenized:
+                    tokens = list(itertools.chain(*(self.tokenize(t, is_pretokenized=True, **kwargs) for t in text)))
+                    return self.convert_tokens_to_ids(tokens)
+                else:
+                    return self.convert_tokens_to_ids(text)
+            elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
+                return text
+            else:
+                raise ValueError(
+                    "Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
+                )
+
+        if return_offsets_mapping:
+            raise NotImplementedError(
+                "return_offset_mapping is not available when using Python tokenizers."
+                "To use this feature, change your tokenizer to one deriving from "
+                "transformers.PreTrainedTokenizerFast."
+            )
+
+        input_ids = []
+        for ids_or_pair_ids in batch_text_or_text_pairs:
+            if not isinstance(ids_or_pair_ids, (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            elif is_pretokenized and not isinstance(ids_or_pair_ids[0], (list, tuple)):
+                ids, pair_ids = ids_or_pair_ids, None
+            else:
+                ids, pair_ids = ids_or_pair_ids
+
+            first_ids = get_input_ids(ids)
+            second_ids = get_input_ids(pair_ids) if pair_ids is not None else None
+            input_ids.append((first_ids, second_ids))
+
+        batch_outputs = self._batch_prepare_for_model(
+            input_ids,
+            add_special_tokens=add_special_tokens,
+            padding_strategy=padding_strategy,
+            truncation_strategy=truncation_strategy,
+            max_length=max_length,
+            stride=stride,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask,
+            return_token_type_ids=return_token_type_ids,
+            return_overflowing_tokens=return_overflowing_tokens,
+            return_special_tokens_mask=return_special_tokens_mask,
+            return_length=return_length,
+            return_tensors=return_tensors,
+            verbose=verbose,
+        )
+
+        return BatchEncoding(batch_outputs)
+
+    @add_end_docstrings(ENCODE_KWARGS_DOCSTRING, ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING)
+    def _batch_prepare_for_model(
+        self,
+        batch_ids_pairs: List[Union[PreTokenizedInputPair, Tuple[List[int], None]]],
+        add_special_tokens: bool = True,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        truncation_strategy: TruncationStrategy = TruncationStrategy.DO_NOT_TRUNCATE,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_tensors: Optional[str] = None,
+        return_token_type_ids: Optional[bool] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+    ) -> BatchEncoding:
+        """ Prepares a sequence of input id, or a pair of sequences of inputs ids so that it can be used by the model.
+        It adds special tokens, truncates sequences if overflowing while taking into account the special tokens and
+        manages a moving window (with user defined stride) for overflowing tokens
+
+        Args:
+            batch_ids_pairs: list of tokenized input ids or input ids pairs
+        """
+
+        batch_outputs = {}
+        for first_ids, second_ids in batch_ids_pairs:
+            outputs = self.prepare_for_model(
+                first_ids,
+                second_ids,
+                add_special_tokens=add_special_tokens,
+                padding=PaddingStrategy.DO_NOT_PAD.value,  # we pad in batch afterward
+                truncation=truncation_strategy.value,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=None,  # we pad in batch afterward
+                return_attention_mask=False,  # we pad in batch afterward
+                return_token_type_ids=return_token_type_ids,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_length=return_length,
+                return_tensors=None,  # We convert the whole batch to tensors at the end
+                prepend_batch_axis=False,
+                verbose=verbose,
+            )
+
+            for key, value in outputs.items():
+                if key not in batch_outputs:
+                    batch_outputs[key] = []
+                batch_outputs[key].append(value)
+
+        batch_outputs = self.pad(
+            batch_outputs,
+            padding=padding_strategy.value,
+            max_length=max_length,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=return_attention_mask,
+        )
+
+        batch_outputs = BatchEncoding(batch_outputs, tensor_type=return_tensors)
+
+        return batch_outputs
+
+    def prepare_for_tokenization(self, text: str, is_pretokenized=False, **kwargs) -> (str, dict):
+        """ Performs any necessary transformations before tokenization.
+
+            This method should pop the arguments from kwargs and return kwargs as well.
+            We test kwargs at the end of the encoding process to be sure all the arguments have been used.
+        """
+        return (text, kwargs)
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List, token_ids_1: Optional[List] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` method.
+
+        Args:
+            token_ids_0: list of ids (must not contain special tokens)
+            token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
+                for sequence pairs
+            already_has_special_tokens: (default False) Set to True if the token list is already formated with
+                special tokens for the model
+
+        Returns:
+            A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        return [0] * ((len(token_ids_1) if token_ids_1 else 0) + len(token_ids_0))
+
+    def convert_ids_to_tokens(
+        self, ids: Union[int, List[int]], skip_special_tokens: bool = False
+    ) -> Union[str, List[str]]:
+        """ Converts a single index or a sequence of indices (integers) in a token "
+            (resp.) a sequence of tokens (str), using the vocabulary and added tokens.
+
+            Args:
+                skip_special_tokens: Don't decode special tokens (self.all_special_tokens). Default: False
+        """
+        if isinstance(ids, int):
+            if ids in self.added_tokens_decoder:
+                return self.added_tokens_decoder[ids]
+            else:
+                return self._convert_id_to_token(ids)
+        tokens = []
+        for index in ids:
+            index = int(index)
+            if skip_special_tokens and index in self.all_special_ids:
+                continue
+            if index in self.added_tokens_decoder:
+                tokens.append(self.added_tokens_decoder[index])
+            else:
+                tokens.append(self._convert_id_to_token(index))
+        return tokens
+
+    def _convert_id_to_token(self, index: int) -> str:
+        raise NotImplementedError
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        """ Converts a sequence of tokens (string) in a single string.
+            The most simple way to do it is ' '.join(self.convert_ids_to_tokens(token_ids))
+            but we often want to remove sub-word tokenization artifacts at the same time.
+        """
+        return " ".join(self.convert_ids_to_tokens(tokens))
+
+    def decode(
+        self, token_ids: List[int], skip_special_tokens: bool = False, clean_up_tokenization_spaces: bool = True
+    ) -> str:
+        filtered_tokens = self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens)
+
+        # To avoid mixing byte-level and unicode for byte-level BPT
+        # we need to build string separatly for added tokens and byte-level tokens
+        # cf. https://github.com/huggingface/transformers/issues/1133
+        sub_texts = []
+        current_sub_text = []
+        for token in filtered_tokens:
+            if skip_special_tokens and token in self.all_special_ids:
+                continue
+            if token in self.added_tokens_encoder:
+                if current_sub_text:
+                    sub_texts.append(self.convert_tokens_to_string(current_sub_text))
+                    current_sub_text = []
+                sub_texts.append(token)
+            else:
+                current_sub_text.append(token)
+        if current_sub_text:
+            sub_texts.append(self.convert_tokens_to_string(current_sub_text))
+        text = " ".join(sub_texts)
+
+        if clean_up_tokenization_spaces:
+            clean_text = self.clean_up_tokenization(text)
+            return clean_text
+        else:
+            return text
+
+    def save_vocabulary(self, save_directory) -> Tuple[str]:
+        """ Save the tokenizer vocabulary to a directory. This method does *NOT* save added tokens
+            and special token mappings.
+
+            Please use :func:`~transformers.PreTrainedTokenizer.save_pretrained` `()` to save the full
+            Tokenizer state if you want to reload it using the :func:`~transformers.PreTrainedTokenizer.from_pretrained`
+            class method.
+        """
+        raise NotImplementedError

criterions/__init__.py

@@ -0,0 +1 @@
+from .label_smoothed_cross_entropy import AdjustLabelSmoothedCrossEntropyCriterion

criterions/label_smoothed_cross_entropy.py

@@ -0,0 +1,394 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+from dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+from fairseq import metrics, utils
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.dataclass import FairseqDataclass
+from omegaconf import II
+
+
+@dataclass
+class AdjustLabelSmoothedCrossEntropyCriterionConfig(FairseqDataclass):
+    label_smoothing: float = field(
+        default=0.0,
+        metadata={"help": "epsilon for label smoothing, 0 means no label smoothing"},
+    )
+    report_accuracy: bool = field(
+        default=False,
+        metadata={"help": "report accuracy metric"},
+    )
+    det_weight: float = field(
+        default=1.0,
+        metadata={"help": "weight of detection loss"},
+    )
+    cls_weight: float = field(
+        default=1.0,
+        metadata={"help": "weight of classification loss"},
+    )
+
+    ignore_prefix_size: int = field(
+        default=0,
+        metadata={"help": "Ignore first N tokens"},
+    )
+    ignore_eos: bool = field(
+        default=False,
+        metadata={"help": "Ignore eos token"},
+    )
+    sentence_avg: bool = II("optimization.sentence_avg")
+    drop_worst_ratio: float = field(
+        default=0.0,
+        metadata={"help": "ratio for discarding bad samples"},
+    )
+    drop_worst_after: int = field(
+        default=0,
+        metadata={"help": "steps for discarding bad samples"},
+    )
+    use_rdrop: bool = field(
+        default=False, metadata={"help": "use R-Drop"}
+    )
+    reg_alpha: float = field(
+        default=1.0, metadata={"help": "weight for R-Drop"}
+    )
+    sample_patch_num: int = field(
+        default=196, metadata={"help": "sample patches for v1"}
+    )
+    constraint_range: Optional[str] = field(
+        default=None,
+        metadata={"help": "constraint range"}
+    )
+
+
+def construct_rdrop_sample(x):
+    if isinstance(x, dict):
+        for key in x:
+            x[key] = construct_rdrop_sample(x[key])
+        return x
+    elif isinstance(x, torch.Tensor):
+        return x.repeat(2, *([1] * (x.dim() - 1)))
+    elif isinstance(x, int):
+        return x * 2
+    elif isinstance(x, np.ndarray):
+        return x.repeat(2)
+    else:
+        raise NotImplementedError
+
+
+def kl_loss(p, q):
+    p_loss = F.kl_div(p, torch.exp(q), reduction='sum')
+    q_loss = F.kl_div(q, torch.exp(p), reduction='sum')
+    loss = (p_loss + q_loss) / 2
+    return loss
+
+
+def label_smoothed_nll_loss(
+        lprobs, target, epsilon, update_num, reduce=True,
+        drop_worst_ratio=0.0, drop_worst_after=0, use_rdrop=False, reg_alpha=1.0,
+        constraint_masks=None, constraint_start=None, constraint_end=None
+):
+    if target.dim() == lprobs.dim() - 1:
+        target = target.unsqueeze(-1)
+    nll_loss = -lprobs.gather(dim=-1, index=target).squeeze(-1)
+    if constraint_masks is not None:
+        smooth_loss = -lprobs.masked_fill(~constraint_masks, 0).sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (constraint_masks.sum(1) - 1 + 1e-6)
+    elif constraint_start is not None and constraint_end is not None:
+        constraint_range = [0, 1, 2, 3] + list(range(constraint_start, constraint_end))
+        smooth_loss = -lprobs[:, constraint_range].sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (len(constraint_range) - 1 + 1e-6)
+    else:
+        smooth_loss = -lprobs.sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (lprobs.size(-1) - 1)
+    loss = (1.0 - epsilon - eps_i) * nll_loss + eps_i * smooth_loss
+    if drop_worst_ratio > 0 and update_num > drop_worst_after:
+        if use_rdrop:
+            true_batch_size = loss.size(0) // 2
+            _, indices = torch.topk(loss[:true_batch_size], k=int(true_batch_size * (1 - drop_worst_ratio)), largest=False)
+            loss = torch.cat([loss[indices], loss[indices+true_batch_size]])
+            nll_loss = torch.cat([nll_loss[indices], nll_loss[indices+true_batch_size]])
+            lprobs = torch.cat([lprobs[indices], lprobs[indices+true_batch_size]])
+        else:
+            loss, indices = torch.topk(loss, k=int(loss.shape[0] * (1 - drop_worst_ratio)), largest=False)
+            nll_loss = nll_loss[indices]
+            lprobs = lprobs[indices]
+
+
+    ntokens = loss.numel()
+    nll_loss = nll_loss.sum()
+
+    loss = loss.sum()
+    if use_rdrop:
+        true_batch_size = lprobs.size(0) // 2
+        p = lprobs[:true_batch_size]
+        q = lprobs[true_batch_size:]
+        if constraint_start is not None and constraint_end is not None:
+            constraint_range = [0, 1, 2, 3] + list(range(constraint_start, constraint_end))
+            p = p[:, constraint_range]
+            q = q[:, constraint_range]
+        loss += kl_loss(p, q) * reg_alpha
+
+    return loss, nll_loss, ntokens
+
+@register_criterion(
+    "adjust_label_smoothed_cross_entropy", dataclass=AdjustLabelSmoothedCrossEntropyCriterionConfig
+)
+class AdjustLabelSmoothedCrossEntropyCriterion(FairseqCriterion):
+    def __init__(
+            self,
+            task,
+            sentence_avg,
+            label_smoothing,
+            ignore_prefix_size=0,
+            ignore_eos=False,
+            report_accuracy=False,
+            drop_worst_ratio=0,
+            drop_worst_after=0,
+            use_rdrop=False,
+            reg_alpha=1.0,
+            sample_patch_num=196,
+            constraint_range=None,
+            det_weight=1.0,
+            cls_weight=1.0
+    ):
+        super().__init__(task)
+        self.sentence_avg = sentence_avg
+        self.eps = label_smoothing
+        self.ignore_prefix_size = ignore_prefix_size
+        self.ignore_eos = ignore_eos
+        self.report_accuracy = report_accuracy
+        self.drop_worst_ratio = drop_worst_ratio
+        self.drop_worst_after = drop_worst_after
+        self.use_rdrop = use_rdrop
+        self.reg_alpha = reg_alpha
+        self.sample_patch_num = sample_patch_num
+
+        self.det_weight = det_weight
+        self.cls_weight = cls_weight
+
+        self.constraint_start = None
+        self.constraint_end = None
+        if constraint_range is not None:
+            constraint_start, constraint_end = constraint_range.split(',')
+            self.constraint_start = int(constraint_start)
+            self.constraint_end = int(constraint_end)
+
+    def forward(self, model, sample, update_num=0, reduce=True):
+        """Compute the loss for the given sample.
+
+        Returns a tuple with three elements:
+        1) the loss
+        2) the sample size, which is used as the denominator for the gradient
+        3) logging outputs to display while training
+        """
+        if isinstance(sample, list):
+            if self.sample_patch_num > 0:
+                sample[0]['net_input']['sample_patch_num'] = self.sample_patch_num
+            loss_v1, sample_size_v1, logging_output_v1 = self.forward(model, sample[0], update_num, reduce)
+            loss_v2, sample_size_v2, logging_output_v2 = self.forward(model, sample[1], update_num, reduce)
+            loss = loss_v1 / sample_size_v1 + loss_v2 / sample_size_v2
+            sample_size = 1
+            logging_output = {
+                "loss": loss.data,
+                "loss_v1": loss_v1.data,
+                "loss_v2": loss_v2.data,
+                "nll_loss": logging_output_v1["nll_loss"].data / sample_size_v1 + logging_output_v2[
+                    "nll_loss"].data / sample_size_v2,
+                "ntokens": logging_output_v1["ntokens"] + logging_output_v2["ntokens"],
+                "nsentences": logging_output_v1["nsentences"] + logging_output_v2["nsentences"],
+                "sample_size": 1,
+                "sample_size_v1": sample_size_v1,
+                "sample_size_v2": sample_size_v2,
+            }
+            return loss, sample_size, logging_output
+
+        if self.use_rdrop:
+            construct_rdrop_sample(sample)
+
+        net_output = model(**sample["net_input"])
+        loss, nll_loss, ntokens = self.compute_loss(model, net_output, sample, update_num, det_weight=self.det_weight,
+                                                    cls_weight=self.cls_weight, reduce=reduce)
+        sample_size = (
+            sample["target"].size(0)
+        )
+        logging_output = {
+            "loss": loss.data,
+            "nll_loss": nll_loss.data,
+            "ntokens": sample["ntokens"],
+            "nsentences": sample["nsentences"],
+            "sample_size": sample_size,
+        }
+        if self.report_accuracy:
+            n_correct, total = self.compute_accuracy(model, net_output, sample)
+            logging_output["n_correct"] = utils.item(n_correct.data)
+            logging_output["total"] = utils.item(total.data)
+        return loss, sample_size, logging_output
+
+    def get_lprobs_and_target(self, model, net_output, sample):
+        conf = sample['conf'][:, None, None] if 'conf' in sample and sample['conf'] is not None else 1
+        constraint_masks = None
+        if "constraint_masks" in sample and sample["constraint_masks"] is not None:
+            constraint_masks = sample["constraint_masks"]
+            net_output[0].masked_fill_(~constraint_masks, -math.inf)
+        if self.constraint_start is not None and self.constraint_end is not None:
+            net_output[0][:, :, 4:self.constraint_start] = -math.inf
+            net_output[0][:, :, self.constraint_end:] = -math.inf
+        lprobs = model.get_normalized_probs(net_output, log_probs=True) * conf
+        target = sample["token_type"]
+        if self.ignore_prefix_size > 0:
+            lprobs = lprobs[:, self.ignore_prefix_size:, :].contiguous()
+            target = target[:, self.ignore_prefix_size:].contiguous()
+            if constraint_masks is not None:
+                constraint_masks = constraint_masks[:, self.ignore_prefix_size:, :].contiguous()
+        if self.ignore_eos:
+            bsz, seq_len, embed_dim = lprobs.size()
+            eos_indices = target.eq(self.task.tgt_dict.eos())
+            lprobs = lprobs[~eos_indices].reshape(bsz, seq_len - 1, embed_dim)
+            target = target[~eos_indices].reshape(bsz, seq_len - 1)
+            if constraint_masks is not None:
+                constraint_masks = constraint_masks[~eos_indices].reshape(bsz, seq_len - 1, embed_dim)
+        if constraint_masks is not None:
+            constraint_masks = constraint_masks.view(-1, constraint_masks.size(-1))
+
+        # index = torch.zeros(lprobs.shape[:2]).to(lprobs.device)
+        # index[:, :4] = 1   # 1 indicates the location of detection results
+
+        return lprobs.view(-1, lprobs.size(-1)), target.view(-1), constraint_masks, None  # index.view(-1)
+
+    def compute_loss(self, model, net_output, sample, update_num, det_weight=1.0, cls_weight=1.0, reduce=True):
+        b = sample['target'].shape[0]
+        lprobs, target, constraint_masks, index = self.get_lprobs_and_target(model, net_output, sample)
+        if constraint_masks is not None:
+            constraint_masks = constraint_masks[target != -1]
+        # index = index[target != self.padding_idx]
+        lprobs = lprobs[target != -1]
+        target = target[target != -1]
+
+        loss_cls, nll_loss, ntokens = label_smoothed_nll_loss(
+            lprobs,
+            target,
+            self.eps,
+            update_num,
+            reduce=reduce,
+            drop_worst_ratio=self.drop_worst_ratio,
+            drop_worst_after=self.drop_worst_after,
+            use_rdrop=self.use_rdrop,
+            reg_alpha=self.reg_alpha,
+            constraint_masks=constraint_masks,
+            constraint_start=self.constraint_start,
+            constraint_end=self.constraint_end
+        )
+        loss_cls = cls_weight * loss_cls/b
+
+        # compute regression loss
+        token_type = sample["token_type"]
+        token_type = torch.stack([token_type, token_type], -1)
+        target = sample["target"]
+        index = torch.zeros_like(target).to(target.device)
+        index[:, :2, :] = 1  # the first two tokens are bbox points; 1 indicates the location of detection results
+
+        target = target[token_type == 0]
+        index = index[token_type == 0]
+        regression_output = net_output[1].squeeze(-1)
+        regression_output = regression_output[token_type == 0]
+
+        loss_reg = F.l1_loss(target[index == 1], regression_output[index == 1]) * det_weight
+        if (index == 0).any():
+            loss_reg += F.l1_loss(target[index == 0], regression_output[index == 0])
+
+        loss = loss_reg + loss_cls
+        if update_num % 5000 == 1:
+            print(f"loss_reg: {loss_reg.item()} loss_cls: {loss_cls.item()}")
+
+        return loss, nll_loss, ntokens
+
+    def compute_accuracy(self, model, net_output, sample):
+        lprobs, target = self.get_lprobs_and_target(model, net_output, sample)
+        mask = target.ne(self.padding_idx)
+        n_correct = torch.sum(
+            lprobs.argmax(1).masked_select(mask).eq(target.masked_select(mask))
+        )
+        total = torch.sum(mask)
+        return n_correct, total
+
+    @classmethod
+    def reduce_metrics(cls, logging_outputs) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        loss_sum = sum(log.get("loss", 0) for log in logging_outputs)
+        loss_sum_v1 = sum(log.get("loss_v1", 0) for log in logging_outputs)
+        loss_sum_v2 = sum(log.get("loss_v2", 0) for log in logging_outputs)
+        nll_loss_sum = sum(log.get("nll_loss", 0) for log in logging_outputs)
+        ntokens = sum(log.get("ntokens", 0) for log in logging_outputs)
+        nsentences = sum(log.get("nsentences", 0) for log in logging_outputs)
+        sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+        sample_size_v1 = sum(log.get("sample_size_v1", 0) for log in logging_outputs)
+        sample_size_v2 = sum(log.get("sample_size_v2", 0) for log in logging_outputs)
+
+        metrics.log_scalar(
+            "loss", loss_sum / sample_size, sample_size, round=3
+        )
+        metrics.log_scalar(
+            "loss_v1", loss_sum_v1 / max(sample_size_v1, 1), max(sample_size_v1, 1), round=3
+        )
+        metrics.log_scalar(
+            "loss_v2", loss_sum_v2 / max(sample_size_v2, 1), max(sample_size_v2, 1), round=3
+        )
+        metrics.log_scalar(
+            "nll_loss", nll_loss_sum / sample_size, ntokens, round=3
+        )
+        metrics.log_derived(
+            "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)
+        )
+
+        metrics.log_scalar(
+            "ntokens", ntokens, 1, round=3
+        )
+        metrics.log_scalar(
+            "nsentences", nsentences, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size", sample_size, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size_v1", sample_size_v1, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size_v2", sample_size_v2, 1, round=3
+        )
+
+        total = utils.item(sum(log.get("total", 0) for log in logging_outputs))
+        if total > 0:
+            metrics.log_scalar("total", total)
+            n_correct = utils.item(
+                sum(log.get("n_correct", 0) for log in logging_outputs)
+            )
+            metrics.log_scalar("n_correct", n_correct)
+            metrics.log_derived(
+                "accuracy",
+                lambda meters: round(
+                    meters["n_correct"].sum * 100.0 / meters["total"].sum, 3
+                )
+                if meters["total"].sum > 0
+                else float("nan"),
+            )
+
+    @staticmethod
+    def logging_outputs_can_be_summed() -> bool:
+        """
+        Whether the logging outputs returned by `forward` can be summed
+        across workers prior to calling `reduce_metrics`. Setting this
+        to True will improves distributed training speed.
+        """
+        return True

data/__init__.py

@@ -0,0 +1,2 @@
+from .refcoco_dataset import RefcocoDataset
+from .refcoco_pretrain_dataset import RefcocoPretrainDataset

data/base_dataset.py

@@ -0,0 +1,84 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import logging
+import re
+import torch.utils.data
+from fairseq.data import FairseqDataset
+
+logger = logging.getLogger(__name__)
+
+
+class BaseDataset(FairseqDataset):
+    def __init__(self, split, dataset, bpe, src_dict, tgt_dict):
+        self.split = split
+        self.dataset = dataset
+        self.bpe = bpe
+        self.src_dict = src_dict
+        self.tgt_dict = tgt_dict
+
+        self.bos = src_dict.bos()
+        self.eos = src_dict.eos()
+        self.pad = src_dict.pad()
+        self.bos_item = torch.LongTensor([self.bos])
+        self.eos_item = torch.LongTensor([self.eos])
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def encode_text(self, text, length=None, append_bos=False, append_eos=False, use_bpe=True):
+        s = self.tgt_dict.encode_line(
+            line=self.bpe.encode(text) if use_bpe else text,
+            add_if_not_exist=False,
+            append_eos=False
+        ).long()
+        if length is not None:
+            s = s[:length]
+        if append_bos:
+            s = torch.cat([self.bos_item, s])
+        if append_eos:
+            s = torch.cat([s, self.eos_item])
+        return s
+
+    def pre_question(self, question, max_ques_words):
+        question = question.lower().lstrip(",.!?*#:;~").replace('-', ' ').replace('/', ' ')
+
+        question = re.sub(
+            r"\s{2,}",
+            ' ',
+            question,
+        )
+        question = question.rstrip('\n')
+        question = question.strip(' ')
+
+        # truncate question
+        question_words = question.split(' ')
+        if len(question_words) > max_ques_words:
+            question = ' '.join(question_words[:max_ques_words])
+
+        return question
+
+    def pre_caption(self, caption, max_words):
+        caption = caption.lower().lstrip(",.!?*#:;~").replace('-', ' ').replace('/', ' ').replace('<person>', 'person')
+
+        caption = re.sub(
+            r"\s{2,}",
+            ' ',
+            caption,
+        )
+        caption = caption.rstrip('\n')
+        caption = caption.strip(' ')
+
+        # truncate caption
+        caption_words = caption.split(' ')
+        if len(caption_words) > max_words:
+            caption = ' '.join(caption_words[:max_words])
+
+        return caption

data/create_finetuning_data.py

@@ -0,0 +1,123 @@
+from refer.refer import REFER
+import numpy as np
+from PIL import Image
+import random
+import os
+from tqdm import tqdm
+
+import pickle
+from poly_utils import is_clockwise, revert_direction, check_length, reorder_points, \
+    approximate_polygons, interpolate_polygons, image_to_base64, polygons_to_string
+
+
+max_length = 400
+
+data_root = './refer/data'
+datasets = ['refcoco', 'refcoco+', 'refcocog']
+
+image_dir = './datasets/images/mscoco/train2014'
+val_test_files = pickle.load(open("data/val_test_files.p", "rb"))
+
+combined_train_data = []
+
+for dataset in datasets:
+    if dataset == 'refcoco':
+        splits = ['train', 'val', 'testA', 'testB']
+        splitBy = 'unc'
+    elif dataset == 'refcoco+':
+        splits = ['train', 'val', 'testA', 'testB']
+        splitBy = 'unc'
+    elif dataset == 'refcocog':
+        splits = ['train', 'val']
+        splitBy = 'umd'
+
+    save_dir = f'datasets/finetune/{dataset}'
+    os.makedirs(save_dir, exist_ok=True)
+    for split in splits:
+        num_pts = []
+        max_num_pts = 0
+        file_name = os.path.join(save_dir, f"{dataset}_{split}.tsv")
+        print("creating ", file_name)
+
+        uniq_ids = []
+        image_ids = []
+        sents = []
+        coeffs_strings = []
+        img_strings = []
+
+        writer = open(file_name, 'w')
+        refer = REFER(data_root, dataset, splitBy)
+
+        ref_ids = refer.getRefIds(split=split)
+
+        for this_ref_id in tqdm(ref_ids):
+            this_img_id = refer.getImgIds(this_ref_id)
+            this_img = refer.Imgs[this_img_id[0]]
+            fn = this_img['file_name']
+            img_id = fn.split(".")[0].split("_")[-1]
+
+            # load image
+            img = Image.open(os.path.join(image_dir, this_img['file_name'])).convert("RGB")
+
+            # convert image to string
+            img_base64 = image_to_base64(img, format='jpeg')
+
+            # load mask
+            ref = refer.loadRefs(this_ref_id)
+            ref_mask = np.array(refer.getMask(ref[0])['mask'])
+            annot = np.zeros(ref_mask.shape)
+            annot[ref_mask == 1] = 1  # 255
+            annot_img = Image.fromarray(annot.astype(np.uint8), mode="P")
+            annot_base64 = image_to_base64(annot_img, format='png')
+
+            polygons = refer.getPolygon(ref[0])['polygon']
+
+            polygons_processed = []
+            for polygon in polygons:
+                # make the polygon clockwise
+                if not is_clockwise(polygon):
+                    polygon = revert_direction(polygon)
+
+                # reorder the polygon so that the first vertex is the one closest to image origin
+                polygon = reorder_points(polygon)
+                polygons_processed.append(polygon)
+
+            polygons = sorted(polygons_processed, key=lambda x: (x[0] ** 2 + x[1] ** 2, x[0], x[1]))
+            polygons_interpolated = interpolate_polygons(polygons)
+
+            polygons = approximate_polygons(polygons, 5, max_length)
+
+            pts_string = polygons_to_string(polygons)
+            pts_string_interpolated = polygons_to_string(polygons_interpolated)
+
+            # load box
+            box = refer.getRefBox(this_ref_id)  # x,y,w,h
+            x, y, w, h = box
+            box_string = f'{x},{y},{x + w},{y + h}'
+
+            max_num_pts = max(max_num_pts, check_length(polygons))
+
+            num_pts.append(check_length(polygons))
+            # load text
+            ref_sent = refer.Refs[this_ref_id]
+            for i, (sent, sent_id) in enumerate(zip(ref_sent['sentences'], ref_sent['sent_ids'])):
+                uniq_id = f"{this_ref_id}_{i}"
+                instance = '\t'.join(
+                    [uniq_id, str(this_img_id[0]), sent['sent'], box_string, pts_string, img_base64, annot_base64,
+                     pts_string_interpolated]) + '\n'
+                writer.write(instance)
+
+                if img_id not in val_test_files and split == 'train':  # filtered out val/test files
+                    combined_train_data.append(instance)
+        writer.close()
+
+random.shuffle(combined_train_data)
+file_name = os.path.join("datasets/finetune/refcoco+g_train_shuffled.tsv")
+print("creating ", file_name)
+writer = open(file_name, 'w')
+writer.writelines(combined_train_data)
+writer.close()
+
+
+
+

data/create_pretraining_data.py

@@ -0,0 +1,80 @@
+import json
+import os
+from tqdm import tqdm
+import random
+import pickle
+
+# set up image paths
+imgsfile = dict(
+    coco='mscoco/train2014',
+    vg='visual-genome',
+    saiaprtc12='saiaprtc12',
+    flickr='flickr30k'
+)
+
+# load annotation files
+f = open("datasets/annotations/instances.json")
+print("Loading annotation file")
+data = json.load(f)
+f.close()
+
+# load the validation and test image list of refcoco, refcoco+, and refcocog
+val_test_files = pickle.load(open("data/val_test_files.p", "rb"))
+
+# create result folder
+os.makedirs("datasets/pretrain", exist_ok=True)
+
+# generate training tsv file
+train_instances = data['train']
+tsv_filename = "datasets/pretrain/train_shuffled.tsv"
+writer = open(tsv_filename, 'w')
+print("generating ", tsv_filename)
+
+lines = []
+for i, data_i in enumerate(tqdm(train_instances)):
+    data_source = data_i['data_source']
+    image_id = data_i['image_id']
+    bbox = data_i['bbox']
+    expressions = data_i['expressions']
+    height, width = data_i['height'], data_i['width']
+    x, y, w, h = bbox
+    box_string = f'{x},{y},{x + w},{y + h}'
+    img_name = "COCO_train2014_%012d.jpg" if "coco" in data_source else "%d.jpg"
+    img_name = img_name % image_id
+    filepath = os.path.join(imgsfile[data_source], img_name)
+    line = '\t'.join([str(i), expressions[0].replace('\n', ''), box_string, filepath]) + '\n'
+    lines.append(line)
+
+# shuffle the training set
+random.shuffle(lines)
+
+# write training tsv file
+writer.writelines(lines)
+writer.close()
+
+# generate validation tsv files
+val_sets = ['val_refcoco_unc', 'val_refcocoplus_unc', 'val_refcocog_umd', 'val_flickr30k', 'val_referitgame_berkeley']
+for val_set in val_sets:
+    val_instances = data[val_set]
+    tsv_filename = f"datasets/pretrain/{val_set}.tsv"
+    writer = open(tsv_filename, 'w')
+    print("generating ", tsv_filename)
+
+    lines = []
+    for i, data_i in enumerate(tqdm(val_instances)):
+        data_source = data_i['data_source']
+        image_id = data_i['image_id']
+        bbox = data_i['bbox']
+        expressions = data_i['expressions']
+        height, width = data_i['height'], data_i['width']
+        x, y, w, h = bbox
+        box_string = f'{x},{y},{x + w},{y + h}'
+        img_name = "COCO_train2014_%012d.jpg" if "coco" in data_source else "%d.jpg"
+        img_name = img_name % image_id
+        filepath = os.path.join(imgsfile[data_source], img_name)
+        line = '\t'.join([str(i), expressions[0].replace('\n', ''), box_string, filepath]) + '\n'
+        lines.append(line)
+
+    # write tsv file
+    writer.writelines(lines)
+    writer.close()

data/data_utils.py

@@ -0,0 +1,606 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+try:
+    from collections.abc import Iterable
+except ImportError:
+    from collections import Iterable
+import contextlib
+import itertools
+import logging
+import re
+import warnings
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+
+from fairseq.file_io import PathManager
+from fairseq import utils
+import os
+
+logger = logging.getLogger(__name__)
+
+
+def infer_language_pair(path):
+    """Infer language pair from filename: <split>.<lang1>-<lang2>.(...).idx"""
+    src, dst = None, None
+    for filename in PathManager.ls(path):
+        parts = filename.split(".")
+        if len(parts) >= 3 and len(parts[1].split("-")) == 2:
+            return parts[1].split("-")
+    return src, dst
+
+
+def collate_tokens(
+    values,
+    pad_idx,
+    eos_idx=None,
+    left_pad=False,
+    move_eos_to_beginning=False,
+    pad_to_length=None,
+    pad_to_multiple=1,
+    pad_to_bsz=None,
+):
+    """Convert a list of 1d tensors into a padded 2d tensor."""
+    size = max(v.size(0) for v in values)
+    size = size if pad_to_length is None else max(size, pad_to_length)
+    if pad_to_multiple != 1 and size % pad_to_multiple != 0:
+        size = int(((size - 0.1) // pad_to_multiple + 1) * pad_to_multiple)
+
+    def copy_tensor(src, dst):
+        assert dst.numel() == src.numel()
+        if move_eos_to_beginning:
+            if eos_idx is None:
+                # if no eos_idx is specified, then use the last token in src
+                dst[0] = src[-1]
+            else:
+                dst[0] = eos_idx
+            dst[1:] = src[:-1]
+        else:
+            dst.copy_(src)
+
+    if values[0].dim() == 1:
+        res = values[0].new(len(values), size).fill_(pad_idx)
+    elif values[0].dim() == 2:
+        assert move_eos_to_beginning is False
+        res = values[0].new(len(values), size, values[0].size(1)).fill_(pad_idx)
+    else:
+        raise NotImplementedError
+
+    for i, v in enumerate(values):
+        copy_tensor(v, res[i][size - len(v) :] if left_pad else res[i][: len(v)])
+    return res
+
+
+def load_indexed_dataset(
+    path, dictionary=None, dataset_impl=None, combine=False, default="cached"
+):
+    """A helper function for loading indexed datasets.
+
+    Args:
+        path (str): path to indexed dataset (e.g., 'data-bin/train')
+        dictionary (~fairseq.data.Dictionary): data dictionary
+        dataset_impl (str, optional): which dataset implementation to use. If
+            not provided, it will be inferred automatically. For legacy indexed
+            data we use the 'cached' implementation by default.
+        combine (bool, optional): automatically load and combine multiple
+            datasets. For example, if *path* is 'data-bin/train', then we will
+            combine 'data-bin/train', 'data-bin/train1', ... and return a
+            single ConcatDataset instance.
+    """
+    import fairseq.data.indexed_dataset as indexed_dataset
+    from fairseq.data.concat_dataset import ConcatDataset
+
+    datasets = []
+    for k in itertools.count():
+        path_k = path + (str(k) if k > 0 else "")
+        try:
+            path_k = indexed_dataset.get_indexed_dataset_to_local(path_k)
+        except Exception as e:
+            if "StorageException: [404] Path not found" in str(e):
+                logger.warning(f"path_k: {e} not found")
+            else:
+                raise e
+
+        dataset_impl_k = dataset_impl
+        if dataset_impl_k is None:
+            dataset_impl_k = indexed_dataset.infer_dataset_impl(path_k)
+        dataset = indexed_dataset.make_dataset(
+            path_k,
+            impl=dataset_impl_k or default,
+            fix_lua_indexing=True,
+            dictionary=dictionary,
+        )
+        if dataset is None:
+            break
+        logger.info("loaded {:,} examples from: {}".format(len(dataset), path_k))
+        datasets.append(dataset)
+        if not combine:
+            break
+    if len(datasets) == 0:
+        return None
+    elif len(datasets) == 1:
+        return datasets[0]
+    else:
+        return ConcatDataset(datasets)
+
+
+@contextlib.contextmanager
+def numpy_seed(seed, *addl_seeds):
+    """Context manager which seeds the NumPy PRNG with the specified seed and
+    restores the state afterward"""
+    if seed is None:
+        yield
+        return
+    if len(addl_seeds) > 0:
+        seed = int(hash((seed, *addl_seeds)) % 1e6)
+    state = np.random.get_state()
+    np.random.seed(seed)
+    try:
+        yield
+    finally:
+        np.random.set_state(state)
+
+
+def collect_filtered(function, iterable, filtered):
+    """
+    Similar to :func:`filter` but collects filtered elements in ``filtered``.
+
+    Args:
+        function (callable): function that returns ``False`` for elements that
+            should be filtered
+        iterable (iterable): iterable to filter
+        filtered (list): list to store filtered elements
+    """
+    for el in iterable:
+        if function(el):
+            yield el
+        else:
+            filtered.append(el)
+
+
+def _filter_by_size_dynamic(indices, size_fn, max_positions, raise_exception=False):
+    def compare_leq(a, b):
+        return a <= b if not isinstance(a, tuple) else max(a) <= b
+
+    def check_size(idx):
+        if isinstance(max_positions, float) or isinstance(max_positions, int):
+            return size_fn(idx) <= max_positions
+        elif isinstance(max_positions, dict):
+            idx_size = size_fn(idx)
+            assert isinstance(idx_size, dict)
+            intersect_keys = set(max_positions.keys()) & set(idx_size.keys())
+            return all(
+                all(
+                    a is None or b is None or a <= b
+                    for a, b in zip(idx_size[key], max_positions[key])
+                )
+                for key in intersect_keys
+            )
+        else:
+            # For MultiCorpusSampledDataset, will generalize it later
+            if not isinstance(size_fn(idx), Iterable):
+                return all(size_fn(idx) <= b for b in max_positions)
+            return all(
+                a is None or b is None or a <= b
+                for a, b in zip(size_fn(idx), max_positions)
+            )
+
+    ignored = []
+    itr = collect_filtered(check_size, indices, ignored)
+    indices = np.fromiter(itr, dtype=np.int64, count=-1)
+    return indices, ignored
+
+
+def filter_by_size(indices, dataset, max_positions, raise_exception=False):
+    """
+    [deprecated] Filter indices based on their size.
+    Use `FairseqDataset::filter_indices_by_size` instead.
+
+    Args:
+        indices (List[int]): ordered list of dataset indices
+        dataset (FairseqDataset): fairseq dataset instance
+        max_positions (tuple): filter elements larger than this size.
+            Comparisons are done component-wise.
+        raise_exception (bool, optional): if ``True``, raise an exception if
+            any elements are filtered (default: False).
+    """
+    warnings.warn(
+        "data_utils.filter_by_size is deprecated. "
+        "Use `FairseqDataset::filter_indices_by_size` instead.",
+        stacklevel=2,
+    )
+    if isinstance(max_positions, float) or isinstance(max_positions, int):
+        if hasattr(dataset, "sizes") and isinstance(dataset.sizes, np.ndarray):
+            ignored = indices[dataset.sizes[indices] > max_positions].tolist()
+            indices = indices[dataset.sizes[indices] <= max_positions]
+        elif (
+            hasattr(dataset, "sizes")
+            and isinstance(dataset.sizes, list)
+            and len(dataset.sizes) == 1
+        ):
+            ignored = indices[dataset.sizes[0][indices] > max_positions].tolist()
+            indices = indices[dataset.sizes[0][indices] <= max_positions]
+        else:
+            indices, ignored = _filter_by_size_dynamic(
+                indices, dataset.size, max_positions
+            )
+    else:
+        indices, ignored = _filter_by_size_dynamic(indices, dataset.size, max_positions)
+
+    if len(ignored) > 0 and raise_exception:
+        raise Exception(
+            (
+                "Size of sample #{} is invalid (={}) since max_positions={}, "
+                "skip this example with --skip-invalid-size-inputs-valid-test"
+            ).format(ignored[0], dataset.size(ignored[0]), max_positions)
+        )
+    if len(ignored) > 0:
+        logger.warning(
+            (
+                "{} samples have invalid sizes and will be skipped, "
+                "max_positions={}, first few sample ids={}"
+            ).format(len(ignored), max_positions, ignored[:10])
+        )
+    return indices
+
+
+def filter_paired_dataset_indices_by_size(src_sizes, tgt_sizes, indices, max_sizes):
+    """Filter a list of sample indices. Remove those that are longer
+        than specified in max_sizes.
+
+    Args:
+        indices (np.array): original array of sample indices
+        max_sizes (int or list[int] or tuple[int]): max sample size,
+            can be defined separately for src and tgt (then list or tuple)
+
+    Returns:
+        np.array: filtered sample array
+        list: list of removed indices
+    """
+    if max_sizes is None:
+        return indices, []
+    if type(max_sizes) in (int, float):
+        max_src_size, max_tgt_size = max_sizes, max_sizes
+    else:
+        max_src_size, max_tgt_size = max_sizes
+    if tgt_sizes is None:
+        ignored = indices[src_sizes[indices] > max_src_size]
+    else:
+        ignored = indices[
+            (src_sizes[indices] > max_src_size) | (tgt_sizes[indices] > max_tgt_size)
+        ]
+    if len(ignored) > 0:
+        if tgt_sizes is None:
+            indices = indices[src_sizes[indices] <= max_src_size]
+        else:
+            indices = indices[
+                (src_sizes[indices] <= max_src_size)
+                & (tgt_sizes[indices] <= max_tgt_size)
+            ]
+    return indices, ignored.tolist()
+
+
+def batch_by_size(
+    indices,
+    num_tokens_fn,
+    num_tokens_vec=None,
+    max_tokens=None,
+    max_sentences=None,
+    required_batch_size_multiple=1,
+    fixed_shapes=None,
+):
+    """
+    Yield mini-batches of indices bucketed by size. Batches may contain
+    sequences of different lengths.
+
+    Args:
+        indices (List[int]): ordered list of dataset indices
+        num_tokens_fn (callable): function that returns the number of tokens at
+            a given index
+        num_tokens_vec (List[int], optional): precomputed vector of the number
+            of tokens for each index in indices (to enable faster batch generation)
+        max_tokens (int, optional): max number of tokens in each batch
+            (default: None).
+        max_sentences (int, optional): max number of sentences in each
+            batch (default: None).
+        required_batch_size_multiple (int, optional): require batch size to
+            be less than N or a multiple of N (default: 1).
+        fixed_shapes (List[Tuple[int, int]], optional): if given, batches will
+            only be created with the given shapes. *max_sentences* and
+            *required_batch_size_multiple* will be ignored (default: None).
+    """
+    try:
+        from fairseq.data.data_utils_fast import (
+            batch_by_size_fn,
+            batch_by_size_vec,
+            batch_fixed_shapes_fast,
+        )
+    except ImportError:
+        raise ImportError(
+            "Please build Cython components with: "
+            "`python setup.py build_ext --inplace`"
+        )
+    except ValueError:
+        raise ValueError(
+            "Please build (or rebuild) Cython components with `python setup.py build_ext --inplace`."
+        )
+
+    # added int() to avoid TypeError: an integer is required
+    max_tokens = (
+        int(max_tokens) if max_tokens is not None else -1
+    )
+    max_sentences = max_sentences if max_sentences is not None else -1
+    bsz_mult = required_batch_size_multiple
+
+    if not isinstance(indices, np.ndarray):
+        indices = np.fromiter(indices, dtype=np.int64, count=-1)
+
+    if num_tokens_vec is not None and not isinstance(num_tokens_vec, np.ndarray):
+        num_tokens_vec = np.fromiter(num_tokens_vec, dtype=np.int64, count=-1)
+
+    if fixed_shapes is None:
+        if num_tokens_vec is None:
+            return batch_by_size_fn(
+                indices,
+                num_tokens_fn,
+                max_tokens,
+                max_sentences,
+                bsz_mult,
+            )
+        else:
+            return batch_by_size_vec(
+                indices,
+                num_tokens_vec,
+                max_tokens,
+                max_sentences,
+                bsz_mult,
+            )
+
+    else:
+        fixed_shapes = np.array(fixed_shapes, dtype=np.int64)
+        sort_order = np.lexsort(
+            [
+                fixed_shapes[:, 1].argsort(),  # length
+                fixed_shapes[:, 0].argsort(),  # bsz
+            ]
+        )
+        fixed_shapes_sorted = fixed_shapes[sort_order]
+        return batch_fixed_shapes_fast(indices, num_tokens_fn, fixed_shapes_sorted)
+
+
+def post_process(sentence: str, symbol: str):
+    if symbol == "sentencepiece":
+        sentence = sentence.replace(" ", "").replace("\u2581", " ").strip()
+    elif symbol == "wordpiece":
+        sentence = sentence.replace(" ", "").replace("_", " ").strip()
+    elif symbol == "letter":
+        sentence = sentence.replace(" ", "").replace("|", " ").strip()
+    elif symbol == "silence":
+        import re
+        sentence = sentence.replace("<SIL>", "")
+        sentence = re.sub(' +', ' ', sentence).strip()
+    elif symbol == "_EOW":
+        sentence = sentence.replace(" ", "").replace("_EOW", " ").strip()
+    elif symbol in {"subword_nmt", "@@ ", "@@"}:
+        if symbol == "subword_nmt":
+            symbol = "@@ "
+        sentence = (sentence + " ").replace(symbol, "").rstrip()
+    elif symbol == "none":
+        pass
+    elif symbol is not None:
+        raise NotImplementedError(f"Unknown post_process option: {symbol}")
+    return sentence
+
+
+def compute_mask_indices(
+    shape: Tuple[int, int],
+    padding_mask: Optional[torch.Tensor],
+    mask_prob: float,
+    mask_length: int,
+    mask_type: str = "static",
+    mask_other: float = 0.0,
+    min_masks: int = 0,
+    no_overlap: bool = False,
+    min_space: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape
+
+    Args:
+        shape: the the shape for which to compute masks.
+            should be of size 2 where first element is batch size and 2nd is timesteps
+        padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+        mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+            number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+            however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+        mask_type: how to compute mask lengths
+            static = fixed size
+            uniform = sample from uniform distribution [mask_other, mask_length*2]
+            normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+            poisson = sample from possion distribution with lambda = mask length
+        min_masks: minimum number of masked spans
+        no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+        min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+    """
+
+    bsz, all_sz = shape
+    mask = np.full((bsz, all_sz), False)
+
+    all_num_mask = int(
+        # add a random number for probabilistic rounding
+        mask_prob * all_sz / float(mask_length)
+        + np.random.rand()
+    )
+
+    all_num_mask = max(min_masks, all_num_mask)
+
+    mask_idcs = []
+    for i in range(bsz):
+        if padding_mask is not None:
+            sz = all_sz - padding_mask[i].long().sum().item()
+            num_mask = int(
+                # add a random number for probabilistic rounding
+                mask_prob * sz / float(mask_length)
+                + np.random.rand()
+            )
+            num_mask = max(min_masks, num_mask)
+        else:
+            sz = all_sz
+            num_mask = all_num_mask
+
+        if mask_type == "static":
+            lengths = np.full(num_mask, mask_length)
+        elif mask_type == "uniform":
+            lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask)
+        elif mask_type == "normal":
+            lengths = np.random.normal(mask_length, mask_other, size=num_mask)
+            lengths = [max(1, int(round(x))) for x in lengths]
+        elif mask_type == "poisson":
+            lengths = np.random.poisson(mask_length, size=num_mask)
+            lengths = [int(round(x)) for x in lengths]
+        else:
+            raise Exception("unknown mask selection " + mask_type)
+
+        if sum(lengths) == 0:
+            lengths[0] = min(mask_length, sz - 1)
+
+        if no_overlap:
+            mask_idc = []
+
+            def arrange(s, e, length, keep_length):
+                span_start = np.random.randint(s, e - length)
+                mask_idc.extend(span_start + i for i in range(length))
+
+                new_parts = []
+                if span_start - s - min_space >= keep_length:
+                    new_parts.append((s, span_start - min_space + 1))
+                if e - span_start - keep_length - min_space > keep_length:
+                    new_parts.append((span_start + length + min_space, e))
+                return new_parts
+
+            parts = [(0, sz)]
+            min_length = min(lengths)
+            for length in sorted(lengths, reverse=True):
+                lens = np.fromiter(
+                    (e - s if e - s >= length + min_space else 0 for s, e in parts),
+                    np.int,
+                )
+                l_sum = np.sum(lens)
+                if l_sum == 0:
+                    break
+                probs = lens / np.sum(lens)
+                c = np.random.choice(len(parts), p=probs)
+                s, e = parts.pop(c)
+                parts.extend(arrange(s, e, length, min_length))
+            mask_idc = np.asarray(mask_idc)
+        else:
+            min_len = min(lengths)
+            if sz - min_len <= num_mask:
+                min_len = sz - num_mask - 1
+
+            mask_idc = np.random.choice(sz - min_len, num_mask, replace=False)
+
+            mask_idc = np.asarray(
+                [
+                    mask_idc[j] + offset
+                    for j in range(len(mask_idc))
+                    for offset in range(lengths[j])
+                ]
+            )
+
+        mask_idcs.append(np.unique(mask_idc[mask_idc < sz]))
+
+    min_len = min([len(m) for m in mask_idcs])
+    for i, mask_idc in enumerate(mask_idcs):
+        if len(mask_idc) > min_len:
+            mask_idc = np.random.choice(mask_idc, min_len, replace=False)
+        mask[i, mask_idc] = True
+
+    return mask
+
+
+def get_mem_usage():
+    try:
+        import psutil
+
+        mb = 1024 * 1024
+        return f"used={psutil.virtual_memory().used / mb}Mb; avail={psutil.virtual_memory().available / mb}Mb"
+    except ImportError:
+        return "N/A"
+
+
+# lens: torch.LongTensor
+# returns: torch.BoolTensor
+def lengths_to_padding_mask(lens):
+    bsz, max_lens = lens.size(0), torch.max(lens).item()
+    mask = torch.arange(max_lens).to(lens.device).view(1, max_lens)
+    mask = mask.expand(bsz, -1) >= lens.view(bsz, 1).expand(-1, max_lens)
+    return mask
+
+
+# lens: torch.LongTensor
+# returns: torch.BoolTensor
+def lengths_to_mask(lens):
+    return ~lengths_to_padding_mask(lens)
+
+
+def get_buckets(sizes, num_buckets):
+    buckets = np.unique(
+        np.percentile(
+            sizes,
+            np.linspace(0, 100, num_buckets + 1),
+            interpolation='lower',
+        )[1:]
+    )
+    return buckets
+
+
+def get_bucketed_sizes(orig_sizes, buckets):
+    sizes = np.copy(orig_sizes)
+    assert np.min(sizes) >= 0
+    start_val = -1
+    for end_val in buckets:
+        mask = (sizes > start_val) & (sizes <= end_val)
+        sizes[mask] = end_val
+        start_val = end_val
+    return sizes
+
+
+
+def _find_extra_valid_paths(dataset_path: str) -> set:
+    paths = utils.split_paths(dataset_path)
+    all_valid_paths = set()
+    for sub_dir in paths:
+        contents = PathManager.ls(sub_dir)
+        valid_paths = [c for c in contents if re.match("valid*[0-9].*", c) is not None]
+        all_valid_paths |= {os.path.basename(p) for p in valid_paths}
+    # Remove .bin, .idx etc
+    roots = {os.path.splitext(p)[0] for p in all_valid_paths}
+    return roots
+
+
+def raise_if_valid_subsets_unintentionally_ignored(train_cfg) -> None:
+    """Raises if there are paths matching 'valid*[0-9].*' which are not combined or ignored."""
+    if (
+        train_cfg.dataset.ignore_unused_valid_subsets
+        or train_cfg.dataset.combine_valid_subsets
+        or train_cfg.dataset.disable_validation
+        or not hasattr(train_cfg.task, "data")
+    ):
+        return
+    other_paths = _find_extra_valid_paths(train_cfg.task.data)
+    specified_subsets = train_cfg.dataset.valid_subset.split(",")
+    ignored_paths = [p for p in other_paths if p not in specified_subsets]
+    if ignored_paths:
+        advice = "Set --combine-val to combine them or --ignore-unused-valid-subsets to ignore them."
+        msg = f"Valid paths {ignored_paths} will be ignored. {advice}"
+        raise ValueError(msg)

data/file_dataset.py

@@ -0,0 +1,112 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+import torch
+import pickle
+
+
+class FileDataset:
+    def __init__(self, file_path, selected_col_ids=None, dtypes=None, separator="\t", cached_index=False):
+        self.file_path = file_path
+        assert os.path.exists(self.file_path), "Error: The local datafile {} not exists!".format(self.file_path)
+
+        self.separator = separator
+        if selected_col_ids is None:
+            # default to all fields
+            self.selected_col_ids = list(
+                range(len(open(self.file_path).readline().rstrip("\n").split(self.separator))))
+        else:
+            self.selected_col_ids = [int(col_id) for col_id in selected_col_ids.split(",")]
+        if dtypes is None:
+            # default to str
+            self.dtypes = [str for col_id in self.selected_col_ids]
+        else:
+            self.dtypes = [eval(col_dtype) for col_dtype in dtypes.split(",")]
+            assert len(self.dtypes) == len(self.selected_col_ids)
+
+        self.data_cnt = 0
+        try:
+            self.slice_id = torch.distributed.get_rank()
+            self.slice_count = torch.distributed.get_world_size()
+        except Exception:
+            self.slice_id = 0
+            self.slice_count = 1
+        self.cached_index = cached_index
+        self._init_seek_index()
+        self._reader = self._get_reader()
+        print("file {} slice_id {} row count {} total row count {}".format(
+            self.file_path, self.slice_id, self.row_count, self.total_row_count)
+        )
+
+    def _init_seek_index(self):
+        if self.cached_index:
+            cache_path = "{}.index".format(self.file_path)
+            assert os.path.exists(cache_path), "cache file {} not exists!".format(cache_path)
+            self.total_row_count, self.lineid_to_offset = pickle.load(open(cache_path, "rb"))
+            print("local datafile {} slice_id {} use cached row_count and line_idx-to-offset mapping".format(
+                self.file_path, self.slice_id))
+        else:
+            # make an iteration over the file to get row_count and line_idx-to-offset mapping
+            fp = open(self.file_path, "r")
+            print("local datafile {} slice_id {} begin to initialize row_count and line_idx-to-offset mapping".format(
+                self.file_path, self.slice_id))
+            self.total_row_count = 0
+            offset = 0
+            self.lineid_to_offset = []
+            for line in fp:
+                self.lineid_to_offset.append(offset)
+                self.total_row_count += 1
+                offset += len(line.encode('utf-8'))
+        self._compute_start_pos_and_row_count()
+        print("local datafile {} slice_id {} finished initializing row_count and line_idx-to-offset mapping".format(
+            self.file_path, self.slice_id))
+
+    def _compute_start_pos_and_row_count(self):
+        self.row_count = self.total_row_count // self.slice_count
+        if self.slice_id < self.total_row_count - self.row_count * self.slice_count:
+            self.row_count += 1
+            self.start_pos = self.row_count * self.slice_id
+        else:
+            self.start_pos = self.row_count * self.slice_id + (self.total_row_count - self.row_count * self.slice_count)
+
+    def _get_reader(self):
+        fp = open(self.file_path, "r")
+        fp.seek(self.lineid_to_offset[self.start_pos])
+        return fp
+
+    def _seek(self, offset=0):
+        try:
+            print("slice_id {} seek offset {}".format(self.slice_id, self.start_pos + offset))
+            self._reader.seek(self.lineid_to_offset[self.start_pos + offset])
+            self.data_cnt = offset
+        except Exception:
+            print("slice_id {} seek offset {}".format(self.slice_id, offset))
+            self._reader.seek(self.lineid_to_offset[offset])
+            self.data_cnt = offset
+
+    def __del__(self):
+        self._reader.close()
+
+    def __len__(self):
+        return self.row_count
+
+    def get_total_row_count(self):
+        return self.total_row_count
+
+    def __getitem__(self, index):
+        if self.data_cnt == self.row_count:
+            print("reach the end of datafile, start a new reader")
+            self.data_cnt = 0
+            self._reader = self._get_reader()
+        column_l = self._reader.readline().rstrip("\n").split(self.separator)
+        self.data_cnt += 1
+        column_l = [dtype(column_l[col_id]) for col_id, dtype in zip(self.selected_col_ids, self.dtypes)]
+        return column_l

data/poly_utils.py

@@ -0,0 +1,294 @@
+import re
+import numpy as np
+from itertools import groupby
+from PIL import Image
+import math
+from math import ceil, floor
+from skimage import draw
+from random import sample
+import base64
+from io import BytesIO
+
+convert = lambda text: int(text) if text.isdigit() else text.lower()
+natrual_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)]
+
+
+def points_to_token_string(box, polygons):
+    polygon_strings = []
+    for polygon in polygons:
+        polygon_string = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in polygon])
+        polygon_strings.append(polygon_string)
+    polygon_string = " <separator> ".join(polygon_strings)
+    box_string = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in box])
+    token_string = " ".join([box_string, polygon_string])
+
+    token_type = []
+    for token in token_string.split(" "):
+        if "bin" in token:
+            token_type.append(0)  # 0 for coordinate tokens
+        else:
+            token_type.append(1)  # 1 for separator tokens
+    return token_string, token_type
+
+
+def resize_binary_mask(array, new_size):
+    image = Image.fromarray(array.astype(np.uint8) * 255)
+    image = image.resize(new_size)
+    return np.asarray(image).astype(np.bool_)
+
+
+def close_contour(contour):
+    if not np.array_equal(contour[0], contour[-1]):
+        contour = np.vstack((contour, contour[0]))
+    return contour
+
+
+def binary_mask_to_rle(binary_mask):
+    rle = {'counts': [], 'size': list(binary_mask.shape)}
+    counts = rle.get('counts')
+    for i, (value, elements) in enumerate(groupby(binary_mask.ravel(order='F'))):
+        if i == 0 and value == 1:
+            counts.append(0)
+        counts.append(len(list(elements)))
+
+    return rle
+
+
+def revert_direction(poly):
+    poly = np.array(poly).reshape(int(len(poly) / 2), 2)
+    poly = poly[::-1, :]
+    return list(poly.flatten())
+
+
+def reorder_points(poly):
+    poly = np.array(poly)
+    xs = poly[::2]
+    ys = poly[1::2]
+    points = np.array(poly).reshape(int(len(poly) / 2), 2)
+    start = np.argmin(xs ** 2 + ys ** 2)  # smallest distance to the origin
+    poly_reordered = np.concatenate([points[start:], points[:start]], 0)
+    return list(poly_reordered.flatten())
+
+
+def convert_pts(coeffs):
+    pts = []
+    for i in range(len(coeffs) // 2):
+        pts.append([coeffs[2 * i + 1], coeffs[2 * i]])  # y, x
+    return np.array(pts, np.int32)
+
+
+def get_mask_from_codes(codes, img_size):
+    masks = [np.zeros(img_size)]
+    for code in codes:
+        if len(code) > 0:
+            mask = draw.polygon2mask(img_size, convert_pts(code))
+            mask = np.array(mask, np.uint8)
+            masks.append(mask)
+    mask = sum(masks)
+    mask = mask > 0
+    return mask.astype(np.uint8)
+
+
+def is_clockwise(poly):
+    n = len(poly) // 2
+    xs = poly[::2]
+    xs.append(xs[0])
+    ys = poly[1::2]
+    ys.append(ys[0])
+    area = 0
+    for i in range(n):
+        x1, y1 = xs[i], ys[i]
+        x2, y2 = xs[i + 1], ys[i + 1]
+        area += (x2 - x1) * (y2 + y1)
+    return area < 0
+
+
+def close_polygon_contour(poly):
+    poly = np.array(poly).reshape(int(len(poly) / 2), 2)
+    x1, y1 = poly[0]
+    x2, y2 = poly[-1]
+    if x1 != x2:
+        poly = np.concatenate([poly, [poly[0]]], 0)
+    return list(poly.flatten())
+
+
+def close_polygons_contour(polygons):
+    polygons_closed = []
+    for polygon in polygons:
+        polygon_closed = close_polygon_contour(polygon)
+        polygons_closed.append(polygon_closed)
+    return polygons_closed
+
+
+def image_to_base64(img, format):
+    output_buffer = BytesIO()
+    img.save(output_buffer, format=format)
+    byte_data = output_buffer.getvalue()
+    base64_str = base64.b64encode(byte_data)
+    base64_str = str(base64_str, encoding='utf-8')
+    return base64_str
+
+
+def process_polygons(polygons, redirection=True, reorder=True, close=False):
+    polygons_processed = []
+    for polygon in polygons:
+        if redirection and not is_clockwise(polygon):
+            polygon = revert_direction(polygon)
+        if reorder:
+            polygon = reorder_points(polygon)
+        if close:
+            polygon = close_polygon_contour(polygon)
+        polygons_processed.append(polygon)
+    polygons = sorted(polygons_processed, key=lambda x: (x[0] ** 2 + x[1] ** 2, x[0], x[1]))
+    return polygons
+
+
+def string_to_polygons(pts_strings):
+    pts_strings = pts_strings.split(" ")[:-1]
+    polygons = []
+    for pts_string in pts_strings:
+        polygon = pts_string.split(",")
+        polygon = [float(p) for p in polygon]
+        polygons.append(polygon)
+    return polygons
+
+
+def downsample_polygon(polygon, ds_rate=25):
+    points = np.array(polygon).reshape(int(len(polygon) / 2), 2)
+    points = points[::ds_rate]
+    return list(points.flatten())
+
+
+def downsample_polygons(polygons, ds_rate=25):
+    polygons_ds = []
+    for polygon in polygons:
+        polygons_ds.append(downsample_polygon(polygon, ds_rate))
+    return polygons_ds
+
+
+def check_length(polygons):
+    length = 0
+    for polygon in polygons:
+        length += len(polygon)
+    return length
+
+
+def approximate_polygon(poly, tolerance=2):
+    poly = np.array(poly).reshape(int(len(poly) / 2), 2)
+    new_poly = [poly[0]]
+    for i in range(1, len(poly)):
+        x1, y1 = new_poly[-1]
+        x2, y2 = poly[i]
+        dist = math.sqrt((x1 - x2) ** 2 + (y1 - y2) ** 2)
+        if dist > tolerance:
+            new_poly.append(poly[i])
+    new_poly = np.array(new_poly)
+    return list(new_poly.flatten())
+
+
+def approximate_polygons(polys, tolerance=1.0, max_length=400):
+    tol = tolerance
+    while check_length(polys) > max_length:
+        polys_new = []
+        for poly in polys:
+            polys_new.append(approximate_polygon(poly, tolerance=tol))
+        polys = polys_new
+        tol += 2.0
+    return polys
+
+
+def random_int(low, high):
+    if low < high:
+        return np.random.randint(low, high)
+    else:
+        return max(low, high)
+
+
+def interpolate_points(ps, pe):
+    xs, ys = ps
+    xe, ye = pe
+    points = []
+    dx = xe - xs
+    dy = ye - ys
+    if dx != 0:
+        scale = dy / dx
+        if xe > xs:
+            x_interpolated = list(range(ceil(xs), floor(xe) + 1))
+        else:
+            x_interpolated = list(range(floor(xs), ceil(xe) - 1, -1))
+        for x in x_interpolated:
+            y = ys + (x - xs) * scale
+            points.append([x, y])
+    if dy != 0:
+        scale = dx / dy
+        if ye > ys:
+            y_interpolated = list(range(ceil(ys), floor(ye) + 1))
+        else:
+            y_interpolated = list(range(floor(ys), ceil(ye) - 1, -1))
+        for y in y_interpolated:
+            x = xs + (y - ys) * scale
+            points.append([x, y])
+    if xe > xs:
+        points = sorted(points, key=lambda x: x[0])
+    else:
+        points = sorted(points, key=lambda x: -x[0])
+    return points
+
+
+def interpolate_polygon(polygon):
+    points = np.array(polygon).reshape(int(len(polygon) / 2), 2)
+    points_interpolated = []
+    points_interpolated.append(points[0])
+    for i in range(0, len(points) - 1):
+        points_i = interpolate_points(points[i], points[i + 1])
+        points_interpolated += points_i
+        points_interpolated.append(points[i + 1])
+    points_interpolated = prune_points(points_interpolated)
+    polygon_interpolated = np.array(points_interpolated)
+    return list(polygon_interpolated.flatten())
+
+
+def prune_points(points, th=0.1):
+    points_pruned = [points[0]]
+    for i in range(1, len(points)):
+        x1, y1 = points_pruned[-1]
+        x2, y2 = points[i]
+        dist = (x2 - x1) ** 2 + (y2 - y1) ** 2
+        if dist > th:
+            points_pruned.append(points[i])
+    return points_pruned
+
+
+def interpolate_polygons(polygons):
+    polygons_i = []
+    for polygon in polygons:
+        polygons_i.append(interpolate_polygon(polygon))
+    return polygons_i
+
+
+def sample_polygon(polygon, sample_rate=0.5):
+    points = np.array(polygon).reshape(int(len(polygon) / 2), 2)
+    k = int(len(points) * sample_rate)
+    index = sorted(sample(list(range(len(points))), k))
+    points_sampled = points[index]
+    return list(np.array(points_sampled).flatten())
+
+
+def sample_polygons(polygons, max_length=400.0):
+    n = check_length(polygons)
+    k = max_length / n
+    polygons_s = []
+    for polygon in polygons:
+        polygons_s.append(sample_polygon(polygon, k))
+    return polygons_s
+
+
+def polygons_to_string(polygons):
+    pts_strings = []
+    for polygon in polygons:
+        pts_string = ','.join([str(num) for num in polygon])
+        pts_string += " "  # separator
+        pts_strings.append(pts_string)
+    pts_strings = "".join(pts_strings)
+    return pts_strings
+

data/refcoco_dataset.py

@@ -0,0 +1,294 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from io import BytesIO
+
+import logging
+import warnings
+
+import numpy as np
+import torch
+import base64
+import utils.transforms as T
+import math
+from PIL import Image, ImageFile
+
+from data import data_utils
+from data.base_dataset import BaseDataset
+from bert.tokenization_bert import BertTokenizer
+from data.poly_utils import string_to_polygons, downsample_polygons, polygons_to_string, points_to_token_string
+import cv2
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+ImageFile.MAX_IMAGE_PIXELS = None
+Image.MAX_IMAGE_PIXELS = None
+
+logger = logging.getLogger(__name__)
+warnings.filterwarnings("ignore", "(Possibly )?corrupt EXIF data", UserWarning)
+
+IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
+
+
+class RefcocoDataset(BaseDataset):
+    def __init__(
+        self,
+        split,
+        dataset,
+        bpe,
+        src_dict,
+        tgt_dict=None,
+        max_src_length=80,
+        max_tgt_length=30,
+        patch_image_size=512,
+        imagenet_default_mean_and_std=False,
+        num_bins=1000,
+        max_image_size=512
+    ):
+        super().__init__(split, dataset, bpe, src_dict, tgt_dict)
+        self.max_src_length = max_src_length
+        self.max_tgt_length = max_tgt_length
+        self.patch_image_size = patch_image_size
+        self.num_bins = num_bins
+
+        if imagenet_default_mean_and_std:
+            mean = IMAGENET_DEFAULT_MEAN
+            std = IMAGENET_DEFAULT_STD
+        else:
+            mean = [0.5, 0.5, 0.5]
+            std = [0.5, 0.5, 0.5]
+
+        # for positioning
+        self.positioning_transform = T.Compose([
+            T.RandomResize([patch_image_size], max_size=patch_image_size),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std, max_image_size=max_image_size)
+        ])
+        self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+    def __getitem__(self, index):
+        data = self.dataset[index]
+        if len(data) == 7:
+            uniq_id, base64_str, seg64_str, text, poly_original, region_coord, poly_interpolated = data
+            train = True
+        else:
+            uniq_id, base64_str, seg64_str, text, poly, region_coord = data
+            train = False
+
+        # load image and segmentation labels
+        image = Image.open(BytesIO(base64.urlsafe_b64decode(base64_str))).convert("RGB")
+        label = Image.open(BytesIO(base64.urlsafe_b64decode(seg64_str)))
+        label = np.asarray(label)
+        label = cv2.resize(label, [self.patch_image_size, self.patch_image_size], interpolation=cv2.INTER_NEAREST)
+
+        w, h = image.size
+        patch_image = self.positioning_transform(image, target=None)
+        resize_h = self.patch_image_size
+        resize_w = self.patch_image_size
+        patch_mask = torch.tensor([True])
+
+        if train:
+            prob = np.random.uniform()
+            if prob < 0.5:
+                polygons_interpolated = string_to_polygons(poly_interpolated)
+                ds_rate = np.random.randint(25, 41)
+                polygons_augmented = downsample_polygons(polygons_interpolated, ds_rate)
+                poly = polygons_to_string(polygons_augmented)
+            else:
+                poly = poly_original
+
+        polygons = string_to_polygons(poly)
+        polygons_scaled = []
+        for polygon in polygons:
+            n_point = len(polygon) // 2
+            scale = np.concatenate([np.array([w, h]) for _ in range(n_point)], 0)
+            polygon = polygon / scale
+            polygon = polygon.reshape(n_point, 2)
+            polygons_scaled.append(polygon)
+
+        x0, y0, x1, y1 = region_coord.strip().split(',')
+        region_points = [float(x0), float(y0), float(x1), float(y1)]
+        region = np.array(region_points)
+
+        region_points = region_points / np.array([w, h, w, h])  # scaled to [0,1]
+        region_points = torch.tensor(region_points.reshape(2, 2))
+
+        quant_box = region_points * (self.num_bins - 1)
+        quant_box11 = [[math.floor(p[0]), math.floor(p[1])] for p in quant_box]
+        quant_box21 = [[math.ceil(p[0]), math.floor(p[1])] for p in quant_box]
+        quant_box12 = [[math.floor(p[0]), math.ceil(p[1])] for p in quant_box]
+        quant_box22 = [[math.ceil(p[0]), math.ceil(p[1])] for p in quant_box]
+
+        quant_poly = [poly * (self.num_bins - 1) for poly in polygons_scaled]
+        quant_poly11 = [[[math.floor(p[0]), math.floor(p[1])] for p in poly] for poly in quant_poly]
+        quant_poly21 = [[[math.ceil(p[0]), math.floor(p[1])] for p in poly] for poly in quant_poly]
+        quant_poly12 = [[[math.floor(p[0]), math.ceil(p[1])] for p in poly] for poly in quant_poly]
+        quant_poly22 = [[[math.ceil(p[0]), math.ceil(p[1])] for p in poly] for poly in quant_poly]
+
+        region_coord11, _ = points_to_token_string(quant_box11, quant_poly11)
+        region_coord21, _ = points_to_token_string(quant_box21, quant_poly21)
+        region_coord12, _ = points_to_token_string(quant_box12, quant_poly12)
+        region_coord22, token_type = points_to_token_string(quant_box22, quant_poly22)
+
+        # compute bilinear interpolation coefficient
+        delta_x1 = [0] + [p[0] - math.floor(p[0]) for p in quant_box]  # [0] for bos token
+        for polygon in quant_poly:
+            delta = [poly_point[0] - math.floor(poly_point[0]) for poly_point in polygon]
+            delta_x1.extend(delta)
+            delta_x1.extend([0])  # for separator token
+        delta_x1 = delta_x1[:-1]  # there is no separator token in the end
+        delta_x1 = torch.tensor(delta_x1)
+        delta_x2 = 1 - delta_x1
+
+        delta_y1 = [0] + [p[1] - math.floor(p[1]) for p in quant_box]  # [0] for bos token
+        for polygon in quant_poly:
+            delta = [poly_point[1] - math.floor(poly_point[1]) for poly_point in polygon]
+            delta_y1.extend(delta)
+            delta_y1.extend([0])  # for separator token
+        delta_y1 = delta_y1[:-1]  # there is no separator token in the end
+        delta_y1 = torch.tensor(delta_y1)
+        delta_y2 = 1 - delta_y1
+
+        token_type.append(2)  # 2 for eos token
+
+        src_caption = self.pre_caption(text, self.max_src_length)
+
+        prompt = ' which region does the text " {} " describe?'.format(src_caption)
+
+        # tgt for input
+        tgt_item11 = self.encode_text(region_coord11, use_bpe=False)
+        tgt_item12 = self.encode_text(region_coord12, use_bpe=False)
+        tgt_item21 = self.encode_text(region_coord21, use_bpe=False)
+        tgt_item22 = self.encode_text(region_coord22, use_bpe=False)
+
+        # tgt for output
+        target_item = region_points
+        for poly in polygons_scaled:
+            target_item = torch.cat([target_item, torch.tensor(poly), torch.tensor([[0, 0]])], dim=0)  # [0, 0] is padding token for separator and eos
+
+        #target_item = torch.cat([tgt_item, self.eos_item])
+        prev_output_item11 = torch.cat([self.bos_item, tgt_item11])
+        prev_output_item12 = torch.cat([self.bos_item, tgt_item12])
+        prev_output_item21 = torch.cat([self.bos_item, tgt_item21])
+        prev_output_item22 = torch.cat([self.bos_item, tgt_item22])
+        example = {
+            "id": uniq_id,
+            "source": prompt,
+            "patch_image": patch_image,
+            "patch_mask": patch_mask,
+            "target": target_item,
+            "prev_output_tokens_11": prev_output_item11,
+            "prev_output_tokens_12": prev_output_item12,
+            "prev_output_tokens_21": prev_output_item21,
+            "prev_output_tokens_22": prev_output_item22,
+            "delta_x1": delta_x1,
+            "delta_y1": delta_y1,
+            "delta_x2": delta_x2,
+            "delta_y2": delta_y2,
+            "w_resize_ratio": torch.tensor(resize_w / w),
+            "h_resize_ratio": torch.tensor(resize_h / h),
+            "region_coord": torch.tensor(region),
+            "token_type": torch.tensor(token_type),
+            "w": torch.tensor(w),
+            "h": torch.tensor(h),
+            "label": label,
+            "n_poly": len(polygons),
+            "text": src_caption
+        }
+        return example
+
+    def collate(self, samples, pad_idx, eos_idx):
+        if len(samples) == 0:
+            return {}
+
+        def merge(key, padding_item):
+            return data_utils.collate_tokens(
+                [s[key] for s in samples],
+                padding_item,
+                eos_idx=eos_idx,
+            )
+
+        id = np.array([s["id"] for s in samples])
+        captions = [s["source"] for s in samples]
+        tokenized = self.tokenizer.batch_encode_plus(captions, padding="longest", return_tensors="pt")
+        src_tokens = tokenized["input_ids"]
+        att_masks = tokenized["attention_mask"]
+        src_lengths = torch.LongTensor(att_masks.ne(0).long().sum())
+
+        patch_images = torch.stack([sample['patch_image'] for sample in samples], dim=0)
+        patch_masks = torch.cat([sample['patch_mask'] for sample in samples])
+
+        w_resize_ratios = torch.stack([s["w_resize_ratio"] for s in samples], dim=0)
+        h_resize_ratios = torch.stack([s["h_resize_ratio"] for s in samples], dim=0)
+
+        delta_x1 = merge("delta_x1", 0)
+        delta_y1 = merge("delta_y1", 0)
+        delta_x2 = merge("delta_x2", 1)
+        delta_y2 = merge("delta_y2", 1)
+
+        region_coords = torch.stack([s['region_coord'] for s in samples], dim=0)
+
+        target = merge("target", pad_idx)
+        tgt_lengths = torch.LongTensor([s["target"].shape[0] for s in samples])
+        ntokens = tgt_lengths.sum().item()
+
+        prev_output_tokens_11 = merge("prev_output_tokens_11", pad_idx)
+        prev_output_tokens_12 = merge("prev_output_tokens_12", pad_idx)
+        prev_output_tokens_21 = merge("prev_output_tokens_21", pad_idx)
+        prev_output_tokens_22 = merge("prev_output_tokens_22", pad_idx)
+
+        token_type = merge("token_type", -1)
+        w = torch.stack([s["w"] for s in samples], dim=0)
+        h = torch.stack([s["h"] for s in samples], dim=0)
+        n_poly = [s['n_poly'] for s in samples]
+
+        labels = np.stack([sample['label'] for sample in samples], 0)
+        text = [s["text"] for s in samples]
+        batch = {
+            "id": id,
+            "nsentences": len(samples),
+            "ntokens": ntokens,
+            "net_input": {
+                "src_tokens": src_tokens,
+                "src_lengths": src_lengths,
+                "att_masks": att_masks,
+                "patch_images": patch_images,
+                "patch_masks": patch_masks,
+                "prev_output_tokens_11": prev_output_tokens_11,
+                "prev_output_tokens_12": prev_output_tokens_12,
+                "prev_output_tokens_21": prev_output_tokens_21,
+                "prev_output_tokens_22": prev_output_tokens_22,
+                "delta_x1": delta_x1,
+                "delta_y1": delta_y1,
+                "delta_x2": delta_x2,
+                "delta_y2": delta_y2
+            },
+            "target": target,
+            "w_resize_ratios": w_resize_ratios,
+            "h_resize_ratios": h_resize_ratios,
+            "region_coords": region_coords,
+            "label": labels,
+            "token_type": token_type,
+            "w": w,
+            "h": h,
+            "n_poly": n_poly,
+            "text": text
+        }
+
+        return batch
+
+    def collater(self, samples, pad_to_length=None):
+        """Merge a list of samples to form a mini-batch.
+        Args:
+            samples (List[dict]): samples to collate
+        Returns:
+            dict: a mini-batch containing the data of the task
+        """
+        return self.collate(samples, pad_idx=self.pad, eos_idx=self.eos)

data/refcoco_pretrain_dataset.py

@@ -0,0 +1,232 @@
+# ------------------------------------------------------------------------
+# Modified from OFA (https://github.com/OFA-Sys/OFA)
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+# ------------------------------------------------------------------------
+# Modifications Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+from io import BytesIO
+
+import logging
+import warnings
+
+import numpy as np
+import torch
+import base64
+import utils.transforms as T
+import math
+import os
+from PIL import Image, ImageFile
+
+from data import data_utils
+from data.base_dataset import BaseDataset
+from bert.tokenization_bert import BertTokenizer
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+ImageFile.MAX_IMAGE_PIXELS = None
+Image.MAX_IMAGE_PIXELS = None
+
+logger = logging.getLogger(__name__)
+warnings.filterwarnings("ignore", "(Possibly )?corrupt EXIF data", UserWarning)
+
+IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
+
+
+class RefcocoPretrainDataset(BaseDataset):
+    def __init__(
+        self,
+        split,
+        dataset,
+        bpe,
+        src_dict,
+        tgt_dict=None,
+        max_src_length=80,
+        max_tgt_length=30,
+        patch_image_size=512,
+        imagenet_default_mean_and_std=False,
+        num_bins=1000,
+        max_image_size=512,
+        image_path="../../datasets/images"
+    ):
+        super().__init__(split, dataset, bpe, src_dict, tgt_dict)
+        self.max_src_length = max_src_length
+        self.max_tgt_length = max_tgt_length
+        self.patch_image_size = patch_image_size
+        self.num_bins = num_bins
+        self.image_path = image_path
+
+        if imagenet_default_mean_and_std:
+            mean = IMAGENET_DEFAULT_MEAN
+            std = IMAGENET_DEFAULT_STD
+        else:
+            mean = [0.5, 0.5, 0.5]
+            std = [0.5, 0.5, 0.5]
+
+        # for positioning
+        self.positioning_transform = T.Compose([
+            T.RandomResize([patch_image_size], max_size=patch_image_size),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std, max_image_size=max_image_size)
+        ])
+        self.tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+    def __getitem__(self, index):
+        uniq_id, img_file, text, region_coord = self.dataset[index]
+
+        img_path = os.path.join(self.image_path, img_file)
+        image = Image.open(img_path).convert("RGB")
+
+        w, h = image.size
+        boxes_target = {"boxes": [], "labels": [], "area": [], "size": torch.tensor([h, w])}
+        x0, y0, x1, y1 = region_coord.strip().split(',')
+        region = torch.tensor([float(x0), float(y0), float(x1), float(y1)])
+        boxes_target["boxes"] = torch.tensor([[float(x0), float(y0), float(x1), float(y1)]])
+        boxes_target["labels"] = np.array([0])
+        boxes_target["area"] = torch.tensor([(float(x1) - float(x0)) * (float(y1) - float(y0))])
+
+        patch_image, patch_boxes = self.positioning_transform(image, boxes_target)
+        resize_h, resize_w = patch_boxes["size"][0], patch_boxes["size"][1]
+        patch_mask = torch.tensor([True])
+
+        quant_box = [patch_boxes["boxes"][0][i] * (self.num_bins - 1) for i in range(4)]
+        quant_box = np.array(quant_box).reshape(2, 2)
+
+        quant_box11 = [[math.floor(p[0]), math.floor(p[1])] for p in quant_box]
+        quant_box21 = [[math.ceil(p[0]), math.floor(p[1])] for p in quant_box]
+        quant_box12 = [[math.floor(p[0]), math.ceil(p[1])] for p in quant_box]
+        quant_box22 = [[math.ceil(p[0]), math.ceil(p[1])] for p in quant_box]
+
+
+        # compute linear interpolation coefficient (0 for bos token)
+        delta_x1 = torch.tensor([0] + [p[0] - math.floor(p[0]) for p in quant_box])
+        delta_y1 = torch.tensor([0] + [p[1] - math.floor(p[1]) for p in quant_box])
+        delta_x2 = 1 - delta_x1
+        delta_y2 = 1 - delta_y1
+
+        region_coord11 = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in quant_box11])
+        region_coord21 = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in quant_box21])
+        region_coord12 = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in quant_box12])
+        region_coord22 = " ".join([f"<bin_{int(p[0])}_{int(p[1])}>" for p in quant_box22])
+
+        src_caption = self.pre_caption(text, self.max_src_length)
+
+        prompt = ' which region does the text " {} " describe?'.format(src_caption)
+
+        # tgt for input
+        tgt_item11 = self.encode_text(region_coord11, use_bpe=False)
+        tgt_item12 = self.encode_text(region_coord12, use_bpe=False)
+        tgt_item21 = self.encode_text(region_coord21, use_bpe=False)
+        tgt_item22 = self.encode_text(region_coord22, use_bpe=False)
+
+        # tgt for output
+        tgt_box = torch.reshape(patch_boxes["boxes"][0], (2, 2))
+        target_item = torch.cat([tgt_box, torch.tensor([[1, 1]])], dim=0)  # [1, 1] is padding token for eos
+
+        #target_item = torch.cat([tgt_item, self.eos_item])
+        prev_output_item11 = torch.cat([self.bos_item, tgt_item11])
+        prev_output_item12 = torch.cat([self.bos_item, tgt_item12])
+        prev_output_item21 = torch.cat([self.bos_item, tgt_item21])
+        prev_output_item22 = torch.cat([self.bos_item, tgt_item22])
+        example = {
+            "id": uniq_id,
+            "source": prompt,
+            "patch_image": patch_image,
+            "patch_mask": patch_mask,
+            "target": target_item,
+            "prev_output_tokens_11": prev_output_item11,
+            "prev_output_tokens_12": prev_output_item12,
+            "prev_output_tokens_21": prev_output_item21,
+            "prev_output_tokens_22": prev_output_item22,
+            "delta_x1": delta_x1,
+            "delta_y1": delta_y1,
+            "delta_x2": delta_x2,
+            "delta_y2": delta_y2,
+            "w_resize_ratio": resize_w / w,
+            "h_resize_ratio": resize_h / h,
+            "region_coord": region,
+            "token_type": torch.tensor([0, 0, 2])
+        }
+        return example
+
+    def collate(self, samples, pad_idx, eos_idx):
+        if len(samples) == 0:
+            return {}
+
+        def merge(key):
+            return data_utils.collate_tokens(
+                [s[key] for s in samples],
+                pad_idx,
+                eos_idx=eos_idx,
+            )
+
+        id = np.array([s["id"] for s in samples])
+        captions = [s["source"] for s in samples]
+        tokenized = self.tokenizer.batch_encode_plus(captions, padding="longest", return_tensors="pt")
+        src_tokens = tokenized["input_ids"]
+        att_masks = tokenized["attention_mask"]
+        src_lengths = torch.LongTensor(att_masks.ne(0).long().sum())
+
+        patch_images = torch.stack([sample['patch_image'] for sample in samples], dim=0)
+        patch_masks = torch.cat([sample['patch_mask'] for sample in samples])
+
+        w_resize_ratios = torch.stack([s["w_resize_ratio"] for s in samples], dim=0)
+        h_resize_ratios = torch.stack([s["h_resize_ratio"] for s in samples], dim=0)
+
+        delta_x1 = torch.stack([s["delta_x1"] for s in samples], dim=0)
+        delta_y1 = torch.stack([s["delta_y1"] for s in samples], dim=0)
+        delta_x2 = torch.stack([s["delta_x2"] for s in samples], dim=0)
+        delta_y2 = torch.stack([s["delta_y2"] for s in samples], dim=0)
+
+        region_coords = torch.stack([s['region_coord'] for s in samples], dim=0)
+
+        target = merge("target")
+        tgt_lengths = torch.LongTensor([s["target"].ne(pad_idx).long().sum() for s in samples])
+        ntokens = tgt_lengths.sum().item()
+
+        prev_output_tokens_11 = merge("prev_output_tokens_11")
+        prev_output_tokens_12 = merge("prev_output_tokens_12")
+        prev_output_tokens_21 = merge("prev_output_tokens_21")
+        prev_output_tokens_22 = merge("prev_output_tokens_22")
+
+        token_type = merge("token_type")
+
+        batch = {
+            "id": id,
+            "nsentences": len(samples),
+            "ntokens": ntokens,
+            "net_input": {
+                "src_tokens": src_tokens,
+                "src_lengths": src_lengths,
+                "att_masks": att_masks,
+                "patch_images": patch_images,
+                "patch_masks": patch_masks,
+                "prev_output_tokens_11": prev_output_tokens_11,
+                "prev_output_tokens_12": prev_output_tokens_12,
+                "prev_output_tokens_21": prev_output_tokens_21,
+                "prev_output_tokens_22": prev_output_tokens_22,
+                "delta_x1": delta_x1,
+                "delta_y1": delta_y1,
+                "delta_x2": delta_x2,
+                "delta_y2": delta_y2
+            },
+            "target": target,
+            "token_type": token_type,
+            "w_resize_ratios": w_resize_ratios,
+            "h_resize_ratios": h_resize_ratios,
+            "region_coords": region_coords
+        }
+
+        return batch
+
+    def collater(self, samples, pad_to_length=None):
+        """Merge a list of samples to form a mini-batch.
+        Args:
+            samples (List[dict]): samples to collate
+        Returns:
+            dict: a mini-batch containing the data of the task
+        """
+        return self.collate(samples, pad_idx=self.pad, eos_idx=self.eos)

demo.py

@@ -0,0 +1,410 @@
+import torch
+import numpy as np
+from fairseq import utils,tasks
+from utils.checkpoint_utils import load_model_ensemble_and_task
+from utils.eval_utils import eval_step
+from tasks.refcoco import RefcocoTask
+from models.polyformer import PolyFormerModel
+from PIL import Image
+import cv2
+import math
+from skimage import draw
+
+
+tasks.register_task('refcoco', RefcocoTask)
+
+# turn on cuda if GPU is available
+use_cuda = torch.cuda.is_available()
+# use fp16 only when GPU is available
+use_fp16 = True
+
+# Load pretrained ckpt & config
+overrides={"bpe_dir":"utils/BPE"}
+models, cfg, task = load_model_ensemble_and_task(
+        utils.split_paths('weights/polyformer_l_refcocog.pt'),
+        arg_overrides=overrides
+    )
+# print(cfg)
+cfg.common.seed = 7
+cfg.generation.beam = 5
+cfg.generation.min_len = 12
+cfg.generation.max_len_a = 0
+cfg.generation.max_len_b = 420
+cfg.generation.no_repeat_ngram_size = 3
+# cfg.max_tgt_length = 256
+#cfg.num_bins = 1000
+cfg.task.patch_image_size = 512
+
+from bert.tokenization_bert import BertTokenizer
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+# Fix seed for stochastic decoding
+if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
+    np.random.seed(cfg.common.seed)
+    utils.set_torch_seed(cfg.common.seed)
+
+# model = ''
+# Move models to GPU
+for model in models:
+    model.eval()
+    if use_fp16:
+        model.half()
+    if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
+        model.cuda()
+    model.prepare_for_inference_(cfg)
+
+# Initialize generator
+generator = task.build_generator(models, cfg.generation)
+
+
+# Image transform
+from torchvision import transforms
+mean = [0.5, 0.5, 0.5]
+std = [0.5, 0.5, 0.5]
+
+patch_resize_transform = transforms.Compose([
+    lambda image: image.convert("RGB"),
+    transforms.Resize((cfg.task.patch_image_size, cfg.task.patch_image_size), interpolation=Image.BICUBIC),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=mean, std=std),
+])
+
+# Text preprocess
+bos_item = torch.LongTensor([task.src_dict.bos()])
+eos_item = torch.LongTensor([task.src_dict.eos()])
+pad_idx = task.src_dict.pad()
+
+
+# Construct input for refcoco task
+patch_image_size = cfg.task.patch_image_size
+def construct_sample(image: Image, text: str):
+    w, h = image.size
+    w_resize_ratio = torch.tensor(patch_image_size / w).unsqueeze(0)
+    h_resize_ratio = torch.tensor(patch_image_size / h).unsqueeze(0)
+    patch_image = patch_resize_transform(image).unsqueeze(0)
+    patch_mask = torch.tensor([True])
+    
+    prompt = ' which region does the text " {} " describe?'.format(text)
+    tokenized = tokenizer.batch_encode_plus([prompt], padding="longest", return_tensors="pt")
+    src_tokens = tokenized["input_ids"]
+    att_masks = tokenized["attention_mask"]
+    src_lengths = torch.LongTensor(att_masks.ne(0).long().sum())
+    
+    sample = {
+        "id":np.array(['42']),
+        "net_input": {
+            "src_tokens": src_tokens,
+            "src_lengths": src_lengths,
+            "att_masks": att_masks,
+            "patch_images": patch_image,
+            "patch_masks": patch_mask,
+        },
+        "w_resize_ratios": w_resize_ratio,
+        "h_resize_ratios": h_resize_ratio,
+        "region_coords": torch.randn(1, 4),
+        "label": np.zeros((512,512)),
+        "poly": 'None',
+        "text": text
+    }
+    return sample
+
+# Function to turn FP32 to FP16
+def apply_half(t):
+    if t.dtype is torch.float32:
+        return t.to(dtype=torch.half)
+    return t
+
+
+from io import BytesIO
+import base64
+import re
+
+def pre_caption(caption):
+    caption = caption.lower().lstrip(",.!?*#:;~").replace('-', ' ').replace('/', ' ').replace('<person>', 'person')
+
+    caption = re.sub(
+        r"\s{2,}",
+        ' ',
+        caption,
+    )
+    caption = caption.rstrip('\n')
+    caption = caption.strip(' ')
+    return caption
+
+
+def convert_pts(coeffs):
+    pts = []
+    for i in range(len(coeffs) // 2):
+        pts.append([coeffs[2 * i + 1], coeffs[2 * i]])  # y, x
+    return np.array(pts, np.int32)
+
+def get_mask_from_codes(codes, img_size):
+    masks = [np.zeros(img_size)]
+    for code in codes:
+        mask = draw.polygon2mask(img_size, convert_pts(code))
+        mask = np.array(mask, np.uint8)
+        masks.append(mask)
+    mask = sum(masks)
+    mask = mask > 0
+    return mask.astype(np.uint8)
+
+
+def overlay_predictions(img, mask=None, polygons=None, bbox=None, color_box=(0, 255, 0), color_mask=[255, 102, 102], color_poly=[255, 0, 0], thickness=3, radius=6):
+    overlayed = img.copy()
+    if bbox is not None:
+        overlayed = draw_bbox(overlayed, bbox, color=color_box, thickness=thickness)
+    if mask is not None:
+        overlayed = overlay_davis(overlayed, mask, colors=[[0, 0, 0], color_mask])
+    if polygons is not None:
+        overlayed = plot_polygons(overlayed, polygons, color=color_poly, radius=radius)
+    return overlayed
+
+
+def overlay_davis(image, mask, colors=[[0, 0, 0], [255, 102, 102]], cscale=1, alpha=0.4):  # [255, 178, 102] orange [102, 178, 255] red
+    from scipy.ndimage.morphology import binary_dilation
+
+    colors = np.reshape(colors, (-1, 3))
+    colors = np.atleast_2d(colors) * cscale
+
+    im_overlay = image.copy()
+    object_ids = np.unique(mask)
+
+    h_i, w_i = image.shape[0:2]
+    h_m, w_m = mask.shape[0:2]
+    if h_i != h_m:
+        mask = cv2.resize(mask, [h_i, w_i], interpolation=cv2.INTER_NEAREST)
+    for object_id in object_ids[1:]:
+        # Overlay color on  binary mask
+        foreground = image*alpha + np.ones(image.shape)*(1-alpha) * np.array(colors[object_id])
+        binary_mask = mask == object_id
+
+        # Compose image
+        im_overlay[binary_mask] = foreground[binary_mask]
+
+    return im_overlay.astype(image.dtype)
+
+
+def draw_bbox(img, box, color=(0, 255, 0), thickness=2):
+    x1, y1, x2, y2 = box
+    return cv2.rectangle(img, (int(x1), int(y1)), (int(x2), int(y2)), color, thickness=thickness)
+
+def plot_polygons(img, polygons, color=(255, 0, 0), radius=7):
+    for polygon in polygons:
+        if len(polygon) > 0:
+            polygon = np.reshape(polygon[:len(polygon)-len(polygon)%2], (len(polygon)//2, 2)).astype(np.int16)
+            for i, point in enumerate(polygon):
+                img = cv2.circle(img, point, radius, color, thickness=-1)
+            img = cv2.circle(img, polygon[0], radius, color, thickness=-1)
+    return img
+
+def plot_arrow(img, polygons, color=(128, 128, 128), thickness=3, tip_length=0.3):
+    for polygon in polygons:
+        if len(polygon) > 0:
+            polygon = np.reshape(polygon[:len(polygon)-len(polygon)%2], (len(polygon)//2, 2)).astype(np.int16)
+            for i, point in enumerate(polygon):
+                if i > 0: 
+                    img = cv2.arrowedLine(img, polygon[i-1], point, color, thickness=thickness, tipLength=tip_length)  
+    return img
+
+def downsample_polygon(polygon, ds_rate=25):
+    points = np.array(polygon).reshape(int(len(polygon) / 2), 2)
+    points = points[::ds_rate]
+    return list(points.flatten())
+
+
+def downsample_polygons(polygons, ds_rate=25):
+    polygons_ds = []
+    for polygon in polygons:
+        polygons_ds.append(downsample_polygon(polygon, ds_rate))
+    return polygons_ds
+
+
+
+def visual_grounding(image, text):
+
+    # Construct input sample & preprocess for GPU if cuda available
+    sample = construct_sample(image, text.lower())
+    sample = utils.move_to_cuda(sample) if use_cuda else sample
+    sample = utils.apply_to_sample(apply_half, sample) if use_fp16 else sample
+
+    with torch.no_grad():
+        if isinstance(models, list):
+            model = models[0]
+        min_len = 6
+        max_len = 210
+        model.eval()
+        img = sample["net_input"]["patch_images"]
+        b = img.shape[0]
+        prev_output_token_11 = [[0] for _ in range(b)]
+        prev_output_token_12 = [[0] for _ in range(b)]
+        prev_output_token_21 = [[0] for _ in range(b)]
+        prev_output_token_22 = [[0] for _ in range(b)]
+        delta_x1 = [[0] for _ in range(b)]
+        delta_y1 = [[0] for _ in range(b)]
+        delta_x2 = [[1] for _ in range(b)]
+        delta_y2 = [[1] for _ in range(b)]
+
+        gen_out = [[] for _ in range(b)]
+
+        n_bins = 64
+
+        unfinish_flag = np.ones(b)
+        i = 0
+
+        encoder_out = model.encoder(
+            sample['net_input']['src_tokens'],
+            src_lengths=sample['net_input']['src_lengths'],
+            att_masks=sample['net_input']['att_masks'],
+            patch_images=sample['net_input']['patch_images'],
+            patch_masks=sample['net_input']['patch_masks'],
+            token_embeddings=None,
+            return_all_hiddens=False,
+            sample_patch_num=None
+        )
+        attn_masks = []
+        while i < max_len and unfinish_flag.any():
+            # print(i)
+            prev_output_tokens_11_tensor = torch.tensor(np.array(prev_output_token_11)).to(img.device).long()
+            prev_output_tokens_12_tensor = torch.tensor(np.array(prev_output_token_12)).to(img.device).long()
+            prev_output_tokens_21_tensor = torch.tensor(np.array(prev_output_token_21)).to(img.device).long()
+            prev_output_tokens_22_tensor = torch.tensor(np.array(prev_output_token_22)).to(img.device).long()
+            delta_x1_tensor = torch.tensor(np.array(delta_x1)).to(img.device)
+            delta_x2_tensor = torch.tensor(np.array(delta_x2)).to(img.device)
+            delta_y1_tensor = torch.tensor(np.array(delta_y1)).to(img.device)
+            delta_y2_tensor = torch.tensor(np.array(delta_y2)).to(img.device)
+
+            net_output = model.decoder(
+                prev_output_tokens_11_tensor,
+                prev_output_tokens_12_tensor,
+                prev_output_tokens_21_tensor,
+                prev_output_tokens_22_tensor,
+                delta_x1_tensor,
+                delta_y1_tensor,
+                delta_x2_tensor,
+                delta_y2_tensor,
+                code_masks=None,
+                encoder_out=encoder_out,
+                features_only=False,
+                alignment_layer=None,
+                alignment_heads=None,
+                src_lengths=sample['net_input']['src_lengths'],
+                return_all_hiddens=False
+            )
+
+            cls_output = net_output[0]
+            cls_type = torch.argmax(cls_output, 2)
+            reg_output = net_output[1].squeeze(-1)
+            attn = net_output[2]['attn']
+            attn_arrays = [att.detach().cpu().numpy() for att in attn]
+            attn_arrays = np.concatenate(attn_arrays, 0)
+            attn_arrays = np.mean(attn_arrays, 0)
+            attn_arrays = attn_arrays[i, :256].reshape(16, 16)
+            h, w = image.size
+            attn_mask = cv2.resize(attn_arrays.astype(np.float32), (h, w))
+            attn_masks.append(attn_mask)
+            
+            for j in range(b):
+                # print(j)
+                if unfinish_flag[j] == 1:  # prediction is not finished
+                    cls_j = cls_type[j, i].item()
+                    if cls_j == 0 or (cls_j == 2 and i < min_len):  # 0 for coordinate tokens; 2 for eos
+                        output_j_x, output_j_y = reg_output[j, i].cpu().numpy()
+                        output_j_x = min(output_j_x, 1)
+                        output_j_y = min(output_j_y, 1)
+
+                        gen_out[j].extend([output_j_x, output_j_y])
+
+                        output_j_x = output_j_x * (n_bins - 1)
+                        output_j_y = output_j_y * (n_bins - 1)
+
+                        output_j_x_floor = math.floor(output_j_x)
+                        output_j_y_floor = math.floor(output_j_y)
+                        output_j_x_ceil = math.ceil(output_j_x)
+                        output_j_y_ceil = math.ceil(output_j_y)
+
+                        # convert to token
+                        prev_output_token_11[j].append(output_j_x_floor * n_bins + output_j_y_floor + 4)
+                        prev_output_token_12[j].append(output_j_x_floor * n_bins + output_j_y_ceil + 4)
+                        prev_output_token_21[j].append(output_j_x_ceil * n_bins + output_j_y_floor + 4)
+                        prev_output_token_22[j].append(output_j_x_ceil * n_bins + output_j_y_ceil + 4)
+
+                        delta_x = output_j_x - output_j_x_floor
+                        delta_y = output_j_y - output_j_y_floor
+                    elif cls_j == 1:  # 1 for separator tokens
+                        gen_out[j].append(2)  # insert 2 indicating separator tokens
+                        prev_output_token_11[j].append(3)
+                        prev_output_token_12[j].append(3)
+                        prev_output_token_21[j].append(3)
+                        prev_output_token_22[j].append(3)
+                        delta_x = 0
+                        delta_y = 0
+                    else:  # eos is predicted and i >= min_len
+                        unfinish_flag[j] = 0
+                        gen_out[j].append(-1)
+                        prev_output_token_11[j].append(2)  # 2 is eos token
+                        prev_output_token_12[j].append(2)  # 2 is eos token
+                        prev_output_token_21[j].append(2)  # 2 is eos token
+                        prev_output_token_22[j].append(2)  # 2 is eos token
+                        delta_x = 0
+                        delta_y = 0
+                else:  # prediction is finished
+                    gen_out[j].append(-1)
+                    prev_output_token_11[j].append(1)  # 1 is padding token
+                    prev_output_token_12[j].append(1)
+                    prev_output_token_21[j].append(1)
+                    prev_output_token_22[j].append(1)
+                    delta_x = 0
+                    delta_y = 0
+                delta_x1[j].append(delta_x)
+                delta_y1[j].append(delta_y)
+                delta_x2[j].append(1 - delta_x)
+                delta_y2[j].append(1 - delta_y)
+            i += 1
+        print("inference step: ", i)
+
+    hyps = []
+    hyps_det = []
+    n_poly_pred = []
+    b = len(gen_out)
+    for i in range(b):
+        gen_out_i = np.array(gen_out[i])
+        gen_out_i = gen_out_i[gen_out_i != -1]  # excluding eos and padding indices
+
+
+        gen_out_i_det = gen_out_i[:4]
+        w, h = image.size
+        gen_out_i_det[::2] *= w
+        gen_out_i_det[1::2] *= h
+
+        polygons_pred = gen_out_i[4:]
+        polygons_pred = np.append(polygons_pred, [2])
+        size = len(polygons_pred)
+        idx_list = [idx for idx, val in
+                    enumerate(polygons_pred) if val == 2]   # 2 indicates separator token
+
+        polygons_pred[::2] *= w
+        polygons_pred[1::2] *= h
+        if len(idx_list) > 0:   # multiple polygons
+            polygons = []
+            pred_idx = 0
+            for idx in idx_list:
+                cur_idx = idx
+                if pred_idx == cur_idx or pred_idx == size:
+                    pass
+                else:
+                    polygons.append(polygons_pred[pred_idx: cur_idx])
+                pred_idx = cur_idx + 1
+        else:
+            polygons = [polygons_pred]
+
+        n_poly_pred.append(len(polygons))
+        hyps.append(polygons)
+        hyps_det.append(gen_out_i_det)
+        
+
+    pred_mask = get_mask_from_codes(hyps[0], (h, w))
+    pred_overlayed = overlay_predictions(np.asarray(image), pred_mask, hyps[0], hyps_det[0])
+
+    return pred_overlayed, np.array(pred_mask*255, dtype=np.uint8)
+
+

evaluate.py

@@ -0,0 +1,185 @@
+#!/usr/bin/env python3 -u
+# Copyright 2022 The OFA-Sys Team. 
+# All rights reserved.
+# This source code is licensed under the Apache 2.0 license 
+# found in the LICENSE file in the root directory.
+
+import logging
+import os
+import sys
+
+import numpy as np
+import torch
+from fairseq import distributed_utils, options, tasks, utils
+from fairseq.dataclass.utils import convert_namespace_to_omegaconf
+from fairseq.logging import progress_bar
+from fairseq.utils import reset_logging
+from omegaconf import DictConfig
+
+from utils import checkpoint_utils
+from utils.eval_utils import eval_step, merge_results
+
+logging.basicConfig(
+    format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    level=os.environ.get("LOGLEVEL", "INFO").upper(),
+    stream=sys.stdout,
+)
+logger = logging.getLogger("ofa.evaluate")
+
+
+def apply_half(t):
+    if t.dtype is torch.float32:
+        return t.to(dtype=torch.half)
+    return t
+
+
+def main(cfg: DictConfig, **kwargs):
+    utils.import_user_module(cfg.common)
+
+    reset_logging()
+    logger.info(cfg)
+
+    assert (
+            cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None
+    ), "Must specify batch size either with --max-tokens or --batch-size"
+
+    # Fix seed for stochastic decoding
+    if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
+        np.random.seed(cfg.common.seed)
+        utils.set_torch_seed(cfg.common.seed)
+
+    use_fp16 = cfg.common.fp16
+    use_cuda = torch.cuda.is_available() and not cfg.common.cpu
+
+    if use_cuda:
+        torch.cuda.set_device(cfg.distributed_training.device_id)
+
+    # Load ensemble
+    overrides = eval(cfg.common_eval.model_overrides)
+    # Deal with beam-search / all-candidate VQA eval
+    if cfg.task._name == "vqa_gen":
+        overrides['val_inference_type'] = "beamsearch" if kwargs['beam_search_vqa_eval'] else "allcand"
+
+    logger.info("loading model(s) from {}".format(cfg.common_eval.path))
+    if kwargs["zero_shot"]:
+        task = tasks.setup_task(cfg.task)
+        models, saved_cfg = checkpoint_utils.load_model_ensemble(
+            utils.split_paths(cfg.common_eval.path),
+            arg_overrides=overrides,
+            task=task,
+            suffix=cfg.checkpoint.checkpoint_suffix,
+            strict=(cfg.checkpoint.checkpoint_shard_count == 1),
+            num_shards=cfg.checkpoint.checkpoint_shard_count,
+        )
+    else:
+        models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+            utils.split_paths(cfg.common_eval.path),
+            arg_overrides=overrides,
+            suffix=cfg.checkpoint.checkpoint_suffix,
+            strict=(cfg.checkpoint.checkpoint_shard_count == 1),
+            num_shards=cfg.checkpoint.checkpoint_shard_count,
+        )
+
+    # loading the dataset should happen after the checkpoint has been loaded so we can give it the saved task config
+    task.load_dataset(cfg.dataset.gen_subset, task_cfg=saved_cfg.task)
+
+    # Move models to GPU
+    for model, ckpt_path in zip(models, utils.split_paths(cfg.common_eval.path)):
+        if kwargs['ema_eval']:
+            logger.info("loading EMA weights from {}".format(ckpt_path))
+            model.load_state_dict(checkpoint_utils.load_ema_from_checkpoint(ckpt_path)['model'])
+        model.eval()
+        if use_fp16:
+            model.half()
+        if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
+            model.cuda()
+        model.prepare_for_inference_(cfg)
+
+    # Load dataset (possibly sharded)
+    itr = task.get_batch_iterator(
+        dataset=task.dataset(cfg.dataset.gen_subset),
+        max_tokens=cfg.dataset.max_tokens,
+        max_sentences=cfg.dataset.batch_size,
+        max_positions=utils.resolve_max_positions(
+            task.max_positions(), *[m.max_positions() for m in models]
+        ),
+        ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test,
+        required_batch_size_multiple=cfg.dataset.required_batch_size_multiple,
+        seed=cfg.common.seed,
+        num_shards=cfg.distributed_training.distributed_world_size,
+        shard_id=cfg.distributed_training.distributed_rank,
+        num_workers=cfg.dataset.num_workers,
+        data_buffer_size=cfg.dataset.data_buffer_size,
+    ).next_epoch_itr(shuffle=False)
+    progress = progress_bar.progress_bar(
+        itr,
+        log_format=cfg.common.log_format,
+        log_interval=cfg.common.log_interval,
+        default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"),
+    )
+
+    # Initialize generator
+    generator = task.build_generator(models, cfg.generation)
+
+    # for sample in progress:
+    #     if "net_input" not in sample:
+    #         continue
+    #     sample = utils.move_to_cuda(sample) if use_cuda else sample
+    #     sample = utils.apply_to_sample(apply_half, sample) if cfg.common.fp16 else sample
+    #     with torch.no_grad():
+    #         eval_step(task, generator, models, sample, **kwargs)
+    #     progress.log({"sentences": sample["nsentences"]})
+    #
+    # merge_results(task, cfg, logger, kwargs['result_dir'])
+
+    results = []
+    prec_list = [.5, .6, .7, .8, .9]
+    prec_score_sum = [torch.FloatTensor([0]).cuda() for _ in prec_list]
+    f_score_sum = torch.FloatTensor([0]).cuda()
+    ap_det_score_sum = torch.FloatTensor([0]).cuda()
+    score_sum = torch.FloatTensor([0]).cuda()
+    score_cnt = torch.FloatTensor([0]).cuda()
+    cum_I_sum = torch.FloatTensor([0]).cuda()
+    cum_U_sum = torch.FloatTensor([0]).cuda()
+    for sample in progress:
+        if "net_input" not in sample:
+            continue
+        sample = utils.move_to_cuda(sample) if use_cuda else sample
+        sample = utils.apply_to_sample(apply_half, sample) if cfg.common.fp16 else sample
+        with torch.no_grad():
+            result, scores, f_scores, ap_scores, cum_I, cum_U = eval_step(task, generator, models, sample, **kwargs)
+        results += result
+        for prec_score, prec in zip(prec_score_sum, prec_list):
+            prec_score += sum(scores >= prec) if scores is not None else 0
+        cum_I_sum += sum(cum_I) if scores is not None else 0
+        cum_U_sum += sum(cum_U) if scores is not None else 0
+        score_sum += sum(scores) if scores is not None else 0
+        f_score_sum += sum(f_scores) if scores is not None else 0
+        ap_det_score_sum += sum(ap_scores) if scores is not None else 0
+        score_cnt += len(scores) if scores is not None else 0
+        progress.log({"sentences": sample["nsentences"]})
+
+    merge_results(task, cfg, logger, score_cnt, score_sum, f_score_sum, ap_det_score_sum,prec_score_sum, cum_I_sum, cum_U_sum, results)
+
+
+def cli_main():
+    parser = options.get_generation_parser()
+    parser.add_argument("--ema-eval", action='store_true', help="Use EMA weights to make evaluation.")
+    parser.add_argument("--beam-search-vqa-eval", action='store_true', help="Use beam search for vqa evaluation (faster inference speed but sub-optimal result), if not specified, we compute scores for each answer in the candidate set, which is slower but can obtain best result.")
+    parser.add_argument("--zero-shot", action='store_true')
+    parser.add_argument("--vis_dir", type=str, default=None)
+    parser.add_argument("--result_dir", type=str, default=None)
+    parser.add_argument("--vis", action='store_true', default=False)
+    args = options.parse_args_and_arch(parser)
+    cfg = convert_namespace_to_omegaconf(args)
+    if args.result_dir is None:
+        args.result_dir = args.vis_dir
+    distributed_utils.call_main(
+        cfg, main, ema_eval=args.ema_eval, beam_search_vqa_eval=args.beam_search_vqa_eval, zero_shot=args.zero_shot,
+        vis_dir=args.vis_dir, vis=args.vis, result_dir=args.result_dir
+    )
+
+
+if __name__ == "__main__":
+    cli_main()

fairseq/.github/ISSUE_TEMPLATE.md

@@ -0,0 +1,3 @@
+## 👉 [Please follow one of these issue templates](https://github.com/pytorch/fairseq/issues/new/choose) 👈
+
+Note: to keep the backlog clean and actionable, issues may be immediately closed if they do not follow one of the above issue templates.

fairseq/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,43 @@
+---
+name: 🐛 Bug Report
+about: Submit a bug report to help us improve
+labels: 'bug, needs triage'
+---
+
+## 🐛 Bug
+
+<!-- A clear and concise description of what the bug is. -->
+
+### To Reproduce
+
+Steps to reproduce the behavior (**always include the command you ran**):
+
+1. Run cmd '....'
+2. See error
+
+<!-- If you have a code sample, error messages, stack traces, please provide it here as well -->
+
+
+#### Code sample
+<!-- Ideally attach a minimal code sample to reproduce the decried issue.
+Minimal means having the shortest code but still preserving the bug. -->
+
+### Expected behavior
+
+<!-- A clear and concise description of what you expected to happen. -->
+
+### Environment
+
+ - fairseq Version (e.g., 1.0 or main):
+ - PyTorch Version (e.g., 1.0)
+ - OS (e.g., Linux):
+ - How you installed fairseq (`pip`, source):
+ - Build command you used (if compiling from source):
+ - Python version:
+ - CUDA/cuDNN version:
+ - GPU models and configuration:
+ - Any other relevant information:
+
+### Additional context
+
+<!-- Add any other context about the problem here. -->

fairseq/.github/ISSUE_TEMPLATE/documentation.md

@@ -0,0 +1,15 @@
+---
+name: 📚 Documentation/Typos
+about: Report an issue related to documentation or a typo
+labels: 'documentation, needs triage'
+---
+
+## 📚 Documentation
+
+For typos and doc fixes, please go ahead and:
+
+1. Create an issue.
+2. Fix the typo.
+3. Submit a PR.
+
+Thanks!

fairseq/.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,24 @@
+---
+name: 🚀 Feature Request
+about: Submit a proposal/request for a new feature
+labels: 'enhancement, help wanted, needs triage'
+---
+
+## 🚀 Feature Request
+<!-- A clear and concise description of the feature proposal -->
+
+### Motivation
+
+<!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too -->
+
+### Pitch
+
+<!-- A clear and concise description of what you want to happen. -->
+
+### Alternatives
+
+<!-- A clear and concise description of any alternative solutions or features you've considered, if any. -->
+
+### Additional context
+
+<!-- Add any other context or screenshots about the feature request here. -->

fairseq/.github/ISSUE_TEMPLATE/how-to-question.md

@@ -0,0 +1,33 @@
+---
+name: ❓ Questions/Help
+about: If you have questions, please first search existing issues and docs
+labels: 'question, needs triage'
+---
+
+## ❓ Questions and Help
+
+### Before asking:
+1. search the issues.
+2. search the docs.
+
+<!-- If you still can't find what you need: -->
+
+#### What is your question?
+
+#### Code
+
+<!-- Please paste a code snippet if your question requires it! -->
+
+#### What have you tried?
+
+#### What's your environment?
+
+ - fairseq Version (e.g., 1.0 or main):
+ - PyTorch Version (e.g., 1.0)
+ - OS (e.g., Linux):
+ - How you installed fairseq (`pip`, source):
+ - Build command you used (if compiling from source):
+ - Python version:
+ - CUDA/cuDNN version:
+ - GPU models and configuration:
+ - Any other relevant information:

fairseq/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,16 @@
+# Before submitting
+
+- [ ] Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
+- [ ] Did you read the [contributor guideline](https://github.com/pytorch/fairseq/blob/main/CONTRIBUTING.md)?
+- [ ] Did you make sure to update the docs?
+- [ ] Did you write any new necessary tests?
+
+## What does this PR do?
+Fixes # (issue).
+
+## PR review
+Anyone in the community is free to review the PR once the tests have passed.
+If we didn't discuss your PR in Github issues there's a high chance it will not be merged.
+
+## Did you have fun?
+Make sure you had fun coding 🙃

fairseq/.github/stale.yml

@@ -0,0 +1,30 @@
+# Configuration for probot-stale - https://github.com/probot/stale
+# Mostly copied from github.com/facebook/react/blob/master/.github/stale.yml
+# Number of days of inactivity before an issue becomes stale
+daysUntilStale: 90
+# Number of days of inactivity before a stale issue is closed
+daysUntilClose: 7
+# Issues with these labels will never be considered stale
+exemptLabels:
+  - bug
+# Label to use when marking an issue as stale
+staleLabel: stale
+issues:
+  # Comment to post when marking an issue as stale.
+  markComment: >
+    This issue has been automatically marked as stale.
+    **If this issue is still affecting you, please leave any comment** (for example, "bump"), and we'll keep it open.
+    We are sorry that we haven't been able to prioritize it yet. If you have any new additional information, please include it with your comment!
+  # Comment to post when closing a stale issue.
+  closeComment: >
+    Closing this issue after a prolonged period of inactivity. If this issue is still present in the latest release, please create a new issue with up-to-date information. Thank you!
+pulls:
+  # Comment to post when marking a pull request as stale.
+  markComment: >
+    This pull request has been automatically marked as stale.
+    **If this pull request is still relevant, please leave any comment** (for example, "bump"), and we'll keep it open.
+    We are sorry that we haven't been able to prioritize reviewing it yet. Your contribution is very much appreciated.
+  # Comment to post when closing a stale pull request.
+  closeComment: >
+    Closing this pull request after a prolonged period of inactivity. If this issue is still present in the latest release, please ask for this pull request to be reopened. Thank you!
+

fairseq/.github/workflows/build.yml

@@ -0,0 +1,55 @@
+name: build
+
+on:
+  # Trigger the workflow on push to main or any pull request
+  push:
+    branches:
+      - main
+  pull_request:
+
+jobs:
+  build:
+
+    strategy:
+      max-parallel: 4
+      matrix:
+        platform: [ubuntu-latest, macos-latest]
+        python-version: [3.6, 3.7]
+
+    runs-on: ${{ matrix.platform }}
+
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Conditionally install pytorch
+      if: matrix.platform == 'windows-latest'
+      run: pip3 install torch -f https://download.pytorch.org/whl/torch_stable.html
+
+    - name: Install locally
+      run: |
+        python -m pip install --upgrade pip
+        git submodule update --init --recursive
+        python setup.py build_ext --inplace
+        python -m pip install --editable .
+
+    - name: Install optional test requirements
+      run: |
+        python -m pip install iopath transformers pyarrow
+        python -m pip install git+https://github.com/facebookresearch/fairscale.git@main
+
+    - name: Lint with flake8
+      run: |
+        pip install flake8
+        # stop the build if there are Python syntax errors or undefined names
+        flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics --extend-exclude fairseq/model_parallel/megatron
+        # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+        flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics --extend-exclude fairseq/model_parallel/megatron
+
+    - name: Run tests
+      run: |
+          python setup.py test

fairseq/.github/workflows/build_wheels.yml

@@ -0,0 +1,41 @@
+name: build_wheels
+
+on:
+  push:
+    branches:
+      - v[0-9]+.[0-9]+.[x0-9]+
+    tags:
+      - v*
+
+jobs:
+  build_wheels:
+    name: Build wheels on ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Install Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: '3.7'
+
+      - name: Install cibuildwheel
+        run: |
+          python -m pip install cibuildwheel
+
+      - name: Build wheels for CPython
+        run: |
+          python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: "cp36-*64 cp37-*64 cp38-*64"
+          CIBW_MANYLINUX_X86_64_IMAGE: manylinux1
+          CIBW_BEFORE_BUILD: git submodule update --init --recursive && pip install .
+
+      - uses: actions/upload-artifact@v2
+        with:
+          name: wheels
+          path: ./dist/*.whl

fairseq/.gitignore

@@ -0,0 +1,136 @@
+# JetBrains PyCharm IDE
+.idea/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# macOS dir files
+.DS_Store
+
+# Distribution / packaging
+.Python
+env/
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Checkpoints
+checkpoints
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# pyenv
+.python-version
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# dotenv
+.env
+
+# virtualenv
+.venv
+venv/
+ENV/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+
+# Generated files
+/fairseq/temporal_convolution_tbc
+/fairseq/modules/*_layer/*_forward.cu
+/fairseq/modules/*_layer/*_backward.cu
+/fairseq/version.py
+
+# data
+data-bin/
+
+# reranking
+/examples/reranking/rerank_data
+
+# Cython-generated C++ source files
+/fairseq/data/data_utils_fast.cpp
+/fairseq/data/token_block_utils_fast.cpp
+
+# VSCODE
+.vscode/ftp-sync.json
+.vscode/settings.json
+
+# Experimental Folder
+experimental/*
+
+# Weights and Biases logs
+wandb/

fairseq/.gitmodules

@@ -0,0 +1,4 @@
+[submodule "fairseq/model_parallel/megatron"]
+    path = fairseq/model_parallel/megatron
+    url = https://github.com/ngoyal2707/Megatron-LM
+    branch = fairseq

fairseq/CODE_OF_CONDUCT.md

@@ -0,0 +1,77 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+  advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+  address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <conduct@pytorch.org>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq
+

fairseq/CONTRIBUTING.md

@@ -0,0 +1,28 @@
+# Contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq)
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+## License
+By contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq),
+you agree that your contributions will be licensed under the LICENSE file in
+the root directory of this source tree.

fairseq/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) Facebook, Inc. and its affiliates.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

fairseq/README.md

@@ -0,0 +1,229 @@
+<p align="center">
+  <img src="docs/fairseq_logo.png" width="150">
+  <br />
+  <br />
+  <a href="https://github.com/pytorch/fairseq/blob/main/LICENSE"><img alt="MIT License" src="https://img.shields.io/badge/license-MIT-blue.svg" /></a>
+  <a href="https://github.com/pytorch/fairseq/releases"><img alt="Latest Release" src="https://img.shields.io/github/release/pytorch/fairseq.svg" /></a>
+  <a href="https://github.com/pytorch/fairseq/actions?query=workflow:build"><img alt="Build Status" src="https://github.com/pytorch/fairseq/workflows/build/badge.svg" /></a>
+  <a href="https://fairseq.readthedocs.io/en/latest/?badge=latest"><img alt="Documentation Status" src="https://readthedocs.org/projects/fairseq/badge/?version=latest" /></a>
+</p>
+
+--------------------------------------------------------------------------------
+
+Fairseq(-py) is a sequence modeling toolkit that allows researchers and
+developers to train custom models for translation, summarization, language
+modeling and other text generation tasks.
+
+We provide reference implementations of various sequence modeling papers:
+
+<details><summary>List of implemented papers</summary><p>
+
+* **Convolutional Neural Networks (CNN)**
+  + [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/conv_lm/README.md)
+  + [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+  + [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+  + [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+  + [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)
+* **LightConv and DynamicConv models**
+  + [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* **Long Short-Term Memory (LSTM) networks**
+  + Effective Approaches to Attention-based Neural Machine Translation (Luong et al., 2015)
+* **Transformer (self-attention) networks**
+  + Attention Is All You Need (Vaswani et al., 2017)
+  + [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+  + [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+  + [Adaptive Input Representations for Neural Language Modeling (Baevski and Auli, 2018)](examples/language_model/README.adaptive_inputs.md)
+  + [Lexically constrained decoding with dynamic beam allocation (Post & Vilar, 2018)](examples/constrained_decoding/README.md)
+  + [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context (Dai et al., 2019)](examples/truncated_bptt/README.md)
+  + [Adaptive Attention Span in Transformers (Sukhbaatar et al., 2019)](examples/adaptive_span/README.md)
+  + [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
+  + [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
+  + [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
+  + [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md )
+  + [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+  + [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+  + [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+  + [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+  + [Generating Medical Reports from Patient-Doctor Conversations Using Sequence-to-Sequence Models (Enarvi et al., 2020)](examples/pointer_generator/README.md)
+  + [Linformer: Self-Attention with Linear Complexity (Wang et al., 2020)](examples/linformer/README.md)
+  + [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+  + [Deep Transformers with Latent Depth (Li et al., 2020)](examples/latent_depth/README.md)
+  + [Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al., 2020)](https://arxiv.org/abs/2006.13979)
+  + [Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (Hsu, et al., 2021)](https://arxiv.org/abs/2104.01027)
+  + [Unsupervised Speech Recognition (Baevski, et al., 2021)](https://arxiv.org/abs/2105.11084)
+* **Non-autoregressive Transformers**
+  + Non-Autoregressive Neural Machine Translation (Gu et al., 2017)
+  + Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al. 2018)
+  + Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al. 2019)
+  + Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)
+  + [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
+* **Finetuning**
+  + [Better Fine-Tuning by Reducing Representational Collapse (Aghajanyan et al. 2020)](examples/rxf/README.md)
+
+</p></details>
+
+### What's New:
+
+* September 2021 [`master` branch renamed to `main`](https://github.com/github/renaming).
+* July 2021 [Released DrNMT code](examples/discriminative_reranking_nmt/README.md)
+* July 2021 [Released Robust wav2vec 2.0 model](examples/wav2vec/README.md)
+* June 2021 [Released XLMR-XL and XLMR-XXL models](examples/xlmr/README.md)
+* May 2021 [Released Unsupervised Speech Recognition code](examples/wav2vec/unsupervised/README.md)
+* March 2021 [Added full parameter and optimizer state sharding + CPU offloading](examples/fully_sharded_data_parallel/README.md)
+* February 2021 [Added LASER training code](examples/laser/README.md)
+* December 2020: [Added Adaptive Attention Span code](examples/adaptive_span/README.md)
+* December 2020: [GottBERT model and code released](examples/gottbert/README.md)
+* November 2020: Adopted the [Hydra](https://github.com/facebookresearch/hydra) configuration framework
+  * [see documentation explaining how to use it for new and existing projects](docs/hydra_integration.md)
+* November 2020: [fairseq 0.10.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.10.0)
+* October 2020: [Added R3F/R4F (Better Fine-Tuning) code](examples/rxf/README.md)
+* October 2020: [Deep Transformer with Latent Depth code released](examples/latent_depth/README.md)
+* October 2020: [Added CRISS models and code](examples/criss/README.md)
+
+<details><summary>Previous updates</summary><p>
+
+* September 2020: [Added Linformer code](examples/linformer/README.md)
+* September 2020: [Added pointer-generator networks](examples/pointer_generator/README.md)
+* August 2020: [Added lexically constrained decoding](examples/constrained_decoding/README.md)
+* August 2020: [wav2vec2 models and code released](examples/wav2vec/README.md)
+* July 2020: [Unsupervised Quality Estimation code released](examples/unsupervised_quality_estimation/README.md)
+* May 2020: [Follow fairseq on Twitter](https://twitter.com/fairseq)
+* April 2020: [Monotonic Multihead Attention code released](examples/simultaneous_translation/README.md)
+* April 2020: [Quant-Noise code released](examples/quant_noise/README.md)
+* April 2020: [Initial model parallel support and 11B parameters unidirectional LM released](examples/megatron_11b/README.md)
+* March 2020: [Byte-level BPE code released](examples/byte_level_bpe/README.md)
+* February 2020: [mBART model and code released](examples/mbart/README.md)
+* February 2020: [Added tutorial for back-translation](https://github.com/pytorch/fairseq/tree/main/examples/backtranslation#training-your-own-model-wmt18-english-german)
+* December 2019: [fairseq 0.9.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.9.0)
+* November 2019: [VizSeq released (a visual analysis toolkit for evaluating fairseq models)](https://facebookresearch.github.io/vizseq/docs/getting_started/fairseq_example)
+* November 2019: [CamemBERT model and code released](examples/camembert/README.md)
+* November 2019: [BART model and code released](examples/bart/README.md)
+* November 2019: [XLM-R models and code released](examples/xlmr/README.md)
+* September 2019: [Nonautoregressive translation code released](examples/nonautoregressive_translation/README.md)
+* August 2019: [WMT'19 models released](examples/wmt19/README.md)
+* July 2019: fairseq relicensed under MIT license
+* July 2019: [RoBERTa models and code released](examples/roberta/README.md)
+* June 2019: [wav2vec models and code released](examples/wav2vec/README.md)
+
+</p></details>
+
+### Features:
+
+* multi-GPU training on one machine or across multiple machines (data and model parallel)
+* fast generation on both CPU and GPU with multiple search algorithms implemented:
+  + beam search
+  + Diverse Beam Search ([Vijayakumar et al., 2016](https://arxiv.org/abs/1610.02424))
+  + sampling (unconstrained, top-k and top-p/nucleus)
+  + [lexically constrained decoding](examples/constrained_decoding/README.md) (Post & Vilar, 2018)
+* [gradient accumulation](https://fairseq.readthedocs.io/en/latest/getting_started.html#large-mini-batch-training-with-delayed-updates) enables training with large mini-batches even on a single GPU
+* [mixed precision training](https://fairseq.readthedocs.io/en/latest/getting_started.html#training-with-half-precision-floating-point-fp16) (trains faster with less GPU memory on [NVIDIA tensor cores](https://developer.nvidia.com/tensor-cores))
+* [extensible](https://fairseq.readthedocs.io/en/latest/overview.html): easily register new models, criterions, tasks, optimizers and learning rate schedulers
+* [flexible configuration](docs/hydra_integration.md) based on [Hydra](https://github.com/facebookresearch/hydra) allowing a combination of code, command-line and file based configuration
+* [full parameter and optimizer state sharding](examples/fully_sharded_data_parallel/README.md)
+* [offloading parameters to CPU](examples/fully_sharded_data_parallel/README.md)
+
+We also provide [pre-trained models for translation and language modeling](#pre-trained-models-and-examples)
+with a convenient `torch.hub` interface:
+
+``` python
+en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model')
+en2de.translate('Hello world', beam=5)
+# 'Hallo Welt'
+```
+
+See the PyTorch Hub tutorials for [translation](https://pytorch.org/hub/pytorch_fairseq_translation/)
+and [RoBERTa](https://pytorch.org/hub/pytorch_fairseq_roberta/) for more examples.
+
+# Requirements and Installation
+
+* [PyTorch](http://pytorch.org/) version >= 1.5.0
+* Python version >= 3.6
+* For training new models, you'll also need an NVIDIA GPU and [NCCL](https://github.com/NVIDIA/nccl)
+* **To install fairseq** and develop locally:
+
+``` bash
+git clone https://github.com/pytorch/fairseq
+cd fairseq
+pip install --editable ./
+
+# on MacOS:
+# CFLAGS="-stdlib=libc++" pip install --editable ./
+
+# to install the latest stable release (0.10.x)
+# pip install fairseq
+```
+
+* **For faster training** install NVIDIA's [apex](https://github.com/NVIDIA/apex) library:
+
+``` bash
+git clone https://github.com/NVIDIA/apex
+cd apex
+pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
+  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
+  --global-option="--fast_multihead_attn" ./
+```
+
+* **For large datasets** install [PyArrow](https://arrow.apache.org/docs/python/install.html#using-pip): `pip install pyarrow`
+* If you use Docker make sure to increase the shared memory size either with `--ipc=host` or `--shm-size`
+ as command line options to `nvidia-docker run` .
+
+# Getting Started
+
+The [full documentation](https://fairseq.readthedocs.io/) contains instructions
+for getting started, training new models and extending fairseq with new model
+types and tasks.
+
+# Pre-trained models and examples
+
+We provide pre-trained models and pre-processed, binarized test sets for several tasks listed below,
+as well as example training and evaluation commands.
+
+* [Translation](examples/translation/README.md): convolutional and transformer models are available
+* [Language Modeling](examples/language_model/README.md): convolutional and transformer models are available
+
+We also have more detailed READMEs to reproduce results from specific papers:
+
+* [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+* [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+* [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+* [Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020)](examples/quant_noise/README.md)
+* [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+* [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+* [Reducing Transformer Depth on Demand with Structured Dropout (Fan et al., 2019)](examples/layerdrop/README.md)
+* [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md)
+* [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
+* [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
+* [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
+* [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)
+* [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
+* [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+* [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+* [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+* [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+* [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+* [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/README.conv.md)
+
+# Join the fairseq community
+
+* Twitter: https://twitter.com/fairseq
+* Facebook page: https://www.facebook.com/groups/fairseq.users
+* Google group: https://groups.google.com/forum/#!forum/fairseq-users
+
+# License
+
+fairseq(-py) is MIT-licensed.
+The license applies to the pre-trained models as well.
+
+# Citation
+
+Please cite as:
+
+``` bibtex
+@inproceedings{ott2019fairseq,
+  title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling},
+  author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli},
+  booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations},
+  year = {2019},
+}
+```

fairseq/examples/.gitignore

@@ -0,0 +1,2 @@
+!*/*.sh
+!*/*.md

fairseq/examples/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+try:
+    from fairseq.version import __version__  # noqa
+except ImportError:
+    pass

fairseq/examples/adaptive_span/README.md

@@ -0,0 +1,90 @@
+# Adaptive Span
+
+Adaptive Span is a novel self-attention mechanism that can learn its optimal
+attention span. This allows us to extend significantly the maximum context size
+used in Transformer, while maintaining control over their memory footprint
+and computational time. It uses the Truncated BPTT technique for training,
+as in [transformerXL](https://github.com/pytorch/fairseq/blob/main/examples/truncated_bptt/README.md).
+
+Adaptive Span was introduced by paper:
+[Adaptive Attention Span in Transformers](https://arxiv.org/abs/1905.07799),
+which achieved state-of-the-art language modeling results at the time of publication.
+
+We manage to reproduce their result in fairseq and keep most of the
+[original implementation](https://github.com/facebookresearch/adaptive-span) untouched.
+You can refer to the their sweep file as well if any combination of hyperparameter is not clear.
+
+##### 0. Setup
+
+First you need to process the Enwik8 dataset, we use the pre-tokenized dataset
+from [adaptive span paper](https://github.com/facebookresearch/adaptive-span/blob/master/get_data.sh).
+You can download the dataset, and then run:
+```bash
+fairseq-preprocess --only-source --trainpref ~/data/enwik8/train.txt \
+    --validpref ~/data/enwik8/valid.txt --testpref ~/data/enwik8/test.txt \
+    --destdir ~/data/enwik8/data-bin/ --joined-dictionary --workers 20
+```
+
+##### 1. Train a Adaptive Span model on Enwik8
+
+We will train a 12-layer Adaptive Span model following the [hyperparameters
+used in the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).
+
+The following command assumes 4 GPUs, so that the total batch size is 64
+sequences (4 x 16). Training should take 2-3 days on 4 V100 GPUs:
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+    --user-dir examples/adaptive_span \
+    --data  ~/data/enwik8/data-bin/ \
+    --fp16 --fp16-no-flatten-grads --max-update 600000 \
+    --task truncated_bptt_lm --tokens-per-sample 512 --arch adaptive_span \
+    --n-layer 12 --d-model 512 --n-head 8 --d-inner 2048 --dropout 0.3 \
+    --attn-span 8192 --optimizer adagrad_with_grad_clip --adagrad-clip 0.03 \
+    --validate-interval-updates 1000 \
+    --lr-scheduler fixed --warmup-updates 32000 --batch-size-valid 32 \
+    --lr 0.07 --criterion adaptive_span_loss --batch-size 16 --update-freq 1 \
+    --seed 2 --log-format json --log-interval 25 --aux-loss-scaler 5e-07
+```
+This should land around 1.05 on validation, 1.03 on test. You can lower the
+--aux-loss-scaler for better performance (longer span). It gives ~0.03 bpc
+improvement to the transformerXL baseline here.
+If training on a single GPU, set `--update-freq=4` to accumulate 4x gradients
+and simulate training on 4 GPUs.
+You can also reproduce the transformerXL result on enwik8 using this code base.
+It should land around 1.06 on test,matching the [original paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_enwik8_base.sh).
+You can try by
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+    --user-dir examples/truncated_bptt \
+    ~/data/enwik8/data-bin/ \
+    --task truncated_bptt_lm  --fp16 --max-update 400000 \
+    --tokens-per-sample 512 --arch transformer_xl --n-layer 12 \
+    --d-model 512 --n-head 8 --d-head 64 --d-inner 2048 --dropout 0.1 \
+    --dropatt 0.0 --mem-len 512 --optimizer adam --clip-norm 0.25 \
+    --lr-scheduler cosine --warmup-updates 0 \
+    --lr 0.0 --lr 0.00025 --batch-size 15 \
+    --update-freq 1 --seed 2 --log-format json --log-interval 25 \
+    --fp16
+```
+
+##### 2. Evaluate
+For Adaptive Span:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+ --user-dir examples/adaptive_span \
+ --task truncated_bptt_lm --batch-size 8 --tokens-per-sample 512 --gen-subset test
+```
+For Transformer-XL evaluation:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+    --user-dir examples/truncated_bptt/ --task truncated_bptt_lm --batch-size 8 \
+    --tokens-per-sample 80 \
+    --model-overrides '{"mem_len":2100,"clamp_len":820,"same_length":True}' \
+    --gen-subset valid
+```
+
+*Note:* During training the model saw 512 tokens of context
+(``--tokens-per-sample=512``), with batch size 8. These settings match the evaluation
+settings from [the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).

fairseq/examples/adaptive_span/__init__.py

@@ -0,0 +1,19 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import importlib
+import os
+
+# automatically import any Python files in the current directory
+cur_dir = os.path.dirname(__file__)
+for file in os.listdir(cur_dir):
+    path = os.path.join(cur_dir, file)
+    if (
+        not file.startswith("_")
+        and not file.startswith(".")
+        and (file.endswith(".py") or os.path.isdir(path))
+    ):
+        mod_name = file[: file.find(".py")] if file.endswith(".py") else file
+        module = importlib.import_module(__name__ + "." + mod_name)