init

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+refTools/evaluation/meteor/meteor-1.5.jar filter=lfs diff=lfs merge=lfs -text
+refTools/evaluation/tokenizer/stanford-corenlp-3.4.1.jar filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 mshukor
+
+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.

README.md

@@ -1,12 +1,151 @@
----
-title: EP ALM
-emoji: 🐢
-colorFrom: pink
-colorTo: gray
-sdk: gradio
-sdk_version: 3.36.1
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# eP-ALM: Efficient Perceptual Augmentation of Language Models
+
+<p align="center">
+    <img src="images/teaser.jpg" width="500"/>
+</p>
+
+Official implementation of the paper: 
+- [eP-ALM: Efficient Perceptual Augmentation of Language Models](https://arxiv.org/abs/2303.11403)
+
+In this repo, you will find the pretrained models and code to train and evaluate eP-ALM on Image/Video/Audio-Text tasks.
+
+## News
+
+* **[June-2023]** A new version of the paper is released on arXiv: 
+    * We re-evaluate the models with greedy decoding.
+    * We add comparison with SoTA.
+    * We add new experiments, including pretraining on CC3M and evaluation in zero-shot (check Appendix).
+* **[May-2023]** The code is optimized to train and evaluate with float16 mixed precision, using the accelerate library 🤗.
+* **[May-2023]** We found greedy decoding with beam search is significantly better than multinomial/random sampling.
+* **[20-March-2023]** The paper is released on arXiv.
+* **[March-2023]** The paper is submitted for publication and currently under review. 
+
+## Summary:
+
+* [Introduction](#introduction)
+* [Download](#download)
+* [Installation](#installation)
+* [Evaluation](#evaluation)
+* [Accelerated Training 🤗](#accelerated-training)
+* [Training](#training)
+* [Citation](#citation)
+* [Acknowledgment](#acknowledgment)
+
+
+## Introduction
+
+
+
+Large Language Models (LLMs) have so far impressed the world, with unprecedented capabilities that emerge in models at large scales. On the vision side, transformer models (i.e., ViT) are following the same trend, achieving the best performance on challenging benchmarks. With the abundance of such unimodal models, a natural question arises; do we need also to follow this trend to tackle multimodal tasks? In this work, we propose to rather direct effort to efficient adaptations of existing models, and propose to augment Language Models with perception. Existing approaches for adapting pretrained models for vision-language tasks still rely on several key components that hinder their efficiency. In particular, they still train a large number of parameters, rely on large multimodal pretraining, use encoders (e.g., CLIP) trained on huge image-text datasets, and add significant inference overhead. In addition, most of these approaches have focused on Zero-Shot and In Context Learning, with little to no effort on direct finetuning. We investigate the minimal computational effort needed to adapt unimodal models for multimodal tasks and propose a new challenging setup, alongside different approaches, that efficiently adapts unimodal pretrained models. We show that by freezing more than 99\% of total parameters, training only one linear projection layer, and prepending only one trainable token, our approach (dubbed eP-ALM) significantly outperforms other baselines on VQA and Captioning across Image, Video, and Audio modalities, following the proposed setup.
+
+
+<p align="center">
+    <img src="images/variants.jpg" width="500"/>
+</p>
+
+
+### Results
+
+> Comparison of eP-ALM with text generation-based SoTA that train significant number of parameters, including methods with large-scale pretraining. Best and next best scores are bolded and underlined respectively. FT: Finetuning. ZS: Zero-shot.
+
+<p align="center">
+    <img src="images/epalm_sota.png" width="700"/>
+</p>
+
+
+
+> Qualitative results of eP-ALM: the model is able to generate accurate answers and coherent descriptions of the image. Ground truth answers are highlighted in green (with multinomial sampling and OPT350M).
+
+<p align="center">
+    <img src="images/qual.jpg" width="1000"/>
+</p>
+
+
+
+## Download
+
+### OPT Model
+First you need to download OPT models and tokenizers. You can use the following (for OPT-2.7B) to automatically download them:
+
+```
+from transformers import AutoTokenizer, OPTModel
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-2.7b")
+model = OPTModel.from_pretrained("facebook/opt-2.7b")
+```
+
+### Pretrained Models
+We provide only the adaptation parameters (linear connection and Soft Prompt). You can download the following models:
+
+For eP-ALM_pt-L (with OPT-6.7B and ViT-L), trained with float16 mixed precision (`accelerate_training`):
+* VQA v2: [ePALM_pt_L](https://nuage.isir.upmc.fr/index.php/s/aSrCTKXsKQxAE72)
+* COCO Caption: [ePALM_pt_L](https://nuage.isir.upmc.fr/index.php/s/PNrytqFbJWJqdt3)
+* GQA: [ePALM_pt_L](https://nuage.isir.upmc.fr/index.php/s/9o7gk5gWY5ZNKLM)
+
+In the following, we provide smaller models that used to obtain the main results in the paper. Note that these models are trained with float32:
+* VQA v2: [ePALM](https://nuage.isir.upmc.fr/index.php/s/SMTJqfL62KC88z5)
+* COCO Caption: [ePALM](https://nuage.isir.upmc.fr/index.php/s/y9KZr9CEpe42443)
+* GQA: [ePALM](https://nuage.isir.upmc.fr/index.php/s/8rS84b4EH56CPZq)
+* MSR-VTT Video Caption: [ePALM](https://nuage.isir.upmc.fr/index.php/s/nCj7mz7NHgeYokP)
+* MSRVTT-QA Video QA: [ePALM](https://nuage.isir.upmc.fr/index.php/s/RysMQzH9sSf5b7P)
+* MSVD-QA: [ePALM](https://nuage.isir.upmc.fr/index.php/s/LCdLN3xg35jGCP2)
+* AudioCaps Audio Captioning: [ePALM](https://nuage.isir.upmc.fr/index.php/s/ZeeZc9zdFSgFTFC)
+
+### Data
+More details on the download and the organization of datasets can be found [here](docs/datasets.md)
+
+## Installation
+Main requirements:
+```
+python >= 3.8+
+torch >= 1.12+
+transformers >= 4.24+
+accelerate >= 0.11.0
+```
+More details can be found [here](docs/installation.md).
+
+
+## Evaluation
+To evaluate the trained models, you can use the same scripts in `run_scripts/`, and just pass the best checkpoint path to the `--evaluate` arguments.
+
+To visualize the results and test on your own images, you can use this notebook `ePALM.ipynb`.
+
+You should use the same script used for training to evaluate the model (e.g., `run_scripts/accelerate_training` for models trained with accelerate).
+
+Note that you can evaluate the models trained with float32 with `run_scripts/accelerate_training`, but the you might obtain slightly different results (e.g., for caption we obtain 102 CIDEr instead of 97 as reported in the paper).
+
+
+## Accelerated Training 🤗
+We optimized the code code based on the [accelerate](https://github.com/huggingface/accelerate) librairy. Mainly we train with mixed precision and keep the precision of the LM in float16, this significantly reduces the memory consumption (/2) and accelerates (x2) the training. 
+
+For example, after specifying the path to `config` file, `data_dir` and `output_dir`,
+to launch a training of eP-ALM_pt-L on VQA v2:
+
+```
+sh run_scripts/accelerate/image/ePALM_pt_L_vqa_acc.sh
+```
+
+To resume training, specify the initialization checkpoint to the `--resume` argument.
+
+
+## Training
+Previous models are trained with float32 precision. You can launch the training/evaluation of eP-ALM using the different scripts in `run_scripts/float32`. For example you can launch a training on VQA v2 from the following script:
+
+```
+sh run_scripts/float32/image/ePALM_vqa.sh
+```
+
+## Citation
+
+```
+@article{shukor2023ep,
+  title={eP-ALM: Efficient Perceptual Augmentation of Language Models},
+  author={Shukor, Mustafa and Dancette, Corentin and Cord, Matthieu},
+  journal={arXiv preprint arXiv:2303.11403},
+  year={2023}
+}
+```
+## Acknowledgment
+
+Some code was borrowed from [timm](https://github.com/rwightman/pytorch-image-models), [transformers](https://github.com/huggingface/transformers), [TimeSformer](https://github.com/facebookresearch/TimeSformer), and [VL-Adapter](https://github.com/ylsung/VL_adapter).
+
+

TimeSformer/.gitignore

@@ -0,0 +1,143 @@
+
+
+
+
+# Docker file from Python is inspired from here :
+# https://github.com/github/gitignore/blob/master/Python.gitignore
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# 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
+tests/report/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+
+# Cython debug symbols
+cython_debug/

TimeSformer/CODE_OF_CONDUCT.md

@@ -0,0 +1,5 @@
+# Code of Conduct
+
+Facebook has adopted a Code of Conduct that we expect project participants to adhere to.
+Please read the [full text](https://code.fb.com/codeofconduct/)
+so that you can understand what actions will and will not be tolerated.

TimeSformer/CONTRIBUTING.md

@@ -0,0 +1,25 @@
+# Contributing to TimeSformer
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `master`.
+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 TimeSformer, you agree that your contributions will be licensed
+under the [LICENSE.md](LICENSE.md) file in the root directory of this source tree.

TimeSformer/LICENSE

@@ -0,0 +1,399 @@
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public:
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

TimeSformer/README.md

@@ -0,0 +1,248 @@
+# TimeSformer
+
+This is an official pytorch implementation of our ICML 2021 paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/pdf/2102.05095.pdf). In this repository, we provide PyTorch code for training and testing our proposed TimeSformer model. TimeSformer provides an efficient video classification framework that achieves state-of-the-art results on several video action recognition benchmarks such as Kinetics-400.
+
+If you find TimeSformer useful in your research, please use the following BibTeX entry for citation.
+
+```BibTeX
+@inproceedings{gberta_2021_ICML,
+    author  = {Gedas Bertasius and Heng Wang and Lorenzo Torresani},
+    title = {Is Space-Time Attention All You Need for Video Understanding?},
+    booktitle   = {Proceedings of the International Conference on Machine Learning (ICML)}, 
+    month = {July},
+    year = {2021}
+}
+```
+
+# Model Zoo
+
+We provide TimeSformer models pretrained on Kinetics-400 (K400), Kinetics-600 (K600), Something-Something-V2 (SSv2), and HowTo100M datasets.
+
+| name | dataset | # of frames | spatial crop | acc@1 | acc@5 | url |
+| --- | --- | --- | --- | --- | --- | --- |
+| TimeSformer | K400 | 8 | 224 | 77.9 | 93.2 | [model](https://www.dropbox.com/s/g5t24we9gl5yk88/TimeSformer_divST_8x32_224_K400.pyth?dl=0) |
+| TimeSformer-HR | K400 | 16 | 448 | 79.6 | 94.0 | [model](https://www.dropbox.com/s/6f0x172lpqy3oxt/TimeSformer_divST_16x16_448_K400.pyth?dl=0) |
+| TimeSformer-L | K400 | 96 | 224 | 80.6 | 94.7 | [model](https://www.dropbox.com/s/r1iuxahif3sgimo/TimeSformer_divST_96x4_224_K400.pyth?dl=0) |
+
+| name | dataset | # of frames | spatial crop | acc@1 | acc@5 | url |
+| --- | --- | --- | --- | --- | --- | --- |
+| TimeSformer | K600 | 8 | 224 | 79.1 | 94.4 | [model](https://www.dropbox.com/s/4h2qt41m2z3aqrb/TimeSformer_divST_8x32_224_K600.pyth?dl=0) |
+| TimeSformer-HR | K600 | 16 | 448 | 81.8 | 95.8 | [model](https://www.dropbox.com/s/ft1e92g2vhvxecv/TimeSformer_divST_16x16_448_K600.pyth?dl=0) |
+| TimeSformer-L | K600 | 96 | 224 | 82.2 | 95.6 | [model](https://www.dropbox.com/s/857rx6xeclxfhdg/TimeSformer_divST_96x4_224_K600.pyth?dl=0) |
+
+| name | dataset | # of frames | spatial crop | acc@1 | acc@5 | url |
+| --- | --- | --- | --- | --- | --- | --- |
+| TimeSformer | SSv2 | 8 | 224 | 59.1 | 85.6 | [model](https://www.dropbox.com/s/tybhuml57y24wpm/TimeSformer_divST_8_224_SSv2.pyth?dl=0) |
+| TimeSformer-HR | SSv2 | 16 | 448 | 61.8 | 86.9 | [model](https://www.dropbox.com/s/9t68uzk8w2fpfnv/TimeSformer_divST_16_448_SSv2.pyth?dl=0) |
+| TimeSformer-L | SSv2 | 64 | 224 | 62.0 | 87.5 | [model](https://www.dropbox.com/s/3f1rm2al8mhprwa/TimeSformer_divST_64_224_SSv2.pyth?dl=0) |
+
+| name | dataset | # of frames | spatial crop | single clip coverage | acc@1 | url |
+| --- | --- | --- | --- | --- | --- | --- |
+| TimeSformer | HowTo100M | 8 | 224 | 8.5s | 56.8 | [model](https://www.dropbox.com/s/9v8hcm88b9tc6ff/TimeSformer_divST_8x32_224_HowTo100M.pyth?dl=0) |
+| TimeSformer | HowTo100M | 32 | 224 | 34.1s | 61.2 | [model](https://www.dropbox.com/s/4roflx4q1gscu85/TimeSformer_divST_32x32_224_HowTo100M.pyth?dl=0) |
+| TimeSformer | HowTo100M | 64 | 448 | 68.3s | 62.2 | [model](https://www.dropbox.com/s/15bvqltl1j5vyp3/TimeSformer_divST_64x32_224_HowTo100M.pyth?dl=0) |
+| TimeSformer | HowTo100M | 96 | 224 | 102.4s | 62.6 | [model](https://www.dropbox.com/s/t2mzgahnfhgakma/TimeSformer_divST_96x32_224_HowTo100M.pyth?dl=0) |
+
+We note that these models were re-trained using a slightly different implementation than the one used in the paper. Therefore, there might be a small difference in performance compared to the results reported in the paper.
+
+You can load the pretrained models as follows:
+
+```python
+import torch
+from timesformer.models.vit import TimeSformer
+
+model = TimeSformer(img_size=224, num_classes=400, num_frames=8, attention_type='divided_space_time',  pretrained_model='/path/to/pretrained/model.pyth')
+
+dummy_video = torch.randn(2, 3, 8, 224, 224) # (batch x channels x frames x height x width)
+
+pred = model(dummy_video,) # (2, 400)
+```
+
+# Installation
+
+First, create a conda virtual environment and activate it:
+```
+conda create -n timesformer python=3.7 -y
+source activate timesformer
+```
+
+Then, install the following packages:
+
+- torchvision: `pip install torchvision` or `conda install torchvision -c pytorch`
+- [fvcore](https://github.com/facebookresearch/fvcore/): `pip install 'git+https://github.com/facebookresearch/fvcore'`
+- simplejson: `pip install simplejson`
+- einops: `pip install einops`
+- timm: `pip install timm`
+- PyAV: `conda install av -c conda-forge`
+- psutil: `pip install psutil`
+- scikit-learn: `pip install scikit-learn`
+- OpenCV: `pip install opencv-python`
+- tensorboard: `pip install tensorboard`
+
+Lastly, build the TimeSformer codebase by running:
+```
+git clone https://github.com/facebookresearch/TimeSformer
+cd TimeSformer
+python setup.py build develop
+```
+
+# Usage
+
+## Dataset Preparation
+
+Please use the dataset preparation instructions provided in [DATASET.md](timesformer/datasets/DATASET.md).
+
+## Training the Default TimeSformer
+
+Training the default TimeSformer that uses divided space-time attention, and operates on 8-frame clips cropped at 224x224 spatial resolution, can be done using the following command:
+
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_divST_8x32_224.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  NUM_GPUS 8 \
+  TRAIN.BATCH_SIZE 8 \
+```
+You may need to pass location of your dataset in the command line by adding `DATA.PATH_TO_DATA_DIR path_to_your_dataset`, or you can simply add
+
+```
+DATA:
+  PATH_TO_DATA_DIR: path_to_your_dataset
+```
+
+To the yaml configs file, then you do not need to pass it to the command line every time.
+
+## Using a Different Number of GPUs
+
+If you want to use a smaller number of GPUs, you need to modify .yaml configuration files in [`configs/`](configs/). Specifically, you need to modify the NUM_GPUS, TRAIN.BATCH_SIZE, TEST.BATCH_SIZE, DATA_LOADER.NUM_WORKERS entries in each configuration file. The BATCH_SIZE entry should be the same or higher as the NUM_GPUS entry. In [`configs/Kinetics/TimeSformer_divST_8x32_224_4gpus.yaml`](configs/Kinetics/TimeSformer_divST_8x32_224_4gpus.yaml), we provide a sample configuration file for a 4 GPU setup.
+
+
+## Using Different Self-Attention Schemes
+
+If you want to experiment with different space-time self-attention schemes, e.g., space-only or joint space-time attention, use the following commands:
+
+
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_spaceOnly_8x32_224.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  NUM_GPUS 8 \
+  TRAIN.BATCH_SIZE 8 \
+```
+
+and
+
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_jointST_8x32_224.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  NUM_GPUS 8 \
+  TRAIN.BATCH_SIZE 8 \
+```
+
+## Training Different TimeSformer Variants
+
+If you want to train more powerful TimeSformer variants, e.g., TimeSformer-HR (operating on 16-frame clips sampled at 448x448 spatial resolution), and TimeSformer-L (operating on 96-frame clips sampled at 224x224 spatial resolution), use the following commands:
+
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_divST_16x16_448.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  NUM_GPUS 8 \
+  TRAIN.BATCH_SIZE 8 \
+```
+
+and
+
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_divST_96x4_224.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  NUM_GPUS 8 \
+  TRAIN.BATCH_SIZE 8 \
+```
+
+Note that for these models you will need a set of GPUs with ~32GB of memory.
+
+## Inference
+
+Use `TRAIN.ENABLE` and `TEST.ENABLE` to control whether training or testing is required for a given run. When testing, you also have to provide the path to the checkpoint model via TEST.CHECKPOINT_FILE_PATH.
+```
+python tools/run_net.py \
+  --cfg configs/Kinetics/TimeSformer_divST_8x32_224_TEST.yaml \
+  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
+  TEST.CHECKPOINT_FILE_PATH path_to_your_checkpoint \
+  TRAIN.ENABLE False \
+```
+
+## Single-Node Training via Slurm
+
+To train TimeSformer via Slurm, please check out our single node Slurm training script [`slurm_scripts/run_single_node_job.sh`](slurm_scripts/run_single_node_job.sh).
+
+
+## Multi-Node Training via Submitit
+
+Distributed training is available via Slurm and submitit
+
+```
+pip install submitit
+```
+
+To train TimeSformer model on Kinetics using 4 nodes with 8 gpus each use the following command:
+```
+python tools/submit.py --cfg configs/Kinetics/TimeSformer_divST_8x32_224.yaml --job_dir  /your/job/dir/${JOB_NAME}/ --num_shards 4 --name ${JOB_NAME} --use_volta32
+```
+
+We provide a script for launching slurm jobs in [`slurm_scripts/run_multi_node_job.sh`](slurm_scripts/run_multi_node_job.sh).
+
+## Finetuning
+
+To finetune from an existing PyTorch checkpoint add the following line in the command line, or you can also add it in the YAML config:
+
+```
+TRAIN.CHECKPOINT_FILE_PATH path_to_your_PyTorch_checkpoint
+TRAIN.FINETUNE True
+```
+
+## HowTo100M Dataset Split
+
+If you want to experiment with the long-term video modeling task on HowTo100M, please download the train/test split files from [here](https://www.dropbox.com/sh/ttvsxwqypijjuda/AACmJx1CnddW6cVBoc21eSuva?dl=0).
+
+
+# Environment
+
+The code was developed using python 3.7 on Ubuntu 20.04. For training, we used four GPU compute nodes each node containing 8 Tesla V100 GPUs (32 GPUs in total). Other platforms or GPU cards have not been fully tested.
+
+# License
+
+The majority of this work is licensed under [CC-NC 4.0 International license](LICENSE). However portions of the project are available under separate license terms: [SlowFast](https://github.com/facebookresearch/SlowFast) and [pytorch-image-models](https://github.com/rwightman/pytorch-image-models) are licensed under the Apache 2.0 license.
+
+# Contributing
+
+We actively welcome your pull requests. Please see [CONTRIBUTING.md](CONTRIBUTING.md) and [CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md) for more info.
+
+# Acknowledgements
+
+TimeSformer is built on top of [PySlowFast](https://github.com/facebookresearch/SlowFast) and [pytorch-image-models](https://github.com/rwightman/pytorch-image-models) by [Ross Wightman](https://github.com/rwightman). We thank the authors for releasing their code. If you use our model, please consider citing these works as well:
+
+```BibTeX
+@misc{fan2020pyslowfast,
+  author =       {Haoqi Fan and Yanghao Li and Bo Xiong and Wan-Yen Lo and
+                  Christoph Feichtenhofer},
+  title =        {PySlowFast},
+  howpublished = {\url{https://github.com/facebookresearch/slowfast}},
+  year =         {2020}
+}
+```
+
+```BibTeX
+@misc{rw2019timm,
+  author = {Ross Wightman},
+  title = {PyTorch Image Models},
+  year = {2019},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  doi = {10.5281/zenodo.4414861},
+  howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
+}
+```

TimeSformer/configs/Kinetics/SLOWFAST_4x16_R50.yaml

@@ -0,0 +1,63 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+  EVAL_PERIOD: 10
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 32
+  SAMPLING_RATE: 2
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 256
+  INPUT_CHANNEL_NUM: [3, 3]
+SLOWFAST:
+  ALPHA: 8
+  BETA_INV: 8
+  FUSION_CONV_CHANNEL_RATIO: 2
+  FUSION_KERNEL_SZ: 5
+RESNET:
+  ZERO_INIT_FINAL_BN: True
+  WIDTH_PER_GROUP: 64
+  NUM_GROUPS: 1
+  DEPTH: 50
+  TRANS_FUNC: bottleneck_transform
+  STRIDE_1X1: False
+  NUM_BLOCK_TEMP_KERNEL: [[3, 3], [4, 4], [6, 6], [3, 3]]
+  SPATIAL_STRIDES: [[1, 1], [2, 2], [2, 2], [2, 2]]
+  SPATIAL_DILATIONS: [[1, 1], [1, 1], [1, 1], [1, 1]]
+NONLOCAL:
+  LOCATION: [[[], []], [[], []], [[], []], [[], []]]
+  GROUP: [[1, 1], [1, 1], [1, 1], [1, 1]]
+  INSTANTIATION: dot_product
+BN:
+  USE_PRECISE_STATS: True
+  NUM_BATCHES_PRECISE: 200
+SOLVER:
+  BASE_LR: 0.8
+  LR_POLICY: cosine
+  MAX_EPOCH: 196
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  WARMUP_EPOCHS: 34.0
+  WARMUP_START_LR: 0.01
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  NUM_CLASSES: 400
+  ARCH: slowfast
+  MODEL_NAME: SlowFast
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/SLOWFAST_8x8_R101.yaml

@@ -0,0 +1,63 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+  EVAL_PERIOD: 10
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR:  /path/to/kinetics/
+  NUM_FRAMES: 32
+  SAMPLING_RATE: 2
+  TRAIN_JITTER_SCALES: [256, 340]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 256
+  INPUT_CHANNEL_NUM: [3, 3]
+SLOWFAST:
+  ALPHA: 4
+  BETA_INV: 8
+  FUSION_CONV_CHANNEL_RATIO: 2
+  FUSION_KERNEL_SZ: 5
+RESNET:
+  ZERO_INIT_FINAL_BN: True
+  WIDTH_PER_GROUP: 64
+  NUM_GROUPS: 1
+  DEPTH: 101
+  TRANS_FUNC: bottleneck_transform
+  STRIDE_1X1: False
+  NUM_BLOCK_TEMP_KERNEL: [[3, 3], [4, 4], [6, 6], [3, 3]]
+  SPATIAL_STRIDES: [[1, 1], [2, 2], [2, 2], [2, 2]]
+  SPATIAL_DILATIONS: [[1, 1], [1, 1], [1, 1], [1, 1]]
+NONLOCAL:
+  LOCATION: [[[], []], [[], []], [[], []], [[], []]]
+  GROUP: [[1, 1], [1, 1], [1, 1], [1, 1]]
+  INSTANTIATION: dot_product
+BN:
+  USE_PRECISE_STATS: True
+  NUM_BATCHES_PRECISE: 200
+SOLVER:
+  BASE_LR: 0.8 ## 8 nodes
+  LR_POLICY: cosine
+  MAX_EPOCH: 196
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  WARMUP_EPOCHS: 34.0
+  WARMUP_START_LR: 0.01
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  NUM_CLASSES: 400
+  ARCH: slowfast
+  MODEL_NAME: SlowFast
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/SLOWFAST_8x8_R50.yaml

@@ -0,0 +1,63 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+  EVAL_PERIOD: 10
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 32
+  SAMPLING_RATE: 2
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 256
+  INPUT_CHANNEL_NUM: [3, 3]
+SLOWFAST:
+  ALPHA: 4
+  BETA_INV: 8
+  FUSION_CONV_CHANNEL_RATIO: 2
+  FUSION_KERNEL_SZ: 7
+RESNET:
+  ZERO_INIT_FINAL_BN: True
+  WIDTH_PER_GROUP: 64
+  NUM_GROUPS: 1
+  DEPTH: 50
+  TRANS_FUNC: bottleneck_transform
+  STRIDE_1X1: False
+  NUM_BLOCK_TEMP_KERNEL: [[3, 3], [4, 4], [6, 6], [3, 3]]
+  SPATIAL_STRIDES: [[1, 1], [2, 2], [2, 2], [2, 2]]
+  SPATIAL_DILATIONS: [[1, 1], [1, 1], [1, 1], [1, 1]]
+NONLOCAL:
+  LOCATION: [[[], []], [[], []], [[], []], [[], []]]
+  GROUP: [[1, 1], [1, 1], [1, 1], [1, 1]]
+  INSTANTIATION: dot_product
+BN:
+  USE_PRECISE_STATS: True
+  NUM_BATCHES_PRECISE: 200
+SOLVER:
+  BASE_LR: 0.8
+  LR_POLICY: cosine
+  MAX_EPOCH: 196
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  WARMUP_EPOCHS: 34.0
+  WARMUP_START_LR: 0.01
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  NUM_CLASSES: 400
+  ARCH: slowfast
+  MODEL_NAME: SlowFast
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 64
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_divST_16x16_448.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 16
+  SAMPLING_RATE: 16
+  TRAIN_JITTER_SCALES: [448, 512]
+  TRAIN_CROP_SIZE: 448
+  TEST_CROP_SIZE: 448
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_divST_8x32_224.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 8
+  SAMPLING_RATE: 32
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_divST_8x32_224_4gpus.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 4
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 8
+  SAMPLING_RATE: 32
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 4
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 4
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_divST_8x32_224_TEST.yaml

@@ -0,0 +1,46 @@
+TRAIN:
+  ENABLE: False
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 8
+  SAMPLING_RATE: 32
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+  CHECKPOINT_FILE_PATH: '/checkpoint/gedas/jobs/timesformer/kinetics_400/TimeSformer_divST_8x32_224/checkpoints/checkpoint_epoch_00025.pyth'
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_divST_96x4_224.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 96
+  SAMPLING_RATE: 4
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_jointST_8x32_224.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 8
+  SAMPLING_RATE: 32
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'joint_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/Kinetics/TimeSformer_spaceOnly_8x32_224.yaml

@@ -0,0 +1,45 @@
+TRAIN:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: /path/to/kinetics/
+  NUM_FRAMES: 8
+  SAMPLING_RATE: 32
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+TIMESFORMER:
+  ATTENTION_TYPE: 'space_only'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 400
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: kinetics
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 8
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/SSv2/SLOWFAST_16x8_R50.yaml

@@ -0,0 +1,83 @@
+TRAIN:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 16
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: " /path/to/ssv2/annotations/"
+  PATH_PREFIX: "/path/to/ssv2/frames/"
+  NUM_FRAMES: 64
+  SAMPLING_RATE: 2
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 256
+  INPUT_CHANNEL_NUM: [3, 3]
+  INV_UNIFORM_SAMPLE: True
+  RANDOM_FLIP: False
+  REVERSE_INPUT_CHANNEL: True
+SLOWFAST:
+  ALPHA: 4
+  BETA_INV: 8
+  FUSION_CONV_CHANNEL_RATIO: 2
+  FUSION_KERNEL_SZ: 7
+RESNET:
+  SPATIAL_STRIDES: [[1, 1], [2, 2], [2, 2], [2, 2]]
+  SPATIAL_DILATIONS: [[1, 1], [1, 1], [1, 1], [1, 1]]
+  ZERO_INIT_FINAL_BN: True
+  WIDTH_PER_GROUP: 64
+  NUM_GROUPS: 1
+  DEPTH: 50
+  TRANS_FUNC: bottleneck_transform
+  STRIDE_1X1: False
+  NUM_BLOCK_TEMP_KERNEL: [[3, 3], [4, 4], [6, 6], [3, 3]]
+NONLOCAL:
+  LOCATION: [[[], []], [[], []], [[], []], [[], []]]
+  GROUP: [[1, 1], [1, 1], [1, 1], [1, 1]]
+  INSTANTIATION: dot_product
+BN:
+  USE_PRECISE_STATS: True
+  NUM_BATCHES_PRECISE: 200
+  NORM_TYPE: sync_batchnorm
+  NUM_SYNC_DEVICES: 4
+SOLVER:
+  BASE_LR: 0.2 #8 nodes
+  LR_POLICY: cosine
+  MAX_EPOCH: 200
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  WARMUP_EPOCHS: 34.0
+  WARMUP_START_LR: 0.01
+  OPTIMIZING_METHOD: sgd
+  #SOLVER:
+  #  BASE_LR: 0.03
+  #  LR_POLICY: steps_with_relative_lrs
+  #  LRS: [1, 0.1, 0.01, 0.001, 0.0001, 0.00001]
+  #  STEPS: [0, 14, 18]
+  #  MAX_EPOCH: 22
+  #  MOMENTUM: 0.9
+  #  WEIGHT_DECAY: 1e-6
+  #  WARMUP_EPOCHS: 0.19
+  #  WARMUP_START_LR: 0.0001
+  #  OPTIMIZING_METHOD: sgd
+MODEL:
+  NUM_CLASSES: 174
+  ARCH: slowfast
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 16
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 1
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .
+#LOG_MODEL_INFO: False
+LOG_MODEL_INFO: True

TimeSformer/configs/SSv2/TimeSformer_divST_16_448.yaml

@@ -0,0 +1,48 @@
+TRAIN:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: " /path/to/ssv2/annotations/"
+  PATH_PREFIX: "/path/to/ssv2/frames/"
+  NUM_FRAMES: 16
+  TRAIN_JITTER_SCALES: [448, 512]
+  TRAIN_CROP_SIZE: 448
+  TEST_CROP_SIZE: 448
+  INPUT_CHANNEL_NUM: [3]
+  INV_UNIFORM_SAMPLE: True
+  RANDOM_FLIP: False
+  REVERSE_INPUT_CHANNEL: True
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 174
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/SSv2/TimeSformer_divST_64_224.yaml

@@ -0,0 +1,48 @@
+TRAIN:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: " /path/to/ssv2/annotations/"
+  PATH_PREFIX: "/path/to/ssv2/frames/"
+  NUM_FRAMES: 64
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+  INV_UNIFORM_SAMPLE: True
+  RANDOM_FLIP: False
+  REVERSE_INPUT_CHANNEL: True
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 174
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/configs/SSv2/TimeSformer_divST_8_224.yaml

@@ -0,0 +1,48 @@
+TRAIN:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  EVAL_PERIOD: 5
+  CHECKPOINT_PERIOD: 5
+  AUTO_RESUME: True
+DATA:
+  PATH_TO_DATA_DIR: " /path/to/ssv2/annotations/"
+  PATH_PREFIX: "/path/to/ssv2/frames/"
+  NUM_FRAMES: 8
+  TRAIN_JITTER_SCALES: [256, 320]
+  TRAIN_CROP_SIZE: 224
+  TEST_CROP_SIZE: 224
+  INPUT_CHANNEL_NUM: [3]
+  INV_UNIFORM_SAMPLE: True
+  RANDOM_FLIP: False
+  REVERSE_INPUT_CHANNEL: True
+TIMESFORMER:
+  ATTENTION_TYPE: 'divided_space_time'
+SOLVER:
+  BASE_LR: 0.005
+  LR_POLICY: steps_with_relative_lrs
+  STEPS: [0, 11, 14]
+  LRS: [1, 0.1, 0.01]
+  MAX_EPOCH: 15
+  MOMENTUM: 0.9
+  WEIGHT_DECAY: 1e-4
+  OPTIMIZING_METHOD: sgd
+MODEL:
+  MODEL_NAME: vit_base_patch16_224
+  NUM_CLASSES: 174
+  ARCH: vit
+  LOSS_FUNC: cross_entropy
+  DROPOUT_RATE: 0.5
+TEST:
+  ENABLE: True
+  DATASET: ssv2
+  BATCH_SIZE: 8
+  NUM_ENSEMBLE_VIEWS: 1
+  NUM_SPATIAL_CROPS: 3
+DATA_LOADER:
+  NUM_WORKERS: 4
+  PIN_MEMORY: True
+NUM_GPUS: 8
+NUM_SHARDS: 1
+RNG_SEED: 0
+OUTPUT_DIR: .

TimeSformer/environment.yml

@@ -0,0 +1,26 @@
+name: timesformer
+channels:
+- pytorch 
+- conda-forge
+- defaults
+dependencies:
+- python>3.7
+- jupyterlab
+- pandas>=1.2
+- numpy>1.19
+- pytorch>=1.6
+- torchvision>=0.7
+- scikit-learn>=0.22
+- opencv>=4.2
+- pyyaml>=5.1
+- yacs>=0.1.6
+- einops>=0.3
+- tensorboard
+- psutil
+- tqdm
+- matplotlib
+- simplejson
+- pip
+- pip:
+  - fvcore
+  - av

TimeSformer/example.ipynb

@@ -0,0 +1,84 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "08fe0c59",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pathlib import Path\n",
+    "\n",
+    "import torch\n",
+    "from timesformer.models.vit import TimeSformer"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "10239d32",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "True"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "model_file = Path.home()/'TimeSformer/models/TimeSformer_divST_8x32_224_K600.pyth'\n",
+    "model_file.exists()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "652fb03e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model = TimeSformer(img_size=224, num_classes=600, num_frames=8, attention_type='divided_space_time',  pretrained_model=str(model_file))\n",
+    "\n",
+    "dummy_video = torch.randn(2, 3, 8, 224, 224) # (batch x channels x frames x height x width)\n",
+    "\n",
+    "pred = model(dummy_video,) # (2, 600)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "83de13c5-791c-4db7-aba4-6d29ce88584e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "assert pred.shape == (2,600)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.4"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

TimeSformer/setup.cfg

@@ -0,0 +1,23 @@
+[isort]
+line_length=100
+multi_line_output=4
+known_standard_library=numpy,setuptools
+known_myself=timesformer
+known_third_party=fvcore,av,torch,pycocotools,yacs,termcolor,scipy,simplejson,matplotlib,torchvision,yaml,tqdm,psutil,opencv-python,pandas,tensorboard,moviepy,sklearn,cv2
+no_lines_before=STDLIB,THIRDPARTY
+sections=FUTURE,STDLIB,THIRDPARTY,myself,FIRSTPARTY,LOCALFOLDER
+default_section=FIRSTPARTY
+
+[mypy]
+python_version=3.6
+ignore_missing_imports = True
+warn_unused_configs = True
+disallow_untyped_defs = True
+check_untyped_defs = True
+warn_unused_ignores = True
+warn_redundant_casts = True
+show_column_numbers = True
+follow_imports = silent
+allow_redefinition = True
+; Require all functions to be annotated
+disallow_incomplete_defs = True

TimeSformer/setup.py

@@ -0,0 +1,23 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from setuptools import find_packages, setup
+
+setup(
+    name="timesformer",
+    version="1.0",
+    author="FBAI",
+    url="unknown",
+    description="TimeSformer",
+    keywords = [
+    'artificial intelligence',
+    'attention mechanism',
+    'transformers',
+    'video classification',
+    ],
+    install_requires=[
+        'einops>=0.3',
+        'torch>=1.6'
+    ],
+    extras_require={"tensorboard_video_visualization": ["moviepy"]},
+    packages=find_packages(exclude=("configs", "tests")),
+)

TimeSformer/slurm_scripts/run_multi_node_job.sh

@@ -0,0 +1,25 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# A script with a list of commands for submitting SLURM jobs
+
+#### Kinetics training
+JOB_NAME=TimeSformer_divST_8x32_224
+python tools/submit.py --cfg configs/Kinetics/TimeSformer_divST_8x32_224.yaml --job_dir  /your/job/dir/${JOB_NAME}/   --num_shards 4 --partition dev --comment "" --name ${JOB_NAME} --use_volta32
+
+#JOB_NAME=TimeSformer_jointST_8x32_224
+#python tools/submit.py --cfg configs/Kinetics/TimeSformer_jointST_8x32_224.yaml --job_dir  /your/job/dir/${JOB_NAME}/   --num_shards 4 --partition learnfair --comment "" --name ${JOB_NAME} --use_volta32
+
+#JOB_NAME=TimeSformer_spaceOnly_8x32_224
+#python tools/submit.py --cfg configs/Kinetics/TimeSformer_spaceOnly_8x32_224.yaml --job_dir  /your/job/dir/${JOB_NAME}/   --num_shards 4 --partition learnfair --comment "" --name ${JOB_NAME} --use_volta32
+
+#### Kinetics inference
+#JOB_NAME=TimeSformer_divST_8x32_224_TEST_3clips
+#python tools/submit.py --cfg configs/Kinetics/TimeSformer_divST_8x32_224_TEST.yaml --job_dir /your/job/dir/${JOB_NAME}/  --num_shards 4 --partition dev --comment "" --name ${JOB_NAME} --use_volta32
+
+
+##### SSv2 training
+#JOB_NAME=TimeSformer_divST_8_224
+#python tools/submit.py --cfg configs/SSv2/TimeSformer_divST_8_224.yaml --job_dir  /your/job/dir/${JOB_NAME}/   --num_shards 4 --partition learnfair --comment "" --name ${JOB_NAME} --use_volta32
+
+##### Sth-Sth_v2 inference
+#JOB_NAME=TimeSformer_divST_8_224_TEST_3clips
+#python tools/submit.py --cfg configs/SSv2/TimeSformer_divST_8_224_TEST.yaml --job_dir  /your/job/dir/${JOB_NAME}/   --num_shards 4 --partition learnfair --comment "" --name ${JOB_NAME} --use_volta32

TimeSformer/slurm_scripts/run_single_node_job.sh

@@ -0,0 +1,35 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+# A script with a list of commands for submitting SLURM jobs
+
+#SBATCH --job-name=timesformer
+#SBATCH --mail-type=END,FAIL,REQUEUE
+#SBATCH --mail-user=name@domain.com
+
+## %j is the job id, %u is the user id
+#SBATCH --output=/path/to/output/logs/slog-%A-%a.out
+
+## filename for job standard error output (stderr)
+#SBATCH --error=/path/to/error/logs/slog-%A-%a.err
+
+#SBATCH --array=1
+#SBATCH --partition=partition_of_your_choice
+#SBATCH --nodes=1 -C volta32gb
+#SBATCH --ntasks-per-node=1
+#SBATCH --gpus-per-node=8
+#SBATCH --cpus-per-task=80
+#SBATCH --mem=480GB
+#SBATCH --signal=USR1@600
+#SBATCH --time=72:00:00
+#SBATCH --open-mode=append
+
+module purge
+module load cuda/10.0
+module load NCCL/2.4.7-1-cuda.10.0
+module load cudnn/v7.4-cuda.10.0
+source activate timesformer
+
+WORKINGDIR=/path/to/TimeSformer
+CURPYTHON=/path/to/python
+
+srun --label ${CURPYTHON} ${WORKINGDIR}/tools/run_net.py --cfg ${WORKINGDIR}/configs/Kinetics/TimeSformer_divST_8x32_224.yaml NUM_GPUS 8 TRAIN.BATCH_SIZE 8
+

TimeSformer/timesformer/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from timesformer.utils.env import setup_environment
+
+setup_environment()

TimeSformer/timesformer/config/__init__.py

@@ -0,0 +1 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.

TimeSformer/timesformer/config/defaults.py

@@ -0,0 +1,820 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""Configs."""
+from fvcore.common.config import CfgNode
+# -----------------------------------------------------------------------------
+# Config definition
+# -----------------------------------------------------------------------------
+_C = CfgNode()
+
+# ---------------------------------------------------------------------------- #
+# Batch norm options
+# ---------------------------------------------------------------------------- #
+_C.BN = CfgNode()
+
+# Precise BN stats.
+_C.BN.USE_PRECISE_STATS = False
+
+# Number of samples use to compute precise bn.
+_C.BN.NUM_BATCHES_PRECISE = 200
+
+# Weight decay value that applies on BN.
+_C.BN.WEIGHT_DECAY = 0.0
+
+# Norm type, options include `batchnorm`, `sub_batchnorm`, `sync_batchnorm`
+_C.BN.NORM_TYPE = "batchnorm"
+
+# Parameter for SubBatchNorm, where it splits the batch dimension into
+# NUM_SPLITS splits, and run BN on each of them separately independently.
+_C.BN.NUM_SPLITS = 1
+
+# Parameter for NaiveSyncBatchNorm3d, where the stats across `NUM_SYNC_DEVICES`
+# devices will be synchronized.
+_C.BN.NUM_SYNC_DEVICES = 1
+
+
+# ---------------------------------------------------------------------------- #
+# Training options.
+# ---------------------------------------------------------------------------- #
+_C.TRAIN = CfgNode()
+
+# If True Train the model, else skip training.
+_C.TRAIN.ENABLE = True
+
+# Dataset.
+_C.TRAIN.DATASET = "kinetics"
+
+##
+_C.TRAIN.FINETUNE = False
+
+# Total mini-batch size.
+_C.TRAIN.BATCH_SIZE = 64
+
+# Evaluate model on test data every eval period epochs.
+_C.TRAIN.EVAL_PERIOD = 10
+
+# Save model checkpoint every checkpoint period epochs.
+_C.TRAIN.CHECKPOINT_PERIOD = 10
+
+# Resume training from the latest checkpoint in the output directory.
+_C.TRAIN.AUTO_RESUME = True
+
+# Path to the checkpoint to load the initial weight.
+_C.TRAIN.CHECKPOINT_FILE_PATH = ""
+
+# Checkpoint types include `caffe2` or `pytorch`.
+_C.TRAIN.CHECKPOINT_TYPE = "pytorch"
+
+# If True, perform inflation when loading checkpoint.
+_C.TRAIN.CHECKPOINT_INFLATE = False
+
+# If True, reset epochs when loading checkpoint.
+_C.TRAIN.CHECKPOINT_EPOCH_RESET = False
+
+# If set, clear all layer names according to the pattern provided.
+_C.TRAIN.CHECKPOINT_CLEAR_NAME_PATTERN = ()  # ("backbone.",)
+
+# ---------------------------------------------------------------------------- #
+# Testing options
+# ---------------------------------------------------------------------------- #
+_C.TEST = CfgNode()
+
+# If True test the model, else skip the testing.
+_C.TEST.ENABLE = True
+
+# Dataset for testing.
+_C.TEST.DATASET = "kinetics"
+
+# Total mini-batch size
+_C.TEST.BATCH_SIZE = 8
+
+# Path to the checkpoint to load the initial weight.
+_C.TEST.CHECKPOINT_FILE_PATH = ""
+
+# Number of clips to sample from a video uniformly for aggregating the
+# prediction results.
+_C.TEST.NUM_ENSEMBLE_VIEWS = 10
+
+# Number of crops to sample from a frame spatially for aggregating the
+# prediction results.
+_C.TEST.NUM_SPATIAL_CROPS = 3
+
+# Checkpoint types include `caffe2` or `pytorch`.
+_C.TEST.CHECKPOINT_TYPE = "pytorch"
+# Path to saving prediction results file.
+_C.TEST.SAVE_RESULTS_PATH = ""
+# -----------------------------------------------------------------------------
+# ResNet options
+# -----------------------------------------------------------------------------
+_C.RESNET = CfgNode()
+
+# Transformation function.
+_C.RESNET.TRANS_FUNC = "bottleneck_transform"
+
+# Number of groups. 1 for ResNet, and larger than 1 for ResNeXt).
+_C.RESNET.NUM_GROUPS = 1
+
+# Width of each group (64 -> ResNet; 4 -> ResNeXt).
+_C.RESNET.WIDTH_PER_GROUP = 64
+
+# Apply relu in a inplace manner.
+_C.RESNET.INPLACE_RELU = True
+
+# Apply stride to 1x1 conv.
+_C.RESNET.STRIDE_1X1 = False
+
+#  If true, initialize the gamma of the final BN of each block to zero.
+_C.RESNET.ZERO_INIT_FINAL_BN = False
+
+# Number of weight layers.
+_C.RESNET.DEPTH = 50
+
+# If the current block has more than NUM_BLOCK_TEMP_KERNEL blocks, use temporal
+# kernel of 1 for the rest of the blocks.
+_C.RESNET.NUM_BLOCK_TEMP_KERNEL = [[3], [4], [6], [3]]
+
+# Size of stride on different res stages.
+_C.RESNET.SPATIAL_STRIDES = [[1], [2], [2], [2]]
+
+# Size of dilation on different res stages.
+_C.RESNET.SPATIAL_DILATIONS = [[1], [1], [1], [1]]
+
+# ---------------------------------------------------------------------------- #
+# X3D  options
+# See https://arxiv.org/abs/2004.04730 for details about X3D Networks.
+# ---------------------------------------------------------------------------- #
+_C.X3D = CfgNode()
+
+# Width expansion factor.
+_C.X3D.WIDTH_FACTOR = 1.0
+
+# Depth expansion factor.
+_C.X3D.DEPTH_FACTOR = 1.0
+
+# Bottleneck expansion factor for the 3x3x3 conv.
+_C.X3D.BOTTLENECK_FACTOR = 1.0  #
+
+# Dimensions of the last linear layer before classificaiton.
+_C.X3D.DIM_C5 = 2048
+
+# Dimensions of the first 3x3 conv layer.
+_C.X3D.DIM_C1 = 12
+
+# Whether to scale the width of Res2, default is false.
+_C.X3D.SCALE_RES2 = False
+
+# Whether to use a BatchNorm (BN) layer before the classifier, default is false.
+_C.X3D.BN_LIN5 = False
+
+# Whether to use channelwise (=depthwise) convolution in the center (3x3x3)
+# convolution operation of the residual blocks.
+_C.X3D.CHANNELWISE_3x3x3 = True
+
+# -----------------------------------------------------------------------------
+# Nonlocal options
+# -----------------------------------------------------------------------------
+_C.NONLOCAL = CfgNode()
+
+# Index of each stage and block to add nonlocal layers.
+_C.NONLOCAL.LOCATION = [[[]], [[]], [[]], [[]]]
+
+# Number of group for nonlocal for each stage.
+_C.NONLOCAL.GROUP = [[1], [1], [1], [1]]
+
+# Instatiation to use for non-local layer.
+_C.NONLOCAL.INSTANTIATION = "dot_product"
+
+
+# Size of pooling layers used in Non-Local.
+_C.NONLOCAL.POOL = [
+    # Res2
+    [[1, 2, 2], [1, 2, 2]],
+    # Res3
+    [[1, 2, 2], [1, 2, 2]],
+    # Res4
+    [[1, 2, 2], [1, 2, 2]],
+    # Res5
+    [[1, 2, 2], [1, 2, 2]],
+]
+
+# -----------------------------------------------------------------------------
+# Model options
+# -----------------------------------------------------------------------------
+_C.MODEL = CfgNode()
+
+# Model architecture.
+_C.MODEL.ARCH = "slowfast"
+
+# Model name
+_C.MODEL.MODEL_NAME = "SlowFast"
+
+# The number of classes to predict for the model.
+_C.MODEL.NUM_CLASSES = 400
+
+# Loss function.
+_C.MODEL.LOSS_FUNC = "cross_entropy"
+
+# Model architectures that has one single pathway.
+_C.MODEL.SINGLE_PATHWAY_ARCH = ["c2d", "i3d", "slow", "x3d"]
+
+# Model architectures that has multiple pathways.
+_C.MODEL.MULTI_PATHWAY_ARCH = ["slowfast"]
+
+# Dropout rate before final projection in the backbone.
+_C.MODEL.DROPOUT_RATE = 0.5
+
+# Randomly drop rate for Res-blocks, linearly increase from res2 to res5
+_C.MODEL.DROPCONNECT_RATE = 0.0
+
+# The std to initialize the fc layer(s).
+_C.MODEL.FC_INIT_STD = 0.01
+
+# Activation layer for the output head.
+_C.MODEL.HEAD_ACT = "softmax"
+
+
+# -----------------------------------------------------------------------------
+# SlowFast options
+# -----------------------------------------------------------------------------
+_C.SLOWFAST = CfgNode()
+
+# Corresponds to the inverse of the channel reduction ratio, $\beta$ between
+# the Slow and Fast pathways.
+_C.SLOWFAST.BETA_INV = 8
+
+# Corresponds to the frame rate reduction ratio, $\alpha$ between the Slow and
+# Fast pathways.
+_C.SLOWFAST.ALPHA = 8
+
+# Ratio of channel dimensions between the Slow and Fast pathways.
+_C.SLOWFAST.FUSION_CONV_CHANNEL_RATIO = 2
+
+# Kernel dimension used for fusing information from Fast pathway to Slow
+# pathway.
+_C.SLOWFAST.FUSION_KERNEL_SZ = 5
+
+####### TimeSformer Options
+_C.TIMESFORMER = CfgNode()
+_C.TIMESFORMER.ATTENTION_TYPE = 'divided_space_time'
+_C.TIMESFORMER.PRETRAINED_MODEL = ''
+
+## MixUp parameters
+_C.MIXUP = CfgNode()
+_C.MIXUP.ENABLED = False
+_C.MIXUP.ALPHA = 0.8
+_C.MIXUP.CUTMIX_ALPHA = 1.0
+_C.MIXUP.CUTMIX_MINMAX = None
+_C.MIXUP.PROB = 1.0
+_C.MIXUP.SWITCH_PROB = 0.5
+_C.MIXUP.MODE = 'batch'
+
+_C.EMA = CfgNode()
+_C.EMA.ENABLED = False
+
+# -----------------------------------------------------------------------------
+# Data options
+# -----------------------------------------------------------------------------
+_C.DATA = CfgNode()
+
+# The path to the data directory.
+_C.DATA.PATH_TO_DATA_DIR = ""
+
+# The separator used between path and label.
+_C.DATA.PATH_LABEL_SEPARATOR = " "
+
+# Video path prefix if any.
+_C.DATA.PATH_PREFIX = ""
+
+# The spatial crop size of the input clip.
+_C.DATA.CROP_SIZE = 224
+
+# The number of frames of the input clip.
+_C.DATA.NUM_FRAMES = 8
+
+# The video sampling rate of the input clip.
+_C.DATA.SAMPLING_RATE = 8
+
+# The mean value of the video raw pixels across the R G B channels.
+_C.DATA.MEAN = [0.45, 0.45, 0.45]
+# List of input frame channel dimensions.
+
+_C.DATA.INPUT_CHANNEL_NUM = [3, 3]
+
+# The std value of the video raw pixels across the R G B channels.
+_C.DATA.STD = [0.225, 0.225, 0.225]
+
+# The spatial augmentation jitter scales for training.
+_C.DATA.TRAIN_JITTER_SCALES = [256, 320]
+
+# The spatial crop size for training.
+_C.DATA.TRAIN_CROP_SIZE = 224
+
+# The spatial crop size for testing.
+_C.DATA.TEST_CROP_SIZE = 256
+
+# Input videos may has different fps, convert it to the target video fps before
+# frame sampling.
+_C.DATA.TARGET_FPS = 30
+
+# Decoding backend, options include `pyav` or `torchvision`
+_C.DATA.DECODING_BACKEND = "pyav"
+
+# if True, sample uniformly in [1 / max_scale, 1 / min_scale] and take a
+# reciprocal to get the scale. If False, take a uniform sample from
+# [min_scale, max_scale].
+_C.DATA.INV_UNIFORM_SAMPLE = False
+
+# If True, perform random horizontal flip on the video frames during training.
+_C.DATA.RANDOM_FLIP = True
+
+# If True, calculdate the map as metric.
+_C.DATA.MULTI_LABEL = False
+
+# Method to perform the ensemble, options include "sum" and "max".
+_C.DATA.ENSEMBLE_METHOD = "sum"
+
+# If True, revert the default input channel (RBG <-> BGR).
+_C.DATA.REVERSE_INPUT_CHANNEL = False
+
+############
+_C.DATA.TEMPORAL_EXTENT = 8
+_C.DATA.DEIT_TRANSFORMS = False
+_C.DATA.COLOR_JITTER = 0.
+_C.DATA.AUTO_AUGMENT = ''
+_C.DATA.RE_PROB = 0.0
+
+# ---------------------------------------------------------------------------- #
+# Optimizer options
+# ---------------------------------------------------------------------------- #
+_C.SOLVER = CfgNode()
+
+# Base learning rate.
+_C.SOLVER.BASE_LR = 0.1
+
+# Learning rate policy (see utils/lr_policy.py for options and examples).
+_C.SOLVER.LR_POLICY = "cosine"
+
+# Final learning rates for 'cosine' policy.
+_C.SOLVER.COSINE_END_LR = 0.0
+
+# Exponential decay factor.
+_C.SOLVER.GAMMA = 0.1
+
+# Step size for 'exp' and 'cos' policies (in epochs).
+_C.SOLVER.STEP_SIZE = 1
+
+# Steps for 'steps_' policies (in epochs).
+_C.SOLVER.STEPS = []
+
+# Learning rates for 'steps_' policies.
+_C.SOLVER.LRS = []
+
+# Maximal number of epochs.
+_C.SOLVER.MAX_EPOCH = 300
+
+# Momentum.
+_C.SOLVER.MOMENTUM = 0.9
+
+# Momentum dampening.
+_C.SOLVER.DAMPENING = 0.0
+
+# Nesterov momentum.
+_C.SOLVER.NESTEROV = True
+
+# L2 regularization.
+_C.SOLVER.WEIGHT_DECAY = 1e-4
+
+# Start the warm up from SOLVER.BASE_LR * SOLVER.WARMUP_FACTOR.
+_C.SOLVER.WARMUP_FACTOR = 0.1
+
+# Gradually warm up the SOLVER.BASE_LR over this number of epochs.
+_C.SOLVER.WARMUP_EPOCHS = 0.0
+
+# The start learning rate of the warm up.
+_C.SOLVER.WARMUP_START_LR = 0.01
+
+# Optimization method.
+_C.SOLVER.OPTIMIZING_METHOD = "sgd"
+
+# Base learning rate is linearly scaled with NUM_SHARDS.
+_C.SOLVER.BASE_LR_SCALE_NUM_SHARDS = False
+
+# ---------------------------------------------------------------------------- #
+# Misc options
+# ---------------------------------------------------------------------------- #
+
+# Number of GPUs to use (applies to both training and testing).
+_C.NUM_GPUS = 1
+
+# Number of machine to use for the job.
+_C.NUM_SHARDS = 1
+
+# The index of the current machine.
+_C.SHARD_ID = 0
+
+# Output basedir.
+_C.OUTPUT_DIR = "./tmp"
+
+# Note that non-determinism may still be present due to non-deterministic
+# operator implementations in GPU operator libraries.
+_C.RNG_SEED = 1
+
+# Log period in iters.
+_C.LOG_PERIOD = 10
+
+# If True, log the model info.
+_C.LOG_MODEL_INFO = False
+
+# Distributed backend.
+_C.DIST_BACKEND = "nccl"
+
+# Global batch size
+_C.GLOBAL_BATCH_SIZE = 64
+
+# ---------------------------------------------------------------------------- #
+# Benchmark options
+# ---------------------------------------------------------------------------- #
+_C.BENCHMARK = CfgNode()
+
+# Number of epochs for data loading benchmark.
+_C.BENCHMARK.NUM_EPOCHS = 5
+
+# Log period in iters for data loading benchmark.
+_C.BENCHMARK.LOG_PERIOD = 100
+
+# If True, shuffle dataloader for epoch during benchmark.
+_C.BENCHMARK.SHUFFLE = True
+
+
+# ---------------------------------------------------------------------------- #
+# Common train/test data loader options
+# ---------------------------------------------------------------------------- #
+_C.DATA_LOADER = CfgNode()
+
+# Number of data loader workers per training process.
+_C.DATA_LOADER.NUM_WORKERS = 8
+
+# Load data to pinned host memory.
+_C.DATA_LOADER.PIN_MEMORY = True
+
+# Enable multi thread decoding.
+_C.DATA_LOADER.ENABLE_MULTI_THREAD_DECODE = False
+
+
+# ---------------------------------------------------------------------------- #
+# Detection options.
+# ---------------------------------------------------------------------------- #
+_C.DETECTION = CfgNode()
+
+# Whether enable video detection.
+_C.DETECTION.ENABLE = False
+
+# Aligned version of RoI. More details can be found at slowfast/models/head_helper.py
+_C.DETECTION.ALIGNED = True
+
+# Spatial scale factor.
+_C.DETECTION.SPATIAL_SCALE_FACTOR = 16
+
+# RoI tranformation resolution.
+_C.DETECTION.ROI_XFORM_RESOLUTION = 7
+
+
+# -----------------------------------------------------------------------------
+# AVA Dataset options
+# -----------------------------------------------------------------------------
+_C.AVA = CfgNode()
+
+# Directory path of frames.
+_C.AVA.FRAME_DIR = "/mnt/fair-flash3-east/ava_trainval_frames.img/"
+
+# Directory path for files of frame lists.
+_C.AVA.FRAME_LIST_DIR = (
+    "/mnt/vol/gfsai-flash3-east/ai-group/users/haoqifan/ava/frame_list/"
+)
+
+# Directory path for annotation files.
+_C.AVA.ANNOTATION_DIR = (
+    "/mnt/vol/gfsai-flash3-east/ai-group/users/haoqifan/ava/frame_list/"
+)
+
+# Filenames of training samples list files.
+_C.AVA.TRAIN_LISTS = ["train.csv"]
+
+# Filenames of test samples list files.
+_C.AVA.TEST_LISTS = ["val.csv"]
+
+# Filenames of box list files for training. Note that we assume files which
+# contains predicted boxes will have a suffix "predicted_boxes" in the
+# filename.
+_C.AVA.TRAIN_GT_BOX_LISTS = ["ava_train_v2.2.csv"]
+_C.AVA.TRAIN_PREDICT_BOX_LISTS = []
+
+# Filenames of box list files for test.
+_C.AVA.TEST_PREDICT_BOX_LISTS = ["ava_val_predicted_boxes.csv"]
+
+# This option controls the score threshold for the predicted boxes to use.
+_C.AVA.DETECTION_SCORE_THRESH = 0.9
+
+# If use BGR as the format of input frames.
+_C.AVA.BGR = False
+
+# Training augmentation parameters
+# Whether to use color augmentation method.
+_C.AVA.TRAIN_USE_COLOR_AUGMENTATION = False
+
+# Whether to only use PCA jitter augmentation when using color augmentation
+# method (otherwise combine with color jitter method).
+_C.AVA.TRAIN_PCA_JITTER_ONLY = True
+
+# Eigenvalues for PCA jittering. Note PCA is RGB based.
+_C.AVA.TRAIN_PCA_EIGVAL = [0.225, 0.224, 0.229]
+
+# Eigenvectors for PCA jittering.
+_C.AVA.TRAIN_PCA_EIGVEC = [
+    [-0.5675, 0.7192, 0.4009],
+    [-0.5808, -0.0045, -0.8140],
+    [-0.5836, -0.6948, 0.4203],
+]
+
+# Whether to do horizontal flipping during test.
+_C.AVA.TEST_FORCE_FLIP = False
+
+# Whether to use full test set for validation split.
+_C.AVA.FULL_TEST_ON_VAL = False
+
+# The name of the file to the ava label map.
+_C.AVA.LABEL_MAP_FILE = "ava_action_list_v2.2_for_activitynet_2019.pbtxt"
+
+# The name of the file to the ava exclusion.
+_C.AVA.EXCLUSION_FILE = "ava_val_excluded_timestamps_v2.2.csv"
+
+# The name of the file to the ava groundtruth.
+_C.AVA.GROUNDTRUTH_FILE = "ava_val_v2.2.csv"
+
+# Backend to process image, includes `pytorch` and `cv2`.
+_C.AVA.IMG_PROC_BACKEND = "cv2"
+
+# ---------------------------------------------------------------------------- #
+# Multigrid training options
+# See https://arxiv.org/abs/1912.00998 for details about multigrid training.
+# ---------------------------------------------------------------------------- #
+_C.MULTIGRID = CfgNode()
+
+# Multigrid training allows us to train for more epochs with fewer iterations.
+# This hyperparameter specifies how many times more epochs to train.
+# The default setting in paper trains for 1.5x more epochs than baseline.
+_C.MULTIGRID.EPOCH_FACTOR = 1.5
+
+# Enable short cycles.
+_C.MULTIGRID.SHORT_CYCLE = False
+# Short cycle additional spatial dimensions relative to the default crop size.
+_C.MULTIGRID.SHORT_CYCLE_FACTORS = [0.5, 0.5 ** 0.5]
+
+_C.MULTIGRID.LONG_CYCLE = False
+# (Temporal, Spatial) dimensions relative to the default shape.
+_C.MULTIGRID.LONG_CYCLE_FACTORS = [
+    (0.25, 0.5 ** 0.5),
+    (0.5, 0.5 ** 0.5),
+    (0.5, 1),
+    (1, 1),
+]
+
+# While a standard BN computes stats across all examples in a GPU,
+# for multigrid training we fix the number of clips to compute BN stats on.
+# See https://arxiv.org/abs/1912.00998 for details.
+_C.MULTIGRID.BN_BASE_SIZE = 8
+
+# Multigrid training epochs are not proportional to actual training time or
+# computations, so _C.TRAIN.EVAL_PERIOD leads to too frequent or rare
+# evaluation. We use a multigrid-specific rule to determine when to evaluate:
+# This hyperparameter defines how many times to evaluate a model per long
+# cycle shape.
+_C.MULTIGRID.EVAL_FREQ = 3
+
+# No need to specify; Set automatically and used as global variables.
+_C.MULTIGRID.LONG_CYCLE_SAMPLING_RATE = 0
+_C.MULTIGRID.DEFAULT_B = 0
+_C.MULTIGRID.DEFAULT_T = 0
+_C.MULTIGRID.DEFAULT_S = 0
+
+# -----------------------------------------------------------------------------
+# Tensorboard Visualization Options
+# -----------------------------------------------------------------------------
+_C.TENSORBOARD = CfgNode()
+
+# Log to summary writer, this will automatically.
+# log loss, lr and metrics during train/eval.
+_C.TENSORBOARD.ENABLE = False
+# Provide path to prediction results for visualization.
+# This is a pickle file of [prediction_tensor, label_tensor]
+_C.TENSORBOARD.PREDICTIONS_PATH = ""
+# Path to directory for tensorboard logs.
+# Default to to cfg.OUTPUT_DIR/runs-{cfg.TRAIN.DATASET}.
+_C.TENSORBOARD.LOG_DIR = ""
+# Path to a json file providing class_name - id mapping
+# in the format {"class_name1": id1, "class_name2": id2, ...}.
+# This file must be provided to enable plotting confusion matrix
+# by a subset or parent categories.
+_C.TENSORBOARD.CLASS_NAMES_PATH = ""
+
+# Path to a json file for categories -> classes mapping
+# in the format {"parent_class": ["child_class1", "child_class2",...], ...}.
+_C.TENSORBOARD.CATEGORIES_PATH = ""
+
+# Config for confusion matrices visualization.
+_C.TENSORBOARD.CONFUSION_MATRIX = CfgNode()
+# Visualize confusion matrix.
+_C.TENSORBOARD.CONFUSION_MATRIX.ENABLE = False
+# Figure size of the confusion matrices plotted.
+_C.TENSORBOARD.CONFUSION_MATRIX.FIGSIZE = [8, 8]
+# Path to a subset of categories to visualize.
+# File contains class names separated by newline characters.
+_C.TENSORBOARD.CONFUSION_MATRIX.SUBSET_PATH = ""
+
+# Config for histogram visualization.
+_C.TENSORBOARD.HISTOGRAM = CfgNode()
+# Visualize histograms.
+_C.TENSORBOARD.HISTOGRAM.ENABLE = False
+# Path to a subset of classes to plot histograms.
+# Class names must be separated by newline characters.
+_C.TENSORBOARD.HISTOGRAM.SUBSET_PATH = ""
+# Visualize top-k most predicted classes on histograms for each
+# chosen true label.
+_C.TENSORBOARD.HISTOGRAM.TOPK = 10
+# Figure size of the histograms plotted.
+_C.TENSORBOARD.HISTOGRAM.FIGSIZE = [8, 8]
+
+# Config for layers' weights and activations visualization.
+# _C.TENSORBOARD.ENABLE must be True.
+_C.TENSORBOARD.MODEL_VIS = CfgNode()
+
+# If False, skip model visualization.
+_C.TENSORBOARD.MODEL_VIS.ENABLE = False
+
+# If False, skip visualizing model weights.
+_C.TENSORBOARD.MODEL_VIS.MODEL_WEIGHTS = False
+
+# If False, skip visualizing model activations.
+_C.TENSORBOARD.MODEL_VIS.ACTIVATIONS = False
+
+# If False, skip visualizing input videos.
+_C.TENSORBOARD.MODEL_VIS.INPUT_VIDEO = False
+
+
+# List of strings containing data about layer names and their indexing to
+# visualize weights and activations for. The indexing is meant for
+# choosing a subset of activations outputed by a layer for visualization.
+# If indexing is not specified, visualize all activations outputed by the layer.
+# For each string, layer name and indexing is separated by whitespaces.
+# e.g.: [layer1 1,2;1,2, layer2, layer3 150,151;3,4]; this means for each array `arr`
+# along the batch dimension in `layer1`, we take arr[[1, 2], [1, 2]]
+_C.TENSORBOARD.MODEL_VIS.LAYER_LIST = []
+# Top-k predictions to plot on videos
+_C.TENSORBOARD.MODEL_VIS.TOPK_PREDS = 1
+# Colormap to for text boxes and bounding boxes colors
+_C.TENSORBOARD.MODEL_VIS.COLORMAP = "Pastel2"
+# Config for visualization video inputs with Grad-CAM.
+# _C.TENSORBOARD.ENABLE must be True.
+_C.TENSORBOARD.MODEL_VIS.GRAD_CAM = CfgNode()
+# Whether to run visualization using Grad-CAM technique.
+_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.ENABLE = True
+# CNN layers to use for Grad-CAM. The number of layers must be equal to
+# number of pathway(s).
+_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.LAYER_LIST = []
+# If True, visualize Grad-CAM using true labels for each instances.
+# If False, use the highest predicted class.
+_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.USE_TRUE_LABEL = False
+# Colormap to for text boxes and bounding boxes colors
+_C.TENSORBOARD.MODEL_VIS.GRAD_CAM.COLORMAP = "viridis"
+
+# Config for visualization for wrong prediction visualization.
+# _C.TENSORBOARD.ENABLE must be True.
+_C.TENSORBOARD.WRONG_PRED_VIS = CfgNode()
+_C.TENSORBOARD.WRONG_PRED_VIS.ENABLE = False
+# Folder tag to origanize model eval videos under.
+_C.TENSORBOARD.WRONG_PRED_VIS.TAG = "Incorrectly classified videos."
+# Subset of labels to visualize. Only wrong predictions with true labels
+# within this subset is visualized.
+_C.TENSORBOARD.WRONG_PRED_VIS.SUBSET_PATH = ""
+
+
+# ---------------------------------------------------------------------------- #
+# Demo options
+# ---------------------------------------------------------------------------- #
+_C.DEMO = CfgNode()
+
+# Run model in DEMO mode.
+_C.DEMO.ENABLE = False
+
+# Path to a json file providing class_name - id mapping
+# in the format {"class_name1": id1, "class_name2": id2, ...}.
+_C.DEMO.LABEL_FILE_PATH = ""
+
+# Specify a camera device as input. This will be prioritized
+# over input video if set.
+# If -1, use input video instead.
+_C.DEMO.WEBCAM = -1
+
+# Path to input video for demo.
+_C.DEMO.INPUT_VIDEO = ""
+# Custom width for reading input video data.
+_C.DEMO.DISPLAY_WIDTH = 0
+# Custom height for reading input video data.
+_C.DEMO.DISPLAY_HEIGHT = 0
+# Path to Detectron2 object detection model configuration,
+# only used for detection tasks.
+_C.DEMO.DETECTRON2_CFG = "COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"
+# Path to Detectron2 object detection model pre-trained weights.
+_C.DEMO.DETECTRON2_WEIGHTS = "detectron2://COCO-Detection/faster_rcnn_R_50_FPN_3x/137849458/model_final_280758.pkl"
+# Threshold for choosing predicted bounding boxes by Detectron2.
+_C.DEMO.DETECTRON2_THRESH = 0.9
+# Number of overlapping frames between 2 consecutive clips.
+# Increase this number for more frequent action predictions.
+# The number of overlapping frames cannot be larger than
+# half of the sequence length `cfg.DATA.NUM_FRAMES * cfg.DATA.SAMPLING_RATE`
+_C.DEMO.BUFFER_SIZE = 0
+# If specified, the visualized outputs will be written this a video file of
+# this path. Otherwise, the visualized outputs will be displayed in a window.
+_C.DEMO.OUTPUT_FILE = ""
+# Frames per second rate for writing to output video file.
+# If not set (-1), use fps rate from input file.
+_C.DEMO.OUTPUT_FPS = -1
+# Input format from demo video reader ("RGB" or "BGR").
+_C.DEMO.INPUT_FORMAT = "BGR"
+# Draw visualization frames in [keyframe_idx - CLIP_VIS_SIZE, keyframe_idx + CLIP_VIS_SIZE] inclusively.
+_C.DEMO.CLIP_VIS_SIZE = 10
+# Number of processes to run video visualizer.
+_C.DEMO.NUM_VIS_INSTANCES = 2
+
+# Path to pre-computed predicted boxes
+_C.DEMO.PREDS_BOXES = ""
+# Whether to run in with multi-threaded video reader.
+_C.DEMO.THREAD_ENABLE = False
+# Take one clip for every `DEMO.NUM_CLIPS_SKIP` + 1 for prediction and visualization.
+# This is used for fast demo speed by reducing the prediction/visualiztion frequency.
+# If -1, take the most recent read clip for visualization. This mode is only supported
+# if `DEMO.THREAD_ENABLE` is set to True.
+_C.DEMO.NUM_CLIPS_SKIP = 0
+# Path to ground-truth boxes and labels (optional)
+_C.DEMO.GT_BOXES = ""
+# The starting second of the video w.r.t bounding boxes file.
+_C.DEMO.STARTING_SECOND = 900
+# Frames per second of the input video/folder of images.
+_C.DEMO.FPS = 30
+# Visualize with top-k predictions or predictions above certain threshold(s).
+# Option: {"thres", "top-k"}
+_C.DEMO.VIS_MODE = "thres"
+# Threshold for common class names.
+_C.DEMO.COMMON_CLASS_THRES = 0.7
+# Theshold for uncommon class names. This will not be
+# used if `_C.DEMO.COMMON_CLASS_NAMES` is empty.
+_C.DEMO.UNCOMMON_CLASS_THRES = 0.3
+# This is chosen based on distribution of examples in
+# each classes in AVA dataset.
+_C.DEMO.COMMON_CLASS_NAMES = [
+    "watch (a person)",
+    "talk to (e.g., self, a person, a group)",
+    "listen to (a person)",
+    "touch (an object)",
+    "carry/hold (an object)",
+    "walk",
+    "sit",
+    "lie/sleep",
+    "bend/bow (at the waist)",
+]
+# Slow-motion rate for the visualization. The visualized portions of the
+# video will be played `_C.DEMO.SLOWMO` times slower than usual speed.
+_C.DEMO.SLOWMO = 1
+
+def _assert_and_infer_cfg(cfg):
+    # BN assertions.
+    if cfg.BN.USE_PRECISE_STATS:
+        assert cfg.BN.NUM_BATCHES_PRECISE >= 0
+    # TRAIN assertions.
+    assert cfg.TRAIN.CHECKPOINT_TYPE in ["pytorch", "caffe2"]
+    assert cfg.TRAIN.BATCH_SIZE % cfg.NUM_GPUS == 0
+
+    # TEST assertions.
+    assert cfg.TEST.CHECKPOINT_TYPE in ["pytorch", "caffe2"]
+    assert cfg.TEST.BATCH_SIZE % cfg.NUM_GPUS == 0
+    assert cfg.TEST.NUM_SPATIAL_CROPS == 3
+
+    # RESNET assertions.
+    assert cfg.RESNET.NUM_GROUPS > 0
+    assert cfg.RESNET.WIDTH_PER_GROUP > 0
+    assert cfg.RESNET.WIDTH_PER_GROUP % cfg.RESNET.NUM_GROUPS == 0
+
+    # Execute LR scaling by num_shards.
+    if cfg.SOLVER.BASE_LR_SCALE_NUM_SHARDS:
+        cfg.SOLVER.BASE_LR *= cfg.NUM_SHARDS
+
+    # General assertions.
+    assert cfg.SHARD_ID < cfg.NUM_SHARDS
+    return cfg
+
+
+def get_cfg():
+    """
+    Get a copy of the default config.
+    """
+    return _assert_and_infer_cfg(_C.clone())

TimeSformer/timesformer/datasets/DATASET.md

@@ -0,0 +1,26 @@
+# Dataset Preparation
+
+## Kinetics
+
+The Kinetics Dataset could be downloaded from the following [link](https://github.com/cvdfoundation/kinetics-dataset):
+
+After all the videos were downloaded, resize the video to the short edge size of 256, then prepare the csv files for training, validation, and testing set as `train.csv`, `val.csv`, `test.csv`. The format of the csv file is:
+
+```
+path_to_video_1 label_1
+path_to_video_2 label_2
+path_to_video_3 label_3
+...
+path_to_video_N label_N
+```
+
+## Something-Something V2
+1. Please download the dataset and annotations from [dataset provider](https://20bn.com/datasets/something-something).
+
+2. Download the *frame list* from the following links: ([train](https://dl.fbaipublicfiles.com/pyslowfast/dataset/ssv2/frame_lists/train.csv), [val](https://dl.fbaipublicfiles.com/pyslowfast/dataset/ssv2/frame_lists/val.csv)).
+
+3. Extract the frames at 30 FPS. (We used ffmpeg-4.1.3 with command
+`ffmpeg -i "${video}" -r 30 -q:v 1 "${out_name}"`
+   in experiments.) Please put the frames in a structure consistent with the frame lists.
+
+Please put all annotation json files and the frame lists in the same folder, and set `DATA.PATH_TO_DATA_DIR` to the path. Set `DATA.PATH_PREFIX` to be the path to the folder containing extracted frames.

TimeSformer/timesformer/datasets/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from .build import DATASET_REGISTRY, build_dataset  # noqa
+from .kinetics import Kinetics  # noqa
+from .ssv2 import Ssv2  # noqa

TimeSformer/timesformer/datasets/build.py

@@ -0,0 +1,30 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from fvcore.common.registry import Registry
+
+DATASET_REGISTRY = Registry("DATASET")
+DATASET_REGISTRY.__doc__ = """
+Registry for dataset.
+
+The registered object will be called with `obj(cfg, split)`.
+The call should return a `torch.utils.data.Dataset` object.
+"""
+
+
+def build_dataset(dataset_name, cfg, split):
+    """
+    Build a dataset, defined by `dataset_name`.
+    Args:
+        dataset_name (str): the name of the dataset to be constructed.
+        cfg (CfgNode): configs. Details can be found in
+            slowfast/config/defaults.py
+        split (str): the split of the data loader. Options include `train`,
+            `val`, and `test`.
+    Returns:
+        Dataset: a constructed dataset specified by dataset_name.
+    """
+    # Capitalize the the first letter of the dataset_name since the dataset_name
+    # in configs may be in lowercase but the name of dataset class should always
+    # start with an uppercase letter.
+    name = dataset_name.capitalize()
+    return DATASET_REGISTRY.get(name)(cfg, split)

TimeSformer/timesformer/datasets/cv2_transform.py

@@ -0,0 +1,796 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import math
+import numpy as np
+import cv2
+
+
+def clip_boxes_to_image(boxes, height, width):
+    """
+    Clip the boxes with the height and width of the image size.
+    Args:
+        boxes (ndarray): bounding boxes to peform crop. The dimension is
+        `num boxes` x 4.
+        height (int): the height of the image.
+        width (int): the width of the image.
+    Returns:
+        boxes (ndarray): cropped bounding boxes.
+    """
+    boxes[:, [0, 2]] = np.minimum(
+        width - 1.0, np.maximum(0.0, boxes[:, [0, 2]])
+    )
+    boxes[:, [1, 3]] = np.minimum(
+        height - 1.0, np.maximum(0.0, boxes[:, [1, 3]])
+    )
+    return boxes
+
+
+def random_short_side_scale_jitter_list(images, min_size, max_size, boxes=None):
+    """
+    Perform a spatial short scale jittering on the given images and
+    corresponding boxes.
+    Args:
+        images (list): list of images to perform scale jitter. Dimension is
+            `height` x `width` x `channel`.
+        min_size (int): the minimal size to scale the frames.
+        max_size (int): the maximal size to scale the frames.
+        boxes (list): optional. Corresponding boxes to images. Dimension is
+            `num boxes` x 4.
+    Returns:
+        (list): the list of scaled images with dimension of
+            `new height` x `new width` x `channel`.
+        (ndarray or None): the scaled boxes with dimension of
+            `num boxes` x 4.
+    """
+    size = int(round(1.0 / np.random.uniform(1.0 / max_size, 1.0 / min_size)))
+
+    height = images[0].shape[0]
+    width = images[0].shape[1]
+    if (width <= height and width == size) or (
+        height <= width and height == size
+    ):
+        return images, boxes
+    new_width = size
+    new_height = size
+    if width < height:
+        new_height = int(math.floor((float(height) / width) * size))
+        if boxes is not None:
+            boxes = [
+                proposal * float(new_height) / height for proposal in boxes
+            ]
+    else:
+        new_width = int(math.floor((float(width) / height) * size))
+        if boxes is not None:
+            boxes = [proposal * float(new_width) / width for proposal in boxes]
+    return (
+        [
+            cv2.resize(
+                image, (new_width, new_height), interpolation=cv2.INTER_LINEAR
+            ).astype(np.float32)
+            for image in images
+        ],
+        boxes,
+    )
+
+
+def scale(size, image):
+    """
+    Scale the short side of the image to size.
+    Args:
+        size (int): size to scale the image.
+        image (array): image to perform short side scale. Dimension is
+            `height` x `width` x `channel`.
+    Returns:
+        (ndarray): the scaled image with dimension of
+            `height` x `width` x `channel`.
+    """
+    height = image.shape[0]
+    width = image.shape[1]
+    if (width <= height and width == size) or (
+        height <= width and height == size
+    ):
+        return image
+    new_width = size
+    new_height = size
+    if width < height:
+        new_height = int(math.floor((float(height) / width) * size))
+    else:
+        new_width = int(math.floor((float(width) / height) * size))
+    img = cv2.resize(
+        image, (new_width, new_height), interpolation=cv2.INTER_LINEAR
+    )
+    return img.astype(np.float32)
+
+
+def scale_boxes(size, boxes, height, width):
+    """
+    Scale the short side of the box to size.
+    Args:
+        size (int): size to scale the image.
+        boxes (ndarray): bounding boxes to peform scale. The dimension is
+        `num boxes` x 4.
+        height (int): the height of the image.
+        width (int): the width of the image.
+    Returns:
+        boxes (ndarray): scaled bounding boxes.
+    """
+    if (width <= height and width == size) or (
+        height <= width and height == size
+    ):
+        return boxes
+
+    new_width = size
+    new_height = size
+    if width < height:
+        new_height = int(math.floor((float(height) / width) * size))
+        boxes *= float(new_height) / height
+    else:
+        new_width = int(math.floor((float(width) / height) * size))
+        boxes *= float(new_width) / width
+    return boxes
+
+
+def horizontal_flip_list(prob, images, order="CHW", boxes=None):
+    """
+    Horizontally flip the list of image and optional boxes.
+    Args:
+        prob (float): probability to flip.
+        image (list): ilist of images to perform short side scale. Dimension is
+            `height` x `width` x `channel` or `channel` x `height` x `width`.
+        order (str): order of the `height`, `channel` and `width`.
+        boxes (list): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        (ndarray): the scaled image with dimension of
+            `height` x `width` x `channel`.
+        (list): optional. Corresponding boxes to images. Dimension is
+            `num boxes` x 4.
+    """
+    _, width, _ = images[0].shape
+    if np.random.uniform() < prob:
+        if boxes is not None:
+            boxes = [flip_boxes(proposal, width) for proposal in boxes]
+        if order == "CHW":
+            out_images = []
+            for image in images:
+                image = np.asarray(image).swapaxes(2, 0)
+                image = image[::-1]
+                out_images.append(image.swapaxes(0, 2))
+            return out_images, boxes
+        elif order == "HWC":
+            return [cv2.flip(image, 1) for image in images], boxes
+    return images, boxes
+
+
+def spatial_shift_crop_list(size, images, spatial_shift_pos, boxes=None):
+    """
+    Perform left, center, or right crop of the given list of images.
+    Args:
+        size (int): size to crop.
+        image (list): ilist of images to perform short side scale. Dimension is
+            `height` x `width` x `channel` or `channel` x `height` x `width`.
+        spatial_shift_pos (int): option includes 0 (left), 1 (middle), and
+            2 (right) crop.
+        boxes (list): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        cropped (ndarray): the cropped list of images with dimension of
+            `height` x `width` x `channel`.
+        boxes (list): optional. Corresponding boxes to images. Dimension is
+            `num boxes` x 4.
+    """
+
+    assert spatial_shift_pos in [0, 1, 2]
+
+    height = images[0].shape[0]
+    width = images[0].shape[1]
+    y_offset = int(math.ceil((height - size) / 2))
+    x_offset = int(math.ceil((width - size) / 2))
+
+    if height > width:
+        if spatial_shift_pos == 0:
+            y_offset = 0
+        elif spatial_shift_pos == 2:
+            y_offset = height - size
+    else:
+        if spatial_shift_pos == 0:
+            x_offset = 0
+        elif spatial_shift_pos == 2:
+            x_offset = width - size
+
+    cropped = [
+        image[y_offset : y_offset + size, x_offset : x_offset + size, :]
+        for image in images
+    ]
+    assert cropped[0].shape[0] == size, "Image height not cropped properly"
+    assert cropped[0].shape[1] == size, "Image width not cropped properly"
+
+    if boxes is not None:
+        for i in range(len(boxes)):
+            boxes[i][:, [0, 2]] -= x_offset
+            boxes[i][:, [1, 3]] -= y_offset
+    return cropped, boxes
+
+
+def CHW2HWC(image):
+    """
+    Transpose the dimension from `channel` x `height` x `width` to
+        `height` x `width` x `channel`.
+    Args:
+        image (array): image to transpose.
+    Returns
+        (array): transposed image.
+    """
+    return image.transpose([1, 2, 0])
+
+
+def HWC2CHW(image):
+    """
+    Transpose the dimension from `height` x `width` x `channel` to
+        `channel` x `height` x `width`.
+    Args:
+        image (array): image to transpose.
+    Returns
+        (array): transposed image.
+    """
+    return image.transpose([2, 0, 1])
+
+
+def color_jitter_list(
+    images, img_brightness=0, img_contrast=0, img_saturation=0
+):
+    """
+    Perform color jitter on the list of images.
+    Args:
+        images (list): list of images to perform color jitter.
+        img_brightness (float): jitter ratio for brightness.
+        img_contrast (float): jitter ratio for contrast.
+        img_saturation (float): jitter ratio for saturation.
+    Returns:
+        images (list): the jittered list of images.
+    """
+    jitter = []
+    if img_brightness != 0:
+        jitter.append("brightness")
+    if img_contrast != 0:
+        jitter.append("contrast")
+    if img_saturation != 0:
+        jitter.append("saturation")
+
+    if len(jitter) > 0:
+        order = np.random.permutation(np.arange(len(jitter)))
+        for idx in range(0, len(jitter)):
+            if jitter[order[idx]] == "brightness":
+                images = brightness_list(img_brightness, images)
+            elif jitter[order[idx]] == "contrast":
+                images = contrast_list(img_contrast, images)
+            elif jitter[order[idx]] == "saturation":
+                images = saturation_list(img_saturation, images)
+    return images
+
+
+def lighting_list(imgs, alphastd, eigval, eigvec, alpha=None):
+    """
+    Perform AlexNet-style PCA jitter on the given list of images.
+    Args:
+        images (list): list of images to perform lighting jitter.
+        alphastd (float): jitter ratio for PCA jitter.
+        eigval (list): eigenvalues for PCA jitter.
+        eigvec (list[list]): eigenvectors for PCA jitter.
+    Returns:
+        out_images (list): the list of jittered images.
+    """
+    if alphastd == 0:
+        return imgs
+    # generate alpha1, alpha2, alpha3
+    alpha = np.random.normal(0, alphastd, size=(1, 3))
+    eig_vec = np.array(eigvec)
+    eig_val = np.reshape(eigval, (1, 3))
+    rgb = np.sum(
+        eig_vec * np.repeat(alpha, 3, axis=0) * np.repeat(eig_val, 3, axis=0),
+        axis=1,
+    )
+    out_images = []
+    for img in imgs:
+        for idx in range(img.shape[0]):
+            img[idx] = img[idx] + rgb[2 - idx]
+        out_images.append(img)
+    return out_images
+
+
+def color_normalization(image, mean, stddev):
+    """
+    Perform color normalization on the image with the given mean and stddev.
+    Args:
+        image (array): image to perform color normalization.
+        mean (float): mean value to subtract.
+        stddev (float): stddev to devide.
+    """
+    # Input image should in format of CHW
+    assert len(mean) == image.shape[0], "channel mean not computed properly"
+    assert len(stddev) == image.shape[0], "channel stddev not computed properly"
+    for idx in range(image.shape[0]):
+        image[idx] = image[idx] - mean[idx]
+        image[idx] = image[idx] / stddev[idx]
+    return image
+
+
+def pad_image(image, pad_size, order="CHW"):
+    """
+    Pad the given image with the size of pad_size.
+    Args:
+        image (array): image to pad.
+        pad_size (int): size to pad.
+        order (str): order of the `height`, `channel` and `width`.
+    Returns:
+        img (array): padded image.
+    """
+    if order == "CHW":
+        img = np.pad(
+            image,
+            ((0, 0), (pad_size, pad_size), (pad_size, pad_size)),
+            mode=str("constant"),
+        )
+    elif order == "HWC":
+        img = np.pad(
+            image,
+            ((pad_size, pad_size), (pad_size, pad_size), (0, 0)),
+            mode=str("constant"),
+        )
+    return img
+
+
+def horizontal_flip(prob, image, order="CHW"):
+    """
+    Horizontally flip the image.
+    Args:
+        prob (float): probability to flip.
+        image (array): image to pad.
+        order (str): order of the `height`, `channel` and `width`.
+    Returns:
+        img (array): flipped image.
+    """
+    assert order in ["CHW", "HWC"], "order {} is not supported".format(order)
+    if np.random.uniform() < prob:
+        if order == "CHW":
+            image = image[:, :, ::-1]
+        elif order == "HWC":
+            image = image[:, ::-1, :]
+        else:
+            raise NotImplementedError("Unknown order {}".format(order))
+    return image
+
+
+def flip_boxes(boxes, im_width):
+    """
+    Horizontally flip the boxes.
+    Args:
+        boxes (array): box to flip.
+        im_width (int): width of the image.
+    Returns:
+        boxes_flipped (array): flipped box.
+    """
+
+    boxes_flipped = boxes.copy()
+    boxes_flipped[:, 0::4] = im_width - boxes[:, 2::4] - 1
+    boxes_flipped[:, 2::4] = im_width - boxes[:, 0::4] - 1
+    return boxes_flipped
+
+
+def crop_boxes(boxes, x_offset, y_offset):
+    """
+    Crop the boxes given the offsets.
+    Args:
+        boxes (array): boxes to crop.
+        x_offset (int): offset on x.
+        y_offset (int): offset on y.
+    """
+    boxes[:, [0, 2]] = boxes[:, [0, 2]] - x_offset
+    boxes[:, [1, 3]] = boxes[:, [1, 3]] - y_offset
+    return boxes
+
+
+def random_crop_list(images, size, pad_size=0, order="CHW", boxes=None):
+    """
+    Perform random crop on a list of images.
+    Args:
+        images (list): list of images to perform random crop.
+        size (int): size to crop.
+        pad_size (int): padding size.
+        order (str): order of the `height`, `channel` and `width`.
+        boxes (list): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        cropped (ndarray): the cropped list of images with dimension of
+            `height` x `width` x `channel`.
+        boxes (list): optional. Corresponding boxes to images. Dimension is
+            `num boxes` x 4.
+    """
+    # explicitly dealing processing per image order to avoid flipping images.
+    if pad_size > 0:
+        images = [
+            pad_image(pad_size=pad_size, image=image, order=order)
+            for image in images
+        ]
+
+    # image format should be CHW.
+    if order == "CHW":
+        if images[0].shape[1] == size and images[0].shape[2] == size:
+            return images, boxes
+        height = images[0].shape[1]
+        width = images[0].shape[2]
+        y_offset = 0
+        if height > size:
+            y_offset = int(np.random.randint(0, height - size))
+        x_offset = 0
+        if width > size:
+            x_offset = int(np.random.randint(0, width - size))
+        cropped = [
+            image[:, y_offset : y_offset + size, x_offset : x_offset + size]
+            for image in images
+        ]
+        assert cropped[0].shape[1] == size, "Image not cropped properly"
+        assert cropped[0].shape[2] == size, "Image not cropped properly"
+    elif order == "HWC":
+        if images[0].shape[0] == size and images[0].shape[1] == size:
+            return images, boxes
+        height = images[0].shape[0]
+        width = images[0].shape[1]
+        y_offset = 0
+        if height > size:
+            y_offset = int(np.random.randint(0, height - size))
+        x_offset = 0
+        if width > size:
+            x_offset = int(np.random.randint(0, width - size))
+        cropped = [
+            image[y_offset : y_offset + size, x_offset : x_offset + size, :]
+            for image in images
+        ]
+        assert cropped[0].shape[0] == size, "Image not cropped properly"
+        assert cropped[0].shape[1] == size, "Image not cropped properly"
+
+    if boxes is not None:
+        boxes = [crop_boxes(proposal, x_offset, y_offset) for proposal in boxes]
+    return cropped, boxes
+
+
+def center_crop(size, image):
+    """
+    Perform center crop on input images.
+    Args:
+        size (int): size of the cropped height and width.
+        image (array): the image to perform center crop.
+    """
+    height = image.shape[0]
+    width = image.shape[1]
+    y_offset = int(math.ceil((height - size) / 2))
+    x_offset = int(math.ceil((width - size) / 2))
+    cropped = image[y_offset : y_offset + size, x_offset : x_offset + size, :]
+    assert cropped.shape[0] == size, "Image height not cropped properly"
+    assert cropped.shape[1] == size, "Image width not cropped properly"
+    return cropped
+
+
+# ResNet style scale jittering: randomly select the scale from
+# [1/max_size, 1/min_size]
+def random_scale_jitter(image, min_size, max_size):
+    """
+    Perform ResNet style random scale jittering: randomly select the scale from
+        [1/max_size, 1/min_size].
+    Args:
+        image (array): image to perform random scale.
+        min_size (int): min size to scale.
+        max_size (int) max size to scale.
+    Returns:
+        image (array): scaled image.
+    """
+    img_scale = int(
+        round(1.0 / np.random.uniform(1.0 / max_size, 1.0 / min_size))
+    )
+    image = scale(img_scale, image)
+    return image
+
+
+def random_scale_jitter_list(images, min_size, max_size):
+    """
+    Perform ResNet style random scale jittering on a list of image: randomly
+        select the scale from [1/max_size, 1/min_size]. Note that all the image
+        will share the same scale.
+    Args:
+        images (list): list of images to perform random scale.
+        min_size (int): min size to scale.
+        max_size (int) max size to scale.
+    Returns:
+        images (list): list of scaled image.
+    """
+    img_scale = int(
+        round(1.0 / np.random.uniform(1.0 / max_size, 1.0 / min_size))
+    )
+    return [scale(img_scale, image) for image in images]
+
+
+def random_sized_crop(image, size, area_frac=0.08):
+    """
+    Perform random sized cropping on the given image. Random crop with size
+        8% - 100% image area and aspect ratio in [3/4, 4/3].
+    Args:
+        image (array): image to crop.
+        size (int): size to crop.
+        area_frac (float): area of fraction.
+    Returns:
+        (array): cropped image.
+    """
+    for _ in range(0, 10):
+        height = image.shape[0]
+        width = image.shape[1]
+        area = height * width
+        target_area = np.random.uniform(area_frac, 1.0) * area
+        aspect_ratio = np.random.uniform(3.0 / 4.0, 4.0 / 3.0)
+        w = int(round(math.sqrt(float(target_area) * aspect_ratio)))
+        h = int(round(math.sqrt(float(target_area) / aspect_ratio)))
+        if np.random.uniform() < 0.5:
+            w, h = h, w
+        if h <= height and w <= width:
+            if height == h:
+                y_offset = 0
+            else:
+                y_offset = np.random.randint(0, height - h)
+            if width == w:
+                x_offset = 0
+            else:
+                x_offset = np.random.randint(0, width - w)
+            y_offset = int(y_offset)
+            x_offset = int(x_offset)
+            cropped = image[y_offset : y_offset + h, x_offset : x_offset + w, :]
+            assert (
+                cropped.shape[0] == h and cropped.shape[1] == w
+            ), "Wrong crop size"
+            cropped = cv2.resize(
+                cropped, (size, size), interpolation=cv2.INTER_LINEAR
+            )
+            return cropped.astype(np.float32)
+    return center_crop(size, scale(size, image))
+
+
+def lighting(img, alphastd, eigval, eigvec):
+    """
+    Perform AlexNet-style PCA jitter on the given image.
+    Args:
+        image (array): list of images to perform lighting jitter.
+        alphastd (float): jitter ratio for PCA jitter.
+        eigval (array): eigenvalues for PCA jitter.
+        eigvec (list): eigenvectors for PCA jitter.
+    Returns:
+        img (tensor): the jittered image.
+    """
+    if alphastd == 0:
+        return img
+    # generate alpha1, alpha2, alpha3.
+    alpha = np.random.normal(0, alphastd, size=(1, 3))
+    eig_vec = np.array(eigvec)
+    eig_val = np.reshape(eigval, (1, 3))
+    rgb = np.sum(
+        eig_vec * np.repeat(alpha, 3, axis=0) * np.repeat(eig_val, 3, axis=0),
+        axis=1,
+    )
+    for idx in range(img.shape[0]):
+        img[idx] = img[idx] + rgb[2 - idx]
+    return img
+
+
+def random_sized_crop_list(images, size, crop_area_fraction=0.08):
+    """
+    Perform random sized cropping on the given list of images. Random crop with
+        size 8% - 100% image area and aspect ratio in [3/4, 4/3].
+    Args:
+        images (list): image to crop.
+        size (int): size to crop.
+        area_frac (float): area of fraction.
+    Returns:
+        (list): list of cropped image.
+    """
+    for _ in range(0, 10):
+        height = images[0].shape[0]
+        width = images[0].shape[1]
+        area = height * width
+        target_area = np.random.uniform(crop_area_fraction, 1.0) * area
+        aspect_ratio = np.random.uniform(3.0 / 4.0, 4.0 / 3.0)
+        w = int(round(math.sqrt(float(target_area) * aspect_ratio)))
+        h = int(round(math.sqrt(float(target_area) / aspect_ratio)))
+        if np.random.uniform() < 0.5:
+            w, h = h, w
+        if h <= height and w <= width:
+            if height == h:
+                y_offset = 0
+            else:
+                y_offset = np.random.randint(0, height - h)
+            if width == w:
+                x_offset = 0
+            else:
+                x_offset = np.random.randint(0, width - w)
+            y_offset = int(y_offset)
+            x_offset = int(x_offset)
+
+            croppsed_images = []
+            for image in images:
+                cropped = image[
+                    y_offset : y_offset + h, x_offset : x_offset + w, :
+                ]
+                assert (
+                    cropped.shape[0] == h and cropped.shape[1] == w
+                ), "Wrong crop size"
+                cropped = cv2.resize(
+                    cropped, (size, size), interpolation=cv2.INTER_LINEAR
+                )
+                croppsed_images.append(cropped.astype(np.float32))
+            return croppsed_images
+
+    return [center_crop(size, scale(size, image)) for image in images]
+
+
+def blend(image1, image2, alpha):
+    return image1 * alpha + image2 * (1 - alpha)
+
+
+def grayscale(image):
+    """
+    Convert the image to gray scale.
+    Args:
+        image (tensor): image to convert to gray scale. Dimension is
+            `channel` x `height` x `width`.
+    Returns:
+        img_gray (tensor): image in gray scale.
+    """
+    # R -> 0.299, G -> 0.587, B -> 0.114.
+    img_gray = np.copy(image)
+    gray_channel = 0.299 * image[2] + 0.587 * image[1] + 0.114 * image[0]
+    img_gray[0] = gray_channel
+    img_gray[1] = gray_channel
+    img_gray[2] = gray_channel
+    return img_gray
+
+
+def saturation(var, image):
+    """
+    Perform color saturation on the given image.
+    Args:
+        var (float): variance.
+        image (array): image to perform color saturation.
+    Returns:
+        (array): image that performed color saturation.
+    """
+    img_gray = grayscale(image)
+    alpha = 1.0 + np.random.uniform(-var, var)
+    return blend(image, img_gray, alpha)
+
+
+def brightness(var, image):
+    """
+    Perform color brightness on the given image.
+    Args:
+        var (float): variance.
+        image (array): image to perform color brightness.
+    Returns:
+        (array): image that performed color brightness.
+    """
+    img_bright = np.zeros(image.shape).astype(image.dtype)
+    alpha = 1.0 + np.random.uniform(-var, var)
+    return blend(image, img_bright, alpha)
+
+
+def contrast(var, image):
+    """
+    Perform color contrast on the given image.
+    Args:
+        var (float): variance.
+        image (array): image to perform color contrast.
+    Returns:
+        (array): image that performed color contrast.
+    """
+    img_gray = grayscale(image)
+    img_gray.fill(np.mean(img_gray[0]))
+    alpha = 1.0 + np.random.uniform(-var, var)
+    return blend(image, img_gray, alpha)
+
+
+def saturation_list(var, images):
+    """
+    Perform color saturation on the list of given images.
+    Args:
+        var (float): variance.
+        images (list): list of images to perform color saturation.
+    Returns:
+        (list): list of images that performed color saturation.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+
+    out_images = []
+    for image in images:
+        img_gray = grayscale(image)
+        out_images.append(blend(image, img_gray, alpha))
+    return out_images
+
+
+def brightness_list(var, images):
+    """
+    Perform color brightness on the given list of images.
+    Args:
+        var (float): variance.
+        images (list): list of images to perform color brightness.
+    Returns:
+        (array): list of images that performed color brightness.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+
+    out_images = []
+    for image in images:
+        img_bright = np.zeros(image.shape).astype(image.dtype)
+        out_images.append(blend(image, img_bright, alpha))
+    return out_images
+
+
+def contrast_list(var, images):
+    """
+    Perform color contrast on the given list of images.
+    Args:
+        var (float): variance.
+        images (list): list of images to perform color contrast.
+    Returns:
+        (array): image that performed color contrast.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+
+    out_images = []
+    for image in images:
+        img_gray = grayscale(image)
+        img_gray.fill(np.mean(img_gray[0]))
+        out_images.append(blend(image, img_gray, alpha))
+    return out_images
+
+
+def color_jitter(image, img_brightness=0, img_contrast=0, img_saturation=0):
+    """
+    Perform color jitter on the given image.
+    Args:
+        image (array): image to perform color jitter.
+        img_brightness (float): jitter ratio for brightness.
+        img_contrast (float): jitter ratio for contrast.
+        img_saturation (float): jitter ratio for saturation.
+    Returns:
+        image (array): the jittered image.
+    """
+    jitter = []
+    if img_brightness != 0:
+        jitter.append("brightness")
+    if img_contrast != 0:
+        jitter.append("contrast")
+    if img_saturation != 0:
+        jitter.append("saturation")
+
+    if len(jitter) > 0:
+        order = np.random.permutation(np.arange(len(jitter)))
+        for idx in range(0, len(jitter)):
+            if jitter[order[idx]] == "brightness":
+                image = brightness(img_brightness, image)
+            elif jitter[order[idx]] == "contrast":
+                image = contrast(img_contrast, image)
+            elif jitter[order[idx]] == "saturation":
+                image = saturation(img_saturation, image)
+    return image
+
+
+def revert_scaled_boxes(size, boxes, img_height, img_width):
+    """
+    Revert scaled input boxes to match the original image size.
+    Args:
+        size (int): size of the cropped image.
+        boxes (array): shape (num_boxes, 4).
+        img_height (int): height of original image.
+        img_width (int): width of original image.
+    Returns:
+        reverted_boxes (array): boxes scaled back to the original image size.
+    """
+    scaled_aspect = np.min([img_height, img_width])
+    scale_ratio = scaled_aspect / size
+    reverted_boxes = boxes * scale_ratio
+    return reverted_boxes

TimeSformer/timesformer/datasets/decoder.py

@@ -0,0 +1,392 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import math
+import numpy as np
+import random
+import torch
+import torchvision.io as io
+
+
+def temporal_sampling(frames, start_idx, end_idx, num_samples):
+    """
+    Given the start and end frame index, sample num_samples frames between
+    the start and end with equal interval.
+    Args:
+        frames (tensor): a tensor of video frames, dimension is
+            `num video frames` x `channel` x `height` x `width`.
+        start_idx (int): the index of the start frame.
+        end_idx (int): the index of the end frame.
+        num_samples (int): number of frames to sample.
+    Returns:
+        frames (tersor): a tensor of temporal sampled video frames, dimension is
+            `num clip frames` x `channel` x `height` x `width`.
+    """
+    index = torch.linspace(start_idx, end_idx, num_samples)
+    index = torch.clamp(index, 0, frames.shape[0] - 1).long()
+    frames = torch.index_select(frames, 0, index)
+    return frames
+
+
+def get_start_end_idx(video_size, clip_size, clip_idx, num_clips):
+    """
+    Sample a clip of size clip_size from a video of size video_size and
+    return the indices of the first and last frame of the clip. If clip_idx is
+    -1, the clip is randomly sampled, otherwise uniformly split the video to
+    num_clips clips, and select the start and end index of clip_idx-th video
+    clip.
+    Args:
+        video_size (int): number of overall frames.
+        clip_size (int): size of the clip to sample from the frames.
+        clip_idx (int): if clip_idx is -1, perform random jitter sampling. If
+            clip_idx is larger than -1, uniformly split the video to num_clips
+            clips, and select the start and end index of the clip_idx-th video
+            clip.
+        num_clips (int): overall number of clips to uniformly sample from the
+            given video for testing.
+    Returns:
+        start_idx (int): the start frame index.
+        end_idx (int): the end frame index.
+    """
+    delta = max(video_size - clip_size, 0)
+    if clip_idx == -1:
+        # Random temporal sampling.
+        start_idx = random.uniform(0, delta)
+    else:
+        # Uniformly sample the clip with the given index.
+        start_idx = delta * clip_idx / num_clips
+    end_idx = start_idx + clip_size - 1
+    return start_idx, end_idx
+
+
+def pyav_decode_stream(
+    container, start_pts, end_pts, stream, stream_name, buffer_size=0
+):
+    """
+    Decode the video with PyAV decoder.
+    Args:
+        container (container): PyAV container.
+        start_pts (int): the starting Presentation TimeStamp to fetch the
+            video frames.
+        end_pts (int): the ending Presentation TimeStamp of the decoded frames.
+        stream (stream): PyAV stream.
+        stream_name (dict): a dictionary of streams. For example, {"video": 0}
+            means video stream at stream index 0.
+        buffer_size (int): number of additional frames to decode beyond end_pts.
+    Returns:
+        result (list): list of frames decoded.
+        max_pts (int): max Presentation TimeStamp of the video sequence.
+    """
+    # Seeking in the stream is imprecise. Thus, seek to an ealier PTS by a
+    # margin pts.
+    margin = 1024
+    seek_offset = max(start_pts - margin, 0)
+
+    container.seek(seek_offset, any_frame=False, backward=True, stream=stream)
+    frames = {}
+    buffer_count = 0
+    max_pts = 0
+    for frame in container.decode(**stream_name):
+        max_pts = max(max_pts, frame.pts)
+        if frame.pts < start_pts:
+            continue
+        if frame.pts <= end_pts:
+            frames[frame.pts] = frame
+        else:
+            buffer_count += 1
+            frames[frame.pts] = frame
+            if buffer_count >= buffer_size:
+                break
+    result = [frames[pts] for pts in sorted(frames)]
+    return result, max_pts
+
+
+def torchvision_decode(
+    video_handle,
+    sampling_rate,
+    num_frames,
+    clip_idx,
+    video_meta,
+    num_clips=10,
+    target_fps=30,
+    modalities=("visual",),
+    max_spatial_scale=0,
+):
+    """
+    If video_meta is not empty, perform temporal selective decoding to sample a
+    clip from the video with TorchVision decoder. If video_meta is empty, decode
+    the entire video and update the video_meta.
+    Args:
+        video_handle (bytes): raw bytes of the video file.
+        sampling_rate (int): frame sampling rate (interval between two sampled
+            frames).
+        num_frames (int): number of frames to sample.
+        clip_idx (int): if clip_idx is -1, perform random temporal
+            sampling. If clip_idx is larger than -1, uniformly split the
+            video to num_clips clips, and select the clip_idx-th video clip.
+        video_meta (dict): a dict contains VideoMetaData. Details can be found
+            at `pytorch/vision/torchvision/io/_video_opt.py`.
+        num_clips (int): overall number of clips to uniformly sample from the
+            given video.
+        target_fps (int): the input video may has different fps, convert it to
+            the target video fps.
+        modalities (tuple): tuple of modalities to decode. Currently only
+            support `visual`, planning to support `acoustic` soon.
+        max_spatial_scale (int): the maximal resolution of the spatial shorter
+            edge size during decoding.
+    Returns:
+        frames (tensor): decoded frames from the video.
+        fps (float): the number of frames per second of the video.
+        decode_all_video (bool): if True, the entire video was decoded.
+    """
+    # Convert the bytes to a tensor.
+    video_tensor = torch.from_numpy(np.frombuffer(video_handle, dtype=np.uint8))
+
+    decode_all_video = True
+    video_start_pts, video_end_pts = 0, -1
+    # The video_meta is empty, fetch the meta data from the raw video.
+    if len(video_meta) == 0:
+        # Tracking the meta info for selective decoding in the future.
+        meta = io._probe_video_from_memory(video_tensor)
+        # Using the information from video_meta to perform selective decoding.
+        video_meta["video_timebase"] = meta.video_timebase
+        video_meta["video_numerator"] = meta.video_timebase.numerator
+        video_meta["video_denominator"] = meta.video_timebase.denominator
+        video_meta["has_video"] = meta.has_video
+        video_meta["video_duration"] = meta.video_duration
+        video_meta["video_fps"] = meta.video_fps
+        video_meta["audio_timebas"] = meta.audio_timebase
+        video_meta["audio_numerator"] = meta.audio_timebase.numerator
+        video_meta["audio_denominator"] = meta.audio_timebase.denominator
+        video_meta["has_audio"] = meta.has_audio
+        video_meta["audio_duration"] = meta.audio_duration
+        video_meta["audio_sample_rate"] = meta.audio_sample_rate
+
+    fps = video_meta["video_fps"]
+    if (
+        video_meta["has_video"]
+        and video_meta["video_denominator"] > 0
+        and video_meta["video_duration"] > 0
+    ):
+        # try selective decoding.
+        decode_all_video = False
+        clip_size = sampling_rate * num_frames / target_fps * fps
+        start_idx, end_idx = get_start_end_idx(
+            fps * video_meta["video_duration"], clip_size, clip_idx, num_clips
+        )
+        # Convert frame index to pts.
+        pts_per_frame = video_meta["video_denominator"] / fps
+        video_start_pts = int(start_idx * pts_per_frame)
+        video_end_pts = int(end_idx * pts_per_frame)
+
+    # Decode the raw video with the tv decoder.
+    v_frames, _ = io._read_video_from_memory(
+        video_tensor,
+        seek_frame_margin=1.0,
+        read_video_stream="visual" in modalities,
+        video_width=0,
+        video_height=0,
+        video_min_dimension=max_spatial_scale,
+        video_pts_range=(video_start_pts, video_end_pts),
+        video_timebase_numerator=video_meta["video_numerator"],
+        video_timebase_denominator=video_meta["video_denominator"],
+    )
+
+    if v_frames.shape == torch.Size([0]):
+        # failed selective decoding
+        decode_all_video = True
+        video_start_pts, video_end_pts = 0, -1
+        v_frames, _ = io._read_video_from_memory(
+            video_tensor,
+            seek_frame_margin=1.0,
+            read_video_stream="visual" in modalities,
+            video_width=0,
+            video_height=0,
+            video_min_dimension=max_spatial_scale,
+            video_pts_range=(video_start_pts, video_end_pts),
+            video_timebase_numerator=video_meta["video_numerator"],
+            video_timebase_denominator=video_meta["video_denominator"],
+        )
+
+    return v_frames, fps, decode_all_video
+
+
+def pyav_decode(
+    container, sampling_rate, num_frames, clip_idx, num_clips=10, target_fps=30, start=None, end=None
+, duration=None, frames_length=None):
+    """
+    Convert the video from its original fps to the target_fps. If the video
+    support selective decoding (contain decoding information in the video head),
+    the perform temporal selective decoding and sample a clip from the video
+    with the PyAV decoder. If the video does not support selective decoding,
+    decode the entire video.
+
+    Args:
+        container (container): pyav container.
+        sampling_rate (int): frame sampling rate (interval between two sampled
+            frames.
+        num_frames (int): number of frames to sample.
+        clip_idx (int): if clip_idx is -1, perform random temporal sampling. If
+            clip_idx is larger than -1, uniformly split the video to num_clips
+            clips, and select the clip_idx-th video clip.
+        num_clips (int): overall number of clips to uniformly sample from the
+            given video.
+        target_fps (int): the input video may has different fps, convert it to
+            the target video fps before frame sampling.
+    Returns:
+        frames (tensor): decoded frames from the video. Return None if the no
+            video stream was found.
+        fps (float): the number of frames per second of the video.
+        decode_all_video (bool): If True, the entire video was decoded.
+    """
+    # Try to fetch the decoding information from the video head. Some of the
+    # videos does not support fetching the decoding information, for that case
+    # it will get None duration.
+    fps = float(container.streams.video[0].average_rate)
+
+    orig_duration = duration
+    tb = float(container.streams.video[0].time_base)
+    frames_length = container.streams.video[0].frames
+    duration = container.streams.video[0].duration
+    if duration is None and orig_duration is not None:
+       duration = orig_duration / tb
+
+    if duration is None:
+        # If failed to fetch the decoding information, decode the entire video.
+        decode_all_video = True
+        video_start_pts, video_end_pts = 0, math.inf
+    else:
+        # Perform selective decoding.
+        decode_all_video = False
+        start_idx, end_idx = get_start_end_idx(
+            frames_length,
+            sampling_rate * num_frames / target_fps * fps,
+            clip_idx,
+            num_clips,
+        )
+        timebase = duration / frames_length
+        video_start_pts = int(start_idx * timebase)
+        video_end_pts = int(end_idx * timebase)
+
+    if start is not None and end is not None:
+        decode_all_video = False
+
+    frames = None
+    # If video stream was found, fetch video frames from the video.
+    if container.streams.video:
+        if start is None and end is None:
+            video_frames, max_pts = pyav_decode_stream(
+                container,
+                video_start_pts,
+                video_end_pts,
+                container.streams.video[0],
+                {"video": 0},
+            )
+        else:
+            timebase = duration / frames_length
+            start_i = start
+            end_i = end
+            video_frames, max_pts = pyav_decode_stream(
+                container,
+                start_i,
+                end_i,
+                container.streams.video[0],
+                {"video": 0},
+            )
+        container.close()
+
+        frames = [frame.to_rgb().to_ndarray() for frame in video_frames]
+        frames = torch.as_tensor(np.stack(frames))
+
+    return frames, fps, decode_all_video
+
+
+def decode(
+    container,
+    sampling_rate,
+    num_frames,
+    clip_idx=-1,
+    num_clips=10,
+    video_meta=None,
+    target_fps=30,
+    backend="pyav",
+    max_spatial_scale=0,
+    start=None,
+    end=None,
+    duration=None,
+    frames_length=None,
+):
+    """
+    Decode the video and perform temporal sampling.
+    Args:
+        container (container): pyav container.
+        sampling_rate (int): frame sampling rate (interval between two sampled
+            frames).
+        num_frames (int): number of frames to sample.
+        clip_idx (int): if clip_idx is -1, perform random temporal
+            sampling. If clip_idx is larger than -1, uniformly split the
+            video to num_clips clips, and select the
+            clip_idx-th video clip.
+        num_clips (int): overall number of clips to uniformly
+            sample from the given video.
+        video_meta (dict): a dict contains VideoMetaData. Details can be find
+            at `pytorch/vision/torchvision/io/_video_opt.py`.
+        target_fps (int): the input video may have different fps, convert it to
+            the target video fps before frame sampling.
+        backend (str): decoding backend includes `pyav` and `torchvision`. The
+            default one is `pyav`.
+        max_spatial_scale (int): keep the aspect ratio and resize the frame so
+            that shorter edge size is max_spatial_scale. Only used in
+            `torchvision` backend.
+    Returns:
+        frames (tensor): decoded frames from the video.
+    """
+    # Currently support two decoders: 1) PyAV, and 2) TorchVision.
+    assert clip_idx >= -1, "Not valied clip_idx {}".format(clip_idx)
+    try:
+        if backend == "pyav":
+            frames, fps, decode_all_video = pyav_decode(
+                container,
+                sampling_rate,
+                num_frames,
+                clip_idx,
+                num_clips,
+                target_fps,
+                start,
+                end,
+                duration,
+                frames_length,
+            )
+        elif backend == "torchvision":
+            frames, fps, decode_all_video = torchvision_decode(
+                container,
+                sampling_rate,
+                num_frames,
+                clip_idx,
+                video_meta,
+                num_clips,
+                target_fps,
+                ("visual",),
+                max_spatial_scale,
+            )
+        else:
+            raise NotImplementedError(
+                "Unknown decoding backend {}".format(backend)
+            )
+    except Exception as e:
+        print("Failed to decode by {} with exception: {}".format(backend, e))
+        return None
+
+    # Return None if the frames was not decoded successfully.
+    if frames is None or frames.size(0) == 0:
+        return None
+
+    clip_sz = sampling_rate * num_frames / target_fps * fps
+    start_idx, end_idx = get_start_end_idx(
+        frames.shape[0],
+        clip_sz,
+        clip_idx if decode_all_video else 0,
+        num_clips if decode_all_video else 1,
+    )
+    # Perform temporal sampling from the decoded video.
+    frames = temporal_sampling(frames, start_idx, end_idx, num_frames)
+    return frames

TimeSformer/timesformer/datasets/kinetics.py

@@ -0,0 +1,294 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import os
+import random
+import torch
+import torch.utils.data
+from fvcore.common.file_io import PathManager
+
+import timesformer.utils.logging as logging
+
+from . import decoder as decoder
+from . import utils as utils
+from . import video_container as container
+from .build import DATASET_REGISTRY
+logger = logging.get_logger(__name__)
+
+
+@DATASET_REGISTRY.register()
+class Kinetics(torch.utils.data.Dataset):
+    """
+    Kinetics video loader. Construct the Kinetics video loader, then sample
+    clips from the videos. For training and validation, a single clip is
+    randomly sampled from every video with random cropping, scaling, and
+    flipping. For testing, multiple clips are uniformaly sampled from every
+    video with uniform cropping. For uniform cropping, we take the left, center,
+    and right crop if the width is larger than height, or take top, center, and
+    bottom crop if the height is larger than the width.
+    """
+
+    def __init__(self, cfg, mode, num_retries=10):
+        """
+        Construct the Kinetics video loader with a given csv file. The format of
+        the csv file is:
+        ```
+        path_to_video_1 label_1
+        path_to_video_2 label_2
+        ...
+        path_to_video_N label_N
+        ```
+        Args:
+            cfg (CfgNode): configs.
+            mode (string): Options includes `train`, `val`, or `test` mode.
+                For the train and val mode, the data loader will take data
+                from the train or val set, and sample one clip per video.
+                For the test mode, the data loader will take data from test set,
+                and sample multiple clips per video.
+            num_retries (int): number of retries.
+        """
+        # Only support train, val, and test mode.
+        assert mode in [
+            "train",
+            "val",
+            "test",
+        ], "Split '{}' not supported for Kinetics".format(mode)
+        self.mode = mode
+        self.cfg = cfg
+
+        self._video_meta = {}
+        self._num_retries = num_retries
+        # For training or validation mode, one single clip is sampled from every
+        # video. For testing, NUM_ENSEMBLE_VIEWS clips are sampled from every
+        # video. For every clip, NUM_SPATIAL_CROPS is cropped spatially from
+        # the frames.
+        if self.mode in ["train", "val"]:
+            self._num_clips = 1
+        elif self.mode in ["test"]:
+            self._num_clips = (
+                cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS
+            )
+
+        logger.info("Constructing Kinetics {}...".format(mode))
+        self._construct_loader()
+
+    def _construct_loader(self):
+        """
+        Construct the video loader.
+        """
+        path_to_file = os.path.join(
+            self.cfg.DATA.PATH_TO_DATA_DIR, "{}.csv".format(self.mode)
+        )
+        assert PathManager.exists(path_to_file), "{} dir not found".format(
+            path_to_file
+        )
+
+        self._path_to_videos = []
+        self._labels = []
+        self._spatial_temporal_idx = []
+        with PathManager.open(path_to_file, "r") as f:
+            for clip_idx, path_label in enumerate(f.read().splitlines()):
+                assert (
+                    len(path_label.split(self.cfg.DATA.PATH_LABEL_SEPARATOR))
+                    == 2
+                )
+                path, label = path_label.split(
+                    self.cfg.DATA.PATH_LABEL_SEPARATOR
+                )
+                for idx in range(self._num_clips):
+                    self._path_to_videos.append(
+                        os.path.join(self.cfg.DATA.PATH_PREFIX, path)
+                    )
+                    self._labels.append(int(label))
+                    self._spatial_temporal_idx.append(idx)
+                    self._video_meta[clip_idx * self._num_clips + idx] = {}
+        assert (
+            len(self._path_to_videos) > 0
+        ), "Failed to load Kinetics split {} from {}".format(
+            self._split_idx, path_to_file
+        )
+        logger.info(
+            "Constructing kinetics dataloader (size: {}) from {}".format(
+                len(self._path_to_videos), path_to_file
+            )
+        )
+
+    def __getitem__(self, index):
+        """
+        Given the video index, return the list of frames, label, and video
+        index if the video can be fetched and decoded successfully, otherwise
+        repeatly find a random video that can be decoded as a replacement.
+        Args:
+            index (int): the video index provided by the pytorch sampler.
+        Returns:
+            frames (tensor): the frames of sampled from the video. The dimension
+                is `channel` x `num frames` x `height` x `width`.
+            label (int): the label of the current video.
+            index (int): if the video provided by pytorch sampler can be
+                decoded, then return the index of the video. If not, return the
+                index of the video replacement that can be decoded.
+        """
+        short_cycle_idx = None
+        # When short cycle is used, input index is a tupple.
+        if isinstance(index, tuple):
+            index, short_cycle_idx = index
+
+        if self.mode in ["train", "val"]:
+            # -1 indicates random sampling.
+            temporal_sample_index = -1
+            spatial_sample_index = -1
+            min_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[0]
+            max_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[1]
+            crop_size = self.cfg.DATA.TRAIN_CROP_SIZE
+            if short_cycle_idx in [0, 1]:
+                crop_size = int(
+                    round(
+                        self.cfg.MULTIGRID.SHORT_CYCLE_FACTORS[short_cycle_idx]
+                        * self.cfg.MULTIGRID.DEFAULT_S
+                    )
+                )
+            if self.cfg.MULTIGRID.DEFAULT_S > 0:
+                # Decreasing the scale is equivalent to using a larger "span"
+                # in a sampling grid.
+                min_scale = int(
+                    round(
+                        float(min_scale)
+                        * crop_size
+                        / self.cfg.MULTIGRID.DEFAULT_S
+                    )
+                )
+        elif self.mode in ["test"]:
+            temporal_sample_index = (
+                self._spatial_temporal_idx[index]
+                // self.cfg.TEST.NUM_SPATIAL_CROPS
+            )
+            # spatial_sample_index is in [0, 1, 2]. Corresponding to left,
+            # center, or right if width is larger than height, and top, middle,
+            # or bottom if height is larger than width.
+            spatial_sample_index = (
+                (
+                    self._spatial_temporal_idx[index]
+                    % self.cfg.TEST.NUM_SPATIAL_CROPS
+                )
+                if self.cfg.TEST.NUM_SPATIAL_CROPS > 1
+                else 1
+            )
+            min_scale, max_scale, crop_size = (
+                [self.cfg.DATA.TEST_CROP_SIZE] * 3
+                if self.cfg.TEST.NUM_SPATIAL_CROPS > 1
+                else [self.cfg.DATA.TRAIN_JITTER_SCALES[0]] * 2
+                + [self.cfg.DATA.TEST_CROP_SIZE]
+            )
+            # The testing is deterministic and no jitter should be performed.
+            # min_scale, max_scale, and crop_size are expect to be the same.
+            assert len({min_scale, max_scale}) == 1
+        else:
+            raise NotImplementedError(
+                "Does not support {} mode".format(self.mode)
+            )
+        sampling_rate = utils.get_random_sampling_rate(
+            self.cfg.MULTIGRID.LONG_CYCLE_SAMPLING_RATE,
+            self.cfg.DATA.SAMPLING_RATE,
+        )
+        # Try to decode and sample a clip from a video. If the video can not be
+        # decoded, repeatly find a random video replacement that can be decoded.
+        for i_try in range(self._num_retries):
+            video_container = None
+            try:
+                video_container = container.get_video_container(
+                    self._path_to_videos[index],
+                    self.cfg.DATA_LOADER.ENABLE_MULTI_THREAD_DECODE,
+                    self.cfg.DATA.DECODING_BACKEND,
+                )
+            except Exception as e:
+                logger.info(
+                    "Failed to load video from {} with error {}".format(
+                        self._path_to_videos[index], e
+                    )
+                )
+            # Select a random video if the current video was not able to access.
+            if video_container is None:
+                logger.warning(
+                    "Failed to meta load video idx {} from {}; trial {}".format(
+                        index, self._path_to_videos[index], i_try
+                    )
+                )
+                if self.mode not in ["test"] and i_try > self._num_retries // 2:
+                    # let's try another one
+                    index = random.randint(0, len(self._path_to_videos) - 1)
+                continue
+
+            # Decode video. Meta info is used to perform selective decoding.
+            frames = decoder.decode(
+                video_container,
+                sampling_rate,
+                self.cfg.DATA.NUM_FRAMES,
+                temporal_sample_index,
+                self.cfg.TEST.NUM_ENSEMBLE_VIEWS,
+                video_meta=self._video_meta[index],
+                target_fps=self.cfg.DATA.TARGET_FPS,
+                backend=self.cfg.DATA.DECODING_BACKEND,
+                max_spatial_scale=min_scale,
+            )
+
+            # If decoding failed (wrong format, video is too short, and etc),
+            # select another video.
+            if frames is None:
+                logger.warning(
+                    "Failed to decode video idx {} from {}; trial {}".format(
+                        index, self._path_to_videos[index], i_try
+                    )
+                )
+                if self.mode not in ["test"] and i_try > self._num_retries // 2:
+                    # let's try another one
+                    index = random.randint(0, len(self._path_to_videos) - 1)
+                continue
+
+
+            label = self._labels[index]
+
+            # Perform color normalization.
+            frames = utils.tensor_normalize(
+                frames, self.cfg.DATA.MEAN, self.cfg.DATA.STD
+            )
+
+            # T H W C -> C T H W.
+            frames = frames.permute(3, 0, 1, 2)
+            # Perform data augmentation.
+            frames = utils.spatial_sampling(
+                frames,
+                spatial_idx=spatial_sample_index,
+                min_scale=min_scale,
+                max_scale=max_scale,
+                crop_size=crop_size,
+                random_horizontal_flip=self.cfg.DATA.RANDOM_FLIP,
+                inverse_uniform_sampling=self.cfg.DATA.INV_UNIFORM_SAMPLE,
+            )
+
+
+            if not self.cfg.MODEL.ARCH in ['vit']:
+                frames = utils.pack_pathway_output(self.cfg, frames)
+            else:
+                # Perform temporal sampling from the fast pathway.
+                frames = torch.index_select(
+                     frames,
+                     1,
+                     torch.linspace(
+                         0, frames.shape[1] - 1, self.cfg.DATA.NUM_FRAMES
+
+                     ).long(),
+                )
+
+            return frames, label, index, {}
+        else:
+            raise RuntimeError(
+                "Failed to fetch video after {} retries.".format(
+                    self._num_retries
+                )
+            )
+
+    def __len__(self):
+        """
+        Returns:
+            (int): the number of videos in the dataset.
+        """
+        return len(self._path_to_videos)

TimeSformer/timesformer/datasets/loader.py

@@ -0,0 +1,134 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""Data loader."""
+
+import itertools
+import numpy as np
+import torch
+from torch.utils.data._utils.collate import default_collate
+from torch.utils.data.distributed import DistributedSampler
+from torch.utils.data.sampler import RandomSampler
+
+from timesformer.datasets.multigrid_helper import ShortCycleBatchSampler
+
+from . import utils as utils
+from .build import build_dataset
+
+
+def detection_collate(batch):
+    """
+    Collate function for detection task. Concatanate bboxes, labels and
+    metadata from different samples in the first dimension instead of
+    stacking them to have a batch-size dimension.
+    Args:
+        batch (tuple or list): data batch to collate.
+    Returns:
+        (tuple): collated detection data batch.
+    """
+    inputs, labels, video_idx, extra_data = zip(*batch)
+    inputs, video_idx = default_collate(inputs), default_collate(video_idx)
+    labels = torch.tensor(np.concatenate(labels, axis=0)).float()
+
+    collated_extra_data = {}
+    for key in extra_data[0].keys():
+        data = [d[key] for d in extra_data]
+        if key == "boxes" or key == "ori_boxes":
+            # Append idx info to the bboxes before concatenating them.
+            bboxes = [
+                np.concatenate(
+                    [np.full((data[i].shape[0], 1), float(i)), data[i]], axis=1
+                )
+                for i in range(len(data))
+            ]
+            bboxes = np.concatenate(bboxes, axis=0)
+            collated_extra_data[key] = torch.tensor(bboxes).float()
+        elif key == "metadata":
+            collated_extra_data[key] = torch.tensor(
+                list(itertools.chain(*data))
+            ).view(-1, 2)
+        else:
+            collated_extra_data[key] = default_collate(data)
+
+    return inputs, labels, video_idx, collated_extra_data
+
+
+def construct_loader(cfg, split, is_precise_bn=False):
+    """
+    Constructs the data loader for the given dataset.
+    Args:
+        cfg (CfgNode): configs. Details can be found in
+            slowfast/config/defaults.py
+        split (str): the split of the data loader. Options include `train`,
+            `val`, and `test`.
+    """
+    assert split in ["train", "val", "test"]
+    if split in ["train"]:
+        dataset_name = cfg.TRAIN.DATASET
+        batch_size = int(cfg.TRAIN.BATCH_SIZE / max(1, cfg.NUM_GPUS))
+        shuffle = True
+        drop_last = True
+    elif split in ["val"]:
+        dataset_name = cfg.TRAIN.DATASET
+        batch_size = int(cfg.TRAIN.BATCH_SIZE / max(1, cfg.NUM_GPUS))
+        shuffle = False
+        drop_last = False
+    elif split in ["test"]:
+        dataset_name = cfg.TEST.DATASET
+        batch_size = int(cfg.TEST.BATCH_SIZE / max(1, cfg.NUM_GPUS))
+        shuffle = False
+        drop_last = False
+
+    # Construct the dataset
+    dataset = build_dataset(dataset_name, cfg, split)
+
+    if cfg.MULTIGRID.SHORT_CYCLE and split in ["train"] and not is_precise_bn:
+        # Create a sampler for multi-process training
+        sampler = utils.create_sampler(dataset, shuffle, cfg)
+        batch_sampler = ShortCycleBatchSampler(
+            sampler, batch_size=batch_size, drop_last=drop_last, cfg=cfg
+        )
+        # Create a loader
+        loader = torch.utils.data.DataLoader(
+            dataset,
+            batch_sampler=batch_sampler,
+            num_workers=cfg.DATA_LOADER.NUM_WORKERS,
+            pin_memory=cfg.DATA_LOADER.PIN_MEMORY,
+            worker_init_fn=utils.loader_worker_init_fn(dataset),
+        )
+    else:
+        # Create a sampler for multi-process training
+        sampler = utils.create_sampler(dataset, shuffle, cfg)
+        # Create a loader
+        loader = torch.utils.data.DataLoader(
+            dataset,
+            batch_size=batch_size,
+            shuffle=(False if sampler else shuffle),
+            sampler=sampler,
+            num_workers=cfg.DATA_LOADER.NUM_WORKERS,
+            pin_memory=cfg.DATA_LOADER.PIN_MEMORY,
+            drop_last=drop_last,
+            collate_fn=detection_collate if cfg.DETECTION.ENABLE else None,
+            worker_init_fn=utils.loader_worker_init_fn(dataset),
+        )
+    return loader
+
+
+def shuffle_dataset(loader, cur_epoch):
+    """ "
+    Shuffles the data.
+    Args:
+        loader (loader): data loader to perform shuffle.
+        cur_epoch (int): number of the current epoch.
+    """
+    sampler = (
+        loader.batch_sampler.sampler
+        if isinstance(loader.batch_sampler, ShortCycleBatchSampler)
+        else loader.sampler
+    )
+    assert isinstance(
+        sampler, (RandomSampler, DistributedSampler)
+    ), "Sampler type '{}' not supported".format(type(sampler))
+    # RandomSampler handles shuffling automatically
+    if isinstance(sampler, DistributedSampler):
+        # DistributedSampler shuffles data based on epoch
+        sampler.set_epoch(cur_epoch)

TimeSformer/timesformer/datasets/multigrid_helper.py

@@ -0,0 +1,78 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""Helper functions for multigrid training."""
+
+import numpy as np
+from torch._six import int_classes as _int_classes
+from torch.utils.data.sampler import Sampler
+
+
+class ShortCycleBatchSampler(Sampler):
+    """
+    Extend Sampler to support "short cycle" sampling.
+    See paper "A Multigrid Method for Efficiently Training Video Models",
+    Wu et al., 2019 (https://arxiv.org/abs/1912.00998) for details.
+    """
+
+    def __init__(self, sampler, batch_size, drop_last, cfg):
+        if not isinstance(sampler, Sampler):
+            raise ValueError(
+                "sampler should be an instance of "
+                "torch.utils.data.Sampler, but got sampler={}".format(sampler)
+            )
+        if (
+            not isinstance(batch_size, _int_classes)
+            or isinstance(batch_size, bool)
+            or batch_size <= 0
+        ):
+            raise ValueError(
+                "batch_size should be a positive integer value, "
+                "but got batch_size={}".format(batch_size)
+            )
+        if not isinstance(drop_last, bool):
+            raise ValueError(
+                "drop_last should be a boolean value, but got "
+                "drop_last={}".format(drop_last)
+            )
+        self.sampler = sampler
+        self.drop_last = drop_last
+
+        bs_factor = [
+            int(
+                round(
+                    (
+                        float(cfg.DATA.TRAIN_CROP_SIZE)
+                        / (s * cfg.MULTIGRID.DEFAULT_S)
+                    )
+                    ** 2
+                )
+            )
+            for s in cfg.MULTIGRID.SHORT_CYCLE_FACTORS
+        ]
+
+        self.batch_sizes = [
+            batch_size * bs_factor[0],
+            batch_size * bs_factor[1],
+            batch_size,
+        ]
+
+    def __iter__(self):
+        counter = 0
+        batch_size = self.batch_sizes[0]
+        batch = []
+        for idx in self.sampler:
+            batch.append((idx, counter % 3))
+            if len(batch) == batch_size:
+                yield batch
+                counter += 1
+                batch_size = self.batch_sizes[counter % 3]
+                batch = []
+        if len(batch) > 0 and not self.drop_last:
+            yield batch
+
+    def __len__(self):
+        avg_batch_size = sum(self.batch_sizes) / 3.0
+        if self.drop_last:
+            return int(np.floor(len(self.sampler) / avg_batch_size))
+        else:
+            return int(np.ceil(len(self.sampler) / avg_batch_size))

TimeSformer/timesformer/datasets/ssv2.py

@@ -0,0 +1,278 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import json
+import numpy as np
+import os
+import random
+from itertools import chain as chain
+import torch
+import torch.utils.data
+from fvcore.common.file_io import PathManager
+
+import timesformer.utils.logging as logging
+
+from . import utils as utils
+from .build import DATASET_REGISTRY
+
+logger = logging.get_logger(__name__)
+
+
+@DATASET_REGISTRY.register()
+class Ssv2(torch.utils.data.Dataset):
+    """
+    Something-Something v2 (SSV2) video loader. Construct the SSV2 video loader,
+    then sample clips from the videos. For training and validation, a single
+    clip is randomly sampled from every video with random cropping, scaling, and
+    flipping. For testing, multiple clips are uniformaly sampled from every
+    video with uniform cropping. For uniform cropping, we take the left, center,
+    and right crop if the width is larger than height, or take top, center, and
+    bottom crop if the height is larger than the width.
+    """
+
+    def __init__(self, cfg, mode, num_retries=10):
+        """
+        Load Something-Something V2 data (frame paths, labels, etc. ) to a given
+        Dataset object. The dataset could be downloaded from Something-Something
+        official website (https://20bn.com/datasets/something-something).
+        Please see datasets/DATASET.md for more information about the data format.
+        Args:
+            cfg (CfgNode): configs.
+            mode (string): Options includes `train`, `val`, or `test` mode.
+                For the train and val mode, the data loader will take data
+                from the train or val set, and sample one clip per video.
+                For the test mode, the data loader will take data from test set,
+                and sample multiple clips per video.
+            num_retries (int): number of retries for reading frames from disk.
+        """
+        # Only support train, val, and test mode.
+        assert mode in [
+            "train",
+            "val",
+            "test",
+        ], "Split '{}' not supported for Something-Something V2".format(mode)
+        self.mode = mode
+        self.cfg = cfg
+
+        self._video_meta = {}
+        self._num_retries = num_retries
+        # For training or validation mode, one single clip is sampled from every
+        # video. For testing, NUM_ENSEMBLE_VIEWS clips are sampled from every
+        # video. For every clip, NUM_SPATIAL_CROPS is cropped spatially from
+        # the frames.
+        if self.mode in ["train", "val"]:
+            self._num_clips = 1
+        elif self.mode in ["test"]:
+            self._num_clips = (
+                cfg.TEST.NUM_ENSEMBLE_VIEWS * cfg.TEST.NUM_SPATIAL_CROPS
+            )
+
+        logger.info("Constructing Something-Something V2 {}...".format(mode))
+        self._construct_loader()
+
+    def _construct_loader(self):
+        """
+        Construct the video loader.
+        """
+        # Loading label names.
+        with PathManager.open(
+            os.path.join(
+                self.cfg.DATA.PATH_TO_DATA_DIR,
+                "something-something-v2-labels.json",
+            ),
+            "r",
+        ) as f:
+            label_dict = json.load(f)
+
+        # Loading labels.
+        label_file = os.path.join(
+            self.cfg.DATA.PATH_TO_DATA_DIR,
+            "something-something-v2-{}.json".format(
+                "train" if self.mode == "train" else "validation"
+            ),
+        )
+        with PathManager.open(label_file, "r") as f:
+            label_json = json.load(f)
+
+        self._video_names = []
+        self._labels = []
+        for video in label_json:
+            video_name = video["id"]
+            template = video["template"]
+            template = template.replace("[", "")
+            template = template.replace("]", "")
+            label = int(label_dict[template])
+            self._video_names.append(video_name)
+            self._labels.append(label)
+
+        path_to_file = os.path.join(
+            self.cfg.DATA.PATH_TO_DATA_DIR,
+            "{}.csv".format("train" if self.mode == "train" else "val"),
+        )
+        assert PathManager.exists(path_to_file), "{} dir not found".format(
+            path_to_file
+        )
+
+        self._path_to_videos, _ = utils.load_image_lists(
+            path_to_file, self.cfg.DATA.PATH_PREFIX
+        )
+
+        assert len(self._path_to_videos) == len(self._video_names), (
+            len(self._path_to_videos),
+            len(self._video_names),
+        )
+
+
+        # From dict to list.
+        new_paths, new_labels = [], []
+        for index in range(len(self._video_names)):
+            if self._video_names[index] in self._path_to_videos:
+                new_paths.append(self._path_to_videos[self._video_names[index]])
+                new_labels.append(self._labels[index])
+
+        self._labels = new_labels
+        self._path_to_videos = new_paths
+
+        # Extend self when self._num_clips > 1 (during testing).
+        self._path_to_videos = list(
+            chain.from_iterable(
+                [[x] * self._num_clips for x in self._path_to_videos]
+            )
+        )
+        self._labels = list(
+            chain.from_iterable([[x] * self._num_clips for x in self._labels])
+        )
+        self._spatial_temporal_idx = list(
+            chain.from_iterable(
+                [
+                    range(self._num_clips)
+                    for _ in range(len(self._path_to_videos))
+                ]
+            )
+        )
+        logger.info(
+            "Something-Something V2 dataloader constructed "
+            " (size: {}) from {}".format(
+                len(self._path_to_videos), path_to_file
+            )
+        )
+
+    def __getitem__(self, index):
+        """
+        Given the video index, return the list of frames, label, and video
+        index if the video frames can be fetched.
+        Args:
+            index (int): the video index provided by the pytorch sampler.
+        Returns:
+            frames (tensor): the frames of sampled from the video. The dimension
+                is `channel` x `num frames` x `height` x `width`.
+            label (int): the label of the current video.
+            index (int): the index of the video.
+        """
+        short_cycle_idx = None
+        # When short cycle is used, input index is a tupple.
+        if isinstance(index, tuple):
+            index, short_cycle_idx = index
+
+        if self.mode in ["train", "val"]: #or self.cfg.MODEL.ARCH in ['resformer', 'vit']:
+            # -1 indicates random sampling.
+            spatial_sample_index = -1
+            min_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[0]
+            max_scale = self.cfg.DATA.TRAIN_JITTER_SCALES[1]
+            crop_size = self.cfg.DATA.TRAIN_CROP_SIZE
+            if short_cycle_idx in [0, 1]:
+                crop_size = int(
+                    round(
+                        self.cfg.MULTIGRID.SHORT_CYCLE_FACTORS[short_cycle_idx]
+                        * self.cfg.MULTIGRID.DEFAULT_S
+                    )
+                )
+            if self.cfg.MULTIGRID.DEFAULT_S > 0:
+                # Decreasing the scale is equivalent to using a larger "span"
+                # in a sampling grid.
+                min_scale = int(
+                    round(
+                        float(min_scale)
+                        * crop_size
+                        / self.cfg.MULTIGRID.DEFAULT_S
+                    )
+                )
+        elif self.mode in ["test"]:
+            # spatial_sample_index is in [0, 1, 2]. Corresponding to left,
+            # center, or right if width is larger than height, and top, middle,
+            # or bottom if height is larger than width.
+            spatial_sample_index = (
+                self._spatial_temporal_idx[index]
+                % self.cfg.TEST.NUM_SPATIAL_CROPS
+            )
+            if self.cfg.TEST.NUM_SPATIAL_CROPS == 1:
+                spatial_sample_index = 1
+
+            min_scale, max_scale, crop_size = [self.cfg.DATA.TEST_CROP_SIZE] * 3
+            # The testing is deterministic and no jitter should be performed.
+            # min_scale, max_scale, and crop_size are expect to be the same.
+            assert len({min_scale, max_scale, crop_size}) == 1
+        else:
+            raise NotImplementedError(
+                "Does not support {} mode".format(self.mode)
+            )
+
+        label = self._labels[index]
+
+        num_frames = self.cfg.DATA.NUM_FRAMES
+        video_length = len(self._path_to_videos[index])
+
+
+        seg_size = float(video_length - 1) / num_frames
+        seq = []
+        for i in range(num_frames):
+            start = int(np.round(seg_size * i))
+            end = int(np.round(seg_size * (i + 1)))
+            if self.mode == "train":
+                seq.append(random.randint(start, end))
+            else:
+                seq.append((start + end) // 2)
+
+        frames = torch.as_tensor(
+            utils.retry_load_images(
+                [self._path_to_videos[index][frame] for frame in seq],
+                self._num_retries,
+            )
+        )
+
+        # Perform color normalization.
+        frames = utils.tensor_normalize(
+            frames, self.cfg.DATA.MEAN, self.cfg.DATA.STD
+        )
+
+        # T H W C -> C T H W.
+        frames = frames.permute(3, 0, 1, 2)
+        frames = utils.spatial_sampling(
+            frames,
+            spatial_idx=spatial_sample_index,
+            min_scale=min_scale,
+            max_scale=max_scale,
+            crop_size=crop_size,
+            random_horizontal_flip=self.cfg.DATA.RANDOM_FLIP,
+            inverse_uniform_sampling=self.cfg.DATA.INV_UNIFORM_SAMPLE,
+        )
+        #if not self.cfg.RESFORMER.ACTIVE:
+        if not self.cfg.MODEL.ARCH in ['vit']:
+            frames = utils.pack_pathway_output(self.cfg, frames)
+        else:
+            # Perform temporal sampling from the fast pathway.
+            frames = torch.index_select(
+                 frames,
+                 1,
+                 torch.linspace(
+                     0, frames.shape[1] - 1, self.cfg.DATA.NUM_FRAMES
+
+                 ).long(),
+            )
+        return frames, label, index, {}
+
+    def __len__(self):
+        """
+        Returns:
+            (int): the number of videos in the dataset.
+        """
+        return len(self._path_to_videos)

TimeSformer/timesformer/datasets/transform.py

@@ -0,0 +1,459 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import math
+import numpy as np
+import torch
+
+
+def random_short_side_scale_jitter(
+    images, min_size, max_size, boxes=None, inverse_uniform_sampling=False
+):
+    """
+    Perform a spatial short scale jittering on the given images and
+    corresponding boxes.
+    Args:
+        images (tensor): images to perform scale jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+        min_size (int): the minimal size to scale the frames.
+        max_size (int): the maximal size to scale the frames.
+        boxes (ndarray): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+        inverse_uniform_sampling (bool): if True, sample uniformly in
+            [1 / max_scale, 1 / min_scale] and take a reciprocal to get the
+            scale. If False, take a uniform sample from [min_scale, max_scale].
+    Returns:
+        (tensor): the scaled images with dimension of
+            `num frames` x `channel` x `new height` x `new width`.
+        (ndarray or None): the scaled boxes with dimension of
+            `num boxes` x 4.
+    """
+    if inverse_uniform_sampling:
+        size = int(
+            round(1.0 / np.random.uniform(1.0 / max_size, 1.0 / min_size))
+        )
+    else:
+        size = int(round(np.random.uniform(min_size, max_size)))
+
+    height = images.shape[2]
+    width = images.shape[3]
+    if (width <= height and width == size) or (
+        height <= width and height == size
+    ):
+        return images, boxes
+    new_width = size
+    new_height = size
+    if width < height:
+        new_height = int(math.floor((float(height) / width) * size))
+        if boxes is not None:
+            boxes = boxes * float(new_height) / height
+    else:
+        new_width = int(math.floor((float(width) / height) * size))
+        if boxes is not None:
+            boxes = boxes * float(new_width) / width
+
+    return (
+        torch.nn.functional.interpolate(
+            images,
+            size=(new_height, new_width),
+            mode="bilinear",
+            align_corners=False,
+        ),
+        boxes,
+    )
+
+
+def crop_boxes(boxes, x_offset, y_offset):
+    """
+    Peform crop on the bounding boxes given the offsets.
+    Args:
+        boxes (ndarray or None): bounding boxes to peform crop. The dimension
+            is `num boxes` x 4.
+        x_offset (int): cropping offset in the x axis.
+        y_offset (int): cropping offset in the y axis.
+    Returns:
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    cropped_boxes = boxes.copy()
+    cropped_boxes[:, [0, 2]] = boxes[:, [0, 2]] - x_offset
+    cropped_boxes[:, [1, 3]] = boxes[:, [1, 3]] - y_offset
+
+    return cropped_boxes
+
+
+def random_crop(images, size, boxes=None):
+    """
+    Perform random spatial crop on the given images and corresponding boxes.
+    Args:
+        images (tensor): images to perform random crop. The dimension is
+            `num frames` x `channel` x `height` x `width`.
+        size (int): the size of height and width to crop on the image.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        cropped (tensor): cropped images with dimension of
+            `num frames` x `channel` x `size` x `size`.
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    if images.shape[2] == size and images.shape[3] == size:
+        return images, None
+    height = images.shape[2]
+    width = images.shape[3]
+    y_offset = 0
+    if height > size:
+        y_offset = int(np.random.randint(0, height - size))
+    x_offset = 0
+    if width > size:
+        x_offset = int(np.random.randint(0, width - size))
+    cropped = images[
+        :, :, y_offset : y_offset + size, x_offset : x_offset + size
+    ]
+
+    cropped_boxes = (
+        crop_boxes(boxes, x_offset, y_offset) if boxes is not None else None
+    )
+
+    return cropped, cropped_boxes
+
+
+def horizontal_flip(prob, images, boxes=None):
+    """
+    Perform horizontal flip on the given images and corresponding boxes.
+    Args:
+        prob (float): probility to flip the images.
+        images (tensor): images to perform horizontal flip, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        images (tensor): images with dimension of
+            `num frames` x `channel` x `height` x `width`.
+        flipped_boxes (ndarray or None): the flipped boxes with dimension of
+            `num boxes` x 4.
+    """
+    if boxes is None:
+        flipped_boxes = None
+    else:
+        flipped_boxes = boxes.copy()
+
+    if np.random.uniform() < prob:
+        images = images.flip((-1))
+
+        width = images.shape[3]
+        if boxes is not None:
+            flipped_boxes[:, [0, 2]] = width - boxes[:, [2, 0]] - 1
+
+    return images, flipped_boxes
+
+
+def uniform_crop(images, size, spatial_idx, boxes=None):
+    """
+    Perform uniform spatial sampling on the images and corresponding boxes.
+    Args:
+        images (tensor): images to perform uniform crop. The dimension is
+            `num frames` x `channel` x `height` x `width`.
+        size (int): size of height and weight to crop the images.
+        spatial_idx (int): 0, 1, or 2 for left, center, and right crop if width
+            is larger than height. Or 0, 1, or 2 for top, center, and bottom
+            crop if height is larger than width.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        cropped (tensor): images with dimension of
+            `num frames` x `channel` x `size` x `size`.
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    assert spatial_idx in [0, 1, 2]
+    height = images.shape[2]
+    width = images.shape[3]
+
+    y_offset = int(math.ceil((height - size) / 2))
+    x_offset = int(math.ceil((width - size) / 2))
+
+    if height > width:
+        if spatial_idx == 0:
+            y_offset = 0
+        elif spatial_idx == 2:
+            y_offset = height - size
+    else:
+        if spatial_idx == 0:
+            x_offset = 0
+        elif spatial_idx == 2:
+            x_offset = width - size
+    cropped = images[
+        :, :, y_offset : y_offset + size, x_offset : x_offset + size
+    ]
+
+    cropped_boxes = (
+        crop_boxes(boxes, x_offset, y_offset) if boxes is not None else None
+    )
+
+    return cropped, cropped_boxes
+
+
+def uniform_crop_2crops(images, size, spatial_idx, boxes=None):
+    """
+    Perform uniform spatial sampling on the images and corresponding boxes.
+    Args:
+        images (tensor): images to perform uniform crop. The dimension is
+            `num frames` x `channel` x `height` x `width`.
+        size (int): size of height and weight to crop the images.
+        spatial_idx (int): 0, 1, or 2 for left, center, and right crop if width
+            is larger than height. Or 0, 1, or 2 for top, center, and bottom
+            crop if height is larger than width.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+    Returns:
+        cropped (tensor): images with dimension of
+            `num frames` x `channel` x `size` x `size`.
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    assert spatial_idx in [0, 1, 2]
+    height = images.shape[2]
+    width = images.shape[3]
+
+
+    if height > width:
+        x_offset = 0
+        if height > size * 2:
+            if spatial_idx == 0:
+                y_offset = int((height -  size * 2) // 2)
+            elif spatial_idx == 1:
+                y_offset = int(height - size - ((height -  size * 2) // 2))
+        else:
+            if spatial_idx == 0:
+                y_offset = 0
+            elif spatial_idx == 1:
+                y_offset = height - size
+    else:
+        y_offset = 0
+        if width > size * 2:
+            if spatial_idx == 0:
+                x_offset = int((width -  size * 2) // 2)
+            elif spatial_idx == 1:
+                x_offset = int(width - size - ((width -  size * 2) // 2))
+        else:
+            if spatial_idx == 0:
+                x_offset = 0
+            elif spatial_idx == 1:
+                x_offset = width - size
+
+    cropped = images[
+        :, :, y_offset : y_offset + size, x_offset : x_offset + size
+    ]
+
+    cropped_boxes = (
+        crop_boxes(boxes, x_offset, y_offset) if boxes is not None else None
+    )
+
+    return cropped, cropped_boxes
+
+def clip_boxes_to_image(boxes, height, width):
+    """
+    Clip an array of boxes to an image with the given height and width.
+    Args:
+        boxes (ndarray): bounding boxes to perform clipping.
+            Dimension is `num boxes` x 4.
+        height (int): given image height.
+        width (int): given image width.
+    Returns:
+        clipped_boxes (ndarray): the clipped boxes with dimension of
+            `num boxes` x 4.
+    """
+    clipped_boxes = boxes.copy()
+    clipped_boxes[:, [0, 2]] = np.minimum(
+        width - 1.0, np.maximum(0.0, boxes[:, [0, 2]])
+    )
+    clipped_boxes[:, [1, 3]] = np.minimum(
+        height - 1.0, np.maximum(0.0, boxes[:, [1, 3]])
+    )
+    return clipped_boxes
+
+
+def blend(images1, images2, alpha):
+    """
+    Blend two images with a given weight alpha.
+    Args:
+        images1 (tensor): the first images to be blended, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+        images2 (tensor): the second images to be blended, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+        alpha (float): the blending weight.
+    Returns:
+        (tensor): blended images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    return images1 * alpha + images2 * (1 - alpha)
+
+
+def grayscale(images):
+    """
+    Get the grayscale for the input images. The channels of images should be
+    in order BGR.
+    Args:
+        images (tensor): the input images for getting grayscale. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+    Returns:
+        img_gray (tensor): blended images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    # R -> 0.299, G -> 0.587, B -> 0.114.
+    img_gray = torch.tensor(images)
+    gray_channel = (
+        0.299 * images[:, 2] + 0.587 * images[:, 1] + 0.114 * images[:, 0]
+    )
+    img_gray[:, 0] = gray_channel
+    img_gray[:, 1] = gray_channel
+    img_gray[:, 2] = gray_channel
+    return img_gray
+
+
+def color_jitter(images, img_brightness=0, img_contrast=0, img_saturation=0):
+    """
+    Perfrom a color jittering on the input images. The channels of images
+    should be in order BGR.
+    Args:
+        images (tensor): images to perform color jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+        img_brightness (float): jitter ratio for brightness.
+        img_contrast (float): jitter ratio for contrast.
+        img_saturation (float): jitter ratio for saturation.
+    Returns:
+        images (tensor): the jittered images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+
+    jitter = []
+    if img_brightness != 0:
+        jitter.append("brightness")
+    if img_contrast != 0:
+        jitter.append("contrast")
+    if img_saturation != 0:
+        jitter.append("saturation")
+
+    if len(jitter) > 0:
+        order = np.random.permutation(np.arange(len(jitter)))
+        for idx in range(0, len(jitter)):
+            if jitter[order[idx]] == "brightness":
+                images = brightness_jitter(img_brightness, images)
+            elif jitter[order[idx]] == "contrast":
+                images = contrast_jitter(img_contrast, images)
+            elif jitter[order[idx]] == "saturation":
+                images = saturation_jitter(img_saturation, images)
+    return images
+
+
+def brightness_jitter(var, images):
+    """
+    Perfrom brightness jittering on the input images. The channels of images
+    should be in order BGR.
+    Args:
+        var (float): jitter ratio for brightness.
+        images (tensor): images to perform color jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+    Returns:
+        images (tensor): the jittered images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+
+    img_bright = torch.zeros(images.shape)
+    images = blend(images, img_bright, alpha)
+    return images
+
+
+def contrast_jitter(var, images):
+    """
+    Perfrom contrast jittering on the input images. The channels of images
+    should be in order BGR.
+    Args:
+        var (float): jitter ratio for contrast.
+        images (tensor): images to perform color jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+    Returns:
+        images (tensor): the jittered images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+
+    img_gray = grayscale(images)
+    img_gray[:] = torch.mean(img_gray, dim=(1, 2, 3), keepdim=True)
+    images = blend(images, img_gray, alpha)
+    return images
+
+
+def saturation_jitter(var, images):
+    """
+    Perfrom saturation jittering on the input images. The channels of images
+    should be in order BGR.
+    Args:
+        var (float): jitter ratio for saturation.
+        images (tensor): images to perform color jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+    Returns:
+        images (tensor): the jittered images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    alpha = 1.0 + np.random.uniform(-var, var)
+    img_gray = grayscale(images)
+    images = blend(images, img_gray, alpha)
+
+    return images
+
+
+def lighting_jitter(images, alphastd, eigval, eigvec):
+    """
+    Perform AlexNet-style PCA jitter on the given images.
+    Args:
+        images (tensor): images to perform lighting jitter. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+        alphastd (float): jitter ratio for PCA jitter.
+        eigval (list): eigenvalues for PCA jitter.
+        eigvec (list[list]): eigenvectors for PCA jitter.
+    Returns:
+        out_images (tensor): the jittered images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    if alphastd == 0:
+        return images
+    # generate alpha1, alpha2, alpha3.
+    alpha = np.random.normal(0, alphastd, size=(1, 3))
+    eig_vec = np.array(eigvec)
+    eig_val = np.reshape(eigval, (1, 3))
+    rgb = np.sum(
+        eig_vec * np.repeat(alpha, 3, axis=0) * np.repeat(eig_val, 3, axis=0),
+        axis=1,
+    )
+    out_images = torch.zeros_like(images)
+    for idx in range(images.shape[1]):
+        out_images[:, idx] = images[:, idx] + rgb[2 - idx]
+
+    return out_images
+
+
+def color_normalization(images, mean, stddev):
+    """
+    Perform color nomration on the given images.
+    Args:
+        images (tensor): images to perform color normalization. Dimension is
+            `num frames` x `channel` x `height` x `width`.
+        mean (list): mean values for normalization.
+        stddev (list): standard deviations for normalization.
+
+    Returns:
+        out_images (tensor): the noramlized images, the dimension is
+            `num frames` x `channel` x `height` x `width`.
+    """
+    assert len(mean) == images.shape[1], "channel mean not computed properly"
+    assert (
+        len(stddev) == images.shape[1]
+    ), "channel stddev not computed properly"
+
+    out_images = torch.zeros_like(images)
+    for idx in range(len(mean)):
+        out_images[:, idx] = (images[:, idx] - mean[idx]) / stddev[idx]
+
+    return out_images

TimeSformer/timesformer/datasets/utils.py

@@ -0,0 +1,380 @@
+#!/usr/bin/env python3
+
+import logging
+import numpy as np
+import os
+import random
+import time
+from collections import defaultdict
+import cv2
+import torch
+from fvcore.common.file_io import PathManager
+from torch.utils.data.distributed import DistributedSampler
+
+from . import transform as transform
+
+logger = logging.getLogger(__name__)
+
+
+def retry_load_images(image_paths, retry=10, backend="pytorch"):
+    """
+    This function is to load images with support of retrying for failed load.
+
+    Args:
+        image_paths (list): paths of images needed to be loaded.
+        retry (int, optional): maximum time of loading retrying. Defaults to 10.
+        backend (str): `pytorch` or `cv2`.
+
+    Returns:
+        imgs (list): list of loaded images.
+    """
+    for i in range(retry):
+        imgs = []
+        for image_path in image_paths:
+            with PathManager.open(image_path, "rb") as f:
+                img_str = np.frombuffer(f.read(), np.uint8)
+                img = cv2.imdecode(img_str, flags=cv2.IMREAD_COLOR)
+            imgs.append(img)
+
+        if all(img is not None for img in imgs):
+            if backend == "pytorch":
+                imgs = torch.as_tensor(np.stack(imgs))
+            return imgs
+        else:
+            logger.warn("Reading failed. Will retry.")
+            time.sleep(1.0)
+        if i == retry - 1:
+            raise Exception("Failed to load images {}".format(image_paths))
+
+
+def get_sequence(center_idx, half_len, sample_rate, num_frames):
+    """
+    Sample frames among the corresponding clip.
+
+    Args:
+        center_idx (int): center frame idx for current clip
+        half_len (int): half of the clip length
+        sample_rate (int): sampling rate for sampling frames inside of the clip
+        num_frames (int): number of expected sampled frames
+
+    Returns:
+        seq (list): list of indexes of sampled frames in this clip.
+    """
+    seq = list(range(center_idx - half_len, center_idx + half_len, sample_rate))
+
+    for seq_idx in range(len(seq)):
+        if seq[seq_idx] < 0:
+            seq[seq_idx] = 0
+        elif seq[seq_idx] >= num_frames:
+            seq[seq_idx] = num_frames - 1
+    return seq
+
+
+def pack_pathway_output(cfg, frames):
+    """
+    Prepare output as a list of tensors. Each tensor corresponding to a
+    unique pathway.
+    Args:
+        frames (tensor): frames of images sampled from the video. The
+            dimension is `channel` x `num frames` x `height` x `width`.
+    Returns:
+        frame_list (list): list of tensors with the dimension of
+            `channel` x `num frames` x `height` x `width`.
+    """
+    if cfg.DATA.REVERSE_INPUT_CHANNEL:
+        frames = frames[[2, 1, 0], :, :, :]
+    if cfg.MODEL.ARCH in cfg.MODEL.SINGLE_PATHWAY_ARCH:
+        frame_list = [frames]
+    elif cfg.MODEL.ARCH in cfg.MODEL.MULTI_PATHWAY_ARCH:
+        fast_pathway = frames
+        # Perform temporal sampling from the fast pathway.
+        slow_pathway = torch.index_select(
+            frames,
+            1,
+            torch.linspace(
+                0, frames.shape[1] - 1, frames.shape[1] // cfg.SLOWFAST.ALPHA
+            ).long(),
+        )
+        frame_list = [slow_pathway, fast_pathway]
+    else:
+        raise NotImplementedError(
+            "Model arch {} is not in {}".format(
+                cfg.MODEL.ARCH,
+                cfg.MODEL.SINGLE_PATHWAY_ARCH + cfg.MODEL.MULTI_PATHWAY_ARCH,
+            )
+        )
+    return frame_list
+
+
+def spatial_sampling(
+    frames,
+    spatial_idx=-1,
+    min_scale=256,
+    max_scale=320,
+    crop_size=224,
+    random_horizontal_flip=True,
+    inverse_uniform_sampling=False,
+):
+    """
+    Perform spatial sampling on the given video frames. If spatial_idx is
+    -1, perform random scale, random crop, and random flip on the given
+    frames. If spatial_idx is 0, 1, or 2, perform spatial uniform sampling
+    with the given spatial_idx.
+    Args:
+        frames (tensor): frames of images sampled from the video. The
+            dimension is `num frames` x `height` x `width` x `channel`.
+        spatial_idx (int): if -1, perform random spatial sampling. If 0, 1,
+            or 2, perform left, center, right crop if width is larger than
+            height, and perform top, center, buttom crop if height is larger
+            than width.
+        min_scale (int): the minimal size of scaling.
+        max_scale (int): the maximal size of scaling.
+        crop_size (int): the size of height and width used to crop the
+            frames.
+        inverse_uniform_sampling (bool): if True, sample uniformly in
+            [1 / max_scale, 1 / min_scale] and take a reciprocal to get the
+            scale. If False, take a uniform sample from [min_scale,
+            max_scale].
+    Returns:
+        frames (tensor): spatially sampled frames.
+    """
+    assert spatial_idx in [-1, 0, 1, 2]
+    if spatial_idx == -1:
+        frames, _ = transform.random_short_side_scale_jitter(
+            images=frames,
+            min_size=min_scale,
+            max_size=max_scale,
+            inverse_uniform_sampling=inverse_uniform_sampling,
+        )
+        frames, _ = transform.random_crop(frames, crop_size)
+        if random_horizontal_flip:
+            frames, _ = transform.horizontal_flip(0.5, frames)
+    else:
+        # The testing is deterministic and no jitter should be performed.
+        # min_scale, max_scale, and crop_size are expect to be the same.
+        #assert len({min_scale, max_scale, crop_size}) == 1
+        frames, _ = transform.random_short_side_scale_jitter(
+            frames, min_scale, max_scale
+        )
+        frames, _ = transform.uniform_crop(frames, crop_size, spatial_idx)
+    return frames
+
+def spatial_sampling_2crops(
+    frames,
+    spatial_idx=-1,
+    min_scale=256,
+    max_scale=320,
+    crop_size=224,
+    random_horizontal_flip=True,
+    inverse_uniform_sampling=False,
+):
+    """
+    Perform spatial sampling on the given video frames. If spatial_idx is
+    -1, perform random scale, random crop, and random flip on the given
+    frames. If spatial_idx is 0, 1, or 2, perform spatial uniform sampling
+    with the given spatial_idx.
+    Args:
+        frames (tensor): frames of images sampled from the video. The
+            dimension is `num frames` x `height` x `width` x `channel`.
+        spatial_idx (int): if -1, perform random spatial sampling. If 0, 1,
+            or 2, perform left, center, right crop if width is larger than
+            height, and perform top, center, buttom crop if height is larger
+            than width.
+        min_scale (int): the minimal size of scaling.
+        max_scale (int): the maximal size of scaling.
+        crop_size (int): the size of height and width used to crop the
+            frames.
+        inverse_uniform_sampling (bool): if True, sample uniformly in
+            [1 / max_scale, 1 / min_scale] and take a reciprocal to get the
+            scale. If False, take a uniform sample from [min_scale,
+            max_scale].
+    Returns:
+        frames (tensor): spatially sampled frames.
+    """
+    assert spatial_idx in [-1, 0, 1, 2]
+    if spatial_idx == -1:
+        frames, _ = transform.random_short_side_scale_jitter(
+            images=frames,
+            min_size=min_scale,
+            max_size=max_scale,
+            inverse_uniform_sampling=inverse_uniform_sampling,
+        )
+        frames, _ = transform.random_crop(frames, crop_size)
+        if random_horizontal_flip:
+            frames, _ = transform.horizontal_flip(0.5, frames)
+    else:
+        # The testing is deterministic and no jitter should be performed.
+        # min_scale, max_scale, and crop_size are expect to be the same.
+        #assert len({min_scale, max_scale, crop_size}) == 1
+        frames, _ = transform.random_short_side_scale_jitter(
+            frames, min_scale, max_scale
+        )
+        frames, _ = transform.uniform_crop_2crops(frames, crop_size, spatial_idx)
+    return frames
+
+
+def as_binary_vector(labels, num_classes):
+    """
+    Construct binary label vector given a list of label indices.
+    Args:
+        labels (list): The input label list.
+        num_classes (int): Number of classes of the label vector.
+    Returns:
+        labels (numpy array): the resulting binary vector.
+    """
+    label_arr = np.zeros((num_classes,))
+
+    for lbl in set(labels):
+        label_arr[lbl] = 1.0
+    return label_arr
+
+
+def aggregate_labels(label_list):
+    """
+    Join a list of label list.
+    Args:
+        labels (list): The input label list.
+    Returns:
+        labels (list): The joint list of all lists in input.
+    """
+    all_labels = []
+    for labels in label_list:
+        for l in labels:
+            all_labels.append(l)
+    return list(set(all_labels))
+
+
+def convert_to_video_level_labels(labels):
+    """
+    Aggregate annotations from all frames of a video to form video-level labels.
+    Args:
+        labels (list): The input label list.
+    Returns:
+        labels (list): Same as input, but with each label replaced by
+        a video-level one.
+    """
+    for video_id in range(len(labels)):
+        video_level_labels = aggregate_labels(labels[video_id])
+        for i in range(len(labels[video_id])):
+            labels[video_id][i] = video_level_labels
+    return labels
+
+
+def load_image_lists(frame_list_file, prefix="", return_list=False):
+    """
+    Load image paths and labels from a "frame list".
+    Each line of the frame list contains:
+    `original_vido_id video_id frame_id path labels`
+    Args:
+        frame_list_file (string): path to the frame list.
+        prefix (str): the prefix for the path.
+        return_list (bool): if True, return a list. If False, return a dict.
+    Returns:
+        image_paths (list or dict): list of list containing path to each frame.
+            If return_list is False, then return in a dict form.
+        labels (list or dict): list of list containing label of each frame.
+            If return_list is False, then return in a dict form.
+    """
+    image_paths = defaultdict(list)
+    labels = defaultdict(list)
+    with PathManager.open(frame_list_file, "r") as f:
+        assert f.readline().startswith("original_vido_id")
+        for line in f:
+            row = line.split()
+            # original_vido_id video_id frame_id path labels
+            assert len(row) == 5
+            video_name = row[0]
+            if prefix == "":
+                path = row[3]
+            else:
+                path = os.path.join(prefix, row[3])
+            image_paths[video_name].append(path)
+            frame_labels = row[-1].replace('"', "")
+            if frame_labels != "":
+                labels[video_name].append(
+                    [int(x) for x in frame_labels.split(",")]
+                )
+            else:
+                labels[video_name].append([])
+
+    if return_list:
+        keys = image_paths.keys()
+        image_paths = [image_paths[key] for key in keys]
+        labels = [labels[key] for key in keys]
+        return image_paths, labels
+    return dict(image_paths), dict(labels)
+
+
+def tensor_normalize(tensor, mean, std):
+    """
+    Normalize a given tensor by subtracting the mean and dividing the std.
+    Args:
+        tensor (tensor): tensor to normalize.
+        mean (tensor or list): mean value to subtract.
+        std (tensor or list): std to divide.
+    """
+    if tensor.dtype == torch.uint8:
+        tensor = tensor.float()
+        tensor = tensor / 255.0
+    if type(mean) == list:
+        mean = torch.tensor(mean)
+    if type(std) == list:
+        std = torch.tensor(std)
+    tensor = tensor - mean
+    tensor = tensor / std
+    return tensor
+
+
+def get_random_sampling_rate(long_cycle_sampling_rate, sampling_rate):
+    """
+    When multigrid training uses a fewer number of frames, we randomly
+    increase the sampling rate so that some clips cover the original span.
+    """
+    if long_cycle_sampling_rate > 0:
+        assert long_cycle_sampling_rate >= sampling_rate
+        return random.randint(sampling_rate, long_cycle_sampling_rate)
+    else:
+        return sampling_rate
+
+
+def revert_tensor_normalize(tensor, mean, std):
+    """
+    Revert normalization for a given tensor by multiplying by the std and adding the mean.
+    Args:
+        tensor (tensor): tensor to revert normalization.
+        mean (tensor or list): mean value to add.
+        std (tensor or list): std to multiply.
+    """
+    if type(mean) == list:
+        mean = torch.tensor(mean)
+    if type(std) == list:
+        std = torch.tensor(std)
+    tensor = tensor * std
+    tensor = tensor + mean
+    return tensor
+
+
+def create_sampler(dataset, shuffle, cfg):
+    """
+    Create sampler for the given dataset.
+    Args:
+        dataset (torch.utils.data.Dataset): the given dataset.
+        shuffle (bool): set to ``True`` to have the data reshuffled
+            at every epoch.
+        cfg (CfgNode): configs. Details can be found in
+            slowfast/config/defaults.py
+    Returns:
+        sampler (Sampler): the created sampler.
+    """
+    sampler = DistributedSampler(dataset) if cfg.NUM_GPUS > 1 else None
+
+    return sampler
+
+
+def loader_worker_init_fn(dataset):
+    """
+    Create init function passed to pytorch data loader.
+    Args:
+        dataset (torch.utils.data.Dataset): the given dataset.
+    """
+    return None

TimeSformer/timesformer/datasets/video_container.py

@@ -0,0 +1,31 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import av
+
+
+def get_video_container(path_to_vid, multi_thread_decode=False, backend="pyav"):
+    """
+    Given the path to the video, return the pyav video container.
+    Args:
+        path_to_vid (str): path to the video.
+        multi_thread_decode (bool): if True, perform multi-thread decoding.
+        backend (str): decoder backend, options include `pyav` and
+            `torchvision`, default is `pyav`.
+    Returns:
+        container (container): video container.
+    """
+    if backend == "torchvision":
+        with open(path_to_vid, "rb") as fp:
+            container = fp.read()
+        return container
+    elif backend == "pyav":
+        #try:
+        container = av.open(path_to_vid)
+        if multi_thread_decode:
+            # Enable multiple threads for decoding.
+            container.streams.video[0].thread_type = "AUTO"
+        #except:
+        #  container = None
+        return container
+    else:
+        raise NotImplementedError("Unknown backend {}".format(backend))

TimeSformer/timesformer/models/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+from .build import MODEL_REGISTRY, build_model  # noqa
+from .custom_video_model_builder import *  # noqa
+from .video_model_builder import ResNet, SlowFast # noqa

TimeSformer/timesformer/models/batchnorm_helper.py

@@ -0,0 +1,217 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""BatchNorm (BN) utility functions and custom batch-size BN implementations"""
+
+from functools import partial
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+from torch.autograd.function import Function
+
+import timesformer.utils.distributed as du
+
+
+def get_norm(cfg):
+    """
+    Args:
+        cfg (CfgNode): model building configs, details are in the comments of
+            the config file.
+    Returns:
+        nn.Module: the normalization layer.
+    """
+    if cfg.BN.NORM_TYPE == "batchnorm":
+        return nn.BatchNorm3d
+    elif cfg.BN.NORM_TYPE == "sub_batchnorm":
+        return partial(SubBatchNorm3d, num_splits=cfg.BN.NUM_SPLITS)
+    elif cfg.BN.NORM_TYPE == "sync_batchnorm":
+        return partial(
+            NaiveSyncBatchNorm3d, num_sync_devices=cfg.BN.NUM_SYNC_DEVICES
+        )
+    else:
+        raise NotImplementedError(
+            "Norm type {} is not supported".format(cfg.BN.NORM_TYPE)
+        )
+
+
+class SubBatchNorm3d(nn.Module):
+    """
+    The standard BN layer computes stats across all examples in a GPU. In some
+    cases it is desirable to compute stats across only a subset of examples
+    (e.g., in multigrid training https://arxiv.org/abs/1912.00998).
+    SubBatchNorm3d splits the batch dimension into N splits, and run BN on
+    each of them separately (so that the stats are computed on each subset of
+    examples (1/N of batch) independently. During evaluation, it aggregates
+    the stats from all splits into one BN.
+    """
+
+    def __init__(self, num_splits, **args):
+        """
+        Args:
+            num_splits (int): number of splits.
+            args (list): other arguments.
+        """
+        super(SubBatchNorm3d, self).__init__()
+        self.num_splits = num_splits
+        num_features = args["num_features"]
+        # Keep only one set of weight and bias.
+        if args.get("affine", True):
+            self.affine = True
+            args["affine"] = False
+            self.weight = torch.nn.Parameter(torch.ones(num_features))
+            self.bias = torch.nn.Parameter(torch.zeros(num_features))
+        else:
+            self.affine = False
+        self.bn = nn.BatchNorm3d(**args)
+        args["num_features"] = num_features * num_splits
+        self.split_bn = nn.BatchNorm3d(**args)
+
+    def _get_aggregated_mean_std(self, means, stds, n):
+        """
+        Calculate the aggregated mean and stds.
+        Args:
+            means (tensor): mean values.
+            stds (tensor): standard deviations.
+            n (int): number of sets of means and stds.
+        """
+        mean = means.view(n, -1).sum(0) / n
+        std = (
+            stds.view(n, -1).sum(0) / n
+            + ((means.view(n, -1) - mean) ** 2).view(n, -1).sum(0) / n
+        )
+        return mean.detach(), std.detach()
+
+    def aggregate_stats(self):
+        """
+        Synchronize running_mean, and running_var. Call this before eval.
+        """
+        if self.split_bn.track_running_stats:
+            (
+                self.bn.running_mean.data,
+                self.bn.running_var.data,
+            ) = self._get_aggregated_mean_std(
+                self.split_bn.running_mean,
+                self.split_bn.running_var,
+                self.num_splits,
+            )
+
+    def forward(self, x):
+        if self.training:
+            n, c, t, h, w = x.shape
+            x = x.view(n // self.num_splits, c * self.num_splits, t, h, w)
+            x = self.split_bn(x)
+            x = x.view(n, c, t, h, w)
+        else:
+            x = self.bn(x)
+        if self.affine:
+            x = x * self.weight.view((-1, 1, 1, 1))
+            x = x + self.bias.view((-1, 1, 1, 1))
+        return x
+
+
+class GroupGather(Function):
+    """
+    GroupGather performs all gather on each of the local process/ GPU groups.
+    """
+
+    @staticmethod
+    def forward(ctx, input, num_sync_devices, num_groups):
+        """
+        Perform forwarding, gathering the stats across different process/ GPU
+        group.
+        """
+        ctx.num_sync_devices = num_sync_devices
+        ctx.num_groups = num_groups
+
+        input_list = [
+            torch.zeros_like(input) for k in range(du.get_local_size())
+        ]
+        dist.all_gather(
+            input_list, input, async_op=False, group=du._LOCAL_PROCESS_GROUP
+        )
+
+        inputs = torch.stack(input_list, dim=0)
+        if num_groups > 1:
+            rank = du.get_local_rank()
+            group_idx = rank // num_sync_devices
+            inputs = inputs[
+                group_idx
+                * num_sync_devices : (group_idx + 1)
+                * num_sync_devices
+            ]
+        inputs = torch.sum(inputs, dim=0)
+        return inputs
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        """
+        Perform backwarding, gathering the gradients across different process/ GPU
+        group.
+        """
+        grad_output_list = [
+            torch.zeros_like(grad_output) for k in range(du.get_local_size())
+        ]
+        dist.all_gather(
+            grad_output_list,
+            grad_output,
+            async_op=False,
+            group=du._LOCAL_PROCESS_GROUP,
+        )
+
+        grads = torch.stack(grad_output_list, dim=0)
+        if ctx.num_groups > 1:
+            rank = du.get_local_rank()
+            group_idx = rank // ctx.num_sync_devices
+            grads = grads[
+                group_idx
+                * ctx.num_sync_devices : (group_idx + 1)
+                * ctx.num_sync_devices
+            ]
+        grads = torch.sum(grads, dim=0)
+        return grads, None, None
+
+
+class NaiveSyncBatchNorm3d(nn.BatchNorm3d):
+    def __init__(self, num_sync_devices, **args):
+        """
+        Naive version of Synchronized 3D BatchNorm.
+        Args:
+            num_sync_devices (int): number of device to sync.
+            args (list): other arguments.
+        """
+        self.num_sync_devices = num_sync_devices
+        if self.num_sync_devices > 0:
+            assert du.get_local_size() % self.num_sync_devices == 0, (
+                du.get_local_size(),
+                self.num_sync_devices,
+            )
+            self.num_groups = du.get_local_size() // self.num_sync_devices
+        else:
+            self.num_sync_devices = du.get_local_size()
+            self.num_groups = 1
+        super(NaiveSyncBatchNorm3d, self).__init__(**args)
+
+    def forward(self, input):
+        if du.get_local_size() == 1 or not self.training:
+            return super().forward(input)
+
+        assert input.shape[0] > 0, "SyncBatchNorm does not support empty inputs"
+        C = input.shape[1]
+        mean = torch.mean(input, dim=[0, 2, 3, 4])
+        meansqr = torch.mean(input * input, dim=[0, 2, 3, 4])
+
+        vec = torch.cat([mean, meansqr], dim=0)
+        vec = GroupGather.apply(vec, self.num_sync_devices, self.num_groups) * (
+            1.0 / self.num_sync_devices
+        )
+
+        mean, meansqr = torch.split(vec, C)
+        var = meansqr - mean * mean
+        self.running_mean += self.momentum * (mean.detach() - self.running_mean)
+        self.running_var += self.momentum * (var.detach() - self.running_var)
+
+        invstd = torch.rsqrt(var + self.eps)
+        scale = self.weight * invstd
+        bias = self.bias - mean * scale
+        scale = scale.reshape(1, -1, 1, 1, 1)
+        bias = bias.reshape(1, -1, 1, 1, 1)
+        return input * scale + bias

TimeSformer/timesformer/models/build.py

@@ -0,0 +1,54 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""Model construction functions."""
+
+import torch
+from fvcore.common.registry import Registry
+
+MODEL_REGISTRY = Registry("MODEL")
+MODEL_REGISTRY.__doc__ = """
+Registry for video model.
+
+The registered object will be called with `obj(cfg)`.
+The call should return a `torch.nn.Module` object.
+"""
+
+
+def build_model(cfg, gpu_id=None):
+    """
+    Builds the video model.
+    Args:
+        cfg (configs): configs that contains the hyper-parameters to build the
+        backbone. Details can be seen in slowfast/config/defaults.py.
+        gpu_id (Optional[int]): specify the gpu index to build model.
+    """
+    if torch.cuda.is_available():
+        assert (
+            cfg.NUM_GPUS <= torch.cuda.device_count()
+        ), "Cannot use more GPU devices than available"
+    else:
+        assert (
+            cfg.NUM_GPUS == 0
+        ), "Cuda is not available. Please set `NUM_GPUS: 0 for running on CPUs."
+
+    # Construct the model
+    name = cfg.MODEL.MODEL_NAME
+    model = MODEL_REGISTRY.get(name)(cfg)
+
+    if cfg.NUM_GPUS:
+        if gpu_id is None:
+            # Determine the GPU used by the current process
+            cur_device = torch.cuda.current_device()
+        else:
+            cur_device = gpu_id
+        # Transfer the model to the current GPU device
+        model = model.cuda(device=cur_device)
+
+
+    # Use multi-process data parallel model in the multi-gpu setting
+    if cfg.NUM_GPUS > 1:
+        # Make model replica operate on the current device
+        model = torch.nn.parallel.DistributedDataParallel(
+            module=model, device_ids=[cur_device], output_device=cur_device
+        )
+    return model

TimeSformer/timesformer/models/conv2d_same.py

@@ -0,0 +1,74 @@
+# Copyright 2020 Ross Wightman
+# Conv2d w/ Same Padding
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Tuple, Optional
+
+import math
+from typing import List, Tuple
+#from .padding import pad_same, get_padding_value
+
+# Dynamically pad input x with 'SAME' padding for conv with specified args
+def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):
+    ih, iw = x.size()[-2:]
+    pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(iw, k[1], s[1], d[1])
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2], value=value)
+    return x
+
+# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution
+def get_same_padding(x: int, k: int, s: int, d: int):
+    return max((math.ceil(x / s) - 1) * s + (k - 1) * d + 1 - x, 0)
+
+def get_padding_value(padding, kernel_size, **kwargs) -> Tuple[Tuple, bool]:
+    dynamic = False
+    if isinstance(padding, str):
+        # for any string padding, the padding will be calculated for you, one of three ways
+        padding = padding.lower()
+        if padding == 'same':
+            # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact
+            if is_static_pad(kernel_size, **kwargs):
+                # static case, no extra overhead
+                padding = get_padding(kernel_size, **kwargs)
+            else:
+                # dynamic 'SAME' padding, has runtime/GPU memory overhead
+                padding = 0
+                dynamic = True
+        elif padding == 'valid':
+            # 'VALID' padding, same as padding=0
+            padding = 0
+        else:
+            # Default to PyTorch style 'same'-ish symmetric padding
+            padding = get_padding(kernel_size, **kwargs)
+    return padding, dynamic
+
+def conv2d_same(
+        x, weight: torch.Tensor, bias: Optional[torch.Tensor] = None, stride: Tuple[int, int] = (1, 1),
+        padding: Tuple[int, int] = (0, 0), dilation: Tuple[int, int] = (1, 1), groups: int = 1):
+    x = pad_same(x, weight.shape[-2:], stride, dilation)
+    return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)
+
+
+class Conv2dSame(nn.Conv2d):
+    """ Tensorflow like 'SAME' convolution wrapper for 2D convolutions
+    """
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1,
+                 padding=0, dilation=1, groups=1, bias=True):
+        super(Conv2dSame, self).__init__(
+            in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias)
+
+    def forward(self, x):
+        return conv2d_same(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
+
+
+def create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):
+    padding = kwargs.pop('padding', '')
+    kwargs.setdefault('bias', False)
+    padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)
+    if is_dynamic:
+        return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)
+    else:
+        return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)

TimeSformer/timesformer/models/custom_video_model_builder.py

@@ -0,0 +1,4 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+
+"""A More Flexible Video models."""

TimeSformer/timesformer/models/features.py

@@ -0,0 +1,266 @@
+# Copyright 2020 Ross Wightman
+
+from collections import OrderedDict, defaultdict
+from copy import deepcopy
+from functools import partial
+from typing import Dict, List, Tuple
+
+import torch
+import torch.nn as nn
+
+
+class FeatureInfo:
+
+    def __init__(self, feature_info: List[Dict], out_indices: Tuple[int]):
+        prev_reduction = 1
+        for fi in feature_info:
+            # sanity check the mandatory fields, there may be additional fields depending on the model
+            assert 'num_chs' in fi and fi['num_chs'] > 0
+            assert 'reduction' in fi and fi['reduction'] >= prev_reduction
+            prev_reduction = fi['reduction']
+            assert 'module' in fi
+        self.out_indices = out_indices
+        self.info = feature_info
+
+    def from_other(self, out_indices: Tuple[int]):
+        return FeatureInfo(deepcopy(self.info), out_indices)
+
+    def get(self, key, idx=None):
+        """ Get value by key at specified index (indices)
+        if idx == None, returns value for key at each output index
+        if idx is an integer, return value for that feature module index (ignoring output indices)
+        if idx is a list/tupple, return value for each module index (ignoring output indices)
+        """
+        if idx is None:
+            return [self.info[i][key] for i in self.out_indices]
+        if isinstance(idx, (tuple, list)):
+            return [self.info[i][key] for i in idx]
+        else:
+            return self.info[idx][key]
+
+    def get_dicts(self, keys=None, idx=None):
+        """ return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)
+        """
+        if idx is None:
+            if keys is None:
+                return [self.info[i] for i in self.out_indices]
+            else:
+                return [{k: self.info[i][k] for k in keys} for i in self.out_indices]
+        if isinstance(idx, (tuple, list)):
+            return [self.info[i] if keys is None else {k: self.info[i][k] for k in keys} for i in idx]
+        else:
+            return self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}
+
+    def channels(self, idx=None):
+        """ feature channels accessor
+        """
+        return self.get('num_chs', idx)
+
+    def reduction(self, idx=None):
+        """ feature reduction (output stride) accessor
+        """
+        return self.get('reduction', idx)
+
+    def module_name(self, idx=None):
+        """ feature module name accessor
+        """
+        return self.get('module', idx)
+
+    def __getitem__(self, item):
+        return self.info[item]
+
+    def __len__(self):
+        return len(self.info)
+
+
+class FeatureHooks:
+    """ Feature Hook Helper
+    This module helps with the setup and extraction of hooks for extracting features from
+    internal nodes in a model by node name. This works quite well in eager Python but needs
+    redesign for torcscript.
+    """
+
+    def __init__(self, hooks, named_modules, out_map=None, default_hook_type='forward'):
+        # setup feature hooks
+        modules = {k: v for k, v in named_modules}
+        for i, h in enumerate(hooks):
+            hook_name = h['module']
+            m = modules[hook_name]
+            hook_id = out_map[i] if out_map else hook_name
+            hook_fn = partial(self._collect_output_hook, hook_id)
+            hook_type = h['hook_type'] if 'hook_type' in h else default_hook_type
+            if hook_type == 'forward_pre':
+                m.register_forward_pre_hook(hook_fn)
+            elif hook_type == 'forward':
+                m.register_forward_hook(hook_fn)
+            else:
+                assert False, "Unsupported hook type"
+        self._feature_outputs = defaultdict(OrderedDict)
+
+    def _collect_output_hook(self, hook_id, *args):
+        x = args[-1]  # tensor we want is last argument, output for fwd, input for fwd_pre
+        if isinstance(x, tuple):
+            x = x[0]  # unwrap input tuple
+        self._feature_outputs[x.device][hook_id] = x
+
+    def get_output(self, device) -> Dict[str, torch.tensor]:
+        output = self._feature_outputs[device]
+        self._feature_outputs[device] = OrderedDict()  # clear after reading
+        return output
+
+
+def _module_list(module, flatten_sequential=False):
+    # a yield/iter would be better for this but wouldn't be compatible with torchscript
+    ml = []
+    for name, module in module.named_children():
+        if flatten_sequential and isinstance(module, nn.Sequential):
+            # first level of Sequential containers is flattened into containing model
+            for child_name, child_module in module.named_children():
+                combined = [name, child_name]
+                ml.append(('_'.join(combined), '.'.join(combined), child_module))
+        else:
+            ml.append((name, name, module))
+    return ml
+
+
+def _get_feature_info(net, out_indices):
+    feature_info = getattr(net, 'feature_info')
+    if isinstance(feature_info, FeatureInfo):
+        return feature_info.from_other(out_indices)
+    elif isinstance(feature_info, (list, tuple)):
+        return FeatureInfo(net.feature_info, out_indices)
+    else:
+        assert False, "Provided feature_info is not valid"
+
+
+def _get_return_layers(feature_info, out_map):
+    module_names = feature_info.module_name()
+    return_layers = {}
+    for i, name in enumerate(module_names):
+        return_layers[name] = out_map[i] if out_map is not None else feature_info.out_indices[i]
+    return return_layers
+
+
+class FeatureDictNet(nn.ModuleDict):
+    """ Feature extractor with OrderedDict return
+    Wrap a model and extract features as specified by the out indices, the network is
+    partially re-built from contained modules.
+    There is a strong assumption that the modules have been registered into the model in the same
+    order as they are used. There should be no reuse of the same nn.Module more than once, including
+    trivial modules like `self.relu = nn.ReLU`.
+    Only submodules that are directly assigned to the model class (`model.feature1`) or at most
+    one Sequential container deep (`model.features.1`, with flatten_sequent=True) can be captured.
+    All Sequential containers that are directly assigned to the original model will have their
+    modules assigned to this module with the name `model.features.1` being changed to `model.features_1`
+    Arguments:
+        model (nn.Module): model from which we will extract the features
+        out_indices (tuple[int]): model output indices to extract features for
+        out_map (sequence): list or tuple specifying desired return id for each out index,
+            otherwise str(index) is used
+        feature_concat (bool): whether to concatenate intermediate features that are lists or tuples
+            vs select element [0]
+        flatten_sequential (bool): whether to flatten sequential modules assigned to model
+    """
+    def __init__(
+            self, model,
+            out_indices=(0, 1, 2, 3, 4), out_map=None, feature_concat=False, flatten_sequential=False):
+        super(FeatureDictNet, self).__init__()
+        self.feature_info = _get_feature_info(model, out_indices)
+        self.concat = feature_concat
+        self.return_layers = {}
+        return_layers = _get_return_layers(self.feature_info, out_map)
+        modules = _module_list(model, flatten_sequential=flatten_sequential)
+        remaining = set(return_layers.keys())
+        layers = OrderedDict()
+        for new_name, old_name, module in modules:
+            layers[new_name] = module
+            if old_name in remaining:
+                # return id has to be consistently str type for torchscript
+                self.return_layers[new_name] = str(return_layers[old_name])
+                remaining.remove(old_name)
+            if not remaining:
+                break
+        assert not remaining and len(self.return_layers) == len(return_layers), \
+            f'Return layers ({remaining}) are not present in model'
+        self.update(layers)
+
+    def _collect(self, x) -> (Dict[str, torch.Tensor]):
+        out = OrderedDict()
+        for name, module in self.items():
+            x = module(x)
+            if name in self.return_layers:
+                out_id = self.return_layers[name]
+                if isinstance(x, (tuple, list)):
+                    # If model tap is a tuple or list, concat or select first element
+                    # FIXME this may need to be more generic / flexible for some nets
+                    out[out_id] = torch.cat(x, 1) if self.concat else x[0]
+                else:
+                    out[out_id] = x
+        return out
+
+    def forward(self, x) -> Dict[str, torch.Tensor]:
+        return self._collect(x)
+
+
+class FeatureListNet(FeatureDictNet):
+    """ Feature extractor with list return
+    See docstring for FeatureDictNet above, this class exists only to appease Torchscript typing constraints.
+    In eager Python we could have returned List[Tensor] vs Dict[id, Tensor] based on a member bool.
+    """
+    def __init__(
+            self, model,
+            out_indices=(0, 1, 2, 3, 4), out_map=None, feature_concat=False, flatten_sequential=False):
+        super(FeatureListNet, self).__init__(
+            model, out_indices=out_indices, out_map=out_map, feature_concat=feature_concat,
+            flatten_sequential=flatten_sequential)
+
+    def forward(self, x) -> (List[torch.Tensor]):
+        return list(self._collect(x).values())
+
+
+class FeatureHookNet(nn.ModuleDict):
+    """ FeatureHookNet
+    Wrap a model and extract features specified by the out indices using forward/forward-pre hooks.
+    If `no_rewrite` is True, features are extracted via hooks without modifying the underlying
+    network in any way.
+    If `no_rewrite` is False, the model will be re-written as in the
+    FeatureList/FeatureDict case by folding first to second (Sequential only) level modules into this one.
+    FIXME this does not currently work with Torchscript, see FeatureHooks class
+    """
+    def __init__(
+            self, model,
+            out_indices=(0, 1, 2, 3, 4), out_map=None, out_as_dict=False, no_rewrite=False,
+            feature_concat=False, flatten_sequential=False, default_hook_type='forward'):
+        super(FeatureHookNet, self).__init__()
+        assert not torch.jit.is_scripting()
+        self.feature_info = _get_feature_info(model, out_indices)
+        self.out_as_dict = out_as_dict
+        layers = OrderedDict()
+        hooks = []
+        if no_rewrite:
+            assert not flatten_sequential
+            if hasattr(model, 'reset_classifier'):  # make sure classifier is removed?
+                model.reset_classifier(0)
+            layers['body'] = model
+            hooks.extend(self.feature_info.get_dicts())
+        else:
+            modules = _module_list(model, flatten_sequential=flatten_sequential)
+            remaining = {f['module']: f['hook_type'] if 'hook_type' in f else default_hook_type
+                         for f in self.feature_info.get_dicts()}
+            for new_name, old_name, module in modules:
+                layers[new_name] = module
+                for fn, fm in module.named_modules(prefix=old_name):
+                    if fn in remaining:
+                        hooks.append(dict(module=fn, hook_type=remaining[fn]))
+                        del remaining[fn]
+                if not remaining:
+                    break
+            assert not remaining, f'Return layers ({remaining}) are not present in model'
+        self.update(layers)
+        self.hooks = FeatureHooks(hooks, model.named_modules(), out_map=out_map)
+
+    def forward(self, x):
+        for name, module in self.items():
+            x = module(x)
+        out = self.hooks.get_output(x.device)
+        return out if self.out_as_dict else list(out.values())

TimeSformer/timesformer/models/head_helper.py

@@ -0,0 +1,235 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+"""ResNe(X)t Head helper."""
+
+import torch
+import torch.nn as nn
+
+class ResNetBasicHead(nn.Module):
+    """
+    ResNe(X)t 3D head.
+    This layer performs a fully-connected projection during training, when the
+    input size is 1x1x1. It performs a convolutional projection during testing
+    when the input size is larger than 1x1x1. If the inputs are from multiple
+    different pathways, the inputs will be concatenated after pooling.
+    """
+
+    def __init__(
+        self,
+        dim_in,
+        num_classes,
+        pool_size,
+        dropout_rate=0.0,
+        act_func="softmax",
+    ):
+        """
+        The `__init__` method of any subclass should also contain these
+            arguments.
+        ResNetBasicHead takes p pathways as input where p in [1, infty].
+
+        Args:
+            dim_in (list): the list of channel dimensions of the p inputs to the
+                ResNetHead.
+            num_classes (int): the channel dimensions of the p outputs to the
+                ResNetHead.
+            pool_size (list): the list of kernel sizes of p spatial temporal
+                poolings, temporal pool kernel size, spatial pool kernel size,
+                spatial pool kernel size in order.
+            dropout_rate (float): dropout rate. If equal to 0.0, perform no
+                dropout.
+            act_func (string): activation function to use. 'softmax': applies
+                softmax on the output. 'sigmoid': applies sigmoid on the output.
+        """
+        super(ResNetBasicHead, self).__init__()
+        assert (
+            len({len(pool_size), len(dim_in)}) == 1
+        ), "pathway dimensions are not consistent."
+        self.num_pathways = len(pool_size)
+
+        for pathway in range(self.num_pathways):
+            if pool_size[pathway] is None:
+                avg_pool = nn.AdaptiveAvgPool3d((1, 1, 1))
+            else:
+                avg_pool = nn.AvgPool3d(pool_size[pathway], stride=1)
+            self.add_module("pathway{}_avgpool".format(pathway), avg_pool)
+
+        if dropout_rate > 0.0:
+            self.dropout = nn.Dropout(dropout_rate)
+        # Perform FC in a fully convolutional manner. The FC layer will be
+        # initialized with a different std comparing to convolutional layers.
+        self.projection = nn.Linear(sum(dim_in), num_classes, bias=True)
+
+        # Softmax for evaluation and testing.
+        if act_func == "softmax":
+            self.act = nn.Softmax(dim=4)
+        elif act_func == "sigmoid":
+            self.act = nn.Sigmoid()
+        else:
+            raise NotImplementedError(
+                "{} is not supported as an activation"
+                "function.".format(act_func)
+            )
+
+    def forward(self, inputs):
+        assert (
+            len(inputs) == self.num_pathways
+        ), "Input tensor does not contain {} pathway".format(self.num_pathways)
+        pool_out = []
+        for pathway in range(self.num_pathways):
+            m = getattr(self, "pathway{}_avgpool".format(pathway))
+            pool_out.append(m(inputs[pathway]))
+        x = torch.cat(pool_out, 1)
+        # (N, C, T, H, W) -> (N, T, H, W, C).
+        x = x.permute((0, 2, 3, 4, 1))
+        # Perform dropout.
+        if hasattr(self, "dropout"):
+            x = self.dropout(x)
+        x = self.projection(x)
+
+        # Performs fully convlutional inference.
+        if not self.training:
+            x = self.act(x)
+            x = x.mean([1, 2, 3])
+
+        x = x.view(x.shape[0], -1)
+        return x
+
+
+class X3DHead(nn.Module):
+    """
+    X3D head.
+    This layer performs a fully-connected projection during training, when the
+    input size is 1x1x1. It performs a convolutional projection during testing
+    when the input size is larger than 1x1x1. If the inputs are from multiple
+    different pathways, the inputs will be concatenated after pooling.
+    """
+
+    def __init__(
+        self,
+        dim_in,
+        dim_inner,
+        dim_out,
+        num_classes,
+        pool_size,
+        dropout_rate=0.0,
+        act_func="softmax",
+        inplace_relu=True,
+        eps=1e-5,
+        bn_mmt=0.1,
+        norm_module=nn.BatchNorm3d,
+        bn_lin5_on=False,
+    ):
+        """
+        The `__init__` method of any subclass should also contain these
+            arguments.
+        X3DHead takes a 5-dim feature tensor (BxCxTxHxW) as input.
+
+        Args:
+            dim_in (float): the channel dimension C of the input.
+            num_classes (int): the channel dimensions of the output.
+            pool_size (float): a single entry list of kernel size for
+                spatiotemporal pooling for the TxHxW dimensions.
+            dropout_rate (float): dropout rate. If equal to 0.0, perform no
+                dropout.
+            act_func (string): activation function to use. 'softmax': applies
+                softmax on the output. 'sigmoid': applies sigmoid on the output.
+            inplace_relu (bool): if True, calculate the relu on the original
+                input without allocating new memory.
+            eps (float): epsilon for batch norm.
+            bn_mmt (float): momentum for batch norm. Noted that BN momentum in
+                PyTorch = 1 - BN momentum in Caffe2.
+            norm_module (nn.Module): nn.Module for the normalization layer. The
+                default is nn.BatchNorm3d.
+            bn_lin5_on (bool): if True, perform normalization on the features
+                before the classifier.
+        """
+        super(X3DHead, self).__init__()
+        self.pool_size = pool_size
+        self.dropout_rate = dropout_rate
+        self.num_classes = num_classes
+        self.act_func = act_func
+        self.eps = eps
+        self.bn_mmt = bn_mmt
+        self.inplace_relu = inplace_relu
+        self.bn_lin5_on = bn_lin5_on
+        self._construct_head(dim_in, dim_inner, dim_out, norm_module)
+
+    def _construct_head(self, dim_in, dim_inner, dim_out, norm_module):
+
+        self.conv_5 = nn.Conv3d(
+            dim_in,
+            dim_inner,
+            kernel_size=(1, 1, 1),
+            stride=(1, 1, 1),
+            padding=(0, 0, 0),
+            bias=False,
+        )
+        self.conv_5_bn = norm_module(
+            num_features=dim_inner, eps=self.eps, momentum=self.bn_mmt
+        )
+        self.conv_5_relu = nn.ReLU(self.inplace_relu)
+
+        if self.pool_size is None:
+            self.avg_pool = nn.AdaptiveAvgPool3d((1, 1, 1))
+        else:
+            self.avg_pool = nn.AvgPool3d(self.pool_size, stride=1)
+
+        self.lin_5 = nn.Conv3d(
+            dim_inner,
+            dim_out,
+            kernel_size=(1, 1, 1),
+            stride=(1, 1, 1),
+            padding=(0, 0, 0),
+            bias=False,
+        )
+        if self.bn_lin5_on:
+            self.lin_5_bn = norm_module(
+                num_features=dim_out, eps=self.eps, momentum=self.bn_mmt
+            )
+        self.lin_5_relu = nn.ReLU(self.inplace_relu)
+
+        if self.dropout_rate > 0.0:
+            self.dropout = nn.Dropout(self.dropout_rate)
+        # Perform FC in a fully convolutional manner. The FC layer will be
+        # initialized with a different std comparing to convolutional layers.
+        self.projection = nn.Linear(dim_out, self.num_classes, bias=True)
+
+        # Softmax for evaluation and testing.
+        if self.act_func == "softmax":
+            self.act = nn.Softmax(dim=4)
+        elif self.act_func == "sigmoid":
+            self.act = nn.Sigmoid()
+        else:
+            raise NotImplementedError(
+                "{} is not supported as an activation"
+                "function.".format(self.act_func)
+            )
+
+    def forward(self, inputs):
+        # In its current design the X3D head is only useable for a single
+        # pathway input.
+        assert len(inputs) == 1, "Input tensor does not contain 1 pathway"
+        x = self.conv_5(inputs[0])
+        x = self.conv_5_bn(x)
+        x = self.conv_5_relu(x)
+        x = self.avg_pool(x)
+
+        x = self.lin_5(x)
+        if self.bn_lin5_on:
+            x = self.lin_5_bn(x)
+        x = self.lin_5_relu(x)
+
+        # (N, C, T, H, W) -> (N, T, H, W, C).
+        x = x.permute((0, 2, 3, 4, 1))
+        # Perform dropout.
+        if hasattr(self, "dropout"):
+            x = self.dropout(x)
+        x = self.projection(x)
+
+        # Performs fully convlutional inference.
+        if not self.training:
+            x = self.act(x)
+            x = x.mean([1, 2, 3])
+
+        x = x.view(x.shape[0], -1)
+        return x