[New Model] Add TimeSformer model (#18908)

* init timesformer

* apply fix-copies

* reformat style

* revert back some incoorect style updates

* init timesformer

* apply fix-copies

* reformat style

* revert back some incoorect style updates

* update timseformer doc

* add some functions and classes

* add new config params

* implement multiple classes

* update TimeSformerLayer

* update TimeSformerModel, TimeSformerPreTrainedModel, TimeSformerEncoder

* several fixes

* reformat

* temporary update

* fix some typos

* fix weight converter

* more fixes

* fix a typo

* fix typo

* remove redundant params

* fix for latest hf-hub

* merge fix

* fix some checks

* video classification works with einops

* add paper info to docs

* merge fix

* remove redundant line

* remove redundant docstring

* update config

* fix some typos

* fix converter

* update some test constants

* refactor einops functions

* reformat

* fix a comment

* remove redundat imports

* reformat

* fix a typo

* remove comment

* remove unused imports

* remove redundant doc line

* reformat

* add missing line

* fix docs

* fix timesformer auto feat ext

* add unittests

* reformat

* fix docs

* some fixes and updates

* fix readme

* fix modeling

* fix readme

* update index

* revert _toctree.yml changes

* update timseformer.mdx

* update drop_path_prob to drop_path_rate

* add dosctring for drop_path_rate

* update TimeSformerPatchEmbed naming

* remove to_2tuple

* explicit use of nn.functional

* reformat

* many updates from review comments

* fix a typo

* reformat

* remove assert, better variable name

* make variable names more explicit

* add some adapted from

* more explicit variable names

* remove redundant docstring

* fix initilaization

* move permute inside embedding

* update class names

* remove unused imports

* add test for video classification

* update PretrainedModel with PreTrainedModel

* remove double permute

* update based on sylvain's review

* aply auto fix

* update image_processing_auto for timesformer

* update hub urls

* reformat

* remove duplicate import

* update doc link

README.md

@@ -390,7 +390,8 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Table Transformer](https://huggingface.co/docs/transformers/model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
-1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)**  (from HuggingFace).
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.

README_es.md

@@ -391,6 +391,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
 1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)**  (from HuggingFace).
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.

README_ja.md

@@ -426,6 +426,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
 1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)**  (from HuggingFace).
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
 1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.

README_ko.md

@@ -341,7 +341,8 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
 1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler

README_zh-hans.md

@@ -365,7 +365,8 @@ conda install -c huggingface transformers
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (来自 Google AI) 伴随论文 [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) 由 Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos 发布。
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (来自 Microsoft Research) 伴随论文 [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) 由 Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou 发布。
 1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (来自 Google/CMU) 伴随论文 [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) 由 Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov 发布。
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (来自 Microsoft) 伴随论文 [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) 由 Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei 发布。
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler

README_zh-hant.md

@@ -376,8 +376,9 @@ conda install -c huggingface transformers
 1. **[Table Transformer](https://huggingface.co/docs/transformers/model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
-1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/model_doc/time_series_transformer)** (from HuggingFace).  
+1. **[TimeSformer](https://huggingface.co/docs/transformers/main/model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft) released with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler

docs/source/en/_toctree.yml

@@ -432,6 +432,8 @@
         title: Swin Transformer V2
       - local: model_doc/table-transformer
         title: Table Transformer
+      - local: model_doc/timesformer
+        title: TimeSformer
       - local: model_doc/van
         title: VAN
       - local: model_doc/videomae

docs/source/en/index.mdx

@@ -178,7 +178,8 @@ The documentation is organized into five sections:
 1. **[Table Transformer](model_doc/table-transformer)** (from Microsoft Research) released with the paper [PubTables-1M: Towards Comprehensive Table Extraction From Unstructured Documents](https://arxiv.org/abs/2110.00061) by Brandon Smock, Rohith Pesala, Robin Abraham.
 1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
-1. **[Time Series Transformer](model_doc/time_series_transformer)**  (from HuggingFace).
+1. **[Time Series Transformer](model_doc/time_series_transformer)** (from HuggingFace).
+1. **[TimeSformer](model_doc/timesformer)** (from Facebook) released with the paper [Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Gedas Bertasius, Heng Wang, Lorenzo Torresani.
 1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
@@ -339,6 +340,7 @@ Flax), PyTorch, and/or TensorFlow.
 |       Table Transformer       |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             TAPAS             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |    Time Series Transformer    |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|          TimeSformer          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |    Trajectory Transformer     |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |        Transformer-XL         |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |             TrOCR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/timesformer.mdx

+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# TimeSformer
+
+## Overview
+
+The TimeSformer model was proposed in [TimeSformer: Is Space-Time Attention All You Need for Video Understanding?](https://arxiv.org/abs/2102.05095) by Facebook Research.
+This work is a milestone in action-recognition field being the first video transformer. It inspired many transformer based video understanding and classification papers.
+
+The abstract from the paper is the following:
+
+*We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: [this https URL](https://github.com/facebookresearch/TimeSformer).*
+
+Tips:
+
+There are many pretrained variants. Select your pretrained model based on the dataset it is trained on. Moreover, the number of input frames per clip changes based on the model size so you should consider this parameter while selecting your pretrained model.
+
+This model was contributed by [fcakyon](https://huggingface.co/fcakyon).
+The original code can be found [here](https://github.com/facebookresearch/TimeSformer).
+
+
+## TimesformerConfig
+
+[[autodoc]] TimesformerConfig
+
+## TimesformerModel
+
+[[autodoc]] TimesformerModel
+    - forward
+
+## TimesformerForVideoClassification
+
+[[autodoc]] TimesformerForVideoClassification
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -379,6 +379,7 @@ _import_structure = {
         "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "TimeSeriesTransformerConfig",
     ],
+    "models.timesformer": ["TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "TimesformerConfig"],
     "models.trajectory_transformer": [
         "TRAJECTORY_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "TrajectoryTransformerConfig",
@@ -2104,6 +2105,14 @@ else:
             "TimeSeriesTransformerPreTrainedModel",
         ]
     )
+    _import_structure["models.timesformer"].extend(
+        [
+            "TIMESFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "TimesformerForVideoClassification",
+            "TimesformerModel",
+            "TimesformerPreTrainedModel",
+        ]
+    )
     _import_structure["models.trajectory_transformer"].extend(
         [
             "TRAJECTORY_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3588,6 +3597,7 @@ if TYPE_CHECKING:
         TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         TimeSeriesTransformerConfig,
     )
+    from .models.timesformer import TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, TimesformerConfig
     from .models.trajectory_transformer import (
         TRAJECTORY_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         TrajectoryTransformerConfig,
@@ -5015,6 +5025,12 @@ if TYPE_CHECKING:
             TimeSeriesTransformerModel,
             TimeSeriesTransformerPreTrainedModel,
         )
+        from .models.timesformer import (
+            TIMESFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TimesformerForVideoClassification,
+            TimesformerModel,
+            TimesformerPreTrainedModel,
+        )
         from .models.trajectory_transformer import (
             TRAJECTORY_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             TrajectoryTransformerModel,

src/transformers/models/__init__.py

@@ -152,6 +152,7 @@ from . import (
     tapas,
     tapex,
     time_series_transformer,
+    timesformer,
     trajectory_transformer,
     transfo_xl,
     trocr,

src/transformers/models/auto/configuration_auto.py

@@ -148,6 +148,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("table-transformer", "TableTransformerConfig"),
         ("tapas", "TapasConfig"),
         ("time_series_transformer", "TimeSeriesTransformerConfig"),
+        ("timesformer", "TimesformerConfig"),
         ("trajectory_transformer", "TrajectoryTransformerConfig"),
         ("transfo-xl", "TransfoXLConfig"),
         ("trocr", "TrOCRConfig"),
@@ -290,6 +291,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("table-transformer", "TABLE_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("tapas", "TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("time_series_transformer", "TIME_SERIES_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("timesformer", "TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("transfo-xl", "TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("unispeech", "UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("unispeech-sat", "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -455,6 +457,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("tapas", "TAPAS"),
         ("tapex", "TAPEX"),
         ("time_series_transformer", "Time Series Transformer"),
+        ("timesformer", "TimeSformer"),
         ("trajectory_transformer", "Trajectory Transformer"),
         ("transfo-xl", "Transformer-XL"),
         ("trocr", "TrOCR"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -77,6 +77,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("swin", "ViTFeatureExtractor"),
         ("swinv2", "ViTFeatureExtractor"),
         ("table-transformer", "DetrFeatureExtractor"),
+        ("timesformer", "VideoMAEFeatureExtractor"),
         ("van", "ConvNextFeatureExtractor"),
         ("videomae", "VideoMAEFeatureExtractor"),
         ("vilt", "ViltFeatureExtractor"),

src/transformers/models/auto/image_processing_auto.py

@@ -74,6 +74,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("swin", "ViTImageProcessor"),
         ("swinv2", "ViTImageProcessor"),
         ("table-transformer", "DetrImageProcessor"),
+        ("timesformer", "VideoMAEImageProcessor"),
         ("van", "ConvNextImageProcessor"),
         ("videomae", "VideoMAEImageProcessor"),
         ("vilt", "ViltImageProcessor"),

src/transformers/models/auto/modeling_auto.py

@@ -144,6 +144,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("table-transformer", "TableTransformerModel"),
         ("tapas", "TapasModel"),
         ("time_series_transformer", "TimeSeriesTransformerModel"),
+        ("timesformer", "TimesformerModel"),
         ("trajectory_transformer", "TrajectoryTransformerModel"),
         ("transfo-xl", "TransfoXLModel"),
         ("unispeech", "UniSpeechModel"),
@@ -429,6 +430,7 @@ MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING_NAMES = OrderedDict(
 
 MODEL_FOR_VIDEO_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
+        ("timesformer", "TimesformerForVideoClassification"),
         ("videomae", "VideoMAEForVideoClassification"),
     ]
 )

src/transformers/models/timesformer/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_timesformer": ["TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "TimesformerConfig"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_timesformer"] = [
+        "TIMESFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TimesformerModel",
+        "TimesformerForVideoClassification",
+        "TimesformerPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_timesformer import TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, TimesformerConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_timesformer import (
+            TIMESFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TimesformerForVideoClassification,
+            TimesformerModel,
+            TimesformerPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/timesformer/configuration_timesformer.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TimeSformer model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+TIMESFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/timesformer": "https://huggingface.co/facebook/timesformer/resolve/main/config.json",
+}
+
+
+class TimesformerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`TimesformerModel`]. It is used to instantiate a
+    TimeSformer model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the TimeSformer
+    [facebook/timesformer](https://huggingface.co/facebook/timesformer-base-finetuned-k600) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        num_frames (`int`, *optional*, defaults to 8):
+            The number of frames in each video.
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-6):
+            The epsilon used by the layer normalization layers.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        attention_type (`str`, *optional*, defaults to `"divided_space_time"`):
+            The attention type to use. Must be one of `"divided_space_time"`, `"space_only"`, `"joint_space_time"`.
+        drop_path_rate (`float`, *optional*, defaults to 0):
+            The dropout ratio for stochastic depth.
+
+    Example:
+
+    ```python
+    >>> from transformers import TimesformerConfig, TimesformerModel
+
+    >>> # Initializing a TimeSformer timesformer-base style configuration
+    >>> configuration = TimesformerConfig()
+
+    >>> # Randomly initializing a model from the configuration
+    >>> model = TimesformerModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "timesformer"
+
+    def __init__(
+        self,
+        image_size=224,
+        patch_size=16,
+        num_channels=3,
+        num_frames=8,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-6,
+        qkv_bias=True,
+        attention_type="divided_space_time",
+        drop_path_rate=0,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.num_frames = num_frames
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.qkv_bias = qkv_bias
+
+        self.attention_type = attention_type
+        self.drop_path_rate = drop_path_rate

src/transformers/models/timesformer/convert_timesformer_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert TimeSformer checkpoints from the original repository: https://github.com/MCG-NJU/TimeSformer"""
+
+import argparse
+import json
+
+import numpy as np
+import torch
+
+import gdown
+from huggingface_hub import hf_hub_download
+from transformers import TimesformerConfig, TimesformerForVideoClassification, VideoMAEFeatureExtractor
+
+
+def get_timesformer_config(model_name):
+    config = TimesformerConfig()
+
+    if "large" in model_name:
+        config.num_frames = 96
+
+    if "hr" in model_name:
+        config.num_frames = 16
+        config.image_size = 448
+
+    repo_id = "huggingface/label-files"
+    if "k400" in model_name:
+        config.num_labels = 400
+        filename = "kinetics400-id2label.json"
+    elif "k600" in model_name:
+        config.num_labels = 600
+        filename = "kinetics600-id2label.json"
+    elif "ssv2" in model_name:
+        config.num_labels = 174
+        filename = "something-something-v2-id2label.json"
+    else:
+        raise ValueError("Model name should either contain 'k400', 'k600' or 'ssv2'.")
+    id2label = json.load(open(hf_hub_download(repo_id, filename, repo_type="dataset"), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    return config
+
+
+def rename_key(name):
+    if "encoder." in name:
+        name = name.replace("encoder.", "")
+    if "cls_token" in name:
+        name = name.replace("cls_token", "timesformer.embeddings.cls_token")
+    if "pos_embed" in name:
+        name = name.replace("pos_embed", "timesformer.embeddings.position_embeddings")
+    if "time_embed" in name:
+        name = name.replace("time_embed", "timesformer.embeddings.time_embeddings")
+    if "patch_embed.proj" in name:
+        name = name.replace("patch_embed.proj", "timesformer.embeddings.patch_embeddings.projection")
+    if "patch_embed.norm" in name:
+        name = name.replace("patch_embed.norm", "timesformer.embeddings.norm")
+    if "blocks" in name:
+        name = name.replace("blocks", "timesformer.encoder.layer")
+    if "attn.proj" in name:
+        name = name.replace("attn.proj", "attention.output.dense")
+    if "attn" in name and "bias" not in name and "temporal" not in name:
+        name = name.replace("attn", "attention.self")
+    if "attn" in name and "temporal" not in name:
+        name = name.replace("attn", "attention.attention")
+    if "temporal_norm1" in name:
+        name = name.replace("temporal_norm1", "temporal_layernorm")
+    if "temporal_attn.proj" in name:
+        name = name.replace("temporal_attn", "temporal_attention.output.dense")
+    if "temporal_fc" in name:
+        name = name.replace("temporal_fc", "temporal_dense")
+    if "norm1" in name and "temporal" not in name:
+        name = name.replace("norm1", "layernorm_before")
+    if "norm2" in name:
+        name = name.replace("norm2", "layernorm_after")
+    if "mlp.fc1" in name:
+        name = name.replace("mlp.fc1", "intermediate.dense")
+    if "mlp.fc2" in name:
+        name = name.replace("mlp.fc2", "output.dense")
+    if "norm.weight" in name and "fc" not in name and "temporal" not in name:
+        name = name.replace("norm.weight", "timesformer.layernorm.weight")
+    if "norm.bias" in name and "fc" not in name and "temporal" not in name:
+        name = name.replace("norm.bias", "timesformer.layernorm.bias")
+    if "head" in name:
+        name = name.replace("head", "classifier")
+
+    return name
+
+
+def convert_state_dict(orig_state_dict, config):
+    for key in orig_state_dict.copy().keys():
+        val = orig_state_dict.pop(key)
+
+        if key.startswith("model."):
+            key = key.replace("model.", "")
+
+        if "qkv" in key:
+            key_split = key.split(".")
+            layer_num = int(key_split[1])
+            prefix = "timesformer.encoder.layer."
+            if "temporal" in key:
+                postfix = ".temporal_attention.attention.qkv."
+            else:
+                postfix = ".attention.attention.qkv."
+            if "weight" in key:
+                orig_state_dict[f"{prefix}{layer_num}{postfix}weight"] = val
+            else:
+                orig_state_dict[f"{prefix}{layer_num}{postfix}bias"] = val
+        else:
+            orig_state_dict[rename_key(key)] = val
+
+    return orig_state_dict
+
+
+# We will verify our results on a video of eating spaghetti
+# Frame indices used: [164 168 172 176 181 185 189 193 198 202 206 210 215 219 223 227]
+def prepare_video():
+    file = hf_hub_download(
+        repo_id="hf-internal-testing/spaghetti-video", filename="eating_spaghetti.npy", repo_type="dataset"
+    )
+    video = np.load(file)
+    return list(video)
+
+
+def convert_timesformer_checkpoint(checkpoint_url, pytorch_dump_folder_path, model_name, push_to_hub):
+    config = get_timesformer_config(model_name)
+
+    model = TimesformerForVideoClassification(config)
+
+    # download original checkpoint, hosted on Google Drive
+    output = "pytorch_model.bin"
+    gdown.cached_download(checkpoint_url, output, quiet=False)
+    files = torch.load(output, map_location="cpu")
+    if "model" in files:
+        state_dict = files["model"]
+    elif "module" in files:
+        state_dict = files["module"]
+    else:
+        state_dict = files["model_state"]
+    new_state_dict = convert_state_dict(state_dict, config)
+
+    model.load_state_dict(new_state_dict)
+    model.eval()
+
+    # verify model on basic input
+    feature_extractor = VideoMAEFeatureExtractor(image_mean=[0.5, 0.5, 0.5], image_std=[0.5, 0.5, 0.5])
+    video = prepare_video()
+    inputs = feature_extractor(video[:8], return_tensors="pt")
+
+    outputs = model(**inputs)
+    logits = outputs.logits
+
+    model_names = [
+        # Kinetics-400 checkpoints (hr = high resolution input of 448px instead of 224px)
+        "timesformer-base-finetuned-k400",
+        "timesformer-large-finetuned-k400",
+        "timesformer-hr-finetuned-k400",
+        # Kinetics-600 checkpoints (hr = high resolution input of 448px instead of 224px)
+        "timesformer-base-finetuned-k600",
+        "timesformer-large-finetuned-k600",
+        "timesformer-hr-finetuned-k600",
+        # Something-Something-v2 checkpoints (hr = high resolution input of 448px instead of 224px)
+        "timesformer-base-finetuned-ssv2",
+        "timesformer-large-finetuned-ssv2",
+        "timesformer-hr-finetuned-ssv2",
+    ]
+
+    # NOTE: logits were tested with image_mean and image_std equal to [0.5, 0.5, 0.5] and [0.5, 0.5, 0.5]
+    if model_name == "timesformer-base-finetuned-k400":
+        expected_shape = torch.Size([1, 400])
+        expected_slice = torch.tensor([-0.3016, -0.7713, -0.4205])
+    elif model_name == "timesformer-base-finetuned-k600":
+        expected_shape = torch.Size([1, 600])
+        expected_slice = torch.tensor([-0.7267, -0.7466, 3.2404])
+    elif model_name == "timesformer-base-finetuned-ssv2":
+        expected_shape = torch.Size([1, 174])
+        expected_slice = torch.tensor([-0.9059, 0.6433, -3.1457])
+    elif model_name == "timesformer-large-finetuned-k400":
+        expected_shape = torch.Size([1, 400])
+        expected_slice = torch.tensor([0, 0, 0])
+    elif model_name == "timesformer-large-finetuned-k600":
+        expected_shape = torch.Size([1, 600])
+        expected_slice = torch.tensor([0, 0, 0])
+    elif model_name == "timesformer-large-finetuned-ssv2":
+        expected_shape = torch.Size([1, 174])
+        expected_slice = torch.tensor([0, 0, 0])
+    elif model_name == "timesformer-hr-finetuned-k400":
+        expected_shape = torch.Size([1, 400])
+        expected_slice = torch.tensor([-0.9617, -3.7311, -3.7708])
+    elif model_name == "timesformer-hr-finetuned-k600":
+        expected_shape = torch.Size([1, 600])
+        expected_slice = torch.tensor([2.5273, 0.7127, 1.8848])
+    elif model_name == "timesformer-hr-finetuned-ssv2":
+        expected_shape = torch.Size([1, 174])
+        expected_slice = torch.tensor([-3.6756, -0.7513, 0.7180])
+    else:
+        raise ValueError(f"Model name not supported. Should be one of {model_names}")
+
+    # verify logits
+    assert logits.shape == expected_shape
+    assert torch.allclose(logits[0, :3], expected_slice, atol=1e-4)
+    print("Logits ok!")
+
+    if pytorch_dump_folder_path is not None:
+        print(f"Saving model and feature extractor to {pytorch_dump_folder_path}")
+        feature_extractor.save_pretrained(pytorch_dump_folder_path)
+        model.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print("Pushing to the hub...")
+        model.push_to_hub(f"fcakyon/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_url",
+        default="https://drive.google.com/u/1/uc?id=17yvuYp9L4mn-HpIcK5Zo6K3UoOy1kA5l&export=download",
+        type=str,
+        help=(
+            "URL of the original PyTorch checkpoint (on Google Drive) you'd like to convert. Should be a direct"
+            " download link."
+        ),
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default="",
+        type=str,
+        help="Path to the output PyTorch model directory.",
+    )
+    parser.add_argument("--model_name", default="timesformer-base-finetuned-k400", type=str, help="Name of the model.")
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_timesformer_checkpoint(
+        args.checkpoint_url, args.pytorch_dump_folder_path, args.model_name, args.push_to_hub
+    )

src/transformers/utils/dummy_pt_objects.py

@@ -5379,6 +5379,30 @@ class TimeSeriesTransformerPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+TIMESFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TimesformerForVideoClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TimesformerModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class TimesformerPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 TRAJECTORY_TRANSFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None