Code + pre-trained models for the paper Keeping Your Eye on the Ball Trajectory Attention in Video Transformers

Overview

Motionformer

This is an official pytorch implementation of paper Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers. In this repository, we provide PyTorch code for training and testing our proposed Motionformer model. Motionformer use proposed trajectory attention to achieve state-of-the-art results on several video action recognition benchmarks such as Kinetics-400 and Something-Something V2.

If you find Motionformer useful in your research, please use the following BibTeX entry for citation.

@misc{patrick2021keeping,
      title={Keeping Your Eye on the Ball: Trajectory Attention in Video Transformers}, 
      author={Mandela Patrick and Dylan Campbell and Yuki M. Asano and Ishan Misra Florian Metze and Christoph Feichtenhofer and Andrea Vedaldi and Jo\ão F. Henriques},
      year={2021},
      eprint={2106.05392},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Model Zoo

We provide Motionformer models pretrained on Kinetics-400 (K400), Kinetics-600 (K600), Something-Something-V2 (SSv2), and Epic-Kitchens datasets.

name dataset # of frames spatial crop [email protected] [email protected] url
Joint K400 16 224 79.2 94.2 model
Divided K400 16 224 78.5 93.8 model
Motionformer K400 16 224 79.7 94.2 model
Motionformer-HR K400 16 336 81.1 95.2 model
Motionformer-L K400 32 224 80.2 94.8 model
name dataset # of frames spatial crop [email protected] [email protected] url
Motionformer K600 16 224 81.6 95.6 model
Motionformer-HR K600 16 336 82.7 96.1 model
Motionformer-L K600 32 224 82.2 96.0 model
name dataset # of frames spatial crop [email protected] [email protected] url
Joint SSv2 16 224 64.0 88.4 model
Divided SSv2 16 224 64.2 88.6 model
Motionformer SSv2 16 224 66.5 90.1 model
Motionformer-HR SSv2 16 336 67.1 90.6 model
Motionformer-L SSv2 32 224 68.1 91.2 model
name dataset # of frames spatial crop A acc N acc url
Motionformer EK 16 224 43.1 56.5 model
Motionformer-HR EK 16 336 44.5 58.5 model
Motionformer-L EK 32 224 44.1 57.6 model

Installation

First, create a conda virtual environment and activate it:

conda create -n motionformer python=3.8.5 -y
source activate motionformer

Then, install the following packages:

  • torchvision: pip install torchvision or conda install torchvision -c pytorch
  • 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
  • matplotlib: pip install matplotlib
  • pandas: pip install pandas
  • ffmeg: pip install ffmpeg-python

OR:

simply create conda environment with all packages just from yaml file:

conda env create -f environment.yml

Lastly, build the Motionformer codebase by running:

git clone https://github.com/facebookresearch/Motionformer
cd Motionformer
python setup.py build develop

Usage

Dataset Preparation

Please use the dataset preparation instructions provided in DATASET.md.

Training the Default Motionformer

Training the default Motionformer that uses trajectory attention, and operates on 16-frame clips cropped at 224x224 spatial resolution, can be done using the following command:

python tools/run_net.py \
  --cfg configs/K400/motionformer_224_16x4.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 modify

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/. 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.

Using Different Self-Attention Schemes

If you want to experiment with different space-time self-attention schemes, e.g., joint space-time attention or divided space-time attention, use the following commands:

python tools/run_net.py \
  --cfg configs/K400/joint_224_16x4.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/K400/divided_224_16x4.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  NUM_GPUS 8 \
  TRAIN.BATCH_SIZE 8 \

Training Different Motionformer Variants

If you want to train more powerful Motionformer variants, e.g., Motionformer-HR (operating on 16-frame clips sampled at 336x336 spatial resolution), and Motionformer-L (operating on 32-frame clips sampled at 224x224 spatial resolution), use the following commands:

python tools/run_net.py \
  --cfg configs/K400/motionformer_336_16x8.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/K400/motionformer_224_32x3.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/K400/motionformer_224_16x4.yaml \
  DATA.PATH_TO_DATA_DIR path_to_your_dataset \
  TEST.CHECKPOINT_FILE_PATH path_to_your_checkpoint \
  TRAIN.ENABLE False \

Alterantively, you can modify provided SLURM script and run following:

sbatch slurm_scripts/test.sh configs/K400/motionformer_224_16x4.yaml path_to_your_checkpoint

Single-Node Training via Slurm

To train Motionformer via Slurm, please check out our single node Slurm training script slurm_scripts/run_single_node_job.sh.

sbatch slurm_scripts/run_single_node_job.sh configs/K400/motionformer_224_16x4.yaml /your/job/dir/${JOB_NAME}/

Multi-Node Training via Submitit

Distributed training is available via Slurm and submitit

pip install submitit

To train Motionformer model on Kinetics using 8 nodes with 8 gpus each use the following command:

python run_with_submitit.py --cfg configs/K400/motionformer_224_16x4.yaml --job_dir  /your/job/dir/${JOB_NAME}/ --partition $PARTITION --num_shards 8 --use_volta32

We provide a script for launching slurm jobs in slurm_scripts/run_multi_node_job.sh.

sbatch slurm_scripts/run_multi_node_job.sh configs/K400/motionformer_224_16x4.yaml /your/job/dir/${JOB_NAME}/

Please note that hyper-parameters in configs were used with 8 nodes with 8 gpus (32 GB). Please scale batch-size, and learning-rate appropriately for your cluster configuration.

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_EPOCH_RESET: True
TRAIN.CHECKPOINT_FILE_PATH path_to_your_PyTorch_checkpoint

Environment

The code was developed using python 3.8.5 on Ubuntu 20.04. For training, we used eight 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. However, portions of the project are available under separate license terms: SlowFast and pytorch-image-models are licensed under the Apache 2.0 license.

Contributing

We actively welcome your pull requests. Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.

Acknowledgements

Motionformer is built on top of PySlowFast, Timesformer and pytorch-image-models by Ross Wightman. We thank the authors for releasing their code. If you use our model, please consider citing these works as well:

@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}
}
@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}
}
@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}}
}
Owner
Facebook Research
Facebook Research
Python SDK for building, training, and deploying ML models

Overview of Kubeflow Fairing Kubeflow Fairing is a Python package that streamlines the process of building, training, and deploying machine learning (

Kubeflow 325 Dec 13, 2022
Bringing sanity to world of messed-up data

Sanitize sanitize is a Python module for making sure various things (e.g. HTML) are safe to use. It was originally written by Mark Pilgrim and is dist

Alireza Savand 63 Oct 26, 2021
UT-Sarulab MOS prediction system using SSL models

UTMOS: UTokyo-SaruLab MOS Prediction System Official implementation of "UTMOS: UTokyo-SaruLab System for VoiceMOS Challenge 2022" submitted to INTERSP

sarulab-speech 58 Nov 22, 2022
[ICCV 2021] Counterfactual Attention Learning for Fine-Grained Visual Categorization and Re-identification

Counterfactual Attention Learning Created by Yongming Rao*, Guangyi Chen*, Jiwen Lu, Jie Zhou This repository contains PyTorch implementation for ICCV

Yongming Rao 90 Dec 31, 2022
Code for the ECCV2020 paper "A Differentiable Recurrent Surface for Asynchronous Event-Based Data"

A Differentiable Recurrent Surface for Asynchronous Event-Based Data Code for the ECCV2020 paper "A Differentiable Recurrent Surface for Asynchronous

Marco Cannici 21 Oct 05, 2022
Atif Hassan 103 Dec 14, 2022
Grounding Representation Similarity with Statistical Testing

Grounding Representation Similarity with Statistical Testing This repo contains code to replicate the results in our paper, which evaluates representa

26 Dec 02, 2022
DCSL - Generalizable Crowd Counting via Diverse Context Style Learning

DCSL Generalizable Crowd Counting via Diverse Context Style Learning Requirement

3 Jun 13, 2022
Yolov5-lite - Minimal PyTorch implementation of YOLOv5

Yolov5-Lite: Minimal YOLOv5 + Deep Sort Overview This repo is a shortened versio

Kadir Nar 57 Nov 28, 2022
Differential fuzzing for the masses!

NEZHA NEZHA is an efficient and domain-independent differential fuzzer developed at Columbia University. NEZHA exploits the behavioral asymmetries bet

147 Dec 05, 2022
Object-aware Contrastive Learning for Debiased Scene Representation

Object-aware Contrastive Learning Official PyTorch implementation of "Object-aware Contrastive Learning for Debiased Scene Representation" by Sangwoo

43 Dec 14, 2022
The official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness.

This repository is the official implementation of A Unified Game-Theoretic Interpretation of Adversarial Robustness. Requirements pip install -r requi

Jie Ren 17 Dec 12, 2022
Scheduling BilinearRewards

Scheduling_BilinearRewards Requirement Python 3 =3.5 Structure main.py This file includes the main function. For getting the results in Figure 1, ple

junghun.kim 0 Nov 25, 2021
Benchmark for Answering Existential First Order Queries with Single Free Variable

EFO-1-QA Benchmark for First Order Query Estimation on Knowledge Graphs This repository contains an entire pipeline for the EFO-1-QA benchmark. EFO-1

HKUST-KnowComp 14 Oct 24, 2022
Image Deblurring using Generative Adversarial Networks

DeblurGAN arXiv Paper Version Pytorch implementation of the paper DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks. Our netwo

Orest Kupyn 2.2k Jan 01, 2023
Step by Step on how to create an vision recognition model using LOBE.ai, export the model and run the model in an Azure Function

Step by Step on how to create an vision recognition model using LOBE.ai, export the model and run the model in an Azure Function

El Bruno 3 Mar 30, 2022
training script for space time memory network

Trainig Script for Space Time Memory Network This codebase implemented training code for Space Time Memory Network with some cyclic features. Requirem

Yuxi Li 100 Dec 20, 2022
An efficient implementation of GPNN

Efficient-GPNN An efficient implementation of GPNN as depicted in "Drop the GAN: In Defense of Patches Nearest Neighbors as Single Image Generative Mo

7 Apr 16, 2022
Raster Vision is an open source Python framework for building computer vision models on satellite, aerial, and other large imagery sets

Raster Vision is an open source Python framework for building computer vision models on satellite, aerial, and other large imagery sets (including obl

Azavea 1.7k Dec 22, 2022
python library for invisible image watermark (blind image watermark)

invisible-watermark invisible-watermark is a python library and command line tool for creating invisible watermark over image.(aka. blink image waterm

Shield Mountain 572 Jan 07, 2023