lucidrains / TimeSformer-pytorch

Implementation of TimeSformer from Facebook AI, a pure attention-based solution for video classification

TimeSformer - Pytorch

Implementation of TimeSformer, from Facebook AI. A pure and simple attention-based solution for reaching SOTA on video classification. This repository will only house the best performing variant, 'Divided Space-Time Attention', which is nothing more than attention along the time axis before the spatial.

Press release

Install

$ pip install timesformer-pytorch

Usage

import torch
from timesformer_pytorch import TimeSformer

model = TimeSformer(
    dim = 512,
    image_size = 224,
    patch_size = 16,
    num_frames = 8,
    num_classes = 10,
    depth = 12,
    heads = 8,
    dim_head =  64,
    attn_dropout = 0.1,
    ff_dropout = 0.1
)

video = torch.randn(2, 8, 3, 224, 224) # (batch x frames x channels x height x width)
mask = torch.ones(2, 8).bool() # (batch x frame) - use a mask if there are variable length videos in the same batch

pred = model(video, mask = mask) # (2, 10)

Citations

@misc{bertasius2021spacetime,
    title   = {Is Space-Time Attention All You Need for Video Understanding?}, 
    author  = {Gedas Bertasius and Heng Wang and Lorenzo Torresani},
    year    = {2021},
    eprint  = {2102.05095},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV}
}

@misc{su2021roformer,
    title   = {RoFormer: Enhanced Transformer with Rotary Position Embedding},
    author  = {Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu},
    year    = {2021},
    eprint  = {2104.09864},
    archivePrefix = {arXiv},
    primaryClass = {cs.CL}
}

@article{tokshift2021,
    title   = {Token Shift Transformer for Video Classification},
    author  = {Hao Zhang, Yanbin Hao, Chong-Wah Ngo},
    journal = {ACM Multimedia 2021},
}