SlowFast

A PyTorch implementation of SlowFast based on ICCV 2019 paper SlowFast Networks for Video Recognition.

Network Architecture

Requirements

conda install pytorch=1.9.1 torchvision cudatoolkit -c pytorch


pip install pytorchvideo

Dataset

kinetics-400 dataset is used in this repo, you could download these datasets from official websites. The data directory structure is shown as follows:

├──data
  ├── train
      ├── abseiling
          ├── _4YTwq0-73Y_000044_000054.mp4
          └── ...
          ...
      ├── archery
          same structure as abseiling
  ├── test
     same structure as train
     ...

Usage

Train Model

python train.py --batch_size 16
optional arguments:
--data_root                   Datasets root path [default value is 'data']
--batch_size                  Number of videos in each mini-batch [default value is 8]
--epochs                      Number of epochs over the model to train [default value is 10]
--save_root                   Result saved root path [default value is 'result']

Test Model

python test.py --video_path data/test/beatboxing/5s_gFWie1Ys_000069_000079.mp4
optional arguments:
--model_path                  Model path [default value is 'result/slow_fast.pth']
--video_path                  Video path [default value is 'data/test/applauding/_V-dzjftmCQ_000023_000033.mp4']

GitHub

https://github.com/leftthomas/SlowFast