SlowFastNetworks

PyTorch implementation of "SlowFast Networks for Video Recognition".

Setup

Environment

conda env create -f environment.yml

Download UCF 101

wget https://www.crcv.ucf.edu/data/UCF101/UCF101.rar
export DATA_DIR=<YOUR_DATA_DIR>
unrar x UCF101.rar $DATA_DIR

Split into train/validation

export N_valid=30 # number of videos in the validatation set per class
mv $DATA_DIR/UCF-101/ $DATA_DIR/train
mkdir $DATA_DIR/validation
for dir in $(ls $DATA_DIR/train); do mkdir $DATA_DIR/validation/$dir; shuf -zn$N_valid -e $DATA_DIR/train/$dir/*.avi | xargs -0 -I{} mv -v {} $DATA_DIR/validation/$dir; done

After running the last command, the dataset will be organized in the following way

<YOUR_DATA_DIR>
│    │ train
│    │    │ ApplyEyeMakeup
│    │    │ ApplyLipstick
│    │    │ ...
│    │ validation
│    │    │ ApplyEyeMakeup
│    │    │ ApplyLipstick
│    │    │ ...

2. Modify the params in config.py (i.e. set params['dataset'] to <YOUR_DATA_DIR> )

Train model

conda activate slowfastnet
python train.py

Code Reference:

[1] https://github.com/Guocode/SlowFast-Networks/

[2] https://github.com/jfzhang95/pytorch-video-recognition

[3] https://github.com/irhumshafkat/R2Plus1D-PyTorch