This project provides a compact self-contained kinetics dataloader based on Pytorch. The codes are modified from SlowFast.
Run the following commands to install the exvironments
conda create -n YOUR_ENV_NAME pip python=3.8
source activate YOUR_ENV_NAME
conda install pytorch==1.6.0 torchvision==0.7.0 -c pytorch
conda install pillow
conda install matplotlib
pip install pyyaml
pip install av
Follow the Dataset Preprocessing to prepare the Kinetics-400/Kinetics-600 datasets.
Run the Demo to visualize the sampled video frames. This project keep the most compact version of dataloader by removing all the frame-level data augmentation.