SlowFast

Why we included this project

PySlowFast is a research codebase from Facebook AI Research that implements many of the video backbones still used in modern work, including SlowFast, I3D, X3D, and the multiscale vision transformers, along with non-local blocks. Its training recipes are tuned on Kinetics and AVA, so you can reproduce published baselines and then swap in your own ideas without reimplementing everything. The repo also includes multigrid training, self-supervised and masked autoencoder projects, and a model zoo of pretrained weights. This is not a plug-and-play library: it expects comfort with YAML configs and PyTorch and gives you fine control over experiments. For researchers or engineers building action recognition or video understanding models, this is a dependable starting point.