Overview
PySlowFast is an open-source video understanding codebase developed by FAIR, designed for efficient training and reproduction of state-of-the-art video models. Written in PyTorch, it supports rapid implementation and evaluation of novel video research ideas across tasks like classification and detection. The architecture supports multiple backbone networks, including SlowFast, Non-local Network, X3D, MViTv1, MViTv2, and Rev-ViT. It also integrates with PyTorchVideo datasets. PySlowFast facilitates unsupervised spatiotemporal representation learning and offers tools for model analysis and inference. The Model Zoo provides pre-trained models for download, and detailed installation instructions are available.