How facebookresearch/slowfast works
Project type and purpose
• PySlowFast is an open-source video understanding framework built on PyTorch.
• It provides end-to-end video model backbones, training, evaluation, and visualization tools for video classification/detection research.
• Key use cases: training new models, running inference/demo, evaluating on AVA/Kinetics datasets, visualizing results.
File structure
• Top-level components:
• slowfast/ – Core library: config, datasets, models, utils, visualization
• configs/ – YAML configurations for experiments (various datasets, backbones)
• tools/ – CLI entrypoints: train_net.py, test_net.py, run_net.py, demo_net.py, benchmark.py, visualization.py
• ava_evaluation/ – Evaluation code for AVA action detection
• demo/ – Demo YAMLs and visualization assets
• projects/ – Research projects/extensions (MAE, maskfeat, MViTv2, etc.)
• setup.py, setup.cfg – Package installation
• scripts & docs: INSTALL.md, GETTING_STARTED.md, MODEL_ZOO.md, VISUALIZATION_TOOLS.md, etc.
README for additional insights
• Architecture is modular:
• Config system decouples code from experiment settings (slowfast/config)
• Datasets abstraction (slowfast/datasets)
• Model builder pattern (slowfast/models/video_model_builder, custom_video_model_builder, ptv_model_builder)
• Utilities for checkpointing, distributed training, logging, lr policy (slowfast/utils)
• Visualization components (slowfast/visualization, tools/visualization.py)
• Supports multiple backbones: SlowFast, I3D, X3D, MViT, Rev-ViT, contrastive SSL, masked SSL.
Final advice
• Be very detailed: break “slowfast/models” further into builders, backbone definitions, loss.
• Essential components; show both training and inference paths.