What it does

Reimplements the Temporal Segment Networks (TSN) paper in PyTorch, matching the original Caffe settings exactly. Trains and tests on UCF101 using RGB frames, optical flow, or RGB differences with a BN-Inception backbone. You still need the original TSN repo to extract optical flow and generate video lists.

The interesting bit

The README opens with the authors pleading for you to migrate to MMAction, their newer toolbox. This repo survives purely as “historical reference” — a rare case of maintainers actively discouraging their own 1,000-star project.

Key highlights

• Supports three input modalities: RGB, optical flow, and RGB difference
• Reproduces original TSN experiments with identical hyperparameters
• Requires --recursive clone for Inception architectures (submodule dependency)
• Single-GPU training via main.py, evaluation via test_models.py
• Pre-dates modern video understanding frameworks; authors now maintain MMAction2

Caveats

• Optical flow extraction and video list generation require the separate original TSN codebase
• No mention of pretrained weights, multi-GPU support, or modern data loaders
• README contains a typo (“bettter”) that has apparently survived unmolested

Verdict

Worth a quick look if you’re tracing the lineage of video action recognition in PyTorch, or need to reproduce a 2016-era baseline exactly. Otherwise, the authors’ recommendation stands: use MMAction instead.