TorchSSL

Why we included this project

TorchSSL is the official PyTorch implementation behind FlexMatch, the curriculum pseudo-labeling paper from NeurIPS 2021, and it bundles nine classic semi-supervised learning methods, including PiModel, Mean-Teacher, MixMatch, and FixMatch, under one interface so head-to-head results are easy to compare. The caveat is real: the project is deprecated, and the README points you to its successor, USB, which trains in roughly an eighth of the time and also covers NLP and audio. Treat TorchSSL as a readable historical reference for how these algorithms are structured and for reproducing published numbers, not as a base for new work. That role still earns its keep, because having the internals of pseudo-labeling and consistency-regularization methods sitting side by side in one codebase makes studying them far simpler.