byol-pytorch

Why we included this project

If you train vision models and are short on labeled data, this library gives you a small, clear way to add self-supervised pretraining. It implements BYOL, which learns from unlabeled images without negative pairs by matching a moving-average target encoder. You wrap any image-based network, like a ResNet, feed it raw images, and train; afterward the same network typically performs better on your supervised downstream tasks. The code is minimal and pip-installable, so it's easy to follow for researchers and easy to use for engineers wanting a quick pretraining step before fine-tuning on a limited labeled set.