PyTorch implementation of Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning.
This repository contains training pipeline for BYOL, and I reimplement it with PyTorch. It support CIFAR-10, CIFAR-100, STL-10, TinyImageNet-200 and ImageNet. It also provides the linear probing code, which I borrow from sthalles. Thanks for their excellent project :)
I recommand you to put your dataset into ./data/xxx:
./data/STL-10
./data/CIFAR-10
...
You need first create environment and then run train.sh:
conda env create --name byol -r requirements.txt
conda activate byol
sh train.shIf all goes well, the final result will be approximately 73.8%.
Feel free to contact me if you have any problem:)