https://github.com/fkodom/byol

PyTorch implementation of BYOL: a fantastically simple method for self-supervised image representation learning with SOTA performance.
https://github.com/fkodom/byol

machine-learning machine-learning-algorithms mit-license python3 pytorch pytorch-implementation

Last synced: about 2 months ago
JSON representation

PyTorch implementation of BYOL: a fantastically simple method for self-supervised image representation learning with SOTA performance.

• Host: GitHub
• URL: https://github.com/fkodom/byol
• Owner: fkodom
• License: mit
• Created: 2020-08-20T02:07:21.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2021-03-24T15:57:40.000Z (over 3 years ago)
• Last Synced: 2024-10-23T03:16:35.687Z (2 months ago)
• Topics: machine-learning, machine-learning-algorithms, mit-license, python3, pytorch, pytorch-implementation
• Language: Python
• Homepage:
• Size: 31.3 KB
• Stars: 9
• Watchers: 2
• Forks: 2
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.txt

Awesome Lists containing this project

README

        
# BYOL: Bootstrap Your Own Latent

PyTorch implementation of [BYOL](https://arxiv.org/abs/2006.07733): a fantastically simple method for self-supervised image representation learning with SOTA performance.  Strongly influenced and inspired by [this Github repo](https://github.com/lucidrains/byol-pytorch), but with a few notable differences:

1. Enables **multi-GPU** training in PyTorch Lightning.

2. (Optionally) Automatically trains a linear classifier, and logs its accuracy after each epoch.

3. All functions and classes are fully type annotated for better usability/hackability with Python>=3.6.

## TO DO

* Enable mixed-precision training in PyTorch Lightning.  `kornia.augmentation.RandomResizedCrop` currently doesn't support this.  I'll need to ensure that our implementation is sufficiently performant, so it doesn't inadvertently slow down training.