shilu10/BYOL-tensorflow2

BYOL - TensorFlow 2

This is an unofficial implementation of "Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning" (BYOL). It is uses two networks namely online and target network,and target network parameter, will updated in exponential moving average fashion. It is similar to DQN.

Architecture

Dataset

For Pretraining the BYOL, CIFAR-10 dataset is used. Also the Linear Evaluation done using the CIFAR-10 dataset.

• Pretraining dataset : CIFAR-10 dataset.

• DownStream dataset : CIFAR-10 dataset.

References

• Jean-Bastien Grill et al., "Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning", 2020

• https://sthalles.github.io/keras-regularizer/
• https://github.com/garder14/byol-tensorflow2
• https://www.kaggle.com/code/shilashm/byol-tensorflow-2-0/edit