Minimal implementation of SimSiam (Exploring Simple Siamese Representation Learning by Xinlei Chen & Kaiming He) in TensorFlow 2.

The purpose of this repository is to demonstrate the workflow of SimSiam and NOT to implement it note to note and at the same time I will try not to miss out on the major bits discussed in the paper. For that matter, I’ll be using the Flowers dataset.

The authors emphasize the stop_gradient  operation that helps the network to avoid collapsing solutions. Further details about this are available in the paper. SimSiam eliminates the need for using large batch sizes, momentum encoders, memory banks, negative samples, etc. that are important components of the modern self-supervised learning frameworks for visual recognition. This makes SimSiam an easily approachable framework for practical problems.

Following depicts the workflow of SimSiam (taken from the paper) -

The authors have also provided a PyTorch-like psuedocode in the paper (how cool!) -

#machine learning

Minimal implementation of SimSiamin TensorFlow 2
1.45 GEEK