I explore SimCLR pre-training framework proposed by Google in arXiv:2002.05709 paper. I explain the SimCLR and its contrastive loss function step by step, starting from naive implementation followed by faster, vectorized one. Then I show how to use SimCLR’s pretraining routine to first build image embeddings using EfficientNet network architecture and finally how to build classifier on top of it.
This post covers:
Don't forget to tag @marrrcin in your comment, otherwise they may not be notified.