Reverse image search, similarity models, and image retrieval
This repository condains a PyTorch implementation of Grafit, which is based on this BYOL implementation. In addition to adding the KNN loss from Grafit, this implementation also uses Albumentation
for augmentation.
Grafit TODO: Create KNN tensor, project cosine similarity operation, make sure it's detached, build KNN loss. Define list of augs, build initial embedding loop, build embedding update for batch, enhance instance loss with N augmentations