Shio-Koube's picture
Create README.md
57bfae5 verified
metadata
license: mit

This is a MAE trained on Anime dataset. The main goal is to have a model efficient for image search, retrival and clustering.

There are 2 parts of this model, the encoder and decoder. The encoder encode the full images into 8x512 embedding and the masked out image into 8 (28x28/10) x 512 embedding. The decoder try to reconstruct that image.

Model arch is LocalViT small but with 16 layers instead of 12, Decoder is a simple transformers model with LocalViT style MLP.