ryanccarelli / ssljax

self-supervised learning in jax
GNU General Public License v2.0
8 stars 0 forks source link

DINO head #72

Closed ryanccarelli closed 2 years ago

ryanccarelli commented 2 years ago

3-layer MLP (2048 hidden dim) with GELU, last layer without GELU l2 norm weight normalized fully connected layer no batchnorm

https://github.com/facebookresearch/dino/blob/cb711401860da580817918b9167ed73e3eef3dcf/vision_transformer.py#L257