mihirneal/walnut-v0-1

Walnut v0.1

Pretrained 3D ViT-MAE checkpoints for structural MRI.

The vitb and vitl directories contain models trained with subject-specific sampling at four dataset scales. These runs used a fixed compute budget of 99,000 optimizer steps with global batch size 256. Each directory contains the final checkpoint and its training configuration.

vitl/sub-52k/ddep8 is the corresponding full-subject ViT-L experiment with an eight-layer decoder.