-
Update Swift pretraining checkpoint
-
Update task1 neurostorm ckpts
zxcvb6958
-
Add pt_neurostorm_mae_5ds: NeuroSTORM MAE pretrained on 5 datasets
zxcvb6958
Pretrained checkpoint for NeuroSTORM with masked autoencoder objective
on 5 fMRI datasets (HCP1200, HCPA, HCPD, ABCD, UKB) using hard_random
reweighting (topk=100 subjects/dataset/epoch).
- Architecture: 4D Swin Transformer + Mamba (embed_dim=36, depths=[2,2,6,2])
- MAE settings: spatial_mask=window, time_mask=random, mask_ratio=0.5
- Trained 20 epochs, best epoch 18, valid_loss=0.50, test_loss=0.46
- Split: 8/1/1 train/val/test -
Reorganize: flatten pretraining/, add neurostorm task1 ckpt
