Skip to content

Release subcortical prediction checkpoint (MaskProjector with 8,808 voxels) #23

@x0rc1st

Description

@x0rc1st

Hi team, great work on TRIBE v2!

We noticed the codebase includes full infrastructure for subcortical predictions:

  • tribev2/plotting/subcortical.py — subcortical mask, labels, voxel-to-mesh conversion
  • tribev2/grids/run_subcortical.py — training grid config with "mask": "subcortical" and
    MaskProjector
  • get_subcortical_labels() returns 16 bilateral structures (hippocampus, amygdala, thalamus,
    caudate, putamen, pallidum, accumbens, lateral ventricle) totaling 8,808 voxels

However, the published checkpoint (best.ckpt on HuggingFace) only includes the cortical surface
head (predictor.weights: [1, 2048, 20484] on fsaverage5).

Request: Would you consider releasing a subcortical checkpoint with the MaskProjector head
(2048 → 8,808)? The shared encoder and low_rank_head weights from the current checkpoint should
transfer — only the final predictor layer would differ.

The training code and data pipeline for subcortical targets appear ready in run_subcortical.py.
If there's a technical reason the subcortical head wasn't included, or if fine-tuning
instructions are available, that would also be helpful.

@sdascoli any insights on this would be highly appreciated....

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions