cortex_mae.flat.1 / README.md
rciric's picture
Initial upload via tools/push_to_hf.py (architecture: ilex.models.cortex_mae.CortexMAE)
8311c51 verified
|
Raw
History Blame Contribute Delete
3.9 kB
metadata
library_name: ilex
tags:
  - jax
  - equinox
  - ilex
  - neuroimaging
  - fmri
license: cc-by-nc-4.0
license_link: https://creativecommons.org/licenses/by-nc/4.0/
gated: auto
extra_gated_prompt: |
  CortexMAE is distributed under the Creative Commons
  Attribution-NonCommercial 4.0 International license
  (CC-BY-NC-4.0). Permission is granted to use, copy, modify,
  and distribute the weights solely for non-commercial
  research and educational purposes. Commercial use --
  including clinical decision support, clinical workflows,
  products, or services for a fee -- requires separate
  written authorisation from MedARC-AI.

  By requesting access you affirm that your intended use falls
  within the CC-BY-NC-4.0 terms.
extra_gated_fields:
  Name: text
  Affiliation: text
  Email: text
  Intended use: text
  I acknowledge the CC-BY-NC v4 terms: checkbox
extra_gated_button_content: Acknowledge and download

CortexMAE (ViT-B MAE foundation model for fMRI) -- CortexMAE flat (2D flat brain map input)

Description

CortexMAE (MedARC-AI, arXiv 2510.13768) is a ViT-B masked autoencoder pretrained on 2.1K hours of HCP fMRI BOLD data. The architecture is the standard MAE (encoder + lighter decoder + MSE loss on reconstructed patches) with separate Q/K/V Linears (the original MAE convention) and separable spatial + temporal positional embeddings (mae-st style).

v0 ships the parcel variant only (cortex_mae_parcel -- Schaefer-400 1D parcellation input). The flat (2D brain map) and volume (3D MNI cortex) variants are queued as follow-ups (#157 + #158) on top of the same shared transformer body.

Intended use

Cortical flat-map 2D backbone. Input (1, 16, 224, 560): 1 BOLD channel, 16 temporal frames, 224x560 flat map. Output: cls embedding (1, 768) + patch embeddings (1960, 768) over 4 temporal x 14x35 spatial patches.

Usage

from ilex.models.cortex_mae import CortexMAE
model = CortexMAE.from_pretrained('ilex-hub/cortex_mae.flat.1')

Authors

MedARC-AI (research collective)

Citation

Adkinson R. et al. (2025). CortexMAE -- A masked autoencoder foundation model for fMRI. arXiv 2510.13768.

References

  • Adkinson R. et al. (2025). CortexMAE -- A masked autoencoder foundation model for fMRI. arXiv 2510.13768.
  • Feichtenhofer C. et al. (2022). Masked autoencoders as spatiotemporal learners (MAE-ST). NeurIPS.
  • Upstream code -- github.com/MedARC-AI/CortexMAE (Apache-2.0); weights -- huggingface.co/medarc/CortexMAE (CC-BY-NC-4.0).

License

HF Hub license tag: cc-by-nc-4.0

Effective terms: CC-BY-NC-4.0 on the released checkpoint weights (huggingface.co/medarc/CortexMAE, LICENSE.models). The upstream code (github.com/MedARC-AI/CortexMAE) is separately Apache-2.0. Non-commercial restriction on the weights requires gated distribution with explicit acknowledgement. The ilex JAX / Equinox port code is separately Apache-2.0 / GPL-3.0; non-commercial use only continues to apply to the weights regardless of the port-code licence.

Upstream license reference: https://creativecommons.org/licenses/by-nc/4.0/

Copyright

CortexMAE upstream code is copyright (c) MedARC-AI, Apache-2.0 on the code (github.com/MedARC-AI/CortexMAE, LICENSE) and CC-BY-NC-4.0 on the released checkpoint weights (huggingface.co/medarc/CortexMAE, LICENSE.models). The non-commercial restriction on the weights is preserved through gated HuggingFace distribution; the ilex JAX / Equinox port code is separately Apache-2.0 / GPL-3.0.

Upstream source

Original weights / reference implementation: https://github.com/MedARC-AI/CortexMAE

Provenance

This artefact was produced by ilex's save/load pipeline. The architecture is implemented in ilex.models.cortex_mae.CortexMAE and the weights have been converted from their upstream format. See the upstream source above for the canonical reference.