krakencoder_pca_stack.scfc-3atlas-pc256.1

Krakencoder PCA stack -- multi-flavor input/output linear projections -- SCFC 3-atlas PCA stack (15 flavors; 256-dim PC space)

Description

Krakencoder (Jamison, Gu, Wang, Tozlu, Sabuncu, Kuceyeski, Nature Methods 2025), ported to JAX / Equinox from the upstream PyTorch release (github.com/kjamison/krakencoder). A linked autoencoder that simultaneously bidirectionally translates between structural and functional brain connectivity across different atlases and processing variants ('flavors') via a common 128-dim L2-normalised latent representation. The Nature Methods 2025 publication's canonical model jointly encodes 15 flavors (3 atlases × {3 functional connectivity types + 2 structural tractography types}) and maps each to / from the shared latent. Architecture (per published recipe): per-flavor 256-dim PCA input transformation -> 256 -> 128 Linear encoder -> 128-dim L2-normalised latent -> 128 -> 256 Linear decoder -> inverse PCA to the destination flavor's full-dim connectivity space. v0 of this port ships the canonical bundle plus its 15-flavor PCA stack (separate krakencoder_pca_stack bundle that the KrakencoderPipeline co-loads).

Intended use

The 15-flavor PCA stack co-loaded by the canonical krakencoder.scfc-fs86_shen268_coco439-pc256.1 bundle. Each flavor's projection has shape (pca_components=(256, input_dim), pca_input_mean=(input_dim,)); input_dim varies per atlas (3,655 / 35,778 / 96,141 for FS86 / Shen268 / Coco439). Bundle weight: ~665 MB; loads in ~1.2s on modern hardware.

Usage

from ilex.models.krakencoder import KrakencoderPCAStack
model = KrakencoderPCAStack.from_pretrained('ilex-hub/krakencoder_pca_stack.scfc-3atlas-pc256.1')

Authors

Keith W. Jamison, Zijin Gu, Qinxin Wang, Ceren Tozlu, Mert R. Sabuncu, Amy Kuceyeski

Citation

Jamison K.W., Gu Z., Wang Q., Tozlu C., Sabuncu M.R., Kuceyeski A. (2025). Krakencoder: a unified brain connectome translation and fusion tool. Nature Methods. DOI: 10.1038/s41592-025-02706-2.

References

• Jamison K.W., Gu Z., Wang Q., Tozlu C., Sabuncu M.R., Kuceyeski A. (2025). Krakencoder: a unified brain connectome translation and fusion tool. Nature Methods. DOI: 10.1038/s41592-025-02706-2.
• Preprint: bioRxiv 10.1101/2024.04.12.589274.
• Upstream code: github.com/kjamison/krakencoder (model.py + fetch.py + per-flavor PCA transforms hosted on OSF: osf.io/dfp92).

License

HF Hub license tag: mit

Effective terms: MIT (copyright (c) 2024 Keith W. Jamison). PCA matrices derived from 710 HCP training subjects, hosted on OSF. The ilex JAX / Equinox port code is separately licensed under Apache-2.0 / GPL-3.0.

Upstream license reference: https://opensource.org/licenses/MIT

Copyright

Network architecture, training code, and pretrained weights -- copyright (c) 2024 Keith W. Jamison; released under the MIT License. JAX / Equinox port code -- copyright (c) the ilex authors, released under the Apache-2.0 / GPL-3.0 dual license used by ilex itself.

Upstream source

Original weights / reference implementation: https://github.com/kjamison/krakencoder

Provenance

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