BridgetteLiu2026/transdiagnostic-nn

Three connectivity classifiers (nn_model.pt checkpoint plus a pickle-free safetensors copy), one folder per model. Each folder is meant to be uploaded as its own Hugging Face model repo; its README.md is that repo's model card (this file is not uploaded).

All are small PCA -> MLP classifiers on resting-state fMRI functional connectivity. Research prototypes, not diagnostic tools.

Folder	Task	Input features	Train n	5-fold CV acc. (majority)	Status
abide/	ASD vs control	19,900 (CC200)	846	64.1% +/- 1.6% (53.8%)	modest signal
cobre/	schizophrenia vs control	7,381 (BASC-122)	146	67.6% +/- 5.4% (50.7%)	modest signal; logistic regression does as well
adhd200/	ADHD vs control	1,128 (Harvard-Oxford 48)	40	57.0% +/- 10.9% (50.0%)	not above chance, published for reproducibility only

CV figures are optimistic (hyperparameters were tuned on the same folds) and do not measure cross-site generalization. See each model card for details.

Folder contents

<dataset>/
  nn_model.pt         training checkpoint (torch pickle; load with weights_only=False)
  model.safetensors   same scaler, PCA and MLP weights, pickle-free
  config.json         n_features, n_components, hidden sizes, classes, threshold
  inference.py        predict(features) -> (P(class 1), labels)
  README.md           model card (YAML front matter + details)

nn_model.pt is copied from output/<dataset>/; model.safetensors and config.json are generated by export_hf.py; inference.py and README.md are hand-written and never overwritten by the export.

Regenerate

From the project root, after training (output/<dataset>/nn_model.pt exists):

python export_hf.py abide      # or adhd200 / cobre (default)

Quick check

cd hf_upload/cobre
python inference.py features.npy     # features.npy shape: (n, 7381)

Requires numpy, torch, safetensors. Features must come from the exact atlas and pipeline listed in the model card; other atlases will not work.

Upload

pip install -U huggingface_hub
huggingface-cli login
huggingface-cli upload <user>/<repo-name> hf_upload/cobre .

Repeat per folder with a different repo name. Notes:

• license in the front matter: cc-by-4.0 for COBRE; other for ABIDE and ADHD-200 (their data-use terms apply, check them before publishing).
• nn_model.pt is a pickle: each model card carries a loading snippet and a trust warning. Drop it from the upload if you want a pickle-free repo.
• No subject-level data is included, but the scaler/PCA parameters are statistics of the training subjects.
• Before uploading, review the model cards; the adhd200 card states the model is unusable and should keep that warning.

Citations

• ABIDE: Di Martino et al., 2014 (INDI / Preprocessed Connectomes Project).
• COBRE: Abraham et al., 2017, Figshare DOI 10.6084/m9.figshare.1450804.
• ADHD-200: ADHD-200 consortium / INDI.