lggnn

Run the reference training script with an atlas name and labels CSV, for example train_reference.py with --atlas aal_116 and --fold 0. It loads PyG brain graphs from the shared BrainGNN dataset directory and runs stratified K-fold cross-validation.

Pass --task regression with a continuous label column such as age, for example using the HCP age labels CSV. The script automatically sets nclass to 1, switches to plain K-fold splitting, and reports MAE instead of accuracy.

Yes, the NeuroClaw reimplementation is compatible with PyG 2.7. The removed topk_pool and filter_adj functions are replaced with pool.select.topk and an inline filter_adj implementation.

It consumes PyG Data objects with node features x of shape [N, N], edge_index, and edge_attr, stored as sub-*.pt files under data/braingnn_input/<atlas>/. This is the same format BrainGNN uses, so no additional preprocessing is needed.

The MI estimate starts near zero because the joint and marginal distributions are initially similar in the SABP pooling module. It should gradually become positive during training, and the loss subtracts 0.1 times the MI term to encourage maximization.

The adaptation removes the original population graph built from demographic data, so it only performs single-subject prediction with an MLP head. It is intended for research use only and depends on preprocessed connectivity data produced by the fMRI preprocessing pipeline.