What problem does it solve? Running the Com-BrainTF model for fMRI phenotype prediction normally requires adapting hydra-based configs and hardcoded community maps; this Skill provides a NeuroClaw reimplementation that auto-derives community partitions from atlas naming and trains the model with a unified classification/regression interface. ## Core Features & Use Cases - Two-level Transformer architecture: Per-community local Transformers with independent CLS tokens feed a global Transformer with DEC pooling for fMRI connectome prediction. - Automatic community partitioning: Derives Yeo 7-network groups for Schaefer atlases, lobe-based groups for AAL/Destrieux/DK/Harvard-Oxford, and hash-based fallback groups for other atlases. - Unified training pipeline: Supports classification and regression with stratified K-fold cross-validation, reusing the BNT dataset format and collate functions. - Use Case: Train Com-BrainTF on Schaefer-200 functional connectivity matrices to predict HCP subject gender, or run regression on AAL-116 data to predict subject age. ## Quick Start Run the training reference script with the schaefer_200_7net atlas, a labels CSV, fold 0, 10 epochs, and batch size 8 to smoke-test Com-BrainTF classification.
