SlepNet

This repository contains SlepNet, a graph neural network designed to process spatiotemporal brain signals by learning localized spectral filters using graph Slepian functions. SlepNet focuses on optimally concentrating signal energy within anatomically meaningful subgraphs while respecting the global graph structure.

---

Overview

The provided Python training script applies SlepNet to neuroscience graph datasets (e.g., fMRI, EEG). It utilizes:

• PyTorch and PyTorch Geometric for GNN construction and training
• Weights & Biases (wandb) for experiment logging
• scikit-learn for data splits and classification utilities
• tqdm for real-time progress bars

Key components:

• Computes graph Laplacian and Slepian basis vectors (with optional neural eigenmapping)
• Learns node masks and localized filters over selected subgraphs
• Outputs graph-level or embedding-level predictions
• Supports Tier 1 (diagnosis) and Tier 2 (downstream) classification tasks

---

Requirements

Install all required Python dependencies with:

pip install -r requirements.txt

---

Arguments

You can customize the behavior of SlepNet using the following command-line arguments:

Argument	Type	Description	Default
--raw_dir	str	Path to the dataset directory	model_paths_ASD
--task	str	Task type: "classification" or "embedding"	classification
--hidden_dim	int	Hidden dimension size for each MLP or spectral layer	250
--num_layers	int	Number of stacked SlepNet layers	3
--num_slepians	int	Number of Slepian basis vectors used per layer	10
--lr	float	Learning rate for optimizer	0.001
--wd	float	Weight decay coefficient	0.0005
--num_epochs	int	Total number of training epochs	100
--batch_size	int	Batch size for training and evaluation	32
--gpu	int	Index of GPU to use; set to -1 to run on CPU	0
--use_neuralmap	flag	If included, use a learnable neural eigenmap instead of standard eigh	(optional)
--num_runs	int	Number of repeated training runs for statistical robustness	10
--layer_type	str	Type of SlepNet layer used per block. Options: batch_energy, batch_distance,	batch_energy
--dataset	str	Dataset to train on.	pvdm

Running the Script

Before running, ensure that --raw_dir points to a directory containing compatible PyTorch Geometric Data objects. The data loading and preparation are handled internally (e.g., by the load_graphs_from_directory utility or a similar function in your pipeline). Consult the data loading code for the expected format.

To run a standard training session on GPU:

python main.py --raw_dir path_to_data

Weights & Biases Integration

We utilize Weights & Biases for visualization of the losses and accuracies. The script automatically logs training metrics and experiment configurations to Weights & Biases (wandb) if you have an account and have run the following command:

wandb login