Skip to content

README.md

Latest commit

 

History

History
57 lines (43 loc) · 2.05 KB

README.md

File metadata and controls

57 lines (43 loc) · 2.05 KB

Graph Neural Network Hyperparameter Optimization

This project focuses on tuning the hyperparameters of various Graph Neural Network (GNN) models and provides implementations for the APPNP, Splineconv, and GAT models using PyTorch Geometric.

Overview

  • Hyperparameter Optimization: Using Bayesian optimization provided by Optuna to find the best hyperparameters that lead to optimal model performance.

  • Model Implementations: Built-in support for APPNP, Splineconv, and GAT architectures. Each model has its own customizable parameters.

Features

  • Model-specific hyperparameter search space.
  • Modular design for easy expansion to new GNN architectures.
  • Pre-defined activation functions for flexibility.
  • Comprehensive evaluation on graph datasets like Cora, PubMed, and CiteSeer.

Requirements

  • Python 3.x
  • torch
  • torch_geometric
  • optuna

Usage

  1. Clone the repository:
git clone https://github.com/darkzard05/planetoid_search_hyperparam.git
  1. Install the requied packages:
pip install -r requirements.txt
  1. Run the main script with desired arguments:
python main.py --dataset=Cora --model==APPNP --n_trials=100 --epochs=100 --batch_size=1024 --num_neighbors=[10,10] --num_workers=0

Arguments

  • model: Model to be used. (APPNP, Splineconv, GAT)
  • dataset: Dataset to be used. (Cora, PubMed, CiteSeer, Reddit)
  • n_trials: Number of trials.
  • epochs: Number of epochs per trial.
  • batch_size: set data per iteration. (Reddit)
  • num_neighbors: neighbors sampled in graph layer. (Reddit)
  • num_workers: how many subprocesses to use for data loading. (default: 0)

Models

APPNP

  • Implements the Approximated Personalized Propagation of Neural Predictions (APPNP) layer.

Splineconv

  • Utilizes B-spline basis functions to hierarchically partition and transform the input graph.

GAT

  • Graph Attention Networks (GAT) use attention mechanisms to weigh neighbor features.