How campusai/hamiltonian-generative-networks works
Project type and purpose
• This is a research-oriented Python codebase implementing Hamiltonian Generative Networks (HGN). It provides:
• Physical environment simulators (“environments/”) for data generation.
• Neural network modules (“networks/”) for the encoder, transformer, Hamiltonian, and decoder.
• Utility functions (“utilities/”) for integration, data loading, logging, and loss computation.
• CLI scripts for dataset generation (generate_data.py), model definition (hamiltonian_generative_network.py), training (train.py), and sampling rollouts (sample_rollouts.py).
• Purpose: generate or load simulated physics datasets, train an HGN model end-to-end, and evaluate it.
File structure
• Top-level CLI scripts:
• generate_data.py → data creation (online/offline)
• train.py → orchestrates training with configs and CLI args
• sample_rollouts.py → model inference / visualization
• Core packages:
• environments/ → physical systems and a factory to instantiate them
• networks/ → PyTorch modules: encoder, transformer, Hamiltonian, decoder
• utilities/ → data conversions, integrator, loader, losses, logger, statistics
• Configuration:
• experiment_params/ → YAML templates for datasets, environments, and training
• Tests:
• tests/ → unit and gradient tests for modules and utilities
README for additional insights
• Dependencies: PyTorch, YAML, numpy, etc.
• Training modes:
• Online: simulate on the fly via environments + DataLoader
• Offline: load pre-generated datasets via loader utilities
• Core architecture described: encoder → transformer → Hamiltonian net → integrator → decoder