Project Overview

Code repo of paper "Learning Generalized Hamiltonian Dynamics with Stability from Noisy Trajectory Data".

This repository provides a flexible, YAML‑driven framework for:

• Generating physics datasets (e.g. pendulum trajectories)
• Training a variety of models (SSGP, HNN, etc.) with different trainers (GDA, Equal, No‑Reg)
• Validating and visualizing results
• Benchmarking multiple methods over multiple systems
• Sweeping hyperparameters in parallel on HPC clusters

Everything is configured via YAML files.

---

Prerequisites & Setup

1. Python 3.8+
2. Clone the repo:

   git clone <repo_url>
   cd <repo>

3. (Optional) Create a virtual environment:

   python -m venv .venv
   source .venv/bin/activate

4. Install dependencies:

   pip install numpy torch torchdiffeq matplotlib tensorboard PyYAML torchjd

---

Configuration

• configs/default.yaml selects one variant from each subgroup:

  dataset:   windy_pendulum.yaml
  model:     ssgp.yaml
  optimizer: adam.yaml
  trainer:   gda.yaml
  
  data_root: "data/"
  mode:      train            # or "visualize"
  save_dir:  "checkpoints/{trainer}/{dataset}"

• Subgroup YAMLs define module, class, and args: for each component.
• Support some CLI overrides let you tweak any field without editing files:

  python main.py     --config configs/default.yaml     -O optimizer.args.lr=0.0005     -O trainer.args.epochs=200     -O save_dir="checkpoints/gda/custom_run"

---

Running Training or Visualization

Train

python main.py --config configs/default.yaml

Dataset are listed in data/ folder, one can verify and regenerate it via scripts/generate_all_datasets.py

Parallel Training

1. Define experiments.csv with columns: dataset, lr, optimizer, model, trainer, epochs, save_dir
2. Run:

   sbatch scrips/run_experiment_array.slurm

Visualize

python main.py   --config configs/default.yaml   -O mode=visualize   -O checkpoint="checkpoints/gda/windy_pendulum_ssgp_adam.pth.tar"

We also provide some standalone visualization scripts (for phase space) in scripts/visualize and visualize_results.py.

TensorBoard

TensorBoard logs are by default written to runs/checkpoints/. To inspect:

tensorboard --logdir runs/checkpoints

---