decon-hnn's Introduction

Deconstructing The Inductive Biases of Hamiltonian Neural Networks

This repo contains the implementation and the experiments for the paper

Deconstructing The Inductive Biases of Hamiltonian Neural Networks by Nate Gruver, Marc Finzi, Sam Stanton, and Andrew Gordon Wilson.

Code

Our repo was constructed from https://github.com/mfinzi/constrained-hamiltonian-neural-networks.git

Installation instructions

Pip

git clone https://github.com/ngruver/decon-hnn.git
cd decon-hnn
pip install -r requirements.txt

Conda

git clone https://github.com/ngruver/decon-hnn.git
cd decon-hnn
conda env create -f decon-hnn.yml

View experimental runs and recreate paper figures

All figures from the paper can be recreated by running the notebooks

energy_plots.ipynb

bar_plots.ipynb

These notebooks process data from the following wandb sweeps

https://wandb.ai/ngruver/physics-uncertainty-exps/sweeps

https://wandb.ai/samuelstanton/physics-uncertainty-exps/sweeps

All experimental data, and the associated configurations, are contained in these sweeps.

Train pendulum models

You can train the models NN (NODE), MechanicsNN (NODE + SO), and HNN using the model_type option as shown below.

python toy_systems.py --system_type "ChainPendulum" --model_type "NN"
python toy_systems.py --system_type "ChainPendulum" --model_type "MechanicsNN"
python toy_systems.py --system_type "ChainPendulum" --model_type "HNN"

The other systems, with and without friction, can be specified as SpringPendulum, FrictionChainPendulum, and FrictionSpringPendulum.

Train Mujoco models

To train models on mujoco, you must first download our saved mujoco trajectories with full state and velocity.

Mac

brew install gdrive
gdrive download 1Vdf8rjPXabfMaCouNfqUYf0ifDW3qAU2 --recursive
mv full_state_mujoco_trajs data

Linux

pip install gshell
gshell init
gshell download --with-id '1Vdf8rjPXabfMaCouNfqUYf0ifDW3qAU2' --recursive
mv full_state_mujoco_trajs data

Once the data has been downloaded, NODE, CoupledNODE(NODE + SO), and MixtureHNN (SymODEN) models can be trained as shown below.

python mujoco.py --model_type "NODE" --task "HopperFull-v0"
python mujoco.py --model_type "CoupledNODE" --task "HopperFull-v0"
python mujoco.py --model_type "MixtureHNN" --task "HopperFull-v0"

The other mujoco tasks included in the paper can be specified as SwimmerFull-v0 and HalfCheetahFull-v0

Citation

If you find our work helpful, please cite it with

@inproceedings{
  gruver2022deconstructing,
  title={Deconstructing the Inductive Biases of Hamiltonian Neural Networks},
  author={Nate Gruver and Marc Anton Finzi and Samuel Don Stanton and Andrew Gordon Wilson},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=EDeVYpT42oS}
}

decon-hnn's People

Contributors

Stargazers

Watchers

decon-hnn's Issues

Some problem with mujoco data downloading

Hello!
It's really an exciting work and we really appreciate your released codes ! But we found some errors when we tried to download the mujoco training data you provided. Maybe something wrong with the authentication and vertification codes of gdrive api but we still cannot solve it after a few days' trying. If it is convenient for you, could you provide the direct link for us please to download the data via google drive instead of gdrive which occurs some bugs for us?
We really appreciate your time and help! Best wishes!

Code can't run

Running the command line python toy_systems.py --system_type "ChainPendulum" --model_type "MechanicsNN" returns the following error, could you please check?

NN
{'C': 5,
 'bs': 200,
 'data_seed': 0,
 'device': 'cuda:0',
 'loss': 'l2',
 'lr': 0.005,
 'n_systems': 1000,
 'net_cfg': {'hidden_size': 128},
 'net_seed': 0,
 'num_epochs': 2000,
 'regen': False,
 'tau': 10.0,
 'weight_decay': 0.0001}
Traceback (most recent call last):
  File "/home/ubuntu/decon-hnn/toy_systems.py", line 68, in <module>
    Fire(main)
  File "/home/ubuntu/anaconda3/envs/check/lib/python3.9/site-packages/fire/core.py", line 141, in Fire
    component_trace = _Fire(component, args, parsed_flag_args, context, name)
  File "/home/ubuntu/anaconda3/envs/check/lib/python3.9/site-packages/fire/core.py", line 466, in _Fire
    component, remaining_args = _CallAndUpdateTrace(
  File "/home/ubuntu/anaconda3/envs/check/lib/python3.9/site-packages/fire/core.py", line 681, in _CallAndUpdateTrace
    component = fn(*varargs, **kwargs)
  File "/home/ubuntu/decon-hnn/toy_systems.py", line 60, in main
    trainer = make_trainer(**cfg, network=net, body=body,
TypeError: make_trainer() got an unexpected keyword argument 'loss'

Recommend Projects

ngruver / decon-hnn Goto Github PK