Learning Physical Constraints with Neural Projections

Dependencies

All the dependencies are standard python packages:

• Pytorch, Numpy, Matplotlib, PIL (used for visualization)

Code Structures:

• _*.py: Some common classes/functions used for training and testing.
  • _constraint_net.py: A simple network used to represent the constraints to be learned
  • _iterative_proj.py: The iterative projection to solve the constrains
  • _training.py: Utils used for training
  • _dataloader.py: Load training data
  • _run_simulation.py: Run simulation use the learned constraint net and the projection operator
• training_*.py: Train the model.
  • Data used for training can be found in this shared google drive folder.
  • The trained models are in \model folder.
• simulation_*.py: Use the trained model to generate simulations.
  • Simulation results will be written to \results folder.
  • In "evaluation" branch, simulation scripts are modified to run multiple simulation samples

Webpage:

https://y-sq.github.io/proj/neural_proj/