javier-rozalen/vfpg

Variational Feynman Path Generator

A Generative-Model Approach to Path Integrals

What is this repository?

The repository contains the code generated for my Master's Thesis.

Repository structure:

.
├── MonteCarlo/                               # Path Integrals with MCMC 
├── example_plots/                            # Assets
├── vae/                                      # VAE code
├── CITATION.cff                              # Citation file
├── LICENSE                                   # License
├── README.md

Requirements

The machine learning part of the code in the files above is written in PyTorch. It does not come with the default Python 3 installation; to install it, go to Official PyTorch page or type:

pip3 install torch

Also, the progress bar tqdm is used. To install it:

pip3 install tqdm

Finally, the numpy library:

pip3 install numpy

Usage guide

---

Step 1. Generating paths with MCMC.

We open the file MonteCarlo/main_mcmc.py, set the desired initial parameters and run the file. An explanation of the adjustable parameters can be found at the beggining of the file. If the saving parameters were set to True, the program will save the data under the MonteCarlo/saved_data/ folder (created automatically).

Example of the results:

Step 2. Training the VAE.

We repeat the process of Step 1, but this time with the file vae/main_vae.py. This will train a VAE using the paths generated in Step 1 and, if desired, save the model for posterior experiments.

Example of the training loss evolution:

Step 3. Generating paths with VAE.

Once we have some generated data, we go to the vae/sampling_from_vae.py file, set the desired initial parameters and run the file. Again, an explanation of the adjustable parameters can be found at the beggining of the file. This will plot a ground-state wave function computed with VAE-generated paths, along with some of these paths.

Example of the GS density and some paths generated by the VAE:

---

Contact & Support

If you have any questions or issues, please contact us at [email protected].