Variational autoencoder implementation using an architecture consisting of Dense blocks.
Please cite "Disentangling Generative Factors of Physical Fields Using VariationalAutoencoders" (Jacobsen 2021). Portions of the code are adapted from "Bayesian Deep Convolutional Encoder-Decoder Networks for Surrogate Modeling and Uncertainty Quantification" (Zhu 2018), see paper for more in depth explanation on architecture.
Data used in paper is too large to post here. Matlab scripts are included to generate data if desired. "Solve_Darcy.m" solves a 2D darcy flow problem and saves the data. The generative parameter distribution and factors of variation can be adjusted.
The PyTorch dataloader framework is used with "load_data_new.py" specifying the dataset. If using other data, ensure it follow this format.
Train a VAE by running "train_VAE.py". This will train a VAE with the configuration found in "config_train.py" and save to the specified location. Do not modify the name of this file.
Evaluation script not currently provided. Training saves the VAE to .pth file. Load weights into initialized model using pytorch method model.load_state_dict(state_dict)