HMC-VAE

This repository contains code for hierarchical convolutional variational autoencoders with Hamiltonian Monte Carlo, which accompanies my MPhil thesis at University of Cambridge.

Directory Structure

The reposiroty contains 7 Python files:

models/hmc.py      # Basic version of Hamiltonian Monte Carlo
models/blocks.py   # Conv/ConvTranspose building blocks for the model
models/vae.py      # VAE with hierarchical latent variables
models/hmc_vae.py  # VAE with hierarchical latent variables and HMC
image_dataset.py   # Download and preprocess image datasets
utils.py           # Utility for turning on/off parts of a model
train.py           # Main entry point for traning

And my thesis thesis.pdf that details theoretical background, implementation details, and experimental results.

Example

The command below trains a model with 2 latent layers each containing 30 units, and 64 filters, on the dataset CIFAR10 (which is stored in ./data), for 100 epochs (100 variational epochs and hard-coded 3 HMC epochs), with the test set evaluated after every 1 epoch, where both trained model and test set scores are written to /path/to/save/.

python3 train.py --data_name CIFAR10 --data_root ./data \
    --hidden_channels 64 --epochs 100 --eval_interval 1 \
    --savedir /path/to/save/ --latent_dims 30 30

Architecture Visualization (1 latent layer)

Encoder

Decoder

Citation

@Misc{HMC-VAE,
    author = {Haoran Peng},
    title = {Outlier Detection with Hierarchical VAEs and Hamiltonian Monte Carlo},
    year = {2022},
    url = "https://github.com/GavinPHR/HMC-VAE"
}