pytorch-vae

A minimal pytorch implementation of VAE, IWAE, and MIWAE. We followed the experimental details of the IWAE paper.

Usage

You should be able to run experiments right away. First create a virtual environment using pipenv:

pipenv install

To run experiments, you simply have to use:

pipenv run python main.py <options>

Example commands

For original VAE:

pipenv run python main.py

To also make figures (reconstruction, samples):

pipenv run python main.py --figs

For IWAE with 5 importance samples:

pipenv run python main.py --importance_num=5

For MIWAE(16, 4):

pipenv run python main.py --mean_num=16 --importance_num=4

See the config file for more options.

Results

Quantitative results on dynamically binarized MNIST