Decoder-based latent variable models through thelens of Fisher divergence

Here is a subset of the code for the paper.

Requirements

To install requirements:

pip install -r requirements.txt

To use the notebook, you should also install the jupyter notebook kernels.

Training

We provide training code for the following experiments:

1. Inference over 2d latent space with Fisher divergence and KL divergence: run_2d.py, two_d.ipynb
2. Training VAE model with objectives proposed in the paper: run_ae.py, denosing_vae.ipynb
3. Bilevel optimization for VAE models: run_bilevel_toy.py

All the scripts above should run without any arguments. See the argument list for specific settings.