Implicit-Vae-Pytorch

This repository has two implementations of Semi-Implicit Variational Autoencoders (not finished yet):

1. The original Semi-Implicit Variational Inference paper and offitial github repo I used to reimplement.
2. Unbiased Implicit Variational Inference and offitial github repo I used to reimplement.

Usage

$ python3 main.py
  -d {bmnist,fashionmnist}, --dataset {bmnist,fashionmnist}
                              Indicate the dataset. It can take on one of these
                              values: [bmnist, fashionmnist]

  -n {sivi,usivi}, --method {sivi,usivi}
                              Specify the method. It can take on one of these
                              values: [sivi, usivi]

  -z Z_DIM, --z-dim Z_DIM
                              Number dimension of the latent space. If none passed,
                              defaults will be used

  -b BURN, --burn BURN        Number of burning iterations for the HMC chain

  -s SAMPLING, --sampling SAMPLING
                              Number of samples obtained in the HMC procedure for
                              the reverse conditional

  --mcmc-samples MS           Number of samples to be drawn from HMCMC

  --batch-size BTCH           Minibatch size

  -e EPOCHES, --epoches EPOCHES
                              Number of epoches to run

  -k K, --K K                 number of samples for importance weight sampling

  -t, --train                 If it is train or test

Dependencies

• numpy >= 1.17
• pytorch >= 1.4

Results (on MNIST only)

1. Training with batch size of 135 and 2000 epochs, the lowest variational bound was 133.39 at epoch 5 and the biggest log likelihood of -76.34 at epoch 1747.
2. None yet.

Problems

2. Implementation of 2. not functional (weird gradient values).