discrete_vae:
A discrete VAE model with N categorial latent variables of size K.
The gradients can be estimated with gumbel-max / gumbel-softmax reparametrizations or without a reparametrization (unbiased).
[Note that the unbiased estimator works only with N=1]
mixture_model_mnist:
contains mixture of gaussian and discrete VAE (unsupervised or semi-supervised) [mnist]
To run the above, use the scripts run_direct_gsm.py, run_mixture.py respectively.
Impementation of a structured-VAE is available only with direct_vae: run_structured_cplex.py, run_structured_maxflow.py
Check out the paper for more details https://arxiv.org/abs/1806.02867
GuyLor/Direct-VAE
Implementation of the paper "Direct Optimization through argmax for discrete Variational Auto-Encoder"
Python