Amortized SVGD is a simple method which can be utilized to train black box inference networks to draw samples.

We demonstrate two examples utilizing amortized SVGD in this repository:

We demonstrate we can draw samples from intractable posteriors by applying amortized svgd to train variational autoencoders with a non-Gaussian encoder, where we add dropout rate 0.3 noise in each layer. The experiments shows that our method can capture the multimodal posterior due to the new structure.

We use the implementation of Wu et al., 2017 to evaluate the test log-likelihood.

VAE	SteinVAE
TestLL: -84.68	Test LL: -84.31

We use a Langevin sampler structrue to sample parameters of bayesian logistic regression.

• Theano
• pylearn2
• matplotlib
• sklearn
• tqdm

We would put the paper on Arxiv soon.