MFreidank / pysgmcmc / 172

90%

push

travis-ci

svgd back in; with notebook to showcase

Stein Variational Gradient Descent was still missing in this repository;
this commit adds it back in.
This is not yet perfectly integrated into our framework, in particular
with respect to our BNN. The following desiderata are not yet met:
    1. SVGD should not take a single tf.Variable with all particles as
       input, but rather a list of particles - each a single
       tf.Variable. We might solve this issue by using tf.pack (or
       something similar) to obtain a single tensor from that list.
    2. BNN needs to add some code to support SVGD. In particular, instead of
       maintaining one network obtained by callable `get_net` it needs
       to maintain a network per SVGD particle.

After fixing these open issues, SVGD can be added to
`sampling.Sampler.get_sampler` and we should be able to fit BNNs using
SVGD, without any overhead for the user and without changing existing
code for any of the other samplers.

53 of 53 new or added lines in 3 files covered. (100.0%)

465 of 723 relevant lines covered (64.32%)

0.64 hits per line