jaberg / montetheano Goto Github PK

Now that we have mixture models - we can test that sampling distributions, cdf, and pdf match. Essentially the tests should draw a bunch of samples - put down gaussians around the samples, and then show that the gaussian mixture has about the same pdf as the original distribution.

There are many distributions in scipy.stats. They come with logpdf, cdf, logcdf, etc. Make these available via a few standard Ops.

Write a generic test that verifies lpdf functions sum to 1 in particular cases where numerical error can be kept under control.