NaN/Inf present in the evaluation of the MoG proposal posterior SEIRb

[ali-akhavan89]

I've reported something like this before on a toy Gaussian task, so I have a hunch of what's happening, but I just wanted to report this FYI. I'm using MDN, so there should be no leakage even though I'm working on SEIRb. I've attached the model.

1. SEIRb NPE MDN vs MAF v02.py.zip