Generated samples are very homogenous for conditional sampler #9
Description
I've noted that generated samples for MNIST are quite homogenous and originally thought this was an implementation issue:
Having looked at this a little more and playing around with the way losses are computed (on an instance basis vs. aggregated across batch) I'm no longer sure. I also noted that the unconditional samples in Tom are also homogenous:
