My Pytorch implementation of Berkeley Deep Unsupervised Learning from Spring 2019
Only notable results have been included

Note that scalar variance data set 1 KL is almost 0

Note how mean of vector variance model is centered at 0

Top: Data colored. Note, how the latent of vector variance is completely scrambled, no data is stored. Reconstructions are similar. This shows latent is not used. This is due to the data being able to be completely modeled by the decoder with no help of the latent.

This time both models behave the same due to the multivariate diagonal covariance gaussian being insufficient to model the data alone.