CS294-158-SP19

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

Homework 1 Results

Exercise 1.2

MADE network on 2d data:

PixelCNN results

Homework 2 Results

Original Data, note colors

Autoregressive Flow results

Autoregressive Flow Latent Space

Simple Real NVP Results

Simple Real NVP Latent and Generated Samples

Simple Real NVP Latent Colored

Homework 3 Results

Model Training Curves.

Note that scalar variance data set 1 KL is almost 0

Results on Data set 1

Note how mean of vector variance model is centered at 0

Latent Space of Data set 1 Models

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.

Results on Data set 1

Latent Space of Data set 1 Models

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

IWAE Training

IWAE Results

IWAE Latent space

SVHN Training Curve

SVHN Training Curve

SVHN Results

SVHN Interpolations

WGAN Training Curves

WGAN Results