Generative Models

SSL-VAE

PyTorch implementation of the method described in Semi-Supervised Learning with Deep Generative Models.

MNIST

	N=100	N=1000
Paper M2	88.03%	96.40%
My impl. M2	92.85%	96.72%
Paper M1+M2	96.67%	97.60%
My impl. M1+M2	97.20%	97.55%

M1+M2 analogies using N=100:

(The original samples are in the first column)

SVHN

I could achieve reliable results only when I increased the classification weight to $\alpha=0.5$. The following result is the average of three runs.

	N=1000
Paper M1+M2	63.98%
My impl. M1+M2	73.20%

M1+M2 analogies using N=1000:

(The original samples are in the first column)

Score Matching, NCE, RealNVP

Score Matching

| |

NCE

| |