bayesian_dnns

Practice & experiment of bayesian deep neural networks with pixyz.

GMVAE

A pixyz implementation of Gaussian Mixture Variational Auto Encoder proposed by Rui Sue.
By using gaussian-mixture prior for the generative model, its robustness for imbalanced data is much higher than Kingma's m2 model.
I also refered to the pytorch implementation by jariasf

Example: Imbalanced MNIST

Data

• labelled
  [label:number of images] [0:1000, 1:10, 2:10, 3:10, 4:10, 5:100, 6:70, 7:40, 8:50, 9:30]
• unlabelled
  Total 50000 images. Sampling ratio of each labels are same as labelled data (imbalanced).
• validation
  Total 10000 images. Sampling ratio of each labels are equal (balanced)

Result

Kingma's M2 model

Latent variables(dimension 0 and 1) and reconstructed images.

GMVAE

Latent variables(dimension 0 and 1) and reconstructed images. You can see that each label seems to have its own distribution.