Replication of Semi-Supervised Learning with Deep Generative Models

• Paper: Semi-Supervised Learning with Deep Generative Models
• Authors: Diederik P. Kingma, Danilo J. Rezende, Shakir Mohamed, Max Welling
• Original Implementation: github

Implements the latent-feature discriminative model (M1) and generative semi-supervised model (M2) from the paper in TensorFlow (python).

Dependencies

• TensorFlow >= 0.8.0 (due to prettytensor, might work with older versions of prettytensor - not tested)
• prettytensor
• numpy
• optionally matplotlib, seaborn for VAE images

Usage

• To train latent-feature model (M1) run train_vae.py. Parameters set in same file.
• To train M1+M2 classifier run train_classifier.py. Parameters set in same file. Location of saved M1 (VAE) model must be specified.
• Using the provided VAE model and the given parameters should produce an accuracy of about 95.4% on the test set using 100 labelled examples.

Example of the style and orientation learnt by the generative semi-supervised model in the latent variable (z) on the MNIST dataset. Generated using this implementation to replicate qualitative results from the paper.

Credits

1. Semi-Supervised Learning with Deep Generative Models
2. Implementation by Authors