IB_GAN

Description

This repository is modified from artemyk/ibsgd to fit the experiment on GAN. GAN is trained on 3 activations (tanh, relu, leakyrelu) to compare their results.

Prerequisites

• python 3.5.6
• six 1.15.0
• keras 2.3.1
• matplotlib 2.0.2
• numpy 1.16.6
• seaborn 0.9.1

Usage

• save_activity_gan.py serves two purposes:

  • train GAN and save generator model
  • use the generator model to generate fake images, then concatenate it with real images, forming a new dataset. Next, save outputs of each layers while training GAN.

• computeMI_gan.py load rawdata and compute Mutual Information of each layers

Step 1

Implement this step to obtain generator model, skip this step if generator_tanh.h5 exist.

python save_activity_gan.py --activation tanh --isTrain 1

Step 2

Save outputs of each layer while training GAN.

python save_activity_gan.py --activation tanh --isTrain 0

Step 3

Compute Mutual Information and plot figures.

python computeMI_gan.py --activation tanh

Results

Browse plots/mnist for all results with different activations. Showing plots below with tanh experiments.

Information Plane

Loss vs Mutual Information (Lower method)

Summary