Imprving the Discriminator by intorducing Vatiational Bottleneck Layer (VDB)

Paper - https://arxiv.org/pdf/1810.00821.pdf

Dependencies

• python>=2.7

• tensorflow>=1.9

Data

Have tested with mnsit data set. It will be automatically downloaded to data folder

Usage

python train-WGAN-GP-VDB.py

Important Implementations

• Generator and Discriminator Architectures are inside model.py script
• Normal wgan-gp loss and discriminator bottleneck layer loss are insode the vdb_losses.py script

Important Paramters

• I_c - This is the information contrain. This is a hyper paramter
• Bottleneck Layer Dimentions
• Alpha - This paramters is to update the adaptive lagrange paramters (Documentation can be found inside the code)

Training Results

• Trained paramters can be found inside train-WGAN-GP-VDB.py script
• There are pre-trained checkpoints in the checkpoints folder
• The images generated during the training progress are insode the sample_images_while_training folder
• Tensorbored Visualizations can be find inside summeries folder

Results:

Tensorbored Visualizations

• You can examine the learning progress by visualizing two loss functions of generator and discriminator
• Also its impotant to undertand the change in beta parameter(Eq(6) in the paper) with adaptive update method. Here we maximize the beta

Visualizations while Training:

Future Work

• The modified descriminator can be easily use with GAIL.
• Cheking the effect of varios hyper paramters and how generator behaves acording to them