This project is by David Bachmann (bacdavid). It is not published or affiliated.

Combining a Variational Auto-Encoder with a modified Adversarial Loss to make GAN training more stable. The main idea is that the generated image is closely related to a sample, it is marely a slight perturbation of it. Exploiting this fact the feature maps of the critic can be matched along the regular adversarial loss, allowing for far more stable training.

RobuGAN

Vanilla-VAE

• Typical VAE network for the generator: Encoder - Sampler - Decoder
• Mainly convolutional layers for the encoder and de-convolutional layers for the decoder with kernel size 5x5 and strides of 2x2
• Batch Norm followed by ReLU after the (de-)convolution
• 64 - 128 - 256, 256 - 128 - 64 - 3 (RGB) feature maps for encoder and decoder, respectively

• Similar to the encoder; use LeakyReLU instead of regular
• 64 - 128 - 256 feature maps

The generator loss is a modified adversarial loss function. It contains 3 terms:

• Reconstruction loss: This guarantees feature consistency between the generator and the samples
• Generator loss: This is the adversarial loss term, minimizing the distance between the critic's real and fake classification
• Regularizer: Regularize the Sampler to match a normal distribution

Train the critic to identify fake images ('0') and real images ('1')

Simply open the file

train.py

and perform the required adjustments.