Code for Variational GAN with critic

This project was developed as a part of my Master's research. The main idea is to use Сritic network to improve GAN's ability to reconstruct original pictures. Instead of simply using L1 loss to evaluate reconstructed objects, Critic approximates joint distribution on objects pairs.

Model architecture

Final Model is composed of:

• Encoder (E)
• Generator (G)
• Discriminator (D)
• Latent space discriminator (C)
• Critic-network (S)

Usage

Data preparation

Download celebA dataset

Training a model

Put path to dataset directory in train.py in CriticGan instance and run:

python train.py

Results

Random sampling

Randomly generated faces	Interpolation in latent space

Reconstruction

Faces from original dataset	Reconstructions

Links

• Paper abstract on variational GAN with critic, in English

• Paper on variational GAN with critic, in Russian

• Model's weights pretrained on celebA dataset