VAE-GAN-Pytorch

Generation of 128x128 bird images using VAE-GAN with additional feature matching loss.

Model Description

Resnet18 based Encoder. Generator and Discriminator architectures are similar to that of DCGAN. Discriminator is trained with traditional loss function and Generator is trained with Heuristic non saturating loss. Encoder is trained with KL-Divergence loss to ensure latent 'z' generated is close to standard normal distribution. In addition, the combination of Encoder and Generator is trained with reconstruction loss and Discriminator's feature matching loss.

Prerequisites

• Python 2.7
• Pytorch 0.4.0
• Torchvision 0.2.1

Data

Download CUB-2011 dataset from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html and copy the images folder to data folder in the repository

Running on GPUs

Enabled running on multiple GPUs. Edit cuda device numbers in main.py

Training

To train the model from the beginning, run:

python main.py

To resume training from a saved model, run:

python main.py --resume_training=True

Samples generated from noise at each epoch can be viewed at data/results folder

Testing

To generate images from saved model, run:

python main.py --to_train=False

Samples generated from noise with weight of feature matching loss = 0.01: