/vaegan-celeba

Primary LanguagePython

vaegan-celeba

This is a keras implementation of the paper "Autoencoding beyond pixels using a learned similarity metric" by Larsen, A., et al.

Link to implementation documentation: https://docs.google.com/presentation/d/1_i0gV0_QhC84PsnBj4jPq4H6qS7n2vRv_pGFt8ZQQmE/edit?usp=sharing

VAEGAN Model

Dependencies

  • keras
  • tensorflow
  • OpenCV
  • numpy

Training:

  • vaegan-celeba.py was used to train the VAEGAN model.
  • vaegan-loader.py is used to produce images from a trained vaegan model.
  • train-celeba.py was used to train the VAE model.
  • gan-celeba.py was used to train the GAN model.
  • add "-w" to load weights (e.g. python train-celeba.py -w vae-model.h5)
  • models_celebA.py contains models used for this project.
  • celebA_loader.py contains the data loader used to load the CelebA dataset.
  • CelebA dataset was used for training the model

Results

VAE (Images from noise input):

VAE Autoencoder (output-left; input-right):

GAN (Images from noise input):

VAEGAN (Images from noise input):

VAEGAN Autoencoder (output-left; input-right):

References