hengliusky/Feature-preserving-Based-Face-Hallucination

Deep Feature-preserving Based Face Hallucination: Feature Discrimination Versus Pixels Approximation

This is a tensorflow implementation of Deep Feature-preserving Based Face Hallucination: Feature Discrimination Versus Pixels Approximation.

File Description

• config.py : include configuration for file paths and hyperparameters for networks.
• train_generator.py : code to train our generator with MSE loss.
• train_gan.py : code to train our generator and discriminator with MSE loss, perceptual loss and adversarial loss.
• train_gan_encoder.py : code to train our generator, discriminator and encoder with MSE loss, adversarial loss and encoder loss.
• train_full.py : code to train our full network with MSE loss, adversarial loss and feature adversarial loss.
• test.py : code to inference our generator with a trained model.

Setup

requires

• python 3.5
• tensorflow 1.8.0
• tensorlayer
• numpy
• scipy
• skimage

run

• Clone this repo:

  > git clone git@github.com:hengliusky/Feature-preserving-Based-Face-Hallucination.git
  > cd Feature-preserving-Based-Face-Hallucination-tensorflow
  

• Download the CelebA dataset. Move the first 18K faces to /data/train, the next 2k faces to /data/test
• Train or test the model.

  > python train_generator.py
  

Resluts

Here is the results generated by this implementation:

Our-generator

Our-GAN

Our-GAN-Encoder

Our-full

Acknowledge

Our code architecture is based on tensorlayer-srgan. Thanks for their excellent work!