UCL Deep Learning Group Project

Working Repository for Deep Learning Group Project

Team:

Project

Comparing the performance of plain GANs, GANs with Reconstruction Loss and DiscoGANs in style transfer in human faces
Style transfer to be investigated along lines of gender, age and emotion
Training data to be a subset from the CelebA dataset http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

In order to train the model, we use data from CelebA
We provide a simple tool (readText.py) to extract datasets of specific feature domains. This can be found in the 'CelebA Data Sort' folder within this repository.
The instructions.txt file, in the same folder, provides detail instructions on how to use the tool

A sample implementation for each of the three GAN models is included in main.py
To select a model, simply uncomment the relevant portions of code, or else use the GANFactory class to create your own
A number of parameters are available when instantiating each of the GANs, including source and target directories, batch, epoch and kernel size and image dimensions. These are designed to enable quick customisation of model implementations
With a GAN object, you may use the .train() method to begin learning the model. With saved weights (i.e. after training), you can test the trained implementation using .test(). This essentially runs through one more pass of the GAN with your selected input image, saving the output image in the desired directory.
Further changes to the architecture are possible by adjusting the code in GANFactory.py

We assume the user has installed anaconda: https://www.anaconda.com/download/#macos

Most packages required are included as part of anaconda. Additionally we require TensorFlow

You should be able to simply run:

pip install TensorFlow

(you may need sudo depending on your local permissions)