About the Project

This project is an implementation of a Generative Adversarial Network, trained to generate faces using the IMDB-WIKI dataset. This project was created for my term project for the University of Saskatchewan's CMPT 498: Learning and Data Analytics course in Fall 2016.

Python Scripts

• NeuralNet.py
  • cointains all tensorflow code building the model, along training, logging, sampling, and other related functions
• Trainer.py
  • loads an instance of the network, and runs training samples through it, printing results
• DataLoader.py
  • filters the IMDB-WIKI dataset to a smaller number of high quality images, and builds an index for quick access
  • contains a function that will load batches of images in a background thread, for use in training the neural network
• Sampler.pt
  • used to generate images from the trained network
• CsvStats.py
  • outputs information about the dataset csv file generated by DataLoader.py
• FaceDetector.py
  • uses OpenCV's Haar face detector to determine whether a face is in an image
  • used to evaluate results, by running the face deterctor on generated images
  • this file is run to train the network
• Visualization.py
  • used to generate a png containing a grid of faces
  • faces are input as a numpy array

Usage

Training

To train the network, modify the "datasetDir" variable in Training.py, and run the script for as may training rounds as desired.

Before training, the IMDB-WIKI dataset will be filtered and indexed, with no input required by the user

Sampling

To obtain generated face images from a trained network, run Sampler.py. A number of sample images will be generated in the working directory

Results

Included in this repository is a file called "Project Paper.pdf". This paper details the results of the project, and provides sample images generated by the network