Image Generation using Deep Convolutional GAN

In this project I use, a deep learning approach to generate human faces. The technique used is called Deep Convolutional Generative Adverserial Network, which uses a Convolutional neural network as a discriminator and a deconvolutional neural network is as a generator. The whole idea behind GAN is to have a zero-sum game framework by using two neural networks contesting with each other. One neural network is the generator and the main task of the generator is generating new items the other neural network is a discriminator and the task of discriminator is to distinguish between real and fake items. The items that would be generated by the generator 2 would be fake items since it is trying to mimic the real data items the main goal of the generator is to make these data items as real as possible so that it can fool the discriminator on the other hand the goal of discriminator is to distinguish these fake these and real items as best as possible so here the discriminator works as a adversary judging the real and the fake items. So at the start the generator produces some fake data items these fake data items are feed into the discriminator along with the real data items and the discriminator is made to learn which are real and fake. The results of the discriminator are than further used to improve both the generator and itself. Backpropagation is used on both the networks so that so that the generator produces better images, while the discriminator becomes more skilled at flagging data items. (Goodfellow I. , 2016) This process continues indefinitely and in the end we get two high trained models one that is highly capable of generating new data items and other that is high capable at distinguishing these data items.

Getting Started

1. Clone the entire repositry.
2. Download the aligned and cropped dataset from here.
3. Open the Data preparation Jupyter notebook and run each cell to compile the entire dataset into a single numpy array.
4. Run the DCGAN.py to start the training process.
5. The results will be stored in the Output folder and the models after every 10,000 epoch will be stored in the models folder.
6. After training, run Plotting Graphs.py file to plot the graphs of generator and discriminator losses and the accuracy of the discriminator.

The size of pre-processing the images can be changed in the Data Preparation.ipynb. Note by changing the size of the pre-processed images architecute of GAN has to be tweaked accordingly. The save interval and the batch size can also be changed in the DCGAN.py file.

The already pre-processed dataset can be found here and the pre-trained models can be found here

Software Requirements

1. OS Windows / Mac OS / Linux

2. Software

Python 3.5 or higher.

Tensorflow --suggested to use Tensorflow-gpu

Keras

OpenCV

Numpy

Matplotlib

Pillow

h5py

Scipy

Requirements required for running

Architecture

License

This project is licensed under the MIT License - see the LICENSE.md file for details