GAN art generator

About and Technologies

This is an implmentation of a Generative adversarial network (GAN). The aim of this project was to train the GAN such that it produces images that are indistinguishable from existing art.
The GAN neural network was built using Python and utilised the PyTorch library for constructing the discriminator and generator neural networks. Tkinter was used for the GUI aspects of the project.

Environment Setup

Before you are able to run this project you need to first setup the proper environment on your computer. As this project was only run on Windows with Pytorch in Cuda mode this guide is only going to describe how to setup for that environment.

Installing Python 3 and Tkinter

The first thing you need to do is to install python3 on your Machine. (If you already have python installed you can skip this step.):

Go to Python’s download page:
Download the appropriate installer for your machine
Make sure to check the box that asks you to add python to your path when prompted
You should now have python installed and added to your path
Tkinter comes prepackeged with Python3 so you now should also have Tkinter installed

Installing PyTorch in Cuda Mode

After Installing Python3 and adding it to your path
Open Command Prompt
run the "pip3 install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html" without the qoutes to install Pytorch
You have now successfully installed PyTorch for Cuda 11
Note that if your GPU device does not support Cuda 11 , you can try other installations of PyTorch here https://pytorch.org/ however this may require minor changes to the code in order to get it to work

Code Documentation

The codebase is small and consists of only 5 classes. Here is a quick run down of these classes and how to use them

Viewer.pyw

This is file responsible for providing the GUI and running the GAN in inference mode. The viewer class allows the user to ask the GAN to generate new Fake images. The user can change which dataset they want the GAN to emulate and which epoch of training the GAN stopped at. The Viewer is a simple class extends a TK window. On initialization it creates 3 radio buttons that allows the user to select which GAN model they want to be loaded. The GAN models differ on which training set they were trained on. It also providers a slider that allows the user to specify which epoch of training to load. It also initializes 2 buttons "Generate New random vector" and "Generate New". "Generate New" instructs the GAN to generate new images using the internal initialization vector and then displays these images on the screen ."Generate New random vector" generates a new random vector to be used by the GAN when genrating images. To see the changes the new vector to produces you need to always click on the "Generate New" button after clicking on the "Generate New random vector" button. You can run the GUI by double clicking on the Viewer.pyw file after installing python. Note that the GUI will not generate new images if trained models do not exist in the specified directory structure (described later) and it also wont generate images if the model with the specified Model Number doesn't exists. Currently the trainer saves every 10 iterations so if you try to load iteration that is not a multiple of 10 the gui will not be able to load it. reading https://docs.python.org/3/library/tk.html May be useful to get a basic understanding on How Tkinter works

Discriminator.py and Genrator.py

These file contain the implementation of the Generator and Discriminator. They are built using PyTorchs sequential class. The generator takes in a random 1d array vector and through 5 convoltional layers transposes that random vector onto the sample space of the training images of size 64x64. On the other hand, the discrimnator also uses 5 convolutional layers but instead it transposes the input image into a binary decision of whether or not an image was produced by the generator.
The following resources may be useful when using these classes:

Configuaration.py and Trainer.py

the Trainer.py file is responsible for training the GAN network on a specified dataset and storing and updated model after a specified number of epochs so that it can be used by GUI. The configuration.py file contains the configuration settings of the Trainer.py class. Here you can change which data set the Trainer.py file will use, how often and where the models will be saved (Other settings should not be changed by the user) and which epoch the trainer should resume training from. When you run the Trainer.py file it will use the settings from configuration.py and start training your model.

Folder Structure

As the models and Training datasets are pretty large they were not uploaded to Github. As such you will need to download the training sets and train your models for a bit before you're are able to use GUI class. Here are the links to the datasets that were used when training this model:

Note that your datasets need to be inserted in a folder structure that looks like this relative to trainer.py "data/datasetname/1/". You would put your training images inside the 1 folder. Similarly when saving your models they will be save in a folder structure that looks like this "models/datasetname/"

hatthawaybriggs / gan-art-generator Goto Github PK