A GAN loosely based on INFOGAN trained to conditionally generate MNIST images. This basicly just means that the discrimnator also attempts to predict class but is only trained to do this for the classes of the training data and the generator aims to fool the discriminator but have the class correctly preidcted.
WARNING: Only Linux is officaly supported
Feedback and pull requests are very welcome
Inspired by https://github.com/eriklindernoren/Keras-GAN/blob/master/infogan/infogan.py
Copyright © 2019 Rory Sharp All rights reserved.
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
You should have received a copy of the GNU General Public License along with this program. If you have not received this, see http://www.gnu.org/licenses/gpl-3.0.html.
For a summary of the licence go to https://tldrlegal.com/license/gnu-general-public-license-v3-(gpl-3)
- Curl
apt-get install curl
- Python 3
- Keras
pip3 install keras - Numpy
pip3 install numpy - TensorFlow
pip3 install tensorflow - h5py
pip3 install h5py - matplotlib
pip3 install matplotlib - tqdm (for training only) (likely to already be installed)
pip3 install tqdm - keract (for intermediate layer visualisation and linear interpolation only)
pip3 install keract - imageio (for linear interpolation only)
pip3 install imageio - libhdf5 (only needed on some systems)
sudo apt-get install libhdf5-serial-dev
curl https://raw.githubusercontent.com/qwertpi/mnist-infogan/master/install.bash | bash
0. Download this repo
1. Create a folder called images that will be used to peridocly save generated imaes to during training
2. Change the if True statements for checkpoint loading to if False when running the code for the first time or it will load the pretrained generator in this repo and spoil your fun
3. Run train.py for as long as you want (I recomend around 40,000 epochs), I trained using google colabatory
4. Run generate.py
5. Marvel at what modern neural networks can do
6. If you run viz.py you will be able to see the outputs of the convolutional layers inbetween the input and output
7. Decide whether you want to be shown the images as a matplotlib plot or as a gif and change the value of save within interpolate.py accordingly
8. Run interpolate.py
9. Provide the starting and finishing classes that the model will smoothly move between (I found that it goes through imbetween classes)
10. If you opted to save find the interpolated.gif file