Derived from the MNIST GAN example https://github.com/apache/incubator-mxnet/blob/master/contrib/clojure-package/examples/gan/src/gan/gan_mnist.clj
This is an example of how to train GANs with your own images. In this example it is flans. To use, you must collect your images and store that in a .rec file. Use this guide to get started https://mxnet.incubator.apache.org/faq/recordio.html
This particular example is for 128x128 images. It takes a bit of memory to run. If you run into troubles try starting smaller like 28x28. There are code comments in the source to help guide you along.
Do lein run and the images generated will be in the results directory. The gout* images are the ones generated, the diff* images are the visualization of the input gradient different fed to the generator
lein run :gpu will run on gpu
You can also work with the repl and explore the pretrained model with
explore-pretrained
Or explore your trained model at a specifice epoch with explore
If you are running on AWS you will need to setup X11 for graphics
sudo apt install xauth x11-apps
then relogin in ssh -X -i creds ubuntu@yourinstance