junyanz/CycleGAN ? reverse-engineered prompt

Build me a working Torch version of CycleGAN that can learn image to image translation from unpaired photos.

I want to be able to train on two folders of images, like horses and zebras, then run a test script that turns new images into the other style and saves the results in a simple webpage or image folder. Please include the basic pieces for training, testing, loading datasets, saving checkpoints, and using pretrained models for common examples like horse to zebra, Monet to photo, and map to satellite.

It should be easy to run from the command line with a few settings like data folder, experiment name, train or test mode, and whether to use GPU or CPU. If there are example scripts or sample datasets, wire those up too so I can try it quickly. Use the current docs online if you need to check the expected setup or model behavior.