CycleGAN Production Version

this repo based on the original implementation of CycleGAN: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.git, in this version I reconstruct some code and made a generate API to simply generate image from your own single image and your trained model.

CycleGAN - Generate Image Like Magic

I have trained apple2orange and horse2zebra for now, here is the real result of convert apple -> orange:

I only trained about 50 epochs, but the result is fair enough for now. Laterly I will finish horse2zebra model, and update some more results.

Here is the horse2zebra result:

Requirements

• Python3+
• PyTorch
• visdom
• PIL

Usage

• For Train

About how to train, simply run this:

python3 train.py --dataroot ./datasets/apple2orange --name apple2orange --model cycle_gan

One things have to mention that, --name indicates the model save dir, and --model is using cycle_gan or pixel2pixel , I only tried cycle_gan.

• For Generate

Train is very simple, but the original repo have not implement predict API, so I managed to write by myself. Here is the way to use:

python3 generate.py --image_path ./apple_test.jpg --name apple2orange --model cycle_gan --gpu_ids -1

As you can see, you only need to specific image path where stores your image to generate, and --name is the same as previous trained, as well as model type. --gpu_ids indicates we are inference using CPU.

OK, that's all.

Research and Discuss

I really love to connect to people, so if you have any question about this repo, you can find me on wechat jintianiloveu, I have some groups which discuss about GANs I will invite you in if you like.

Copyright

(c) 2017 Jin Fagang under LICENSE Apache 2.0