MXNET-Scala CycleGAN

MXNet-scala module implementation of cycleGAN[1].

Based on: https://github.com/junyanz/CycleGAN

So far, I have tried all the tasks and done a lot of experiments,

but just success in two tasks: "apple2orange" and "photo2ukiuoe".

I think I followed the torch implementation completely and couldn't locate the problem :( .

Results:

Requirements

sbt 0.13 http://www.scala-sbt.org/
Mxnet https://github.com/dmlc/mxnet

steps

1, compile MXNet with CUDA, then compile the scala-pkg，doc： https://github.com/dmlc/mxnet/tree/master/scala-package

2, under the Mxnet-Scala/CycleGAN folder

 mkdir lib;

3, copy the compiled mxnet-full_2.11-INTERNAL.jar into lib folder;

4, run sbt then compile the project

Datasets

you can download the datasets with the datasets/download_dataset.sh. you can refer to https://github.com/junyanz/CycleGAN for how to use this script.

Testing

you can try the pretrained model of "apple2orange" and "photo2ukiuoe" with the scripts/test_cycle_gan.sh script. The pretrain model zip file is under models folder.

you need to provide the input image file.

If you use the apple2orange model, A means apple, B means orange.

If you use the photo2ukiuoe model which has only one direction, BtoA, B means photo.

# pretrain models are under the $ROOT/datas/pretrain_models directory
PREAREIN_G_MODEL=$ROOT/models/pretrain_models/

INPUT_IMAGE=

# -1 for cpu
GPU=0

# "AtoB" or "BtoA"
DIRECTION="BtoA"

Training new models

use the train_cycle_gan.sh script under scripts folder. If you keep all the default settings, you just need to provide the domainA images path and domainB images path:

DOMAIN_A_PATH=
DOMAIN_B_PATH=

License

MIT

Reference

[1] Zhu, Jun Yan, et al. "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks." 2017.