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