CycleGAN --> now with Cloud superpower

---

Forked to enable the training of CycleGANs models on Google Cloud ML.

Improvement:

• improved input pipelines now using only native tensorflow ops
• improved test code
• removed unused stuff
• everything is faster
• add some input parameters, removed a lot of unused ones

---

Tensorflow implementation for learning an image-to-image translation without input-output pairs. (Not completed!) The method is proposed by Jun-Yan Zhu in Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networkssee. For example:

Prerequisites

• tensorflow r1.0 or higher version
• numpy 1.11.0
• scipy 0.17.0
• pillow 3.3.0

Getting Started

Installation

• Install tensorflow from https://github.com/tensorflow/tensorflow
• Clone this repo:

git clone https://github.com/xhujoy/CycleGAN-tensorflow
cd CycleGAN-tensorflow

Train

• Download a dataset (e.g. zebra and horse images from ImageNet):

bash ./download_dataset.sh horse2zebra

• Train a model:

python main.py --phase=train --dataset_dir=$PWD/datasets/horse2zebra

• Use tensorboard to visualize the training details:

tensorboard --logdir=./checkpoints

Test

Assuming a pretrained model is saved in ./checkpoints

python main.py --phase test  --which_direction AtoB --dataset_dir=$PWD/datasets/horse2zebra --checkpoint_dir=$PWD/checkpoints --test_dir=$PWD/test_result

• dataset_dir is a directory containing all the image to be transformed by Cycle GAN.

• test_result is the folder were will be saved the transformed image and an html for side by side comparison between original and transformed images. The directory structure of dataset_dir will be manteined.

Training and Test Details

To see additional parameters

python main.py -h 

Datasets

Download the datasets using the following script:

bash ./download_dataset.sh dataset_name

• facades: 400 images from the CMP Facades dataset.
• cityscapes: 2975 images from the Cityscapes training set.
• maps: 1096 training images scraped from Google Maps.
• horse2zebra: 939 horse images and 1177 zebra images downloaded from ImageNet using keywords wild horse and zebra
• apple2orange: 996 apple images and 1020 orange images downloaded from ImageNet using keywords apple and navel orange.
• summer2winter_yosemite: 1273 summer Yosemite images and 854 winter Yosemite images were downloaded using Flickr API. See more details in our paper.
• monet2photo, vangogh2photo, ukiyoe2photo, cezanne2photo: The art images were downloaded from Wikiart. The real photos are downloaded from Flickr using combination of tags landscape and landscapephotography. The training set size of each class is Monet:1074, Cezanne:584, Van Gogh:401, Ukiyo-e:1433, Photographs:6853.
• iphone2dslr_flower: both classe of images were downlaoded from Flickr. The training set size of each class is iPhone:1813, DSLR:3316. See more details in our paper.

Reference

• The torch implementation of CycleGAN, https://github.com/junyanz/CycleGAN
• The tensorflow implementation of pix2pix, https://github.com/yenchenlin/pix2pix-tensorflow