CycleGAN-TensorFlow

• Tensorflow implementation of CycleGAN in the paper "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks" [Paper]
• Borrowed code and ideas from junyanz's pytorch-CycleGAN-and-pix2pix: https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix.

Install Required Packages

First ensure that you have installed the following required packages:

• TensorFlow1.4.0 (instructions). Maybe other version is ok.
• Opencv (instructions). Here is opencv-2.4.9.

See requirements.txt for details.

More Qualitative results

The datasets used in experiment is apple2orange, vangogh2photo and summer_winter, you can utilize other datasets for style transform using this codes. If so, you should change the training datasets. Here is the training results.

apple <-> orange

vangogh <-> photo

summer <-> winter

Datasets

• In the implementation of the CycleGAN, there are some datasets. Forexample, ae_photos, apple2orange, vangogh2photo and so on.
• Run the download_dataset.sh to download these datasets, you can put all the datasets in the datasets folder.
• Run the scripts/build_tfrecords.py to generate training data, data format is tfrecords. In the process of generating TFRecords file, it will generate two TFRecord files, one is the trainA, the other is trainB. When training CycleGAN, all the training data are converted to BGR images and resized to 256x256.

Training a Model

• Run the following script to train the model, in the process of training, will save the training images every 500 steps. See the model/cyclegan.py for details.

sh train.sh

You can change the arguments in train.sh depend on your machine config.

• Run the following script to test the trained model. The test.sh will transform the datasets.

sh test.sh

The script will load the trained CycleGAN model to generate the transformed images. You could change the arguments in test.sh depend on your machine config.

Downloading data/trained model

• Run the download_dataset.sh to download these datasets, see the download_dataset.sh for details.
• Pretrained model: [download].

Discussion

• When use vangogh2photo dataset as the training data, you will see the result from vangogh to photo is not good. I guess it relate to the vangogh dataset is too smalle, the vangogh dataset has only 400 images. So, in the process of training vangogh2photo, the network with vangogh2photo is not trained well.
• The testing phase is also not good when using apple2orange testing data, maybe is the problem of save images.