OT-CycleGAN (Optimal Transport-driven CycleGAN)

Pytorch implementation of Optimal Transport driven CycleGAN for Unsupervised Learning in Inverse Problems [Paper]

Train on custom dataset

python train.py --data_path directory_for_custom_dataset --domain_A subdirctory_for_domain_A \\
                --domain_B subdirectory_for_domain_B --extension png

Details about implementation

1. data_path should contain domain_A and domain_B as its subdirectories.
2. extension should be able to open with numpy.load or SimpleITK.ReadImage. If your file extension is not supported by these functions, modify [utils.py].
3. noramlize only supports one of minmax, tanh, CT, None. If you want to normalize with different method, modify [utils.py]. CT stands for computed tomography in medical imaging.
4. This repository is only implemented with Figure 7 and Table 1(d) of the original Paper. Especially, I only implemented discriminator loss for Equation 3.24.
5. With my personal experience, I changed Algorithm 3.1 to train the generator prior to the discriminator. Training generator prior to discriminator makes model more stable.
6. If you train model with CT option for normalize configuration, you may want to convert numpy ndarray into DICOM format. It might be helpful referencing this code
7. Alert! This is not official implementation.

Training Result

I only experimented CT image for model training. However, for legal issue - such as privacy of the patient, I cannot upload converted CT image.

In further commits, I would be happy to share other images that have no legal issue.

Changes

2021.01.07 : Added skip connection in ConvBlock