From Image-to-Image Translation with Conditional Adversarial Networks
Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, Alexei A. Efros. In CVPR 2017
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

List of changes

• Adding of L2 binary loss by thresholding the image outputted during the training of the model.

pip install -r requirements.txt

python train.py --dataroot ./datasets/pptbf/ --name training --model pix2pix --direction BtoA --lambda_bin 50

/datasets/pptbf/ must contains a directory named train of concatenated 512*256 images, with the C-PPTBF image on the left side and the structure map on the right side. The threshold value of the structure map must be indicated in the file names, with the first 2 decimal digits just before the image format. For example pptbf25.png for an image with a threshold value of 0.25.
--no_thresh must be specified otherwise, but the L2 binary loss will not be applied.
--lambda_bin is the weight for the L2 binary loss, not indicating it will train the model with only the L1 loss.

Other options such as the use of GPUs can be find in /options/.

python test.py --dataroot ./datasets/pptbf/ --name training --model pix2pix --direction BtoA --num_test N

/datasets/pptbf/ must contains a directory named test of concatenated 512*256 images, with the real or the C-PPTBF image on the left side and the structure map on the right side.

--num_test is the number of images in the directory.

To test our trained model, download on [link] latest_net_D.pth and latest_net_G.pth, put them in /checkpoints/training/ and then run the above test command line. The results will be saved in /results/training/.

python filter.py --input image.png --iterations 40

image.png is the binary image on which we want to apply the filter.
--iterations is the number of iterations of the filter to apply on the image, 40 by default.
It will output the filtered image in current directory.