Pytorch-cyclegan-and-pix2pix: Parameter selection for gray image?

[John1231983]

I am using cycle GAN with UNet generator to translation from T1 to T2 in the medical image. Both are gray images with size of 1x256x256 (CxHxW).

I am wondering about the parameters opt.input_nc, opt.output_nc, opt.ngf in the

https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/690470c80bb7021d96c9d3f82e416e633dbb2da3/models/cycle_gan_model.py#L73

What should I choose? In my opinion, it should be 1 for opt.input_nc=1, opt.output_nc=1, and opt.ngf=8. Thanks

[junyanz]

You can set --input_nc 1 and --output_nc and leave opt.ngf as default. You may want to try Resnet architecture as well.

[John1231983]

Thanks. It worked now. Just for your suggestion, unet has more parameters and larger checkpoints MB. What is the benefit of using resnet comparison with unet? When should I use resnet?

[taesungp]

We do not have a definite answer for that. Resnet tends to make less changes and work better when two domains are similar.