search

junyanz / pytorch-CycleGAN-and-pix2pix

Image-to-Image Translation in PyTorch
Other
22.69k stars 6.27k forks source link

Which generator networks to use when? #549

Open elpequeno opened 5 years ago

elpequeno commented 5 years ago

Hey everybody. I was trying CycleGan on my medical image dataset and I was wondering if there are any recommendations on which generator to use under which conditions? How do you decide whether you use resnet_6blocks, resnet_9blocks, unet_128 oder unet_256.

Some hint, recommended paper or blog post would be appreciated.

Thank you very much.

taesungp commented 5 years ago

Resnet_6blocks and Resnet_9blocks produce similar results, but resnet_9blocks has larger receptive field, which means it should learn relationship between objects further apart. I would use resnet_6blocks for idea iteration and resnet_9blocks for the best result if your GPU allows it.

Similarly for unet_128 and unet_256. They were designed to work with images of sizes 128x128 and 256x256, respectively, although the fully convolutional architecture makes them available for many other difference input sizes.