search

junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.23k stars 1.93k forks source link

Training Resolution #62

Closed CJHFUTURE closed 6 years ago

CJHFUTURE commented 6 years ago

Hi, could you confirm the flags and parameters to set for training at a resolution of 512x512? (Both image pairs are 512x512) And testing at the same and higher resolution.

Cheers:)

junyanz commented 6 years ago

It actually depends on your applications and GPU memory. If you have a GPU with 20GB, you can just set --loadSize 572 fineSize 512 for both training and test. If you have a GPU with 10GB, you can train a model on cropped patches. In this case, you can set --loadSize 512 fineSize 256 for training, and set --loadSize 512 fineSize 512 for test. The model consumes much less memory during the test time, and the fully convolutional network can work for images with arbitrary sizes