search

junyanz / CycleGAN

Software that can generate photos from paintings, turn horses into zebras, perform style transfer, and more.
Other
12.34k stars 1.94k forks source link

Validation #94

Closed doantientai closed 6 years ago

doantientai commented 6 years ago

Why don't we have val data and validation loss along with train and test data such like in pix2pix and other networks?

junyanz commented 6 years ago

Pix2pix can be evaluated using ground truth input-output pairs. But in the unpaired case, we don't even have the ground truth output for a given input. One can calculate the GAN loss and cycle loss on a validation set. But the loss itself may not tell you much.

doantientai commented 6 years ago

Oh I got it. That's right. Thanks!

gimletkim commented 1 year ago

@junyanz If the validation loss itseelf may not tell me much, dose it have no meaning? If so, when I split datasets to use cycleGAN, is it unncessary to have validation sets?