search

csjliang / LPTN

Official implementation of the paper 'High-Resolution Photorealistic Image Translation in Real-Time: A Laplacian Pyramid Translation Network' in CVPR 2021
Apache License 2.0
413 stars 47 forks source link

train PairedImageDataset on fivek #14

Closed semchan closed 3 years ago

semchan commented 3 years ago

I follow here to use your code to train PairedImageDataset on fivek . After 20000 iters, I check the training results and find the phenomenon of "over exposure" (as shown in the figure below, the right is the input, the middle is the prediction result, and the left is GT). All training super parameters are the default configuration in your code. Can you help to guide me on how to reduce the "over exposure"? thanks a lot.

D1645D64@F008A51E F0C5C260 jpg_recompress

csjliang commented 3 years ago

Hi I have replied your email several days ago (you send me the same figure and the same question). Please try to remove or decrease the weight of the Discriminator loss when you use paired training data. Thanks.

semchan commented 3 years ago

Hi I have replied your email several days ago (you send me the same figure and the same question). Please try to remove or decrease the weight of the Discriminator loss when you use paired training data. Thanks.

Thanks for your quick reply. Sorry, I didn't receive the email (maybe system error^_^), so I'm asking here. OK, I will follow your suggestion to try. Thanks again.

semchan commented 3 years ago

Hi, according to your guidance, I removed GAN loss and only used pixel loss (MSE constraint), which looks good. Thank you very much. a0002-dgw_005_400