search

qibao77 / CFSNet

pytorch code of "CFSNet: Toward a Controllable Feature Space for Image Restoration"（ICCV2019）
62 stars 4 forks source link

Some questions about the training #2

Open raywzy opened 4 years ago

raywzy commented 4 years ago

During training, only control parameters α =1 and α = 0 will be input into the network?

qibao77 commented 4 years ago

Yes. In training step 1, we keep α_in= 0, while in training step 2, we keep α_in = 1. I am sorry for the late reply.

LaterBetterThanNever commented 4 years ago

During testing, how can we choose the control parameters α to input into the network?

qibao77 commented 4 years ago

It depends on your training settings. For example, for the JPEG image deblocking task in our experiment, we set quality factor as 10 in the first training stage and change it to 40 in the second training stage. Therefore, during testing, we change α_in from 0 to 1 with an interval of 0.1 to get different results of any input image with quality factor between 10 and 40. And we can choose the best α_in based on visual quality or PSNR.

LaterBetterThanNever commented 4 years ago

Thanks for replying!