Train on the CUFED5 about dcsr HOT 4 CLOSED

Hi,
The initial lr is 1e-4, decay at 7-13-20-28-36 epoches. We usually train 20 epoches for a CUFED5 model.

from dcsr.

The initial lr is 1e-4, decay at 7-13-20-28-36 epoches. We usually train 20 epoches for a CUFED5 model.

Is your description really complete and correct? You mean you usually train 20 epoches for a CUFED5 model，but your learning rate strategy continues to 36 epoches. By the way ，according to the paper, training should have two stages. Thank you!

from dcsr.

@jiaxinxie97 @Tengfei-Wang

from dcsr.

Hi,
We found the model converged within 15~16 epochs for CUFED5 (with --num_patch=300). So we used the max epoch of 20 in our experiments, though we initially set it to 36.
For the concern about two-stage training, we only employ SRA on CameraFusion dataset, as we wanna obtain 8K images but we do not have ground truth. For CUFED5, the training is one-stage as we can obtain the GT HR images in the datasets.

from dcsr.