Why does PSNR keep dropping when training a SRResnet with DIV2K dataset?

[1conglin]

I trained the SRResnet with DIV2K training dataset( 800 images) and test set was DIV2K validation dataset(100 images) , croppped each image to 512512 with step 256, and GT size was set to 256256, train SRResnet 90 epoches with batch size of 16. But as the training progressed, PSNR kept dropping. Here are tensorboard logger figure: why did this happened?

[Lornatang]

The Loss curve shows that the model trained normally try set lr to half

[1conglin]

The Loss curve shows that the model trained normally try set lr to half

It didn't work. I changed lr from 0.001 to 0.0005, but PSNR still kept dropping as training progressed. Could you please tell me if there are any other possible reasons?

[1conglin]

The Loss curve shows that the model trained normally try set lr to half

I have found the reason, the train lr is different from test lr. train lr is generated from a complex function , but my test lr is just bicubic downsampled from test hr. I changed the train lr to bicubic downsamppling, then the problem resolved: Thank you for your attention and suggestion.

[Lornatang]

Glad to hear that you have solved the problem. Good luck in your life