search

Maclory / SPSR

Pytorch implementation of Structure-Preserving Super Resolution with Gradient Guidance (CVPR 2020 & TPAMI 2021)
447 stars 83 forks source link

blank image #37

Open sangjun04 opened 3 years ago

sangjun04 commented 3 years ago

I am training with my custom data. I have successfully trained on custom data and the validation result looks good. However, when I run the exact same model on a test set, all images are black. I have checked pixel values, and only one channel among three has values between 0 and 20 even after the normalization.

Any clue what might have caused this issue?

Maclory commented 3 years ago

Hi, I think this may be caused by incorrect preprocessing of testing images. You can check whether the input pattern of the testing set is similar to the pattern of other images.

sangjun04 commented 3 years ago

To make sure the model is working, I used the same image from the validation set for testing.

Maclory commented 3 years ago

Does the model work well on the validation set?

sangjun04 commented 3 years ago

The validation result created during training looks valid.

stdinR commented 1 year ago

Hello, I want to ask about custom training. Since you said, you used your own dataset. Is HR size in your dataset different from default (128) given by the author? I found when I changed the HR_size = 128 to my own HR size (which is 480) in train_spsr.json, the training didn't work. However, when I used the default HR_size (128) without any change, it could train. This makes me wonder would this impact the training result?

leotuTu commented 7 months ago

Hello, I want to ask about custom training. Since you said, you used your own dataset. Is HR size in your dataset different from default (128) given by the author? I found when I changed the HR_size = 128 to my own HR size (which is 480) in train_spsr.json, the training didn't work. However, when I used the default HR_size (128) without any change, it could train. This makes me wonder would this impact the training result?

I'm having the same problem. Have you solved it now?