About 4x start about dcsr

[Lkinyuu]

For example, I set the scale of self.aa1() and FeatureMatching to 4 ,but failed

[Tengfei-Wang]

Hi,
For 4x SR, the scale of ref image is 4x of that of LR images. So we also need to modify the feature extractor (pretrained vgg) accordingly. The key is to make sure the feature map sizes of Ref and LR are the same so that we can calculate patch-wise cosine similarity. For example, we modified 7 to 11 in this line:

DCSR/model/attention.py

Line 26 in 70ff906

for x in range(7):

And other lines specified on 2x SR should also be modified, like shape_out*4. It can be a little tedious.

DCSR/model/attention.py

Line 110 in 70ff906

warpped_features = F.fold(warpped_value, output_size=(shape_out[2]*2, shape_out[3]*2), kernel_size=(kernel,kernel), padding=0, stride=self.scale)

[Lkinyuu]

Hi, For 4x SR, the scale of ref image is 4x of that of LR images. So we also need to modify the feature extractor (pretrained vgg) accordingly. The key is to make sure the feature map sizes of Ref and LR are the same so that we can calculate patch-wise cosine similarity. For example, we modified 7 to 11 in this line:

DCSR/model/attention.py

Line 26 in 70ff906

for x in range(7):

And other lines specified on 2x SR should also be modified, like shape_out*4. It can be a little tedious.

DCSR/model/attention.py

Line 110 in 70ff906

warpped_features = F.fold(warpped_value, output_size=(shape_out[2]*2, shape_out[3]*2), kernel_size=(kernel,kernel), padding=0, stride=self.scale)

Maybe you mean " for x in range(10):"? Because when modified 7 to 11 ，it will cause channels error .Besides, if I follow the above method to modify parameters, can I train on the CUFED5x4 to get the same results as in the paper?

[Lkinyuu]

Hi, For 4x SR, the scale of ref image is 4x of that of LR images. So we also need to modify the feature extractor (pretrained vgg) accordingly. The key is to make sure the feature map sizes of Ref and LR are the same so that we can calculate patch-wise cosine similarity. For example, we modified 7 to 11 in this line:

DCSR/model/attention.py

Line 26 in 70ff906

for x in range(7):

And other lines specified on 2x SR should also be modified, like shape_out*4. It can be a little tedious.

DCSR/model/attention.py

Line 110 in 70ff906

warpped_features = F.fold(warpped_value, output_size=(shape_out[2]*2, shape_out[3]*2), kernel_size=(kernel,kernel), padding=0, stride=self.scale)

May I ask for the settings that train on the CUFED5 ? I want to reproduce the same result as in the paper.

[Lkinyuu]

@Tengfei-Wang