NAFNet

NAFNet copied to clipboard

Hello!

Thank you for this excellent work!

In the implementation of TLC, the AvgPool2d is used to replace the AdaptiveAvgPool2d, whose output size is $1\times 1$. However, in the AvgPool2d, the output size seems to identical with the size of input.

I think the modified function actually reweighs the feature map in spatial domain instead of in channel domain, and it's not the standard channel attention.

Could you help me with this question?

Thanks.

Hello, I'm actually experiencing this problem and when inserting a channel attention using AdaptiveAvgPool2d, the attention map is reverted to the original input size and an error occurs.