search

c-yn / SFNet

[ICLR 2023] Selective Frequency Network for Image Restoration
MIT License
97 stars 9 forks source link

Why set learnable parameters in function GAP #1

Closed dbzqyyy closed 11 months ago

dbzqyyy commented 1 year ago

Hi, I observed that setting the parameter fscale_d and fscale_h in the function and it has the values 0 and 1 respectively, why is this? Will this ensure that the gradient propagates properly?

c-yn commented 11 months ago

Hi, thanks for the question.

We just want the network to learn more high-frequency signals by setting 1 for fscale_h.

You could try to set the same value for both fscale_d/h.

Thanks.