How to implement spatial attention module?

[imyangs]

Dear author

Do you have any plans to relase the code for spatial attention? If not, please give some ideas about how to implement.
Many thanks.

[Christian-lyc]

Hi, thank you so much for your interest.
The spatial part is same as the channel part. you basically need to reshape the H and W to 1 D. BN just measure the importance of each pixel.

[imyangs]

Hi, so many thanks for your quick reply.

As I cannot map the default dimension of scale factor of BN to spatial dims ( C != H X W ? ), so my real concern is how to implement the "pixel normalization" module. My idea is set "self.channels = HxW", and than the forward() func should implement as below:

please help me figure out the right soultion. Many thanks.

[Christian-lyc]

Hi, It seems this is the same as the paper said.

[imyangs]

Thank you so much. I will test it later.

[Ronky123]

Thank you so much. I will test it later.

hi,do you test the spatial attention?does it work?

[imyangs]

Thank you so much. I will test it later.

hi,do you test the spatial attention?does it work?

Sorry, I have not. Maybe you can just paste my above code and run the train script for a result.

[haikunzhang95]

Thank you so much. I will test it later.

Hi, did you test your spatial attention codes? Does it work well?
Thank you!

[23jisuper]

非常感谢。我稍后会测试它。

嗨，你测试空间注意力吗？它有效吗？

对不起，我没有。也许你可以粘贴我上面的代码并运行训练脚本以获得结果。

您好 关于空间注意力您实现了吗？ 我的想法是能不能先采用池化将H*W进行归一化 在计算BN