Qibin Hou comments

Results 40 comments of


                                            Qibin Hou

According to my experiences, CA has a slightly better results on ResNet than SE.

Sorry, could not really get your point.

I think the answer should be both branches do so.

Have not tried that yet. CA is more friendly to downstream tasks with dense prediction, such as semantic segmentation.

Maybe this is caused by the adaptive_avg_pool. If the spatial size of the input feature map is available, you can use avg_pool instead.

Have you pretrained the model on ImageNet?

Exactly. Otherwise, it does not work. Recently, I will release the pretrained models.

Check it out now.

Given the code you provided, it is hard to figure out the problem.

CoordAtt(planes * self.expansion,reduction) shoud be CoordAtt(planes * self.expansion, planes * self.expansion, reduction)