Questions about detailed attention map during test phase

[zzw1123]

Hi, thanks for your contribution! In Section 4.2 in the paper, you mentioned that

we randomly select a channel of attention maps in each iteration in training stage, and conduct average pooling over attention maps for testing

If we do average pooling during testing, then there is no difference between structure attention map and detail attention map. So can you explain the detailed operations in testing process?

[JingyunLiang]

According to my understanding, just as the authors said in Section 1, efficient inference can be obtained as we can use master-net to perform recognition in the testing stage, I think only master-net branch is used for inference. conduct average pooling over attention maps may means that all generated attention maps are averaged to form an average map for master-net, which is exactly the same as the training process if we ignore part-net.

[vb123er951]

Hi, I also notice that can use master-net to do testing, but I wonder how to do this in code? I need to use model.tasn() or just use model.part_master_net()?