search

LeiCheng-no / CDIL-CNN

31 stars 5 forks source link

Classifier part #1

Closed anonymousz97 closed 2 years ago

anonymousz97 commented 2 years ago

https://github.com/LeiCheng-no/CDIL-CNN/blob/main/figures/cdil.png image Hello , as i see the picture and the code in the demo.py i think the classifier self.classifier(torch.mean(y_conv, dim=2)) is not true. Can you explain that for me? The picture is we first pass it through linear layers and aggregate the result. Thanks

image

LeiCheng-no commented 2 years ago

Hello. As we explained in the paper, the average module and linear classification both are linear operations, so they are exchangeable. We found it will speed up the network if we apply the average first. image