The weight for 4 extra_conv to get hierarchical feature in IPE

[Zhengyang1995]

Hey, according to your paper and code, there are 4 extra_conv to get hierarchical feature in IPE, I guess their main effort is to reduce dimensionality.
My question is how do you update their weights when only IPE works, (as mentioned in table 4, miou reaches 53.2% ). I could understand that in the total framework, the weight could be updated by the Lgsc loss, but in only IPE, there is only Lcls, how could you update their weights then?

[chenqi1126]

Hi @Zhengyang1995 ,

Thanks for your attention. IS-CAM with only IPE (53.2%) uses the semantic feature instead of the hierarchical feature for lacking Lgsc loss.

[Zhengyang1995]

Okay, I get it, thank you~

[Zhengyang1995]

Do you publish the weight for only IPE or the codes? How could I only run this part in your available codes? Thanks~

[chenqi1126]

You can train a vanilla Resnet50 with only cross-entropy loss, and then replace the hierarchical feature (hie_fea) with semantic feature (x4) in the following two lines: https://github.com/chenqi1126/SIPE/blob/e0c6a3bc578a6bade2cf85fe34a1793dce71170d/network/resnet50_SIPE.py#L149-L150

[Zhengyang1995]

Okay, thanks~