auxiliary loss

[lds92]

Hello! I am trying to create my own training script for PSPNet, and I do not understand how to apply the auxiliary loss. I have to do another pyramid pooling for res4b22, then softmax and sum them???? If someone can help me, I will appreciate it :)

[krishvishal]

After reading the paper, I found the loss function has two softmax loss functions addedup together with the weight of 0.4. Take pspnet101_VOC2012_473.prototxt as an example, I guess the loss function should be added to conv6(not conv6_interp ) and conv4_23 * 0.4, L= conv6 + 0.4*conv4_23, then softmax with the label.

[zhipeng-fan]

@krishnavishalv @hszhao What's the label here? it will be the segmentation map or a class in pre-train stage?

[dheerajpai]

Can you also explain why is it necessary?

[littletomatodonkey]

@krishnavishalv @hszhao What's the label here? it will be the segmentation map or a class in pre-train stage?

aux layer和最后的output layer的feature size大小都是相同的，因此也是经过softmax之后与最后的label进行比较，具体可以参见gluoncv中PSPNet的实现方法。