Do we need to reweight the output of max pooling layer by the probability of presence when testing? In theory, the max-pooling will be impacted by the dropout ratio.
The background is that I'm reading the paper(https://arxiv.org/pdf/2111.14973.pdf). In chapter 3.2, it uses max-pooling for the variable size objects as the input.
For normal max pooling. In the test phase, we should reweight the output by (1-p). But for the struct, I have no idea.reweight
