HRNet
about production of lower resolution representation
in paper‘s figure, a lower resolution representation in new stage seems to be produced from all streams in previous stage(a bit like fuse layer), but in code:
x_list.append(self.transition2[i](y_list[-1]))
it seems to be produced from the last stream in previous stage(y_list[-1]).
So, which is right? I am so confused about this problem.
in paper‘s figure：
a lower resolution representation is connected to all representations in previous stage by wire, but in code it seems should be connected to tha last representation in previous stage.
likes:
I am so confused about this! Can someone help me! Please!
Have you figured this out ? I was getting familiar with this repository not that long ago and for me it seemed like the code does not completely reflect paper's illustration, and in fact it does the very thing you illustrated in the last image, at least that is what I thought it does.
@mishazakharov I haven't figured this out yet. I found all official repositories about HRNet do the thing i illustrated in the last image. Perhaps the diagrams in the paper are meant to be easier for the reader to understand. As a rookie, I'm still confused.
