Problems when add dsntnn into HourglassNet. about dsntnn HOT 6 CLOSED

Thanks to your quick replay!
The mistake you point out is indeed the cause that 'LM', that means loss for js regularisation, increases during the training process

tmpOutMap = self.tmpOut[i](ll) heatmaps = dsntnn.flat_softmax(tmpOutMap) outMap.append(tmpOutMap) # <-- Should beoutMap.append(heatmaps)?
But the problem still unresolved is that 'LR', that means Euclidean loss for the numerical coordinates, keeps coverging very slowly. And of course, the performance is still bad.

I think I need go deeper for some other bugs.
Thanks for your answer again!

from dsntnn.

Hi @anibali Thanks for the response, I found the bug, it's indeed on my side. Thanks for this awesome work.

from dsntnn.

One potential problem I can see is this:

      tmpOutMap = self.tmpOut[i](ll)
      heatmaps = dsntnn.flat_softmax(tmpOutMap)
      outMap.append(tmpOutMap)  # <-- Should be `outMap.append(heatmaps)`?

You seem to be adding the non-softmaxed version of the heatmap to outMap, which means that you end up applying the JS regularisation to the non-normalised heatmaps.

from dsntnn.

@Hellomodo @anibali Hi, have you found the bug? I used this layer in another task, and the loss for numerical coordinates doesn't decrease at all.

from dsntnn.

@sunshineatnoon I currently don't have any reason to believe that there's a bug in dsntnn. It's more likely that the bug is in your usage or your other code. If I'm wrong and you do manage to narrow down the problem to dsntnn with a minimal example, please feel free to open a new issue.

from dsntnn.

problem about LR， have you sloved it?

from dsntnn.