dumyy/handpose

Resnet-50 basemodel structure

Closed this issue · 2 comments

Hi,
Thanks again for sharing your code. While going through your codebase, i can see that you created some resnet blocks like this:
resnet-q1
Where base depth is 512 for ''block4'.
But you implemented it in a slightly different way:
resnet-q2
Last 2 blocks have 64*4 each if i am correct. Is there a reason behind that?

dumyy commented

in fact, the input is only a depth map, so it's easier compared to other tasks. Here we simply the resnet, and make its end channel only 256, and of course it's enough.

Got it! Thanks a lot :-)
Now i realize why you never needed to use a deeper resnet