A PyTorch version

[bearpaw]

Thanks for sharing your code!

I wrote a pytorch version of hourglass network. Hope this could be helpful for who are not familiar with Torch. Many codes for data processing are brought from your code (src/pypose). Thank the author again!

However, the code cannot reproduce the results perfectly (83.58 [email protected] score for the simplified 4-stack hourglass). Some details might be missed, especially the post processing part (e.g., coordinates mapping, etc). Anyone interested in this project is welcomed to contribute to this project!

[wangzheallen]

Do u have comparison experiments on using different number of hourglass?

[obwang49]

I noticed in both lua and pytorch implementation, the residual modules structure is bn->relu->conv. But the resnet modules are conv->bn->relu. Is there any particular reason why you modify the sequence this way?

[obwang49]

Okay, so actually, this repo's residual module is not the same as the demo residual module. I feel the final demo version makes more sense, since it is the same with resnet implementations.

[anewell]

The change in the resnet module is based on this paper: https://arxiv.org/pdf/1603.05027.pdf which discusses the effects of switching between pre- and post-activation. Can't remember now if there was much difference when I changed the code.

[obwang49]

Thanks for the note! I was wondering why resent in pytorch model zoo didn't adopt the bn->relu->conv which everybody else seems to be using.