fxia22 / stn.pytorch

pytorch version of spatial transformer networks
Other
587 stars 87 forks source link

STN added to Pytorch by Soumith Chintala #19

Open Cadene opened 7 years ago

Cadene commented 7 years ago

pytorch: Adding Spatial Transformers w/CuDNN support

fxia22 commented 7 years ago

Thanks @Cadene, I am aware of it. I am going to fix the existing issues for this repo when I have a chance for current users. But given STN is added to pytorch with higher performance, it's unlikely I will develop this repo further.

jwyang commented 7 years ago

@fxia22 ,

Hi, Fei, thanks a lot for sharing this code. I compared with Pytorch official STN and your implementation, and found the gradients are different when the number of channel of feature map > 1. I am wondering whether you also found this, or it's just me having something wrong?

thanks!

jwyang commented 7 years ago

@fxia22

One more point, when channel > 1, the gradients of the first channel are the same, the other channels are different.

jwyang commented 7 years ago

I found the problem, it is not about the cuda code, it on my own code. never mind.