search

fxia22 / stn.pytorch

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

fix a critical mistake in my_lib_cuda_kernel.cu #4

Closed lim0606 closed 7 years ago

lim0606 commented 7 years ago

I made a critical mistake in my_lib_cuda_kernel.cu, and fixed it.

I apologize for my mistake!

Best,

Jae Hyun Lim

fxia22 commented 7 years ago

merged, thanks

fxia22 commented 7 years ago

@jj0mst This needs to be done, the reason for the performance drop is likely due to how you handle out of FOV? Also, you can play with the learning rates of STN, they are known to be difficult to train.

jj0mst commented 7 years ago

I don't use the entire network, but only the bilinear sampler.

I have sampling grids at different scales as output of a part of my network (which is not related to stn), so i just use the bilinear sampler for sampling from an input image.

Since there's no learnable parameter in the portion i use, i ran some simple tests to check that it does what i need and everything seems correct.

I just wanted to be sure that this change was correct.

Thank you both for your work.

fxia22 commented 7 years ago

@jj0mst your use case is similar to mine, I did similar tests and things seem to work fine. Thanks for your feedback.