Open JarveeLee opened 5 years ago
my main modification is the bp for d-dimension like here,now my offset grad error is around 1e-3... grad img , grad mask , grad bias has been lower than 1e-5....
But offset grad error from your 2D implementation can reach 1e-5...I don't know what is wrong here...
@JarveeLee have you come up with an adequate 3D solution that could be shared?
But offset grad error from your 2D implementation can reach 1e-5...I don't know what is wrong here...
Hi @CharlesShang @Jarveelee, I was trying to do the same thing but stuck in function dmcn_im2col_bilinear in deform_im2col_cuda.cuh, any hint about it?
any update? Any one knows does the im2col implementation suitable for 3d conv? (will it be too big for memory)
Could you please write a 3D version of deformable convolution ... I am dealing with medical data which require 3D conv but my modified 3D deformable conv from your code can not pass gradient check if threshold is small.