I tried to re-implement your solution on PyTorch with using F.grid_sample operation, which can do slicing and then use affine transformation on coefs from slices to build and image. But somehow network doesn't train as it should.
All nn stuff checked many times, and it looks similar to the original code. But grid_sample under the question.
Maybe you or you colleagues who worked with pytorch can say how bilateral slice and grid_sample can be different? Cause im totally not an expert in Cuda code.
I tried to re-implement your solution on PyTorch with using F.grid_sample operation, which can do slicing and then use affine transformation on coefs from slices to build and image. But somehow network doesn't train as it should. All nn stuff checked many times, and it looks similar to the original code. But grid_sample under the question.
Maybe you or you colleagues who worked with pytorch can say how bilateral slice and grid_sample can be different? Cause im totally not an expert in Cuda code.
Thanks.