variable radius with depth feature request in torch_cluster.radius_cuda

[imransai]

Hi Matthias,
I was working on an algorithm which would collect all pixels within in a radius that varies with depth/range. Your radius () seems to accept a fixed radius to collect neighbors within a certain 3D point. But I am not sure how to implement a variable radius with cuda. Is the variable radius idea easier to integrate? Or can you show me how to modify your code to make it variable radius implementation in cuda?

Let me know if you need more info. Really waiting for your reply. :)

[rusty1s]

Instead of passing a fixed scalar to radius, you would need to pass in a vector, holding the radius for each point. The radius check then happens on a per point basis.

[imransai]

I see that makes sense. I guess that is the only .cu file that needs to be changed? The question is how to integrate it as a python wrapper for use with python? I am newbie in wrapping up a cuda code in python module. Is there any practice or suggestions you would like me to follow?

[rusty1s]

The simplest way would be to fork torch-sparse and modify the code accordingly, i.e., csrc/radius.cpp, csrc/cuda/radius_cuda.h and csrc/cuda/radius_cuda.cu now take in a torch::Tensor for r instead of a single scalar. Then, in csrc/cuda/radius_cuda.cu, you need to pass in r.data_ptr<scalar_t>() to radius_kernel. In the radius_kernel, you simply access radius[n_y] instead of radius here.

The Python bindings are handled here, in which r should be replaced by a torch.Tensor as well.

Let me know if this information helps :)