GPU based voxel subsampling for KPConvFPN

[SajjadPSavoji]

The current implementation performs the subsampling on cpu through the collate func registration_collate_fn_stack_mode. I am attempting to use your pre-trained backbone on a batch of point clouds which are already on GPU.

1. Does your source code support subsampling on GPU, if so where can I find it?

2. Is there other implementations that support GPU subsampling ?

I think there might be support for such operation as the sampling process is only performed when precompute_data is True.

    if precompute_data:
        input_dict = precompute_data_stack_mode(points, lengths, num_stages, voxel_size, search_radius, neighbor_limits)
        collated_dict.update(input_dict)

[HernandoR]

if you mean the radius_search function, you can check out my impl, where i replaced it with open3D's API