Does sparse_quantize support tensor input?

[WilliamKRobert]

Everything works great when the function takes as input a numpy array. But it does not work for torch tensor. The error comes from this line in MinkowskiEngine/utils/quantization.py:

discrete_coordinates = np.floor(discrete_coordinates)

To resolve the issue, I tried to replace the line with

if isinstance(discrete_coordinates, np.ndarray):
        discrete_coordinates = np.floor(discrete_coordinates)
else:
        discrete_coordinates = torch.floor(discrete_coordinates)

And then another error was triggered in the following line of code:

elif "cuda" in device:
        manager = MEB.CoordinateMapManagerGPU_c10()

Looks like the input tensor must reside on 'cuda:0' which is the not the case for my application.

---

To Reproduce

mapping = ME.utils.sparse_quantize(coords, return_index=True, device=['cuda:1'])

---

Desktop (please complete the following information):

• OS: Ubuntu 18.04
• Python version: 3.8.10
• Pytorch version: 1.8.1
• CUDA version: 10.2
• Minkowski Engine version: 0.5.4