`thrust::system::system_error` when `sparse_quantize` called with label

[nbgbn]

Hi,

When I run codes below

import MinkowskiEngine as ME, MinkowskiEngine.utils as ME_utils
import numpy as np

coords = np.random.rand(10000, 3) * 100
feats = np.random.rand(10000, 3)
labels = np.ones((10000, 1))
print(ME_utils.sparse_quantize(coords, feats=feats, labels=labels))

I got error:

terminate called after throwing an instance of 'thrust::system::system_error'
  what():  parallel_for failed: invalid argument

But code above works fine after I reduced the size of coords (100, such as) or just removed the argument 'labels'.

Is there any suggestion?

THX!

[chrischoy]

Thanks for submitting the issue. I can confirm that this is reproducible in my setup as well.

For now, you can use like the examples/indoor.py#L107.

inds = ME_utils.sparse_quantize(coords)
return coords[inds], feats[inds], labels[inds]

This is how I trained all my recent models that is top-ranked on the ScanNet 3D semantic segmentation benchmark. However, I'll fix this issue in few days.

[chrischoy]

It turns out that thrust 1.9 is not compatible with the old code for some reasons that are not specified in the API. The simplest sort_by_key without a tuple iterator does not work giving the error

CUDA error 11 [/usr/local/cuda/include/thrust/system/cuda/detail/cub/device/dispatch/dispatch_radix_sort.cuh, 976]: invalid argument
CUDA error 11 [/usr/local/cuda/include/thrust/system/cuda/detail/cub/device/dispatch/dispatch_radix_sort.cuh, 1183]: invalid argument
CUDA error 11 [/usr/local/cuda/include/thrust/system/cuda/detail/cub/device/dispatch/dispatch_radix_sort.cuh, 1283]: invalid argument

This would take more longer term investigation and would require working with thrust. I'll make an alternative CPU based method as the default voxelization method for now.