`labels_to_contours` selects CuPy/Dask backend and crashes inside scikit-image

[Erwin-Jo]

Hi Ultrack team,

first I want to say my thanks for the awesome software.

On Ultrack version 0.6.3, calling ultrack.utils.edge.labels_to_contours(...) with NumPy label stacks on a machine/ JupyterHub cluter where CuPy is available (or when inputs originate from Dask with GPU meta data) causes Ultrack to choose the CuPy array backend internally. The function subsequently calls scikit-image/SciPy CPU-only morphology routines, which attempt to coerce a cupy.ndarray to NumPy and fail with:

TypeError: Implicit conversion to a NumPy array is not allowed. Please use .get() to construct a NumPy array explicitly.

It occurs despite calling compute() on a dask.array or creating a fresh sample Numpy.ndarray. This error is quite persistent and the only round away solution I found is to put the following before labels_to_contours:

import ultrack.utils.edge as edge
edge.xp = np

I wanted to bring this bug to your attention to know if there is a better way to circumvent this problem and understand what is the cause behind it.

[JoOkuma]

Hi @Erwin-Jo, can you try installing cucim?
I think I need to improve how I guess whether the operation should be numpy or cupy.

[Erwin-Jo]

Hi,
I just installed cucim-cu12-25.8.0 and the error seems to be resolved as the code now executes without the my previous Numpy hack. But if the code now runs on the CPU or GPU, can I as of right now not determine.

Nonetheless, thank you for your answer and this quick fix.