Numpy v1 compatibility broken on main with `DimArray`

[JamesWrigley]

Affects: PythonCall

Describe the bug ArrayValue.__array__() now supports the copy argument and defaults to copy=None, but this is not supported on numpy v1 and breaks when using a DimArray from DimensionalData.jl:

# Works
@pyeval (x=rand(10), np) => "np.asarray(x)"

# Fails
@pyeval (x=rand(X(10)), np) => "np.asarray(x)"
Python: ValueError: NoneType copy mode not allowed.
Python stacktrace:
 [1] __array__
   @ /path/to/julia-depot/packages/PythonCall/syXxf/src/JlWrap/array.jl:371
 [2] <module>
   @ In[44]:1:1

Not sure why it only breaks with DimArray but works with Array.

Your system Please provide detailed information about your system:

• Linux
• Julia 1.12.2, PythonCall main, Python 3.11.13, Numpy 1.26.4.

[cjdoris]

Thanks. I think the fix is to call np.asarray instead of np.array when copy=None because np.array(x, copy=None) is not supported in numpy v1.

[JamesWrigley]

Hmm, but that doesn't explain why it works with Array and not DimArray 🤔 I think something else is happening.

[cjdoris]

My guess is that Base.unsafe_convert(::Type{Ptr{T}}, ::DimArray) is not defined, which is required to be wrappable as a python array, since this is how we get a pointer to the underlying data.

In this case we fall back to converting the array to an Array, and perhaps the above issue only arises in this branch?

[JamesWrigley]

Is that a new requirement? Because it did work in the last stable release.

[cjdoris]

No it's not new. Can you give an MWE showing it working and not working?