When should npyinitialize be called?

[JobJob]

Am I, mr user, supposed to call npyinitialize()? I'm guessing not, so this might be a bug.

Seems asscalar, and in turn convert(Int, numpy_array_element::PyObject), depends on npy_number being initialized

using PyCall
@pyimport numpy as np
pyoarr = pycall(np.array, PyObject, 1:3)
el = get(pyoarr, PyObject, 0)

@show pytypeof(el) pytypeof(PyCall.asscalar(el)) PyCall.npy_number
convert(Int, el)

outputs

pytypeof(el) = PyObject <class 'numpy.int64'>
pytypeof(PyCall.asscalar(el)) = PyObject <class 'numpy.int64'>  (should be int)
PyCall.npy_number = PyObject NULL  (shouldn't be null for asscalar to work)
ERROR: PyError (ccall(@pysym(PyInt_AsSsize_t), Int, (PyPtr,), asscalar(po))) <class 'TypeError'>
TypeError('an integer is required',)

[stevengj]

No, you are never supposed to call it. You generally shouldn't be calling undocumented internal functions like PyCall.asscalar either.

You can just use

julia> pycall(np.asscalar, PyObject, el)
PyObject 1

for example.

[JobJob]

but is

using PyCall
@pyimport numpy as np
pyoarr = pycall(np.array, PyObject, 1:3)
el = get(pyoarr, PyObject, 0)
convert(Int, el)

supposed to work? That calls asscalar, and PyCall.npy_number is PyNULL then

[stevengj]

Oh, right, I see the problem.

It actually works fine for me in Python 2.7:

julia> convert(Int, el)
1

but my recollection from 9304bbc was that the asscalar is needed in Python 3. The current code works if you get numpy arrays via Julia's automatic conversions, but not if you import numpy manually yourself.

[JobJob]

Ok cheers, yep I'm on Python 3. Feel free to close this.

[stevengj]

No, Python 3 is supported too, so I need to fix this.

[JobJob]

Ok, just fyi the actual use case that landed me here was fairly contrived. I was just testing converting an array to a tuple, it wasn't part of any actual work/project.

using PyCall
@pyimport numpy as np
arrpytpl = pycall(np.array, Tuple{Vararg{Int}}, 1:10)

[stevengj]

Should be fixed now.

[JobJob]

Thanks for almost fixing this - just tried it and unsafe_asscalar(o) isn't calling o.item(), it's just returning the function o.item.

(Edit: Changing it to unsafe_asscalar(o) = pycall(o["item"], PyObject) makes tests pass for me on Julia 0.6.2, Python 3)

While I'm here, just regarding my PRs - I'm happy enough with them all so you can take another look. I've just rebased them all off master. This issue here is breaking tests locally for them, so we'll need to check again once that's fixed (and I merge/rebase the fix into each of them).

[stevengj]

Whoops! Will make a PR shortly.