Strange result (and no error/warning) for np.min() on integer array with initial=np.nan

[mgeier]

I know that np.nan is not an integer, but I was still surprised about this behavior:

Reproducing code example:

import numpy as np
print(np.min([1, 2, 3], initial=np.nan))
print(np.max([1, 2, 3], initial=np.nan))
print(np.max([1, 2, 3], initial=9.9))

Result:

-9223372036854775808
3
9

I'm not sure what I was expecting, but somehow I felt that I have to report this ...

Error message:

No error, no warning.

Numpy/Python version information:

1.15.0 3.6.7 (default, Oct 21 2018, 08:08:16)
[GCC 8.2.0]

[seberg]

I agree, initial was added rather recently. I think we should add a visible Deprecation warning maybe now, and then change the casting to "safe" for it. Just disallowing unsafe casts seems the most straight forward to me.

[eric-wieser]

@hameerabbasi, want to take a go at this at some point?

[hameerabbasi]

Is the arbitrary object as identity PR already merged? If so, I can change the casting to safe, and it shouldn’t affect anything. If not, we’ll need that to be merged first.

[eric-wieser]

Yes, that PR is merged - although I'm not sure I see the connection.

[hameerabbasi]

If we changed the casting rule to safe without that, the casting from 1 to True and so on would be forbidden, would it not?

[eric-wieser]

Oh, maybe. Ultimately I think we're going to want a different identity for each loop type, but that's probably not going to happen any time soon

Can we apply the casting rule to just the initial argument, and let the casting continue to be safe when the identity is used?

[hameerabbasi]

The error seems to be at

numpy/numpy/core/src/umath/reduction.c

Line 489 in 28e5c07

if (PyArray_FillWithScalar(result, identity) < 0) {

, but I have no idea how to change the casting rule to safe inside the PyArray_FillWithScalar function, and I'm not sure we'd want to. If we changed it before and after there would be thread safety concerns.

[hameerabbasi]

We could try casting before and raise an error if a failure occurs.

[hameerabbasi]

I found PyArray_CanCastSafely... But I don't know how to get the PyArray_Descr from the two arrays.

[seberg]

This now gives a warning at least after gh-21437. That doesn't quite fix the issue for initial=, since it is not clear that initial= (as a ufunc argument!) should allow unsafe casting.

However, we also have the more recent gh-21668 for the same issue, so going to close this one.