BUG: np.full_like casts float to int for non-finite values

[onnoeberhard]

Describe the issue:

Please see the code snippet below. I create an integer array a, then create a new one (b) which is supposed to have the same shape as a, but filled with the float value np.nan. Here, I don't explicitly set the dtype for b. NumPy converts these float values to integer.

This is unexpected behaviour, and I assume is not intended as such. I would have expected either

1. b is automatically made a float-array (the dtype information from a is not used), or
2. A TypeError is thrown, notifying me that I am trying to fill an int-array with float values.

Reproduce the code example:

>>> import numpy as np
>>> a = np.array([1, 2, 3])
>>> b = np.full_like(a, np.nan)
>>> b
array([-9223372036854775808, -9223372036854775808, -9223372036854775808])

Error message:

No response

NumPy/Python version information:

1.22.3 3.8.12 (default, Oct 22 2021, 18:39:35)
[Clang 13.0.0 (clang-1300.0.29.3)]

[onnoeberhard]

I have looked again into the docs and it seems like this behaviour is indeed intended.. But would it ever really make sense in this context? I mean casting non-finite float values to int without even throwing a warning?

[seberg]

See also gh-21437 which would ensure a warning for this example. (Not for a value like 2.5 though, we would have to add a cast-safety check explicitly.)

[seberg]

This now gives a warning on the main branch (it does not fail, unless np.errstate is used). I am not exactly sure whether we should modify the behavior to be stricter about doing unsafe casts.

I suspect this is a good step in the right direction and the next step would require a larger thought including other functions. So I am tempted to close the issue, but holding off for now.

[seberg]

Actually, let me close it. Please don't hesitate to reopen an issue though. Xref gh-21437 which added the warnings.