Deprecate/remove silent truncation on assignment of floats into an integer array

[markrichardson]

Ubuntu 14.04
Python 2.7.11

In [1]: import numpy as np

In [2]: np.__version__
Out[2]: '1.10.4'

In [3]: v = 4 - 0.1

In [4]: np.array([v])
Out[4]: array([ 3.9])

In [5]: a = np.array([1]); a[0] = v; a
Out[5]: array([3])

[seberg]

The behaviour is of course correct, a is autotyped to be an integer array, so the float is cast to an integer on assignment. This could have a warning in principle for the unsafe cast, but it is expected and has to work that way since numpy arrays are (normally) typed.

[charris]

Note that numpy doesn't round in these cases, it truncates. @seberg I thought this was deprecated, but it doesn't seem so.

[seberg]

There is only a deprecation for the in place ufuncs, maybe we should think about making this deprecate too?

[charris]

@seberg Sounds like a good idea, we should raise it on the list.

[njsmith]

https://mail.scipy.org/pipermail/numpy-discussion/2016-June/075679.html

[njsmith]

Sounds like everyone is in favor, so I renamed the bug report. I guess concretely the desired change should follow the same path we used for in-place operations? Which IIRC was that we switched it to same_kind casting, with a deprecation period.

[seberg]

Yeah, would have to be a pretty long deprecation just like there. I am a little worried that the workaround with copyto or casting the right hand side is sometimes a bit annoying. But I suppose that it should be rare that a bit of explicit casting/more convolved stuff does not hurt and the error can easily tell you the solution.

[stsievert]

I think this is an important usability issue. One of my labmates @sumeetsk wasted a month of research effort on this bug (and just before his defense too). It boiled down to

>>> a = np.array([0, 1, 2])
>>> a[0] = 1.1
>>> print(a)
... array([1, 1, 2])

which is unexpected. This issue is also mentioned in a textbook:

if you attempt to insert a floating-point value to an integer array, the value will be silently truncated. Don't be caught unaware by this behavior!

— @jakevdp's Data Science Handbook, Chapter 2.2 "The Basics of NumPy Arrays"

I'll try to submit a PR for this issue. This is still around in NumPy 1.14.3, and the recent release notes don't seem to mention it.

[sumeetsk]

What makes this especially unexpected / counter-intuitive is that python does not behave this way.

[mattip]

Overflow and truncation seem to be mentioned often as issues. Maybe we could have a verbose mode that would report overflow and truncation as warnings, at a cost in performance

[stevengj]

(Note that this issue and #8733 are basically duplicates.)