nan_to_num silently fails (does nothing) on object arrays

[gdementen]

Annoyingly, object arrays occur frequently because there is some text in the array when loading it, and in that context something like nan_to_num is often needed.

>>> arr = ndtest((2, 3)).astype(object)
>>> arr['a0', 'b1'] = la.nan
>>> arr
a\b  b0   b1  b2
 a0   0  nan   2
 a1   3    4   5
>>> nan_to_num(arr)
a\b  b0   b1  b2
 a0   0  nan   2
 a1   3    4   5

The workaround is to convert to float:

>>> nan_to_num(arr.astype(float))
a\b   b0   b1   b2
 a0  0.0  0.0  2.0
 a1  3.0  4.0  5.0

I think I remember there was another issue regarding nan and object arrays, but I cannot find it anymore. I guess I fixed that problem without creating an issue about it, because general_isnan and obj_isnan in larray code were created exactly to alleviate this issue.

[gdementen]

bumped priority because the failure is silent

[gdementen]

The work-around cited above (convert to float) does not work in the presence of actual strings in the data, but the following code should work:

def nan_to_num2(arr, copy=True, nan=0, posinf=None, neginf=None):
    if arr.dtype.kind == 'O':
        import sys
        if posinf is None:
            posinf = sys.float_info.max
        if neginf is None:
            neginf = -sys.float_info.max
        res = arr.copy() if copy else arr
        res[arr != arr] = nan
        res[arr == inf] = posinf
        res[arr == -inf] = neginf
        return res
    else:
        return nan_to_num(arr, copy=copy, nan=nan, posinf=posinf, neginf=neginf)

[gdementen]

I think I remember there was another issue regarding nan and object arrays, but I cannot find it anymore. I guess I fixed that problem without creating an issue about it, because general_isnan and obj_isnan in larray code were created exactly to alleviate this issue.

I now think that it was that currently (as of larray 0.34.6), la.isnan does not use obj_isnan and so does not support object arrays.

Note also that:

• obj_isnan displays nan warnings (which is kinda silly) and it returns a numpy array (so should stay an internal function).
• general_isnan is only a private function in Array.eq

>>> arr = ndtest((2, 3)).astype(object)
>>> arr['a0', 'b1'] = la.nan
>>> la.isnan(arr)
TypeError: ufunc 'isnan' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''
>>> la.core.array.obj_isnan(arr)
<...>\Lib\site-packages\numpy\lib\_function_base_impl.py:2537: RuntimeWarning: invalid value (NaN) encountered during operation (this is typically caused by a 0 / 0)
  return self._vectorize_call(func=func, args=vargs)

I don't remember what is the status of this in the dev branch though.