FLOAT_CMP isn't powerful enough to handle whitespace that is added in fixed-width printing of floats

[ayshih]

See title. Here's a minimal example with ndarray, where I vary the least significant digit of the element the first row:

>>> import numpy as np
>>> np.array([[0.12345678], [0.11111111]])
array([[0.12345678],
       [0.11111111]])

>>> np.array([[0.12345671], [0.11111111]])  # doctest: +FLOAT_CMP
array([[0.12345678],
       [0.11111111]])

>>> np.array([[0.12345670], [0.11111111]])  # doctest: +FLOAT_CMP
array([[0.12345678],
       [0.11111111]])

The second test passes with FLOAT_CMP because the change is well under the relative tolerance. The third test fails with:

Expected:
    array([[0.12345678],
           [0.11111111]])
Got:
    array([[0.1234567 ],
           [0.11111111]])

This failure isn't because of exceeding the tolerance, but rather because of the space between "7" and "]". (I'm hitting this issue with a SkyCoord transformation, where the 32-bit calculation ends in a zero and the 64-bit calculation does not.)

The "obvious" workaround is to turn off FLOAT_CMP and rely on ELLIPSIS to handle the slight discrepancy. But, that could mean adding a lot of ellipsis markers. Is there a better solution?

[ayshih]

I could change the NumPy printing. First setup with:

>>> import numpy
>>> old_floatmode = np.get_printoptions()['floatmode']
>>> np.set_printoptions(floatmode='fixed')

and then teardown with:

>>> np.set_printoptions(floatmode=old_floatmode)

I'll have to mull over whether that's a better workaround.

[bsipocz]

ideally NORMALIZE_WHITESPACE should take care of this, I think. So likely this may be a bug?

[ayshih]

I guess it's technically out of the scope of NORMALIZE_WHITESPACE because that treats all non-zero whitespace the same (and different from zero whitespace), but what I want is for zero whitespace to be treated the same as non-zero whitespace when a comma is involved.

Anyway, I decided that being able to do proper float comparisons is more desirable than using ellipsis markers, so I added (hidden) testsetup/testcleanup blocks as in my earlier comment.

[bsipocz]

I guess it's technically out of the scope of NORMALIZE_WHITESPACE because that treats all non-zero whitespace the same (and different from zero whitespace), but what I want is for zero whitespace to be treated the same as non-zero whitespace when a comma is involved.

yes, thus I think ideally NORMALIZE_WHITESPACE should deal with this, or should at least work with other directives better.

[pllim]

I just hit this very same problem over at astropy/astropy#10511 . Glad I am not going crazy. I also resorted to manually setting numpy print options, which isn't pretty in the user example.