Python 3.5 NumPy 1.10.0 rounding (issue?)

[jakirkham]

Not sure if this is a bug or there is simply a discrepancy between NumPy and Python. Though this is what I am seeing in some (not all) cases. FYI, similar (but not the exact same) behavior for np.float128.

>>> import numpy as np
>>> a = np.float64(0.47140452079103168)
>>> round(a, 4)
0.47139999999999999
>>> a = np.float64(0.47140452079103168)
>>> round(float(a), 4)
0.4714
>>> a = 0.47140452079103168
>>> round(a, 4)
0.4714

[njsmith]

It's a discrepancy in repr implementations:

In [1]: a = np.float64(0.47140452079103168)

In [2]: round(a, 4)
Out[2]: 0.47139999999999999

In [3]: type(_2)
Out[3]: numpy.float64

In [4]: float(_2)
Out[4]: 0.4714

np.float64.__repr__ is showing more digits than float.__repr__.

Still arguably a bug -- notice that it's only the float64 scalar type that's affected; the array printer works correctly:

In [5]: np.array(round(a, 4))
Out[5]: array(0.4714)

[jakirkham]

The float128 scalar type has this issue as well it just goes the other direction. Though I see your point about arrays. So, maybe just a floating point/complex representation issue then?

[anntzer]

See #6136 too.

[NeilGirdhar]

I think this is a duplicate of #6136. One of these should be closed. The other should be renamed: string representation of floating point numbers in arrays should not display extra digits.

Simpler test case:

In [37]: print(np.ones(1, dtype=np.float32) * (0.12345678901234567890))
[ 0.12345679104328155517578125]

In [38]: print(np.ones(1, dtype=np.float64) * (0.12345678901234567890))
[ 0.12345678901234567736988623209981597028672695159912109375]

First one should be "[0.12345678]". Second one should be "[0.123456789012345678]".

[ron819]

I think this one is no longer an issue

[rgommers]

Thanks @ron819, indeed fixed.

[charris]

Seems broken?

It's just how arrays are printed. The printing precision can be set, but by default it isn't large in order that more array elements may be displayed. The value is fine.

In [4]: (np.ones(1, dtype=np.float64) * (0.12345678901234567890))[0]
Out[4]: 0.12345678901234568

[SamWildmo]

@charris better to open a new ticket this one is very old