BUG: conversion of float32 to string shows too much precision

[jorisvandenbossche]

On master (but also on 1.1):

In [4]: pd.Series([0.1], dtype="float64").astype("string")
Out[4]: 
0    0.1
dtype: string

In [5]: pd.Series([0.1], dtype="float32").astype("string")
Out[5]: 
0    0.10000000149011612
dtype: string

When converting to the object-dtype string, it works as expected:

In [6]: pd.Series([0.1], dtype="float64").astype("str")
Out[6]: 
0    0.1
dtype: object

In [7]: pd.Series([0.1], dtype="float32").astype("str")
Out[7]: 
0    0.1
dtype: object

cc @topper-123

[jorisvandenbossche]

Correction: the above behaviour is on released 1.1.0, and on 1.1.2 and master we have this bug for astype(str) as well:

In [1]: pd.Series([0.1], dtype="float64").astype("string")  
Out[1]: 
0    0.1
dtype: string

In [2]: pd.Series([0.1], dtype="float32").astype("string")
Out[2]: 
0    0.10000000149011612
dtype: string

In [3]: pd.Series([0.1], dtype="float64").astype("str")
Out[3]: 
0    0.1
dtype: object

In [4]: pd.Series([0.1], dtype="float32").astype("str") 
Out[4]: 
0    0.10000000149011612
dtype: object

[dsaxton]

ref #35519 which I think caused the regression for str

[topper-123]

This is a bit odd, seems like a straight str conversion (as is done in #35519) should work right:

>>> str(np.float64(0.1))
0.1  # ok
str(np.float32(0.1))
0.1  # ok

But I see the function added in #35519 gives a different string:

>>> f64 = np.array([0.1], dtype="float64")
>>> pd._libs.lib.ensure_string_array(f64)
array(['0.1'], dtype=object)  # ok
>>> f32 = np.array([0.1], dtype="float32")
>>> pd._libs.lib.ensure_string_array(f32)
array(['0.10000000149011612'], dtype=object)  # not ok

so something in ensure_string_array. I'll look into it.

EDIT: I can see @dsaxton has already figured it out and pushed a PR, Thanks!

[jorisvandenbossche]

Ah, so the difference is:

In [38]: str(np.float32(0.1))                                                                                                                                                                                      
Out[38]: '0.1'

In [39]: str(float(np.float32(0.1)))                                                                                                                                                                               
Out[39]: '0.10000000149011612'

(because converting to object dtype converts to builtin float type)

So I suppose because float32 has less precision, the stdlib float shows those decimals. While the numpy.float32 behaves more as we want.