series.astype(float) not returning series of floats

[rwijtvliet]

Start: series with dimensionless values

import pandas as pd
import pint
import pint_pandas

s0 = pd.Series([1.0, 2.5])
s1 = s0.astype('pint[dimensionless]')

s1
# 0    1.0
# 1    2.5
# dtype: pint[]

Action causing issue: conversion back to floats

s2 = s1.astype(float)

Issue 1: changed dtype

Previously, the resulting Series had a float data type. Now, the datatype is slightly different, which causes the comparison with s0 to fail.

s2.dtype
# PandasDtype('float64')

pd.testing.assert_series_equal(s0, s2)  #"dtypes are different"

Issue 2: conversion to string

Another issue is, that the __repr__ and __str__ methods of s2 fail:

str(s2)
#AttributeError: 'numpy.ndarray' object has no attribute '_formatter'

Excuse me as I currently cannot test this with the latest pandas version; I've used 2.0 instead. pint-pandas 0.5 , and pint 0.22.

[andrewgsavage]

Yea that doesn't seem right and could be better.

Is this something you've noticed since pint-pandas 0.5?

[andrewgsavage]

https://github.com/hgrecco/pint-pandas/blob/ef8a1209699d4533299303b800982578e8322242/pint_pandas/pint_array.py#L420 return pd.array(self.quantity, dtype, copy) should be something like return np.array(self.quantity, dtype, copy) or return pd.array(self.quantity.m, dtype, copy)

[rwijtvliet]

Yea that doesn't seem right and could be better.

Is this something you've noticed since pint-pandas 0.5?

Thanks for your fast reply. I'm not sure if it's because I upgraded pandas or pint-pandas - if that is useful information let me know and I'll investigate.

My workaround is to use s1.pint.m if the dtype is pint[]

[andrewgsavage]

there's a fix, would be great to have someone look over it before meging.

[MichaelTiemannOSC]

I think I just ran into this issue in some of my code. Taking a look now...

[MichaelTiemannOSC]

I was working around a related problem, which is that pint arrays don't support cumprod for pint[dimensionless]. Here's an accumulate function that does:

    def _accumulate(self, name: str, *, skipna: bool = True, **kwds):
        if name == "cumprod" and self.dtype!="pint[dimensionless]":
            raise TypeError("cumprod not supported for pint arrays")
        functions = {
            "cummin": np.minimum.accumulate,
        "cummax": np.maximum.accumulate,
            "cumsum": np.cumsum,
            "cumprod": np.cumprod,
    }

    if isinstance(self._data, ExtensionArray):
            try:
                result = self._data._accumulate(name, **kwds)
            except NotImplementedError:
                result = functions[name](self.numpy_data, **kwds)

        return self._from_sequence(result, self.units)

However, my workaround was to use pint.dequantify() / pint.quantify(), not astype("float") so no triggers. I need to round-trip through dequantify/quantify as I don't want to expose uncertainties to certain disaster in a "float" conversion.