BUG: AttributeError: 'bool' object has no attribute 'to_numpy' in "mask_missing" method of core/missing.py

[rainnnnny]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import pandas
import numpy
from qpython.qcollection import qlist

res = pandas.Series(data=qlist([True, True, False, False], qtype=1, adjust_dtype=False)).replace(False, numpy.NaN)

Issue Description

new_mask = arr == x

this expected to get an array-like object with bool elements, like [False, False, True, ...] , but when arr being an special array-like object, like the qlist object in example , the expression returns a single False, which is a bool object, thus we'll get the exception shown above

Expected Behavior

returns an array-like object instead of a bool object

Installed Versions

1.4.1

[simonjayhawkins]

Thanks @rainnnnny for the report.

labelling as constructor issue as I believe that the pd.Series constructor should be converting/materializing array-likes passed as data into standard numpy arrays for the in-memory storage and not storing the Qlist object directly.

>>> res._mgr.array
QList([ True,  True, False, False])

as a workaround

pandas.Series(
    data=np.asarray(qlist([True, True, False, False], qtype=1, adjust_dtype=False))
).replace(False, numpy.NaN)

gives

0    True
1    True
2     NaN
3     NaN
dtype: object

If there is a reason that you would want to retain the QList object in memory, should investigate creating an Extension Array https://pandas.pydata.org/pandas-docs/dev/development/extending.html#extensionarray

[bcmyguest1]

import pandas as pd
# case 1
df = pd.DataFrame({"d":[pd.NA]})
df.replace('',pd.NA) # throws the same error
# case 2
df = pd.DataFrame({"d":[None]})
df.replace('',pd.NA) # works fine

FYI this is not only related to QList.

new_mask ends up being [[False]] in case 2, but False in case 1.

[simonjayhawkins]

FYI this is not only related to QList.

Thanks @bcmyguest1 for this example.

It does raise a warning FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison, so I guess this issue will resolve itself in a future pandas.

However, this example did work in pandas 1.3.5, so labelling as regression pending further investigation.

import pandas as pd

# case 1
df = pd.DataFrame({"d": [pd.NA]})
print(df.replace("", pd.NA))

      d
0  <NA>

Note: the FutureWarning was also given in pandas 1.3.5, but I think that it should probably have been suppressed.

[simonjayhawkins]

However, this example did work in pandas 1.3.5, so labelling as regression pending further investigation.

first bad commit: [b2d54d9] BUG: IntegerArray/FloatingArray ufunc with 'out' kwd (#45122)

cc @jbrockmendel

[jbrockmendel]

1. agreed with Simon that qlist shouldn't be held directly, or should be made an EA
2. if it is made an EA, arr == x should be array-like
3. we can change the check for ndarray to be specific to EA/BooleanArray
   3b) this goes in the pile of bugs created indirectly by pd.NA-workarounds

[simonjayhawkins]

moving to 1.4.4

[CloseChoice]

The reason why this goes wrong is that qlist inherits from np.ndarray therefore all our checks for this are passed and the qlist is stored directly. Maybe we could sanitize_array a bit to handle qlists correctly but maybe we can't get around an explicit cast to np.ndarray

[simonjayhawkins]

3. we can change the check for ndarray to be specific to EA/BooleanArray
3b) this goes in the pile of bugs created indirectly by pd.NA-workarounds

@jbrockmendel I've opened a PR #48313 but just checked for bool also instead (and also suppressed the numpy warnings while making changes here, see #47101 (comment))

[simonjayhawkins]

removing milestone, xref #47485 (comment)

[ancri]

As an ugly workaround until this gets patched, this seemed to work for me, at pandas/core/missing.py:112

[ChrisMLikesMath]

I've run into this on pandas 2.0/2.1 as well on line 118 of missing.py,

new_mask = arr ==x
...
new_mask = new_mask.to_numpy(dtype=bool,na_value=False)

It definitely feels like a bug here. This is what appears to be happening for me:

1. x is of type integer
2. arr is of type list of integers (in my case a QList)
3. The comparison arr == x checks if the integer is the list, or QList, which it is not.
4. We expect an array but get a single bool since arr was never an array in the first place.

Generally, np.array() overrides the == operator to do this elementwise which is what is expected here.

I think the hack above won't quite handle that enumeration although it'll at least not throw an exception.

Probably it'd be best to check if arr is in fact iterable then convert to a numpy array or run the comparison elementwise at the point where we call arr == x since the fundamental assumption here as that arr performs that check on x elementwise and returns an array.

[ChrisMLikesMath]

I think replacing arr == x with np.equal(arr, x) might do the trick with relative ease.

I think np.equal will check if either argument is an array/iterable and do that comparison elementwise.

Fundamentally, in my case, I feel like perhaps the issue might also partially fall on qpython.qcollect.QList which claims to be an np.ndarray but doesn't implement the comparison elementwise.