BUG: number.__int__ and int(number) are not the same #9972

Activity

1. eric-wieser

   added

   00 - Bug

   component: numpy._core

   on Nov 6, 2017
2. mattip commented on Nov 15, 2017

   

   mattip

   on Nov 15, 2017

   Member

   More actions

   Not only because of reassignment after PyType_Ready, also because of CPython quirks. For example, this code has different paths due to CPython not inheriting nb_int but inheriting __int__ via mro lookup.

   It successfully calls int(a) via int_new which eventually calls int_from_string in cpython2.7 abstract.c since PyStringArrType_Type.tp_from_number.nb_int is NULL.

   It fails with a.__int__() since type(a).mro() hits numpy.generic which has gentype_int as its nb_int

   >>> a = np.array(['100'])[0]
   >>> int(a)
   100
   >>> a.__int__()
   Traceback (most recent call last):
     File "<stdin>", line 1, in <module>
   TypeError: don't know how to convert scalar number to int
   
3. mattip commented on Nov 15, 2017

   

   mattip

   on Nov 15, 2017

   Member

   More actions

   Just to clarify - PyStringArrType_Type.tp_as_number is not overwritten after PyType_Ready in addscalar_math, only the numeric types are overwritten, so this is purely a CPython method lookup quirk
4. eric-wieser commented on Nov 17, 2017

   

   eric-wieser

   on Nov 17, 2017

   MemberAuthor

   More actions

   I'm wondering if the culprit is ndarray.__int__, which for some reason doesn't fall back on PyNumber_Int, instead trying to call the members directly.
5. mattip commented on Nov 17, 2017

   

   mattip

   on Nov 17, 2017

   Member

   More actions

   That would add another layer of abstraction, slowing down the conversion for the 99% of the cases when a nb_int exists
6. eric-wieser commented on Nov 17, 2017

   

   eric-wieser

   on Nov 17, 2017

   Last edited by eric-wieser

   MemberAuthor

   More actions

   Right, but it also means that our implementation right now is incorrect. int(x) is also supposed to try __trunc__ if __int__ doesn't exist. #10042 fixes that.
7. eric-wieser commented on Nov 17, 2017

   

   eric-wieser

   on Nov 17, 2017

   Last edited by eric-wieser

   MemberAuthor

   More actions

   Another example:

   >>> a = np.array(None)
   >>> a[()] = bytearray('10')
   >>> int(a)
   TypeError: cannot convert to an int; scalar object is not a number
   >>> int(bytearray('10'))
   10
8. mattip commented on Feb 20, 2019

   

   mattip

   on Feb 20, 2019

   Member

   More actions

   #10042 fixed these. These now work

   >>> a = np.array([b'100'])[0]
   >>> int(a)
   100
   >>> a.__int__()
   100
   

   and

   >>> a = np.array(None)
   >>> a[()] = bytearray(b'10')
   >>> int(a)
   10
   
9. bsipocz commented on Oct 15, 2019

   

   bsipocz

   on Oct 15, 2019

   Member

   More actions

   It looks like everything works here now. Can this be closed @mattip?
10. mattip commented on Dec 8, 2019

    

    mattip

    on Dec 8, 2019

    Member

    More actions

    Closing. Please reopen if more discrepancies appear
11. mattip

    closed this as completedon Dec 8, 2019