Repository navigation
Why does recarray assignment fail: "cannot call setfield on an object array"? (Trac #2006) #2599
Description
Original ticket http://projects.scipy.org/numpy/ticket/2006 on 2011-12-23 by trac user kevin000, assigned to unknown.
Using a recarray with member assignment fails if any column uses the object data type. I don't understand why this fails, especially since the dictionary style assignment works.
Brief example showing what recarray assignments succeed or fail
import numpy as np
recarray with integer datatypes
dt = np.dtype([('foo', 'i8'), ('bar', 'i8')])
r = np.zeros((1,3), dtype=dt).view(np.recarray)
r['foo'] = np.array([1, 2, 3]) # OK
r.foo = np.array([1, 2, 3]) # OK
recarray with an object datatype
dt = np.dtype([('foo', 'i8'), ('bar', 'O8')])
r = np.zeros((1,3), dtype=dt).view(np.recarray)
r['foo'] = np.array([1, 2, 3]) # OK
r.foo = np.array([1, 2, 3]) # RuntimeError
For the relevant C code search for "cannot call setfield on an object array" in this file:
https://github.com/numpy/numpy/blob/master/numpy/core/src/multiarray/methods.c
Thanks!
Kevin
Activity
cmarquardt commentedon Dec 15, 2012
The same problem occurs in 1.6.2 and 1.7.0b2.
The (big) change that introduced this behavior is 2635398 and apparently has introduced the error condition on purpose; there's even a test checking that using the setfield() method on an (ordinary) object array throws a RuntimeError.
As modifying object sub-arrays using the dictionary approach works fine, the easiest solution seems to be to replace the setfield() method in the setattr() method of recarray with ndarray's setitem() method; I just attached a patch against the 1.7.0b2 distribution of numpy which runns successfully through the same tests as the original distribution does.
I know way too little about numpy internals to judge if this patch is consistent with the numpy conventions etc.; but keeping this bug in recarrays somehow make the whole recarray class superfluous if not fixed IMHO.
Thanks for considering the patch (which is below),
Christian.
*** numpy-1.7.0b2.orig/numpy/core//records.py 2012-09-04 23:31:31.000000000 +0200 --- numpy-1.7.0b2/numpy/core//records.py 2012-12-02 23:20:00.757234770 +0100 *************** *** 451,457 **** res = fielddict[attr][:2] except (TypeError, KeyError): raise AttributeError("record array has no attribute %s" % attr) ! return self.setfield(val, *res) def __getitem__(self, indx): obj = ndarray.__getitem__(self, indx) --- 451,457 ---- res = fielddict[attr][:2] except (TypeError, KeyError): raise AttributeError("record array has no attribute %s" % attr) ! return ndarray.__setitem__(self, attr, val) def __getitem__(self, indx): obj = ndarray.__getitem__(self, indx)charris commentedon Feb 21, 2014
There are a bunch of tickets that look related to this. The patch has not been applied as of 1.9-devel.
charris commentedon Feb 21, 2014
Easy fix if the patch is good and we want that behavior.
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
ENH: structured datatype safety checks
Removed the numpy-path.h script.