test file for this bug in attachment.
old = True fails with numpy trunk, old=False works
Benny
2010/7/16 Benny Malengier <[email protected]>
>
>
> 2010/7/16 Benny Malengier <[email protected]>
>
>>
>>
>> 2010/7/16 Benny Malengier <[email protected]>
>>
>>> I hit a much more problematic thing with this numpy.
>>>
>>> The code
>>>
>>> from fipy.meshes.numMesh.grid2D import Grid2D
>>> mesh = Grid2D(dx=0.1, dy=0.1, nx=10, ny=10)
>>> ap = CellVariable(mesh=mesh, value=0.01)
>>> coeff = mesh._getFaceAreas() * mesh._getCellDistances() / \
>>> ap.getArithmeticFaceValue()
>>> print 'coval', type(coeff), type(coeff.getValue())
>>>
>>> Now fails with:
>>> AttributeError: 'MaskedArray' object has no attribute 'getValue'
>>>
>>> So coeff is a MaskedArray, while on older numpy it is the expected binOp
>>>
>>> I'm investigating what is causing it.
>>>
>>
>> Note that above fragment is from the stokescavity example. Changing the
>> code to:
>>
>> coeff = 1/ ap.getArithmeticFaceValue() * mesh._getFaceAreas() *
>> mesh._getCellDistances()
>>
>> works around it. It might be a bug in numpy trunk or it is due to new
>> features.
>>
>
> I'll stop investigating. My guess is that in numpy/ma/core.py the function:
>
> get_data
> does not fail, so the wrong operation is done. mesh._getFaceAreas() is a
> masked array after all. This can be seen by looking at the code:
>
> try:
> data = a._data
> except AttributeError:
> data = np.array(a, copy=False, subok=subok)
>
> where here a is the ap.getArithmeticFaceValue() , which will work, as the
> variable allows downcasting to an np.array.
> So the fix would be to not allow casting to array automatically for a
> variable. Not idea if that is a problem for fipy.
>
> As some proof for this statement,
> see get_data in 1.2.1
>
> http://projects.scipy.org/numpy/browser/tags/1.2.1/numpy/ma/core.py?rev=5972
> as opposed to trunk version:
> http://projects.scipy.org/numpy/browser/trunk/numpy/ma/core.py
>
> The AttributeError catch is new.
>
> Benny
>
>
>> Benny
>>
>>
>>
>>> Benny
>>>
>>> 2010/7/15 Jonathan Guyer <[email protected]>
>>>
>>>
>>>>
>>>> On Jul 15, 2010, at 3:02 AM, Benny Malengier wrote:
>>>>
>>>> > It seems ndarray has a dot function now in numpy trunk, updated my
>>>> numpy and fipy crashes like ...
>>>>
>>>> Thanks for the alert
>>>>
>>>>
>>>>
>>>>
>>>
>>
>
from fipy import *
from fipy.meshes.numMesh.grid2D import Grid2D
import numpy as np
old = True
mesh = Grid2D(dx=0.1, dy=0.1, nx=10, ny=10)
ap = CellVariable(mesh=mesh, value=0.01)
if old:
#fails on new numpy
coeff = mesh._getFaceAreas() * mesh._getCellDistances() / ap.getArithmeticFaceValue()
else:
#works
coeff = 1/ap.getArithmeticFaceValue() * mesh._getFaceAreas() * mesh._getCellDistances()
print np.array(ap.getArithmeticFaceValue(), copy=False, subok=True)
print type(mesh._getFaceAreas()), type(mesh._getCellDistances())
print 'coval', type(coeff), type(coeff.getValue())
result = coeff * numerix.identity(mesh.getDim())
