numpy migration (also posted to numpy-discussion)

[Duncan Smith]

Hello,
Since moving to numpy I've had a few problems with my existing
code. It basically revolves around the numpy scalar types. e.g.

------------------------------------------------

>>import Numeric as N
a = N.array([[0,1],[2,3]])
a

array([[0, 1],
[2, 3]])

>>i = a[0,0]
1/i

Traceback (most recent call last):
File "<pyshell#30>", line 1, in -toplevel-
1/i
ZeroDivisionError: integer division or modulo by zero

>>b = a * 1.5
b

array([[ 0. , 1.5],
[ 3. , 4.5]])

>>N.floor(b)

array([[ 0., 1.],
[ 3., 4.]])

>>================================ RESTART

================================

>>import numpy as N
a = N.array([[0,1],[2,3]])
a

array([[0, 1],
[2, 3]])

>>i = a[0,0]
1/i

0

>>b = a * 1.5
b

array([[ 0. , 1.5],
[ 3. , 4.5]])

>>N.floor(b)

array([[ 0., 1.],
[ 3., 4.]])

>>a = N.array([[0,1],[2,3]], dtype='O')
a

array([[0, 1],
[2, 3]], dtype=object)

>>i = a[0,0]
1/i

Traceback (most recent call last):
File "<pyshell#45>", line 1, in -toplevel-
1/i
ZeroDivisionError: integer division or modulo by zero

>>b = a * 1.5
b

array([[0.0, 1.5],
[3.0, 4.5]], dtype=object)

>>N.floor(b)

Traceback (most recent call last):
File "<pyshell#48>", line 1, in -toplevel-
N.floor(b)
AttributeError: 'float' object has no attribute 'floor'

>>>

----------------------------------------------

An additional problem involves classes that have e.g. __rmul__ methods
defined and are sufficiently similar to numpy arrays that my classes'
__rmul__ methods are not invoked when using numpy scalars.
Using the 'O' dtype gives me Python types that raise zero division
errors appropriately (for my code) and the desired calls to e.g.
__rmul__ methods, but reduced functionality in other repects.

I might (I hope) be missing something obvious; but it seems like, to be
safe, I'm going to have to do a lot of explicit conversions to Python
types (or abandon catching zero division errors, and documenting some of
my classes to highlight that whether scalar * a equals a * scalar
depends on whether a.__rmul__ is called, which depends on the type of
scalar).

I suppose I might get round both issues by subclassing existing numpy
dtypes. Any ideas? Cheers. TIA.

Duncan

[Travis E. Oliphant]

Duncan Smith wrote:

Hello,
Since moving to numpy I've had a few problems with my existing
code. It basically revolves around the numpy scalar types. e.g.

You will probably get more help on the numpy discussion list:

nu**************@scipy.org
You are encountering problems because numpy scalar types don't raise
errors (unless you have set the appropriate hardware flag using
numpy.seterr).

You can get Python scalars out of NumPy arrays if you really want them
using (for example...)

a.item(0,0)

>
An additional problem involves classes that have e.g. __rmul__ methods
defined and are sufficiently similar to numpy arrays that my classes'
__rmul__ methods are not invoked when using numpy scalars.

Could you please post an example showing the problem?

>
I might (I hope) be missing something obvious; but it seems like, to be
safe, I'm going to have to do a lot of explicit conversions to Python
types (or abandon catching zero division errors, and documenting some of
my classes to highlight that whether scalar * a equals a * scalar
depends on whether a.__rmul__ is called, which depends on the type of
scalar).

numpy scalars are try a lot more things before giving up on
multiplication and letting the other class have a stab at it.

Post your problems to the numpy discussion list for better help and more
discussion.
-Travis

[Duncan Smith]

Travis E. Oliphant wrote:

Duncan Smith wrote:

>Hello,
Since moving to numpy I've had a few problems with my existing
code. It basically revolves around the numpy scalar types. e.g.

You will probably get more help on the numpy discussion list:

nu**************@scipy.org
You are encountering problems because numpy scalar types don't raise
errors (unless you have set the appropriate hardware flag using
numpy.seterr).

Aha!

You can get Python scalars out of NumPy arrays if you really want them
using (for example...)

a.item(0,0)

>>
An additional problem involves classes that have e.g. __rmul__ methods
defined and are sufficiently similar to numpy arrays that my classes'
__rmul__ methods are not invoked when using numpy scalars.

Could you please post an example showing the problem?

I'll try to post a minimal example tomorrow. But they are classes that
have an ndarray as an attribute, and with __getitem__ and __setitem__
methods which simply call the corresponding array methods. Maybe that's
enough to account for the behaviour? I'll check tomorrow.

>>
I might (I hope) be missing something obvious; but it seems like, to be
safe, I'm going to have to do a lot of explicit conversions to Python
types (or abandon catching zero division errors, and documenting some of
my classes to highlight that whether scalar * a equals a * scalar
depends on whether a.__rmul__ is called, which depends on the type of
scalar).

numpy scalars are try a lot more things before giving up on
multiplication and letting the other class have a stab at it.

Post your problems to the numpy discussion list for better help and more
discussion.

Yes, I have done. But it's awaiting moderation; presumably because I
posted using a different e-mail address than the one I registered with
(I wasn't thinking). Thanks for the reply.

Duncan

[Duncan Smith]

Travis E. Oliphant wrote:

Duncan Smith wrote:

>Hello,
Since moving to numpy I've had a few problems with my existing
code. It basically revolves around the numpy scalar types. e.g.

You will probably get more help on the numpy discussion list:

nu**************@scipy.org
You are encountering problems because numpy scalar types don't raise
errors (unless you have set the appropriate hardware flag using
numpy.seterr).

Unfortunately it seems to raise a FloatingPointError.

>>import numpy as N
N.__version__

'1.0.1'

>>a = N.array([[0,1],[2,3]])
a

array([[0, 1],
[2, 3]])

>>i = a[0,0]
1/i

0

>>N.seterr(divide='raise')

{'over': 'print', 'divide': 'print', 'invalid': 'print', 'under': 'ignore'}

>>1/i

Traceback (most recent call last):
File "<pyshell#9>", line 1, in <module>
1/i
FloatingPointError: divide by zero encountered in long_scalars

You can get Python scalars out of NumPy arrays if you really want them
using (for example...)

a.item(0,0)

>>
An additional problem involves classes that have e.g. __rmul__ methods
defined and are sufficiently similar to numpy arrays that my classes'
__rmul__ methods are not invoked when using numpy scalars.

Could you please post an example showing the problem?

[snip]

-----------------example.py--------------------

from __future__ import division

import numpy

class MyClass(object):

def __init__(self, arr, labels):
self.arr = arr
self.labels = labels

def __repr__(self):
return numpy.array2string(self.arr, separator=', ') +
repr(self.labels)

def __len__(self):
return len(self.labels)

def __getitem__(self, key):
return self.arr[key]

def __setitem__(self, key, item):
self.arr[key] = item

def __mul__(self, other):
return self.__class__(self.arr * other, self.labels)

__rmul__ = __mul__

----------------------------------------------------

>>import example
import numpy as N
ex = example.MyClass(N.array([[6,7],[8,9]]), ['axis0', 'axis1'])
i = ex.arr[0,0]
ex

[[6, 7],
[8, 9]]['axis0', 'axis1']

>>ex * i

[[36, 42],
[48, 54]]['axis0', 'axis1']

>>i * ex

array([[36, 42],
[48, 54]])

>>>

It seems that it requires having __len__, __setitem__ and __getitem__
defined to get the undesired behaviour. Cheers.

Duncan