[Numpy-discussion] Upgrading from numpy 0.9.7.2416 to 0.9.9.2683
Travis Oliphant
oliphant.travis at ieee.org
Tue Jun 27 13:50:05 CDT 2006
Stefan van der Walt wrote:
> On Tue, Jun 27, 2006 at 09:45:57AM -0700, Keith Goodman wrote:
>
> With r2691 I see
>
> In [7]: x = N.asmatrix(N.zeros((3,2)),float)
>
> In [8]: y = N.asmatrix(N.rand(3,1))
>
> In [12]: x[:,1] = y > 0.5
>
> In [13]: x
> Out[13]:
> matrix([[ 0., 1.],
> [ 0., 1.],
> [ 0., 1.]])
>
This was a bug, indirectly caused by the move to broadcasted copying and
casting and the use of a matrix here. Previously the shapes didn't
matter as long as the total size was the same.
Thus internally x[:,1] was creating a (1,3) matrix referencing the last
column of x (call it xp) and y>0.5 was a (3,1) matrix (call it yp)
Thus the resulting casting code was repeatedly filling in x with
(y>0.5). Thus, the last entry of (y>0.5) was the one that resulted.
Previously, this would have worked because the shape of the arrays
didn't matter, but now they do. The real culprit was not allowing the
matrices "getitem" method to be called (which would have correctly
obtained a (3,1) matrix from x[:,1] and thus avoided the strange result.
Thus, in SVN, now PyObject_GetItem is used instead of the default
ndarray getitem. The upshot, is that this should now work --- and
there is now a unit-test to check for it.
Thanks to Keith for exposing this bug.
-Travis
More information about the Numpy-discussion
mailing list