On 11/15/2011 07:03 PM, Gael Varoquaux wrote: > On Tue, Nov 15, 2011 at 05:57:14PM +0000, Robert Kern wrote: >> Actually, last time I suggested it, it was brought up that the online >> algorithms can be worse numerically. I'll try to find the thread. > Indeed, especially for smallish datasets where the memory overhead is not > an issue. I think that this is a common situation for which there is not > _one_ good algorithm to solve the problem. IMHO the best way forward is > to propose several options to the user, either with several function, or > with a keyword argument. > I thought it would be possible to do the "two pass algorithm" (as wikipedia calls it) without copying the dataset. I guess I didn't really think it through. That should be possible doing a custom reduce operation, right? I don't know in how far this would have numerical problems.
_______________________________________________ NumPy-Discussion mailing list [email protected] http://mail.scipy.org/mailman/listinfo/numpy-discussion
