Since you asked...from my perspective, the easiest argument against it is
pnormalized = p ./ sum(p, 1)
If you drop the summed dimension, that won't work anymore. One can write
pnormalized = p ./ reshape(sum(p, 1), 1, size(p, 2))
but that's a little uglier than the converse, calling `squeeze` when you'd
rather get rid of the dimension.
That said, I recognize that your point has validity. I offer the example just
to make sure it's known.
Best,
--Tim
On Wednesday, May 25, 2016 06:49:53 PM a. kramer wrote:
> That's a solution, yes, but my feeling is that the "default" slicing and
> reduction behavior should play nice with one another. In 0.5 it seems
> dropping dimensions with A[1,:] is default, which I do prefer. It seems
> natural for sum(A,1) to be equivalent to A[1,:] + A[2,:] + A[3,:] + ...
>
> In my workflow (primarily data analysis), I often mix these slicing and
> reduction operations, but it would be unusual for a situation to appear
> where I would want one behavior for slicing and a different one for array
> reduction. In situations where array dimensions correspond to, say,
> repeated observations, it is common to compare slices to means or maxima
> across dimensions. However, I would be interested to hear arguments
> against.
>
> On Wednesday, May 25, 2016 at 9:05:04 PM UTC-4, Gabriel Gellner wrote:
> > Does it bother you to do the A[[1], :] to keep the old behavior? I haven't
> > thought enough about the role of sum, mean etc.
> >
> > On Wednesday, May 25, 2016 at 5:11:20 PM UTC-7, a. kramer wrote:
> >> Apologies in advance if this is something that has been discussed at
> >> length already, I wasn't able to find it.
> >>
> >> In Julia 0.5, if A is a 5x5 matrix, the behavior of A[1,:] will be
> >> changed to return a 5-element array instead of a 1x5 array. However, at
> >> least in the current build, sum(A,1) still gives a 1x5 array as it does
> >> in
> >> earlier versions, and similarly for other array reductions like mean,
> >> maximum, etc.
> >>
> >> I understand that these functions and slicing are fundamentally different
> >> things, but I find this a little counter intuitive (at the very least
> >> different from numpy's behavior). I often find myself interested in
> >> quantities such as A[1,:] ./ mean(A,1) (the first row of a matrix
> >> normalized by the average of each column's entries). In 0.5, this gives
> >> something quite different from what I'm expecting (in fact it gives a 5x5
> >> matrix).
> >>
> >> So my questions are: Is there a discussion of the rationale behind doing
> >> things this way? Is this something that may be changed in the future?
> >> If
> >> not, is there an alternative to the standard sum, mean, etc. functions
> >> that
> >> is recommended for this? Just a liberal use of squeeze()?
