Hmm, I see your point. Looks like we need a "good" version of this. Would you
please file an issue requesting this feature?
In the meantime, if you need this for only special cases (say, in 3
dimensions, where you're always taking the max along the 3rd dimension), it
would only be a handful of lines to write a good version using loops. Or, as
you did before, use mapslices.
--Tim
On Thursday, May 01, 2014 09:40:09 AM Florian Oswald wrote:
> actually, the ismaxfun is not doing that I want. I want to have the maximum
> value and its index in the given dimension. while the maximum value is
> fine, I can't get the index:
>
> julia> function ismaxfun(x::Array{Float64},d::Int)
> z = maximum(x,d)
> ismax = x .== z
> (z,ismax)
> end
>
> A = rand(2,2,2)
> x = ismaxfun(A,3)
> maxval = squeeze(x[1],3)
> maxind = find(x[2])
>
> maxind contains the linear indices of the maxima in dimension 3. I want to
> map those indices into an array Z of size(maxval)=(2,2), where the value at
> Z(i,j) tells me what the maximal index into along A[i,j,:] is. I can get
> this
>
> julia> map(x->ind2sub((2,2,2),x),maxind)
> 4-element Array{(Int64,Int64,Int64),1}:
> (2,1,1)
> (1,1,2)
> (1,2,2)
> (2,2,2)
>
> which is close, but really I want that:
>
> ms=mapslices(findmax,A,3)
> julia> map(x->x[2],ms)
> 2x2x1 Array{Int64,3}:
> [:, :, 1] =
> 2 2
> 1 2
>
>
> Any suggestion on how to achieve this using ismaxfun above? thanks.
>
> On Wednesday, 30 April 2014 15:58:16 UTC+1, Florian Oswald wrote:
> > jeez, you are right! sorry i'm doing my first steps here - forget about
> > the compile step. it IS much faster than 2x. after changing ismaxfun to
> > output something similar to slicefun, it's 10x faster. but maybe my
> > reshaping is not the smartest idea either. anyway, that's pretty good!
> >
> > julia> function ismaxfun(x::Array{Float64,6})
> >
> > z = maximum(x,6)
> > ismax = x .== z
> > id = reshape(find(ismax),size(x)[1:5])
> > end
> >
> > A = rand(10,10,10,10,10,50)
> >
> > julia> @time k = slicefun(A)
> > elapsed time: 0.912256696 seconds (213759264 bytes allocated)
> >
> > julia> @time k = ismaxfun(A)
> > elapsed time: 0.089194917 seconds (3484324 bytes allocated)
> >
> > On Wednesday, 30 April 2014 15:47:30 UTC+1, Tim Holy wrote:
> >> I time it at 50x faster:
> >>
> >> julia> @time slicefun(A);
> >> elapsed time: 1.135524683 seconds (181765472 bytes allocated)
> >>
> >> julia> @time ismaxfun(A);
> >> elapsed time: 0.020300699 seconds (931120 bytes allocated)
> >>
> >>
> >> Did you run it a second time? (You don't want to include the JIT timing.)
> >>
> >> --Tim
> >>
> >> On Wednesday, April 30, 2014 07:38:07 AM Florian Oswald wrote:
> >> > I see. thanks!
> >> > you are (of course) right with the timing:
> >> >
> >> > julia> function slicefun(x::Array{Float64,6})
> >> >
> >> > z = mapslices(findmax,x,6)
> >> > end
> >> >
> >> > julia> function ismaxfun(x::Array{Float64,6})
> >> >
> >> > z = maximum(x,6)
> >> > ismax = x .== z
> >> > end
> >> >
> >> > times as follows. more than twice as fast!
> >> >
> >> > julia> @time k = ismaxfun(A)
> >> > elapsed time: 0.53359244 seconds (14418060 bytes allocated)
> >> > julia> @time k = slicefun(A)
> >> > elapsed time: 1.396534023 seconds (226762376 bytes allocated)
> >> >
> >> > On Wednesday, 30 April 2014 15:25:25 UTC+1, Tim Holy wrote:
> >> > > There is
> >> > >
> >> > > maximum(A, 3)
> >> > >
> >> > > but I gather you also want the index? If so, mapslices is your best
> >> > > one-liner
> >> > > as of now. But if you're performance-sensitive, you might also look
> >>
> >> into
> >>
> >> > > this:
> >> > > M = maximum(A, dims)
> >> > > ismax = A .== M
> >> > >
> >> > > and then find the `true` elements of ismax. I'll bet that's quite a
> >>
> >> lot
> >>
> >> > > faster
> >> > > than using mapslices, even though it requires two traversals of the
> >>
> >> array.
> >>
> >> > > As
> >> > > a bonus, you'll also learn about ties.
> >> > >
> >> > > --Tim
> >> > >
> >> > > On Wednesday, April 30, 2014 04:04:47 AM Florian Oswald wrote:
> >> > > > I looked at the issue open at
> >> > > >
> >> > > > https://github.com/JuliaLang/julia/issues/3893
> >> > > >
> >> > > > but couldn't figure out what's the best thing to do. I gather that
> >>
> >> there
> >>
> >> > > > will be a
> >> > > >
> >> > > > findmax(A,dims)
> >> > > >
> >> > > > in some future version of julia? that's certainly what im looking
> >>
> >> for
> >>
> >> > > here.
> >> > >
> >> > > > in the meantime, is this the best I can do?
> >> > > >
> >> > > > A = randn(3,3,3)
> >> > > > mapslices(findmax,A,3)
> >> > > >
> >> > > > cheers
