[Numpy-discussion] Simple multi-arg wrapper for dot()

Bill Baxter wbaxter@gmail....
Sat Mar 24 14:03:11 CDT 2007


On 3/24/07, Anne Archibald <peridot.faceted@gmail.com> wrote:
> On 24/03/07, Bill Baxter <wbaxter@gmail.com> wrote:
>
> > Nice, but how does that fare on things like mdot(a,(b,c),d) ?  I'm
> > pretty sure it doesn't handle it.
> > I think an mdot that can only multiply things left to right comes up
> > short compared to an infix operator that can easily use parentheses to
> > control order.
>
> Well, since exact dotting is associative, the parenthesization doesn't
> matter there; if you're worried about roundoff or efficiency, you're
> going to have to be more explicit.

Yes performance was the thing I was thinking about.  You often know a
little something about the sizes of your inputs that can help make
things more efficient.  Like
          BigMatrix1 * BigMatrix2 * LittleVector
Is much faster to calculate as
          BigMatrix1 * (BigMatrix2 * LittleVector)
than
          (BigMatrix1 * BigMatrix2) * LittleVector

> Unfortunately your approach requires tuples to be treated differently from arrays.

Yeh, that was a conscious decision (which I noted in the docstring).
You are right that it is different from what Numpy normally does,
which is one reason I didn't say "lets replace dot with this
function".

> Most functions
> in numpy will happily treat a tuple of length n, a list of length n,
> and an array of length n the same way.

Yep, it's nice in general.  But in practice I very rarely do math on
tuples.  The only thing I can think of is I used to do prod(arr.shape)
before I realized that arr.size give the same thing.  I'm positive
I've never used dot on tuples.

> You could do this, and for your
> own code maybe it's worth it, but I think it would be confusing in the
> library.

Could be.  Doesn't seem so confusing to me as long as it's documented
clearly in the docstring, but YMMV.

--bb


More information about the Numpy-discussion mailing list