Fold (reduce)? about l1

[mlajtos]

This is something I would love to support the right way, but I am afraid it won't happen. But there will be a +/ operator for Sum and */ for Product.

I have really no idea how to support generalized reductions. :(

4666e21

[suhr]

Well, that's unfortunate. But what's the problem with generalized reductions?

[mlajtos]

Ok, I might have been quick to judge.

So far, the only problem right now is that scan is not supported by the TF.js. When it will be available, it should not be a problem to implement it.

[mlajtos]

Until tf.scan is available, there will be fake +/ and */ and then I will revisit the reductions in a proper way. So closing for now...

[mlajtos]

::Sum – introduced in be02a2e

I lied. Sum is an unary +, not +/. Same will go for a Product as unary *.

(Originally, I intended unary * as an operator for Sign function, but this is much better.)

[mlajtos]

I read my comment again, and I realized I mixed up reduction and scan... As for cumsum (+\) I have no idea what would be a good notation, or how to support tf.scan in the future. But the question was about the reduction...

I think that prefix notation is not as powerful as reduction notation from APL – one obvious example is -[1 2 3] – but it seems good enough to me.

So, non-commutative binary operators in unary position:
-[1 2 3] = 0 - [1 2 3] – works
/[1 2 3] = 1 / [1 2 3] – works
^[1 2 3] = e ^ [1 2 3] – need work
%[1 2 3] = ??? – no idea, (as postfix it could be (a => a / 100), so prefix in reverse?, or something more useful :))
@[1 2 3] = ??? – no idea

Non-existing (random ideas):
<[1 2 3] = Min [1 2 3] – breaks rule of non-commutative bi-op in un-pos (because it reduces)
>[1 2 3] = Max [1 2 3] – same