Introduce add!!

[oxinabox]

This closes #113
at least the most important part.
I don't think we need store!,
and it doesn't do multi-arg accumulation, on the basis that we can add that later, lets not gold-plate this with features we don't yet need.
We can open a follow up issue about that after this is merged

It is named accumulate!! after the BangBang.jl convention.
a suffix of !! means I might mute the input, or i might not, but i will return definately return the thing

[nickrobinson251]

One ! not enough hype?

[nickrobinson251]

the BangBang.jl convention.
a suffix of !! means I might mute the input, or i might not, but i will return definately return the thing

I don't see this convention documented this way (at least not so succinctly) in the BangBang.jl docs... I'm sure you're right about the convention, but i'd feel better if it were explained in the BangBang docs so we could point to it

[willtebbutt]

This is nice.

Regarding naming: is there a reason we can't call this function add! or sum! or something? While accumulate!! does correctly convey the intuition that we're somehow combining a collection of (co)tangents, it doesn't naturally convey the idea that the particular way in which we're "accumulating" these objects is via addition.

[oxinabox]

I would be down with add!!, I will make that change.
Much better since it means it doesn't conflict with defination of accumulate exported by Base that does take an operation.

Going for add reather than sum as for sum it should take a vector to things to be summed, rather than discrete arguments

@tkf confirmed for me that this was the correct convention on Slack.

[oxinabox]

I am still calling the file accumulation.jl as that refers to gradient accumultion.
and if we want to bring back store! etc then that would also go here.

[nickrobinson251]

straying away from this PR... should we have an issue to adopt !! naming convention throughoutt ChainRules? e.g. it;s apply here i think: https://github.com/JuliaDiff/ChainRules.jl/blob/master/src/rulesets/LinearAlgebra/utils.jl#L29-L36

[nickrobinson251]

some questions but either answer is fine, so LGTM

[nickrobinson251]

do we have/need an issue for this?

[oxinabox]

we don't, I am not sure we need one yet.

[oxinabox]

should we have an issue to adopt !! naming convention throughoutt ChainRules? e.g. it;s apply here i think: https://github.com/JuliaDiff/ChainRules.jl/blob/master/src/rulesets/LinearAlgebra/utils.jl#L29-L36

That particular case should just be a call to add!!
I will change that and look for others