Rename Backward to Reverse

[blegat]

What's the correct terminology ? The classical terminology is forward mode AD vs reverse mode AD, as in, frule and rrule so we should have Reverse... attributes and DiffOpt.reverse instead of Backward... and DiffOpt.backward.

[matbesancon]

Sounds reasonable, let's do this now more than later then

[joaquimg]

Not a big fan of DiffOpt.reverse because there is a function with that name in base. Also, very personally, I hate the asymmetry between forward X reverse instead of forward X backward, but I agree that if it will make it better for users we change.

Maybe DiffOpt.compute_reverse_diff and consequently DiffOpt.compute_forward_diff.

We can probably bikeshed a bit on the attribute names also...

DiffOpt.BackwardInVariablePrimal() could be DiffOpt.BackwardInputVariable()
and
DiffOpt.BackwardOutConstraint() could be DiffOpt.BackwardOutputConstraint()

Very open to any suggestions...

[blegat]

I would rename forward to forward_differentiate! and backward to reverse_differentiate!.
Then, I would rename BackwardInVariablePrimal to ReverseVariablePrimal. The fact that it's the input is implied by the fact that it's the variable primal in reverse mode so I don't think we need to add that distinction in the name.
It's also clear when you read code since you see that it is being set.

[matbesancon]

For searchability and clarity for users, I think keeping In or Input is beneficial

[joaquimg]

Makes sense, nowadays I leaning more towards Input. In might be slighty confusing.

[joaquimg]

On second thought, I am between Input or nothing (since it is implied), we can also throw nice errors for wrong combinations of set/get and the atributes

[joaquimg]

I like forward_differentiate! and backward to reverse_differentiate!

[blegat]

Having both a ForwardIn... and ReverseOut... with one you can set and the other you can get is kind of weird.
It just means use In whenever you use set and use Out whenever you use get. So it's completely redundant.
The information of whether you want In or Out is already encoded in the fact that you are using get or set.

[joaquimg]

For @matbesancon point, we might be able to solve that via making it crystal clear on docs.

[joaquimg]

What about VariablePrimal versus Variable, since we have Constraint instead of ConstraintPrimal?

[matbesancon]

yeah we can indicate it in the doc but it's handy to just do DiffOpt.For and get immediately what you want without a doubt

[blegat]

What about VariablePrimal versus Variable, since we have Constraint instead of ConstraintPrimal?

For the constraint, we are modifying the constraint function so it should be ConstraintFunction and ObjectiveFunction and VariablePrimal

yeah we can indicate it in the doc but it's handy to just do DiffOpt.For and get immediately what you want without a doubt

That workflow would still work. The complete specification of what you want always need the full method signature.

[matbesancon]

we are modifying the constraint function

Not really no? We are modifying both constraint function and set, since parameters can occur in both, we just normalize by switching the set constant to the function part

[blegat]

For the constant, we can modify the function constant. The fact it is rewritten internally as a modification to the set is abstracted out.
I think we could have either Variable/Constraint/Objective or VariablePrimal/ConstraintFunction/ObjectiveFunction but not a mix.
I have a preference for the second one as it is more consistent with the MOI attributes.

[blegat]

I think ReverseObjective should be ReverseObjectiveFunction then

[matbesancon]

Done