Can we make at-expression more general?

[IainNZ]

At the moment if only works for linear expressions, which means it can't be used for extensions (or even quadratics). One option would be just to use JuMPDicts of Anys, I guess, but that could hurt you if you wanted to use it for matrix ops later I suppose.

[mlubin]

The scalar @defExpr should work for generic expressions on 0.4 (but not 0.3).
For collections of expressions, we should probably choose the type of the JuMP container optimistically (like map) and widen it if we need to. Will be easier to implement once the JuMPArray code becomes nicer.

[IainNZ]

Yeah I was thinking of collections.

[mlubin]

This still doesn't work:

julia> @expression(m, [i=1:2], x^2)
ERROR: Collection of expressions with @expression must be linear. For quadratic expressions, use your own array.

[blegat]

We should make containers infer the element type by creating the first element and looking at its value. This would also remove the need for JuMP.constraint_type which is bad since it does not return a concrete type

[mlubin]

We should make containers infer the element type by creating the first element and looking at its value.

We shouldn't assume that the element type is homogeneous. We should use the same approach as map and widen/promote if needed:

julia> map(+, Any[1, 2, 3], Any[10, 20, 30])
3-element Array{Int64,1}:
 11
 22
 33

julia> map(+, Any[1, 2, 3], Any[10, 20, 30.0])
3-element Array{Real,1}:
 11  
 22  
 33.0

[blegat]

We shouldn't assume that the element type is homogeneous. We should use the same approach as map and widen/promote if needed

I agree, normally, this is what is done by Base.collect_to_with_first! for Vector, implementing Base.collect_to_with_first! for Containers might be the way to go.
The broad idea is instead of creating nested loops, we create a function that given the indices, return the element for these indices, we feed this function to a Base.Generator and then we collect. That is what map does.