Ai2{Hyperrectangle} Id layer returns Zonotope

[SebastianGuadalupe]

The current implementation is

function forward_layer(solver::Ai2, L::Layer{Id}, input)
    return affine_map(L, input)
end

but what I understand is correct would be

forward_layer(solver::Ai2, L::Layer{Id}, input) = input

Edit: This is not true, but there is a type instability in the function, when passing a Hyperrectangle, it returns a Zonotope

[tomerarnon]

Technically this is not a type instability, since the return type of affine_map(L, ::Hyperrectangle) is inferable (always a Zonotope). I don't think this is really a problem, except that it is maybe surprising to get a Zonotope out of the Box solver (even though that is more precise than overapproximating the output as a hyperrectangle). Although it loses a bit of precision, I wouldn't mind changing it to return a hyperrectangle, which I think matches your original implementation.

I did note a bug looking into this that we haven't run into because we haven't tested with any networks that have identity layers in the middle (always at the end)

NeuralVerification.jl/src/reachability/ai2.jl

Lines 77 to 80 in f36efbe

	function forward_layer(solver::Box, L::Layer{ReLU}, input::AbstractZonotope)
	Ẑ = approximate_affine_map(L, input)
	return rectify(Ẑ)
	end

approximate_affine_map above wouldn't actually work for AbstractZonotope.