Define == for Bijector

[mohamed82008]

This PR defines == for bijectors. This is needed to check that the support of a distribution hasn't changed in DynamicPPL.

[torfjelde]

Uhhh a heads up: most of these should just work because Julia, e.g.

julia> using Bijectors: Exp, Log, Logit

julia> Exp() == Exp()
true

julia> Logit(0., 1.) == Logit(1., 2.)
false

julia> Logit(0., 1.) == Logit(0., 1.)
true

julia> b1 = Exp() ∘ Logit(0., 1.); b2 = Exp() ∘ Logit(0., 1.)
Composed{Tuple{Logit{0,Float64},Exp{0}},0}((Logit{0,Float64}(0.0, 1.0), Exp{0}()))

julia> b1 == b2
true

julia> b1 = Exp() ∘ Logit(0., 1.); b2 = Exp() ∘ Logit(0., 2.)
Composed{Tuple{Logit{0,Float64},Exp{0}},0}((Logit{0,Float64}(0.0, 2.0), Exp{0}()))

julia> b1 == b2
false

Are there any particular ones for which it doesn't work?

[torfjelde]

Probably only needed for mutable ones, e.g. PlanarLayer, RadialLayer and Normalize (I'm not entirely sure why we need these to be mutable though; could just make the mutations be performed elementwise using .= since the fields are all arrays anyways)

[mohamed82008]

Uhhh a heads up: most of these should just work because Julia, e.g.
Are there any particular ones for which it doesn't work?

Hmm let me see.

[torfjelde]

Related: JuliaLang/julia#4648

Immutable structs are compared by recursively applying ==, but this is not the case for mutable (because issues: JuliaLang/julia#4648 (comment))

[mohamed82008]

Permute seems to need an == method too.

[torfjelde]

Curious. So from the issue it seems like == will recursively call isequal and, lo and behold,

julia> b1 = Permute(2, 1 => 2, 2 => 1)
Permute{SparseArrays.SparseMatrixCSC{Float64,Int64}}(
  [2, 1]  =  1.0
  [1, 2]  =  1.0)

julia> b2 = Permute(2, 1 => 2, 2 => 1)
Permute{SparseArrays.SparseMatrixCSC{Float64,Int64}}(
  [2, 1]  =  1.0
  [1, 2]  =  1.0)

julia> b1 == b2
false

julia> b1.A == b2.A
true

julia> isequal(b1.A) == isequal(b2.A)
false

Not clear to me why this happens. Might be an issue with SparseArrays.jl, because there's nothing in the docstring of isequal that tells me they shouldn't be equal.

[mohamed82008]

I think the recursive "rule" doesn't hold for arrays in general.

julia> struct T
           a
       end

julia> T(ones(2)) == T(ones(2))
false

[mohamed82008]

So we need == methods for any bijector that can have an array field.

[torfjelde]

Ah, yeah seems like you're right 👍

[mohamed82008]

@torfjelde I think we need == for all bijectors with numeric fields for this to return true:

julia> using Bijectors, ReverseDiff

julia> Bijectors.Logit(0, 1) == Bijectors.Logit(ReverseDiff.track(0), ReverseDiff.track(1))
false

This is not super important as the AD number types showing up in the bijector is already a sign that the support of the distribution is changing so we might as well get false here. But I would still like to see true in the above case.