Improvements to Composed and TruncatedBijector fixing #40

[torfjelde]

Overview

This PR introduces most of the performance improvements to Composed discussed in #46 and the TruncatedBijector discussed in #40 making bijector(d::Distribution) type-stable and providing a good "default" bijector.

In addition to the above, Test.@inferred is now used extensively in test/interface.jl to try and catch instability for the <:Tuple versions of Composed and Stacked.

Finally the project structure has been changed a bit (for the better). Implementations of Bijector are now found in the src/bijectors directory, e.g. src/bijectors/scale.jl. This makes it much easier for people unfamiliar with the code to contribute, plus easier to reason about in general.

Changes

Composed

From a user-perspective, the major change is that ∘ will now "flatten" the composition structure in the sense that

julia> using Bijectors; using Bijectors: Exp

julia> b = Exp()
Exp{0}()

julia> cb1 = b ∘ b;

julia> cb2 = b ∘ b;

julia> (cb1 ∘ cb2).ts # <= different
(Exp{0}(), Exp{0}(), Exp{0}(), Exp{0}())

julia> (cb1 ∘ cb2).ts isa NTuple{4, Exp{0}}
true

while in master it looks like

julia> using Bijectors; using Bijectors: Exp
[ Info: Recompiling stale cache file /home/tor/.julia/compiled/v1.1/Bijectors/39uFz.ji for Bijectors [76274a88-744f-5084-9051-94815aaf08c4]

julia> b = Exp()
Exp{0}()

julia> cb1 = b ∘ b;

julia> cb2 = b ∘ b;

julia> (cb1 ∘ cb2).ts # <= different
(Composed{Tuple{Exp{0},Exp{0}},0}((Exp{0}(), Exp{0}())), Composed{Tuple{Exp{0},Exp{0}},0}((Exp{0}(), Exp{0}())))

julia> (cb1 ∘ cb2).ts isa Tuple{Composed, Composed}
true

Note that this is only the case for ∘. In some cases it might be desirable to keep certain operations together, e.g. Composed{Tuple{Scale, Shift}}} represents an affine transformation. Therefore composel and composer will not do this "flattening", but instead preserve the structure of the compositions.

The above is also described in the docstring of Composed.

[xukai92]

Looks good to me overall!

[xukai92]

@torfjelde Can I merge this now?

[torfjelde]

@torfjelde Can I merge this now?

Yup! Though maybe I should bump the version number? I guess this would technically be a breaking change since the circ-operation now has different behavior. Should I bump the minor- or patch-version?

EDIT: Actually, let me do a "review" myself tomorrow and make sure everything is good. Shouldn't be any bugs, just filling in some simply additions e.g. just realized I had forgotten to implement the transform and logabsdetjac for Composed using the @generated approach we're currently doing for forward.

[torfjelde]

Should be good now @kai!:)

[xukai92]

Thanks for the great work!