Gibbs does not use initialisation strategies for component samplers

[penelopeysm]

Minimal working example

using Turing, Bijectors

dist = Beta(1, 1)
binv = Bijectors.inverse(bijector(dist))

@model function f()
    h ~ dist
    m ~ dist
end

spl = Gibbs(@varname(h) => HMC(0.1, 5), @varname(m) => MH())

chn = sample(f(), spl, MCMCThreads(), 1, 100)
all(h -> binv(-2) < h < binv(2), chn[:h]) # false

chn2 = sample(f(), spl, MCMCThreads(), 1, 100; initial_params=fill(InitFromUniform(), 100))
all(h -> binv(-2) < h < binv(2), chn2[:h]) # true

Description

The Gibbs sampler does not make use of initialisation strategies for component samplers. That's because it uses Turing.Inference.init_strategy(::Gibbs) which resolves to InitFromPrior, and init_strategy(::HMC) is never called.

In the above example the HMC variable should always be initialised to (binv(-2), binv(2)) but it isn't as can be seen from the check.

The solution to this is to write a custom AbstractInitStrategy for Gibbs which, I think, maps varnames to init strategies of the component samplers: so this would wrap @varname(h) => InitFromUniform(), @varname(m) => InitFromPrior(). Then inside DynamicPPL.init it would check the varname and dispatch to the appropriate strategy.

Julia version info

1.11.7

Manifest

0.41.0 (currently using #2689)

Note the same is true on previous versions as the first call to all(...) also returns false on 0.40.