DPPL 0.37 compat for particle MCMC (#2625)

* Progress in DPPL 0.37 compat for particle MCMC

* WIP PMCMC work

* Gibbs fixes for DPPL 0.37 (plus tiny bugfixes for ESS + HMC) (#2628)

* Obviously this single commit will make Gibbs work

* Fixes for ESS

* Fix HMC call

* improve some comments

* Fixes to ProduceLogLikelihoodAccumulator

* Use LogProbAccumulator for ProduceLogLikelihoodAccumulator

* use get_conditioned_gibbs

---------

Co-authored-by: Penelope Yong <[email protected]>

Author: mhauru

src/mcmc/ess.jl

@@ -54,7 +54,7 @@ function AbstractMCMC.step(
 
     # update sample and log-likelihood
     vi = DynamicPPL.unflatten(vi, sample)
-    vi = setloglikelihood!!(vi, state.loglikelihood)
+    vi = DynamicPPL.setloglikelihood!!(vi, state.loglikelihood)
 
     return Transition(model, vi), vi
 end
@@ -88,6 +88,11 @@ function Base.rand(rng::Random.AbstractRNG, p::ESSPrior)
     # p.varinfo, PriorInit())` after TuringLang/DynamicPPL.jl#984. The reason
     # why we had to use the 'del' flag before this was because
     # SampleFromPrior() wouldn't overwrite existing variables.
+    # The main problem I'm rather unsure about is ESS-within-Gibbs. The
+    # current implementation I think makes sure to only resample the variables
+    # that 'belong' to the current ESS sampler. InitContext on the other hand
+    # would resample all variables in the model (??) Need to think about this
+    # carefully.
     vns = keys(varinfo)
     for vn in vns
         set_flag!(varinfo, vn, "del")
@@ -102,13 +107,13 @@ Distributions.mean(p::ESSPrior) = p.μ
 # Evaluate log-likelihood of proposals. We need this struct because
 # EllipticalSliceSampling.jl expects a callable struct / a function as its
 # likelihood.
-struct ESSLikelihood{M<:Model,V<:AbstractVarInfo}
-    ldf::DynamicPPL.LogDensityFunction{M,V}
+struct ESSLikelihood{L<:DynamicPPL.LogDensityFunction}
+    ldf::L
 
     # Force usage of `getloglikelihood` in inner constructor
     function ESSLikelihood(model::Model, varinfo::AbstractVarInfo)
         ldf = DynamicPPL.LogDensityFunction(model, DynamicPPL.getloglikelihood, varinfo)
-        return new{typeof(model),typeof(varinfo)}(ldf)
+        return new{typeof(ldf)}(ldf)
     end
 end
 

src/mcmc/gibbs.jl

@@ -177,13 +177,15 @@ function DynamicPPL.tilde_assume(context::GibbsContext, right, vn, vi)
         # Fall back to the default behavior.
         DynamicPPL.tilde_assume(child_context, right, vn, vi)
     elseif has_conditioned_gibbs(context, vn)
-        # Short-circuit the tilde assume if `vn` is present in `context`.
-        # TODO(mhauru) Fix accumulation here. In this branch anything that gets
-        # accumulated just gets discarded with `_`.
-        value, _ = DynamicPPL.tilde_assume(
-            child_context, right, vn, get_global_varinfo(context)
-        )
-        value, vi
+        # This branch means that a different sampler is supposed to handle this
+        # variable. From the perspective of this sampler, this variable is
+        # conditioned on, so we can just treat it as an observation.
+        # The only catch is that the value that we need is to be obtained from
+        # the global VarInfo (since the local VarInfo has no knowledge of it).
+        # Note that tilde_observe!! will trigger resampling in particle methods
+        # for variables that are handled by other Gibbs component samplers.
+        val = get_conditioned_gibbs(context, vn)
+        DynamicPPL.tilde_observe!!(child_context, right, val, vn, vi)
     else
         # If the varname has not been conditioned on, nor is it a target variable, its
         # presumably a new variable that should be sampled from its prior. We need to add
@@ -210,13 +212,25 @@ function DynamicPPL.tilde_assume(
     vn, child_context = DynamicPPL.prefix_and_strip_contexts(child_context, vn)
 
     return if is_target_varname(context, vn)
+        # This branch means that that `sampler` is supposed to handle
+        # this variable. We can thus use its default behaviour, with
+        # the 'local' sampler-specific VarInfo.
         DynamicPPL.tilde_assume(rng, child_context, sampler, right, vn, vi)
     elseif has_conditioned_gibbs(context, vn)
-        value, _ = DynamicPPL.tilde_assume(
-            child_context, right, vn, get_global_varinfo(context)
-        )
-        value, vi
+        # This branch means that a different sampler is supposed to handle this
+        # variable. From the perspective of this sampler, this variable is
+        # conditioned on, so we can just treat it as an observation.
+        # The only catch is that the value that we need is to be obtained from
+        # the global VarInfo (since the local VarInfo has no knowledge of it).
+        # Note that tilde_observe!! will trigger resampling in particle methods
+        # for variables that are handled by other Gibbs component samplers.
+        val = get_conditioned_gibbs(context, vn)
+        DynamicPPL.tilde_observe!!(child_context, right, val, vn, vi)
     else
+        # If the varname has not been conditioned on, nor is it a target variable, its
+        # presumably a new variable that should be sampled from its prior. We need to add
+        # this new variable to the global `varinfo` of the context, but not to the local one
+        # being used by the current sampler.
         value, new_global_vi = DynamicPPL.tilde_assume(
             rng,
             child_context,

src/mcmc/hmc.jl

@@ -162,7 +162,9 @@ function find_initial_params(
         # Resample and try again.
         # NOTE: varinfo has to be linked to make sure this samples in unconstrained space
         varinfo = last(
-            DynamicPPL.evaluate!!(model, rng, varinfo, DynamicPPL.SampleFromUniform())
+            DynamicPPL.evaluate_and_sample!!(
+                rng, model, varinfo, DynamicPPL.SampleFromUniform()
+            ),
         )
     end