Merge c439885 into a7902a3

Project.toml

@@ -3,13 +3,15 @@ uuid = "7a57a42e-76ec-4ea3-a279-07e840d6d9cf"
 keywords = ["probablistic programming"]
 license = "MIT"
 desc = "Common interfaces for probabilistic programming"
-version = "0.3.1"
+version = "0.4"
 
 [deps]
 AbstractMCMC = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
+DensityInterface = "b429d917-457f-4dbc-8f4c-0cc954292b1d"
 Setfield = "efcf1570-3423-57d1-acb7-fd33fddbac46"
 
 [compat]
 AbstractMCMC = "2, 3"
+DensityInterface = "0.3"
 Setfield = "0.7.1, 0.8"
 julia = "1"

interface.md

@@ -24,12 +24,14 @@ There are three interrelating aspects that this interface intends to standardize
 - Sampling
 - “Conversions” between different conditionings of models
 
-Therefore, the interface consists of:
+Therefore, the interface consists of an `AbstractProbabilisticProgram` supertype, together with
+functions
 
 - `condition(::Model, ::Trace) -> ConditionedModel`
 - `decondition(::ConditionedModel) -> GenerativeModel`
 - `sample(::Model, ::Sampler = Exact(), [Int])` (from `AbstractMCMC.sample`)
-- `logdensity(::Model, ::Trace)`
+- `logdensityof(::Model, ::Trace)` and `density(::Model, ::Trace)` (from
+  [DensityInterface.jl](https://github.com/JuliaMath/DensityInterface.jl))
 
 
 ### Traces & probability expressions
@@ -198,32 +200,45 @@ Not all variants need to be supported – for example, a posterior model might n
 model.
 
 
-### Density Calculation
+### Density Evaluation
 
 Since the different “versions” of how a model is to be understood as generative or conditioned are
 to be expressed in the type or dispatch they support, there should be no need for separate functions
-`logjoint`, `loglikelihood`, etc., which force these semantic distinctions on the implementor; one
-`logdensity` should suffice for all, with the distinction being made by the capabilities of the
-concrete model instance.
+`logjoint`, `loglikelihood`, etc., which force these semantic distinctions on the implementor; we
+therefore adapt the interface of
+[DensityInterface.jl](https://github.com/JuliaMath/DensityInterface.jl).  Its main function
+`logdensityof` should suffice for variants, with the distinction being made by the capabilities of
+the concrete model instance.
 
-Note that this generalizes `logpdf`, too, since the posterior density will of course in general be
-unnormalized and hence not a probability density.
+ DensityInterface.jl also requires the trait function `hasdensity`, which is set to `true` for the
+`AbstractProbabilisticProgram` type.  Additional functions
 
-The evaluation will usually work with the internal, concrete trace type, like `VarInfo` in Turing.jl:
+```
+DensityInterface.densityof(d, x) = exp(logdensityof(d, x))
+DensityInterface.logdensity(d) = Base.Fix1(logdensity, d)
+```
+
+are provided automatically. 
+
+Note that `logdensityof` strictly generalizes `logpdf`, since the posterior density will of course
+in general be unnormalized and hence not a probability density.
+
+The evaluation will usually work with the internal, concrete trace type, like `VarInfo` in
+Turing.jl:
 
 ```julia
-logdensity(m, vi)
+logdensityof(m, vi)
 ```
 
 But the user will more likely work on the interface using probability expressions:
 
 ```julia
-logdensity(m, @T(X = ...))
+logdensityof(m, @T(X = …))
 ```
 
 (Note that this would replace the current `prob` string macro in Turing.jl.)
 
-Densities need not be normalized.
+Densities need (and usually, will) not be normalized.
 
 
 #### Implementation notes 
@@ -232,15 +247,15 @@ It should be able to make this fall back on the internal method with the right d
 implementation of `maketrace`:
 
 ```julia
-logdensity(m, t::ProbabilityExpression) = logdensity(m, maketrace(m, t))
+logdensityof(m, t::ProbabilityExpression) = logdensityof(m, maketrace(m, t))
 ```
 
 There is one open question – should normalized and unnormalized densities be able to be
 distinguished?  This could be done by dispatch as well, e.g., if the caller wants to make sure
 normalization:
 
 ```
-logdensity(g, @T(X = ..., Y = ..., Z = ...); normalized=Val{true})
+logdensityof(g, @T(X = …, Y = …, Z = …); normalized=Val{true})
 ```
 
 Although there is proably a better way through traits; maybe like for arrays, with

src/AbstractPPL.jl

@@ -5,7 +5,7 @@ export VarName, getsym, getindexing, getlens, inspace, subsumes, varname, vsym,
 
 
 # Abstract model functions
-export AbstractProbabilisticProgram, condition, decondition, logdensity
+export AbstractProbabilisticProgram, condition, decondition, logdensityof, densityof
 
 
 # Abstract traces

src/abstractprobprog.jl

@@ -1,4 +1,5 @@
 using AbstractMCMC
+using DensityInterface
 
 
 """
@@ -8,19 +9,21 @@ Common base type for models expressed as probabilistic programs.
 """
 abstract type AbstractProbabilisticProgram <: AbstractMCMC.AbstractModel end
 
+DensityInterface.hasdensity(::AbstractProbabilisticProgram) = true
+
 
 """
-    logdensity(model, trace)
+    logdensityof(model, trace)
 
 Evaluate the (possibly unnormalized) density of the model specified by the probabilistic program
 in `model`, at specific values for the random variables given through `trace`.
 
 `trace` can be of any supported internal trace type, or a fixed probability expression.
 
-`logdensity` should interact with conditioning and deconditioning in the way required by probability
-theory.
+`logdensityof` should interact with conditioning and deconditioning in the way required by
+probability theory.
 """
-function logdensity end
+DensityInterface.logdensityof
 
 
 """