add run!

src/ApproxInferenceBase.jl

@@ -1,5 +1,49 @@
 module ApproxInferenceBase
-    using Distributions
-    using Random
-    include("priors.jl")
+using Distributions
+using Random
+export run!
+include("priors.jl")
+
+"""
+    run!(method, model, data;
+         verbosity = 0, callback = () -> nothing, rng = Random.GLOBAL_RNG)
+
+Run approximate inference `method` on `model` and `data`.
+The `model` should be a callable object (function or functor) with one argument
+and return something that can be compared to the `data`. The comparison metric is
+defined in the `method`.
+Handling of constants and extraction of summary statistics should be done in
+the `model` (see examples below).
+Verbosity levels are `verbosity = 0` (silent), `verbosity = 1` (progress),
+`verbosity = 2` (detailed).
+Callbacks `callback` are callable objects with no argument that are called after
+every iteration of an iterative `method`. Custom random number generators can be
+given through the argument `rng`.
+
+# Model examples
+```
+# simple model
+model(params) = sum(params)
+
+# complex model with constants
+complex_model(params, constants) = sum(params) + sum(constants)
+model(params) = let constants = [1, 2, 3]
+        complex_model(params, constants)
+    end
+
+# extracting summary statistics
+raw_model(params) = rand(4, 3)
+model(params) = mean(raw_model(params), dims = 2)
+
+# functor
+struct Model
+    options
+end
+(m::Model)(params) = sum(params) + sum(m.options)
+```
+"""
+function run!(method, model, data;
+              verbosity = 0, callback = () -> nothing, rng = Random.GLOBAL_RNG)
+    throw(MethodError(run!, (method, model, data)))
+end
 end # module