feature/laplace dynamic delta

autofit/graphical/expectation_propagation/factor_optimiser.py

@@ -16,11 +16,12 @@ class AbstractFactorOptimiser(ABC):
     """
     logger = logger.debug
 
-    def __init__(self, initial_values=None, deltas=None, inplace=False, delta=1):
+    def __init__(self, initial_values=None, deltas=None, inplace=False, delta=1, dynamic_delta=False):
         self.initial_values = initial_values or {}
         self.inplace = inplace
         self.delta = delta
         self.deltas = deltas or {}
+        self.dynamic_delta = dynamic_delta
 
     def update_model_approx(
             self,
@@ -30,14 +31,26 @@ def update_model_approx(
             status: Optional[Status] = Status(),
             delta: Optional[float] = None,
     ) -> Tuple[EPMeanField, Status]:
-        delta = delta or self.deltas.get(factor_approx.factor) or self.delta
-        new_approx, status = model_approx.project_mean_field(
+
+        variable_message_count = model_approx.variable_message_count
+        min_value = min(variable_message_count.values())
+
+        delta = delta or self.delta
+
+        if factor_approx.factor in self.deltas:
+            delta = self.deltas[factor_approx.factor]
+        elif self.dynamic_delta:
+            delta = MeanField({
+                variable: self.delta * (min_value / message_count)
+                for variable, message_count in variable_message_count.items()
+            })
+
+        return model_approx.project_mean_field(
             new_model_dist,
             factor_approx,
             delta=delta,
             status=status,
         )
-        return new_approx, status
 
     @abstractmethod
     def optimise(

autofit/non_linear/abstract_search.py

@@ -101,7 +101,7 @@ def __init__(
         session
             An SQLAlchemy session instance so the results of the model-fit are written to an SQLite database.
         """
-        super().__init__(delta=kwargs.get("delta", 1.0))
+        super().__init__(delta=kwargs.get("delta", 1.0), dynamic_delta=kwargs.get("dynamic_delta", True))
 
         from autofit.non_linear.paths.database import DatabasePaths
 
@@ -204,15 +204,13 @@ def __init__(
 
         self.optimisation_counter = Counter()
 
-        self.dynamic_delta = kwargs.get("dynamic_delta", True)
-
     __identifier_fields__ = tuple()
 
     def optimise(
             self,
             factor: Factor,
             model_approx: EPMeanField,
-            status: Optional[Status] = None
+            status: Status = Status(),
     ) -> Tuple[EPMeanField, Status]:
         """
         Perform optimisation for expectation propagation. Currently only
@@ -289,25 +287,12 @@ def optimise(
             result.projected_model.priors
         )
 
-        variable_message_count = model_approx.variable_message_count
-        min_value = min(variable_message_count.values())
-
-        if self.dynamic_delta:
-            delta = MeanField({
-                variable: self.delta * (min_value / message_count)
-                for variable, message_count in variable_message_count.items()
-            })
-        else:
-            delta = self.delta
-
-        projection, status = factor_approx.project(
-            new_model_dist,
-            delta=delta
+        model_approx, status = self.update_model_approx(
+            new_model_dist, factor_approx, model_approx, status
         )
-
         status.result = result
 
-        return model_approx.project(projection, status)
+        return model_approx, status
 
     @property
     def name(self):

test_autofit/graphical/regression/test_static.py

@@ -11,6 +11,7 @@ def __init__(self):
         self._paths = af.DirectoryPaths()
         self.delta = 1.0
         self.dynamic_delta = False
+        self.deltas = {}
 
     def fit(
             self,