Resolve #430 (Add normalisation flag to constrained prior)

This MR resolves issue #430 (closed) to properly normalize ConstrainedPriorDict.prob() when sampling the prior. This MR allows for users to input a normalisation flag when creating an instance of a Constraint prior and when calling ConstrainedPriorDict.prob() if there are normalisation constants given will return the normalisation*prob and otherwise use the ratio of total samples to kept samples as this normalisation constant. Pass normalisation=1 when creating a Constraint Prior for behaviour before this MR.

bilby/core/prior/dict.py

@@ -3,7 +3,6 @@ from io import open as ioopen
 import json
 import numpy as np
 import os
-from functools import lru_cache
 
 from future.utils import iteritems
 from matplotlib.cbook import flatten
@@ -42,6 +41,7 @@ class PriorDict(dict):
             self.from_file(filename)
         elif dictionary is not None:
             raise ValueError("PriorDict input dictionary not understood")
+        self._cached_normalizations= {}
 
         self.convert_floats_to_delta_functions()
 
@@ -379,18 +379,27 @@ class PriorDict(dict):
                            if not isinstance(self[key], Constraint)}
             return all_samples
 
-    @lru_cache()
     def normalize_constraint_factor(self, keys):
-        min_accept = 50
-        sampling_chunk = 250
-        samples = self.sample_subset(keys=keys, size=sampling_chunk)
-        keep = np.array(self.evaluate_constraints(samples))
-        while np.count_nonzero(keep) < min_accept:
-            new_samples = self.sample_subset(keys=keys, size=sampling_chunk)
-            for key in samples:
-                samples[key] = np.concatenate(samples[key], new_samples[key])
-            keep = np.array(self.evaluate_constraints(samples))
-        return len(keep) / np.count_nonzero(keep)
+        if tuple(keys) in self._cached_normalizations.keys():
+            return self._cached_normalizations[tuple(keys)]
+        else:
+            min_accept = 1000
+            sampling_chunk = 5000
+            samples = self.sample_subset(keys=keys, size=sampling_chunk)
+            keep = np.atleast_1d(self.evaluate_constraints(samples))
+            if len(keep) == 1:
+                return 1
+            all_samples = {key: np.array([]) for key in keys}
+            _first_key = list(all_samples.keys())[0]
+            while np.count_nonzero(keep) < min_accept:
+                samples = self.sample_subset(keys=keys, size=sampling_chunk)
+                for key in samples:
+                    all_samples[key] = np.hstack(
+                        [all_samples[key], samples[key].flatten()])
+                keep = np.array(self.evaluate_constraints(all_samples), dtype=bool)
+            factor = len(keep) / np.count_nonzero(keep)
+            self._cached_normalizations[tuple(keys)] = factor
+            return factor
 
     def prob(self, sample, **kwargs):
         """
@@ -410,14 +419,7 @@ class PriorDict(dict):
         prob = np.product([self[key].prob(sample[key])
                            for key in sample], **kwargs)
 
-        ratio = 1
-        outsample = self.conversion_function(sample)
-        # Check if there is a constraint in sample/outsample
-        if (np.any(isinstance([self[key] for key in sample.keys()], Constraint)) or
-                np.any(isinstance([self[key] for key in outsample.keys()], Constraint))):
-            # If constraint exists in keys, caclulate the cached normalization constant
-            ratio = self.normalize_constraint_factor(sample.keys())
-
+        ratio = self.normalize_constraint_factor(tuple(sample.keys()))
         if np.all(prob == 0.):
             return prob
         else:
@@ -451,14 +453,7 @@ class PriorDict(dict):
         ln_prob = np.sum([self[key].ln_prob(sample[key])
                           for key in sample], axis=axis)
 
-        ratio = 1
-        outsample = self.conversion_function(sample)
-        # Check if there is a constraint in sample/outsample
-        if (np.any(isinstance([self[key] for key in sample.keys()], Constraint)) or
-                np.any(isinstance([self[key] for key in outsample.keys()], Constraint))):
-            # If constraint exists in keys, caclulate the cached normalization constant
-            ratio = self.normalize_constraint_factor(sample.keys())
-
+        ratio = self.normalize_constraint_factor(tuple(sample.keys()))
         if np.all(np.isinf(ln_prob)):
             return ln_prob
         else: