Fixed the epsilon selectors so they correctly handle the rounds.

sciope/inference/smc_abc.py

@@ -181,10 +181,10 @@ def infer(self, num_samples, batch_size,
         abc_history.append(abc_results)
 
         # SMC iterations
-        round = 1
+        abc_round = 2 # Indicates the round getting ready to start
         while not terminate:
 
-            tol, relative, terminate = eps_selector.get_epsilon(round, abc_history)
+            tol, relative, terminate = eps_selector.get_epsilon(abc_round, abc_history)
 
             print("Starting epsilon = {}".format(tol))
             if self.use_logger:
@@ -228,7 +228,7 @@ def infer(self, num_samples, batch_size,
                 if self.parameters is not None:
                     abc_results = InferenceRound.build_from_inference_round(abc_results, list(self.parameters.keys()))
                 abc_history.append(abc_results)
-                round += 1
+                abc_round += 1
 
             except KeyboardInterrupt:
                 if self.parameters is None:

sciope/utilities/epsilonselectors/absolute_epsilon_selector.py

@@ -34,7 +34,7 @@ def __init__(self, epsilon_sequence):
 
         assert (len(epsilon_sequence) > 0)
         self.epsilon_sequence = epsilon_sequence
-        self.last_round = len(self.epsilon_sequence) - 1
+        self.last_round = len(self.epsilon_sequence)
 
     def get_initial_epsilon(self):
         """Gets the first epsilon in the sequence.
@@ -48,14 +48,14 @@ def get_initial_epsilon(self):
         has_more : bool
             Whether there are more epsilons after this one
         """
-        return self.epsilon_sequence[0], False, len(self.epsilon_sequence) == 1
+        return self.epsilon_sequence[0], False, self.last_round == 1
 
-    def get_epsilon(self, round, abc_history):
+    def get_epsilon(self, abc_round, abc_history):
         """Returns the n-th epsilon in the seqeunce.
 
         Parameters
         ----------
-        round : int
+        abc_round : int
             the round to get the epsilon for
         abc_history : type
             A list of dictionaries with keys `accepted_samples` and `distances`
@@ -70,4 +70,4 @@ def get_epsilon(self, round, abc_history):
         terminate : bool
             Whether to stop after this epsilon
         """
-        return self.epsilon_sequence[round], False, round == self.last_round
+        return self.epsilon_sequence[abc_round - 1], False, abc_round >= self.last_round

sciope/utilities/epsilonselectors/relative_epsilon_selector.py

@@ -35,6 +35,8 @@ def __init__(self, epsilon_percentile, max_rounds=None):
         max_rounds : int
             The maximum number of rounds before stopping. If None, doesn't end.
         """
+        if max_rounds == 0:
+            raise ValueError("max_rounds must be greater than 0.")
 
         self.epsilon_percentile = epsilon_percentile
         self.max_rounds = max_rounds
@@ -52,14 +54,14 @@ def get_initial_epsilon(self):
         has_more : bool
             Whether there are more epsilons after this one
         """
-        return self.epsilon_percentile, True, self.max_rounds == 0
+        return self.epsilon_percentile, True, self.max_rounds == 1
 
-    def get_epsilon(self, round, abc_history):
+    def get_epsilon(self, abc_round, abc_history):
         """Returns the new epsilon based on the n-th round.
 
         Parameters
         ----------
-        round : int
+        abc_round : int
             the n-th round of the sequence
         abc_history : type
             A list of dictionaries with keys `accepted_samples` and `distances`
@@ -74,8 +76,8 @@ def get_epsilon(self, round, abc_history):
         terminate : bool
             Whether to stop after this epsilon
         """
-        if round > len(abc_history):
+        if abc_round > len(abc_history):
             epsilon = np.percentile(abc_history[-1]['distances'], self.epsilon_percentile)
         else:
-            epsilon = np.percentile(abc_history[round - 1]['distances'], self.epsilon_percentile)
-        return epsilon, False, self.max_rounds and round + 1 == self.max_rounds
+            epsilon = np.percentile(abc_history[abc_round - 1]['distances'], self.epsilon_percentile)
+        return epsilon, False, self.max_rounds and abc_round >= self.max_rounds