Merge pull request #39 from alan-turing-institute/mlda_develop

Improve stats calculation

Author: gmingas

pymc3/step_methods/metropolis.py

@@ -981,7 +981,8 @@ class MLDA(ArrayStepShared):
     stats_dtypes = [{
         'accept': np.float64,
         'accepted': np.bool,
-        'tune': np.bool
+        'tune': np.bool,
+        'base_scaling': object
     }]
 
     def __init__(self, coarse_models, vars=None, base_S=None, base_proposal_dist=None,
@@ -1102,13 +1103,6 @@ def __init__(self, coarse_models, vars=None, base_S=None, base_proposal_dist=Non
                                                  self.tune,
                                                  self.subsampling_rates[-1])
 
-        # Update stats data types dictionary given vars and base_blocked
-        if self.base_blocked or len(self.vars) == 1:
-            self.stats_dtypes[0]['base_scaling'] = np.float64
-        else:
-            for name in self.var_names:
-                self.stats_dtypes[0]['base_scaling_' + name] = np.float64
-
     def astep(self, q0):
         """One MLDA step, given current sample q0"""
         # Check if the tuning flag has been changed and if yes,
@@ -1137,11 +1131,15 @@ def astep(self, q0):
         # do not calculate likelihood, just set accept to 0.0
         if (q == q0).all():
             accept = np.float(0.0)
+            skipped_logp = True
         else:
             accept = self.delta_logp(q, q0) + self.delta_logp_next(q0, q)
+            skipped_logp = False
 
         # Accept/reject sample - next sample is stored in q_new
         q_new, accepted = metrop_select(accept, q, q0)
+        if skipped_logp:
+            accepted = False
 
         # Update acceptance counter
         self.accepted += accepted
@@ -1155,12 +1153,15 @@ def astep(self, q0):
         # Capture latest base chain scaling stats from next step method
         self.base_scaling_stats = {}
         if isinstance(self.next_step_method, CompoundStep):
+            scaling_list = []
             for method in self.next_step_method.methods:
-                self.base_scaling_stats["base_scaling_" + method.vars[0].name] = method.scaling
-        elif isinstance(self.next_step_method, Metropolis):
-            self.base_scaling_stats["base_scaling"] = self.next_step_method.scaling
+                scaling_list.append(method.scaling)
+            self.base_scaling_stats = {"base_scaling": np.array(scaling_list)}
+        elif not isinstance(self.next_step_method, MLDA):
+            # next method is any block sampler
+            self.base_scaling_stats = {"base_scaling": np.array(self.next_step_method.scaling)}
         else:
-            # next method is MLDA
+            # next method is MLDA - propagate dict from lower levels
             self.base_scaling_stats = self.next_step_method.base_scaling_stats
         stats = {**stats, **self.base_scaling_stats}
 

pymc3/tests/test_step.py

@@ -1126,10 +1126,10 @@ def test_acceptance_rate_against_coarseness(self):
             Normal("x", 5.0, 1.0)
 
         with Model() as coarse_model_1:
-            Normal("x", 5.5, 1.5)
+            Normal("x", 6.0, 2.0)
 
         with Model() as coarse_model_2:
-            Normal("x", 6.0, 2.0)
+            Normal("x", 20.0, 5.0)
 
         possible_coarse_models = [coarse_model_0,
                                   coarse_model_1,
@@ -1139,9 +1139,9 @@ def test_acceptance_rate_against_coarseness(self):
         with Model():
             Normal("x", 5.0, 1.0)
             for coarse_model in possible_coarse_models:
-                step = MLDA(coarse_models=[coarse_model], subsampling_rates=1,
-                            tune=False)
-                trace = sample(chains=1, draws=500, tune=0, step=step)
+                step = MLDA(coarse_models=[coarse_model], subsampling_rates=3,
+                            tune=True)
+                trace = sample(chains=1, draws=500, tune=100, step=step)
                 acc.append(trace.get_sampler_stats('accepted').mean())
             assert acc[0] > acc[1] > acc[2], "Acceptance rate is not " \
                                              "strictly increasing when" \
@@ -1197,10 +1197,10 @@ def test_tuning_and_scaling_on(self):
         assert trace.get_sampler_stats('tune', chains=0)[ts - 1]
         assert not trace.get_sampler_stats('tune', chains=0)[ts]
         assert not trace.get_sampler_stats('tune', chains=0)[-1]
-        assert trace.get_sampler_stats('base_scaling_x', chains=0)[0] == 100.
-        assert trace.get_sampler_stats('base_scaling_y_logodds__', chains=0)[0] == 100.
-        assert trace.get_sampler_stats('base_scaling_x', chains=0)[-1] < 100.
-        assert trace.get_sampler_stats('base_scaling_y_logodds__', chains=0)[-1] < 100.
+        assert trace.get_sampler_stats('base_scaling', chains=0)[0][0] == 100.
+        assert trace.get_sampler_stats('base_scaling', chains=0)[0][1] == 100.
+        assert trace.get_sampler_stats('base_scaling', chains=0)[-1][0] < 100.
+        assert trace.get_sampler_stats('base_scaling', chains=0)[-1][1] < 100.
 
     def test_tuning_and_scaling_off(self):
         """Test that tuning is deactivated when sample()'s tune=0 and that
@@ -1239,17 +1239,19 @@ def test_tuning_and_scaling_off(self):
 
         assert not trace_0.get_sampler_stats('tune', chains=0)[0]
         assert not trace_0.get_sampler_stats('tune', chains=0)[-1]
-        assert trace_0.get_sampler_stats('base_scaling_x', chains=0)[0] == \
-               trace_0.get_sampler_stats('base_scaling_x', chains=0)[-1] == 100.
+        assert trace_0.get_sampler_stats('base_scaling', chains=0)[0][0] == \
+               trace_0.get_sampler_stats('base_scaling', chains=0)[-1][0] == \
+               trace_0.get_sampler_stats('base_scaling', chains=0)[0][1] == \
+               trace_0.get_sampler_stats('base_scaling', chains=0)[-1][1] == 100.
 
         assert trace_1.get_sampler_stats('tune', chains=0)[0]
         assert trace_1.get_sampler_stats('tune', chains=0)[ts_1 - 1]
         assert not trace_1.get_sampler_stats('tune', chains=0)[ts_1]
         assert not trace_1.get_sampler_stats('tune', chains=0)[-1]
-        assert trace_1.get_sampler_stats('base_scaling_x', chains=0)[0] == 100.
-        assert trace_1.get_sampler_stats('base_scaling_y_logodds__', chains=0)[0] == 100.
-        assert trace_1.get_sampler_stats('base_scaling_x', chains=0)[-1] < 100.
-        assert trace_1.get_sampler_stats('base_scaling_y_logodds__', chains=0)[-1] < 100.
+        assert trace_1.get_sampler_stats('base_scaling', chains=0)[0][0] == 100.
+        assert trace_1.get_sampler_stats('base_scaling', chains=0)[0][1] == 100.
+        assert trace_1.get_sampler_stats('base_scaling', chains=0)[-1][0] < 100.
+        assert trace_1.get_sampler_stats('base_scaling', chains=0)[-1][1] < 100.
 
     def test_trace_length(self):
         """Check if trace length is as expected."""