log_prior -> log_prior_density (#1329)

gpflow/base.py

@@ -84,8 +84,8 @@ def __init__(
                 unconstrained_value, dtype=dtype, name=name, trainable=trainable
             )
 
-    def log_prior(self):
-        """ Prior probability density of the constrained variable. """
+    def log_prior_density(self):
+        """ Log of the prior probability density of the constrained variable. """
 
         if self.prior is None:
             return tf.convert_to_tensor(0.0, dtype=self.dtype)

gpflow/models/model.py

@@ -25,7 +25,7 @@
 from ..kernels import Kernel
 from ..likelihoods import Likelihood
 from ..mean_functions import MeanFunction, Zero
-from ..utilities import ops
+from ..utilities import ops, to_default_float
 
 Data = TypeVar("Data", Tuple[tf.Tensor, tf.Tensor], tf.Tensor)
 DataPoint = tf.Tensor
@@ -45,14 +45,16 @@ def neg_log_marginal_likelihood(self, *args, **kwargs) -> tf.Tensor:
         return -self.log_marginal_likelihood(*args, **kwargs)
 
     def log_marginal_likelihood(self, *args, **kwargs) -> tf.Tensor:
-        return self.log_likelihood(*args, **kwargs) + self.log_prior()
+        return self.log_likelihood(*args, **kwargs) + self.log_prior_density()
 
-    def log_prior(self) -> tf.Tensor:
-        log_priors = [p.log_prior() for p in self.trainable_parameters]
-        if log_priors:
-            return tf.add_n(log_priors)
+    def log_prior_density(self) -> tf.Tensor:
+        """
+        Sum of the log prior probability densities of all (constrained) variables in this model.
+        """
+        if self.trainable_parameters:
+            return tf.add_n([p.log_prior_density() for p in self.trainable_parameters])
         else:
-            return tf.convert_to_tensor(0.0, dtype=default_float())
+            return to_default_float(0.0)
 
     @abc.abstractmethod
     def log_likelihood(self, *args, **kwargs) -> tf.Tensor:

tests/gpflow/models/test_gpr.py

@@ -44,4 +44,4 @@ def test_non_trainable_model_objective():
     set_trainable(model, False)
 
     _ = model.log_marginal_likelihood()
-    assert model.log_prior() == 0.0
+    assert model.log_prior_density() == 0.0

tests/gpflow/test_base_prior.py

@@ -48,7 +48,7 @@ def test_log_prior_with_no_prior():
     even if it has a transform to constrain it.
     """
     param = gpflow.Parameter(5.3, transform=gpflow.utilities.positive())
-    assert param.log_prior().numpy() == 0.0
+    assert param.log_prior_density().numpy() == 0.0
 
 
 def test_log_prior_for_uniform_prior():
@@ -59,9 +59,9 @@ def test_log_prior_for_uniform_prior():
 
     uniform_prior = Uniform(low=np.float64(0), high=np.float64(100))
     param = gpflow.Parameter(1.0, transform=gpflow.utilities.positive(), prior=uniform_prior)
-    low_value = param.log_prior().numpy()
+    low_value = param.log_prior_density().numpy()
     param.assign(10.0)
-    high_value = param.log_prior().numpy()
+    high_value = param.log_prior_density().numpy()
 
     assert np.isclose(low_value, high_value)
 
@@ -78,9 +78,9 @@ def test_log_prior_on_unconstrained():
     param = gpflow.Parameter(
         initial_value, transform=Exp(), prior=uniform_prior, prior_on=PriorOn.UNCONSTRAINED,
     )
-    low_value = param.log_prior().numpy()
+    low_value = param.log_prior_density().numpy()
     param.assign(scale_factor * initial_value)
-    high_value = param.log_prior().numpy()
+    high_value = param.log_prior_density().numpy()
 
     assert np.isclose(low_value, high_value + np.log(scale_factor))