Merge 209c32e into 071a4ba

edward/inferences/hmc.py

@@ -128,7 +128,7 @@ def build_update(self):
     # Update Empirical random variables.
     assign_ops = []
     for z_unconstrained, qz_unconstrained in six.iteritems(
-        self.latent_vars_unconstrained):
+            self.latent_vars_unconstrained):
       variable = qz_unconstrained.get_variables()[0]
       assign_ops.append(tf.scatter_update(
           variable, self.t, sample[z_unconstrained]))
@@ -139,7 +139,7 @@ def build_update(self):
 
   def _log_joint_unconstrained(self, z_sample):
     """
-    Given a sample in unconstrained latent space, transform it back into 
+    Given a sample in unconstrained latent space, transform it back into
     the original space, and compute the log joint density with appropriate
     Jacobian correction.
     """
@@ -151,17 +151,17 @@ def _log_joint_unconstrained(self, z_sample):
     z_sample_transformed = {}
     log_det_jacobian = 0.0
     for z_unconstrained, qz_unconstrained in z_sample.items():
-      z = (unconstrained_to_z[z_unconstrained] 
-           if z_unconstrained in unconstrained_to_z 
+      z = (unconstrained_to_z[z_unconstrained]
+           if z_unconstrained in unconstrained_to_z
            else z_unconstrained)
 
       try:
         bij = self.transformations[z].bijector
         z_sample_transformed[z] = bij.inverse(qz_unconstrained)
         log_det_jacobian += tf.reduce_sum(
-          bij.inverse_log_det_jacobian(qz_unconstrained))
-      except: # if z not in self.transformations, 
-              # or is not a TransformedDist w/ bijector
+            bij.inverse_log_det_jacobian(qz_unconstrained))
+      except:  # if z not in self.transformations,
+               # or is not a TransformedDist w/ bijector
         z_sample_transformed[z] = qz_unconstrained
 
     return self._log_joint(z_sample_transformed) + log_det_jacobian

edward/inferences/inference.py

@@ -15,6 +15,7 @@
 
 from tensorflow.contrib.distributions import bijectors
 
+
 @six.add_metaclass(abc.ABCMeta)
 class Inference(object):
   """Abstract base class for inference. All inference algorithms in
@@ -222,11 +223,13 @@ def initialize(self, n_iter=1000, n_print=None, scale=None,
     self.transformations = {}
     if auto_transform:
       latent_vars = self.latent_vars.copy()
-      self.latent_vars = {} # maps original latent vars to constrained Q's
-      self.latent_vars_unconstrained = {} # maps unconstrained vars to unconstrained Q's
+      # latent_vars maps original latent vars to constrained Q's.
+      # latent_vars_unconstrained maps unconstrained vars to unconstrained Q's.
+      self.latent_vars = {}
+      self.latent_vars_unconstrained = {}
       for z, qz in six.iteritems(latent_vars):
         if hasattr(z, 'support') and hasattr(qz, 'support') and \
-              z.support != qz.support and qz.support != 'point':
+                z.support != qz.support and qz.support != 'point':
 
           # transform z to an unconstrained space
           z_unconstrained = transform(z)
@@ -243,12 +246,12 @@ def initialize(self, n_iter=1000, n_print=None, scale=None,
           # back into the original constrained space
           if z_unconstrained != z:
             qz_constrained = transform(
-              qz_unconstrained, bijectors.Invert(z_unconstrained.bijector))
+                qz_unconstrained, bijectors.Invert(z_unconstrained.bijector))
 
-            try: # attempt to pushforward the params of Empirical distributions
+            try:  # attempt to pushforward the params of Empirical distributions
               qz_constrained.params = z_unconstrained.bijector.inverse(
-                qz_unconstrained.params)
-            except: # qz_unconstrained is not an Empirical distribution
+                  qz_unconstrained.params)
+            except:  # qz_unconstrained is not an Empirical distribution
               pass
 
           else:

tests/inferences/test_inference_auto_transform.py

@@ -181,7 +181,7 @@ def test_hmc_betabernoulli(self):
       qz_constrained = z_unconstrained.bijector.inverse(qz_samples)
       qz_mean, qz_var = sess.run(tf.nn.moments(qz_constrained, 0))
 
-      true_posterior = Beta(1. + np.sum(x_obs), 1. + np.sum(1-x_obs))      
+      true_posterior = Beta(1. + np.sum(x_obs), 1. + np.sum(1 - x_obs))
       pz_mean, pz_var = sess.run((true_posterior.mean(),
                                   true_posterior.variance()))
       self.assertAllClose(qz_mean, pz_mean, rtol=5e-2, atol=5e-2)
@@ -199,16 +199,19 @@ def test_klqp_betabernoulli(self):
                                 initializer=tf.random_normal(()))
       qz_std = tf.nn.softplus(tf.get_variable(name="qz_prestd",
                                               initializer=tf.random_normal(())))
-      qz_unconstrained = ed.models.Normal(loc=qz_mean, scale=qz_std, name="z_posterior")
+      qz_unconstrained = ed.models.Normal(
+          loc=qz_mean, scale=qz_std, name="z_posterior")
 
-      inference_klqp = ed.inferences.KLqp({z: qz_unconstrained}, data={xs: x_obs})
+      inference_klqp = ed.inferences.KLqp(
+          {z: qz_unconstrained}, data={xs: x_obs})
       inference_klqp.run(n_iter=500, auto_transform=True)
 
       z_unconstrained = inference_klqp.transformations[z]
-      qz_constrained = z_unconstrained.bijector.inverse(qz_unconstrained.sample(1000))
+      qz_constrained = z_unconstrained.bijector.inverse(
+          qz_unconstrained.sample(1000))
       qz_mean, qz_var = sess.run(tf.nn.moments(qz_constrained, 0))
 
-      true_posterior = Beta(np.sum(x_obs) + 1., np.sum(1-x_obs) + 1.)
+      true_posterior = Beta(np.sum(x_obs) + 1., np.sum(1 - x_obs) + 1.)
       pz_mean, pz_var = sess.run((true_posterior.mean(),
                                   true_posterior.variance()))
       self.assertAllClose(qz_mean, pz_mean, rtol=5e-2, atol=5e-2)