remove need for rvs for internal random_ops (blei-lab#829)

edward/inferences/hmc.py

@@ -10,11 +10,6 @@
 from edward.models import RandomVariable
 from edward.util import copy
 
-try:
-  from edward.models import Normal, Uniform
-except Exception as e:
-  raise ImportError("{0}. Your TensorFlow version is not supported.".format(e))
-
 
 class HMC(MonteCarlo):
   """Hamiltonian Monte Carlo, also known as hybrid Monte Carlo
@@ -93,9 +88,7 @@ def build_update(self):
     old_r_sample = OrderedDict()
     for z, qz in six.iteritems(self.latent_vars_unconstrained):
       event_shape = qz.event_shape
-      normal = Normal(loc=tf.zeros(event_shape, dtype=qz.dtype),
-                      scale=tf.ones(event_shape, dtype=qz.dtype))
-      old_r_sample[z] = normal.sample()
+      old_r_sample[z] = tf.random_normal(event_shape, dtype=qz.dtype)
 
     # Simulate Hamiltonian dynamics.
     new_sample, new_r_sample = leapfrog(old_sample, old_r_sample,
@@ -112,8 +105,7 @@ def build_update(self):
     ratio -= self._log_joint_unconstrained(old_sample)
 
     # Accept or reject sample.
-    u = Uniform(low=tf.constant(0.0, dtype=ratio.dtype),
-                high=tf.constant(1.0, dtype=ratio.dtype)).sample()
+    u = tf.random_uniform([], dtype=ratio.dtype)
     accept = tf.log(u) < ratio
     sample_values = tf.cond(accept, lambda: list(six.itervalues(new_sample)),
                             lambda: list(six.itervalues(old_sample)))

edward/inferences/metropolis_hastings.py

@@ -10,11 +10,6 @@
 from edward.models import RandomVariable
 from edward.util import check_latent_vars, copy
 
-try:
-  from edward.models import Uniform
-except Exception as e:
-  raise ImportError("{0}. Your TensorFlow version is not supported.".format(e))
-
 
 class MetropolisHastings(MonteCarlo):
   """Metropolis-Hastings [@metropolis1953equation; @hastings1970monte].
@@ -139,8 +134,7 @@ def build_update(self):
         ratio -= tf.reduce_sum(x_zold.log_prob(dict_swap[x]))
 
     # Accept or reject sample.
-    u = Uniform(low=tf.constant(0.0, dtype=ratio.dtype),
-                high=tf.constant(1.0, dtype=ratio.dtype)).sample()
+    u = tf.random_uniform([], dtype=ratio.dtype)
     accept = tf.log(u) < ratio
     sample_values = tf.cond(accept, lambda: list(six.itervalues(new_sample)),
                             lambda: list(six.itervalues(old_sample)))

edward/inferences/sghmc.py

@@ -9,11 +9,6 @@
 from edward.models import RandomVariable, Empirical
 from edward.util import copy
 
-try:
-  from edward.models import Normal
-except Exception as e:
-  raise ImportError("{0}. Your TensorFlow version is not supported.".format(e))
-
 
 class SGHMC(MonteCarlo):
   """Stochastic gradient Hamiltonian Monte Carlo [@chen2014stochastic].
@@ -82,14 +77,12 @@ def build_update(self):
     for z, grad_log_p in zip(six.iterkeys(old_sample), grad_log_joint):
       qz = self.latent_vars[z]
       event_shape = qz.event_shape
-      normal = Normal(
-          loc=tf.zeros(event_shape, dtype=qz.dtype),
-          scale=(tf.sqrt(tf.cast(learning_rate * self.friction, qz.dtype)) *
-                 tf.ones(event_shape, dtype=qz.dtype)))
+      stddev = tf.sqrt(tf.cast(learning_rate * self.friction, qz.dtype))
+      normal = tf.random_normal(event_shape, dtype=qz.dtype)
       sample[z] = old_sample[z] + old_v_sample[z]
       v_sample[z] = ((1.0 - 0.5 * self.friction) * old_v_sample[z] +
                      learning_rate * tf.convert_to_tensor(grad_log_p) +
-                     normal.sample())
+                     stddev * normal)
 
     # Update Empirical random variables.
     assign_ops = []

edward/inferences/sgld.py

@@ -9,11 +9,6 @@
 from edward.models import RandomVariable
 from edward.util import copy
 
-try:
-  from edward.models import Normal
-except Exception as e:
-  raise ImportError("{0}. Your TensorFlow version is not supported.".format(e))
-
 
 class SGLD(MonteCarlo):
   """Stochastic gradient Langevin dynamics [@welling2011bayesian].
@@ -75,13 +70,11 @@ def build_update(self):
     for z, grad_log_p in zip(six.iterkeys(old_sample), grad_log_joint):
       qz = self.latent_vars[z]
       event_shape = qz.event_shape
-      normal = Normal(
-          loc=tf.zeros(event_shape, dtype=qz.dtype),
-          scale=(tf.sqrt(tf.cast(learning_rate, qz.dtype)) *
-                 tf.ones(event_shape, dtype=qz.dtype)))
-      sample[z] = old_sample[z] + \
-          0.5 * learning_rate * tf.convert_to_tensor(grad_log_p) + \
-          normal.sample()
+      stddev = tf.sqrt(tf.cast(learning_rate, qz.dtype))
+      normal = tf.random_normal(event_shape, dtype=qz.dtype)
+      sample[z] = (old_sample[z] +
+                   0.5 * learning_rate * tf.convert_to_tensor(grad_log_p) +
+                   stddev * normal)
 
     # Update Empirical random variables.
     assign_ops = []

edward/inferences/wgan_inference.py

@@ -8,11 +8,6 @@
 from edward.inferences.gan_inference import GANInference
 from edward.util import get_session
 
-try:
-  from edward.models import Uniform
-except Exception as e:
-  raise ImportError("{0}. Your TensorFlow version is not supported.".format(e))
-
 
 class WGANInference(GANInference):
   """Parameter estimation with GAN-style training
@@ -81,9 +76,7 @@ def build_loss_and_gradients(self, var_list):
     if self.penalty is None or self.penalty == 0:
       penalty = 0.0
     else:
-      eps = Uniform().sample(x_true.shape[0])
-      while eps.shape.ndims < x_true.shape.ndims:
-        eps = tf.expand_dims(eps, -1)
+      eps = tf.random_uniform(tf.shape(x_true))
       x_interpolated = eps * x_true + (1.0 - eps) * x_fake
       with tf.variable_scope("Disc", reuse=True):
         d_interpolated = self.discriminator(x_interpolated)