use variable._ref(); fix empirical.py, test_beta_log_prob.py

edward/inferences/klpq.py

@@ -143,6 +143,6 @@ def build_loss_and_gradients(self, var_list):
     loss = tf.reduce_mean(w_norm * log_w)
     grads = tf.gradients(
         -tf.reduce_mean(q_log_prob * tf.stop_gradient(w_norm)),
-        var_list)
+        [v._ref() for v in var_list])
     grads_and_vars = list(zip(grads, var_list))
     return loss, grads_and_vars

edward/inferences/klqp.py

@@ -366,7 +366,7 @@ def build_reparam_loss_and_gradients(inference, var_list):
   if var_list is None:
     var_list = tf.trainable_variables()
 
-  grads = tf.gradients(loss, [v.ref() for v in var_list])
+  grads = tf.gradients(loss, [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars
 
@@ -439,7 +439,7 @@ def build_reparam_kl_loss_and_gradients(inference, var_list):
   if var_list is None:
     var_list = tf.trainable_variables()
 
-  grads = tf.gradients(loss, [v.ref() for v in var_list])
+  grads = tf.gradients(loss, [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars
 
@@ -511,7 +511,7 @@ def build_reparam_entropy_loss_and_gradients(inference, var_list):
   if var_list is None:
     var_list = tf.trainable_variables()
 
-  grads = tf.gradients(loss, [v.ref() for v in var_list])
+  grads = tf.gradients(loss, [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars
 
@@ -582,7 +582,7 @@ def build_score_loss_and_gradients(inference, var_list):
 
   grads = tf.gradients(
       -tf.reduce_mean(q_log_prob * tf.stop_gradient(losses)),
-      var_list)
+      [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars
 
@@ -657,7 +657,7 @@ def build_score_kl_loss_and_gradients(inference, var_list):
   loss = -(tf.reduce_mean(p_log_lik) - kl)
   grads = tf.gradients(
       -(tf.reduce_mean(q_log_prob * tf.stop_gradient(p_log_lik)) - kl),
-      var_list)
+      [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars
 
@@ -732,6 +732,6 @@ def build_score_entropy_loss_and_gradients(inference, var_list):
   grads = tf.gradients(
       -(tf.reduce_mean(q_log_prob * tf.stop_gradient(p_log_prob)) +
           q_entropy),
-      var_list)
+      [v._ref() for v in var_list])
   grads_and_vars = list(zip(grads, var_list))
   return loss, grads_and_vars

edward/inferences/map.py

@@ -147,7 +147,7 @@ def build_loss_and_gradients(self, var_list):
     if var_list is None:
       var_list = tf.trainable_variables()
 
-    grads = tf.gradients(loss, [v.ref() for v in var_list])
+    grads = tf.gradients(loss, [v._ref() for v in var_list])
     grads_and_vars = list(zip(grads, var_list))
     return loss, grads_and_vars
 

edward/models/empirical.py

@@ -78,12 +78,12 @@ def _std(self):
   def _variance(self):
     return math_ops.square(self.std())
 
-  def sample_n(self, n, seed=None):
+  def _sample_n(self, n, seed=None):
     if self.n != 1:
       logits = logit(tf.ones(self.n, dtype=tf.float32) /
                      tf.cast(self.n, dtype=tf.float32))
       cat = tf.contrib.distributions.Categorical(logits=logits)
-      indices = cat.sample_n(n, seed)
+      indices = cat._sample_n(n, seed)
       return tf.gather(self._params, indices)
     else:
       multiples = tf.concat(

examples/iwvi.py

@@ -79,7 +79,7 @@ def build_loss_and_gradients(self, var_list):
       log_w += [p_log_prob - q_log_prob]
 
     loss = -log_mean_exp(log_w)
-    grads = tf.gradients(loss, [v.ref() for v in var_list])
+    grads = tf.gradients(loss, [v._ref() for v in var_list])
     grads_and_vars = list(zip(grads, var_list))
     return loss, grads_and_vars
 

tests/test-models/test_beta_log_prob.py

@@ -18,7 +18,7 @@ def _test(a, b, n):
   a = a.eval()
   b = b.eval()
   assert np.allclose(rv.log_prob(x_tf).eval(),
-                     stats.beta.logpdf(x, a, b))
+                     stats.beta.logpdf(x, a, b), atol=1e-3)
 
 
 class test_beta_log_prob_class(tf.test.TestCase):