Make compatible with TensorFlow 0.12

* `ones_initializer` is now callable.
* Replaced `tf.initialize_all_variables` with `tf.global_variables_initializer`.
* Replaced `tf.initialize_variables` with `tf.variables_initializer`.

README.md

@@ -1,5 +1,8 @@
 # [Learning to Learn](https://arxiv.org/abs/1606.04474) in TensorFlow
 
+Compatible with TensorFlow 0.12.
+
+
 ## Training
 
 ```

convergence_test.py

@@ -53,7 +53,7 @@ def testSimple(self):
  minimize_ops = optimizer.meta_minimize(problem, 20, learning_rate=1e-2)
  # L2L should solve the simple problem is less than 500 epochs.
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  cost, _ = train(sess, minimize_ops, 500, 5)
  self.assertLess(cost, 1e-5)
 

evaluate.py

@@ -56,10 +56,10 @@ def main(_):
  if FLAGS.optimizer == "Adam":
  cost_op = problem()
  problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
- problem_reset = tf.initialize_variables(problem_vars)
+ problem_reset = tf.variables_initializer(problem_vars)
 
  optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
- optimizer_reset = tf.initialize_variables(optimizer.get_slot_names())
+ optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
  update = optimizer.minimize(cost_op)
  reset = [problem_reset, optimizer_reset]
  elif FLAGS.optimizer == "L2L":

meta.py

@@ -367,7 +367,7 @@ def time_step(t, fx_array, x, state):
  variables = (nest.flatten(state) +
  x + constants)
  # Empty array as part of the reset process.
- reset = [tf.initialize_variables(variables), fx_array.close()]
+ reset = [tf.variables_initializer(variables), fx_array.close()]
 
  # Operator to update the parameters and the RNN state after our loop, but
  # during an epoch.

meta_test.py

@@ -57,7 +57,7 @@ def testResults(self):
  }))
  minimize_ops = optimizer.meta_minimize(problem, 5)
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  cost, final_x = train(sess, minimize_ops, 1, 2)
 
  # Torch results
@@ -121,7 +121,7 @@ def testMultiOptimizer(self, net_assignments, net_config):
  minimize_ops = optimizer.meta_minimize(problem, 3,
  net_assignments=net_assignments)
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  train(sess, minimize_ops, 1, 2)
 
  def testSecondDerivatives(self):
@@ -133,7 +133,7 @@ def testSecondDerivatives(self):
  minimize_ops = optimizer.meta_minimize(problem, 3,
  second_derivatives=True)
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  train(sess, minimize_ops, 1, 2)
 
  def testConvolutional(self):
@@ -162,13 +162,13 @@ def convolutional_problem():
  net_assignments=[("conv", ["conv/w"])]
  )
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  train(sess, minimize_ops, 1, 2)
 
  def testWhileLoopProblem(self):
  """Tests L2L applied to problem with while loop."""
  def while_loop_problem():
- x = tf.get_variable("x", shape=[], initializer=tf.ones_initializer)
+ x = tf.get_variable("x", shape=[], initializer=tf.ones_initializer())
 
  # Strange way of squaring the variable.
  _, x_squared = tf.while_loop(
@@ -183,7 +183,7 @@ def while_loop_problem():
  net_options={"layers": ()}))
  minimize_ops = optimizer.meta_minimize(while_loop_problem, 3)
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  train(sess, minimize_ops, 1, 2)
 
  def testSaveAndLoad(self):
@@ -203,7 +203,7 @@ def testSaveAndLoad(self):
  minimize_ops = optimizer.meta_minimize(problem, 3)
 
  with self.test_session(graph=g1) as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  train(sess, minimize_ops, 1, 2)
 
  # Save optimizer.
@@ -222,7 +222,7 @@ def testSaveAndLoad(self):
  minimize_ops = optimizer.meta_minimize(problem, 3)
 
  with self.test_session(graph=g2) as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  cost_loaded, x_loaded = train(sess, minimize_ops, num_unrolls, num_epochs)
 
  # The last cost should be the same.

networks_test.py

@@ -64,7 +64,7 @@ def testResults(self, initializer):
  update, _ = net(gradients, state)
 
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  update_np = sess.run(update)
  self.assertAllEqual(update_np, np.zeros(shape))
 
@@ -111,7 +111,7 @@ def testResults(self, initializer):
  update, _ = net(gradients, state)
 
  with self.test_session() as sess:
- sess.run(tf.initialize_all_variables())
+ sess.run(tf.global_variables_initializer())
  update_np = sess.run(update)
  self.assertAllEqual(update_np, np.zeros(shape))
 

nn/batch_norm.py

@@ -129,7 +129,7 @@ def _set_default_initializer(self, var_name):
  """
  if var_name not in self._initializers:
  if var_name == self.GAMMA:
- self._initializers[self.GAMMA] = tf.ones_initializer
+ self._initializers[self.GAMMA] = tf.ones_initializer()
  elif var_name == self.BETA:
  self._initializers[self.BETA] = tf.zeros_initializer
 
@@ -160,7 +160,7 @@ def _build_statistics_variance(self, input_batch,
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
  tf.GraphKeys.VARIABLES],
- initializer=tf.ones_initializer,
+ initializer=tf.ones_initializer(),
  trainable=False)
 
  def build_batch_stats():
@@ -226,7 +226,7 @@ def _build_statistics_second_moment(self, input_batch,
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
  tf.GraphKeys.VARIABLES],
- initializer=tf.ones_initializer,
+ initializer=tf.ones_initializer(),
  trainable=False)
 
  self._moving_variance = tf.sub(self._moving_second_moment,

problems.py

@@ -46,7 +46,7 @@ def build():
  "x",
  shape=[],
  dtype=tf.float32,
- initializer=tf.ones_initializer)
+ initializer=tf.ones_initializer())
  return tf.square(x, name="x_squared")
 
  return build
@@ -59,7 +59,7 @@ def get_coordinate(i):
  return tf.get_variable("x_{}".format(i),
  shape=[],
  dtype=tf.float32,
- initializer=tf.ones_initializer)
+ initializer=tf.ones_initializer())
 
  def build():
  coordinates = [get_coordinate(i) for i in xrange(num_dims)]