Refactor code for TF 1.0

Signed-off-by: Norman Heckscher <[email protected]>

README.md

@@ -1,6 +1,6 @@
 # [Learning to Learn](https://arxiv.org/abs/1606.04474) in TensorFlow
 
-Compatible with TensorFlow 0.12.
+Compatible with TensorFlow 1.0
 
 
 ## Training

meta.py

@@ -360,7 +360,7 @@ def time_step(t, fx_array, x, state):
  fx_final = _make_with_custom_variables(make_loss, x_final)
  fx_array = fx_array.write(len_unroll, fx_final)
 
- loss = tf.reduce_sum(fx_array.pack(), name="loss")
+ loss = tf.reduce_sum(fx_array.stack(), name="loss")
 
  # Reset the state; should be called at the beginning of an epoch.
  with tf.name_scope("reset"):

networks.py

@@ -88,7 +88,7 @@ def _convert_to_initializer(initializer):
  """
 
  if isinstance(initializer, str):
- return getattr(tf, initializer + "_initializer")
+ return getattr(tf, initializer + "_initializer")(dtype=tf.float32)
  elif isinstance(initializer, np.ndarray):
  return tf.constant_initializer(initializer)
  else:
@@ -182,7 +182,7 @@ def __init__(self, output_size, layers, preprocess_name="identity",
  else:
  self._preprocess = getattr(tf, preprocess_name)
 
- with tf.variable_scope(self._template.var_scope):
+ with tf.variable_scope(self._template.variable_scope):
  self._cores = []
  for i, size in enumerate(layers, start=1):
  name = "lstm_{}".format(i)

nn/base.py

@@ -145,30 +145,30 @@ def __call__(self, *args, **kwargs):
  return out
 
  @property
- def var_scope(self):
+ def variable_scope(self):
  """Returns the variable_scope declared by the module.
 
- It is valid for library users to access the internal templated var_scope,
+ It is valid for library users to access the internal templated variable_scope,
  but only makes sense to do so after connection. Therefore we raise an error
- here if the var_scope is requested before connection.
+ here if the variable_scope is requested before connection.
 
- The only case where it does make sense to access the var_scope before
+ The only case where it does make sense to access the variable_scope before
  connection is to get the post-uniquification name, which we support using
  the separate .name property.
 
  Returns:
- var_scope: `tf.VariableScope` instance of the internal `tf.Template`.
+ variable_scope: `tf.VariableScope` instance of the internal `tf.Template`.
 
  Raises:
  NotConnectedError: If the module is not connected to the Graph.
  """
  self._ensure_is_connected()
- return self._template.var_scope
+ return self._template.variable_scope
 
  @property
  def name(self):
  """Returns the name of the Module."""
- return self._template.var_scope.name
+ return self._template.variable_scope.name
 
  @property
  def is_connected(self):

nn/batch_norm.py

@@ -131,7 +131,7 @@ def _set_default_initializer(self, var_name):
  if var_name == self.GAMMA:
  self._initializers[self.GAMMA] = tf.ones_initializer()
  elif var_name == self.BETA:
- self._initializers[self.BETA] = tf.zeros_initializer
+ self._initializers[self.BETA] = tf.zeros_initializer()
 
  def _build_statistics_variance(self, input_batch,
  reduction_indices, use_batch_stats):
@@ -151,15 +151,15 @@ def _build_statistics_variance(self, input_batch,
  "moving_mean",
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
- tf.GraphKeys.VARIABLES],
- initializer=tf.zeros_initializer,
+ tf.GraphKeys.GLOBAL_VARIABLES],
+ initializer=tf.zeros_initializer(),
  trainable=False)
 
  self._moving_variance = tf.get_variable(
  "moving_variance",
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
- tf.GraphKeys.VARIABLES],
+ tf.GraphKeys.GLOBAL_VARIABLES],
  initializer=tf.ones_initializer(),
  trainable=False)
 
@@ -217,15 +217,15 @@ def _build_statistics_second_moment(self, input_batch,
  "moving_mean",
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
- tf.GraphKeys.VARIABLES],
- initializer=tf.zeros_initializer,
+ tf.GraphKeys.GLOBAL_VARIABLES],
+ initializer=tf.zeros_initializer(),
  trainable=False)
 
  self._moving_second_moment = tf.get_variable(
  "moving_second_moment",
  shape=self._mean_shape,
  collections=[tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
- tf.GraphKeys.VARIABLES],
+ tf.GraphKeys.GLOBAL_VARIABLES],
  initializer=tf.ones_initializer(),
  trainable=False)
 

nn/convnet.py

@@ -159,7 +159,7 @@ def __init__(self,
  def _instantiate_layers(self):
  """Instantiates all the convolutional modules used in the network."""
 
- with tf.variable_scope(self._template.var_scope):
+ with tf.variable_scope(self._template.variable_scope):
  self._layers = tuple(conv.Conv2D(name="conv_2d_{}".format(i),
  output_channels=self._output_channels[i],
  kernel_shape=self._kernel_shapes[i],
@@ -387,7 +387,7 @@ def __init__(self,
  def _instantiate_layers(self):
  """Instantiates all the convolutional modules used in the network."""
 
- with tf.variable_scope(self._template.var_scope):
+ with tf.variable_scope(self._template.variable_scope):
  self._layers = tuple(
  conv.Conv2DTranspose(name="conv_2d_transpose_{}".format(i),
  output_channels=self._output_channels[i],

nn/gated_rnn.py

@@ -42,7 +42,7 @@
 import tensorflow as tf
 
 from tensorflow.python.ops import array_ops
-from tensorflow.python.ops import rnn_cell
+from tensorflow.contrib import rnn
 
 from nn import base
 from nn import basic
@@ -318,11 +318,11 @@ def _build(self, inputs, prev_state, is_training=True, test_local_stats=True):
  gates = gates_h + gates_x + self._b
  else:
  # Parameters of gates are concatenated into one multiply for efficiency.
- inputs_and_hidden = tf.concat(1, [inputs, prev_hidden])
+ inputs_and_hidden = tf.concat_v2([inputs, prev_hidden], 1)
  gates = tf.matmul(inputs_and_hidden, self._w_xh) + self._b
 
  # i = input_gate, j = new_input, f = forget_gate, o = output_gate
- i, j, f, o = array_ops.split(1, 4, gates)
+ i, j, f, o = array_ops.split(gates, 4, 1)
 
  if self._use_peepholes: # diagonal connections
  self._create_peephole_variables(inputs.dtype)
@@ -559,7 +559,7 @@ def create_batch_norm():
  else:
  return create_batch_norm()
 
- class CellWithExtraInput(rnn_cell.RNNCell):
+ class CellWithExtraInput(rnn.RNNCell):
  """Wraps an RNNCell to create a new RNNCell with extra input appended.
 
  This will pass the additional input `args` and `kwargs` to the __call__

nn/mlp.py

@@ -97,7 +97,7 @@ def _instantiate_layers(self):
  connected to the graph.
  """
 
- with tf.variable_scope(self._template.var_scope):
+ with tf.variable_scope(self._template.variable_scope):
  self._layers = [basic.Linear(self._output_sizes[i],
  name="linear_{}".format(i),
  initializers=self._initializers,

nn/rnn_core.py

@@ -30,7 +30,7 @@
 import tensorflow as tf
 
 from tensorflow.python.framework import tensor_shape
-from tensorflow.python.ops import rnn_cell
+from tensorflow.contrib import rnn
 from tensorflow.python.util import nest
 
 from nn import base
@@ -142,7 +142,7 @@ def trainable_initial_state(batch_size, state_size, dtype, initializers=None):
 
 
 @six.add_metaclass(abc.ABCMeta)
-class RNNCore(base.AbstractModule, rnn_cell.RNNCell):
+class RNNCore(base.AbstractModule, rnn.RNNCell):
  """Superclass for Recurrent Neural Network Cores.
 
  This class defines the basic functionality that every core should implement,

nn/util.py

@@ -62,7 +62,7 @@ def get_variables_in_module(module,
  Raises:
  NotConnectedError: If the module is not connected to the Graph.
  """
- return get_variables_in_scope(module.var_scope, collection=collection)
+ return get_variables_in_scope(module.variable_scope, collection=collection)
 
 
 def check_initializers(initializers, keys):

preprocess.py

@@ -68,4 +68,4 @@ def _build(self, gradients):
  clamped_log = Clamp(min_value=-1.0)(log / self._k) # pylint: disable=not-callable
  sign = Clamp(min_value=-1.0, max_value=1.0)(gradients * np.exp(self._k)) # pylint: disable=not-callable
 
- return tf.concat(ndims - 1, [clamped_log, sign])
+ return tf.concat_v2([clamped_log, sign], ndims - 1)

problems.py

@@ -63,7 +63,7 @@ def get_coordinate(i):
 
  def build():
  coordinates = [get_coordinate(i) for i in xrange(num_dims)]
- x = tf.concat(0, [tf.expand_dims(c, 0) for c in coordinates])
+ x = tf.concat_v2([tf.expand_dims(c, 0) for c in coordinates], 0)
  return tf.reduce_sum(tf.square(x, name="x_squared"))
 
  return build
@@ -94,7 +94,7 @@ def build():
  initializer=tf.random_uniform_initializer(),
  trainable=False)
 
- product = tf.squeeze(tf.batch_matmul(w, tf.expand_dims(x, -1)))
+ product = tf.squeeze(tf.matmul(w, tf.expand_dims(x, -1)))
  return tf.reduce_mean(tf.reduce_sum((product - y) ** 2, 1))
 
  return build
@@ -134,7 +134,7 @@ def build():
 
 
 def _xent_loss(output, labels):
- loss = tf.nn.sparse_softmax_cross_entropy_with_logits(output, labels)
+ loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=output, labels=labels)
  return tf.reduce_mean(loss)