def add_variable(self, name, shape=None, dtype=None, initializer=None):
   """***Only for use by descendants of Metric***."""
   if self._built:
     raise RuntimeError("Can't call add_variable() except in build().")
   if context.in_eager_mode():
     collections = None
   else:
     if self._use_global_variables:
       collections = [ops.GraphKeys.GLOBAL_VARIABLES]
     else:
       collections = [ops.GraphKeys.LOCAL_VARIABLES]
     collections += [ops.GraphKeys.METRIC_VARIABLES]
   # Variables are Checkpointable dependencies of Metrics regardless of the
   # global/local distinction. Users can avoid saving variables by not adding a
   # dependency on the Metric.
   v = self._add_variable_with_custom_getter(
       name=name,
       shape=shape,
       dtype=dtype,
       initializer=initializer,
       trainable=False,
       collections=collections,
       use_resource=True,
       getter=variable_scope.get_variable,
       # Raise duplicate variable exceptions from get_variable rather than
       # Checkpointable.
       overwrite=True)
   self._vars.append(v)
   if context.in_eager_mode():
     self._initial_values[v] = v.value()
   return v

    def infer_step(recent_output, recent_logits, unused_loss):
      """Inference step."""
      if not context.in_eager_mode():
        recent_output.set_shape([None, None, None, 1])
      padded = tf.pad(recent_output, [[0, 0], [0, 1], [0, 0], [0, 0]])
      features["targets"] = padded
      # This is inefficient in that it generates samples at all timesteps,
      # not just the last one, except if target_modality is pointwise.
      samples, logits, losses = self.sample(features)
      # Concatenate the already-generated recent_output with last timestep
      # of the newly-generated samples.
      if target_modality.top_is_pointwise:
        cur_sample = samples[:, -1, :, :]
      else:
        cur_sample = samples[:,
                             common_layers.shape_list(recent_output)[1], :, :]
      cur_sample = tf.to_int64(tf.expand_dims(cur_sample, axis=1))
      samples = tf.concat([recent_output, cur_sample], axis=1)
      if not context.in_eager_mode():
        samples.set_shape([None, None, None, 1])

      # Assuming we have one shard for logits.
      logits = tf.concat([recent_logits, logits[:, -1:]], 1)
      loss = sum([l for l in losses.values() if l is not None])
      return samples, logits, loss

  def testVariablesAcrossGraphs(self):
    optimizer = momentum_lib.MomentumOptimizer(0.01, 0.5)
    with ops.Graph().as_default():
      var0 = resource_variable_ops.ResourceVariable(
          [1.0, 2.0], dtype=dtypes.float32, name="var0")
      var1 = resource_variable_ops.ResourceVariable(
          [3.0, 4.0], dtype=dtypes.float32, name="var1")
      if context.in_eager_mode():
        loss = lambda: math_ops.reduce_sum(var0 + var1)
      else:
        loss = math_ops.reduce_sum(var0 + var1)
      optimizer.minimize(loss)
      optimizer_variables = optimizer.variables()
      self.assertStartsWith(optimizer_variables[0].name, "var0")
      self.assertStartsWith(optimizer_variables[1].name, "var1")
      self.assertEquals(2, len(optimizer_variables))

    with ops.Graph().as_default():
      var2 = resource_variable_ops.ResourceVariable(
          [1.0, 2.0], dtype=dtypes.float32, name="var2")
      var3 = resource_variable_ops.ResourceVariable(
          [3.0, 4.0], dtype=dtypes.float32, name="var3")
      if context.in_eager_mode():
        loss = lambda: math_ops.reduce_sum(var2 + var3)
      else:
        loss = math_ops.reduce_sum(var2 + var3)
      optimizer.minimize(loss)
      optimizer_variables = optimizer.variables()
      self.assertStartsWith(optimizer_variables[0].name, "var2")
      self.assertStartsWith(optimizer_variables[1].name, "var3")
      self.assertEquals(2, len(optimizer_variables))

  def test_zero_padding_1d(self):
    num_samples = 2
    input_dim = 2
    num_steps = 5
    shape = (num_samples, num_steps, input_dim)
    inputs = np.ones(shape)

    # basic test
    with self.test_session(use_gpu=True):
      testing_utils.layer_test(
          keras.layers.ZeroPadding1D,
          kwargs={'padding': 2},
          input_shape=inputs.shape)
      testing_utils.layer_test(
          keras.layers.ZeroPadding1D,
          kwargs={'padding': (1, 2)},
          input_shape=inputs.shape)

    # correctness test
    with self.test_session(use_gpu=True):
      layer = keras.layers.ZeroPadding1D(padding=2)
      layer.build(shape)
      output = layer(keras.backend.variable(inputs))
      if context.in_eager_mode():
        np_output = output.numpy()
      else:
        np_output = keras.backend.eval(output)
      for offset in [0, 1, -1, -2]:
        np.testing.assert_allclose(np_output[:, offset, :], 0.)
      np.testing.assert_allclose(np_output[:, 2:-2, :], 1.)

      layer = keras.layers.ZeroPadding1D(padding=(1, 2))
      layer.build(shape)
      output = layer(keras.backend.variable(inputs))
      if context.in_eager_mode():
        np_output = output.numpy()
      else:
        np_output = keras.backend.eval(output)
      for left_offset in [0]:
        np.testing.assert_allclose(np_output[:, left_offset, :], 0.)
      for right_offset in [-1, -2]:
        np.testing.assert_allclose(np_output[:, right_offset, :], 0.)
      np.testing.assert_allclose(np_output[:, 1:-2, :], 1.)
      layer.get_config()

    # test incorrect use
    with self.assertRaises(ValueError):
      keras.layers.ZeroPadding1D(padding=(1, 1, 1))
    with self.assertRaises(ValueError):
      keras.layers.ZeroPadding1D(padding=None)

  def test_averagepooling_2d(self):
    testing_utils.layer_test(
        keras.layers.AveragePooling2D,
        kwargs={'strides': (2, 2),
                'padding': 'same',
                'pool_size': (2, 2)},
        input_shape=(3, 5, 6, 4))
    testing_utils.layer_test(
        keras.layers.AveragePooling2D,
        kwargs={'strides': (2, 2),
                'padding': 'valid',
                'pool_size': (3, 3)},
        input_shape=(3, 5, 6, 4))

    # This part of the test can only run on GPU but doesn't appear
    # to be properly assigned to a GPU when running in eager mode.
    if not context.in_eager_mode():
      # Only runs on GPU with CUDA, channels_first is not supported on CPU.
      # TODO(b/62340061): Support channels_first on CPU.
      if test.is_gpu_available(cuda_only=True):
        testing_utils.layer_test(
            keras.layers.AveragePooling2D,
            kwargs={
                'strides': (1, 1),
                'padding': 'valid',
                'pool_size': (2, 2),
                'data_format': 'channels_first'
            },
            input_shape=(3, 4, 5, 6))

def assert_integer(x, message=None, name=None):
  """Assert that `x` is of integer dtype.

  Example of adding a dependency to an operation:

  ```python
  with tf.control_dependencies([tf.assert_integer(x)]):
    output = tf.reduce_sum(x)
  ```

  Args:
    x: `Tensor` whose basetype is integer and is not quantized.
    message: A string to prefix to the default message.
    name: A name for this operation (optional).  Defaults to "assert_integer".

  Raises:
    TypeError:  If `x.dtype` is anything other than non-quantized integer.

  Returns:
    A `no_op` that does nothing.  Type can be determined statically.
  """
  message = message or ''
  with ops.name_scope(name, 'assert_integer', [x]):
    x = ops.convert_to_tensor(x, name='x')
    if not x.dtype.is_integer:
      if context.in_eager_mode():
        name = 'tensor'
      else:
        name = x.name
      err_msg = (
          '%s  Expected "x" to be integer type.  Found: %s of dtype %s'
          % (message, name, x.dtype))
      raise TypeError(err_msg)

    return control_flow_ops.no_op('statically_determined_was_integer')

def _defun_internal(name, func, args, kwds):
  """Defines and returns graph-mode version of func."""
  graph_key = ops.get_default_graph()._graph_key  # pylint: disable=protected-access
  with context.graph_mode():
    captures = {}
    tmp_graph = CapturingGraph(captures)
    # Inherit the graph key, since this is used for matching variables in
    # optimizers.
    tmp_graph._graph_key = graph_key  # pylint: disable=protected-access
    # Copy the graph collections to ensure summaries and other things work. This
    # lets the function access (but not mutate) collections of the containing
    # graph, such as the global step and the summary writer collections.
    curr_graph = ops.get_default_graph()
    for collection in curr_graph.collections:
      tmp_graph.get_collection_ref(collection)[:] = curr_graph.get_collection(
          collection)
    with tmp_graph.as_default():
      func_inputs = _get_defun_inputs(args)

      with capture_tensors(captures):
        this_tape = tape.push_new_tape()
        try:
          func_outputs = func(*func_inputs, **kwds)
        finally:
          tape.pop_tape(this_tape)
        variables = this_tape.watched_variables()

        # Returning a closed-over tensor as an output does not trigger a
        # call to convert_to_tensor, so we manually capture all such tensors.
        outputs_list = _flatten(func_outputs)
        func_def_outputs = [
            _convert_to_graph_tensor(x) for x in outputs_list if x is not None
        ]

      ids = list(sorted(captures.keys()))
      if ids:
        extra_inputs, extra_placeholders = zip(* [captures[x] for x in ids])
      else:
        extra_inputs = []
        extra_placeholders = []
      output_shapes = tuple(
          x.shape if isinstance(x, ops.Tensor) else None
          for x in outputs_list)

  flat_inputs = [x for x in nest.flatten(func_inputs)
                 if isinstance(x, ops.Tensor)]
  all_inputs = flat_inputs + list(extra_placeholders)
  all_ignored_ops = frozenset(x.op for x in all_inputs)
  fname = _inference_name(name)
  operations = tuple(x for x in tmp_graph.get_operations()
                     if x not in all_ignored_ops)
  # Register any other functions defined in the graph
  # TODO(ashankar): Oh lord, forgive me for this lint travesty.
  if context.in_eager_mode():
    for f in tmp_graph._functions.values():  # pylint: disable=protected-access
      # TODO(ashankar): What about the gradient registry?
      _register(f._c_func)  # pylint: disable=protected-access
  return GraphModeFunction(
      fname, all_inputs, extra_inputs, tmp_graph, operations, func_def_outputs,
      func_outputs, output_shapes, variables)

  def start_queue_runners(self, sess, queue_runners=None):
    """Start threads for `QueueRunners`.

    Note that the queue runners collected in the graph key `QUEUE_RUNNERS`
    are already started automatically when you create a session with the
    supervisor, so unless you have non-collected queue runners to start
    you do not need to call this explicitly.

    Args:
      sess: A `Session`.
      queue_runners: A list of `QueueRunners`. If not specified, we'll use the
        list of queue runners gathered in the graph under the key
        `GraphKeys.QUEUE_RUNNERS`.

    Returns:
      The list of threads started for the `QueueRunners`.

    Raises:
      RuntimeError: If called with eager execution enabled.

    @compatibility(eager)
    Queues are not compatible with eager execution. To ingest data when eager
    execution is enabled, use the `tf.data` API.
    @end_compatibility
    """
    if context.in_eager_mode():
      raise RuntimeError("Queues are not compatible with eager execution.")
    if queue_runners is None:
      queue_runners = self._graph.get_collection(ops.GraphKeys.QUEUE_RUNNERS)
    threads = []
    for qr in queue_runners:
      threads.extend(qr.create_threads(sess, coord=self._coord, daemon=True,
                                       start=True))
    return threads

  def losses(self):
    """Retrieve the network's losses.

    Will only include losses that are either
    unconditional, or conditional on inputs to this model
    (e.g. will not include losses that depend on tensors
    that aren't inputs to this model).

    Returns:
        A list of loss tensors.
    """
    losses = []
    if context.in_eager_mode():
      for layer in self.layers:
        losses += layer.losses
      return losses

    for layer in self.layers:
      losses += layer.losses

    relevant_inputs = []
    for i in range(len(self._inbound_nodes)):
      inputs = self.get_input_at(i)
      if isinstance(inputs, list):
        relevant_inputs += inputs
      else:
        relevant_inputs.append(inputs)
    reachable = layers_util.get_reachable_from_inputs(relevant_inputs, losses)
    relevant_conditional_losses = [x for x in losses if x in reachable]
    unconditional_losses = [
        x for x in losses if x._unconditional_loss]  # pylint: disable=protected-access
    return list(set(
        relevant_conditional_losses + unconditional_losses + self._losses))

 def decorator(hp_fn, registration_name=None):
   """Registers & returns hp_fn with registration_name or default name."""
   hp_name = registration_name or default_name(hp_fn)
   if hp_name in _HPARAMS and not context.in_eager_mode():
     raise LookupError("HParams set %s already registered." % hp_name)
   _HPARAMS[hp_name] = hp_fn
   return hp_fn

def _convert_to_graph_constant(value, dtype=None, name=None, as_ref=False):
  """Captures a tfe Tensor while building a graph mode function.

  Creates a placeholder to pass the tensor as an argument.

  Arguments:
    value: A tfe.Tensor object
    dtype: The datatype of the value produced by the node in the graph.
    name:  Name of the node in the graph.
    as_ref: Ignored (required by register_tensor_conversion_function).

  Returns:
    A placeholder which will, at runtime, have the value of this tensor.

  Raises:
    ValueError: if called outside a defun context.
  """
  if context.in_eager_mode():
    return value
  _ = as_ref
  tensor_map = _scoped_captures.tensors
  if tensor_map is None:
    raise ValueError(
        "Trying to use tfe.Tensor objects in a graph outside graph mode. "
        "To build a graph use tfe.defun or tfe.make_template.")
  captured_value = tensor_map.get(ops.tensor_id(value), None)
  if captured_value is None:
    captured_value = graph_placeholder(
        dtype=dtype or value.dtype, shape=value.shape, name=name)
    if captured_value.dtype == dtypes.resource:
      captured_value._handle_data = value._handle_data  # pylint: disable=protected-access
    tensor_map[ops.tensor_id(value)] = (value, captured_value)
  else:
    captured_value = captured_value[1]
  return captured_value

  def zero_state(self, batch_size, dtype):
    """Return zero-filled state tensor(s).

    Args:
      batch_size: int, float, or unit Tensor representing the batch size.
      dtype: the data type to use for the state.

    Returns:
      If `state_size` is an int or TensorShape, then the return value is a
      `N-D` tensor of shape `[batch_size, state_size]` filled with zeros.

      If `state_size` is a nested list or tuple, then the return value is
      a nested list or tuple (of the same structure) of `2-D` tensors with
      the shapes `[batch_size, s]` for each s in `state_size`.
    """
    # Try to use the last cached zero_state. This is done to avoid recreating
    # zeros, especially when eager execution is enabled.
    state_size = self.state_size
    is_eager = context.in_eager_mode()
    if is_eager and hasattr(self, "_last_zero_state"):
      (last_state_size, last_batch_size, last_dtype,
       last_output) = getattr(self, "_last_zero_state")
      if (last_batch_size == batch_size and
          last_dtype == dtype and
          last_state_size == state_size):
        return last_output
    with ops.name_scope(type(self).__name__ + "ZeroState", values=[batch_size]):
      output = _zero_state_tensors(state_size, batch_size, dtype)
    if is_eager:
      self._last_zero_state = (state_size, batch_size, dtype, output)
    return output

  def testMultiIONetworkbuilding(self):
    input_a = network_layers.Input(shape=(32,))
    input_b = network_layers.Input(shape=(16,))
    a = core_layers.Dense(16)(input_a)

    class AddLayer(base_layers.Layer):

      def call(self, inputs):
        return inputs[0] + inputs[1]

      def compute_output_shape(self, input_shape):
        return input_shape[0]

    c = AddLayer()([a, input_b])  # pylint: disable=not-callable
    c = core_layers.Dense(2)(c)

    network = network_layers.GraphNetwork([input_a, input_b], [a, c])
    if context.in_eager_mode():
      a_val = constant_op.constant(
          np.random.random((10, 32)).astype('float32'))
      b_val = constant_op.constant(
          np.random.random((10, 16)).astype('float32'))
      outputs = network([a_val, b_val])
      self.assertEqual(len(outputs), 2)
      self.assertEqual(outputs[0].shape.as_list(), [10, 16])
      self.assertEqual(outputs[1].shape.as_list(), [10, 2])

def _convert_to_graph_tensor(value, dtype=None, name=None, as_ref=False):
  """Captures a Tensor while building a graph mode function.

  Arguments:
    value: A Tensor object.
    dtype: The datatype of the value produced by the node in the graph.
    name:  Name of the node in the graph.
    as_ref: Ignored (required by register_tensor_conversion_function).

  Returns:
    Returns a constant (the current value of the tensor) if capturing
    is not enabled. A placeholder which will have the value of the
    tensor at runtime otherwise.
  """
  if context.in_eager_mode():
    return value
  _ = as_ref
  tensor_map = _scoped_captures.tensors
  if tensor_map is None:
    # Capturing is not enabled.
    return constant_op.constant(value.numpy())
  captured_value = tensor_map.get(ops.tensor_id(value), None)
  if captured_value is None:
    captured_value = graph_placeholder(
        dtype=dtype or value.dtype, shape=value.shape, name=name)
    if captured_value.dtype == dtypes.resource:
      captured_value._handle_data = value._handle_data  # pylint: disable=protected-access
    tensor_map[ops.tensor_id(value)] = (value, captured_value)
  else:
    captured_value = captured_value[1]
  tape.record_operation("captured_value", [captured_value], [value], [],
                        lambda x: x)
  return captured_value

  def losses(self):
    """Retrieve the network's losses.

    Will only include losses that are either
    unconditional, or conditional on inputs to this model
    (e.g. will not include losses that depend on tensors
    that aren't inputs to this model).

    Returns:
        A list of loss tensors.
    """
    losses = []
    if context.in_eager_mode():
      for layer in self.layers:
        losses += layer.losses
      return losses

    # Retrieve losses for all internal layers.
    for layer in self.layers:
      if hasattr(layer, 'losses'):
        # Collect losses that are dependent on inputs
        # that are part of the model.
        for node_index, node in enumerate(layer._inbound_nodes):  # pylint: disable=protected-access
          node_key = _make_node_key(layer.name, node_index)
          if node_key in self._network_nodes:
            # The model owns this layer node.
            inputs = node.input_tensors
            losses += layer.get_losses_for(inputs)
        # Collect unconditional losses.
        losses += layer.get_losses_for(None)
    # Add any potential unconditional model-level loss.
    losses += self.get_losses_for(None)
    return losses

  def _get_weights(self, hidden_dim=None):
    """Create or get concatenated embedding or softmax variable.

    Args:
      hidden_dim: dim of the variable. Defaults to self._body_input_depth

    Returns:
       a list of self._num_shards Tensors.
    """
    if hidden_dim is None:
      hidden_dim = self._body_input_depth
    num_shards = self._model_hparams.symbol_modality_num_shards
    shards = []
    for i in xrange(num_shards):
      shard_size = (self._vocab_size // num_shards) + (
          1 if i < self._vocab_size % num_shards else 0)
      var_name = "weights_%d" % i
      shards.append(
          tf.get_variable(
              var_name, [shard_size, hidden_dim],
              initializer=tf.random_normal_initializer(0.0, hidden_dim**-0.5)))
    if num_shards == 1:
      ret = shards[0]
    else:
      ret = tf.concat(shards, 0)
    # Convert ret to tensor.
    if not context.in_eager_mode():
      ret = eu.convert_gradient_to_tensor(ret)
    return ret

  def variables(self):
    """A list of variables which encode the current state of `Optimizer`.

    Includes slot variables and additional global variables created by the
    optimizer in the current default graph.

    Returns:
      A list of variables.
    """
    executing_eagerly = context.in_eager_mode()
    current_graph = ops.get_default_graph()

    def _from_current_graph(variable):
      if executing_eagerly:
        # No variable.op in eager mode. We don't expect lots of eager graphs,
        # but behavior should be consistent with graph mode.
        return variable._container_prefix == current_graph._container_prefix  # pylint: disable=protected-access
      else:
        return variable.op.graph is current_graph

    optimizer_variables = [v for v in self._non_slot_variables()
                           if _from_current_graph(v)]
    for _, variable_dict in self._slots.items():
      for _, slot_for_variable in variable_dict.items():
        if _from_current_graph(slot_for_variable):
          optimizer_variables.append(slot_for_variable)
    # Sort variables by name so that the return is deterministic.
    return sorted(optimizer_variables, key=lambda v: v.name)

  def __init__(self, dataset):
    """Creates a new iterator over the given dataset.

    For example:
    ```python
    dataset = tf.contrib.data.Dataset.range(4)
    for x in Iterator(dataset):
      print(x)
    ```

    Args:
      dataset: A `tf.contrib.data.Dataset` object.

    Raises:
      RuntimeError: When invoked without eager execution enabled.
    """

    if not context.in_eager_mode():
      raise RuntimeError(
          "{} objects only make sense when eager execution is enabled".format(
              type(self)))
    ds_variant = dataset._as_variant_tensor()  # pylint: disable=protected-access
    self._output_types = dataset.output_types
    self._flat_output_types = nest.flatten(dataset.output_types)
    self._flat_output_shapes = nest.flatten(dataset.output_shapes)
    self._resource = gen_dataset_ops.iterator(
        container="",
        shared_name=_iterator_shared_name(),
        output_types=self._flat_output_types,
        output_shapes=self._flat_output_shapes)
    gen_dataset_ops.make_iterator(ds_variant, self._resource)

  def run_evaluation(init_op, call_op, results_op, sess=None):
    """Convenience method for running the ops returned by evaluate_on_dataset.

    Args:
      init_op: An op that initializes/resets evaluation state.
      call_op: An op that updates evaluation state on a mini-batch of examples.
        Must generate an tf.errors.OutOfRangeError when done.
      results_op: A dictionary of tensors that compute the final evaluation
        results from the evaluation state.
      sess: The Session to run the evaluation in. Defaults to the default
        Session.

    Returns:
      A dictionary of values, parallel to results_op.

    Raises:
      RuntimeError: if eager execution is enabled.

    @compatibility(eager)
    Only for graph execution.
    @end_compatibility
    """
    if context.in_eager_mode():
      raise RuntimeError("Evaluator.run_evaluation() not supported when "
                         "eager execution is enabled.")
    sess = sess or ops.get_default_session()
    sess.run(init_op)
    try:
      while True:
        sess.run(call_op)
    except errors_impl.OutOfRangeError:
      pass
    return sess.run(results_op)

def global_step(sess, global_step_tensor):
  """Small helper to get the global step.

  ```python
  # Creates a variable to hold the global_step.
  global_step_tensor = tf.Variable(10, trainable=False, name='global_step')
  # Creates a session.
  sess = tf.Session()
  # Initializes the variable.
  print('global_step: %s' % tf.train.global_step(sess, global_step_tensor))

  global_step: 10
  ```

  Args:
    sess: A TensorFlow `Session` object.
    global_step_tensor:  `Tensor` or the `name` of the operation that contains
      the global step.

  Returns:
    The global step value.
  """
  if context.in_eager_mode():
    return int(global_step_tensor.numpy())
  return int(sess.run(global_step_tensor))