def get_updates(self, loss, params):
    grads = self.get_gradients(loss, params)
    self.updates = [K.update_add(self.iterations, 1)]

    lr = self.lr
    if self.initial_decay > 0:
      lr *= (1. / (1. + self.decay * K.cast(self.iterations,
                                            K.dtype(self.decay))))

    t = K.cast(self.iterations, K.floatx()) + 1
    lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
                 (1. - K.pow(self.beta_1, t)))

    ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
    vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
    self.weights = [self.iterations] + ms + vs

    for p, g, m, v in zip(params, grads, ms, vs):
      m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
      v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
      p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)

      self.updates.append(K.update(m, m_t))
      self.updates.append(K.update(v, v_t))
      new_p = p_t

      # Apply constraints.
      if getattr(p, 'constraint', None) is not None:
        new_p = p.constraint(new_p)

      self.updates.append(K.update(p, new_p))
    return self.updates

  def get_updates(self, params, constraints, loss):
    grads = self.get_gradients(loss, params)
    shapes = [K.int_shape(p) for p in params]
    accumulators = [K.zeros(shape) for shape in shapes]
    delta_accumulators = [K.zeros(shape) for shape in shapes]
    self.weights = accumulators + delta_accumulators
    self.updates = []

    lr = self.lr
    if self.initial_decay > 0:
      lr *= (1. / (1. + self.decay * self.iterations))
      self.updates.append(K.update_add(self.iterations, 1))

    for p, g, a, d_a in zip(params, grads, accumulators, delta_accumulators):
      # update accumulator
      new_a = self.rho * a + (1. - self.rho) * K.square(g)
      self.updates.append(K.update(a, new_a))

      # use the new accumulator and the *old* delta_accumulator
      update = g * K.sqrt(d_a + self.epsilon) / K.sqrt(new_a + self.epsilon)

      new_p = p - lr * update
      # apply constraints
      if p in constraints:
        c = constraints[p]
        new_p = c(new_p)
      self.updates.append(K.update(p, new_p))

      # update delta_accumulator
      new_d_a = self.rho * d_a + (1 - self.rho) * K.square(update)
      self.updates.append(K.update(d_a, new_d_a))
    return self.updates

  def call(self, inputs, mask=None):
    input_shape = K.int_shape(inputs)
    if input_shape[0]:
      # batch size matters, use rnn-based implementation
      def step(x, _):
        output = self.layer.call(x)
        return output, []

      _, outputs, _ = K.rnn(
          step,
          inputs,
          initial_states=[],
          input_length=input_shape[1],
          unroll=False)
      y = outputs
    else:
      # No batch size specified, therefore the layer will be able
      # to process batches of any size.
      # We can go with reshape-based implementation for performance.
      input_length = input_shape[1]
      if not input_length:
        input_length = K.shape(inputs)[1]
      # Shape: (num_samples * timesteps, ...)
      inputs = K.reshape(inputs, (-1,) + input_shape[2:])
      y = self.layer.call(inputs)  # (num_samples * timesteps, ...)
      # Shape: (num_samples, timesteps, ...)
      output_shape = self._compute_output_shape(input_shape).as_list()  # pylint: disable=protected-access
      y = K.reshape(y, [-1, input_length] + output_shape[2:])

    # Apply activity regularizer if any:
    if (hasattr(self.layer, 'activity_regularizer') and
        self.layer.activity_regularizer is not None):
      regularization_loss = self.layer.activity_regularizer(y)
      self.add_loss(regularization_loss, inputs)
    return y

  def get_constants(self, inputs, training=None):
    constants = []
    if self.implementation != 0 and 0 < self.dropout < 1:
      input_shape = K.int_shape(inputs)
      input_dim = input_shape[-1]
      ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
      ones = K.tile(ones, (1, int(input_dim)))

      def dropped_inputs():
        return K.dropout(ones, self.dropout)

      dp_mask = [
          K.in_train_phase(dropped_inputs, ones, training=training)
          for _ in range(3)
      ]
      constants.append(dp_mask)
    else:
      constants.append([K.cast_to_floatx(1.) for _ in range(3)])

    if 0 < self.recurrent_dropout < 1:
      ones = K.ones_like(K.reshape(inputs[:, 0, 0], (-1, 1)))
      ones = K.tile(ones, (1, self.units))

      def dropped_inputs():  # pylint: disable=function-redefined
        return K.dropout(ones, self.recurrent_dropout)

      rec_dp_mask = [
          K.in_train_phase(dropped_inputs, ones, training=training)
          for _ in range(3)
      ]
      constants.append(rec_dp_mask)
    else:
      constants.append([K.cast_to_floatx(1.) for _ in range(3)])
    return constants

  def call(self, inputs, mask=None, initial_state=None, training=None):
    # input shape: `(samples, time (padded with zeros), input_dim)`
    # note that the .build() method of subclasses MUST define
    # self.input_spec and self.state_spec with complete input shapes.
    if initial_state is not None:
      if not isinstance(initial_state, (list, tuple)):
        initial_states = [initial_state]
      else:
        initial_states = list(initial_state)
    if isinstance(inputs, list):
      initial_states = inputs[1:]
      inputs = inputs[0]
    elif self.stateful:
      initial_states = self.states
    else:
      initial_states = self.get_initial_states(inputs)

    if len(initial_states) != len(self.states):
      raise ValueError('Layer has ' + str(len(self.states)) +
                       ' states but was passed ' + str(len(initial_states)) +
                       ' initial states.')
    input_shape = K.int_shape(inputs)
    if self.unroll and input_shape[1] is None:
      raise ValueError('Cannot unroll a RNN if the '
                       'time dimension is undefined. \n'
                       '- If using a Sequential model, '
                       'specify the time dimension by passing '
                       'an `input_shape` or `batch_input_shape` '
                       'argument to your first layer. If your '
                       'first layer is an Embedding, you can '
                       'also use the `input_length` argument.\n'
                       '- If using the functional API, specify '
                       'the time dimension by passing a `shape` '
                       'or `batch_shape` argument to your Input layer.')
    constants = self.get_constants(inputs, training=None)
    preprocessed_input = self.preprocess_input(inputs, training=None)
    last_output, outputs, states = K.rnn(
        self.step,
        preprocessed_input,
        initial_states,
        go_backwards=self.go_backwards,
        mask=mask,
        constants=constants,
        unroll=self.unroll)
    if self.stateful:
      updates = []
      for i in range(len(states)):
        updates.append((self.states[i], states[i]))
      self.add_update(updates, inputs)

    # Properly set learning phase
    if 0 < self.dropout + self.recurrent_dropout:
      last_output._uses_learning_phase = True
      outputs._uses_learning_phase = True

    if self.return_sequences:
      return outputs
    else:
      return last_output

  def call(self, inputs, training=None, mask=None):
    kwargs = {}
    if has_arg(self.layer.call, 'training'):
      kwargs['training'] = training
    uses_learning_phase = False  # pylint: disable=redefined-outer-name

    input_shape = K.int_shape(inputs)
    if input_shape[0]:
      # batch size matters, use rnn-based implementation
      def step(x, _):
        global uses_learning_phase  # pylint: disable=global-variable-undefined
        output = self.layer.call(x, **kwargs)
        if hasattr(output, '_uses_learning_phase'):
          uses_learning_phase = (output._uses_learning_phase or
                                 uses_learning_phase)
        return output, []

      _, outputs, _ = K.rnn(
          step,
          inputs,
          initial_states=[],
          unroll=False)
      y = outputs
    else:
      # No batch size specified, therefore the layer will be able
      # to process batches of any size.
      # We can go with reshape-based implementation for performance.
      input_length = input_shape[1]
      if not input_length:
        input_length = K.shape(inputs)[1]
      # Shape: (num_samples * timesteps, ...). And track the
      # transformation in self._input_map.
      input_uid = tf_base_layers._object_list_uid(inputs)
      inputs = K.reshape(inputs, (-1,) + input_shape[2:])
      self._input_map[input_uid] = inputs
      # (num_samples * timesteps, ...)
      y = self.layer.call(inputs, **kwargs)
      if hasattr(y, '_uses_learning_phase'):
        uses_learning_phase = y._uses_learning_phase
      # Shape: (num_samples, timesteps, ...)
      output_shape = self._compute_output_shape(input_shape).as_list()
      y = K.reshape(y, (-1, input_length) + tuple(output_shape[2:]))

    # Apply activity regularizer if any:
    if (hasattr(self.layer, 'activity_regularizer') and
        self.layer.activity_regularizer is not None):
      regularization_loss = self.layer.activity_regularizer(y)
      self.add_loss(regularization_loss, inputs)

    if uses_learning_phase:
      y._uses_learning_phase = True
    return y

  def pop(self):
    """Removes the last layer in the model.

    Raises:
        TypeError: if there are no layers in the model.
    """
    if not self.layers:
      raise TypeError('There are no layers in the model.')

    self.layers.pop()
    if not self.layers:
      self.outputs = []
      self.inbound_nodes = []
      self.outbound_nodes = []
    else:
      self.layers[-1].outbound_nodes = []
      self.outputs = [self.layers[-1].output]
      # update self.inbound_nodes
      self.inbound_nodes[0].output_tensors = self.outputs
      self.inbound_nodes[0].output_shapes = [K.int_shape(self.outputs[0])]
    self.built = False

  def add(self, layer):
    """Adds a layer instance on top of the layer stack.

    Arguments:
        layer: layer instance.

    Raises:
        TypeError: If `layer` is not a layer instance.
        ValueError: In case the `layer` argument does not
            know its input shape.
        ValueError: In case the `layer` argument has
            multiple output tensors, or is already connected
            somewhere else (forbidden in `Sequential` models).
    """
    if not isinstance(layer, Layer):
      raise TypeError('The added layer must be '
                      'an instance of class Layer. '
                      'Found: ' + str(layer))
    if not self.outputs:
      # first layer in model: check that it is an input layer
      if not layer.inbound_nodes:
        # create an input layer
        if not hasattr(layer, 'batch_input_shape'):
          raise ValueError('The first layer in a '
                           'Sequential model must '
                           'get an `input_shape` or '
                           '`batch_input_shape` argument.')
        # Instantiate the input layer.
        x = Input(
            batch_shape=layer.batch_input_shape,
            dtype=layer.dtype,
            name=layer.name + '_input')
        # This will build the current layer
        # and create the node connecting the current layer
        # to the input layer we just created.
        layer(x)

      if len(layer.inbound_nodes) != 1:
        raise ValueError('A layer added to a Sequential model must '
                         'not already be connected somewhere else. '
                         'Model received layer ' + layer.name + ' which has ' +
                         str(len(layer.inbound_nodes)) +
                         ' pre-existing inbound connections.')

      if len(layer.inbound_nodes[0].output_tensors) != 1:
        raise ValueError('All layers in a Sequential model '
                         'should have a single output tensor. '
                         'For multi-output layers, '
                         'use the functional API.')

      self.outputs = [layer.inbound_nodes[0].output_tensors[0]]
      self.inputs = topology.get_source_inputs(self.outputs[0])

      # We create an input node, which we will keep updated
      # as we add more layers
      topology.Node(
          outbound_layer=self,
          inbound_layers=[],
          node_indices=[],
          tensor_indices=[],
          input_tensors=self.inputs,
          output_tensors=self.outputs,
          # no model-level masking for now
          input_masks=[None for _ in self.inputs],
          output_masks=[None])
    else:
      output_tensor = layer(self.outputs[0])
      if isinstance(output_tensor, list):
        raise TypeError('All layers in a Sequential model '
                        'should have a single output tensor. '
                        'For multi-output layers, '
                        'use the functional API.')
      self.outputs = [output_tensor]
      # update self.inbound_nodes
      self.inbound_nodes[0].output_tensors = self.outputs
      self.inbound_nodes[0].output_shapes = [K.int_shape(self.outputs[0])]

    self.layers.append(layer)
    self.built = False

  def set_model(self, model):
    self.model = model
    self.sess = K.get_session()
    if self.histogram_freq and self.merged is None:
      for layer in self.model.layers:
        for weight in layer.weights:
          mapped_weight_name = weight.name.replace(':', '_')
          tf_summary.histogram(mapped_weight_name, weight)
          if self.write_grads:
            grads = model.optimizer.get_gradients(model.total_loss, weight)
            tf_summary.histogram('{}_grad'.format(mapped_weight_name), grads)
          if self.write_images:
            w_img = array_ops.squeeze(weight)
            shape = K.int_shape(w_img)
            if len(shape) == 2:  # dense layer kernel case
              if shape[0] > shape[1]:
                w_img = array_ops.transpose(w_img)
                shape = K.int_shape(w_img)
              w_img = array_ops.reshape(w_img, [1, shape[0], shape[1], 1])
            elif len(shape) == 3:  # convnet case
              if K.image_data_format() == 'channels_last':
                # switch to channels_first to display
                # every kernel as a separate image
                w_img = array_ops.transpose(w_img, perm=[2, 0, 1])
                shape = K.int_shape(w_img)
              w_img = array_ops.reshape(w_img,
                                        [shape[0], shape[1], shape[2], 1])
            elif len(shape) == 1:  # bias case
              w_img = array_ops.reshape(w_img, [1, shape[0], 1, 1])
            else:
              # not possible to handle 3D convnets etc.
              continue

            shape = K.int_shape(w_img)
            assert len(shape) == 4 and shape[-1] in [1, 3, 4]
            tf_summary.image(mapped_weight_name, w_img)

        if hasattr(layer, 'output'):
          tf_summary.histogram('{}_out'.format(layer.name), layer.output)
    self.merged = tf_summary.merge_all()

    if self.write_graph:
      self.writer = tf_summary.FileWriter(self.log_dir, self.sess.graph)
    else:
      self.writer = tf_summary.FileWriter(self.log_dir)

    if self.embeddings_freq:
      embeddings_layer_names = self.embeddings_layer_names

      if not embeddings_layer_names:
        embeddings_layer_names = [
            layer.name for layer in self.model.layers
            if type(layer).__name__ == 'Embedding'
        ]

      embeddings = {
          layer.name: layer.weights[0]
          for layer in self.model.layers if layer.name in embeddings_layer_names
      }

      self.saver = saver_lib.Saver(list(embeddings.values()))

      embeddings_metadata = {}

      if not isinstance(self.embeddings_metadata, str):
        embeddings_metadata = self.embeddings_metadata
      else:
        embeddings_metadata = {
            layer_name: self.embeddings_metadata
            for layer_name in embeddings.keys()
        }

      config = projector.ProjectorConfig()
      self.embeddings_ckpt_path = os.path.join(self.log_dir,
                                               'keras_embedding.ckpt')

      for layer_name, tensor in embeddings.items():
        embedding = config.embeddings.add()
        embedding.tensor_name = tensor.name

        if layer_name in embeddings_metadata:
          embedding.metadata_path = embeddings_metadata[layer_name]

      projector.visualize_embeddings(self.writer, config)