def get_updates(self, loss, params):
    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 = [state_ops.assign_add(self.iterations, 1)]

    lr = self.lr
    if self.initial_decay > 0:
      lr = lr * (  # pylint: disable=g-no-augmented-assignment
          1. / (1. + self.decay * math_ops.cast(self.iterations,
                                                K.dtype(self.decay))))

    for p, g, a, d_a in zip(params, grads, accumulators, delta_accumulators):
      # update accumulator
      new_a = self.rho * a + (1. - self.rho) * math_ops.square(g)
      self.updates.append(state_ops.assign(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 getattr(p, 'constraint', None) is not None:
        new_p = p.constraint(new_p)

      self.updates.append(state_ops.assign(p, new_p))

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

  def get_updates(self, loss, params):
    grads = self.get_gradients(loss, params)
    shapes = [K.int_shape(p) for p in params]
    accumulators = [K.zeros(shape) for shape in shapes]
    self.weights = accumulators
    self.updates = [state_ops.assign_add(self.iterations, 1)]

    lr = self.lr
    if self.initial_decay > 0:
      lr = lr * (  # pylint: disable=g-no-augmented-assignment
          1. /
          (1. +
           self.decay * math_ops.cast(self.iterations, K.dtype(self.decay))))

    for p, g, a in zip(params, grads, accumulators):
      new_a = a + math_ops.square(g)  # update accumulator
      self.updates.append(state_ops.assign(a, new_a))
      new_p = p - lr * g / (K.sqrt(new_a) + self.epsilon)

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

      self.updates.append(state_ops.assign(p, new_p))
    return self.updates

  def get_initial_state(self, inputs):
    # (samples, timesteps, rows, cols, filters)
    initial_state = K.zeros_like(inputs)
    # (samples, rows, cols, filters)
    initial_state = K.sum(initial_state, axis=1)
    shape = list(self.cell.kernel_shape)
    shape[-1] = self.cell.filters
    initial_state = self.cell.input_conv(initial_state,
                                         K.zeros(tuple(shape)),
                                         padding=self.cell.padding)

    if hasattr(self.cell.state_size, '__len__'):
      return [initial_state for _ in self.cell.state_size]
    else:
      return [initial_state]

  def reset_states(self, states=None):
    if not self.stateful:
      raise AttributeError('Layer must be stateful.')
    input_shape = self.input_spec[0].shape
    state_shape = self.compute_output_shape(input_shape)
    if self.return_state:
      state_shape = state_shape[0]
    if self.return_sequences:
      state_shape = state_shape[:1].concatenate(state_shape[2:])
    if None in state_shape:
      raise ValueError('If a RNN is stateful, it needs to know '
                       'its batch size. Specify the batch size '
                       'of your input tensors: \n'
                       '- If using a Sequential model, '
                       'specify the batch size by passing '
                       'a `batch_input_shape` '
                       'argument to your first layer.\n'
                       '- If using the functional API, specify '
                       'the time dimension by passing a '
                       '`batch_shape` argument to your Input layer.\n'
                       'The same thing goes for the number of rows and '
                       'columns.')

    # helper function
    def get_tuple_shape(nb_channels):
      result = list(state_shape)
      if self.cell.data_format == 'channels_first':
        result[1] = nb_channels
      elif self.cell.data_format == 'channels_last':
        result[3] = nb_channels
      else:
        raise KeyError
      return tuple(result)

    # initialize state if None
    if self.states[0] is None:
      if hasattr(self.cell.state_size, '__len__'):
        self.states = [K.zeros(get_tuple_shape(dim))
                       for dim in self.cell.state_size]
      else:
        self.states = [K.zeros(get_tuple_shape(self.cell.state_size))]
    elif states is None:
      if hasattr(self.cell.state_size, '__len__'):
        for state, dim in zip(self.states, self.cell.state_size):
          K.set_value(state, np.zeros(get_tuple_shape(dim)))
      else:
        K.set_value(self.states[0],
                    np.zeros(get_tuple_shape(self.cell.state_size)))
    else:
      if not isinstance(states, (list, tuple)):
        states = [states]
      if len(states) != len(self.states):
        raise ValueError('Layer ' + self.name + ' expects ' +
                         str(len(self.states)) + ' states, ' +
                         'but it received ' + str(len(states)) +
                         ' state values. Input received: ' + str(states))
      for index, (value, state) in enumerate(zip(states, self.states)):
        if hasattr(self.cell.state_size, '__len__'):
          dim = self.cell.state_size[index]
        else:
          dim = self.cell.state_size
        if value.shape != get_tuple_shape(dim):
          raise ValueError('State ' + str(index) +
                           ' is incompatible with layer ' +
                           self.name + ': expected shape=' +
                           str(get_tuple_shape(dim)) +
                           ', found shape=' + str(value.shape))
        # TODO(anjalisridhar): consider batch calls to `set_value`.
        K.set_value(state, value)