def _renorm_correction_and_moments(self, mean, variance, training):
    """Returns the correction and update values for renorm."""
    stddev = math_ops.sqrt(variance + self.epsilon)
    # Compute the average mean and standard deviation, as if they were
    # initialized with this batch's moments.
    mixed_renorm_mean = (self.renorm_mean +
                         (1. - self.renorm_mean_weight) * mean)
    mixed_renorm_stddev = (self.renorm_stddev +
                           (1. - self.renorm_stddev_weight) * stddev)
    # Compute the corrections for batch renorm.
    r = stddev / mixed_renorm_stddev
    d = (mean - mixed_renorm_mean) / mixed_renorm_stddev
    # Ensure the corrections use pre-update moving averages.
    with ops.control_dependencies([r, d]):
      mean = array_ops.identity(mean)
      stddev = array_ops.identity(stddev)
    rmin, rmax, dmax = [self.renorm_clipping.get(key)
                        for key in ['rmin', 'rmax', 'dmax']]
    if rmin is not None:
      r = math_ops.maximum(r, rmin)
    if rmax is not None:
      r = math_ops.minimum(r, rmax)
    if dmax is not None:
      d = math_ops.maximum(d, -dmax)
      d = math_ops.minimum(d, dmax)
    # When not training, use r=1, d=0, and decay=1 meaning no updates.
    r = utils.smart_cond(training, lambda: r, lambda: array_ops.ones_like(r))
    d = utils.smart_cond(training, lambda: d, lambda: array_ops.zeros_like(d))
    decay = utils.smart_cond(training, lambda: self.renorm_momentum, lambda: 1.)

    def _update_renorm_variable(var, weight, value):
      """Updates a moving average and weight, returns the unbiased value."""
      # Update the variables without zero debiasing. The debiasing will be
      # accomplished by dividing the exponential moving average by the weight.
      # For example, after a single update, the moving average would be
      # (1-decay) * value. and the weight will be 1-decay, with their ratio
      # giving value.
      # Make sure the weight is not updated until before r and d computation.
      value = array_ops.identity(value)
      with ops.control_dependencies([value]):
        weight_value = array_ops.constant(1., dtype=weight.dtype)
      new_var = moving_averages.assign_moving_average(
          var, value, decay, zero_debias=False)
      new_weight = moving_averages.assign_moving_average(
          weight, weight_value, decay, zero_debias=False)
      return new_var / new_weight

    with ops.colocate_with(self.moving_mean):
      new_mean = _update_renorm_variable(self.renorm_mean,
                                         self.renorm_mean_weight,
                                         mean)
    with ops.colocate_with(self.moving_variance):
      new_stddev = _update_renorm_variable(self.renorm_stddev,
                                           self.renorm_stddev_weight,
                                           stddev)
      # Make sqrt(moving_variance + epsilon) = new_stddev.
      new_variance = math_ops.square(new_stddev) - self.epsilon

    return (r, d, new_mean, new_variance)

  def _fused_batch_norm(self, inputs, training):
    """Returns the output of fused batch norm."""
    # TODO(reedwm): Add support for fp16 inputs.
    beta = self.beta if self.center else self._beta_const
    gamma = self.gamma if self.scale else self._gamma_const

    def _fused_batch_norm_training():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          epsilon=self.epsilon,
          data_format=self._data_format)

    def _fused_batch_norm_inference():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          mean=self.moving_mean,
          variance=self.moving_variance,
          epsilon=self.epsilon,
          is_training=False,
          data_format=self._data_format)

    output, mean, variance = utils.smart_cond(
        training, _fused_batch_norm_training, _fused_batch_norm_inference)
    mean = array_ops.reshape(mean, shape=self.moving_mean.get_shape())
    variance = array_ops.reshape(variance,
                                 shape=self.moving_variance.get_shape())
    if not self._bessels_correction_test_only:
      # Remove Bessel's correction to be consistent with non-fused batch norm.
      # Note that the variance computed by fused batch norm is
      # with Bessel's correction.
      sample_size = math_ops.cast(
          array_ops.size(inputs) / array_ops.size(variance), variance.dtype)
      factor = (sample_size - math_ops.cast(1.0, variance.dtype)) / sample_size
      variance *= factor

    training_value = utils.constant_value(training)
    if training_value is None:
      one_minus_decay = utils.smart_cond(training,
                                         lambda: self._one_minus_decay,
                                         lambda: 0.)
    else:
      one_minus_decay = ops.convert_to_tensor(self._one_minus_decay)
    if training_value or training_value is None:
      mean_update = self._assign_moving_average(self.moving_mean, mean,
                                                one_minus_decay)
      variance_update = self._assign_moving_average(self.moving_variance,
                                                    variance, one_minus_decay)
      if context.in_graph_mode():
        # Note that in Eager mode, the updates are already executed when running
        # assign_moving_averages. So we do not need to put them into
        # collections.
        self.add_update(mean_update, inputs=inputs)
        self.add_update(variance_update, inputs=inputs)

    return output

  def _fused_batch_norm(self, inputs, training):
    """Returns the output of fused batch norm."""
    beta = self.beta if self.center else self._beta_const
    gamma = self.gamma if self.scale else self._gamma_const

    def _fused_batch_norm_training():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          epsilon=self.epsilon,
          data_format=self._data_format)

    def _fused_batch_norm_inference():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          mean=self.moving_mean,
          variance=self.moving_variance,
          epsilon=self.epsilon,
          is_training=False,
          data_format=self._data_format)

    output, mean, variance = utils.smart_cond(
        training, _fused_batch_norm_training, _fused_batch_norm_inference)
    if not self._bessels_correction_test_only:
      # Remove Bessel's correction to be consistent with non-fused batch norm.
      # Note that the variance computed by fused batch norm is
      # with Bessel's correction.
      sample_size = math_ops.cast(
          array_ops.size(inputs) / array_ops.size(variance), variance.dtype)
      factor = (sample_size - math_ops.cast(1.0, variance.dtype)) / sample_size
      variance *= factor

    training_value = utils.constant_value(training)
    if training_value is None:
      momentum = utils.smart_cond(training, lambda: self.momentum, lambda: 1.0)
    else:
      momentum = ops.convert_to_tensor(self.momentum)
    if training_value or training_value is None:
      mean_update = self._assign_moving_average(self.moving_mean, mean,
                                                momentum)
      variance_update = self._assign_moving_average(self.moving_variance,
                                                    variance, momentum)
      self.add_update(mean_update, inputs=inputs)
      self.add_update(variance_update, inputs=inputs)

    return output

  def _fused_batch_norm_op(self, input_batch, mean, variance, use_batch_stats):
    """Creates a fused batch normalization op."""
    # The fused batch norm expects the mean, variance, gamma and beta
    # tensors to have dimension 1, so we flatten them to remove the
    # extra dimensions.
    gamma_flatten = tf.reshape(self._gamma, shape=(-1,))
    beta_flatten = tf.reshape(self._beta, shape=(-1,))
    flatten_mean = tf.reshape(mean, shape=(-1,))
    flatten_variance = tf.reshape(variance, shape=(-1,))
    use_batch_stats = tf.convert_to_tensor(use_batch_stats)

    common_args = {
        "scale": gamma_flatten,
        "offset": beta_flatten,
        "epsilon": self._eps,
        "data_format": self._infer_fused_data_format(input_batch),
        "name": "batch_norm"
    }

    def use_batch_stats_fused_batch_norm():
      return tf.nn.fused_batch_norm(input_batch, mean=None, variance=None,
                                    is_training=True, **common_args)

    def moving_average_fused_batch_norm():
      return tf.nn.fused_batch_norm(input_batch, mean=flatten_mean,
                                    variance=flatten_variance,
                                    is_training=False, **common_args)

    batch_norm_op, mean, variance = utils.smart_cond(
        use_batch_stats, use_batch_stats_fused_batch_norm,
        moving_average_fused_batch_norm)

    return batch_norm_op, mean, variance

def summary_writer_function(name, tensor, function, family=None):
  """Helper function to write summaries.

  Args:
    name: name of the summary
    tensor: main tensor to form the summary
    function: function taking a tag and a scope which writes the summary
    family: optional, the summary's family

  Returns:
    The result of writing the summary.
  """
  def record():
    with summary_op_util.summary_scope(
        name, family, values=[tensor]) as (tag, scope):
      with ops.control_dependencies([function(tag, scope)]):
        return constant_op.constant(True)

  if context.context().summary_writer_resource is None:
    return control_flow_ops.no_op()
  with ops.device("cpu:0"):
    op = utils.smart_cond(
        should_record_summaries(), record, _nothing, name="")
    ops.add_to_collection(ops.GraphKeys._SUMMARY_COLLECTION, op)  # pylint: disable=protected-access
  return op

  def _build_statistics(self, input_batch, axis, use_batch_stats, dtype):
    """Builds the statistics part of the graph when using moving variance.

    Args:
      input_batch: Input batch Tensor.
      axis: Indices of `input_batch` to reduce over.
      use_batch_stats: Boolean to indicate if batch statistics should be
        calculated, otherwise moving averages are returned.
      dtype: TensorFlow datatype to use for the moving mean and variance.

    Returns:
      Tuple of (mean, variance).
    """
    # Set up our moving statistics. When connecting in parallel, this is shared.
    if self.MOVING_MEAN not in self._initializers:
      self._initializers[self.MOVING_MEAN] = create_mean_initializer()
    self._moving_mean = tf.get_variable(
        "moving_mean",
        dtype=dtype,
        shape=self._mean_shape,
        collections=[
            tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
            tf.GraphKeys.GLOBAL_VARIABLES,
        ],
        initializer=self._initializers[self.MOVING_MEAN],
        trainable=False)

    if self.MOVING_VARIANCE not in self._initializers:
      self._initializers[self.MOVING_VARIANCE] = create_variance_initializer()
    self._moving_variance = tf.get_variable(
        "moving_variance",
        dtype=dtype,
        shape=self._mean_shape,
        collections=[
            tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
            tf.GraphKeys.GLOBAL_VARIABLES,
        ],
        initializer=self._initializers[self.MOVING_VARIANCE],
        trainable=False)

    def build_batch_stats():
      """Builds the batch statistics calculation ops."""
      mean, variance = tf.nn.moments(input_batch, axis,
                                     keep_dims=True, name="normalize_moments")

      return mean, variance

    def build_moving_stats():
      return (
          tf.identity(self._moving_mean),
          tf.identity(self._moving_variance),
      )

    mean, variance = utils.smart_cond(
        use_batch_stats,
        build_batch_stats,
        build_moving_stats,
    )

    return mean, variance

 def call(self, inputs, training=False):
   def dropped_inputs():
     return nn.dropout(inputs, 1  - self.rate,
                       noise_shape=self.noise_shape,
                       seed=self.seed)
   return utils.smart_cond(training,
                           dropped_inputs,
                           lambda: array_ops.identity(inputs))

def crf_log_norm(inputs, sequence_lengths, transition_params):
  """Computes the normalization for a CRF.

  Args:
    inputs: A [batch_size, max_seq_len, num_tags] tensor of unary potentials
        to use as input to the CRF layer.
    sequence_lengths: A [batch_size] vector of true sequence lengths.
    transition_params: A [num_tags, num_tags] transition matrix.
  Returns:
    log_norm: A [batch_size] vector of normalizers for a CRF.
  """
  # Split up the first and rest of the inputs in preparation for the forward
  # algorithm.
  first_input = array_ops.slice(inputs, [0, 0, 0], [-1, 1, -1])
  first_input = array_ops.squeeze(first_input, [1])

  # If max_seq_len is 1, we skip the algorithm and simply reduce_logsumexp over
  # the "initial state" (the unary potentials).
  def _single_seq_fn():
    log_norm = math_ops.reduce_logsumexp(first_input, [1])
    # Mask `log_norm` of the sequences with length <= zero.
    log_norm = array_ops.where(math_ops.less_equal(sequence_lengths, 0),
                               array_ops.zeros_like(log_norm),
                               log_norm)
    return log_norm

  def _multi_seq_fn():
    """Forward computation of alpha values."""
    rest_of_input = array_ops.slice(inputs, [0, 1, 0], [-1, -1, -1])

    # Compute the alpha values in the forward algorithm in order to get the
    # partition function.
    forward_cell = CrfForwardRnnCell(transition_params)
    # Sequence length is not allowed to be less than zero.
    sequence_lengths_less_one = math_ops.maximum(
        constant_op.constant(0, dtype=sequence_lengths.dtype),
        sequence_lengths - 1)
    _, alphas = rnn.dynamic_rnn(
        cell=forward_cell,
        inputs=rest_of_input,
        sequence_length=sequence_lengths_less_one,
        initial_state=first_input,
        dtype=dtypes.float32)
    log_norm = math_ops.reduce_logsumexp(alphas, [1])
    # Mask `log_norm` of the sequences with length <= zero.
    log_norm = array_ops.where(math_ops.less_equal(sequence_lengths, 0),
                               array_ops.zeros_like(log_norm),
                               log_norm)
    return log_norm

  return utils.smart_cond(
      pred=math_ops.equal(
          tensor_shape.dimension_value(
              inputs.shape[1]) or array_ops.shape(inputs)[1],
          1),
      true_fn=_single_seq_fn,
      false_fn=_multi_seq_fn)

  def testConstantValue(self):
    f1 = lambda: constant_op.constant(5)
    f2 = lambda: constant_op.constant(32)

    # Boolean pred
    self.assertEqual(5, utils.constant_value(utils.smart_cond(True, f1, f2)))
    self.assertEqual(32, utils.constant_value(utils.smart_cond(False, f1, f2)))

    # Integer pred
    self.assertEqual(5, utils.constant_value(utils.smart_cond(1, f1, f2)))
    self.assertEqual(32, utils.constant_value(utils.smart_cond(0, f1, f2)))

    # Unknown pred
    pred = array_ops.placeholder_with_default(True, shape=())
    self.assertIsNone(utils.constant_value(utils.smart_cond(pred, f1, f2)))

    #Error case
    with self.assertRaises(TypeError):
      utils.constant_value(5)

def distort_color(image, batch_position=0, distort_color_in_yiq=False,
                  scope=None):
    """Distort the color of the image.

    Each color distortion is non-commutative and thus ordering of the color ops
    matters. Ideally we would randomly permute the ordering of the color ops.
    Rather then adding that level of complication, we select a distinct ordering
    of color ops based on the position of the image in a batch.

    Args:
        image: float32 Tensor containing single image. Tensor values should be in
            range [0, 1].
        batch_position: the position of the image in a batch. NOTE: this argument
            can be an integer or a tensor
        distort_color_in_yiq: distort color of input images in YIQ space.
        scope: Optional scope for op_scope.
    Returns:
        color-distorted image
    """
    with tf.name_scope(scope or 'distort_color'):

        def distort_fn_0(image=image):
            """Variant 0 of distort function."""
            image = tf.image.random_brightness(image, max_delta=32. / 255.)
            if distort_color_in_yiq:
                image = distort_image_ops.random_hsv_in_yiq(
                    image, lower_saturation=0.5, upper_saturation=1.5,
                    max_delta_hue=0.2 * math.pi)
            else:
                image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
                image = tf.image.random_hue(image, max_delta=0.2)
            image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
            return image

        def distort_fn_1(image=image):
            """Variant 1 of distort function."""
            image = tf.image.random_brightness(image, max_delta=32. / 255.)
            image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
            if distort_color_in_yiq:
                image = distort_image_ops.random_hsv_in_yiq(
                    image, lower_saturation=0.5, upper_saturation=1.5,
                    max_delta_hue=0.2 * math.pi)
            else:
                image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
                image = tf.image.random_hue(image, max_delta=0.2)
            return image

        image = utils.smart_cond(batch_position % 2 == 0, distort_fn_0,
                                 distort_fn_1)
        # The random_* ops do not necessarily clamp.
        image = tf.clip_by_value(image, 0.0, 1.0)
        return image

def crf_multitag_sequence_score(inputs, tag_bitmap, sequence_lengths,
                                transition_params):
  """Computes the unnormalized score of all tag sequences matching tag_bitmap.

  tag_bitmap enables more than one tag to be considered correct at each time
  step. This is useful when an observed output at a given time step is
  consistent with more than one tag, and thus the log likelihood of that
  observation must take into account all possible consistent tags.

  Using one-hot vectors in tag_bitmap gives results identical to
  crf_sequence_score.

  Args:
    inputs: A [batch_size, max_seq_len, num_tags] tensor of unary potentials
        to use as input to the CRF layer.
    tag_bitmap: A [batch_size, max_seq_len, num_tags] boolean tensor
        representing all active tags at each index for which to calculate the
        unnormalized score.
    sequence_lengths: A [batch_size] vector of true sequence lengths.
    transition_params: A [num_tags, num_tags] transition matrix.
  Returns:
    sequence_scores: A [batch_size] vector of unnormalized sequence scores.
  """

  # If max_seq_len is 1, we skip the score calculation and simply gather the
  # unary potentials of all active tags.
  def _single_seq_fn():
    filtered_inputs = array_ops.where(
        tag_bitmap, inputs,
        array_ops.fill(array_ops.shape(inputs), float("-inf")))
    return math_ops.reduce_logsumexp(
        filtered_inputs, axis=[1, 2], keepdims=False)

  def _multi_seq_fn():
    # Compute the logsumexp of all scores of sequences matching the given tags.
    filtered_inputs = array_ops.where(
        tag_bitmap, inputs,
        array_ops.fill(array_ops.shape(inputs), float("-inf")))
    return crf_log_norm(
        inputs=filtered_inputs,
        sequence_lengths=sequence_lengths,
        transition_params=transition_params)

  return utils.smart_cond(
      pred=math_ops.equal(
          tensor_shape.dimension_value(
              inputs.shape[1]) or array_ops.shape(inputs)[1],
          1),
      true_fn=_single_seq_fn,
      false_fn=_multi_seq_fn)

  def _build_update_ops(self, mean, variance, is_training):
    """Builds the moving average update ops when using moving variance.

    Args:
      mean: The mean value to update with.
      variance: The variance value to update with.
      is_training: Boolean Tensor to indicate if we're currently in
        training mode.

    Returns:
      Tuple of `(update_mean_op, update_variance_op)` when `is_training` is or
      could be `True`. Returns `None` when `is_training=False`.
    """

    def build_update_ops():
      """Builds the exponential moving average update ops."""

      update_mean_op = moving_averages.assign_moving_average(
          variable=self._moving_mean,
          value=mean,
          decay=self._decay_rate,
          zero_debias=False,
          name="update_moving_mean").op

      update_variance_op = moving_averages.assign_moving_average(
          variable=self._moving_variance,
          value=variance,
          decay=self._decay_rate,
          zero_debias=False,
          name="update_moving_variance").op

      return update_mean_op, update_variance_op

    def build_no_ops():
      return (tf.no_op(), tf.no_op())

    # Only make the ops if we know that `is_training=True`, or the value of
    # `is_training` is unknown.
    is_training_const = utils.constant_value(is_training)
    if is_training_const is None or is_training_const:
      update_mean_op, update_variance_op = utils.smart_cond(
          is_training,
          build_update_ops,
          build_no_ops,
      )
      return (update_mean_op, update_variance_op)
    else:
      return None

def _get_examples(file_name_queue, reader, num_threads, read_batch_size,
                  filter_fn, parse_fn):
  """Get example filenames matching.

  Args:
    file_name_queue: A queue implementation that dequeues elements in
      first-in first-out order.
    reader: A function or class that returns an object with
      `read` method, (filename tensor) -> (example tensor).
    num_threads: The number of threads enqueuing examples.
    read_batch_size: An int or scalar `Tensor` specifying the number of
      records to read at once.
    filter_fn: Filtering function, takes both keys as well as an `Example`
      Tensors and returns a boolean mask of the same shape as the input Tensors
      to be applied for filtering. If `None`, no filtering is done.
    parse_fn: Parsing function, takes `Example` Tensor returns parsed
      representation. If `None`, no parsing is done.

  Returns:
    List of example file names matching `file_name_queue`.
  """
  with ops.name_scope('read'):
    example_list = []
    for _ in range(num_threads):
      keys, examples_proto = utils.smart_cond(
          read_batch_size > 1,
          lambda: reader().read_up_to(file_name_queue, read_batch_size),
          lambda: reader().read(file_name_queue))

      if filter_fn:
        mask = filter_fn(keys, examples_proto)
        keys = array_ops.boolean_mask(keys, mask)
        examples_proto = array_ops.boolean_mask(examples_proto, mask)
      if parse_fn:
        parsed_examples = parse_fn(examples_proto)
        # Map keys into example map because batch_join doesn't support
        # tuple of Tensor + dict.
        if isinstance(parsed_examples, dict):
          parsed_examples[KEY_FEATURE_NAME] = keys
          example_list.append(parsed_examples)
        else:
          example_list.append((keys, parsed_examples))
      else:
        example_list.append((keys, examples_proto))
    return example_list

def crf_sequence_score(inputs, tag_indices, sequence_lengths,
                       transition_params):
  """Computes the unnormalized score for a tag sequence.

  Args:
    inputs: A [batch_size, max_seq_len, num_tags] tensor of unary potentials
        to use as input to the CRF layer.
    tag_indices: A [batch_size, max_seq_len] matrix of tag indices for which we
        compute the unnormalized score.
    sequence_lengths: A [batch_size] vector of true sequence lengths.
    transition_params: A [num_tags, num_tags] transition matrix.
  Returns:
    sequence_scores: A [batch_size] vector of unnormalized sequence scores.
  """
  # If max_seq_len is 1, we skip the score calculation and simply gather the
  # unary potentials of the single tag.
  def _single_seq_fn():
    batch_size = array_ops.shape(inputs, out_type=tag_indices.dtype)[0]
    example_inds = array_ops.reshape(
        math_ops.range(batch_size, dtype=tag_indices.dtype), [-1, 1])
    sequence_scores = array_ops.gather_nd(
        array_ops.squeeze(inputs, [1]),
        array_ops.concat([example_inds, tag_indices], axis=1))
    sequence_scores = array_ops.where(math_ops.less_equal(sequence_lengths, 0),
                                      array_ops.zeros_like(sequence_scores),
                                      sequence_scores)
    return sequence_scores

  def _multi_seq_fn():
    # Compute the scores of the given tag sequence.
    unary_scores = crf_unary_score(tag_indices, sequence_lengths, inputs)
    binary_scores = crf_binary_score(tag_indices, sequence_lengths,
                                     transition_params)
    sequence_scores = unary_scores + binary_scores
    return sequence_scores

  return utils.smart_cond(
      pred=math_ops.equal(
          tensor_shape.dimension_value(
              inputs.shape[1]) or array_ops.shape(inputs)[1],
          1),
      true_fn=_single_seq_fn,
      false_fn=_multi_seq_fn)

def summary_writer_function(name, tensor, function, family=None):
  """Helper function to write summaries.

  Args:
    name: name of the summary
    tensor: main tensor to form the summary
    function: function taking a tag and a scope which writes the summary
    family: optional, the summary's family

  Returns:
    The result of writing the summary.
  """
  def record():
    with summary_op_util.summary_scope(
        name, family, values=[tensor]) as (tag, scope):
      function(tag, scope)
      return True

  return utils.smart_cond(
      should_record_summaries(), record, _nothing, name="")

 def _update_renorm_variable(var, weight, value):
   """Updates a moving average and weight, returns the unbiased value."""
   value = array_ops.identity(value)
   def _do_update():
     # Update the variables without zero debiasing. The debiasing will be
     # accomplished by dividing the exponential moving average by the weight.
     # For example, after a single update, the moving average would be
     # (1-decay) * value. and the weight will be 1-decay, with their ratio
     # giving the value.
     # Make sure the weight is not updated until before r and d computation.
     with ops.control_dependencies([value]):
       weight_value = array_ops.constant(1., dtype=weight.dtype)
     new_var = moving_averages.assign_moving_average(
         var, value, self.renorm_momentum, zero_debias=False)
     new_weight = moving_averages.assign_moving_average(
         weight, weight_value, self.renorm_momentum, zero_debias=False)
     return new_var / new_weight
   def _fake_update():
     return array_ops.identity(var)
   return utils.smart_cond(training, _do_update, _fake_update)

def dropout_selu(x, rate, alpha= -1.7580993408473766, fixedPointMean=0.0, fixedPointVar=1.0,
                 noise_shape=None, seed=None, name=None, training=False):
    from tensorflow.python.framework import ops, tensor_shape, tensor_util
    from tensorflow.python.ops import array_ops, random_ops, math_ops
    from tensorflow.python.layers import utils
    import numbers
    import tensorflow as tf
    """Dropout to a value with rescaling."""

    def dropout_selu_impl(x, rate, alpha, noise_shape, seed, name):
        keep_prob = 1.0 - rate
        x = ops.convert_to_tensor(x, name="x")
        if isinstance(keep_prob, numbers.Real) and not 0. < keep_prob <= 1.:
            raise ValueError("keep_prob must be a scalar tensor or a float in the "
                                             "range (0, 1], got %g" % keep_prob)
        keep_prob = ops.convert_to_tensor(keep_prob, dtype=x.dtype, name="keep_prob")
        keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())

        alpha = ops.convert_to_tensor(alpha, dtype=x.dtype, name="alpha")
        keep_prob.get_shape().assert_is_compatible_with(tensor_shape.scalar())

        if tensor_util.constant_value(keep_prob) == 1:
            return x

        noise_shape = noise_shape if noise_shape is not None else array_ops.shape(x)
        random_tensor = keep_prob
        random_tensor += random_ops.random_uniform(noise_shape, seed=seed, dtype=x.dtype)
        binary_tensor = math_ops.floor(random_tensor)
        ret = x * binary_tensor + alpha * (1-binary_tensor)

        a = tf.sqrt(fixedPointVar / (keep_prob *((1-keep_prob) * tf.pow(alpha-fixedPointMean,2) + fixedPointVar)))

        b = fixedPointMean - a * (keep_prob * fixedPointMean + (1 - keep_prob) * alpha)
        ret = a * ret + b
        ret.set_shape(x.get_shape())
        return ret

    with ops.name_scope(name, "dropout", [x]) as name:
        return utils.smart_cond(training,
            lambda: dropout_selu_impl(x, rate, alpha, noise_shape, seed, name),
            lambda: array_ops.identity(x))

  def _fused_batch_norm(self, inputs, training):
    """Returns the output of fused batch norm."""
    beta = self.beta if self.center else self._beta_const
    gamma = self.gamma if self.scale else self._gamma_const

    def _fused_batch_norm_training():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          epsilon=self.epsilon,
          data_format=self._data_format)

    def _fused_batch_norm_inference():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          mean=self.moving_mean,
          variance=self.moving_variance,
          epsilon=self.epsilon,
          is_training=False,
          data_format=self._data_format)

    output, mean, variance = utils.smart_cond(
        training, _fused_batch_norm_training, _fused_batch_norm_inference)

    training_value = utils.constant_value(training)
    if training_value is not False:
      decay = _smart_select(training, lambda: self.momentum, lambda: 1.)
      mean_update = moving_averages.assign_moving_average(
          self.moving_mean, mean, decay, zero_debias=False)
      variance_update = moving_averages.assign_moving_average(
          self.moving_variance, variance, decay, zero_debias=False)
      self.add_update(mean_update, inputs=inputs)
      self.add_update(variance_update, inputs=inputs)

    return output

    def _update_renorm_variable(var, weight, value):
      """Updates a moving average and weight, returns the unbiased value."""
      value = array_ops.identity(value)
      def _do_update():
        """Updates the var and weight, returns their updated ratio."""
        # Update the variables without zero debiasing. The debiasing will be
        # accomplished by dividing the exponential moving average by the weight.
        # For example, after a single update, the moving average would be
        # (1-decay) * value. and the weight will be 1-decay, with their ratio
        # giving the value.
        # Make sure the weight is not updated until before r and d computation.
        with ops.control_dependencies([value]):
          weight_value = array_ops.constant(1., dtype=weight.dtype)
        new_var = self._assign_moving_average(var, value, self.renorm_momentum)
        new_weight = self._assign_moving_average(weight, weight_value,
                                                 self.renorm_momentum)
        # TODO(yuefengz): the updates to var and weighted can not be batched
        # together if we fetch their updated values here. Consider calculating
        # new values and delaying the updates.
        return new_var / new_weight

      def _fake_update():
        return array_ops.identity(var)
      return utils.smart_cond(training, _do_update, _fake_update)

 def resize_method_1():
     return utils.smart_cond(batch_position % len(resize_methods) == 1,
                             lambda: lookup(1), resize_method_2)