def _compute_weighted_loss(losses, weight):
    """Computes the weighted loss.

  Args:
    losses: A tensor of size [batch_size, d1, ... dN].
    weight: A tensor of size [1] or [batch_size, d1, ... dK] where K < N.

  Returns:
    A scalar `Tensor` that returns the weighted loss.

  Raises:
    ValueError: If the weight shape is not compatible with the losses shape or
      if the number of dimensions (rank) of either losses or weight is missing.
  """
    losses = math_ops.to_float(losses)
    weight = math_ops.to_float(ops.convert_to_tensor(weight))

    if losses.get_shape().ndims is None:
        raise ValueError("losses.get_shape().ndims cannot be None")
    if weight.get_shape().ndims is None:
        raise ValueError("weight.get_shape().ndims cannot be None")

    total_loss = _scale_losses(losses, weight)
    num_present = _num_present(losses, weight)
    mean_loss = _safe_mean(total_loss, num_present)
    ops.add_to_collection(ops.GraphKeys.LOSSES, mean_loss)
    return mean_loss

def huber_loss(labels, predictions, weights=1.0, delta=1.0, scope=None,
               loss_collection=ops.GraphKeys.LOSSES,
               reduction=Reduction.WEIGHTED_SUM_BY_NONZERO_WEIGHTS):
  """Adds a Huber Loss term to the training procedure.

  For each value x in `error=labels-predictions`, the following is calculated:

  ```
    0.5 * x^2                  if |x| <= d
    0.5 * d^2 + d * (|x| - d)  if |x| > d
  ```

  where d is `delta`.

  See: https://en.wikipedia.org/wiki/Huber_loss

  `weights` acts as a coefficient for the loss. If a scalar is provided, then
  the loss is simply scaled by the given value. If `weights` is a tensor of size
  [batch_size], then the total loss for each sample of the batch is rescaled
  by the corresponding element in the `weights` vector. If the shape of
  `weights` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weights`.

  Args:
    labels: The ground truth output tensor, same dimensions as 'predictions'.
    predictions: The predicted outputs.
    weights: Optional `Tensor` whose rank is either 0, or the same rank as
      `labels`, and must be broadcastable to `labels` (i.e., all dimensions must
      be either `1`, or the same as the corresponding `losses` dimension).
    delta: `float`, the point where the huber loss function
      changes from a quadratic to linear.
    scope: The scope for the operations performed in computing the loss.
    loss_collection: collection to which the loss will be added.
    reduction: Type of reduction to apply to loss.

  Returns:
    A scalar `Tensor` that returns the weighted loss.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `labels` or
      if the shape of `weights` is invalid.
  """
  with ops.name_scope(scope, "huber_loss",
                      (predictions, labels, weights)) as scope:
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    error = math_ops.subtract(predictions, labels)
    abs_error = math_ops.abs(error)
    quadratic = math_ops.minimum(abs_error, delta)
    # The following expression is the same in value as
    # tf.maximum(abs_error - delta, 0), but importantly the gradient for the
    # expression when abs_error == delta is 0 (for tf.maximum it would be 1).
    # This is necessary to avoid doubling the gradient, since there is already a
    # nonzero contribution to the gradient from the quadratic term.
    linear = (abs_error - quadratic)
    losses = 0.5 * quadratic**2 + delta * linear
    return compute_weighted_loss(
        losses, weights, scope, loss_collection, reduction=reduction)

def absolute_difference(
    labels, predictions, weights=1.0, scope=None,
    loss_collection=ops.GraphKeys.LOSSES):
  """Adds an Absolute Difference loss to the training procedure.

  `weights` acts as a coefficient for the loss. If a scalar is provided, then
  the loss is simply scaled by the given value. If `weights` is a tensor of
  size [batch_size], then the total loss for each sample of the batch is
  rescaled by the corresponding element in the `weight` vector. If the shape of
  `weight` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weight`.

  Args:
    labels: The ground truth output tensor, same dimensions as 'predictions'.
    predictions: The predicted outputs.
    weights: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.
    loss_collection: collection to which this loss will be added.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `labels` or
      if the shape of `weight` is invalid.
  """
  with ops.name_scope(scope, "absolute_difference",
                      [predictions, labels, weights]) as scope:
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    losses = math_ops.abs(math_ops.sub(predictions, labels))
    return compute_weighted_loss(losses, weights, scope, loss_collection)

def cosine_distance(
    predictions, labels=None, dim=None, weights=1.0, scope=None):
  """Adds a cosine-distance loss to the training procedure.

  Note that the function assumes that `predictions` and `labels` are already
  unit-normalized.

  Args:
    predictions: An arbitrary matrix.
    labels: A `Tensor` whose shape matches 'predictions'
    dim: The dimension along which the cosine distance is computed.
    weights: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If `predictions` shape doesn't match `labels` shape, or
      `weights` is `None`.
  """
  if dim is None:
    raise ValueError("`dim` cannot be None.")
  with ops.name_scope(scope, "cosine_distance_loss",
                      [predictions, labels, weights]) as scope:
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())

    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)

    radial_diffs = math_ops.multiply(predictions, labels)
    losses = 1 - math_ops.reduce_sum(radial_diffs, reduction_indices=[dim,])
    return compute_weighted_loss(losses, weights, scope=scope)

def compute_weighted_loss(losses, weight=1.0):
  """Computes the weighted loss.

  Args:
    losses: A tensor of size [batch_size, d1, ... dN].
    weight: A tensor of size [1] or [batch_size, d1, ... dK] where K < N.

  Returns:
    A scalar `Tensor` that returns the weighted loss.

  Raises:
    ValueError: If the weight is None or the shape is not compatible with the
      losses shape or if the number of dimensions (rank) of either losses or
      weight is missing.
  """
  if weight is None:
    raise ValueError("`weight` cannot be None")
  input_dtype = losses.dtype
  losses = math_ops.to_float(losses)
  weight = math_ops.to_float(ops.convert_to_tensor(weight))

  if losses.get_shape().ndims is None:
    raise ValueError("losses.get_shape().ndims cannot be None")
  if weight.get_shape().ndims is None:
    raise ValueError("weight.get_shape().ndims cannot be None")

  total_loss = _scale_losses(losses, weight)
  num_present = _num_present(losses, weight)
  mean_loss = _safe_mean(total_loss, num_present)
  # convert the result back to the input type
  mean_loss = math_ops.cast(mean_loss, input_dtype)
  add_loss(mean_loss)
  return mean_loss

def compute_weighted_loss(
    losses, weights=1.0, scope=None, loss_collection=ops.GraphKeys.LOSSES):
  """Computes the weighted loss.

  Args:
    losses: `Tensor` of shape `[batch_size, d1, ... dN]`.
    weights: Optional `Tensor` whose rank is either 0, or the same rank as
      `losses`, and must be broadcastable to `losses` (i.e., all dimensions must
      be either `1`, or the same as the corresponding `losses` dimension).
    scope: the scope for the operations performed in computing the loss.
    loss_collection: the loss will be added to these collections.

  Returns:
    A scalar `Tensor` that returns the weighted loss.

  Raises:
    ValueError: If `weights` is `None` or the shape is not compatible with
      `losses`, or if the number of dimensions (rank) of either `losses` or
      `weights` is missing.
  """
  with ops.name_scope(scope, "weighted_loss", (losses, weights)):
    with ops.control_dependencies((
        weights_broadcast_ops.assert_broadcastable(weights, losses),)):
      losses = ops.convert_to_tensor(losses)
      input_dtype = losses.dtype
      losses = math_ops.to_float(losses)
      weights = math_ops.to_float(weights)
      total_loss = _scale_losses(losses, weights)
      num_present = _num_present(losses, weights)
      mean_loss = _safe_mean(total_loss, num_present)
      # Convert the result back to the input type.
      mean_loss = math_ops.cast(mean_loss, input_dtype)
      util.add_loss(mean_loss, loss_collection)
      return mean_loss

def ParseLabelTensorOrDict(labels):
  """Return a tensor to use for input labels to tensor_forest.

  The incoming targets can be a dict where keys are the string names of the
  columns, which we turn into a single 1-D tensor for classification or
  2-D tensor for regression.

  Converts sparse tensors to dense ones.

  Args:
    labels: `Tensor` or `dict` of `Tensor` objects.

  Returns:
    A 2-D tensor for labels/outputs.
  """
  if isinstance(labels, dict):
    return math_ops.to_float(
        array_ops.concat(
            [
                sparse_ops.sparse_tensor_to_dense(
                    labels[k], default_value=-1) if isinstance(
                        labels, sparse_tensor.SparseTensor) else labels[k]
                for k in sorted(labels.keys())
            ],
            1))
  else:
    if isinstance(labels, sparse_tensor.SparseTensor):
      return math_ops.to_float(sparse_ops.sparse_tensor_to_dense(
          labels, default_value=-1))
    else:
      return math_ops.to_float(labels)

def huber_loss(labels, predictions, weight=1.0, k=1.0, scope=None):
    """Define a huber loss  https://en.wikipedia.org/wiki/Huber_loss
      tensor: tensor to regularize.
      k: value of k in the huber loss
      scope: Optional scope for op_scope.

    Huber loss:
    f(x) = if |x| <= k:
              0.5 * x^2
           else:
              k * |x| - 0.5 * k^2

    Returns:
      the L1 loss op.

    http://concise-bio.readthedocs.io/en/latest/_modules/concise/tf_helper.html
    """
    with ops.name_scope(scope, "absolute_difference",
                        [predictions, labels]) as scope:
        predictions.get_shape().assert_is_compatible_with(labels.get_shape())
        if weight is None:
            raise ValueError("`weight` cannot be None")
        predictions = math_ops.to_float(predictions)
        labels = math_ops.to_float(labels)
        diff = math_ops.subtract(predictions, labels)
        abs_diff = tf.abs(diff)
        losses = tf.where(abs_diff < k,
                          0.5 * tf.square(diff),
                          k * abs_diff - 0.5 * k ** 2)
        return tf.losses.compute_weighted_loss(losses, weight)

def mean_squared_error(labels, predictions, weights=1.0, scope=None,
                       loss_collection=ops.GraphKeys.LOSSES):
  """Adds a Sum-of-Squares loss to the training procedure.

  `weights` acts as a coefficient for the loss. If a scalar is provided, then
  the loss is simply scaled by the given value. If `weights` is a tensor of size
  [batch_size], then the total loss for each sample of the batch is rescaled
  by the corresponding element in the `weights` vector. If the shape of
  `weights` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weights`.

  Args:
    labels: The ground truth output tensor, same dimensions as 'predictions'.
    predictions: The predicted outputs.
    weights: Optional `Tensor` whose rank is either 0, or the same rank as
      `labels`, and must be broadcastable to `labels` (i.e., all dimensions must
      be either `1`, or the same as the corresponding `losses` dimension).
    scope: The scope for the operations performed in computing the loss.
    loss_collection: collection to which the loss will be added.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `labels` or
      if the shape of `weights` is invalid.
  """
  with ops.name_scope(scope, "mean_squared_error",
                      (predictions, labels, weights)) as scope:
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    losses = math_ops.square(math_ops.subtract(predictions, labels))
    return compute_weighted_loss(losses, weights, scope, loss_collection)

def cosine_distance(predictions, targets, dim, weight=1.0, scope=None):
  """Adds a cosine-distance loss to the training procedure.

  Note that the function assumes that the predictions and targets are already
  unit-normalized.

  Args:
    predictions: An arbitrary matrix.
    targets: A `Tensor` whose shape matches 'predictions'
    dim: The dimension along which the cosine distance is computed.
    weight: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If predictions.shape doesn't match targets.shape, if the ignore
                mask is provided and its shape doesn't match targets.shape or if
                the ignore mask is not boolean valued.
  """
  with ops.name_scope(scope, "cosine_distance_loss",
                      [predictions, targets]) as scope:
    predictions.get_shape().assert_is_compatible_with(targets.get_shape())
    if weight is None:
      raise ValueError("`weight` cannot be None")

    predictions = math_ops.to_float(predictions)
    targets = math_ops.to_float(targets)

    radial_diffs = math_ops.mul(predictions, targets)
    losses = 1 - math_ops.reduce_sum(radial_diffs, reduction_indices=[dim,])
    return compute_weighted_loss(losses, weight)

def _mask_probs(probs, eos_token, finished):
  """Masks log probabilities.

  The result is that finished beams allocate all probability mass to eos and
  unfinished beams remain unchanged.

  Args:
    probs: Log probabiltiies of shape `[batch_size, beam_width, vocab_size]`
    eos_token: An int32 id corresponding to the EOS token to allocate
      probability to.
    finished: A boolean tensor of shape `[batch_size, beam_width]` that
      specifies which
      elements in the beam are finished already.

  Returns:
    A tensor of shape `[batch_size, beam_width, vocab_size]`, where unfinished
    beams stay unchanged and finished beams are replaced with a tensor with all
    probability on the EOS token.
  """
  vocab_size = array_ops.shape(probs)[2]
  finished_mask = array_ops.expand_dims(
      math_ops.to_float(1. - math_ops.to_float(finished)), 2)
  # These examples are not finished and we leave them
  non_finished_examples = finished_mask * probs
  # All finished examples are replaced with a vector that has all
  # probability on EOS
  finished_row = array_ops.one_hot(
      eos_token,
      vocab_size,
      dtype=dtypes.float32,
      on_value=0.,
      off_value=dtypes.float32.min)
  finished_examples = (1. - finished_mask) * finished_row
  return finished_examples + non_finished_examples

def hinge_loss(labels, logits, weights=1.0, scope=None,
               loss_collection=ops.GraphKeys.LOSSES):
  """Adds a hinge loss to the training procedure.

  WARNING: `weights` also supports dimensions of 1, but the broadcasting does
  not work as advertised, you'll wind up with weighted sum instead of weighted
  mean for any but the last dimension. This will be cleaned up soon, so please
  do not rely on the current behavior for anything but the shapes documented for
  `weights` below.

  Args:
    labels: The ground truth output tensor. Its shape should match the shape of
      logits. The values of the tensor are expected to be 0.0 or 1.0.
    logits: The logits, a float tensor.
    weights: Coefficients for the loss a scalar, a tensor of shape
      `[batch_size]` or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.
    loss_collection: collection to which the loss will be added.

  Returns:
    A scalar `Tensor` of the loss value.

  Raises:
    ValueError: If the shapes of `logits` and `labels` don't match.
  """
  with ops.name_scope(scope, "hinge_loss", (logits, labels)) as scope:
    logits = math_ops.to_float(logits)
    labels = math_ops.to_float(labels)
    logits.get_shape().assert_is_compatible_with(labels.get_shape())
    # We first need to convert binary labels to -1/1 labels (as floats).
    all_ones = array_ops.ones_like(labels)
    labels = math_ops.subtract(2 * labels, all_ones)
    losses = nn_ops.relu(
        math_ops.subtract(all_ones, math_ops.multiply(labels, logits)))
    return compute_weighted_loss(losses, weights, scope, loss_collection)

def log_loss(predictions, labels=None, weights=1.0, epsilon=1e-7, scope=None):
  """Adds a Log Loss term to the training procedure.

  `weights` acts as a coefficient for the loss. If a scalar is provided, then
  the loss is simply scaled by the given value. If `weights` is a tensor of size
  [batch_size], then the total loss for each sample of the batch is rescaled
  by the corresponding element in the `weights` vector. If the shape of
  `weights` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weights`.

  Args:
    predictions: The predicted outputs.
    labels: The ground truth output tensor, same dimensions as 'predictions'.
    weights: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    epsilon: A small increment to add to avoid taking a log of zero.
    scope: The scope for the operations performed in computing the loss.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `labels` or
      if the shape of `weights` is invalid.
  """
  with ops.name_scope(scope, "log_loss",
                      [predictions, labels, weights]) as scope:
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    losses = -math_ops.multiply(
        labels, math_ops.log(predictions + epsilon)) - math_ops.multiply(
            (1 - labels), math_ops.log(1 - predictions + epsilon))
    return compute_weighted_loss(losses, weights, scope=scope)

def hinge_loss(labels, logits, weights=1.0, scope=None,
               loss_collection=ops.GraphKeys.LOSSES,
               reduction=Reduction.SUM_BY_NONZERO_WEIGHTS):
  """Adds a hinge loss to the training procedure.

  Args:
    labels: The ground truth output tensor. Its shape should match the shape of
      logits. The values of the tensor are expected to be 0.0 or 1.0.
    logits: The logits, a float tensor.
    weights: Optional `Tensor` whose rank is either 0, or the same rank as
      `labels`, and must be broadcastable to `labels` (i.e., all dimensions must
      be either `1`, or the same as the corresponding `losses` dimension).
    scope: The scope for the operations performed in computing the loss.
    loss_collection: collection to which the loss will be added.
    reduction: Type of reduction to apply to loss.

  Returns:
    Weighted loss float `Tensor`. If `reduction` is `NONE`, this has the same
    shape as `labels`; otherwise, it is scalar.

  Raises:
    ValueError: If the shapes of `logits` and `labels` don't match.
  """
  with ops.name_scope(scope, "hinge_loss", (logits, labels, weights)) as scope:
    logits = math_ops.to_float(logits)
    labels = math_ops.to_float(labels)
    logits.get_shape().assert_is_compatible_with(labels.get_shape())
    # We first need to convert binary labels to -1/1 labels (as floats).
    all_ones = array_ops.ones_like(labels)
    labels = math_ops.subtract(2 * labels, all_ones)
    losses = nn_ops.relu(
        math_ops.subtract(all_ones, math_ops.multiply(labels, logits)))
    return compute_weighted_loss(
        losses, weights, scope, loss_collection, reduction=reduction)

def absolute_difference(predictions, targets, weight=1.0, scope=None):
  """Adds an Absolute Difference loss to the training procedure.

  `weight` acts as a coefficient for the loss. If a scalar is provided, then the
  loss is simply scaled by the given value. If `weight` is a tensor of size
  [batch_size], then the total loss for each sample of the batch is rescaled
  by the corresponding element in the `weight` vector. If the shape of
  `weight` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weight`.

  Args:
    predictions: The predicted outputs.
    targets: The ground truth output tensor, same dimensions as 'predictions'.
    weight: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `targets` or
      if the shape of `weight` is invalid.
  """
  with ops.name_scope(scope, "absolute_difference",
                      [predictions, targets]) as scope:
    predictions.get_shape().assert_is_compatible_with(targets.get_shape())
    if weight is None:
      raise ValueError("`weight` cannot be None")
    predictions = math_ops.to_float(predictions)
    targets = math_ops.to_float(targets)
    losses = math_ops.abs(math_ops.sub(predictions, targets))
    return compute_weighted_loss(losses, weight)

def contrastive_loss(labels, embeddings_anchor, embeddings_positive,
                     margin=1.0):
  """Computes the contrastive loss.

  This loss encourages the embedding to be close to each other for
    the samples of the same label and the embedding to be far apart at least
    by the margin constant for the samples of different labels.
  See: http://yann.lecun.com/exdb/publis/pdf/hadsell-chopra-lecun-06.pdf

  Args:
    labels: 1-D tf.int32 `Tensor` with shape [batch_size] of
      binary labels indicating positive vs negative pair.
    embeddings_anchor: 2-D float `Tensor` of embedding vectors for the anchor
      images. Embeddings should be l2 normalized.
    embeddings_positive: 2-D float `Tensor` of embedding vectors for the
      positive images. Embeddings should be l2 normalized.
    margin: margin term in the loss definition.

  Returns:
    contrastive_loss: tf.float32 scalar.
  """
  # Get per pair distances
  distances = math_ops.sqrt(
      math_ops.reduce_sum(
          math_ops.square(embeddings_anchor - embeddings_positive), 1))

  # Add contrastive loss for the siamese network.
  #   label here is {0,1} for neg, pos.
  return math_ops.reduce_mean(
      math_ops.to_float(labels) * math_ops.square(distances) +
      (1. - math_ops.to_float(labels)) *
      math_ops.square(math_ops.maximum(margin - distances, 0.)),
      name='contrastive_loss')

def mean_squared_error(predictions, labels=None, weights=1.0, scope=None):
  """Adds a Sum-of-Squares loss to the training procedure.

  `weights` acts as a coefficient for the loss. If a scalar is provided, then
  the loss is simply scaled by the given value. If `weights` is a tensor of size
  [batch_size], then the total loss for each sample of the batch is rescaled
  by the corresponding element in the `weights` vector. If the shape of
  `weights` matches the shape of `predictions`, then the loss of each
  measurable element of `predictions` is scaled by the corresponding value of
  `weights`.

  Args:
    predictions: The predicted outputs.
    labels: The ground truth output tensor, same dimensions as 'predictions'.
    weights: Coefficients for the loss a scalar, a tensor of shape
      [batch_size] or a tensor whose shape matches `predictions`.
    scope: The scope for the operations performed in computing the loss.

  Returns:
    A scalar `Tensor` representing the loss value.

  Raises:
    ValueError: If the shape of `predictions` doesn't match that of `labels` or
      if the shape of `weights` is invalid.
  """
  with ops.name_scope(scope, "mean_squared_error",
                      [predictions, labels, weights]) as scope:
    predictions.get_shape().assert_is_compatible_with(labels.get_shape())
    predictions = math_ops.to_float(predictions)
    labels = math_ops.to_float(labels)
    losses = math_ops.square(math_ops.subtract(predictions, labels))
    return compute_weighted_loss(losses, weights, scope=scope)

def add_image_comparison_summaries(gan_model, num_comparisons=2,
                                   display_diffs=False):
  """Adds image summaries to compare triplets of images.

  The first image is the generator input, the second is the generator output,
  and the third is the real data. This style of comparison is useful for
  image translation problems, where the generator input is a corrupted image,
  the generator output is the reconstruction, and the real data is the target.

  Args:
    gan_model: A GANModel tuple.
    num_comparisons: The number of image triplets to display.
    display_diffs: Also display the difference between generated and target.

  Raises:
    ValueError: If real data, generated data, and generator inputs aren't
      images.
    ValueError: If the generator input, real, and generated data aren't all the
      same size.
  """
  if isinstance(gan_model, namedtuples.CycleGANModel):
    saved_params = locals()
    saved_params.pop('gan_model', None)
    with ops.name_scope('cyclegan_x2y_image_comparison_summaries'):
      add_image_comparison_summaries(gan_model.model_x2y, **saved_params)
    with ops.name_scope('cyclegan_y2x_image_comparison_summaries'):
      add_image_comparison_summaries(gan_model.model_y2x, **saved_params)
    return

  _assert_is_image(gan_model.generator_inputs)
  _assert_is_image(gan_model.generated_data)
  _assert_is_image(gan_model.real_data)

  gan_model.generated_data.shape.assert_is_compatible_with(
      gan_model.generator_inputs.shape)
  gan_model.real_data.shape.assert_is_compatible_with(
      gan_model.generated_data.shape)

  image_list = []
  image_list.extend(
      array_ops.unstack(gan_model.generator_inputs[:num_comparisons]))
  image_list.extend(
      array_ops.unstack(gan_model.generated_data[:num_comparisons]))
  image_list.extend(array_ops.unstack(gan_model.real_data[:num_comparisons]))
  if display_diffs:
    generated_list = array_ops.unstack(
        gan_model.generated_data[:num_comparisons])
    real_list = array_ops.unstack(gan_model.real_data[:num_comparisons])
    diffs = [
        math_ops.abs(math_ops.to_float(generated) - math_ops.to_float(real)) for
        generated, real in zip(generated_list, real_list)]
    image_list.extend(diffs)

  # Reshape image and display.
  summary.image(
      'image_comparison',
      eval_utils.image_reshaper(image_list, num_cols=num_comparisons),
      max_outputs=1)

 def linear_decay_fn(global_step):
   if global_step is None:
     raise ValueError("global_step is required for linear_decay.")
   global_step = math_ops.minimum(global_step, decay_steps)
   remaining_steps = math_ops.to_int32(decay_steps) - math_ops.to_int32(
       global_step)
   decayed = math_ops.to_float(remaining_steps) / math_ops.to_float(
       decay_steps)
   return math_ops.maximum(0.0, decayed)

def compute_weighted_loss(
    losses, weights=1.0, scope=None, loss_collection=ops.GraphKeys.LOSSES,
    reduction=Reduction.SUM_BY_NONZERO_WEIGHTS):
  """Computes the weighted loss.

  Args:
    losses: `Tensor` of shape `[batch_size, d1, ... dN]`.
    weights: Optional `Tensor` whose rank is either 0, or the same rank as
      `losses`, and must be broadcastable to `losses` (i.e., all dimensions must
      be either `1`, or the same as the corresponding `losses` dimension).
    scope: the scope for the operations performed in computing the loss.
    loss_collection: the loss will be added to these collections.
    reduction: Type of reduction to apply to loss.

  Returns:
    Weighted loss `Tensor` of the same type as `losses`. If `reduction` is
    `NONE`, this has the same shape as `losses`; otherwise, it is scalar.

  Raises:
    ValueError: If `weights` is `None` or the shape is not compatible with
      `losses`, or if the number of dimensions (rank) of either `losses` or
      `weights` is missing.

  Note:
    When calculating the gradient of a weighted loss contributions from
    both `losses` and `weights` are considered. If your `weights` depend
    on some model parameters but you do not want this to affect the loss
    gradient, you need to apply @{tf.stop_gradient} to `weights` before
    passing them to `compute_weighted_loss`.
  """
  Reduction.validate(reduction)
  with ops.name_scope(scope, "weighted_loss", (losses, weights)):
    with ops.control_dependencies((
        weights_broadcast_ops.assert_broadcastable(weights, losses),)):
      losses = ops.convert_to_tensor(losses)
      input_dtype = losses.dtype
      losses = math_ops.to_float(losses)
      weights = math_ops.to_float(weights)
      weighted_losses = math_ops.multiply(losses, weights)
      if reduction == Reduction.NONE:
        loss = weighted_losses
      else:
        loss = math_ops.reduce_sum(weighted_losses)
        if reduction == Reduction.MEAN:
          loss = _safe_mean(
              loss,
              math_ops.reduce_sum(array_ops.ones_like(losses) * weights))
        elif (reduction == Reduction.SUM_BY_NONZERO_WEIGHTS or
              reduction == Reduction.SUM_OVER_NONZERO_WEIGHTS):
          loss = _safe_mean(loss, _num_present(losses, weights))
        elif reduction == Reduction.SUM_OVER_BATCH_SIZE:
          loss = _safe_mean(loss, _num_elements(losses))

      # Convert the result back to the input type.
      loss = math_ops.cast(loss, input_dtype)
      util.add_loss(loss, loss_collection)
      return loss