Python tensorflow.python.ops.math_ops.truediv() Examples

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _estimate_data_distribution(c, num_examples_per_class_seen):
  """Estimate data distribution as labels are seen.

  Args:
    c: The class labels.  Type `int32`, shape `[batch_size]`.
    num_examples_per_class_seen: A `ResourceVariable` containing counts.
      Type `int64`, shape `[num_classes]`.

  Returns:
    dist: The updated distribution.  Type `float32`, shape `[num_classes]`.
  """
  num_classes = num_examples_per_class_seen.get_shape()[0].value
  # Update the class-count based on what labels are seen in
  # batch.  But do this asynchronously to avoid performing a
  # cross-device round-trip.  Just use the cached value.
  num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
      math_ops.reduce_sum(
          array_ops.one_hot(c, num_classes, dtype=dtypes.int64),
          0))
  init_prob_estimate = math_ops.truediv(
      num_examples_per_class_seen,
      math_ops.reduce_sum(num_examples_per_class_seen))
  return math_ops.cast(init_prob_estimate, dtypes.float32) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return math_ops.select(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two tensors element-wise, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  t = math_ops.truediv(numerator, denominator)
  zero = array_ops.zeros_like(t, dtype=denominator.dtype)
  condition = math_ops.greater(denominator, zero)
  zero = math_ops.cast(zero, t.dtype)
  return array_ops.where(condition, t, zero, name=name) 

def _safe_div(numerator, denominator, name):
  """Divides two values, returning 0 if the denominator is <= 0.

  Args:
    numerator: A real `Tensor`.
    denominator: A real `Tensor`, with dtype matching `numerator`.
    name: Name for the returned op.

  Returns:
    0 if `denominator` <= 0, else `numerator` / `denominator`
  """
  return array_ops.where(
      math_ops.greater(denominator, 0),
      math_ops.truediv(numerator, denominator),
      0,
      name=name) 

def _estimate_data_distribution(labels, num_classes, smoothing_constant=10):
  """Estimate data distribution as labels are seen."""
  # Variable to track running count of classes. Smooth by a nonzero value to
  # avoid division-by-zero. Higher values provide more stability at the cost of
  # slower convergence.
  if smoothing_constant <= 0:
    raise ValueError('smoothing_constant must be nonzero.')
  num_examples_per_class_seen = variables.Variable(
      initial_value=[smoothing_constant] * num_classes,
      trainable=False,
      name='class_count',
      dtype=dtypes.int64)

  # Update the class-count based on what labels are seen in batch.
  num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
      math_ops.reduce_sum(
          array_ops.one_hot(
              labels, num_classes, dtype=dtypes.int64), 0))

  # Normalize count into a probability.
  # NOTE: Without the `+= 0` line below, the test
  # `testMultiThreadedEstimateDataDistribution` fails. The reason is that
  # before this line, `num_examples_per_class_seen` is a Tensor that shares a
  # buffer with an underlying `ref` object. When the `ref` is changed by another
  # thread, `num_examples_per_class_seen` changes as well. Since this can happen
  # in the middle of the normalization computation, we get probabilities that
  # are very far from summing to one. Adding `+= 0` copies the contents of the
  # tensor to a new buffer, which will be consistent from the start to the end
  # of the normalization computation.
  num_examples_per_class_seen += 0
  init_prob_estimate = math_ops.truediv(
      num_examples_per_class_seen,
      math_ops.reduce_sum(num_examples_per_class_seen))

  # Must return float32 (not float64) to agree with downstream `_verify_input`
  # checks.
  return math_ops.cast(init_prob_estimate, dtypes.float32) 

def _estimate_data_distribution(labels, num_classes, smoothing_constant=10):
  """Estimate data distribution as labels are seen."""
  # Variable to track running count of classes. Smooth by a nonzero value to
  # avoid division-by-zero. Higher values provide more stability at the cost of
  # slower convergence.
  if smoothing_constant <= 0:
    raise ValueError('smoothing_constant must be nonzero.')
  num_examples_per_class_seen = variables.Variable(
      initial_value=[smoothing_constant] * num_classes, trainable=False,
      name='class_count', dtype=dtypes.int64)

  # Update the class-count based on what labels are seen in batch.
  num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
      math_ops.reduce_sum(array_ops.one_hot(labels, num_classes,
                                            dtype=dtypes.int64), 0))

  # Normalize count into a probability.
  # NOTE: Without the `+= 0` line below, the test
  # `testMultiThreadedEstimateDataDistribution` fails. The reason is that
  # before this line, `num_examples_per_class_seen` is a Tensor that shares a
  # buffer with an underlying `ref` object. When the `ref` is changed by another
  # thread, `num_examples_per_class_seen` changes as well. Since this can happen
  # in the middle of the normalization computation, we get probabilities that
  # are very far from summing to one. Adding `+= 0` copies the contents of the
  # tensor to a new buffer, which will be consistent from the start to the end
  # of the normalization computation.
  num_examples_per_class_seen += 0
  init_prob_estimate = math_ops.truediv(
      num_examples_per_class_seen,
      math_ops.reduce_sum(num_examples_per_class_seen))

  # Must return float32 (not float64) to agree with downstream `_verify_input`
  # checks.
  return math_ops.cast(init_prob_estimate, dtypes.float32) 

def _auc_convert_hist_to_auc(hist_true_acc, hist_false_acc, nbins):
  """Convert histograms to auc.

  Args:
    hist_true_acc:  `Tensor` holding accumulated histogram of scores for records
      that were `True`.
    hist_false_acc:  `Tensor` holding accumulated histogram of scores for
      records that were `False`.
    nbins:  Integer number of bins in the histograms.

  Returns:
    Scalar `Tensor` estimating AUC.
  """
  # Note that this follows the "Approximating AUC" section in:
  # Efficient AUC learning curve calculation, R. R. Bouckaert,
  # AI'06 Proceedings of the 19th Australian joint conference on Artificial
  # Intelligence: advances in Artificial Intelligence
  # Pages 181-191.
  # Note that the above paper has an error, and we need to re-order our bins to
  # go from high to low score.

  # Normalize histogram so we get fraction in each bin.
  normed_hist_true = math_ops.truediv(hist_true_acc,
                                      math_ops.reduce_sum(hist_true_acc))
  normed_hist_false = math_ops.truediv(hist_false_acc,
                                       math_ops.reduce_sum(hist_false_acc))

  # These become delta x, delta y from the paper.
  delta_y_t = array_ops.reverse_v2(normed_hist_true, [0], name='delta_y_t')
  delta_x_t = array_ops.reverse_v2(normed_hist_false, [0], name='delta_x_t')

  # strict_1d_cumsum requires float32 args.
  delta_y_t = math_ops.cast(delta_y_t, dtypes.float32)
  delta_x_t = math_ops.cast(delta_x_t, dtypes.float32)

  # Trapezoidal integration, \int_0^1 0.5 * (y_t + y_{t-1}) dx_t
  y_t = _strict_1d_cumsum(delta_y_t, nbins)
  first_trap = delta_x_t[0] * y_t[0] / 2.0
  other_traps = delta_x_t[1:] * (y_t[1:] + y_t[:nbins - 1]) / 2.0
  return math_ops.add(first_trap, math_ops.reduce_sum(other_traps), name='auc') 

def _estimate_data_distribution(labels, num_classes, smoothing_constant=10):
  """Estimate data distribution as labels are seen."""
  # Variable to track running count of classes. Smooth by a nonzero value to
  # avoid division-by-zero. Higher values provide more stability at the cost of
  # slower convergence.
  if smoothing_constant <= 0:
    raise ValueError('smoothing_constant must be nonzero.')
  num_examples_per_class_seen = variable_scope.variable(
      initial_value=[smoothing_constant] * num_classes,
      trainable=False,
      name='class_count',
      dtype=dtypes.int64)

  # Update the class-count based on what labels are seen in batch.
  num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
      math_ops.reduce_sum(
          array_ops.one_hot(
              labels, num_classes, dtype=dtypes.int64), 0))

  # Normalize count into a probability.
  # NOTE: Without the `+= 0` line below, the test
  # `testMultiThreadedEstimateDataDistribution` fails. The reason is that
  # before this line, `num_examples_per_class_seen` is a Tensor that shares a
  # buffer with an underlying `ref` object. When the `ref` is changed by another
  # thread, `num_examples_per_class_seen` changes as well. Since this can happen
  # in the middle of the normalization computation, we get probabilities that
  # are very far from summing to one. Adding `+= 0` copies the contents of the
  # tensor to a new buffer, which will be consistent from the start to the end
  # of the normalization computation.
  num_examples_per_class_seen += 0
  init_prob_estimate = math_ops.truediv(
      num_examples_per_class_seen,
      math_ops.reduce_sum(num_examples_per_class_seen))

  # Must return float32 (not float64) to agree with downstream `_verify_input`
  # checks.
  return math_ops.cast(init_prob_estimate, dtypes.float32) 

def _estimate_data_distribution(labels, num_classes, smoothing_constant=10):
  """Estimate data distribution as labels are seen."""
  # Variable to track running count of classes. Smooth by a nonzero value to
  # avoid division-by-zero. Higher values provide more stability at the cost of
  # slower convergence.
  if smoothing_constant <= 0:
    raise ValueError('smoothing_constant must be nonzero.')
  num_examples_per_class_seen = variables.Variable(
      initial_value=[smoothing_constant] * num_classes,
      trainable=False,
      name='class_count',
      dtype=dtypes.int64)

  # Update the class-count based on what labels are seen in batch.
  num_examples_per_class_seen = num_examples_per_class_seen.assign_add(
      math_ops.reduce_sum(
          array_ops.one_hot(
              labels, num_classes, dtype=dtypes.int64), 0))

  # Normalize count into a probability.
  # NOTE: Without the `+= 0` line below, the test
  # `testMultiThreadedEstimateDataDistribution` fails. The reason is that
  # before this line, `num_examples_per_class_seen` is a Tensor that shares a
  # buffer with an underlying `ref` object. When the `ref` is changed by another
  # thread, `num_examples_per_class_seen` changes as well. Since this can happen
  # in the middle of the normalization computation, we get probabilities that
  # are very far from summing to one. Adding `+= 0` copies the contents of the
  # tensor to a new buffer, which will be consistent from the start to the end
  # of the normalization computation.
  num_examples_per_class_seen += 0
  init_prob_estimate = math_ops.truediv(
      num_examples_per_class_seen,
      math_ops.reduce_sum(num_examples_per_class_seen))

  # Must return float32 (not float64) to agree with downstream `_verify_input`
  # checks.
  return math_ops.cast(init_prob_estimate, dtypes.float32) 

def _auc_convert_hist_to_auc(hist_true_acc, hist_false_acc, nbins):
  """Convert histograms to auc.

  Args:
    hist_true_acc:  `Tensor` holding accumulated histogram of scores for records
      that were `True`.
    hist_false_acc:  `Tensor` holding accumulated histogram of scores for
      records that were `False`.
    nbins:  Integer number of bins in the histograms.

  Returns:
    Scalar `Tensor` estimating AUC.
  """
  # Note that this follows the "Approximating AUC" section in:
  # Efficient AUC learning curve calculation, R. R. Bouckaert,
  # AI'06 Proceedings of the 19th Australian joint conference on Artificial
  # Intelligence: advances in Artificial Intelligence
  # Pages 181-191.
  # Note that the above paper has an error, and we need to re-order our bins to
  # go from high to low score.

  # Normalize histogram so we get fraction in each bin.
  normed_hist_true = math_ops.truediv(hist_true_acc,
                                      math_ops.reduce_sum(hist_true_acc))
  normed_hist_false = math_ops.truediv(hist_false_acc,
                                       math_ops.reduce_sum(hist_false_acc))

  # These become delta x, delta y from the paper.
  delta_y_t = array_ops.reverse_v2(normed_hist_true, [0], name='delta_y_t')
  delta_x_t = array_ops.reverse_v2(normed_hist_false, [0], name='delta_x_t')

  # strict_1d_cumsum requires float32 args.
  delta_y_t = math_ops.cast(delta_y_t, dtypes.float32)
  delta_x_t = math_ops.cast(delta_x_t, dtypes.float32)

  # Trapezoidal integration, \int_0^1 0.5 * (y_t + y_{t-1}) dx_t
  y_t = _strict_1d_cumsum(delta_y_t, nbins)
  first_trap = delta_x_t[0] * y_t[0] / 2.0
  other_traps = delta_x_t[1:] * (y_t[1:] + y_t[:nbins - 1]) / 2.0
  return math_ops.add(first_trap, math_ops.reduce_sum(other_traps), name='auc')


# TODO(langmore) Remove once a faster cumsum (accumulate_sum) Op is available.
# Also see if cast to float32 above can be removed with new cumsum.
# See:  https://github.com/tensorflow/tensorflow/issues/813 

def weighted_moving_average(value,
                            decay,
                            weight,
                            truediv=True,
                            collections=None,
                            name=None):
  """Compute the weighted moving average of `value`.

  Conceptually, the weighted moving average is:
    `moving_average(value * weight) / moving_average(weight)`,
  where a moving average updates by the rule
    `new_value = decay * old_value + (1 - decay) * update`
  Internally, this Op keeps moving average variables of both `value * weight`
  and `weight`.

  Args:
    value: A numeric `Tensor`.
    decay: A float `Tensor` or float value.  The moving average decay.
    weight:  `Tensor` that keeps the current value of a weight.
      Shape should be able to multiply `value`.
    truediv:  Boolean, if `True`, dividing by `moving_average(weight)` is
      floating point division.  If `False`, use division implied by dtypes.
    collections:  List of graph collections keys to add the internal variables
      `value * weight` and `weight` to.  Defaults to `[GraphKeys.VARIABLES]`.
    name: Optional name of the returned operation.
      Defaults to "WeightedMovingAvg".

  Returns:
    An Operation that updates and returns the weighted moving average.
  """
  # Unlike assign_moving_average, the weighted moving average doesn't modify
  # user-visible variables. It is the ratio of two internal variables, which are
  # moving averages of the updates.  Thus, the signature of this function is
  # quite different than assign_moving_average.
  if collections is None:
    collections = [ops.GraphKeys.VARIABLES]
  with variable_scope.variable_op_scope(
      [value, weight, decay], name, "WeightedMovingAvg") as scope:
    value_x_weight_var = variable_scope.get_variable(
        "value_x_weight",
        initializer=init_ops.zeros_initializer(value.get_shape(),
                                               dtype=value.dtype),
        trainable=False,
        collections=collections)
    weight_var = variable_scope.get_variable(
        "weight",
        initializer=init_ops.zeros_initializer(weight.get_shape(),
                                               dtype=weight.dtype),
        trainable=False,
        collections=collections)
    numerator = assign_moving_average(value_x_weight_var, value * weight, decay)
    denominator = assign_moving_average(weight_var, weight, decay)

    if truediv:
      return math_ops.truediv(numerator, denominator, name=scope.name)
    else:
      return math_ops.div(numerator, denominator, name=scope.name) 

def weighted_resample(inputs, weights, overall_rate, scope=None,
                      mean_decay=0.999, warmup=10, seed=None):
  """Performs an approximate weighted resampling of `inputs`.

  This method chooses elements from `inputs` where each item's rate of
  selection is proportional to its value in `weights`, and the average
  rate of selection across all inputs (and many invocations!) is
  `overall_rate`.

  Args:
    inputs: A list of tensors whose first dimension is `batch_size`.
    weights: A `[batch_size]`-shaped tensor with each batch member's weight.
    overall_rate: Desired overall rate of resampling.
    scope: Scope to use for the op.
    mean_decay: How quickly to decay the running estimate of the mean weight.
    warmup: Until the resulting tensor has been evaluated `warmup`
      times, the resampling menthod uses the true mean over all calls
      as its weight estimate, rather than a decayed mean.
    seed: Random seed.

  Returns:
    A list of tensors exactly like `inputs`, but with an unknown (and
      possibly zero) first dimension.
    A tensor containing the effective resampling rate used for each output.

  """
  # Algorithm: Just compute rates as weights/mean_weight *
  # overall_rate. This way the the average weight corresponds to the
  # overall rate, and a weight twice the average has twice the rate,
  # etc.
  with ops.name_scope(scope, 'weighted_resample', inputs) as opscope:
    # First: Maintain a running estimated mean weight, with decay
    # adjusted (by also maintaining an invocation count) during the
    # warmup period so that at the beginning, there aren't too many
    # zeros mixed in, throwing the average off.

    with variable_scope.variable_scope(scope, 'estimate_mean', inputs):
      count_so_far = variable_scope.get_local_variable(
          'resample_count', initializer=0)

      estimated_mean = variable_scope.get_local_variable(
          'estimated_mean', initializer=0.0)

      count = count_so_far.assign_add(1)
      real_decay = math_ops.minimum(
          math_ops.truediv((count - 1), math_ops.minimum(count, warmup)),
          mean_decay)

      batch_mean = math_ops.reduce_mean(weights)
      mean = moving_averages.assign_moving_average(
          estimated_mean, batch_mean, real_decay, zero_debias=False)

    # Then, normalize the weights into rates using the mean weight and
    # overall target rate:
    rates = weights * overall_rate / mean

    results = resample_at_rate([rates] + inputs, rates,
                               scope=opscope, seed=seed, back_prop=False)

    return (results[1:], results[0]) 

def _auc_convert_hist_to_auc(hist_true_acc, hist_false_acc, nbins):
  """Convert histograms to auc.

  Args:
    hist_true_acc:  `Tensor` holding accumulated histogram of scores for records
      that were `True`.
    hist_false_acc:  `Tensor` holding accumulated histogram of scores for
      records that were `False`.
    nbins:  Integer number of bins in the histograms.

  Returns:
    Scalar `Tensor` estimating AUC.
  """
  # Note that this follows the "Approximating AUC" section in:
  # Efficient AUC learning curve calculation, R. R. Bouckaert,
  # AI'06 Proceedings of the 19th Australian joint conference on Artificial
  # Intelligence: advances in Artificial Intelligence
  # Pages 181-191.
  # Note that the above paper has an error, and we need to re-order our bins to
  # go from high to low score.

  # Normalize histogram so we get fraction in each bin.
  normed_hist_true = math_ops.truediv(hist_true_acc,
                                      math_ops.reduce_sum(hist_true_acc))
  normed_hist_false = math_ops.truediv(hist_false_acc,
                                       math_ops.reduce_sum(hist_false_acc))

  # These become delta x, delta y from the paper.
  delta_y_t = array_ops.reverse(normed_hist_true, [True], name='delta_y_t')
  delta_x_t = array_ops.reverse(normed_hist_false, [True], name='delta_x_t')

  # strict_1d_cumsum requires float32 args.
  delta_y_t = math_ops.cast(delta_y_t, dtypes.float32)
  delta_x_t = math_ops.cast(delta_x_t, dtypes.float32)

  # Trapezoidal integration, \int_0^1 0.5 * (y_t + y_{t-1}) dx_t
  y_t = _strict_1d_cumsum(delta_y_t, nbins)
  first_trap = delta_x_t[0] * y_t[0] / 2.0
  other_traps = delta_x_t[1:] * (y_t[1:] + y_t[:nbins - 1]) / 2.0
  return math_ops.add(first_trap, math_ops.reduce_sum(other_traps), name='auc')


# TODO(langmore) Remove once a faster cumsum (accumulate_sum) Op is available.
# Also see if cast to float32 above can be removed with new cumsum.
# See:  https://github.com/tensorflow/tensorflow/issues/813 

def weighted_moving_average(value,
                            decay,
                            weight,
                            truediv=True,
                            collections=None,
                            name=None):
  """Compute the weighted moving average of `value`.

  Conceptually, the weighted moving average is:
    `moving_average(value * weight) / moving_average(weight)`,
  where a moving average updates by the rule
    `new_value = decay * old_value + (1 - decay) * update`
  Internally, this Op keeps moving average variables of both `value * weight`
  and `weight`.

  Args:
    value: A numeric `Tensor`.
    decay: A float `Tensor` or float value.  The moving average decay.
    weight:  `Tensor` that keeps the current value of a weight.
      Shape should be able to multiply `value`.
    truediv:  Boolean, if `True`, dividing by `moving_average(weight)` is
      floating point division.  If `False`, use division implied by dtypes.
    collections:  List of graph collections keys to add the internal variables
      `value * weight` and `weight` to.
      Defaults to `[GraphKeys.GLOBAL_VARIABLES]`.
    name: Optional name of the returned operation.
      Defaults to "WeightedMovingAvg".

  Returns:
    An Operation that updates and returns the weighted moving average.
  """
  # Unlike assign_moving_average, the weighted moving average doesn't modify
  # user-visible variables. It is the ratio of two internal variables, which are
  # moving averages of the updates.  Thus, the signature of this function is
  # quite different than assign_moving_average.
  if collections is None:
    collections = [ops.GraphKeys.GLOBAL_VARIABLES]
  with variable_scope.variable_scope(name, "WeightedMovingAvg",
                                     [value, weight, decay]) as scope:
    value_x_weight_var = variable_scope.get_variable(
        "value_x_weight",
        initializer=init_ops.zeros_initializer(value.get_shape(),
                                               dtype=value.dtype),
        trainable=False,
        collections=collections)
    weight_var = variable_scope.get_variable(
        "weight",
        initializer=init_ops.zeros_initializer(weight.get_shape(),
                                               dtype=weight.dtype),
        trainable=False,
        collections=collections)
    numerator = assign_moving_average(value_x_weight_var, value * weight, decay)
    denominator = assign_moving_average(weight_var, weight, decay)

    if truediv:
      return math_ops.truediv(numerator, denominator, name=scope.name)
    else:
      return math_ops.div(numerator, denominator, name=scope.name) 

def _auc_convert_hist_to_auc(hist_true_acc, hist_false_acc, nbins):
  """Convert histograms to auc.

  Args:
    hist_true_acc:  `Tensor` holding accumulated histogram of scores for records
      that were `True`.
    hist_false_acc:  `Tensor` holding accumulated histogram of scores for
      records that were `False`.
    nbins:  Integer number of bins in the histograms.

  Returns:
    Scalar `Tensor` estimating AUC.
  """
  # Note that this follows the "Approximating AUC" section in:
  # Efficient AUC learning curve calculation, R. R. Bouckaert,
  # AI'06 Proceedings of the 19th Australian joint conference on Artificial
  # Intelligence: advances in Artificial Intelligence
  # Pages 181-191.
  # Note that the above paper has an error, and we need to re-order our bins to
  # go from high to low score.

  # Normalize histogram so we get fraction in each bin.
  normed_hist_true = math_ops.truediv(hist_true_acc,
                                      math_ops.reduce_sum(hist_true_acc))
  normed_hist_false = math_ops.truediv(hist_false_acc,
                                       math_ops.reduce_sum(hist_false_acc))

  # These become delta x, delta y from the paper.
  delta_y_t = array_ops.reverse_v2(normed_hist_true, [0], name='delta_y_t')
  delta_x_t = array_ops.reverse_v2(normed_hist_false, [0], name='delta_x_t')

  # strict_1d_cumsum requires float32 args.
  delta_y_t = math_ops.cast(delta_y_t, dtypes.float32)
  delta_x_t = math_ops.cast(delta_x_t, dtypes.float32)

  # Trapezoidal integration, \int_0^1 0.5 * (y_t + y_{t-1}) dx_t
  y_t = _strict_1d_cumsum(delta_y_t, nbins)
  first_trap = delta_x_t[0] * y_t[0] / 2.0
  other_traps = delta_x_t[1:] * (y_t[1:] + y_t[:nbins - 1]) / 2.0
  return math_ops.add(first_trap, math_ops.reduce_sum(other_traps), name='auc')


# TODO(langmore) Remove once a faster cumsum (accumulate_sum) Op is available.
# Also see if cast to float32 above can be removed with new cumsum.
# See:  https://github.com/tensorflow/tensorflow/issues/813 

def _auc_convert_hist_to_auc(hist_true_acc, hist_false_acc, nbins):
  """Convert histograms to auc.

  Args:
    hist_true_acc:  `Tensor` holding accumulated histogram of scores for records
      that were `True`.
    hist_false_acc:  `Tensor` holding accumulated histogram of scores for
      records that were `False`.
    nbins:  Integer number of bins in the histograms.

  Returns:
    Scalar `Tensor` estimating AUC.
  """
  # Note that this follows the "Approximating AUC" section in:
  # Efficient AUC learning curve calculation, R. R. Bouckaert,
  # AI'06 Proceedings of the 19th Australian joint conference on Artificial
  # Intelligence: advances in Artificial Intelligence
  # Pages 181-191.
  # Note that the above paper has an error, and we need to re-order our bins to
  # go from high to low score.

  # Normalize histogram so we get fraction in each bin.
  normed_hist_true = math_ops.truediv(hist_true_acc,
                                      math_ops.reduce_sum(hist_true_acc))
  normed_hist_false = math_ops.truediv(hist_false_acc,
                                       math_ops.reduce_sum(hist_false_acc))

  # These become delta x, delta y from the paper.
  delta_y_t = array_ops.reverse_v2(normed_hist_true, [0], name='delta_y_t')
  delta_x_t = array_ops.reverse_v2(normed_hist_false, [0], name='delta_x_t')

  # strict_1d_cumsum requires float32 args.
  delta_y_t = math_ops.cast(delta_y_t, dtypes.float32)
  delta_x_t = math_ops.cast(delta_x_t, dtypes.float32)

  # Trapezoidal integration, \int_0^1 0.5 * (y_t + y_{t-1}) dx_t
  y_t = _strict_1d_cumsum(delta_y_t, nbins)
  first_trap = delta_x_t[0] * y_t[0] / 2.0
  other_traps = delta_x_t[1:] * (y_t[1:] + y_t[:nbins - 1]) / 2.0
  return math_ops.add(first_trap, math_ops.reduce_sum(other_traps), name='auc')


# TODO(langmore) Remove once a faster cumsum (accumulate_sum) Op is available.
# Also see if cast to float32 above can be removed with new cumsum.
# See:  https://github.com/tensorflow/tensorflow/issues/813