Function streaming_concat

tensorflow/contrib/metrics/python/ops/metric_ops.py:3558–3669

Concatenate values along an axis across batches. The function `streaming_concat` creates two local variables, `array` and `size`, that are used to store concatenated values. Internally, `array` is used as storage for a dynamic array (if `maxsize` is `None`), which ensures that updates can b

(values,
                     axis=0,
                     max_size=None,
                     metrics_collections=None,
                     updates_collections=None,
                     name=None)

def streaming_concat(values,
                     axis=0,
                     max_size=None,
                     metrics_collections=None,
                     updates_collections=None,
                     name=None):
  """Concatenate values along an axis across batches.

  The function `streaming_concat` creates two local variables, `array` and
  `size`, that are used to store concatenated values. Internally, `array` is
  used as storage for a dynamic array (if `maxsize` is `None`), which ensures
  that updates can be run in amortized constant time.

  For estimation of the metric over a stream of data, the function creates an
  `update_op` operation that appends the values of a tensor and returns the
  length of the concatenated axis.

  This op allows for evaluating metrics that cannot be updated incrementally
  using the same framework as other streaming metrics.

  Args:
    values: `Tensor` to concatenate. Rank and the shape along all axes other
      than the axis to concatenate along must be statically known.
    axis: optional integer axis to concatenate along.
    max_size: optional integer maximum size of `value` along the given axis.
      Once the maximum size is reached, further updates are no-ops. By default,
      there is no maximum size: the array is resized as necessary.
    metrics_collections: An optional list of collections that `value` should be
      added to.
    updates_collections: An optional list of collections `update_op` should be
      added to.
    name: An optional variable_scope name.

  Returns:
    value: A `Tensor` representing the concatenated values.
    update_op: An operation that concatenates the next values.

  Raises:
    ValueError: if `values` does not have a statically known rank, `axis` is
      not in the valid range or the size of `values` is not statically known
      along any axis other than `axis`.
  """
  with variable_scope.variable_scope(name, 'streaming_concat', (values,)):
    # pylint: disable=invalid-slice-index
    values_shape = values.get_shape()
    if values_shape.dims is None:
      raise ValueError('`values` must have known statically known rank')

    ndim = len(values_shape)
    if axis < 0:
      axis += ndim
    if not 0 <= axis < ndim:
      raise ValueError('axis = %r not in [0, %r)' % (axis, ndim))

    fixed_shape = [dim.value for n, dim in enumerate(values_shape) if n != axis]
    if any(value is None for value in fixed_shape):
      raise ValueError('all dimensions of `values` other than the dimension to '
                       'concatenate along must have statically known size')