# tensors of the SparseTensor are ([None], [None, r], [r]). NOTE: The shape
# invariant here is the shape of the SparseTensor.dense_shape property. It
# must be the shape of a vector.
if shape.ndims is not None and shape.ndims != 1:
raise ValueError("Expected a shape with 1 dimension")
rank = tensor_shape.dimension_value(shape[0])
return SparseTensorSpec(tensor_shape.unknown_shape(rank), self.dtype)
def consumers(self):
return self._consumers()
SparseTensorValue = collections.namedtuple(
"SparseTensorValue", ["indices", "values", "dense_shape"])
tf_export(v1=["SparseTensorValue"])(SparseTensorValue)
_pywrap_utils.RegisterType("SparseTensorValue", SparseTensorValue)
@tf_export("SparseTensorSpec")
class SparseTensorSpec(type_spec.BatchableTypeSpec):
"""Type specification for a `tf.sparse.SparseTensor`."""
__slots__ = ["_shape", "_dtype"]
value_type = property(lambda self: SparseTensor)
def __init__(self, shape=None, dtype=dtypes.float32):
"""Constructs a type specification for a `tf.sparse.SparseTensor`.
Args:
shape: The dense shape of the `SparseTensor`, or `None` to allow any dense
Returns:
A list of `Operation`s.
Raises:
RuntimeError: If this method is called while executing eagerly.
"""
consumers = nest.flatten([
component.consumers()
for component in nest.flatten(self, expand_composites=True)
if getattr(component, "graph", None) is not None
])
return list(set(consumers))
_pywrap_utils.RegisterType("CompositeTensor", CompositeTensor)
def replace_composites_with_components(structure):
"""Recursively replaces CompositeTensors with their components.
Args:
structure: A `nest`-compatible structure, possibly containing composite
tensors.
Returns:
A copy of `structure`, where each composite tensor has been replaced by
its components. The result will contain no composite tensors.
Note that `nest.flatten(replace_composites_with_components(structure))`
returns the same value as `nest.flatten(structure)`.
"""
def __repr__(self):
s = "BoundedTensorSpec(shape={}, dtype={}, name={}, minimum={}, maximum={})"
return s.format(self.shape, repr(self.dtype), repr(self.name),
repr(self.minimum), repr(self.maximum))
def __eq__(self, other):
tensor_spec_eq = super(BoundedTensorSpec, self).__eq__(other)
return (tensor_spec_eq and np.allclose(self.minimum, other.minimum)
and np.allclose(self.maximum, other.maximum))
def __hash__(self):
return hash((self._shape_tuple, self.dtype))
def __reduce__(self):
return BoundedTensorSpec, (self._shape, self._dtype, self._minimum,
self._maximum, self._name)
def _serialize(self):
return (self._shape, self._dtype, self._minimum, self._maximum,
self._name)
_pywrap_utils.RegisterType("TensorSpec", TensorSpec)
# Note: we do not include Tensor names when constructing TypeSpecs.
type_spec.register_type_spec_from_value_converter(
ops.Tensor, lambda tensor: TensorSpec(tensor.shape, tensor.dtype))
type_spec.register_type_spec_from_value_converter(
np.ndarray, lambda array: TensorSpec(array.shape, array.dtype))
return None
_TYPE_CONVERSION_FUNCTION_REGISTRY = []
def register_type_spec_from_value_converter(type_object,
converter_fn,
allow_subclass=False):
"""Registers a function for converting values with a given type to TypeSpecs.
If multiple registered `type_object`s match a value, then the most recent
registration takes precedence. Custom converters should not be defined for
`CompositeTensor`s; use `CompositeTensor._type_spec` instead.
Args:
type_object: A Python `type` object representing the type of values
accepted by `converter_fn`.
converter_fn: A function that takes one argument (an instance of the
type represented by `type_object`) and returns a `TypeSpec`.
allow_subclass: If true, then use `isinstance(value, type_object)` to
check for matches. If false, then use `type(value) is type_object`.
"""
_, type_object = tf_decorator.unwrap(type_object)
_TYPE_CONVERSION_FUNCTION_REGISTRY.append(
(type_object, converter_fn, allow_subclass))
_pywrap_utils.RegisterType("TypeSpec", TypeSpec)
