# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Type specifications for TensorFlow APIs."""
import abc
import functools
import re
from typing import List, Optional, Sequence, Any
import warnings
import numpy as np
from tensorflow.python.framework import composite_tensor
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import tensor_shape
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.types import trace
from tensorflow.python.util import _pywrap_utils
from tensorflow.python.util import compat
from tensorflow.python.util import deprecation
from tensorflow.python.util import nest
from tensorflow.python.util import tf_decorator
from tensorflow.python.util.lazy_loader import LazyLoader
from tensorflow.python.util.tf_export import tf_export
# Use LazyLoader to avoid circular dependencies.
tensor_spec = LazyLoader(
"tensor_spec", globals(),
"tensorflow.python.framework.tensor_spec")
ops = LazyLoader("ops", globals(),
"tensorflow.python.framework.ops")
[文档]@tf_export("TypeSpec", v1=["TypeSpec", "data.experimental.Structure"])
class TypeSpec(trace.TraceType, metaclass=abc.ABCMeta):
"""Specifies a TensorFlow value type.
A `tf.TypeSpec` provides metadata describing an object accepted or returned
by TensorFlow APIs. Concrete subclasses, such as `tf.TensorSpec` and
`tf.RaggedTensorSpec`, are used to describe different value types.
For example, `tf.function`'s `input_signature` argument accepts a list
(or nested structure) of `TypeSpec`s.
Creating new subclasses of `TypeSpec` (outside of TensorFlow core) is not
currently supported. In particular, we may make breaking changes to the
private methods and properties defined by this base class.
Example:
>>> spec = tf.RaggedTensorSpec(shape=[None, None], dtype=tf.int32)
>>> @tf.function(input_signature=[spec])
... def double(x):
... return x * 2
>>> print(double(tf.ragged.constant([[1, 2], [3]])))
<tf.RaggedTensor [[2, 4], [6]]>
"""
# === Subclassing ===
#
# Each `TypeSpec` subclass must define:
#
# * A "component encoding" for values.
# * A "serialization" for types.
#
# The component encoding for a value is a nested structure of `tf.Tensor`
# or `CompositeTensor` that can be used by the `TypeSpec` to reconstruct
# the value. Each individual `TypeSpec` must use the same nested structure
# for all values -- this structure is defined by the `component_specs`
# attribute. Decomposing values into components, and reconstructing them
# from those components, should be inexpensive. In particular, it should
# *not* require any TensorFlow ops.
#
# The serialization for a `TypeSpec` is a nested tuple of values that can
# be used to reconstruct the `TypeSpec`. See the documentation for
# `_serialize()` for more information.
__slots__ = []
@abc.abstractproperty
def value_type(self):
"""The Python type for values that are compatible with this TypeSpec.
In particular, all values that are compatible with this TypeSpec must be an
instance of this type.
"""
raise NotImplementedError("%s.value_type" % type(self).__name__)
[文档] def is_subtype_of(self, other: trace.TraceType) -> bool:
"""Returns True if `self` is a subtype of `other`.
Implements the tf.types.experimental.func.TraceType interface.
If not overridden by a subclass, the default behavior is to assume the
TypeSpec is covariant upon attributes that implement TraceType and
invariant upon rest of the attributes as well as the structure and type
of the TypeSpec.
Args:
other: A TraceType object.
"""
if type(self) is not type(other):
return False
is_subtype = True
def check_attribute(attribute_self, attribute_other):
nonlocal is_subtype
if not is_subtype:
return
if isinstance(attribute_self, trace.TraceType):
if not attribute_self.is_subtype_of(attribute_other):
is_subtype = False
return
else:
if attribute_self != attribute_other:
is_subtype = False
try:
# TODO(b/217959193): Replace _serialize with parameter decomposition.
nest.map_structure(check_attribute, self._serialize(),
other._serialize()) # pylint: disable=protected-access
except (ValueError, TypeError):
return False
return is_subtype
[文档] def most_specific_common_supertype(
self,
others: Sequence[trace.TraceType]) -> Optional["TypeSpec"]:
"""Returns the most specific supertype TypeSpec of `self` and `others`.
Implements the tf.types.experimental.func.TraceType interface.
If not overridden by a subclass, the default behavior is to assume the
TypeSpec is covariant upon attributes that implement TraceType and
invariant upon rest of the attributes as well as the structure and type
of the TypeSpec.
Args:
others: A sequence of TraceTypes.
"""
if any(type(self) is not type(other) for other in others):
return None
has_supertype = True
def make_supertype_attribute(attribute_self, *attribute_others):
nonlocal has_supertype
if not has_supertype:
return
if isinstance(attribute_self, trace.TraceType):
attribute_supertype = attribute_self.most_specific_common_supertype(
attribute_others)
if attribute_supertype is None:
has_supertype = False
return
return attribute_supertype
else:
if not all(attribute_self == attribute_other
for attribute_other in attribute_others):
has_supertype = False
return
return attribute_self
try:
# TODO(b/217959193): Replace _serialize with parameter decomposition.
serialized_supertype = nest.map_structure(
make_supertype_attribute, self._serialize(),
*(o._serialize() for o in others)) # pylint: disable=protected-access
except (ValueError, TypeError):
return None
return self._deserialize(serialized_supertype) if has_supertype else None
# TODO(b/223659753): Return the actual Tensor-based value instead of spec.
def _placeholder_value(self) -> "TypeSpec":
"""Value used for tracing a function signature with this TraceType."""
return self
# TODO(b/225058047): Reconsider semantics.
[文档] def is_compatible_with(self, spec_or_value):
"""Returns true if `spec_or_value` is compatible with this TypeSpec.
Prefer using "is_subtype_of" and "most_specific_common_supertype" wherever
possible.
Args:
spec_or_value: A TypeSpec or TypeSpec associated value to compare against.
"""
# === Subclassing ===
# If not overridden by subclasses, the default behavior is to convert
# `spec_or_value` to a `TypeSpec` (if it isn't already); and then to
# consider two `TypeSpec`s compatible if they have the same type, and
# the values returned by `_serialize` are compatible (where
# `tf.TensorShape`, `tf.TensorSpec`, and `tf.DType` are checked for
# compatibility using their `is_compatible_with` method; and all other
# types are considered compatible if they are equal).
if not isinstance(spec_or_value, TypeSpec):
spec_or_value = type_spec_from_value(spec_or_value)
if type(self) is not type(spec_or_value):
return False
return self.__is_compatible(self._serialize(), spec_or_value._serialize()) # pylint: disable=protected-access
[文档] @deprecation.deprecated(None, "Use most_specific_common_supertype instead.")
def most_specific_compatible_type(self, other: "TypeSpec") -> "TypeSpec":
"""Returns the most specific TypeSpec compatible with `self` and `other`.
Deprecated. Please use `most_specific_common_supertype` instead.
Do not override this function.
Args:
other: A `TypeSpec`.
Raises:
ValueError: If there is no TypeSpec that is compatible with both `self`
and `other`.
"""
result = self.most_specific_common_supertype([other])
if result is None:
raise ValueError("No TypeSpec is compatible with both %s and %s" %
(self, other))
return result
# TODO(b/226395276): Delete after removing usages.
def _with_tensor_ranks_only(self) -> "TypeSpec":
"""Returns a TypeSpec compatible with `self`, with tensor shapes relaxed.
Returns:
A `TypeSpec` that is compatible with `self`, where any `TensorShape`
information has been relaxed to include only tensor rank (and not
the dimension sizes for individual axes).
"""
# === Subclassing ===
# If not overridden by a subclass, the default behavior is to serialize
# this TypeSpec, relax any TensorSpec or TensorShape values, and
# deserialize the result.
def relax(value):
if isinstance(value, TypeSpec):
return value._with_tensor_ranks_only() # pylint: disable=protected-access
elif (isinstance(value, tensor_shape.TensorShape) and
value.rank is not None):
return tensor_shape.TensorShape([None] * value.rank)
else:
return value
return self._deserialize(nest.map_structure(relax, self._serialize()))
# TODO(b/206014848): Helper function to support logic that does not consider
# Tensor name. Will be removed once load-bearing usages of Tensor name are
# fixed.
def _without_tensor_names(self) -> "TypeSpec":
"""Returns a TypeSpec compatible with `self`, with tensor names removed.
Returns:
A `TypeSpec` that is compatible with `self`, where the name of any
`TensorSpec` is set to `None`.
"""
# === Subclassing ===
# If not overridden by a subclass, the default behavior is to serialize
# this TypeSpec, set the TensorSpecs' names to None, and deserialize the
# result.
def rename(value):
if isinstance(value, TypeSpec):
return value._without_tensor_names() # pylint: disable=protected-access
return value
return self._deserialize(nest.map_structure(rename, self._serialize()))
# === Component encoding for values ===
@abc.abstractmethod
def _to_components(self, value):
"""Encodes `value` as a nested structure of `Tensor` or `CompositeTensor`.
Args:
value: A value compatible with this `TypeSpec`. (Caller is responsible
for ensuring compatibility.)
Returns:
A nested structure of `tf.Tensor` or `tf.CompositeTensor` compatible with
`self._component_specs`, which can be used to reconstruct `value`.
"""
# === Subclassing ===
# This method must be inexpensive (do not call TF ops).
raise NotImplementedError("%s._to_components()" % type(self).__name__)
@abc.abstractmethod
def _from_components(self, components):
"""Reconstructs a value from a nested structure of Tensor/CompositeTensor.
Args:
components: A nested structure of `tf.Tensor` or `tf.CompositeTensor`,
compatible with `self._component_specs`. (Caller is responsible for
ensuring compatibility.)
Returns:
A value that is compatible with this `TypeSpec`.
"""
# === Subclassing ===
# This method must be inexpensive (do not call TF ops).
raise NotImplementedError("%s._from_components()" % type(self).__name__)
@abc.abstractproperty
def _component_specs(self):
"""A nested structure of TypeSpecs for this type's components.
Returns:
A nested structure describing the component encodings that are returned
by this TypeSpec's `_to_components` method. In particular, for a
TypeSpec `spec` and a compatible value `value`:
```
nest.map_structure(lambda t, c: assert t.is_compatible_with(c),
spec._component_specs, spec._to_components(value))
```
"""
raise NotImplementedError("%s._component_specs()" % type(self).__name__)
# === Tensor list encoding for values ===
def _to_tensor_list(self, value) -> List["ops.Tensor"]:
"""Encodes `value` as a flat list of `tf.Tensor`.
By default, this just flattens `self._to_components(value)` using
`nest.flatten`. However, subclasses may override this to return a
different tensor encoding for values. In particular, some subclasses
of `BatchableTypeSpec` override this method to return a "boxed" encoding
for values, which then can be batched or unbatched. See
`BatchableTypeSpec` for more details.
Args:
value: A value with compatible this `TypeSpec`. (Caller is responsible
for ensuring compatibility.)
Returns:
A list of `tf.Tensor`, compatible with `self._flat_tensor_specs`, which
can be used to reconstruct `value`.
"""
return nest.flatten(self._to_components(value), expand_composites=True)
def _from_tensor_list(self, tensor_list: List["ops.Tensor"]) -> Any:
"""Reconstructs a value from a flat list of `tf.Tensor`.
Args:
tensor_list: A flat list of `tf.Tensor`, compatible with
`self._flat_tensor_specs`.
Returns:
A value that is compatible with this `TypeSpec`.
Raises:
ValueError: If `tensor_list` is not compatible with
`self._flat_tensor_specs`.
"""
self.__check_tensor_list(tensor_list)
return self._from_compatible_tensor_list(tensor_list)
def _from_compatible_tensor_list(
self, tensor_list: List["ops.Tensor"]) -> Any:
"""Reconstructs a value from a compatible flat list of `tf.Tensor`.
Args:
tensor_list: A flat list of `tf.Tensor`, compatible with
`self._flat_tensor_specs`. (Caller is responsible for ensuring
compatibility.)
Returns:
A value that is compatible with this `TypeSpec`.
"""
return self._from_components(
nest.pack_sequence_as(
self._component_specs, tensor_list, expand_composites=True))
@property
def _flat_tensor_specs(self):
"""A list of TensorSpecs compatible with self._to_tensor_list(v)."""
return nest.flatten(self._component_specs, expand_composites=True)
# === Serialization for types ===
@abc.abstractmethod
def _serialize(self):
"""Returns a nested tuple containing the state of this TypeSpec.
The serialization may contain the following value types: boolean,
integer, string, float, None, `TensorSpec`, `tf.TensorShape`, `tf.DType`,
`np.ndarray`, `TypeSpec`, and nested tuples, namedtuples, dicts, and
OrderedDicts of any of the above.
This method is used to provide default definitions for: equality
testing (__eq__, __ne__), hashing (__hash__), pickling (__reduce__),
string representation (__repr__), `self.is_compatible_with()`,
`self.most_specific_compatible_type()`, and protobuf serialization
(e.g. TensorInfo and StructuredValue).
"""
raise NotImplementedError("%s._serialize()" % type(self).__name__)
@classmethod
def _deserialize(cls, serialization):
"""Reconstructs a TypeSpec from a value returned by `serialize`.
Args:
serialization: A value returned by _serialize. In some contexts,
`namedtuple`s in `serialization` may not have the identical type that
was returned by `_serialize` (but its type will still be a `namedtuple`
type with the same type name and field names). For example, the code
that loads a SavedModel does not have access to the original
`namedtuple` type, so it dynamically creates a new `namedtuple` type
with the same type name and field names as the original one. If
necessary, you can check `serialization` for these duck-typed
`nametuple` types, and restore them to the original type. (E.g., this
would be necessary if you rely on type checks such as `isinstance` for
this `TypeSpec`'s member variables).
Returns:
A `TypeSpec` of type `cls`.
"""
return cls(*serialization) # pytype: disable=not-instantiable # trace-all-classes
# === Operators ===
def __eq__(self, other) -> bool:
# pylint: disable=protected-access
return (type(other) is type(self) and
self.__get_cmp_key() == other.__get_cmp_key())
def __ne__(self, other) -> bool:
return not self == other
def __hash__(self) -> int:
return hash(self.__get_cmp_key())
def __reduce__(self):
return type(self), self._serialize()
def __repr__(self) -> str:
return "%s%r" % (type(self).__name__, self._serialize())
# === Legacy Output ===
# TODO(b/133606651) Document and/or deprecate the legacy_output methods.
# (These are used by tf.data.)
def _to_legacy_output_types(self):
raise NotImplementedError("%s._to_legacy_output_types()" %
type(self).__name__)
def _to_legacy_output_shapes(self):
raise NotImplementedError("%s._to_legacy_output_shapes()" %
type(self).__name__)
def _to_legacy_output_classes(self):
return self.value_type
# === Private Helper Methods ===
# TODO(b/154541175): Currently this usage is used to represent a Tensor
# argument not a TensorSpec argument as it should be.
def __tf_tracing_type__(self,
context: trace.TracingContext) -> trace.TraceType:
if context.include_tensor_ranks_only:
return self._with_tensor_ranks_only()
else:
return self
def __check_tensor_list(self, tensor_list):
"""Raises an exception if tensor_list incompatible w/ flat_tensor_specs."""
expected = self._flat_tensor_specs
specs = [type_spec_from_value(t) for t in tensor_list]
if len(specs) != len(expected):
raise ValueError(f"Cannot create a {self.value_type.__name__} from the "
f"tensor list because the TypeSpec expects "
f"{len(expected)} items, but the provided tensor list "
f"has {len(specs)} items.")
for i, (s1, s2) in enumerate(zip(specs, expected)):
if not s1.is_compatible_with(s2):
raise ValueError(f"Cannot create a {self.value_type.__name__} from the "
f"tensor list because item {i} ({tensor_list[i]!r}) "
f"is incompatible with the expected TypeSpec {s2}.")
def __get_cmp_key(self):
"""Returns a hashable eq-comparable key for `self`."""
# TODO(b/133606651): Decide whether to cache this value.
return (type(self), self.__make_cmp_key(self._serialize()))
def __make_cmp_key(self, value):
"""Converts `value` to a hashable key."""
if isinstance(value, (int, float, bool, np.generic, dtypes.DType, TypeSpec,
tensor_shape.TensorShape)):
return value
if isinstance(value, compat.bytes_or_text_types):
return value
if value is None:
return value
if isinstance(value, dict):
return tuple([
tuple([self.__make_cmp_key(key),
self.__make_cmp_key(value[key])])
for key in sorted(value.keys())
])
if isinstance(value, tuple):
return tuple([self.__make_cmp_key(v) for v in value])
if isinstance(value, list):
return (list, tuple([self.__make_cmp_key(v) for v in value]))
if isinstance(value, np.ndarray):
return (np.ndarray, value.shape,
TypeSpec.__nested_list_to_tuple(value.tolist()))
raise ValueError(f"Cannot generate a hashable key for {self} because "
f"the _serialize() method "
f"returned an unsupproted value of type {type(value)}")
@staticmethod
def __nested_list_to_tuple(value):
"""Converts a nested list to a corresponding nested tuple."""
if isinstance(value, list):
return tuple(TypeSpec.__nested_list_to_tuple(v) for v in value)
return value
@staticmethod
def __same_types(a, b):
"""Returns whether a and b have the same type, up to namedtuple equivalence.
Consistent with tf.nest.assert_same_structure(), two namedtuple types
are considered the same iff they agree in their class name (without
qualification by module name) and in their sequence of field names.
This makes namedtuples recreated by nested_structure_coder compatible with
their original Python definition.
Args:
a: a Python object.
b: a Python object.
Returns:
A boolean that is true iff type(a) and type(b) are the same object
or equivalent namedtuple types.
"""
if nest.is_namedtuple(a) and nest.is_namedtuple(b):
return nest.same_namedtuples(a, b)
else:
return type(a) is type(b)
@staticmethod
def __is_compatible(a, b):
"""Returns true if the given type serializations compatible."""
if isinstance(a, TypeSpec):
return a.is_compatible_with(b)
if not TypeSpec.__same_types(a, b):
return False
if isinstance(a, (list, tuple)):
return (len(a) == len(b) and
all(TypeSpec.__is_compatible(x, y) for (x, y) in zip(a, b)))
if isinstance(a, dict):
return (len(a) == len(b) and sorted(a.keys()) == sorted(b.keys()) and
all(TypeSpec.__is_compatible(a[k], b[k]) for k in a.keys()))
if isinstance(a, (tensor_shape.TensorShape, dtypes.DType)):
return a.is_compatible_with(b)
return a == b
class TypeSpecBatchEncoder(object, metaclass=abc.ABCMeta):
"""Class used to encode and decode composite tensor values for batching.
In order to be batched and unbatched by APIs such as `tf.data.Dataset` and
`tf.map_fn`, composite tensors must be encoded using flat tensors that can
themselves be batched or unbatched. `TypeSpecBatchEncoder`s are
responsible for implementing this encoding.
If a composite tensor's shape is a prefix of the shape of all of its
component tensors, then this encoding can usually be performed by just
returning those component tensors as a list. But if the composite tensor
has components whose shape has a more complex relationship to the shape
of the composite tensor, then a custom `TypeSpecBatchEncoder` may
need to be implemented.
"""
@abc.abstractmethod
def batch(self, spec, batch_size):
"""Returns the TypeSpec representing a batch of values described by `spec`.
Args:
spec: The `TypeSpec` for an individual value.
batch_size: An `int` indicating the number of values that are batched
together, or `None` if the batch size is not known.
Returns:
A `TypeSpec` for a batch of values.
"""
raise NotImplementedError(f"{type(self).__name__}.batch")
@abc.abstractmethod
def unbatch(self, spec):
"""Returns the TypeSpec for a single unbatched element in `spec`.
Args:
spec: The `TypeSpec` for a batch of values.
Returns:
A `TypeSpec` for an individual value.
"""
raise NotImplementedError(f"{type(self).__name__}.unbatch")
@abc.abstractmethod
def encode(self, spec, value, minimum_rank=0):
"""Encodes `value` as a nest of batchable `Tensor` or `CompositeTensor`.
Args:
spec: The TypeSpec of the value to encode.
value: A value compatible with `spec`.
minimum_rank: The minimum rank for the returned Tensors, CompositeTensors,
and ExtensionType values. This can be used to ensure that the encoded
values can be unbatched this number of times. If `minimum_rank>0`,
then `t.shape[:minimum_rank]` must be compatible for all values `t`
returned by `encode`.
Returns:
A nest (as defined by `tf.nest`) of `tf.Tensor`s, batchable
`tf.CompositeTensor`s, or `tf.ExtensionType`s. Stacking, unstacking, or
concatenating these encoded values and then decoding the result must be
equivalent to stacking, unstacking, or concatenating the original values.
"""
raise NotImplementedError(f"{type(self).__name__}.encode")
@abc.abstractmethod
def decode(self, spec, encoded_value):
"""Decodes `value` from a batchable tensor encoding.
Args:
spec: The TypeSpec for the result value. If encoded values with spec `s`
were batched, then `spec` should be `s.batch(batch_size)`; or if encoded
values with spec `s` were unbatched, then `spec` should be
`s.unbatch()`.
encoded_value: A nest of values returned by `encode`; or a nest of
values that was formed by stacking, unstacking, or concatenating the
corresponding elements of values returned by `encode`.
Returns:
A value compatible with `type_spec`.
"""
raise NotImplementedError(f"{type(self).__name__}.decode")
@abc.abstractmethod
def encoding_specs(self, spec):
"""Returns a nest of `TypeSpec`(s) describing the encoding for `spec`.
Args:
spec: The TypeSpec whose encoding should be described.
Returns:
A nest (as defined by `tf.nest) of `tf.TypeSpec`, describing the values
that are returned by `self.encode(spec, ...)`. All TypeSpecs in this
nest must be batchable.
"""
raise NotImplementedError(f"{type(self).__name__}.encoding_specs")
class LegacyTypeSpecBatchEncoder(TypeSpecBatchEncoder):
"""TypeSpecBatchEncoder for legacy composite tensor classes.
TODO(edloper): Update existing composite tensors to use non-legacy
CompositTensorBatchEncoders.
"""
def batch(self, type_spec, batch_size):
return type_spec._batch(batch_size) # pylint: disable=protected-access
def unbatch(self, type_spec):
return type_spec._unbatch() # pylint: disable=protected-access
def encode(self, type_spec, value, minimum_rank=0):
if minimum_rank == 0:
return type_spec._to_tensor_list(value) # pylint: disable=protected-access
elif minimum_rank == 1:
if not isinstance(type_spec, BatchableTypeSpec):
raise ValueError(f"{type_spec.__name__}.encode does not support "
"minimum_rank>0.")
return type_spec._to_batched_tensor_list(value) # pylint: disable=protected-access
else:
raise ValueError(f"{type_spec.__name__}.encode does not support "
"minimum_rank>1.")
def decode(self, type_spec, encoded_value):
return type_spec._from_tensor_list(encoded_value) # pylint: disable=protected-access
def encoding_specs(self, spec):
return spec._flat_tensor_specs # pylint: disable=protected-access
class BatchableTypeSpec(TypeSpec, metaclass=abc.ABCMeta):
"""TypeSpec with a batchable tensor encoding.
The batchable tensor encoding is a list of `tf.Tensor`s that supports
batching and unbatching. In particular, stacking (or unstacking)
values with the same `TypeSpec` must be equivalent to stacking (or
unstacking) each of their tensor lists. Unlike the component encoding
(returned by `self._to_components)`, the batchable tensor encoding
may require using encoding/decoding ops.
If a subclass's batchable tensor encoding is not simply a flattened version
of the component encoding, then the subclass must override `_to_tensor_list`,
`_from_tensor_list`, and _flat_tensor_specs`.
"""
__slots__ = []
__batch_encoder__ = LegacyTypeSpecBatchEncoder()
@abc.abstractmethod
def _batch(self, batch_size) -> TypeSpec:
"""Returns a TypeSpec representing a batch of objects with this TypeSpec.
Args:
batch_size: An `int` representing the number of elements in a batch, or
`None` if the batch size may vary.
Returns:
A `TypeSpec` representing a batch of objects with this TypeSpec.
"""
raise NotImplementedError(f"{type(self).__name__}._batch")
@abc.abstractmethod
def _unbatch(self) -> TypeSpec:
"""Returns a TypeSpec representing a single element this TypeSpec.
Returns:
A `TypeSpec` representing a single element of objects with this TypeSpec.
"""
raise NotImplementedError(f"{type(self).__name__}._unbatch")
@property
def _flat_tensor_specs(self) -> List[TypeSpec]:
"""A list of TensorSpecs compatible with self._to_tensor_list(v)."""
component_flat_tensor_specs = nest.map_structure(
functools.partial(get_batchable_flat_tensor_specs, context_spec=self),
self._component_specs)
return nest.flatten(component_flat_tensor_specs)
def _to_tensor_list(
self,
value: composite_tensor.CompositeTensor) -> List["ops.Tensor"]:
"""Encodes `value` as a flat list of `ops.Tensor`."""
component_tensor_lists = nest.map_structure(
batchable_to_tensor_list,
self._component_specs,
self._to_components(value))
return nest.flatten(component_tensor_lists)
def _to_batched_tensor_list(
self,
value: composite_tensor.CompositeTensor) -> List["ops.Tensor"]:
"""Encodes `value` as a flat list of `ops.Tensor` each with rank>0."""
get_spec_tensor_list = lambda spec, v: ( # pylint: disable=g-long-lambda
batchable_to_tensor_list(spec, v, minimum_rank=1)
if isinstance(spec, BatchableTypeSpec) else spec._to_tensor_list(v)) # pylint: disable=protected-access
component_batched_tensor_lists = nest.map_structure(
get_spec_tensor_list, self._component_specs, self._to_components(value))
tensor_list = nest.flatten(component_batched_tensor_lists)
if any(t.shape.ndims == 0 for t in tensor_list):
raise ValueError(
f"While converting {value} to a list of tensors for batching, "
f"found a scalar item which cannot be batched.")
return tensor_list
def _from_compatible_tensor_list(
self, tensor_list: List["ops.Tensor"]
) -> composite_tensor.CompositeTensor:
"""Reconstructs a value from a compatible flat list of `ops.Tensor`."""
flat_specs = nest.map_structure(
functools.partial(get_batchable_flat_tensor_specs, context_spec=self),
self._component_specs)
nested_tensor_list = nest.pack_sequence_as(flat_specs, tensor_list)
components = nest.map_structure_up_to(
self._component_specs,
batchable_from_tensor_list,
self._component_specs,
nested_tensor_list)
return self._from_components(components)
def get_batchable_flat_tensor_specs(spec, context_spec=None):
"""Returns the flat tensor specs for `spec`."""
if isinstance(spec, tensor_spec.TensorSpec):
return [spec]
elif hasattr(spec, "__batch_encoder__"):
encoding_specs = nest.map_structure(
functools.partial(get_batchable_flat_tensor_specs,
context_spec=context_spec),
spec.__batch_encoder__.encoding_specs(spec))
return nest.flatten(encoding_specs)
else:
# TODO(edloper) Fix existing CompositeTensors that permit this, and
# then turn this warning into an error.
warnings.warn(f"Batchable type {context_spec} contains non-batchable "
f"field or component with type {spec}.")
return spec._flat_tensor_specs # pylint: disable=protected-access
def batchable_to_tensor_list(spec, value, minimum_rank=0):
"""Returns a list of tensors encoding `value`, whose type is `spec`."""
if isinstance(spec, tensor_spec.TensorSpec):
return [value]
elif hasattr(spec, "__batch_encoder__"):
encoded_value = spec.__batch_encoder__.encode(spec, value, minimum_rank)
encoded_specs = spec.__batch_encoder__.encoding_specs(spec)
encoded_flats = nest.map_structure(
functools.partial(batchable_to_tensor_list, minimum_rank=minimum_rank),
encoded_specs,
encoded_value)
return nest.flatten(encoded_flats)
else:
return spec._to_tensor_list(value) # pylint: disable=protected-access
def batchable_from_tensor_list(spec, tensor_list):
"""Returns a value with type `spec` decoded from `tensor_list`."""
if isinstance(spec, tensor_spec.TensorSpec):
assert len(tensor_list) == 1
return tensor_list[0]
elif hasattr(spec, "__batch_encoder__"):
encoded_specs = spec.__batch_encoder__.encoding_specs(spec)
flat_specs = nest.map_structure(get_batchable_flat_tensor_specs,
encoded_specs)
encoded_flats = nest.pack_sequence_as(flat_specs, tensor_list)
encoded_value = nest.map_structure_up_to(
encoded_specs,
batchable_from_tensor_list,
encoded_specs,
encoded_flats)
return spec.__batch_encoder__.decode(spec, encoded_value)
else:
return spec._from_compatible_tensor_list(tensor_list) # pylint: disable=protected-access
[文档]@tf_export("type_spec_from_value")
def type_spec_from_value(value) -> TypeSpec:
"""Returns a `tf.TypeSpec` that represents the given `value`.
Examples:
>>> tf.type_spec_from_value(tf.constant([1, 2, 3]))
TensorSpec(shape=(3,), dtype=tf.int32, name=None)
>>> tf.type_spec_from_value(np.array([4.0, 5.0], np.float64))
TensorSpec(shape=(2,), dtype=tf.float64, name=None)
>>> tf.type_spec_from_value(tf.ragged.constant([[1, 2], [3, 4, 5]]))
RaggedTensorSpec(TensorShape([2, None]), tf.int32, 1, tf.int64)
>>> example_input = tf.ragged.constant([[1, 2], [3]])
>>> @tf.function(input_signature=[tf.type_spec_from_value(example_input)])
... def f(x):
... return tf.reduce_sum(x, axis=1)
Args:
value: A value that can be accepted or returned by TensorFlow APIs. Accepted
types for `value` include `tf.Tensor`, any value that can be converted to
`tf.Tensor` using `tf.convert_to_tensor`, and any subclass of
`CompositeTensor` (such as `tf.RaggedTensor`).
Returns:
A `TypeSpec` that is compatible with `value`.
Raises:
TypeError: If a TypeSpec cannot be built for `value`, because its type
is not supported.
"""
spec = _type_spec_from_value(value)
if spec is not None:
return spec
# Fallback: try converting value to a tensor.
try:
tensor = ops.convert_to_tensor(value)
spec = _type_spec_from_value(tensor)
if spec is not None:
return spec
except (ValueError, TypeError) as e:
logging.vlog(
3, "Failed to convert %r to tensor: %s" % (type(value).__name__, e))
raise TypeError(f"Could not build a TypeSpec for {value} of "
f"unsupported type {type(value)}.")
def _type_spec_from_value(value) -> TypeSpec:
"""Returns a `TypeSpec` that represents the given `value`."""
if isinstance(value, ops.Tensor):
# Note: we do not include Tensor names when constructing TypeSpecs.
return tensor_spec.TensorSpec(value.shape, value.dtype)
if isinstance(value, composite_tensor.CompositeTensor):
return value._type_spec # pylint: disable=protected-access
# If `value` is a list and all of its elements can be represented by the same
# batchable type spec, then we can represent the entire list using a single
# type spec that captures the type accurately (unlike the `convert_to_tensor`
# fallback).
if isinstance(value, list) and value:
subspecs = [_type_spec_from_value(v) for v in value]
if isinstance(subspecs[0], BatchableTypeSpec):
merged_subspec = subspecs[0].most_specific_common_supertype(subspecs[1:])
if merged_subspec is not None:
return merged_subspec._batch(len(subspecs)) # pylint: disable=protected-access
for entry in reversed(_TYPE_CONVERSION_FUNCTION_REGISTRY):
type_object, converter_fn, allow_subclass = entry
if ((type(value) is type_object) or # pylint: disable=unidiomatic-typecheck
(allow_subclass and isinstance(value, type_object))):
return converter_fn(value)
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)
_TYPE_SPEC_TO_NAME = {}
_NAME_TO_TYPE_SPEC = {}
# Regular expression for valid TypeSpec names.
_REGISTERED_NAME_RE = re.compile(r"^(\w+\.)+\w+$")
# TODO(b/173744905) tf_export this as "tf.register_type_spec". (And add a
# usage example to the docstring, once the API is public.)
#
# TODO(b/173744905) Update this decorator to apply to ExtensionType rather than
# TypeSpec (once we do refactoring to move to_components/from_components from
# TypeSpec to ExtensionType).
def register(name):
"""Decorator used to register a globally unique name for a TypeSpec subclass.
Args:
name: The name of the type spec. Must be globally unique. Must have the
form `"{project_name}.{type_name}"`. E.g. `"my_project.MyTypeSpec"`.
Returns:
A class decorator that registers the decorated class with the given name.
"""
if not isinstance(name, str):
raise TypeError("Expected `name` to be a string; got %r" % (name,))
if not _REGISTERED_NAME_RE.match(name):
raise ValueError(
"Registered name must have the form '{project_name}.{type_name}' "
"(e.g. 'my_project.MyTypeSpec'); got %r." % name)
def decorator_fn(cls):
if not (isinstance(cls, type) and issubclass(cls, TypeSpec)):
raise TypeError("Expected `cls` to be a TypeSpec; got %r" % (cls,))
if cls in _TYPE_SPEC_TO_NAME:
raise ValueError("Class %s.%s has already been registered with name %s." %
(cls.__module__, cls.__name__, _TYPE_SPEC_TO_NAME[cls]))
if name in _NAME_TO_TYPE_SPEC:
raise ValueError("Name %s has already been registered for class %s.%s." %
(name, _NAME_TO_TYPE_SPEC[name].__module__,
_NAME_TO_TYPE_SPEC[name].__name__))
_TYPE_SPEC_TO_NAME[cls] = name
_NAME_TO_TYPE_SPEC[name] = cls
return cls
return decorator_fn
# TODO(edloper) tf_export this as "tf.get_type_spec_name" (or some similar name)
def get_name(cls):
"""Returns the registered name for TypeSpec `cls`."""
if not (isinstance(cls, type) and issubclass(cls, TypeSpec)):
raise TypeError("Expected `cls` to be a TypeSpec; got %r" % (cls,))
if cls not in _TYPE_SPEC_TO_NAME:
raise ValueError("TypeSpec %s.%s has not been registered." %
(cls.__module__, cls.__name__))
return _TYPE_SPEC_TO_NAME[cls]
# TODO(edloper) tf_export this as "tf.lookup_type_spec" (or some similar name)
def lookup(name):
"""Returns the TypeSpec that has been registered with name `name`."""
if not isinstance(name, str):
raise TypeError("Expected `name` to be a string; got %r" % (name,))
if name not in _NAME_TO_TYPE_SPEC:
raise ValueError("No TypeSpec has been registered with name %r" % (name,))
return _NAME_TO_TYPE_SPEC[name]