# Copyright 2021 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. # ============================================================================== """tf.function tracing types. See `core.PolymorphicFunction` and `core.ConcreteFunction`. `PolymorphicFunction` assigns types to call arguments, forming a signature. Function signatures are used to match arguments to `ConcreteFunction`s. For example, when a new `ConcreteFunction` is traced, it is assigned a the signature of the arguments it was traced with. Subsequent call arguments which match its signature will be dispatched to the same `ConcreteFunction`. If no `ConcreteFunction` with a matching signature is found, a new one may be traced (a process known as retracing). """ import abc from typing import Any, Iterator, List, Optional, Sequence from typing_extensions import Protocol from typing_extensions import runtime_checkable from tensorflow.python.types import core from tensorflow.python.util.tf_export import tf_export from tensorflow.tools.docs import doc_controls @tf_export("types.experimental.TraceType", v1=[]) class TraceType(metaclass=abc.ABCMeta): """Represents the type of object(s) for tf.function tracing purposes. `TraceType` is an abstract class that other classes might inherit from to provide information regarding associated class(es) for the purposes of tf.function tracing. The typing logic provided through this mechanism will be used to make decisions regarding usage of cached concrete functions and retracing. For example, if we have the following tf.function and classes: ```python @tf.function def get_mixed_flavor(fruit_a, fruit_b): return fruit_a.flavor + fruit_b.flavor class Fruit: flavor = tf.constant([0, 0]) class Apple(Fruit): flavor = tf.constant([1, 2]) class Mango(Fruit): flavor = tf.constant([3, 4]) ``` tf.function does not know when to re-use an existing concrete function in regards to the `Fruit` class so naively it retraces for every new instance. ```python get_mixed_flavor(Apple(), Mango()) # Traces a new concrete function get_mixed_flavor(Apple(), Mango()) # Traces a new concrete function again ``` However, we, as the designers of the `Fruit` class, know that each subclass has a fixed flavor and we can reuse an existing traced concrete function if it was the same subclass. Avoiding such unnecessary tracing of concrete functions can have significant performance benefits. ```python class FruitTraceType(tf.types.experimental.TraceType): def __init__(self, fruit): self.fruit_type = type(fruit) self.fruit_value = fruit def is_subtype_of(self, other): return (type(other) is FruitTraceType and self.fruit_type is other.fruit_type) def most_specific_common_supertype(self, others): return self if all(self == other for other in others) else None def placeholder_value(self, placeholder_context=None): return self.fruit_value class Fruit: def __tf_tracing_type__(self, context): return FruitTraceType(self) ``` Now if we try calling it again: ```python get_mixed_flavor(Apple(), Mango()) # Traces a new concrete function get_mixed_flavor(Apple(), Mango()) # Re-uses the traced concrete function ``` """ @abc.abstractmethod def is_subtype_of(self, other: "TraceType") -> bool: """Returns True if `self` is a subtype of `other`. For example, `tf.function` uses subtyping for dispatch: if `a.is_subtype_of(b)` is True, then an argument of `TraceType` `a` can be used as argument to a `ConcreteFunction` traced with an a `TraceType` `b`. Args: other: A TraceType object to be compared against. Example: ```python class Dimension(TraceType): def __init__(self, value: Optional[int]): self.value = value def is_subtype_of(self, other): # Either the value is the same or other has a generalized value that # can represent any specific ones. return (self.value == other.value) or (other.value is None) ``` """ @abc.abstractmethod def most_specific_common_supertype( self, others: Sequence["TraceType"] ) -> Optional["TraceType"]: """Returns the most specific supertype of `self` and `others`, if exists. The returned `TraceType` is a supertype of `self` and `others`, that is, they are all subtypes (see `is_subtype_of`) of it. It is also most specific, that is, there it has no subtype that is also a common supertype of `self` and `others`. If `self` and `others` have no common supertype, this returns `None`. Args: others: A sequence of TraceTypes. Example: ```python class Dimension(TraceType): def __init__(self, value: Optional[int]): self.value = value def most_specific_common_supertype(self, other): # Either the value is the same or other has a generalized value that # can represent any specific ones. if self.value == other.value: return self.value else: return Dimension(None) ``` """ @abc.abstractmethod def placeholder_value(self, placeholder_context) -> Any: """Creates a placeholder for tracing. tf.funcion traces with the placeholder value rather than the actual value. For example, a placeholder value can represent multiple different actual values. This means that the trace generated with that placeholder value is more general and reusable which saves expensive retracing. Args: placeholder_context: A context reserved for internal/future usage. For the `Fruit` example shared above, implementing: ```python class FruitTraceType: def placeholder_value(self, placeholder_context): return Fruit() ``` instructs tf.function to trace with the `Fruit()` objects instead of the actual `Apple()` and `Mango()` objects when it receives a call to `get_mixed_flavor(Apple(), Mango())`. For example, Tensor arguments are replaced with Tensors of similar shape and dtype, output from a tf.Placeholder op. More generally, placeholder values are the arguments of a tf.function, as seen from the function's body: ```python @tf.function def foo(x): # Here `x` is be the placeholder value ... foo(x) # Here `x` is the actual value ``` """ def to_tensors(self, value: Any) -> List[core.Tensor]: """Breaks down a value of this type into Tensors. For a TraceType instance, the number of tensors generated for corresponding value should be constant. Args: value: A value belonging to this TraceType Returns: List of Tensors. """ del value return [] def from_tensors(self, tensors: Iterator[core.Tensor]) -> Any: """Generates a value of this type from Tensors. Must use the same fixed amount of tensors as `to_tensors`. Args: tensors: An iterator from which the tensors can be pulled. Returns: A value of this type. """ del tensors return self.placeholder_value(PlaceholderContext()) def flatten(self) -> List["TraceType"]: """Returns a list of TensorSpecs corresponding to `to_tensors` values.""" return [] def cast(self, value, cast_context) -> Any: """Cast value to this type. Args: value: An input value belonging to this TraceType. cast_context: A context reserved for internal/future usage. Returns: The value casted to this TraceType. Raises: AssertionError: When _cast is not overloaded in subclass, the value is returned directly, and it should be the same to self.placeholder_value(). """ del cast_context assert value == self.placeholder_value( PlaceholderContext()), f"Can not cast {value!r} to type {self!r}" return value @abc.abstractmethod def __hash__(self) -> int: pass @abc.abstractmethod def __eq__(self, other) -> bool: pass class TracingContext(metaclass=abc.ABCMeta): """Contains information scoped to the tracing of multiple objects. `TracingContext` is a container class for flags and variables that have any kind of influence on the tracing behaviour of the class implementing the __tf_tracing_type__. This context will be shared across all __tf_tracing_type__ calls while constructing the TraceType for a particular set of objects. """ class PlaceholderContext(): """Contains context information for generating placeholders within a scope.""" class CastContext(): """Contains context info and rules for casting values to a TypeSpec.""" @runtime_checkable class SupportsTracingProtocol(Protocol): """A protocol allowing custom classes to control tf.function retracing.""" @doc_controls.doc_private @abc.abstractmethod def __tf_tracing_type__(self, context: TracingContext) -> TraceType: """Returns the tracing type of this object. The tracing type is used to build the signature of a tf.function when traced, and to match arguments with existing signatures. When a Function object is called, tf.function looks at the tracing type of the call arguments. If an existing signature of matching type exists, it will be used. Otherwise, a new function is traced, and its signature will use the tracing type of the call arguments. Args: context: a context reserved for internal/future usage. Returns: The tracing type of this object. """ # TODO(b/219556836): Direct tf_export decorator adds non-method members to the # Protocol which breaks @runtime_checkable since it does not support them. tf_export("types.experimental.SupportsTracingProtocol", v1=[]).export_constant( __name__, "SupportsTracingProtocol" )