tf.types.experimental.TraceType

View source on GitHub

Represents the type of object(s) for tf.function tracing purposes.

Used in the notebooks

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:

@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.

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.

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:

get_mixed_flavor(Apple(), Mango()) # Traces a new concrete function
get_mixed_flavor(Apple(), Mango()) # Re-uses the traced concrete function

Methods

cast

View source

cast(
    value, cast_context
) -> Any

Cast value to this type.

flatten

View source

flatten() -> List['TraceType']

Returns a list of TensorSpecs corresponding to to_tensors values.

from_tensors

View source

from_tensors(
    tensors: Iterator[core.Tensor]
) -> Any

Generates a value of this type from Tensors.

Must use the same fixed amount of tensors as to_tensors.

is_subtype_of

View source

@abc.abstractmethod
is_subtype_of(
    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.

Example:

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)

most_specific_common_supertype

View source

@abc.abstractmethod
most_specific_common_supertype(
    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.

Example:

 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)

placeholder_value

View source

@abc.abstractmethod
placeholder_value(
    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.

For the Fruit example shared above, implementing:

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:

@tf.function
def foo(x):
  # Here `x` is be the placeholder value
  ...

foo(x) # Here `x` is the actual value

to_tensors

View source

to_tensors(
    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.

__eq__

View source

@abc.abstractmethod
__eq__(
    other
) -> bool

Return self==value.