tf.function

View source on GitHub

Compiles a function into a callable TensorFlow graph. (deprecated arguments) (deprecated arguments)

tf.function(
    func=None,
    input_signature=None,
    autograph=True,
    jit_compile=None,
    reduce_retracing=False,
    experimental_implements=None,
    experimental_autograph_options=None,
    experimental_relax_shapes=None,
    experimental_compile=None,
    experimental_follow_type_hints=None
) -> tf.types.experimental.GenericFunction

tf.function constructs a tf.types.experimental.GenericFunction that executes a TensorFlow graph (tf.Graph) created by trace-compiling the TensorFlow operations in func. More information on the topic can be found in Introduction to Graphs and tf.function.

See Better Performance with tf.function for tips on performance and known limitations.

Example usage:

@tf.function
def f(x, y):
  return x ** 2 + y
x = tf.constant([2, 3])
y = tf.constant([3, -2])
f(x, y)
<tf.Tensor: ... numpy=array([7, 7], ...)>

The trace-compilation allows non-TensorFlow operations to execute, but under special conditions. In general, only TensorFlow operations are guaranteed to run and create fresh results whenever the GenericFunction is called.

Features

func may use data-dependent Python control flow statements, including if, for, while break, continue and return:

@tf.function
def f(x):
  if tf.reduce_sum(x) > 0:
    return x * x
  else:
    return -x // 2
f(tf.constant(-2))
<tf.Tensor: ... numpy=1>

func's closure may include tf.Tensor and tf.Variable objects:

@tf.function
def f():
  return x ** 2 + y
x = tf.constant([-2, -3])
y = tf.Variable([3, -2])
f()
<tf.Tensor: ... numpy=array([7, 7], ...)>

func may also use ops with side effects, such as tf.print, tf.Variable and others:

v = tf.Variable(1)
@tf.function
def f(x):
  for i in tf.range(x):
    v.assign_add(i)
f(3)
v
<tf.Variable ... numpy=4>

l = []
@tf.function
def f(x):
  for i in x:
    l.append(i + 1)    # Caution! Will only happen once when tracing
f(tf.constant([1, 2, 3]))
l
[<tf.Tensor ...>]

Instead, use TensorFlow collections like tf.TensorArray:

@tf.function
def f(x):
  ta = tf.TensorArray(dtype=tf.int32, size=0, dynamic_size=True)
  for i in range(len(x)):
    ta = ta.write(i, x[i] + 1)
  return ta.stack()
f(tf.constant([1, 2, 3]))
<tf.Tensor: ..., numpy=array([2, 3, 4], ...)>

tf.function creates polymorphic callables

Internally, tf.types.experimental.GenericFunction may contain multiple tf.types.experimental.ConcreteFunctions, each specialized to arguments with different data types or shapes, since TensorFlow can perform more optimizations on graphs of specific shapes, dtypes and values of constant arguments. tf.function treats any pure Python values as opaque objects (best thought of as compile-time constants), and builds a separate tf.Graph for each set of Python arguments that it encounters. For more information, see the tf.function guide

Executing a GenericFunction will select and execute the appropriate ConcreteFunction based on the argument types and values.

To obtain an individual ConcreteFunction, use the GenericFunction.get_concrete_function method. It can be called with the same arguments as func and returns a tf.types.experimental.ConcreteFunction. ConcreteFunctions are backed by a single tf.Graph:

@tf.function
def f(x):
  return x + 1
isinstance(f.get_concrete_function(1).graph, tf.Graph)
True

ConcreteFunctions can be executed just like GenericFunctions, but their input is resticted to the types to which they're specialized.

Retracing

ConcreteFunctions are built (traced) on the fly, as the GenericFunction is called with new TensorFlow types or shapes, or with new Python values as arguments. When GenericFunction builds a new trace, it is said that func is retraced. Retracing is a frequent performance concern for tf.function as it can be considerably slower than executing a graph that's already been traced. It is ideal to minimize the amount of retracing in your code.

@tf.function
def f(x):
  return tf.abs(x)
f1 = f.get_concrete_function(1)
f2 = f.get_concrete_function(2)  # Slow - compiles new graph
f1 is f2
False
f1 = f.get_concrete_function(tf.constant(1))
f2 = f.get_concrete_function(tf.constant(2))  # Fast - reuses f1
f1 is f2
True

Python numerical arguments should only be used when they take few distinct values, such as hyperparameters like the number of layers in a neural network.

Input signatures

For Tensor arguments, GenericFunctioncreates a new ConcreteFunction for every unique set of input shapes and datatypes. The example below creates two separate ConcreteFunctions, each specialized to a different shape:

@tf.function
def f(x):
  return x + 1
vector = tf.constant([1.0, 1.0])
matrix = tf.constant([[3.0]])
f.get_concrete_function(vector) is f.get_concrete_function(matrix)
False

An "input signature" can be optionally provided to tf.function to control this process. The input signature specifies the shape and type of each Tensor argument to the function using a tf.TensorSpec object. More general shapes can be used. This ensures only one ConcreteFunction is created, and restricts the GenericFunction to the specified shapes and types. It is an effective way to limit retracing when Tensors have dynamic shapes.

@tf.function(
    input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
def f(x):
  return x + 1
vector = tf.constant([1.0, 1.0])
matrix = tf.constant([[3.0]])
f.get_concrete_function(vector) is f.get_concrete_function(matrix)
True

Variables may only be created once

tf.function only allows creating new tf.Variable objects when it is called for the first time:

class MyModule(tf.Module):
  def __init__(self):
    self.v = None

  @tf.function
  def __call__(self, x):
    if self.v is None:
      self.v = tf.Variable(tf.ones_like(x))
    return self.v * x

In general, it is recommended to create tf.Variables outside of tf.function. In simple cases, persisting state across tf.function boundaries may be implemented using a pure functional style in which state is represented by tf.Tensors passed as arguments and returned as return values.

Contrast the two styles below:

state = tf.Variable(1)
@tf.function
def f(x):
  state.assign_add(x)
f(tf.constant(2))  # Non-pure functional style
state
<tf.Variable ... numpy=3>

state = tf.constant(1)
@tf.function
def f(state, x):
  state += x
  return state
state = f(state, tf.constant(2))  # Pure functional style
state
<tf.Tensor: ... numpy=3>

Python operations execute only once per trace

func may contain TensorFlow operations mixed with pure Python operations. However, when the function is executed, only the TensorFlow operations will run. The Python operations run only once, at trace time. If TensorFlow operations depend on results from Python operations, those results will be frozen into the graph.

@tf.function
def f(a, b):
  print('this runs at trace time; a is', a, 'and b is', b)
  return b
f(1, tf.constant(1))
this runs at trace time; a is 1 and b is Tensor("...", shape=(), dtype=int32)
<tf.Tensor: shape=(), dtype=int32, numpy=1>

f(1, tf.constant(2))
<tf.Tensor: shape=(), dtype=int32, numpy=2>

f(2, tf.constant(1))
this runs at trace time; a is 2 and b is Tensor("...", shape=(), dtype=int32)
<tf.Tensor: shape=(), dtype=int32, numpy=1>

f(2, tf.constant(2))
<tf.Tensor: shape=(), dtype=int32, numpy=2>

Using type annotations to improve performance

experimental_follow_type_hints can be used along with type annotations to reduce retracing by automatically casting any Python values to tf.Tensor (something that is not done by default, unless you use input signatures).

@tf.function(experimental_follow_type_hints=True)
def f_with_hints(x: tf.Tensor):
  print('Tracing')
  return x
@tf.function(experimental_follow_type_hints=False)
def f_no_hints(x: tf.Tensor):
  print('Tracing')
  return x
f_no_hints(1)
Tracing
<tf.Tensor: shape=(), dtype=int32, numpy=1>
f_no_hints(2)
Tracing
<tf.Tensor: shape=(), dtype=int32, numpy=2>
f_with_hints(1)
Tracing
<tf.Tensor: shape=(), dtype=int32, numpy=1>
f_with_hints(2)
<tf.Tensor: shape=(), dtype=int32, numpy=2>