View on TensorFlow.org | Run in Google Colab | View source on GitHub | Download notebook |

In TensorFlow 2, eager execution is turned on by default. The user interface is intuitive and flexible (running one-off operations is much easier and faster), but this can come at the expense of performance and deployability.

You can use `tf.function`

to make graphs out of your programs. It is a transformation tool that creates Python-independent dataflow graphs out of your Python code. This will help you create performant and portable models, and it is required to use `SavedModel`

.

This guide will help you conceptualize how `tf.function`

works under the hood so you can use it effectively.

The main takeaways and recommendations are:

- Debug in eager mode, then decorate with
`@tf.function`

. - Don't rely on Python side effects like object mutation or list appends.
`tf.function`

works best with TensorFlow ops; NumPy and Python calls are converted to constants.

## Setup

```
import tensorflow as tf
```

Define a helper function to demonstrate the kinds of errors you might encounter:

```
import traceback
import contextlib
# Some helper code to demonstrate the kinds of errors you might encounter.
@contextlib.contextmanager
def assert_raises(error_class):
try:
yield
except error_class as e:
print('Caught expected exception \n {}:'.format(error_class))
traceback.print_exc(limit=2)
except Exception as e:
raise e
else:
raise Exception('Expected {} to be raised but no error was raised!'.format(
error_class))
```

## Basics

### Usage

A `Function`

you define is just like a core TensorFlow operation: You can execute it eagerly; you can compute gradients; and so on.

```
@tf.function
def add(a, b):
return a + b
add(tf.ones([2, 2]), tf.ones([2, 2])) # [[2., 2.], [2., 2.]]
```

<tf.Tensor: shape=(2, 2), dtype=float32, numpy= array([[2., 2.], [2., 2.]], dtype=float32)>

```
v = tf.Variable(1.0)
with tf.GradientTape() as tape:
result = add(v, 1.0)
tape.gradient(result, v)
```

<tf.Tensor: shape=(), dtype=float32, numpy=1.0>

You can use `Function`

s inside other `Function`

s.

```
@tf.function
def dense_layer(x, w, b):
return add(tf.matmul(x, w), b)
dense_layer(tf.ones([3, 2]), tf.ones([2, 2]), tf.ones([2]))
```

<tf.Tensor: shape=(3, 2), dtype=float32, numpy= array([[3., 3.], [3., 3.], [3., 3.]], dtype=float32)>

`Function`

s can be faster than eager code, especially for graphs with many small ops. But for graphs with a few expensive ops (like convolutions), you may not see much speedup.

```
import timeit
conv_layer = tf.keras.layers.Conv2D(100, 3)
@tf.function
def conv_fn(image):
return conv_layer(image)
image = tf.zeros([1, 200, 200, 100])
# warm up
conv_layer(image); conv_fn(image)
print("Eager conv:", timeit.timeit(lambda: conv_layer(image), number=10))
print("Function conv:", timeit.timeit(lambda: conv_fn(image), number=10))
print("Note how there's not much difference in performance for convolutions")
```

Eager conv: 0.0026629049998518894 Function conv: 0.0034744160000172997 Note how there's not much difference in performance for convolutions

### Tracing

Python's dynamic typing means that you can call functions with a variety of argument types, and Python can do something different in each scenario.

Yet, to create a TensorFlow Graph, static `dtypes`

and shape dimensions are required. `tf.function`

bridges this gap by wrapping a Python function to create a `Function`

object. Based on the given inputs, the `Function`

selects the appropriate graph for the given inputs, retracing the Python function as necessary. Once you understand why and when tracing happens, it's much easier to use `tf.function`

effectively!

You can call a `Function`

with arguments of different types to see this polymorphic behavior in action.

```
@tf.function
def double(a):
print("Tracing with", a)
return a + a
print(double(tf.constant(1)))
print()
print(double(tf.constant(1.1)))
print()
print(double(tf.constant("a")))
print()
```

Tracing with Tensor("a:0", shape=(), dtype=int32) tf.Tensor(2, shape=(), dtype=int32) Tracing with Tensor("a:0", shape=(), dtype=float32) tf.Tensor(2.2, shape=(), dtype=float32) Tracing with Tensor("a:0", shape=(), dtype=string) tf.Tensor(b'aa', shape=(), dtype=string)

Note that if you repeatedly call a `Function`

with the same argument type, TensorFlow will reuse a previously traced graph, as the generated graph would be identical.

```
# This doesn't print 'Tracing with ...'
print(double(tf.constant("b")))
```

tf.Tensor(b'bb', shape=(), dtype=string)

(The following change is available in TensorFlow nightly, and will be available in TensorFlow 2.3.)

You can use `pretty_printed_concrete_signatures()`

to see all of the available traces:

```
print(double.pretty_printed_concrete_signatures())
```

double(a) Args: a: float32 Tensor, shape=() Returns: float32 Tensor, shape=() double(a) Args: a: int32 Tensor, shape=() Returns: int32 Tensor, shape=() double(a) Args: a: string Tensor, shape=() Returns: string Tensor, shape=()

So far, you've seen that `tf.function`

creates a cached, dynamic dispatch layer over TensorFlow's graph tracing logic. To be more specific about the terminology:

- A
`tf.Graph`

is the raw, language-agnostic, portable representation of your computation. - A
`ConcreteFunction`

is an eagerly-executing wrapper around a`tf.Graph`

. - A
`Function`

manages a cache of`ConcreteFunction`

s and picks the right one for your inputs. `tf.function`

wraps a Python function, returning a`Function`

object.

### Obtaining concrete functions

Every time a function is traced, a new concrete function is created. You can directly obtain a concrete function, by using `get_concrete_function`

.

```
print("Obtaining concrete trace")
double_strings = double.get_concrete_function(tf.constant("a"))
print("Executing traced function")
print(double_strings(tf.constant("a")))
print(double_strings(a=tf.constant("b")))
```

Obtaining concrete trace Executing traced function tf.Tensor(b'aa', shape=(), dtype=string) tf.Tensor(b'bb', shape=(), dtype=string)

```
# You can also call get_concrete_function on an InputSpec
double_strings_from_inputspec = double.get_concrete_function(tf.TensorSpec(shape=[], dtype=tf.string))
print(double_strings_from_inputspec(tf.constant("c")))
```

Tracing with Tensor("a:0", shape=(), dtype=string) tf.Tensor(b'cc', shape=(), dtype=string)

(The following change is available in TensorFlow nightly, and will be available in TensorFlow 2.3.)

Printing a `ConcreteFunction`

displays a summary of its input arguments (with types) and its output type.

```
print(double_strings)
```

ConcreteFunction double(a) Args: a: string Tensor, shape=() Returns: string Tensor, shape=()

You can also directly retrieve a concrete function's signature.

```
print(double_strings.structured_input_signature)
print(double_strings.structured_outputs)
```

((TensorSpec(shape=(), dtype=tf.string, name='a'),), {}) Tensor("Identity:0", shape=(), dtype=string)

Using a concrete trace with incompatible types will throw an error

```
with assert_raises(tf.errors.InvalidArgumentError):
double_strings(tf.constant(1))
```

Caught expected exception <class 'tensorflow.python.framework.errors_impl.InvalidArgumentError'>: Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-15-e4e2860a4364>", line 2, in <module> double_strings(tf.constant(1)) tensorflow.python.framework.errors_impl.InvalidArgumentError: cannot compute __inference_double_168 as input #0(zero-based) was expected to be a string tensor but is a int32 tensor [Op:__inference_double_168]

You may notice that Python arguments are given special treatment in a concrete function's input signature. Prior to TensorFlow 2.3, Python arguments were simply removed from the concrete function's signature. Starting with TensorFlow 2.3, Python arguments remain in the signature, but are constrained to take the value set during tracing.

```
@tf.function
def pow(a, b):
return a ** b
square = pow.get_concrete_function(a=tf.TensorSpec(None, tf.float32), b=2)
print(square)
```

ConcreteFunction pow(a, b=2) Args: a: float32 Tensor, shape=<unknown> Returns: float32 Tensor, shape=<unknown>

```
assert square(tf.constant(10.0)) == 100
with assert_raises(TypeError):
square(tf.constant(10.0), b=3)
```

Caught expected exception <class 'TypeError'>: Traceback (most recent call last): File "/tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/eager/function.py", line 1669, in _call_impl cancellation_manager) File "/tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/eager/function.py", line 1714, in _call_with_flat_signature self._flat_signature_summary(), ", ".join(sorted(kwargs)))) TypeError: pow(a) got unexpected keyword arguments: b. During handling of the above exception, another exception occurred: Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-17-d163f3d206cb>", line 4, in <module> square(tf.constant(10.0), b=3) TypeError: ConcreteFunction pow(a, b) was constructed with int value 2 in b, but was called with int value 3

### Obtaining graphs

Each concrete function is a callable wrapper around a `tf.Graph`

. Although retrieving the actual `tf.Graph`

object is not something you'll normally need to do, you can obtain it easily from any concrete function.

```
graph = double_strings.graph
for node in graph.as_graph_def().node:
print(f'{node.input} -> {node.name}')
```

[] -> a ['a', 'a'] -> add ['add'] -> Identity

### Debugging

In general, debugging code is easier in eager mode than inside `tf.function`

. You should ensure that your code executes error-free in eager mode before decorating with `tf.function`

. To assist in the debugging process, you can call `tf.config.run_functions_eagerly(True)`

to globally disable and reenable `tf.function`

.

When tracking down issues that only appear within `tf.function`

, here are some tips:

- Plain old Python
`print`

calls only execute during tracing, helping you track down when your function gets (re)traced. `tf.print`

calls will execute every time, and can help you track down intermediate values during execution.`tf.debugging.enable_check_numerics`

is an easy way to track down where NaNs and Inf are created.`pdb`

can help you understand what's going on during tracing. (Caveat: PDB will drop you into AutoGraph-transformed source code.)

## Tracing semantics

### Cache key rules

A `Function`

determines whether to reuse a traced concrete function by computing a cache key from an input's args and kwargs.

- The key generated for a
`tf.Tensor`

argument is its shape and dtype. - Starting in TensorFlow 2.3, the key generated for a
`tf.Variable`

argument is its`id()`

. - The key generated for a Python primitive is its value. The key generated for nested
`dict`

s,`list`

s,`tuple`

s,`namedtuple`

s, and`attr`

s is the flattened tuple. (As a result of this flattening, calling a concrete function with a different nesting structure than the one used during tracing will result in a TypeError). - For all other Python types, the keys are based on the object
`id()`

so that methods are traced independently for each instance of a class.

### Controlling retracing

Retracing helps ensures that TensorFlow generates correct graphs for each set of inputs. However, tracing is an expensive operation! If your `Function`

retraces a new graph for every call, you'll find that your code executes more slowly than if you didn't use `tf.function`

.

To control the tracing behavior, you can use the following techniques:

- Specify
`input_signature`

in`tf.function`

to limit tracing.

```
@tf.function(input_signature=(tf.TensorSpec(shape=[None], dtype=tf.int32),))
def next_collatz(x):
print("Tracing with", x)
return tf.where(x % 2 == 0, x // 2, 3 * x + 1)
print(next_collatz(tf.constant([1, 2])))
# We specified a 1-D tensor in the input signature, so this should fail.
with assert_raises(ValueError):
next_collatz(tf.constant([[1, 2], [3, 4]]))
# We specified an int32 dtype in the input signature, so this should fail.
with assert_raises(ValueError):
next_collatz(tf.constant([1.0, 2.0]))
```

Tracing with Tensor("x:0", shape=(None,), dtype=int32) tf.Tensor([4 1], shape=(2,), dtype=int32) Caught expected exception <class 'ValueError'>: Caught expected exception <class 'ValueError'>: Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-19-20f544b8adbf>", line 9, in <module> next_collatz(tf.constant([[1, 2], [3, 4]])) ValueError: Python inputs incompatible with input_signature: inputs: ( tf.Tensor( [[1 2] [3 4]], shape=(2, 2), dtype=int32)) input_signature: ( TensorSpec(shape=(None,), dtype=tf.int32, name=None)) Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-19-20f544b8adbf>", line 13, in <module> next_collatz(tf.constant([1.0, 2.0])) ValueError: Python inputs incompatible with input_signature: inputs: ( tf.Tensor([1. 2.], shape=(2,), dtype=float32)) input_signature: ( TensorSpec(shape=(None,), dtype=tf.int32, name=None))

Specify a [None] dimension in

`tf.TensorSpec`

to allow for flexibility in trace reuse.Since TensorFlow matches tensors based on their shape, using a

`None`

dimension as a wildcard will allow`Function`

s to reuse traces for variably-sized input. Variably-sized input can occur if you have sequences of different length, or images of different sizes for each batch (See Transformer and Deep Dream tutorials for example).

```
@tf.function(input_signature=(tf.TensorSpec(shape=[None], dtype=tf.int32),))
def g(x):
print('Tracing with', x)
return x
# No retrace!
print(g(tf.constant([1, 2, 3])))
print(g(tf.constant([1, 2, 3, 4, 5])))
```

Tracing with Tensor("x:0", shape=(None,), dtype=int32) tf.Tensor([1 2 3], shape=(3,), dtype=int32) tf.Tensor([1 2 3 4 5], shape=(5,), dtype=int32)

Cast Python arguments to Tensors to reduce retracing.

Often, Python arguments are used to control hyperparameters and graph constructions - for example,

`num_layers=10`

or`training=True`

or`nonlinearity='relu'`

. So if the Python argument changes, it makes sense that you'd have to retrace the graph.However, it's possible that a Python argument is not being used to control graph construction. In these cases, a change in the Python value can trigger needless retracing. Take, for example, this training loop, which AutoGraph will dynamically unroll. Despite the multiple traces, the generated graph is actually identical, so retracing is unnecessary.

```
def train_one_step():
pass
@tf.function
def train(num_steps):
print("Tracing with num_steps = ", num_steps)
tf.print("Executing with num_steps = ", num_steps)
for _ in tf.range(num_steps):
train_one_step()
print("Retracing occurs for different Python arguments.")
train(num_steps=10)
train(num_steps=20)
print()
print("Traces are reused for Tensor arguments.")
train(num_steps=tf.constant(10))
train(num_steps=tf.constant(20))
```

Retracing occurs for different Python arguments. Tracing with num_steps = 10 Executing with num_steps = 10 Tracing with num_steps = 20 Executing with num_steps = 20 Traces are reused for Tensor arguments. Tracing with num_steps = Tensor("num_steps:0", shape=(), dtype=int32) Executing with num_steps = 10 Executing with num_steps = 20

If you need to force retracing, create a new `Function`

. Separate `Function`

objects are guaranteed not to share traces.

```
def f():
print('Tracing!')
tf.print('Executing')
tf.function(f)()
tf.function(f)()
```

Tracing! Executing Tracing! Executing

### Python side effects

Python side effects like printing, appending to lists, and mutating globals only happen the first time you call a `Function`

with a set of inputs. Afterwards, the traced `tf.Graph`

is reexecuted, without executing the Python code.

The general rule of thumb is to only use Python side effects to debug your traces. Otherwise, TensorFlow ops like `tf.Variable.assign`

, `tf.print`

, and `tf.summary`

are the best way to ensure your code will be traced and executed by the TensorFlow runtime with each call.

```
@tf.function
def f(x):
print("Traced with", x)
tf.print("Executed with", x)
f(1)
f(1)
f(2)
```

Traced with 1 Executed with 1 Executed with 1 Traced with 2 Executed with 2

Many Python features, such as generators and iterators, rely on the Python runtime to keep track of state. In general, while these constructs work as expected in eager mode, many unexpected things can happen inside a `Function`

.

To give one example, advancing iterator state is a Python side effect and therefore only happens during tracing.

```
external_var = tf.Variable(0)
@tf.function
def buggy_consume_next(iterator):
external_var.assign_add(next(iterator))
tf.print("Value of external_var:", external_var)
iterator = iter([0, 1, 2, 3])
buggy_consume_next(iterator)
# This reuses the first value from the iterator, rather than consuming the next value.
buggy_consume_next(iterator)
buggy_consume_next(iterator)
```

Value of external_var: 0 Value of external_var: 0 Value of external_var: 0

Some iteration constructs are supported through AutoGraph. See the section on AutoGraph Transformations for an overview.

If you would like to execute Python code during each invocation of a `Function`

, `tf.py_function`

is an exit hatch. The drawback of `tf.py_function`

is that it's not portable or particularly performant, nor does it work well in distributed (multi-GPU, TPU) setups. Also, since `tf.py_function`

has to be wired into the graph, it casts all inputs/outputs to tensors.

APIs like `tf.gather`

, `tf.stack`

, and `tf.TensorArray`

can help you implement common looping patterns in native TensorFlow.

```
external_list = []
def side_effect(x):
print('Python side effect')
external_list.append(x)
@tf.function
def f(x):
tf.py_function(side_effect, inp=[x], Tout=[])
f(1)
f(1)
f(1)
# The list append happens all three times!
assert len(external_list) == 3
# The list contains tf.constant(1), not 1, because py_function casts everything to tensors.
assert external_list[0].numpy() == 1
```

Python side effect Python side effect Python side effect

### Variables

You may encounter an error when creating a new `tf.Variable`

in a function. This error guards against behavior divergence on repeated calls: In eager mode, a function creates a new variable with each call, but in a `Function`

, a new variable may not be created due to trace reuse.

```
@tf.function
def f(x):
v = tf.Variable(1.0)
v.assign_add(x)
return v
with assert_raises(ValueError):
f(1.0)
```

Caught expected exception <class 'ValueError'>: Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-26-73e410646579>", line 8, in <module> f(1.0) ValueError: in user code: <ipython-input-26-73e410646579>:3 f * v = tf.Variable(1.0) /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/ops/variables.py:262 __call__ ** return cls._variable_v2_call(*args, **kwargs) /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/ops/variables.py:256 _variable_v2_call shape=shape) /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/ops/variables.py:67 getter return captured_getter(captured_previous, **kwargs) /tmpfs/src/tf_docs_env/lib/python3.6/site-packages/tensorflow/python/eager/def_function.py:702 invalid_creator_scope "tf.function-decorated function tried to create " ValueError: tf.function-decorated function tried to create variables on non-first call.

You can create variables inside a `Function`

as long as those variables are only created the first time the function is executed.

```
class Count(tf.Module):
def __init__(self):
self.count = None
@tf.function
def __call__(self):
if self.count is None:
self.count = tf.Variable(0)
return self.count.assign_add(1)
c = Count()
print(c())
print(c())
```

tf.Tensor(1, shape=(), dtype=int32) tf.Tensor(2, shape=(), dtype=int32)

Another error you may encounter is a garbage-collected variable. Unlike normal Python functions, concrete functions only retain WeakRefs to the variables they close over, so you must retain a reference to any variables.

```
external_var = tf.Variable(3)
@tf.function
def f(x):
return x * external_var
traced_f = f.get_concrete_function(4)
print("Calling concrete function...")
print(traced_f(4))
del external_var
print()
print("Calling concrete function after garbage collecting its closed Variable...")
with assert_raises(tf.errors.FailedPreconditionError):
traced_f(4)
```

Calling concrete function... tf.Tensor(12, shape=(), dtype=int32) Calling concrete function after garbage collecting its closed Variable... Caught expected exception <class 'tensorflow.python.framework.errors_impl.FailedPreconditionError'>: Traceback (most recent call last): File "<ipython-input-3-73d0ca52e838>", line 8, in assert_raises yield File "<ipython-input-28-304a18524b57>", line 14, in <module> traced_f(4) tensorflow.python.framework.errors_impl.FailedPreconditionError: 2 root error(s) found. (0) Failed precondition: Error while reading resource variable _AnonymousVar4 from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/_AnonymousVar4/N10tensorflow3VarE does not exist. [[node ReadVariableOp (defined at <ipython-input-28-304a18524b57>:4) ]] (1) Failed precondition: Error while reading resource variable _AnonymousVar4 from Container: localhost. This could mean that the variable was uninitialized. Not found: Resource localhost/_AnonymousVar4/N10tensorflow3VarE does not exist. [[node ReadVariableOp (defined at <ipython-input-28-304a18524b57>:4) ]] [[ReadVariableOp/_2]] 0 successful operations. 0 derived errors ignored. [Op:__inference_f_514] Function call stack: f -> f

## AutoGraph Transformations

AutoGraph is a library that is on by default in `tf.function`

, and transforms a subset of Python eager code into graph-compatible TensorFlow ops. This includes control flow like `if`

, `for`

, `while`

.

TensorFlow ops like `tf.cond`

and `tf.while_loop`

continue to work, but control flow is often easier to write and understand when written in Python.

```
# Simple loop
@tf.function
def f(x):
while tf.reduce_sum(x) > 1:
tf.print(x)
x = tf.tanh(x)
return x
f(tf.random.uniform([5]))
```

[0.592976809 0.128855109 0.13305068 0.379838109 0.429846764] [0.532033205 0.128146663 0.132271096 0.362566859 0.405193239] [0.486933738 0.127449796 0.131505072 0.347472966 0.384383708] [0.451779395 0.126764178 0.130752221 0.334132522 0.366508394] [0.423360586 0.126089528 0.130012169 0.322229117 0.35093388] [0.399757773 0.125425547 0.129284561 0.311521083 0.337203503] [0.379741699 0.124771953 0.128569037 0.301820248 0.324978501] [0.362483114 0.124128476 0.127865285 0.292977482 0.31400153] [0.347399354 0.123494864 0.127172977 0.284872979 0.304073036] [0.334067136 0.12287087 0.1264918 0.277409106 0.295035541] [0.322170526 0.122256249 0.125821471 0.270505458 0.286762893] [0.311468184 0.12165077 0.125161693 0.264095098 0.279152662] [0.301772147 0.121054202 0.124512196 0.258121818 0.272120655] [0.292933524 0.120466337 0.123872712 0.252537966 0.265597] [0.284832567 0.119886965 0.123243 0.247302905 0.259523094] [0.277371794 0.119315878 0.122622795 0.242381677 0.253849417] <tf.Tensor: shape=(5,), dtype=float32, numpy= array([0.27047086, 0.11875288, 0.12201188, 0.23774408, 0.24853374], dtype=float32)>

If you're curious you can inspect the code autograph generates.

```
print(tf.autograph.to_code(f.python_function))
```

def tf__f(x): with ag__.FunctionScope('f', 'fscope', ag__.ConversionOptions(recursive=True, user_requested=True, optional_features=(), internal_convert_user_code=True)) as fscope: do_return = False retval_ = ag__.UndefinedReturnValue() def get_state(): return (x,) def set_state(vars_): nonlocal x (x,) = vars_ def loop_body(): nonlocal x ag__.converted_call(ag__.ld(tf).print, (ag__.ld(x),), None, fscope) x = ag__.converted_call(ag__.ld(tf).tanh, (ag__.ld(x),), None, fscope) def loop_test(): return (ag__.converted_call(ag__.ld(tf).reduce_sum, (ag__.ld(x),), None, fscope) > 1) ag__.while_stmt(loop_test, loop_body, get_state, set_state, ('x',), {}) try: do_return = True retval_ = ag__.ld(x) except: do_return = False raise return fscope.ret(retval_, do_return)

### Conditionals

AutoGraph will convert some `if <condition>`

statements into the equivalent `tf.cond`

calls. This substitution is made if `<condition>`

is a Tensor. Otherwise, the `if`

statement is executed as a Python conditional.

A Python conditional executes during tracing, so exactly one branch of the conditional will be added to the graph. Without AutoGraph, this traced graph would be unable to take the alternate branch if there is data-dependent control flow.

`tf.cond`

traces and adds both branches of the conditional to the graph, dynamically selecting a branch at execution time. Tracing can have unintended side effects; see AutoGraph tracing effects for more.

```
@tf.function
def fizzbuzz(n):
for i in tf.range(1, n + 1):
print('Tracing for loop')
if i % 15 == 0:
print('Tracing fizzbuzz branch')
tf.print('fizzbuzz')
elif i % 3 == 0:
print('Tracing fizz branch')
tf.print('fizz')
elif i % 5 == 0:
print('Tracing buzz branch')
tf.print('buzz')
else:
print('Tracing default branch')
tf.print(i)
fizzbuzz(tf.constant(5))
fizzbuzz(tf.constant(20))
```

Tracing for loop Tracing fizzbuzz branch Tracing fizz branch Tracing buzz branch Tracing default branch 1 2 fizz 4 buzz 1 2 fizz 4 buzz fizz 7 8 fizz buzz 11 fizz 13 14 fizzbuzz 16 17 fizz 19 buzz

See the reference documentation for additional restrictions on AutoGraph-converted if statements.

### Loops

AutoGraph will convert some `for`

and `while`

statements into the equivalent TensorFlow looping ops, like `tf.while_loop`

. If not converted, the `for`

or `while`

loop is executed as a Python loop.

This substitution is made in the following situations:

`for x in y`

: if`y`

is a Tensor, convert to`tf.while_loop`

. In the special case where`y`

is a`tf.data.Dataset`

, a combination of`tf.data.Dataset`

ops are generated.`while <condition>`

: if`<condition>`

is a Tensor, convert to`tf.while_loop`

.

A Python loop executes during tracing, adding additional ops to the `tf.Graph`

for every iteration of the loop.

A TensorFlow loop traces the body of the loop, and dynamically selects how many iterations to run at execution time. The loop body only appears once in the generated `tf.Graph`

.

See the reference documentation for additional restrictions on AutoGraph-converted `for`

and `while`

statements.

#### Looping over Python data

A common pitfall is to loop over Python/Numpy data within a `tf.function`

. This loop will execute during the tracing process, adding a copy of your model to the `tf.Graph`

for each iteration of the loop.

If you want to wrap the entire training loop in `tf.function`

, the safest way to do this is to wrap your data as a `tf.data.Dataset`

so that AutoGraph will dynamically unroll the training loop.

```
def measure_graph_size(f, *args):
g = f.get_concrete_function(*args).graph
print("{}({}) contains {} nodes in its graph".format(
f.__name__, ', '.join(map(str, args)), len(g.as_graph_def().node)))
@tf.function
def train(dataset):
loss = tf.constant(0)
for x, y in dataset:
loss += tf.abs(y - x) # Some dummy computation.
return loss
small_data = [(1, 1)] * 3
big_data = [(1, 1)] * 10
measure_graph_size(train, small_data)
measure_graph_size(train, big_data)
measure_graph_size(train, tf.data.Dataset.from_generator(
lambda: small_data, (tf.int32, tf.int32)))
measure_graph_size(train, tf.data.Dataset.from_generator(
lambda: big_data, (tf.int32, tf.int32)))
```

train([(1, 1), (1, 1), (1, 1)]) contains 11 nodes in its graph train([(1, 1), (1, 1), (1, 1), (1, 1), (1, 1), (1, 1), (1, 1), (1, 1), (1, 1), (1, 1)]) contains 32 nodes in its graph train(<FlatMapDataset shapes: (<unknown>, <unknown>), types: (tf.int32, tf.int32)>) contains 8 nodes in its graph train(<FlatMapDataset shapes: (<unknown>, <unknown>), types: (tf.int32, tf.int32)>) contains 8 nodes in its graph

When wrapping Python/Numpy data in a Dataset, be mindful of `tf.data.Dataset.from_generator`

versus `tf.data.Dataset.from_tensors`

. The former will keep the data in Python and fetch it via `tf.py_function`

which can have performance implications, whereas the latter will bundle a copy of the data as one large `tf.constant()`

node in the graph, which can have memory implications.

Reading data from files via TFRecordDataset/CsvDataset/etc. is the most effective way to consume data, as then TensorFlow itself can manage the asynchronous loading and prefetching of data, without having to involve Python. To learn more, see the tf.data guide.

#### Accumulating values in a loop

A common pattern is to accumulate intermediate values from a loop. Normally, this is accomplished by appending to a Python list or adding entries to a Python dictionary. However, as these are Python side effects, they will not work as expected in a dynamically unrolled loop. Use `tf.TensorArray`

to accumulate results from a dynamically unrolled loop.

```
batch_size = 2
seq_len = 3
feature_size = 4
def rnn_step(inp, state):
return inp + state
@tf.function
def dynamic_rnn(rnn_step, input_data, initial_state):
# [batch, time, features] -> [time, batch, features]
input_data = tf.transpose(input_data, [1, 0, 2])
max_seq_len = input_data.shape[0]
states = tf.TensorArray(tf.float32, size=max_seq_len)
state = initial_state
for i in tf.range(max_seq_len):
state = rnn_step(input_data[i], state)
states = states.write(i, state)
return tf.transpose(states.stack(), [1, 0, 2])
dynamic_rnn(rnn_step,
tf.random.uniform([batch_size, seq_len, feature_size]),
tf.zeros([batch_size, feature_size]))
```

<tf.Tensor: shape=(2, 3, 4), dtype=float32, numpy= array([[[0.3888433 , 0.2078135 , 0.3843341 , 0.5707482 ], [1.1050591 , 0.5968373 , 0.982028 , 0.7987355 ], [1.614423 , 0.8602725 , 1.4517052 , 1.6631885 ]], [[0.64148533, 0.67286134, 0.07972229, 0.9772469 ], [1.5103817 , 0.8244705 , 0.747108 , 1.7505025 ], [2.5020413 , 1.5157869 , 0.8045732 , 2.4830637 ]]], dtype=float32)>

## Further reading

To learn about how to export and load a `Function`

, see the SavedModel guide. To learn more about graph optimizations that are performed after tracing, see the Grappler guide. To learn how to optimize your data pipeline and profile your model, see the Profiler guide.