Shortcuts

TorchScript

TorchScript is a way to create serializable and optimizable models from PyTorch code. Any code written in TorchScript can be saved from a Python process and loaded in a process where there is no Python dependency.

We provide tools to incrementally transition a model from a pure Python program to a TorchScript program that can be run independently from Python, for instance, in a standalone C++ program. This makes it possible to train models in PyTorch using familiar tools and then export the model via TorchScript to a production environment where it is not a good idea to run models as Python programs for performance and multi-threading reasons.

Creating TorchScript Code

class torch.jit.ScriptModule(optimize=True, _qualified_name=None, _compilation_unit=None)[source]

The core data structure in TorchScript is the ScriptModule. It is an analogue of torch’s nn.Module and represents an entire model as a tree of submodules. Like normal modules, each individual module in a ScriptModule can have submodules, parameters, and methods. In nn.Modules methods are implemented as Python functions, but in ScriptModules methods are implemented as TorchScript functions, a statically-typed subset of Python that contains all of PyTorch’s built-in Tensor operations. This difference allows your ScriptModules code to run without the need for a Python interpreter.

ScriptModules be created in two ways:

Tracing:

Using torch.jit.trace and torch.jit.trace_module, you can turn an existing module or Python function into a TorchScript torch._C.Function or ScriptModule. You must provide example inputs, and we run the function, recording the operations performed on all the tensors. * The resulting recording of a standalone function produces torch._C.Function. * The resulting recording of forward function of nn.Module or nn.Module produces ScriptModule. This module also contains any parameters that the original module had as well.

Example (tracing a function):

import torch
def foo(x, y):
    return 2 * x + y
traced_foo = torch.jit.trace(foo, (torch.rand(3), torch.rand(3)))

Note

Tracing a standalone function will construct a torch._C.Function Tracing nn.Module``s ``forward will construct a ScriptModule

Example (tracing an existing module):

import torch
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv = nn.Conv2d(1, 1, 3)

    def forward(self, x):
        return self.conv(x)

    def weighted_kernel_sum(self, weight):
        return weight * self.conv.weight


n = Net()
example_weight = torch.rand(1, 1, 3, 3)
example_forward_input = torch.rand(1, 1, 3, 3)

# all three trace calls below are equivalent
# and construct `ScriptModule` with a single `forward` method
module = torch.jit.trace(n.forward, example_forward_input) # produces ScriptModule with `forward`
module = torch.jit.trace(n, example_forward_input) # produces ScriptModule with `forward`
module = torch.jit.trace_module(n, inputs) # produces ScriptModule with `forward`

inputs = {'forward' : example_forward_input, 'weighted_kernel_sum' : example_weight}
# trace_module produces `ScriptModule` with two methods:
# `forward` and `weighted_kernel_sum`
module = torch.jit.trace_module(n, inputs, True, True)

Note

  • The first three trace/trace_module calls are equivalent and return ScriptModule

with a single forward method. * The last trace_module call produces a ScriptModule with two methods. Tracing only records operations done when the given function is run on the given tensors. Therefore, the returned ScriptModule will always run the same traced graph on any input. This has some important implications when your module is expected to run different sets of operations, depending on the input and/or the module state. For example,

  • Tracing will not record any control-flow like if-statements or loops. When this control-flow is constant across your module, this is fine and it often inlines the control-flow decisions. But sometimes the control-flow is actually part of the model itself. For instance, a recurrent network is a loop over the (possibly dynamic) length of an input sequence.

  • In the returned ScriptModule, operations that have different behaviors in training and eval modes will always behave as if it is in the mode it was in during tracing, no matter which mode the ScriptModule is in.

In cases like these, tracing would not be appropriate and scripting is a better choice.

Scripting:

You can write TorchScript code directly using Python syntax. You do this using the @torch.jit.script decorator (for functions) or @torch.jit.script_method decorator (for methods) on subclasses of ScriptModule. With this decorator the body of the annotated function is directly translated into TorchScript. TorchScript itself is a subset of the Python language, so not all features in Python work, but we provide enough functionality to compute on tensors and do control-dependent operations.

Example (scripting a function):

import torch
@torch.jit.script
def foo(x, y):
    if x.max() > y.max():
        r = x
    else:
        r = y
    return r

Note

A @torch.jit.script decorator will construct a ScriptModule with a single forward method that implements the function. The resulting ScriptModule has no parameters or attributes.

Example (scripting a simple module with a Parameter):

import torch
class MyModule(torch.jit.ScriptModule):
    def __init__(self, N, M):
        super(MyModule, self).__init__()
        self.weight = torch.nn.Parameter(torch.rand(N, M))

    @torch.jit.script_method
    def forward(self, input):
        return self.weight.mv(input)

Example (scripting a module with traced submodules):

import torch
import torch.nn as nn
import torch.nn.functional as F

class MyScriptModule(torch.jit.ScriptModule):
    def __init__(self):
        super(MyScriptModule, self).__init__()
        # torch.jit.trace produces a ScriptModule's conv1 and conv2
        self.conv1 = torch.jit.trace(nn.Conv2d(1, 20, 5), torch.rand(1, 1, 16, 16))
        self.conv2 = torch.jit.trace(nn.Conv2d(20, 20, 5), torch.rand(1, 20, 16, 16))

    @torch.jit.script_method
    def forward(self, input):
      input = F.relu(self.conv1(input))
      input = F.relu(self.conv2(input))
      return input
torch.jit.save(m, f, _extra_files=ExtraFilesMap{})[source]

Save an offline version of this module for use in a separate process. The saved module serializes all of the methods, submodules, parameters, and attributes of this module. It can be loaded into the C++ API using torch::jit::load(filename) or into the Python API with torch.jit.load(filename).

To be able to save a module, it must not make any calls to native Python functions. This means that all submodules must be subclasses of torch.jit.ScriptModule as well.

Danger

All modules, no matter their device, are always loaded onto the CPU during loading. This is different from torch.load()’s semantics and may change in the future.

Parameters
  • m – a ScriptModule to save

  • f – a file-like object (has to implement write and flush) or a string containing a file name

  • _extra_files – Map from filename to contents which will be stored as part of ‘f’

Warning

If you are using Python 2, torch.save does NOT support StringIO.StringIO as a valid file-like object. This is because the write method should return the number of bytes written; StringIO.write() does not do this.

Please use something like io.BytesIO instead.

Example:

m = torch.jit.ScriptModule()

# Save to file
torch.jit.save(m, 'scriptmodule.pt')

# Save to io.BytesIO buffer
buffer = io.BytesIO()
torch.jit.save(m, buffer)

# Save with extra files
extra_files = torch._C.ExtraFilesMap()
extra_files['foo.txt'] = 'bar'
torch.jit.save(m, 'scriptmodule.pt', _extra_files=extra_files)
torch.jit.load(f, map_location=None, _extra_files=ExtraFilesMap{})[source]

Load a ScriptModule previously saved with save

All previously saved modules, no matter their device, are first loaded onto CPU, and then are moved to the devices they were saved from. If this fails (e.g. because the run time system doesn’t have certain devices), an exception is raised. However, storages can be dynamically remapped to an alternative set of devices using the map_location argument. Comparing to torch.load(), map_location in this function is simplified, which only accepts a string (e.g., ‘cpu’, ‘cuda:0’), or torch.device (e.g., torch.device(‘cpu’))

Parameters
  • f – a file-like object (has to implement read, readline, tell, and seek), or a string containing a file name

  • map_location – can a string (e.g., ‘cpu’, ‘cuda:0’), a device (e.g., torch.device(‘cpu’))

  • _extra_files – map from filename to content. The extra filenames given in the map would be loaded and their content would be stored in the provided map.

Returns

A ScriptModule object.

Example:

torch.jit.load('scriptmodule.pt')

# Load ScriptModule from io.BytesIO object
with open('scriptmodule.pt', 'rb') as f:
    buffer = io.BytesIO(f.read())

# Load all tensors to the original device
torch.jit.load(buffer)

# Load all tensors onto CPU, using a device
torch.jit.load(buffer, map_location=torch.device('cpu'))

# Load all tensors onto CPU, using a string
torch.jit.load(buffer, map_location='cpu')

# Load with extra files.
files = {'metadata.json' : ''}
torch.jit.load('scriptmodule.pt', _extra_files = files)
print (files['metadata.json'])
torch.jit.trace(func, example_inputs, optimize=True, check_trace=True, check_inputs=None, check_tolerance=1e-05, _force_outplace=False, _module_class=None)[source]

Trace a function and return an executable ScriptModule or torch.jit._C.Function that will be optimized using just-in-time compilation.

Warning

Tracing only correctly records functions and modules which are not data dependent (e.g., do not have conditionals on data in tensors) and do not have any untracked external dependencies (e.g., perform input/output or access global variables). If you trace such models, you may silently get incorrect results on subsequent invocations of the model. The tracer will try to emit warnings when doing something that may cause an incorrect trace to be produced.

Parameters
  • func (callable or torch.nn.Module) – a Python function or torch.nn.Module that will be run with example_inputs. arguments and returns to func must be tensors or (possibly nested) tuples that contain tensors.

  • example_inputs (tuple) – a tuple of example inputs that will be passed to the function while tracing. The resulting trace can be run with inputs of different types and shapes assuming the traced operations support those types and shapes. example_inputs may also be a single Tensor in which case it is automatically wrapped in a tuple

Keyword Arguments
  • optimize (bool, optional) – whether or not to apply optimizations. Default: True.

  • check_trace (bool, optional) – check if the same inputs run through traced code produce the same outputs. Default: True. You might want to disable this if, for example, your network contains non- deterministic ops or if you are sure that the network is correct despite a checker failure.

  • check_inputs (list of tuples, optional) – A list of tuples of input arguments that should be used to check the trace against what is expected. Each tuple is equivalent to a set of input arguments that would be specified in example_inputs. For best results, pass in a set of checking inputs representative of the space of shapes and types of inputs you expect the network to see. If not specified, the original example_inputs are used for checking

  • check_tolerance (float, optional) – Floating-point comparison tolerance to use in the checker procedure. This can be used to relax the checker strictness in the event that results diverge numerically for a known reason, such as operator fusion.

Returns

if callable is nn.Module or forward() of nn.Module, trace returns a ScriptModule object with a single forward() method containing the traced code. The returned ScriptModule will have the same set of sub-modules and parameters as the original nn.Module. If callable is a standalone function, trace returns torch.jit._C.Function

Example:

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv = nn.Conv2d(1, 1, 3)

    def forward(self, x):
        return self.conv(x)

    def weighted_kernel_sum(self, weight):
        return weight * self.conv.weight

example_weight = torch.rand(1, 1, 3, 3)
example_forward_input = torch.rand(1, 1, 3, 3)
n = Net()
# the following two calls are equivalent
module = torch.jit.trace_module(n, example_forward_input)
module = torch.jit.trace_module(n.forward, example_forward_input)

Mixing Tracing and Scripting

In many cases either tracing or scripting is an easier approach for converting a model to TorchScript. We allow you to compose tracing and scripting to suit the particular requirements of a part of a model.

Scripted functions can call traced functions. This is particularly useful when you need to use control-flow around a simple feed-forward model. For instance the beam search of a sequence to sequence model will typically be written in script but can call an encoder module generated using tracing.

Example:

import torch

def foo(x, y):
    return 2 * x + y
traced_foo = torch.jit.trace(foo, (torch.rand(3), torch.rand(3)))

@torch.jit.script
def bar(x):
    return traced_foo(x, x)

Traced functions can call script functions. This is useful when a small part of a model requires some control-flow even though most of the model is just a feed-forward network. Control-flow inside of a script function called by a traced function is preserved correctly:

Example:

import torch

@torch.jit.script
def foo(x, y):
    if x.max() > y.max():
        r = x
    else:
        r = y
    return r


def bar(x, y, z):
    return foo(x, y) + z

traced_bar = torch.jit.trace(bar, (torch.rand(3), torch.rand(3), torch.rand(3)))

This composition also works for ScriptModules as well, where it can be used to generate a submodule using tracing that can be called from the methods of a script module:

Example:

import torch
import torchvision

class MyScriptModule(torch.jit.ScriptModule):
    def __init__(self):
        super(MyScriptModule, self).__init__()
        self.means = torch.nn.Parameter(torch.tensor([103.939, 116.779, 123.68])
                                        .resize_(1, 3, 1, 1))
        self.resnet = torch.jit.trace(torchvision.models.resnet18(),
                                      torch.rand(1, 3, 224, 224))

    @torch.jit.script_method
    def forward(self, input):
        return self.resnet(input - self.means)

TorchScript Language Reference

TorchScript is a statically typed subset of Python that can either be written directly (using the @torch.jit.script decorator) or generated automatically from Python code via tracing. When using tracing, code is automatically converted into this subset of Python by recording only the actual operators on tensors and simply executing and discarding the other surrounding Python code.

When writing TorchScript directly using @torch.jit.script decorator, the programmer must only use the subset of Python supported in TorchScript. This section documents what is supported in TorchScript as if it were a language reference for a stand alone language. Any features of Python not mentioned in this reference are not part of TorchScript.

As a subset of Python any valid TorchScript function is also a valid Python function. This makes it possible to remove the @torch.jit.script decorator and debug the function using standard Python tools like pdb. The reverse is not true: there are many valid python programs that are not valid TorchScript programs. Instead, TorchScript focuses specifically on the features of Python that are needed to represent neural network models in Torch.

PYTORCH_JIT=1

Setting the environment variable PYTORCH_JIT=0 will disable all script and tracing annotations. If there is hard-to-debug error in one of your ScriptModules, you can use this flag to force everything to run using native Python. This allows the use of tools like pdb to debug code.

Types

The largest difference between TorchScript and the full Python language is that TorchScript only supports a small set of types that are needed to express neural net models. In particular, TorchScript supports:

Type

Description

Tensor

A PyTorch tensor of any dtype, dimension, or backend

Tuple[T0, T1, ...]

A tuple containing subtypes T0, T1, etc. (e.g. Tuple[Tensor, Tensor])

bool

A boolean value

int

A scalar integer

float

A scalar floating point number

List[T]

A list of which all members are type T

Optional[T]

A value which is either None or type T

Dict[K, V]

A dict with key type K and value type V. Only str, int, and float are allowed as key types.

Unlike Python, each variable in TorchScript function must have a single static type. This makes it easier to optimize TorchScript functions.

Example (a type mismatch):

@torch.jit.script
def an_error(x):
    if x:
        r = torch.rand(1)
    else:
        r = 4
    return r # Type mismatch: r is set to type Tensor in the true branch
             # and type int in the false branch

Default Types

By default, all parameters to a TorchScript function are assumed to be Tensor. To specify that an argument to a TorchScript function is another type, it is possible to use MyPy-style type annotations using the types listed above:

Example:

@torch.jit.script
def foo(x, tup):
    # type: (int, Tuple[Tensor, Tensor]) -> Tensor
    t0, t1 = tup
    return t0 + t1 + x

print(foo(3, (torch.rand(3), torch.rand(3))))

Note

It is also possible to annotate types with Python 3 type annotations. In our examples, we use comment-based annotations to ensure Python 2 compatibility as well.

An empty list is assumed to be List[Tensor] and empty dicts Dict[str, Tensor]. To instantiate an empty list or dict of other types, use torch.jit.annotate.

Example:

import torch
from torch.jit import Tensor
from typing import List, Tuple

class EmptyDataStructures(torch.jit.ScriptModule):
    def __init__(self):
        super(EmptyDataStructures, self).__init__()

    @torch.jit.script_method
    def forward(self, x):
        # type: (Tensor) -> Tuple[List[Tuple[int, float]], Dict[str, int]]

        # This annotates the list to be a `List[Tuple[int, float]]`
        my_list = torch.jit.annotate(List[Tuple[int, float]], [])
        for i in range(10):
            my_list.append((x, x))

        my_dict = torch.jit.annotate(Dict[str, int], {})
        return my_list, my_dict

Optional Type Refinement

TorchScript will refine the type of a variable of type Optional[T] when a comparison to None is made inside the conditional of an if-statement. The compiler can reason about multiple None checks that are combined with and, or, and not. Refinement will also occur for else blocks of if-statements that are not explicitly written.

The expression must be emitted within the conditional; assigning a None check to a variable and using it in the conditional will not refine types. An attribute like self.x will not be refined, but assigning self.x to a local variable first will work.

Example:

@torch.jit.script_method
def optional_unwrap(self, x, y):
  # type: (Optional[int], Optional[int]) -> int
  if x is None:
    x = 1
  x = x + 1

  z = self.z
  if y is not None and z is not None:
    x = y + z
  return x

User Defined Types

Python classes can be used in TorchScript if they are annotated with @torch.jit.script, similar to how you would declare a TorchScript function:

@torch.jit.script
class Foo:
  def __init__(self, x, y):
    self.x = x

  def aug_add_x(self, inc):
    self.x += inc

This subset is restricted:

  • All functions must be valid TorchScript functions (including __init__())

  • Classes must be new-style classes, as we use __new__() to construct them with pybind11

  • TorchScript classes are statically typed. Members are declared by assigning to self in the __init__() method

    For example, assigning outside of the __init__() method:

    @torch.jit.script
    class Foo:
      def assign_x(self):
        self.x = torch.rand(2, 3)
    

    Will result in:

    RuntimeError:
    Tried to set nonexistent attribute: x. Did you forget to initialize it in __init__()?:
    def assign_x(self):
      self.x = torch.rand(2, 3)
      ~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    
  • No expressions except method definitions are allowed in the body of the class

  • No support for inheritance or any other polymorphism strategy, except for inheriting from object to specify a new-style class

After a class is defined, it can be used in both TorchScript and Python interchangeably like any other TorchScript type:

@torch.jit.script
class Pair:
  def __init__(self, first, second):
    self.first = first
    self.second = second

@torch.jit.script
def sum_pair(p):
  # type: (Pair) -> Tensor
  return p.first + p.second

p = Pair(torch.rand(2, 3), torch.rand(2, 3))
print(sum_pair(p))

Expressions

The following Python Expressions are supported

Literals

True, False, None, 'string literals', "string literals", number literals 3 (interpreted as int) 3.4 (interpreted as a float)

List Construction

[3, 4], [], [torch.rand(3), torch.rand(4)]

Note

An empty list is assumed have type List[Tensor]. The types of other list literals are derived from the type of the members. To denote an empty list of another type, use torch.jit.annotate.

Tuple Construction

(3, 4), (3,)

Dict Construction

{'hello': 3}, {}, {'a': torch.rand(3), 'b': torch.rand(4)}

Note

An empty dict is assumed have type Dict[str, Tensor]. The types of other dict literals are derived from the type of the members. To denote an empty dict of another type, use torch.jit.annotate.

Variables

my_variable_name

Note

See Variable Resolution for how variables are resolved.

Arithmetic Operators

a + b

a - b

a * b

a / b

a ^ b

a @ b

Comparison Operators

a == b

a != b

a < b

a > b

a <= b

a >= b

Logical Operators

a and b

a or b

not b

Subscripts

t[0]

t[-1]

t[0:2]

t[1:]

t[:1]

t[:]

t[0, 1]

t[0, 1:2]

t[0, :1]

t[-1, 1:, 0]

t[1:, -1, 0]

t[i:j, i]

Function Calls

Calls to built-in functions: torch.rand(3, dtype=torch.int)

Calls to other script functions:

import torch

@torch.jit.script
def foo(x):
  return x + 1

@torch.jit.script
def bar(x):
  return foo(x)

Method Calls

Calls to methods of builtin types like tensor: x.mm(y)

When defining a Script method inside of a ScriptModule, the @script_method annotation is used. Inside of these methods it is possible to call other methods of this class or access methods on the submodules.

Calling a submodule directly (e.g. self.resnet(input)) is equivalent to calling its forward method (e.g. self.resnet.forward(input))

import torch

class MyScriptModule(torch.jit.ScriptModule):
    def __init__(self):
        super(MyScriptModule, self).__init__()
        self.means = torch.nn.Parameter(torch.tensor([103.939, 116.779, 123.68])
                                        .resize_(1, 3, 1, 1))
        self.resnet = torch.jit.trace(torchvision.models.resnet18(),
                                      torch.rand(1, 3, 224, 224))

    @torch.jit.script_method
    def helper(self, input):
      return self.resnet(input - self.means)

    @torch.jit.script_method
    def forward(self, input):
        return self.helper(input)

Ternary Expressions

x if x > y else y

Casts

float(ten)

int(3.5)

bool(ten)

Accessing Module Parameters

self.my_parameter

self.my_submodule.my_parameter

Statements

TorchScript supports the following types of statements:

Simple Assignments
a = b
a += b # short-hand for a = a + b, does not operate in-place on a
a -= b
Pattern Matching Assignments
a, b = tuple_or_list
a, b, *c = a_tuple

Print Statements

print("the result of an add:", a + b)

If Statements

if a < 4:
    r = -a
elif a < 3:
    r = a + a
else:
    r = 3 * a

In addition to bools, floats, ints, and Tensors can be used in a conditional and will be implicitly casted to a boolean.

While Loops

a = 0
while a < 4:
    print(a)
    a += 1

For loops with range

x = 0
for i in range(10):
    x *= i

For loops over tuples:

tup = (3, torch.rand(4))
for x in tup:
    print(x)

Note

for loops over tuples will unroll the loop, generating a body for each member of the tuple. The body must type-check correctly for each member.

For loops over constant torch.nn.ModuleList

class SubModule(torch.jit.ScriptModule):
    def __init__(self):
        super(Sub, self).__init__()
        self.weight = nn.Parameter(torch.randn(2))

    @torch.jit.script_method
    def forward(self, input):
        return self.weight + input

class MyModule(torch.jit.ScriptModule):
    __constants__ = ['mods']

    def __init__(self):
        super(MyModule, self).__init__()
        self.mods = torch.nn.ModuleList([SubModule() for i in range(10)])

    @torch.jit.script_method
    def forward(self, v):
        for module in self.mods:
            v = m(v)
        return v

Note

To use a nn.ModuleList inside a @script_method it must be marked constant by adding the name of the attribute to the __constants__ list for the type. For loops over a nn.ModuleList will unroll the body of the loop at compile time, with each member of the constant module list.

Return

return a, b

Note

TorchScript allows returns in the following circumstances:
  1. At the end of a function

  2. In an if-statement where <true> and <false> both return

  3. In an if-statement where <true> returns and <false> is empty (an early return)

Variable Resolution

TorchScript supports a subset of Python’s variable resolution (i.e. scoping) rules. Local variables behave the same as in Python, except for the restriction that a variable must have the same type along all paths through a function. If a variable has a different type on different sides of an if statement, it is an error to use it after the end of the if statement.

Similarly, a variable is not allowed to be used if it is only defined along some paths through the function.

Example:

@torch.jit.script
def foo(x):
    if x < 0:
        y = 4
    print(y) # Error: undefined value y

Non-local variables are resolved to Python values at compile time when the function is defined. These values are then converted into TorchScript values using the rules described in Use of Python Values.

Use of Python Values

To make writing TorchScript more convenient, we allow script code to refer to Python values in the surrounding scope. For instance, any time there is a reference to torch, the TorchScript compiler is actually resolving it to the torch Python module when the function is declared. These Python values are not a first class part of TorchScript. Instead they are de-sugared at compile-time into the primitive types that TorchScript supports. This depends on the dynamic type of the Python valued referenced when compilation occurs. This section describes the rules that are used when accessing Python values in TorchScript.

Functions

TorchScript can call Python functions. This functionality is very useful when incrementally converting a model to TorchScript. The model can be moved function-by-function to TorchScript, leaving calls to Python functions in place. This way you can incrementally check the correctness of the model as you go.

Example:

def foo(x):
  print("I am called with {}".format(x))
  import pdb; pdb.set_trace()
  return x

@torch.jit.script
def bar(x)
  return foo(x + 1)

Attempting to call save on a ScriptModule that contains calls to Python functions will fail. The intention is that this pathway is used for debugging and the calls removed or turned into script functions before saving. If you want to export a module with a Python function, add the @torch.jit.ignore decorator to the function which will replace these function calls with an exception when the model is saved:

class M(torch.jit.ScriptModule):
  def __init__(self):
    super(M, self).__init__()

  @torch.jit.script_method
  def forward(self, x):
    self.ignored_code(x)
    return x + 2

  @torch.jit.ignore
  def ignored_code(self, x):
    # non-TorchScript code
    import pdb; pdb.set_trace()

m = M()
# Runs, makes upcall to Python to run `ignored_code`
m(torch.ones(2, 2))

# Replaces all calls to `ignored_code` with a `raise`
m.save("m.pt")
loaded = torch.jit.load("m.pt")

# This runs `ignored_code` after saving which will raise an Exception!
loaded(torch.ones(2, 2))

Attribute Lookup On Python Modules

TorchScript can lookup attributes on modules. Builtin functions like torch.add are accessed this way. This allows TorchScript to call functions defined in other modules.

Python-defined Constants

TorchScript also provides a way to use constants that are defined in Python. These can be used to hard-code hyper-parameters into the function, or to define universal constants. There are two ways of specifying that a Python value should be treated as a constant.

  1. Values looked up as attributes of a module are assumed to be constant. Example: math.pi

  2. Attributes of a ScriptModule can be marked constant by listing them as a member of the __constants__ property of the class:

    Example:

    class Foo(torch.jit.ScriptModule):
        __constants__ = ['a']
    
        def __init__(self):
            super(Foo, self).__init__(False)
            self.a = 1 + 4
    
       @torch.jit.script_method
       def forward(self, input):
           return self.a + input
    

Supported constant Python Values are

  • int

  • float

  • bool

  • torch.device

  • torch.layout

  • torch.dtype

  • tuples containing supported types

  • torch.nn.ModuleList which can be used in a TorchScript for loop

Module Attributes

The torch.nn.Parameter wrapper and register_buffer can be used to assign tensors to a ScriptModule. In a similar vein, attributes of any type can be assign on a ScriptModule by wrapping them with torch.jit.Attribute and specifying the type. All types available in TorchScript are supported. These attributes are mutable and are saved in a separate archive in the serialized model binary. Tensor attributes are semantically the same as buffers.

Example:

class Foo(torch.jit.ScriptModule):
  def __init__(self, a_dict):
    super(Foo, self).__init__(False)
    self.words = torch.jit.Attribute([], List[str])
    self.some_dict = torch.jit.Attribute(a_dict, Dict[str, int])

  @torch.jit.script_method
  def forward(self, input):
    # type: (str) -> int
    self.words.append(input)
    return self.some_dict[input]

Debugging

Disable JIT for Debugging

If you want to disable all JIT modes (tracing and scripting) so you can debug your program in raw Python, you can use the PYTORCH_JIT environment variable. PYTORCH_JIT can be used to globally disable the JIT by setting its value to 0. Given an example script:

@torch.jit.script
def scripted_fn(x : torch.Tensor):
    for i in range(12):
        x = x + x
    return x


def fn(x):
    x = torch.neg(x)
    import pdb; pdb.set_trace()
    return scripted_fn(x)

traced_fn = torch.jit.trace(fn, (torch.rand(4, 5),))

traced_fn(torch.rand(3, 4))

Debugging this script with PDB works except for when we invoke the @torch.jit.script function. We can globally disable JIT, so that we can call the @torch.jit.script function as a normal python function and not compile it. If the above script is called disable_jit_example.py, we can invoke it like so:

$ PYTORCH_JIT=0 python disable_jit_example.py

and we will be able to step into the @torch.jit.script function as a normal Python function.

Inspecting Code

TorchScript provides a code pretty-printer for all ScriptModule instances. This pretty-printer gives an interpretation of the script method’s code as valid Python syntax. For example:

@torch.jit.script
def foo(len):
    # type: (int) -> torch.Tensor
    rv = torch.zeros(3, 4)
    for i in range(len):
        if i < 10:
            rv = rv - 1.0
        else:
            rv = rv + 1.0
        return rv

print(foo.code)

A ScriptModule with a single forward method will have an attribute code, which you can use to inspect the ScriptModule’s code. If the ScriptModule has more than one method, you will need to access .code on the method itself and not the module. We can inspect the code of a method named bar on a ScriptModule by accessing .bar.code.

The example script above produces the code:

def forward(self,
            len: int) -> Tensor:
    rv = torch.zeros([3, 4], dtype=None, layout=None, device=None)
    rv0 = rv
    for i in range(len):
        if torch.lt(i, 10):
            rv1 = torch.sub(rv0, 1., 1)
        else:
            rv1 = torch.add(rv0, 1., 1)
        rv0 = rv1
    return rv0

This is TorchScript’s compilation of the code for the forward method. You can use this to ensure TorchScript (tracing or scripting) has captured your model code correctly.

Interpreting Graphs

TorchScript also has a representation at a lower level than the code pretty- printer, in the form of IR graphs.

TorchScript uses a static single assignment (SSA) intermediate representation (IR) to represent computation. The instructions in this format consist of ATen (the C++ backend of PyTorch) operators and other primitive operators, including control flow operators for loops and conditionals. As an example:

@torch.jit.script
def foo(len):
  # type: (int) -> torch.Tensor
  rv = torch.zeros(3, 4)
  for i in range(len):
    if i < 10:
        rv = rv - 1.0
    else:
        rv = rv + 1.0
  return rv

print(foo.graph)

.graph follows the same rules described in the Inspecting Code section with regard to forward method lookup.

The example script above produces the graph:

graph(%len : int) {
  %15 : int = prim::Constant[value=1]()
  %9 : bool = prim::Constant[value=1]()
  %7 : Device = prim::Constant[value="cpu"]()
  %6 : int = prim::Constant[value=0]()
  %5 : int = prim::Constant[value=6]()
  %1 : int = prim::Constant[value=3]()
  %2 : int = prim::Constant[value=4]()
  %11 : int = prim::Constant[value=10]()
  %14 : float = prim::Constant[value=1]()
  %4 : int[] = prim::ListConstruct(%1, %2)
  %rv.1 : Tensor = aten::zeros(%4, %5, %6, %7)
  %rv : Tensor = prim::Loop(%len, %9, %rv.1)
    block0(%i : int, %13 : Tensor) {
      %12 : bool = aten::lt(%i, %11)
      %rv.4 : Tensor = prim::If(%12)
        block0() {
          %rv.2 : Tensor = aten::sub(%13, %14, %15)
          -> (%rv.2)
        }
        block1() {
          %rv.3 : Tensor = aten::add(%13, %14, %15)
          -> (%rv.3)
        }
      -> (%9, %rv.4)
    }
  return (%rv);
}

Take the instruction %rv.1 : Dynamic = aten::zeros(%3, %4, %5, %6) for example. %rv.1 : Dynamic means we assign the output to a (unique) value named rv.1, and that value is of Dynamic type, i.e. we do not know its concrete shape. aten::zeros is the operator (equivalent to torch.zeros) and the input list (%3, %4, %5, %6) specifies which values in scope should be passed as inputs. The schema for built-in functions like aten::zeros can be found at Builtin Functions.

Notice that operators can also have associated blocks, namely the prim::Loop and prim::If operators. In the graph print-out, these operators are formatted to reflect their equivalent source code forms to facilitate easy debugging.

Graphs can be inspected as shown to confirm that the computation described by a ScriptModule is correct, in both automated and manual fashion, as described below.

Tracing Edge Cases

There are some edge cases that exist where the trace of a given Python function/module will not be representative of the underlying code. These cases can include:

  • Tracing of control flow that is dependent on inputs (e.g. tensor shapes)

  • Tracing of in-place operations of tensor views (e.g. indexing on the left-hand side of an assignment)

Note that these cases may in fact be traceable in the future.

Automatic Trace Checking

One way to automatically catch many errors in traces is by using check_inputs on the torch.jit.trace() API. check_inputs takes a list of tuples of inputs that will be used to re-trace the computation and verify the results. For example:

def loop_in_traced_fn(x):
    result = x[0]
    for i in range(x.size(0)):
        result = result * x[i]
    return result

inputs = (torch.rand(3, 4, 5),)
check_inputs = [(torch.rand(4, 5, 6),), (torch.rand(2, 3, 4),)]

traced = torch.jit.trace(loop_in_traced_fn, inputs, check_inputs=check_inputs)
Gives us the following diagnostic information::

ERROR: Graphs differed across invocations! Graph diff:

  graph(%x : Tensor) {
    %1 : int = prim::Constant[value=0]()
    %2 : int = prim::Constant[value=0]()
    %result.1 : Tensor = aten::select(%x, %1, %2)
    %4 : int = prim::Constant[value=0]()
    %5 : int = prim::Constant[value=0]()
    %6 : Tensor = aten::select(%x, %4, %5)
    %result.2 : Tensor = aten::mul(%result.1, %6)
    %8 : int = prim::Constant[value=0]()
    %9 : int = prim::Constant[value=1]()
    %10 : Tensor = aten::select(%x, %8, %9)
-   %result : Tensor = aten::mul(%result.2, %10)
+   %result.3 : Tensor = aten::mul(%result.2, %10)
?          ++
    %12 : int = prim::Constant[value=0]()
    %13 : int = prim::Constant[value=2]()
    %14 : Tensor = aten::select(%x, %12, %13)
+   %result : Tensor = aten::mul(%result.3, %14)
+   %16 : int = prim::Constant[value=0]()
+   %17 : int = prim::Constant[value=3]()
+   %18 : Tensor = aten::select(%x, %16, %17)
-   %15 : Tensor = aten::mul(%result, %14)
?     ^                                 ^
+   %19 : Tensor = aten::mul(%result, %18)
?     ^                                 ^
-   return (%15);
?             ^
+   return (%19);
?             ^
  }

This message indicates to us that the computation differed between when we first traced it and when we traced it with the check_inputs. Indeed, the loop within the body of loop_in_traced_fn depends on the shape of the input x, and thus when we try another x with a different shape, the trace differs.

In this case, data-dependent control flow like this can be captured using script instead:

def fn(x):
    result = x[0]
    for i in range(x.size(0)):
        result = result * x[i]
    return result

inputs = (torch.rand(3, 4, 5),)
check_inputs = [(torch.rand(4, 5, 6),), (torch.rand(2, 3, 4),)]

scripted_fn = torch.jit.script(fn)
print(scripted_fn.graph)

for input_tuple in [inputs] + check_inputs:
    torch.testing.assert_allclose(fn(*input_tuple), scripted_fn(*input_tuple))

Which produces:

graph(%x : Tensor) {
  %5 : bool = prim::Constant[value=1]()
  %1 : int = prim::Constant[value=0]()
  %result.1 : Tensor = aten::select(%x, %1, %1)
  %4 : int = aten::size(%x, %1)
  %result : Tensor = prim::Loop(%4, %5, %result.1)
    block0(%i : int, %7 : Tensor) {
      %10 : Tensor = aten::select(%x, %1, %i)
      %result.2 : Tensor = aten::mul(%7, %10)
      -> (%5, %result.2)
    }
  return (%result);
}

Tracer Warnings

The tracer produces warnings for several problematic patterns in traced computation. As an example, take a trace of a function that contains an in-place assignment on a slice (a view) of a Tensor:

def fill_row_zero(x):
    x[0] = torch.rand(*x.shape[1:2])
    return x

traced = torch.jit.trace(fill_row_zero, (torch.rand(3, 4),))
print(traced.graph)

Produces several warnings and a graph which simply returns the input:

fill_row_zero.py:4: TracerWarning: There are 2 live references to the data region being modified when tracing in-place operator copy_ (possibly due to an assignment). This might cause the trace to be incorrect, because all other views that also reference this data will not reflect this change in the trace! On the other hand, if all other views use the same memory chunk, but are disjoint (e.g. are outputs of torch.split), this might still be safe.
  x[0] = torch.rand(*x.shape[1:2])
fill_row_zero.py:6: TracerWarning: Output nr 1. of the traced function does not match the corresponding output of the Python function. Detailed error:
Not within tolerance rtol=1e-05 atol=1e-05 at input[0, 1] (0.09115803241729736 vs. 0.6782537698745728) and 3 other locations (33.00%)
  traced = torch.jit.trace(fill_row_zero, (torch.rand(3, 4),))
graph(%0 : Float(3, 4)) {
  return (%0);
}

We can fix this by modifying the code to not use the in-place update, but rather build up the result tensor out-of-place with torch.cat:

def fill_row_zero(x):
    x = torch.cat((torch.rand(1, *x.shape[1:2]), x[1:2]), dim=0)
    return x

traced = torch.jit.trace(fill_row_zero, (torch.rand(3, 4),))
print(traced.graph)

Frequently Asked Questions

Q: I would like to train a model on GPU and do inference on CPU. What are the best practices?

First convert your model from GPU to CPU and then save it, like so:

cpu_model = gpu_model.cpu()
sample_input_cpu = sample_input_gpu.cpu()
traced_cpu = torch.jit.trace(traced_cpu, sample_input_cpu)
torch.jit.save(traced_cpu, "cpu.pth")

traced_gpu = torch.jit.trace(traced_gpu, sample_input_gpu)
torch.jit.save(traced_gpu, "gpu.pth")

# ... later, when using the model:

if use_gpu:
  model = torch.jit.load("gpu.pth")
else:
  model = torch.jit.load("cpu.pth")

model(input)

This is recommended because the tracer may witness tensor creation on a specific device, so casting an already-loaded model may have unexpected effects. Casting the model before saving it ensures that the tracer has the correct device information.

Q: How do I store attributes on a ScriptModule?

Say we have a model like:

class Model(torch.jit.ScriptModule):
  def __init__(self):
    super(Model, self).__init__()
    self.x = 2

  @torch.jit.script_method
  def forward(self):
    return self.x

If Model is instantiated it will result in a compilation error since the compiler doesn’t know about x. There are 4 ways to inform the compiler of attributes on ScriptModule:

1. nn.Parameter - values wrapped in nn.Parameter will work as they do on nn.Modules

2. register_buffer - values wrapped in register_buffer will work as they do on nn.Modules

3. __constants__ - adding a list called __constants__ at the class definition level will mark the contained names as constants. Constants are saved directly in the code of the model. See Python-defined Constants.

4. torch.jit.Attribute - values wrapped in torch.jit.Attribute can be any TorchScript type, be mutated and are saved outside of the code of the model. See Module Attributes.

Q: I would like to trace module’s method but I keep getting this error:

RuntimeError: Cannot insert a Tensor that requires grad as a constant. Consider making it a parameter or input, or detaching the gradient

This error usually means that, the method you are tracing, uses module’s parameters and you are passing module’s method instead of a module instance (e.g. my_module_instance.forward vs my_module_instance).

  • Invoking trace with module’s method captures module parameters (which may require gradients) as constants.

  • On the other hand, invoking trace with module’s instance (e.g. my_module) creates a new module and correctly copies parameters into the new module, so they can accumulate gradients if required.

Given that trace treats my_module_instance.forward as a standalone function, it also means there is not currently a way to trace arbitrary methods in the module except for forward that use module’s parameters. Version 1.1.1 will add a new API trace_module that will allow users to trace any method in the module and more than one method

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv = nn.Conv2d(1, 1, 3)

    def forward(self, x):
        return self.conv(x)

    def weighted_kernel_sum(self, weight):
        return weight * self.conv.weight

example_weight = torch.rand(1, 1, 3, 3)
example_forward_input = torch.rand(1, 1, 3, 3)
n = Net()
inputs = {'forward' : example_forward_input, 'weighted_kernel_sum' : example_weight}
module = torch.jit.trace_module(n, inputs)

Builtin Functions

TorchScript supports a subset of the builtin tensor and neural network functions that PyTorch provides. Most methods on Tensor as well as functions in the torch namespace, all functions in torch.nn.functional and all modules from torch.nn are supported in TorchScript, excluding those in the table below. For unsupported modules, we suggest using torch.jit.trace().

Unsupported torch.nn Modules

torch.nn.modules.adaptive.AdaptiveLogSoftmaxWithLoss
torch.nn.modules.normalization.CrossMapLRN2d
torch.nn.modules.fold.Fold
torch.nn.modules.fold.Unfold
torch.nn.modules.rnn.GRU
torch.nn.modules.rnn.RNN

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources