.. currentmodule:: torch.fx
torch.fx
=============
Overview
--------
.. automodule:: torch.fx
.. _Writing Transformations:
Writing Transformations
-----------------------
What is an FX transform? Essentially, it's a function that looks like this.
::
import torch
import torch.fx
def transform(m: nn.Module,
tracer_class : type = torch.fx.Tracer) -> torch.nn.Module:
# Step 1: Acquire a Graph representing the code in `m`
# NOTE: torch.fx.symbolic_trace is a wrapper around a call to
# fx.Tracer.trace and constructing a GraphModule. We'll
# split that out in our transform to allow the caller to
# customize tracing behavior.
graph : torch.fx.Graph = tracer_class().trace(m)
# Step 2: Modify this Graph or create a new one
graph = ...
# Step 3: Construct a Module to return
return torch.fx.GraphModule(m, graph)
Your transform will take in an :class:`torch.nn.Module`, acquire a :class:`Graph`
from it, do some modifications, and return a new
:class:`torch.nn.Module`. You should think of the :class:`torch.nn.Module` that your FX
transform returns as identical to a regular :class:`torch.nn.Module` -- you can pass it to another
FX transform, you can pass it to TorchScript, or you can
run it. Ensuring that the inputs and outputs of your FX transform are a
:class:`torch.nn.Module` will allow for composability.
.. note::
It is also possible to modify an existing :class:`GraphModule` instead of
creating a new one, like so::
import torch
import torch.fx
def transform(m : nn.Module) -> nn.Module:
gm : torch.fx.GraphModule = torch.fx.symbolic_trace(m)
# Modify gm.graph
# <...>
# Recompile the forward() method of `gm` from its Graph
gm.recompile()
return gm
Note that you MUST call :meth:`GraphModule.recompile` to bring the generated
``forward()`` method on the ``GraphModule`` in sync with the modified :class:`Graph`.
Given that you’ve passed in a :class:`torch.nn.Module` that has been traced into a
:class:`Graph`, there are now two primary approaches you can take to building a new
:class:`Graph`.
A Quick Primer on Graphs
^^^^^^^^^^^^^^^^^^^^^^^^
Full treatment of the semantics of graphs can be found in the :class:`Graph`
documentation, but we are going to cover the basics here. A :class:`Graph` is
a data structure that represents a method on a :class:`GraphModule`. The
information that this requires is:
- What are the inputs to the method?
- What are the operations that run inside the method?
- What is the output (i.e. return) value from the method?
All three of these concepts are represented with :class:`Node` instances.
Let's see what we mean by that with a short example:
::
import torch
import torch.fx
class MyModule(torch.nn.Module):
def __init__(self):
super().__init__()
self.param = torch.nn.Parameter(torch.rand(3, 4))
self.linear = torch.nn.Linear(4, 5)
def forward(self, x):
return torch.topk(torch.sum(
self.linear(x + self.linear.weight).relu(), dim=-1), 3)
m = MyModule()
gm = torch.fx.symbolic_trace(m)
gm.graph.print_tabular()
Here we define a module ``MyModule`` for demonstration purposes, instantiate it,
symbolically trace it, then call the :meth:`Graph.print_tabular` method to print
out a table showing the nodes of this :class:`Graph`:
+---------------+---------------+----------------------------+--------------------+-------------+
| opcode | name | target | args | kwargs |
+===============+===============+============================+====================+=============+
| placeholder | x | x | () | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| get_attr | linear_weight | linear.weight | () | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| call_function | add_1 | <built-in function add> | (x, linear_weight) | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| call_module | linear_1 | linear | (add_1,) | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| call_method | relu_1 | relu | (linear_1,) | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| call_function | sum_1 | <built-in method sum ...> | (relu_1,) | {'dim': -1} |
+---------------+---------------+----------------------------+--------------------+-------------+
| call_function | topk_1 | <built-in method topk ...> | (sum_1, 3) | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
| output | output | output | (topk_1,) | {} |
+---------------+---------------+----------------------------+--------------------+-------------+
We can use this information to answer the questions we posed above.
- What are the inputs to the method? In FX, method inputs are specified
via special ``placeholder`` nodes. In this case, we have a single
``placeholder`` node with a ``target`` of ``x``, meaning we have
a single (non-self) argument named x.
- What are the operations within the method? The ``get_attr``,
``call_function``, ``call_module``, and ``call_method`` nodes
represent the operations in the method. A full treatment of
the semantics of all of these can be found in the :class:`Node`
documentation.
- What is the return value of the method? The return value in a
:class:`Graph` is specified by a special ``output`` node.
Given that we now know the basics of how code is represented in
FX, we can now explore how we would edit a :class:`Graph`.
Graph Manipulation
^^^^^^^^^^^^^^^^^^
Direct Graph Manipulation
~~~~~~~~~~~~~~~~~~~~~~~~~
One approach to building this new :class:`Graph` is to directly manipulate your old
one. To aid in this, we can simply take the :class:`Graph` we obtain from symbolic
tracing and modify it. For example, let’s say we desire to replace
:func:`torch.add` calls with :func:`torch.mul` calls.
::
import torch
import torch.fx
# Sample module
class M(torch.nn.Module):
def forward(self, x, y):
return torch.add(x, y)
def transform(m: torch.nn.Module,
tracer_class : type = fx.Tracer) -> torch.nn.Module:
graph : fx.Graph = tracer_class().trace(m)
# FX represents its Graph as an ordered list of
# nodes, so we can iterate through them.
for node in graph.nodes:
# Checks if we're calling a function (i.e:
# torch.add)
if node.op == 'call_function':
# The target attribute is the function
# that call_function calls.
if node.target == torch.add:
node.target = torch.mul
graph.lint() # Does some checks to make sure the
# Graph is well-formed.
return fx.GraphModule(m, graph)
We can also do more involved :class:`Graph` rewrites, such as
deleting or appending nodes. To aid in these transformations,
FX has utility functions for transforming the graph that can
be found in the :class:`Graph` documentation. An
example of using these APIs to append a :func:`torch.relu` call
can be found below.
::
# Specifies the insertion point. Any nodes added to the
# Graph within this scope will be inserted after `node`
with traced.graph.inserting_after(node):
# Insert a new `call_function` node calling `torch.relu`
new_node = traced.graph.call_function(
torch.relu, args=(node,))
# We want all places that used the value of `node` to
# now use that value after the `relu` call we've added.
# We use the `replace_all_uses_with` API to do this.
node.replace_all_uses_with(new_node)
For simple transformations that only consist of substitutions, you can also
make use of the `subgraph rewriter. <https://github.com/pytorch/pytorch/blob/master/torch/fx/subgraph_rewriter.py>`__
Subgraph Rewriting With replace_pattern()
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
FX also provides another level of automation on top of direct graph manipulation.
The :func:`replace_pattern` API is essentially a "find/replace" tool for editing
:class:`Graph`\s. It allows you to specify a ``pattern`` and ``replacement`` function
and it will trace through those functions, find instances of the group of operations
in the ``pattern`` graph, and replace those instances with copies of the ``replacement``
graph. This can help to greatly automate tedious graph manipulation code, which can
get unwieldy as the transformations get more complex.
Graph Manipulation Examples
~~~~~~~~~~~~~~~~~~~~~~~~~~~
- `Replace one
op <https://github.com/pytorch/examples/blob/master/fx/replace_op.py>`__
- `Conv/Batch Norm
fusion <https://github.com/pytorch/pytorch/blob/40cbf342d3c000712da92cfafeaca651b3e0bd3e/torch/fx/experimental/optimization.py#L50>`__
- `replace_pattern: Basic usage <https://github.com/pytorch/examples/blob/master/fx/subgraph_rewriter_basic_use.py>`__
- `Quantization <https://pytorch.org/docs/master/quantization.html#prototype-fx-graph-mode-quantization>`__
- `Invert Transformation <https://github.com/pytorch/examples/blob/master/fx/invert.py>`__
Proxy/Retracing
^^^^^^^^^^^^^^^
Another way of manipulating :class:`Graph`\s is by reusing the :class:`Proxy`
machinery used in symbolic tracing. For example, let’s
imagine that we wanted to write a transformation that decomposed
PyTorch functions into smaller operations. It would transform every
``F.relu(x)`` call into ``(x > 0) * x``. One possibility would be to
perform the requisite graph rewriting to insert the comparison and
multiplication after the ``F.relu``, and then clean up the original
``F.relu``. However, we can automate this process by using :class:`Proxy`
objects to automatically record operations into the :class:`Graph`.
To use this method, we write the operations that we want inserted as regular
PyTorch code and invoke that code with :class:`Proxy` objects as arguments.
These :class:`Proxy` objects will capture the operations that are performed
on them and append them to the :class:`Graph`.
::
# Note that this decomposition rule can be read as regular Python
def relu_decomposition(x):
return (x > 0) * x
decomposition_rules = {}
decomposition_rules[F.relu] = relu_decomposition
def decompose(model: torch.nn.Module,
tracer_class : type = fx.Tracer) -> torch.nn.Module:
"""
Decompose `model` into smaller constituent operations.
Currently,this only supports decomposing ReLU into its
mathematical definition: (x > 0) * x
"""
graph : fx.Graph = tracer_class().trace(model)
new_graph = fx.Graph()
env = {}
tracer = torch.fx.proxy.GraphAppendingTracer(new_graph)
for node in graph.nodes:
if node.op == 'call_function' and node.target in decomposition_rules:
# By wrapping the arguments with proxies,
# we can dispatch to the appropriate
# decomposition rule and implicitly add it
# to the Graph by symbolically tracing it.
proxy_args = [
fx.Proxy(env[x.name], tracer) if isinstance(x, fx.Node) else x for x in node.args]
output_proxy = decomposition_rules[node.target](*proxy_args)
# Operations on `Proxy` always yield new `Proxy`s, and the
# return value of our decomposition rule is no exception.
# We need to extract the underlying `Node` from the `Proxy`
# to use it in subsequent iterations of this transform.
new_node = output_proxy.node
env[node.name] = new_node
else:
# Default case: we don't have a decomposition rule for this
# node, so just copy the node over into the new graph.
new_node = new_graph.node_copy(node, lambda x: env[x.name])
env[node.name] = new_node
return fx.GraphModule(model, new_graph)
In addition to avoiding explicit graph manipulation, using :class:`Proxy`\s
also allows you to specify your rewrite rules as native Python code.
For transformations that require a large amount of rewrite rules
(such as vmap or grad), this can often improve readability and
maintainability of the rules. Note that while calling :class:`Proxy` we also
passed a tracer pointing to the underlying variable `graph`. This is done so
if in case the operations in graph are n-ary (e.g. add is a binary operator)
the call to :class:`Proxy` does not create multiple instances of a graph
tracer which can lead to unexpected runtime errors. We recommend this method
of using :class:`Proxy` especially when the underlying operators can not be
safely assumed to be unary.
A worked example of using :class:`Proxy`\s for :class:`Graph` manipulation
can be found
`here <https://github.com/pytorch/examples/blob/master/fx/proxy_based_graph_creation.py>`__.
The Interpreter Pattern
^^^^^^^^^^^^^^^^^^^^^^^
A useful code organizational pattern in FX is to loop over all the :class:`Node`\s
in a :class:`Graph` and execute them. This can be used for several things including
runtime analysis of values flowing through the graph or transformation of the code
via retracing with :class:`Proxy`\s. For example, suppose we want to run a
:class:`GraphModule` and record the :class:`torch.Tensor` shape and dtype
properties on the nodes as we see them at runtime. That might look like:
::
import torch
import torch.fx
from torch.fx.node import Node
from typing import Dict
class ShapeProp:
"""
Shape propagation. This class takes a `GraphModule`.
Then, its `propagate` method executes the `GraphModule`
node-by-node with the given arguments. As each operation
executes, the ShapeProp class stores away the shape and
element type for the output values of each operation on
the `shape` and `dtype` attributes of the operation's
`Node`.
"""
def __init__(self, mod):
self.mod = mod
self.graph = mod.graph
self.modules = dict(self.mod.named_modules())
def propagate(self, *args):
args_iter = iter(args)
env : Dict[str, Node] = {}
def load_arg(a):
return torch.fx.graph.map_arg(a, lambda n: env[n.name])
def fetch_attr(target : str):
target_atoms = target.split('.')
attr_itr = self.mod
for i, atom in enumerate(target_atoms):
if not hasattr(attr_itr, atom):
raise RuntimeError(f"Node referenced nonexistant target {'.'.join(target_atoms[:i])}")
attr_itr = getattr(attr_itr, atom)
return attr_itr
for node in self.graph.nodes:
if node.op == 'placeholder':
result = next(args_iter)
elif node.op == 'get_attr':
result = fetch_attr(node.target)
elif node.op == 'call_function':
result = node.target(*load_arg(node.args), **load_arg(node.kwargs))
elif node.op == 'call_method':
self_obj, *args = load_arg(node.args)
kwargs = load_arg(node.kwargs)
result = getattr(self_obj, node.target)(*args, **kwargs)
elif node.op == 'call_module':
result = self.modules[node.target](*load_arg(node.args), **load_arg(node.kwargs))
# This is the only code specific to shape propagation.
# you can delete this `if` branch and this becomes
# a generic GraphModule interpreter.
if isinstance(result, torch.Tensor):
node.shape = result.shape
node.dtype = result.dtype
env[node.name] = result
return load_arg(self.graph.result)
As you can see, a full interpreter for FX is not that complicated
but it can be very useful. To ease using this pattern, we provide
the :class:`Interpreter` class, which encompasses the above logic
in a way that certain aspects of the interpreter's execution can
be overridden via method overrides.
In addition to executing operations, we can also generate a new
`Graph` by feeding :class:`Proxy` values through an interpreter.
Similarly, we provide the :class:`Transformer` class to encompass
this pattern. :class:`Transformer` behaves similarly to
:class:`Interpreter`, but instead of calling the ``run`` method to
get a concrete output value from the Module, you would call the
:meth:`Transformer.transform` method to return a new
:class:`GraphModule` which was subject to any transformation rules
you installed as overridden methods.
Examples of the Interpreter Pattern
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- `Shape
Propagation <https://github.com/pytorch/pytorch/blob/master/torch/fx/passes/shape_prop.py>`__
- `Performance Profiler <https://github.com/pytorch/tutorials/pull/1319>`__
Debugging
-----------
Introduction
^^^^^^^^^^^^^^^^
Often in the course of authoring transformations, our code will not be quite right.
In this case, we may need to do some debugging. The key is to work
backwards: first, check the results of invoking the generated module to prove or
disprove correctness. Then, inspect and debug the generated code. Then, debug the
process of transformations that led to the generated code.
If you’re not familiar with debuggers, please see the auxiliary section
:ref:`Available Debuggers`.
Common Pitfalls in Transform Authoring
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
* Nondeterministic ``set`` iteration order. In Python, the ``set`` datatype is
unordered. Using ``set`` to contain collections of objects like ``Node``\ s,
for example, can cause unexpected nondeterminism. An example is iterating
over a set of ``Node``\ s to insert them into a ``Graph``. Because the
``set`` data type is unordered, the ordering of the operations in the output
program will be nondeterministic and can change across program invocations.
The recommended alternative is to use a ``dict`` data type, which is
`insertion ordered <https://mail.python.org/pipermail/python-dev/2017-December/151283.html>`_
as of Python 3.7 (and as of cPython 3.6). A ``dict`` can be used equivalently
to a set by storing values to be deduplicated in the keys of the ``dict``.
Checking Correctness of Modules
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Because the output of most deep learning modules consists of floating
point :class:`torch.Tensor` instances, checking for equivalence between
the results of two :class:`torch.nn.Module` is not as straightforward
as doing a simple equality check. To motivate this, let's use an
example:
::
import torch
import torch.fx
import torchvision.models as models
def transform(m : torch.nn.Module) -> torch.nn.Module:
gm = torch.fx.symbolic_trace(m)
# Imagine we're doing some transforms here
# <...>
gm.recompile()
return gm
resnet18 = models.resnet18()
transformed_resnet18 = transform(resnet18)
input_image = torch.randn(5, 3, 224, 224)
assert resnet18(input_image) == transformed_resnet18(input_image)
"""
RuntimeError: Boolean value of Tensor with more than one value is ambiguous
"""
Here, we've tried to check equality of the values of two deep learning
models with the ``==`` equality operator. However, this is not well-
defined both due to the issue of that operator returning a tensor
and not a bool, but also because comparison of floating point values
should use a margin of error (or epsilon) to account for the
non-commutativity of floating point operations (see
`here <https://floating-point-gui.de/errors/comparison/>`__ for more
details). We can use :func:`torch.allclose` instead, which will give
us an approximate comparison taking into account a relative and
absolute tolerance threshold:
::
assert torch.allclose(resnet18(input_image), transformed_resnet18(input_image))
This is the first tool in our toolbox to check if transformed modules are
behaving as we expect compared to a reference implementation.
Debugging the Generated Code
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Because FX generates the ``forward()`` function on :class:`GraphModule`\s, using
traditional debugging techniques like ``print`` statements or ``pdb`` is
not as straightforward. Luckily, we have several techniques we can use
for debugging the generated code.
Use ``pdb``
~~~~~~~~~~~~~
Invoke ``pdb`` to step into the running program. Although the code that
represents the :class:`Graph` is not in any source file, we can still step
into it manually using ``pdb`` when the forward pass is invoked.
::
import torch
import torch.fx
import torchvision.models as models
def my_pass(inp: torch.nn.Module, tracer_class : type = fx.Tracer) -> torch.nn.Module:
graph = tracer_class().trace(inp)
# Transformation logic here
# <...>
# Return new Module
return fx.GraphModule(inp, graph)
my_module = models.resnet18()
my_module_transformed = my_pass(my_module)
input_value = torch.randn(5, 3, 224, 224)
# When this line is executed at runtime, we will be dropped into an
# interactive `pdb` prompt. We can use the `step` or `s` command to
# step into the execution of the next line
import pdb; pdb.set_trace()
my_module_transformed(input_value)
.. _Print the Generated Code:
Print the Generated Code
~~~~~~~~~~~~~~~~~~~~~~~~~~~
If you’d like to run the same code multiple times, then it can be
a bit tedious to step to the right code with ``pdb``. In that case, one
approach is to simply copy-paste the generated ``forward`` pass into
your code and examine it from there.
::
# Assume that `traced` is a GraphModule that has undergone some
# number of transforms
# Copy this code for later
print(traced)
# Print the code generated from symbolic tracing. This outputs:
"""
def forward(self, y):
x = self.x
add_1 = x + y; x = y = None
return add_1
"""
# Subclass the original Module
class SubclassM(M):
def __init__(self):
super().__init__()
# Paste the generated `forward` function (the one we printed and
# copied above) here
def forward(self, y):
x = self.x
add_1 = x + y; x = y = None
return add_1
# Create an instance of the original, untraced Module. Then, create an
# instance of the Module with the copied `forward` function. We can
# now compare the output of both the original and the traced version.
pre_trace = M()
post_trace = SubclassM()
Use the ``to_folder`` Function From ``GraphModule``
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
:meth:`GraphModule.to_folder` is a method in ``GraphModule`` that allows
you to dump out the generated FX code to a folder. Although copying the
forward pass into the code often suffices as in :ref:`Print the Generated Code`,
it may be easier to examine modules and parameters using ``to_folder``.
::
m = symbolic_trace(M())
m.to_folder("foo", "Bar")
from foo import Bar
y = Bar()
After running the above example, we can then look at the code within
``foo/module.py`` and modify it as desired (e.g. adding ``print``
statements or using ``pdb``) to debug the generated code.
Debugging the Transformation
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Now that we've identified that a transformation is creating incorrect
code, it's time to debug the transformation itself. First, we'll check
the :ref:`Limitations of Symbolic Tracing` section in the documentation.
Once we verify that tracing is working as expected, the goal
becomes figuring out what went wrong during our ``GraphModule``
transformation. There may be a quick answer in
:ref:`Writing Transformations`, but, if not, there are several ways to
examine our traced module:
::
# Sample Module
class M(torch.nn.Module):
def forward(self, x, y):
return x + y
# Create an instance of `M`
m = M()
# Symbolically trace an instance of `M` (returns a GraphModule). In
# this example, we'll only be discussing how to inspect a
# GraphModule, so we aren't showing any sample transforms for the
# sake of brevity.
traced = symbolic_trace(m)
# Print the code produced by tracing the module.
print(traced)
# The generated `forward` function is:
"""
def forward(self, x, y):
add = x + y; x = y = None
return add
"""
# Print the internal Graph.
print(traced.graph)
# This print-out returns:
"""
graph():
%x : [#users=1] = placeholder[target=x]
%y : [#users=1] = placeholder[target=y]
%add : [#users=1] = call_function[target=operator.add](args = (%x, %y), kwargs = {})
return add
"""
# Print a tabular representation of the internal Graph.
traced.graph.print_tabular()
# This gives us:
"""
opcode name target args kwargs
------------- ------ ----------------------- ------ --------
placeholder x x () {}
placeholder y y () {}
call_function add <built-in function add> (x, y) {}
output output output (add,) {}
"""
Using the utility functions above, we can compare our traced Module
before and after we've applied our transformations. Sometimes, a
simple visual comparison is enough to trace down a bug. If it's still
not clear what's going wrong, a debugger like ``pdb`` can be a good
next step.
Going off of the example above, consider the following code:
::
# Sample user-defined function
def transform_graph(module: torch.nn.Module, tracer_class : type = fx.Tracer) -> torch.nn.Module:
# Get the Graph from our traced Module
g = tracer_class().trace(module)
"""
Transformations on `g` go here
"""
return fx.GraphModule(module, g)
# Transform the Graph
transformed = transform_graph(traced)
# Print the new code after our transforms. Check to see if it was
# what we expected
print(transformed)
Using the above example, let’s say that the call to ``print(traced)``
showed us that there was an error in our transforms. We want to find
what goes wrong using a debugger. We start a ``pdb`` session. We can see
what’s happening during the transform by breaking on
``transform_graph(traced)``, then pressing ``s`` to “step into” the call
to ``transform_graph(traced)``.
We may also have good luck by editing the ``print_tabular`` method to print
different attributes of the Nodes in the Graph. (For example, we might
want to see the Node’s ``input_nodes`` and ``users``.)
.. _Available Debuggers:
Available Debuggers
^^^^^^^^^^^^^^^^^^^^^^
The most common Python debugger is
`pdb <https://docs.python.org/3/library/pdb.html>`__. You can start
your program in “debug mode” with ``pdb`` by typing
``python -m pdb FILENAME.py`` into the command line, where ``FILENAME``
is the name of the file you want to debug. After that, you can use the
``pdb`` `debugger commands
<https://docs.python.org/3/library/pdb.html#debugger-commands>`__
to move through your running program stepwise. It’s common to set a
breakpoint (``b LINE-NUMBER``) when you start ``pdb``, then call ``c`` to
run the program until that point. This prevents you from having to step
through each line of execution (using ``s`` or ``n``) to get to the part
of the code you want to examine. Alternatively, you can write
``import pdb; pdb.set_trace()`` before the line you want to break at.
If you add ``pdb.set_trace()``, your program will automatically start
in debug mode when you run it. (In other words, you can just type
``python FILENAME.py`` into the command line instead of
``python -m pdb FILENAME.py``.) Once you're running your file in
debug mode, you can step through the code and examine your program's
internal state using certain commands. There are many excellent
tutorials on ``pdb`` online, including RealPython’s
`“Python Debugging With Pdb” <https://realpython.com/python-debugging-pdb/>`__.
IDEs like PyCharm or VSCode usually have a debugger built in. In your
IDE, you can choose to either a) use ``pdb`` by pulling up a terminal
window in your IDE (e.g. View → Terminal in VSCode), or b) use the
built-in debugger (usually a graphical wrapper around ``pdb``).
.. _Limitations of Symbolic Tracing:
Limitations of Symbolic Tracing
-------------------------------
FX uses a system of **symbolic tracing** (a.k.a `symbolic
execution <https://en.wikipedia.org/wiki/Symbolic_execution>`__)
to capture the semantics of programs in a transformable/analyzable form.
The system is **tracing** in that it executes the program (really a
:class:`torch.nn.Module` or function) to record operations. It is
**symbolic** in that the data flowing through the program during this
execution is not real data, but rather symbols (:class:`Proxy` in FX parlance).
Although symbolic tracing works for most neural net code, it has some
limitations.
Dynamic Control Flow
^^^^^^^^^^^^^^^^^^^^
The main limitation of symbolic tracing is it does not currently support
*dynamic control flow*. That is, loops or ``if`` statements where the
condition may depend on the input values of the program.
For example, let’s examine the following program:
::
def func_to_trace(x):
if x.sum() > 0:
return torch.relu(x)
else:
return torch.neg(x)
traced = torch.fx.symbolic_trace(func_to_trace)
"""
<...>
File "dyn.py", line 6, in func_to_trace
if x.sum() > 0:
File "pytorch/torch/fx/proxy.py", line 155, in __bool__
return self.tracer.to_bool(self)
File "pytorch/torch/fx/proxy.py", line 85, in to_bool
raise TraceError('symbolically traced variables cannot be used as inputs to control flow')
torch.fx.proxy.TraceError: symbolically traced variables cannot be used as inputs to control flow
"""
The condition to the ``if`` statement relies on the value of ``x.sum()``,
which relies on the value of ``x``, a function input. Since
``x`` can change (i.e. if you pass a new input tensor to the traced
function), this is *dynamic control flow*. The traceback walks back up
through your code to show you where this situation happens.
Static Control Flow
~~~~~~~~~~~~~~~~~~~
On the other hand, so-called *static control flow* is supported. Static
control flow is loops or ``if`` statements whose value cannot change
across invocations. Typically, in PyTorch programs, this control flow
arises for code making decisions about a model’s architecture based on
hyper-parameters. As a concrete example:
::
import torch
import torch.fx
class MyModule(torch.nn.Module):
def __init__(self, do_activation : bool = False):
super().__init__()
self.do_activation = do_activation
self.linear = torch.nn.Linear(512, 512)
def forward(self, x):
x = self.linear(x)
# This if-statement is so-called static control flow.
# Its condition does not depend on any input values
if self.do_activation:
x = torch.relu(x)
return x
without_activation = MyModule(do_activation=False)
with_activation = MyModule(do_activation=True)
traced_without_activation = torch.fx.symbolic_trace(without_activation)
print(traced_without_activation.code)
"""
def forward(self, x):
linear_1 = self.linear(x); x = None
return linear_1
"""
traced_with_activation = torch.fx.symbolic_trace(with_activation)
print(traced_with_activation.code)
"""
import torch
def forward(self, x):
linear_1 = self.linear(x); x = None
relu_1 = torch.relu(linear_1); linear_1 = None
return relu_1
"""
The if-statement ``if self.do_activation`` does not depend on any
function inputs, thus it is static. ``do_activation`` can be considered
to be a hyper-parameter, and the traces of different instances of
``MyModule`` with different values for that parameter have different
code. This is a valid pattern that is supported by symbolic tracing.
Many instances of dynamic control flow are semantically static control
flow. These instances can be made to support symbolic tracing by
removing the data dependencies on input values, for example by moving
values to ``Module`` attributes or by binding concrete values to arguments
during symbolic tracing:
::
def f(x, flag):
if flag: return x
else: return x*2
fx.symbolic_trace(f) # Fails!
fx.symbolic_trace(f, concrete_args={'flag': True})
In the case of truly dynamic control flow, the sections of the program
that contain this code can be traced as calls to the Method (see
:ref:`Customizing Tracing`) or function (see
:func:`wrap`) rather than tracing through them.
Non-\ ``torch`` Functions
^^^^^^^^^^^^^^^^^^^^^^^^^
FX uses ``__torch_function__`` as the mechanism by which it intercepts
calls (see the `technical
overview <https://github.com/pytorch/pytorch/blob/master/torch/fx/OVERVIEW.md#technical-details>`__
for more information about this). Some functions, such as builtin Python
functions or those in the ``math`` module, are not covered by
``__torch_function__``, but we would still like to capture them in
symbolic tracing. For example:
::
import torch
import torch.fx
from math import sqrt
def normalize(x):
"""
Normalize `x` by the size of the batch dimension
"""
return x / sqrt(len(x))
# It's valid Python code
normalize(torch.rand(3, 4))
traced = torch.fx.symbolic_trace(normalize)
"""
<...>
File "sqrt.py", line 9, in normalize
return x / sqrt(len(x))
File "pytorch/torch/fx/proxy.py", line 161, in __len__
raise RuntimeError("'len' is not supported in symbolic tracing by default. If you want "
RuntimeError: 'len' is not supported in symbolic tracing by default. If you want this call to be recorded, please call torch.fx.wrap('len') at module scope
"""
The error tells us that the built-in function ``len`` is not supported.
We can make it so that functions like this are recorded in the trace as
direct calls using the :func:`wrap` API:
::
torch.fx.wrap('len')
torch.fx.wrap('sqrt')
traced = torch.fx.symbolic_trace(normalize)
print(traced.code)
"""
import math
def forward(self, x):
len_1 = len(x)
sqrt_1 = math.sqrt(len_1); len_1 = None
truediv = x / sqrt_1; x = sqrt_1 = None
return truediv
"""
.. _Customizing Tracing:
Customizing Tracing with the ``Tracer`` class
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
The :class:`Tracer` class is the class that underlies the
implementation of ``symbolic_trace``. The behavior of tracing can be
customized by subclassing Tracer, like so:
::
class MyCustomTracer(torch.fx.Tracer):
# Inside here you can override various methods
# to customize tracing. See the `Tracer` API
# reference
pass
# Let's use this custom tracer to trace through this module
class MyModule(torch.nn.Module):
def forward(self, x):
return torch.relu(x) + torch.ones(3, 4)
mod = MyModule()
traced_graph = MyCustomTracer().trace(mod)
# trace() returns a Graph. Let's wrap it up in a
# GraphModule to make it runnable
traced = torch.fx.GraphModule(mod, traced_graph)
Leaf Modules
~~~~~~~~~~~~
Leaf Modules are the modules that appear as calls in the symbolic trace
rather than being traced through. The default set of leaf modules is the
set of standard ``torch.nn`` module instances. For example:
::
class MySpecialSubmodule(torch.nn.Module):
def forward(self, x):
return torch.neg(x)
class MyModule(torch.nn.Module):
def __init__(self):
super().__init__()
self.linear = torch.nn.Linear(3, 4)
self.submod = MySpecialSubmodule()
def forward(self, x):
return self.submod(self.linear(x))
traced = torch.fx.symbolic_trace(MyModule())
print(traced.code)
# `linear` is preserved as a call, yet `submod` is traced though.
# This is because the default set of "Leaf Modules" includes all
# standard `torch.nn` modules.
"""
import torch
def forward(self, x):
linear_1 = self.linear(x); x = None
neg_1 = torch.neg(linear_1); linear_1 = None
return neg_1
"""
The set of leaf modules can be customized by overriding
:meth:`Tracer.is_leaf_module`.
Miscellanea
^^^^^^^^^^^
- Tensor constructors (e.g. ``torch.zeros``, ``torch.ones``,
``torch.rand``, ``torch.randn``, ``torch.sparse_coo_tensor``)
are currently not traceable.
- The deterministic constructors (``zeros``, ``ones``) can be used
and the value they produce will be embedded in the trace as a
constant. This is only problematic if the arguments to these
constructors refers to dynamic input sizes. In this case,
``ones_like`` or ``zeros_like`` may be a viable substitute.
- Nondeterministic constructors (``rand``, ``randn``) will have a
single random value embedded in the trace. This is likely not the
intended behavior. One workaround is to wrap ``torch.randn`` in a ``torch.fx.wrap`` function and call that instead.
::
@torch.fx.wrap
def torch_randn(x, shape):
return torch.randn(shape)
def f(x):
return x + torch_randn(x, 5)
fx.symbolic_trace(f)
- This behavior may be fixed in a future release.
- Type annotations
- Python 3-style type annotations (e.g.
``func(x : torch.Tensor, y : int) -> torch.Tensor``) are supported
and will be preserved by symbolic tracing.
- Python 2-style comment type annotations
``# type: (torch.Tensor, int) -> torch.Tensor`` are not currently
supported.
- Annotations on local names within a function are not currently
supported.
- Gotcha around ``training`` flag and submodules
- When using functionals like ``torch.nn.functional.dropout``, it will be common for the training argument to be passed in as ``self.training``. During FX tracing, this will likely be baked in as a constant value.
::
import torch
import torch.fx
class DropoutRepro(torch.nn.Module):
def forward(self, x):
return torch.nn.functional.dropout(x, training=self.training)
traced = torch.fx.symbolic_trace(DropoutRepro())
print(traced.code)
"""
def forward(self, x):
dropout = torch.nn.functional.dropout(x, p = 0.5, training = True, inplace = False); x = None
return dropout
"""
traced.eval()
x = torch.randn(5, 3)
torch.testing.assert_allclose(traced(x), x)
"""
AssertionError: Tensor-likes are not close!
Mismatched elements: 15 / 15 (100.0%)
Greatest absolute difference: 1.6207983493804932 at index (0, 2) (up to 1e-05 allowed)
Greatest relative difference: 1.0 at index (0, 0) (up to 0.0001 allowed)
"""
- However, when the standard ``nn.Dropout()`` submodule is used, the training flag is encapsulated and--because of the preservation of the ``nn.Module`` object model--can be changed.
::
class DropoutRepro2(torch.nn.Module):
def __init__(self):
super().__init__()
self.drop = torch.nn.Dropout()
def forward(self, x):
return self.drop(x)
traced = torch.fx.symbolic_trace(DropoutRepro2())
print(traced.code)
"""
def forward(self, x):
drop = self.drop(x); x = None
return drop
"""
traced.eval()
x = torch.randn(5, 3)
torch.testing.assert_allclose(traced(x), x)
- Because of this difference, consider marking modules that interact with the ``training`` flag dynamically as leaf modules.
API Reference
-------------
.. autofunction:: torch.fx.symbolic_trace
.. autofunction:: torch.fx.wrap
.. autoclass:: torch.fx.GraphModule
:members:
.. automethod:: __init__
.. autoclass:: torch.fx.Graph
:members:
.. automethod:: __init__
.. autoclass:: torch.fx.Node
:members:
.. autoclass:: torch.fx.Tracer
:members:
:inherited-members:
.. autoclass:: torch.fx.Proxy
.. autoclass:: torch.fx.Interpreter
:members:
.. autoclass:: torch.fx.Transformer
:members:
.. autofunction:: torch.fx.replace_pattern
.. The experimental and passes submodules are missing docs.
.. Adding it here for coverage but this doesn't add anything to the
.. rendered doc.
.. py:module:: torch.fx.passes
.. py:module:: torch.fx.passes.infra
.. py:module:: torch.fx.passes.backends
.. py:module:: torch.fx.passes.utils
.. py:module:: torch.fx.passes.tests
.. py:module:: torch.fx.experimental
.. py:module:: torch.fx.experimental.unification
.. py:module:: torch.fx.experimental.unification.multipledispatch
.. py:module:: torch.fx.experimental.migrate_gradual_types
.. py:module:: torch.fx.passes.dialect
.. py:module:: torch.fx.passes.dialect.common