第一篇——什麼是torch.fx
今天聊一下比較重要的torch.fx，也趁著這次機會把之前的torch.fx筆記整理下，筆記大概拆成三份，分別對應三篇：
- 什麼是torch.fx
- 基於torch.fx做量化
- 基於torch.fx量化部署到TensorRT
本文對應第一篇，主要介紹torch.fx和基本使用方法。廢話不多說，直接開始吧！
什麼是Torch.FX
torch.fx是Pytorch 1.8出來的一套工具或者說一個庫，是做python-to-python code transformation，大意就是可以把pytorch中的python前向程式碼轉換為你想要的樣子，官方介紹如下：
We apply this principle in torch.fx, a program capture and
transformation library for PyTorch written entirely in Python and optimized for high developer productivity by ML practitioners
上述來源於FX的論文，感興趣的可以看TORCH.FX: PRACTICAL PROGRAM CAPTURE AND TRANSFORMATION FOR DEEP LEARNING IN PYTHON這篇，知乎上也有一篇不錯的解讀，這裡就不復述了。不過本文也會介紹論文中的內容，更多的是以實踐的角度。
核心的關鍵詞是program capture和transformation library，這兩個概念很重要。
那麼FX怎麼用呢？直觀瞭解一下，我們定義了一個pytorch.nn.module：
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 self.linear(x + self.param).clamp(min=0.0, max=1.0)
很簡單地繼承於torch.nn.Module的Module（熟悉pytorch的應該都懂）。其中前向forward函式也記錄了這個module的具體操作邏輯。
如果我們想把這個Module中forward中的一部分操作邏輯self.linear(x + self.param).clamp(min=0.0, max=1.0)的clamp部分替換為sigmoid，應該怎麼搞呢？
當然可以直接改程式碼麼，但是如果這些操作很多，或者說你寫了很多模組，或者說你想要做很多實驗（某些模組中改某些模組中不改），再這樣就比較煩瑣了。
這時候就需要FX，不需要我們手動修改程式碼（就是自己改這個forward實現），只需要設定好規則，使用torch.fx，帶入這個模型例項進去，跑一下程式碼。然後你的這個MyModule中forward部分就會變為self.linear(x + self.param).sigmoid()：
module = MyModule()
from torch.fx import symbolic_trace
# Symbolic tracing frontend - captures the semantics of the module
symbolic_traced : torch.fx.GraphModule = symbolic_trace(module)
# High-level intermediate representation (IR) - Graph representation
# 列印檢視FX的IR
print(symbolic_traced.graph)
"""
graph():
%x : [#users=1] = placeholder[target=x]
%param : [#users=1] = get_attr[target=param]
%add : [#users=1] = call_function[target=operator.add](args = (%x, %param), kwargs = {})
%linear : [#users=1] = call_module[target=linear](args = (%add,), kwargs = {})
%clamp : [#users=1] = call_method[target=clamp](args = (%linear,), kwargs = {min: 0.0, max: 1.0})
return clamp
"""
# Code generation - valid Python code
# 通過FX生成的程式碼，可以視為module中的forward程式碼
print(symbolic_traced.code)
"""
def forward(self, x):
param = self.param
add = x + param; x = param = None
linear = self.linear(add); add = None
clamp = linear.clamp(min = 0.0, max = 1.0); linear = None
return clamp
"""
這樣，FX會幫助你修改這個Module，並且修改好的這個model就和平常一樣使用就可以，注意這裡，FX capture 了你寫的forward程式碼，然後進行了transform，修改了其中的操作。
當然這只是很簡單很簡單的fx的一個功能，我們還可以通過fx：
- 融合兩個op，比如conv和bn
- 去掉某些op
- 替換某些op
- 在某些op後插入一些op或者其他操作
等等等等。
可能大家會疑惑，這些操作是不是很像AI編譯器中的PASS，而操作物件也是神經網路這種DAG（有向無環圖）。其實吧，FX你也可以理解為是一種編譯器，不過這個編譯器最終產生的可執行檔案，而是python->python，最終的產物還是基於Pytorch規則的python程式碼，也就是為什麼FX一直說自己是Python-to-Python (or Module-to-Module) transformation toolkit而不是compiler了。
FX目前大部分API已經穩定（在torch-1.10中正式釋出），使用起來歷史包袱不大。
fx的官方介紹：
torch.fx與量化的關係
FX的出現第一利好是基於Pytorch的量化工具，這也是我介紹FX的一個原因。藉助FX可以很方便地對pytorch模型做量化操作，之前商湯就出了一個基於fx的量化工具MQBench。
對於量化來說，不論是PTQ（需要插入觀察op來收集每一層的啟用分佈以及權重分佈）還是QTA（需要插入fake量化節點來模擬量化），都會涉及到fx的功能。所以如果想基於Pytorch框架來做量化，建議直接上手torch.fx。
fx在pytorch-1.10中已經處於stable狀態，大部分API已經穩定了，我也拿torch.fx量化了幾個模型，最終搞到TensorRT上，涉及到卷積、BN、反摺積、add、concat等基本操作，使用的版本是Pytorch-1.10和TensorRT-8.2。
其中fx部分自己修改了下原始碼，補充了一些op。這裡我是直接把最新release的pytorch中的fx部分摘出來，然後pip安裝torch-1.10.0+cu113-cp38-cp38-linux_x86_64.whl，兩者搭配食用。
與TorchScript的區別
其實一開始torch.fx出現的時候也想過這兩個有啥區別，都是先解析模型、然後生成IR、然後基於IR做一些優化，最後生成一個最終版的優化後的模型，難道一個是python版本的一個是C++版？肯定沒有這麼簡單。當你FX用多了，會發現FX和torchscript的定位是不一樣的，FX更側重於對模型進行一些功能性的改變（比如批量增加、修改某個操作，比如增加統計操作，比如量化）；而torchscript更側重於優化當前模型的效能，並且可以脫離python，僅在C++環境執行。
借一句官方大佬的回答：
torch.fx is different from TorchScript in that it is a platform for Python-to-Python transformations of PyTorch code. TorchScript, on the other hand, is more targeted at moving PyTorch programs outside of Python for deployment purposes. In this sense, FX and TorchScript are orthogonal to each other, and can even be composed with each other (e.g. transform PyTorch programs with FX, then subsequently export to TorchScript for deployment).
大意就是，FX僅僅是做Python2Python的轉換，不像Torchscript一樣是為了做部署(脫離Python這個環境，在C++中執行)而做轉換。兩者沒什麼關係，不衝突，用FX轉換後的模型也可以用torchscript繼續轉換，兩者是正交的。
Python to Python?
不過需要注意的是，FX的程式碼生成式由Python到Python。也就是說，FX生成的程式碼，和我們平常使用nn.Module搭建的網路沒區別，可以直接使用Pytorch的eager mode跑，不像torchscript一樣，是另一套runtime（我們跑torchscript的時候其實呼叫的是一個VM，也就是虛擬機器，通過VM在C++中跑通過torchscript匯出的模型）。
因此fx轉換後的模型型別和nn.Module一毛一樣，所以對nn.Module能做的，對轉換後的模型也能做，我們們可以連續套娃：
- 自己寫的Module -> fx後還是Module -> 連續fx變化 -> 得到最終的fx模型
FX的IR和Jit的IR
這倆IR不一樣，FX的IR相較Jit的來說，有兩個優點：
- FX緊密地整合到Python的runtime中，因為FX可以更加精準地捕獲
prograim representations，不像jit.trace有時候會出錯。 - FX的Graph和
torch.nn.module沒啥區別，其IR沒有那麼底層，所以說用起來更簡單，效率也會提升。
這裡簡單列一下FX的IR，很簡單，只有六種，大概功能就是調函式、提取attr、獲取輸入輸出等:
placeholderrepresents a function input. Thenameattribute specifies the name this value will take on.targetis similarly the name of the argument.argsholds either: 1) nothing, or 2) a single argument denoting the default parameter of the function input.kwargsis don't-care. Placeholders correspond to the function parameters (e.g.x) in the graph printout.get_attrretrieves a parameter from the module hierarchy.nameis similarly the name the result of the fetch is assigned to.targetis the fully-qualified name of the parameter's position in the module hierarchy.argsandkwargsare don't-carecall_functionapplies a free function to some values.nameis similarly the name of the value to assign to.targetis the function to be applied.argsandkwargsrepresent the arguments to the function, following the Python calling conventioncall_moduleapplies a module in the module hierarchy'sforward()method to given arguments.nameis as previous.targetis the fully-qualified name of the module in the module hierarchy to call.argsandkwargsrepresent the arguments to invoke the module on, including the self argument.call_methodcalls a method on a value.nameis as similar.targetis the string name of the method to apply to theselfargument.argsandkwargsrepresent the arguments to invoke the module on, including the self argumentoutputcontains the output of the traced function in itsargs[0]attribute. This corresponds to the "return" statement in the Graph printout.
相比torchscript的IR，FX的可就簡單多了，我們理解使用起來也很簡單。
symbolic tracer
回到一開頭示例的那段程式碼，其中有一行是symbolic_traced : torch.fx.GraphModule = symbolic_trace(module)，這裡核心就是symbolic_trace函式，也就是FX解析、轉換模型的起點。這個函式其實內部是這樣的：
@compatibility(is_backward_compatible=True)
def symbolic_trace(root : Union[torch.nn.Module, Callable[..., Any]], concrete_args: Optional[Dict[str, Any]] = None,
enable_cpatching: bool = False) -> GraphModule:
"""
Symbolic tracing API
Given an ``nn.Module`` or function instance ``root``, this function will return a ``GraphModule``
constructed by recording operations seen while tracing through ``root``.
...
"""
tracer = Tracer(enable_cpatching=enable_cpatching)
graph = tracer.trace(root, concrete_args)
name = root.__class__.__name__ if isinstance(root, torch.nn.Module) else root.__name__
return GraphModule(tracer.root, graph, name)
首先會建立一個Tracer類然後使用成員函式trace我們的torch.nn.Module。我們在trace這個模型之後，就可以對這個模型進行修改了：
def transform(m: nn.Module,
tracer_class : type = torch.fx.Tracer) -> torch.nn.Module:
# Step 1: Acquire a Graph representing the code in `m`
# 使用 Tracer 類物件去trace模型 m
# 這邊是拆開了，這個transform函式就是實現torch.fx.symbolic_trace的功能
graph : torch.fx.Graph = tracer_class().trace(m)
# Step 2: 這裡就可以任意修改模型了，這也是重點
graph = ...
# Step 3: Construct a Module to return
return torch.fx.GraphModule(m, graph)
修改之後的模型可以直接拿來用，也可以通過graph_module.to_folder，把這個模型摘出來當做單獨的模組去使用（這個之後說）。整體的流程大概就是這樣：
symbolic tracing -> intermediate representation -> transforms -> Python code generation。
各自的功能為：
- symbolic
The symbolic tracer performs “symbolic execution” of the Python code. It feeds fake values, called Proxies, through the code. Operations on theses Proxies are recorded. More information about symbolic tracing can be found in the symbolic_trace() and Tracer documentation.
- intermediate representation
The intermediate representation is the container for the operations that were recorded during symbolic tracing. It consists of a list of Nodes that represent function inputs, callsites (to functions, methods, or torch.nn.Module instances), and return values. More information about the IR can be found in the documentation for Graph. The IR is the format on which transformations are applied.
- Python code generation
Python code generation is what makes FX a Python-to-Python (or Module-to-Module) transformation toolkit. For each Graph IR, we can create valid Python code matching the Graph’s semantics. This functionality is wrapped up in GraphModule, which is a torch.nn.Module instance that holds a Graph as well as a forward method generated from the Graph.
上述就是FX的三個核心功能。
Proxy/Retracing是symbolic trace的核心。因為我對Proxy/Retracing的理解還不是很深，這裡就不擅自描述了，摘一下官方的介紹：
Proxy objects are Node wrappers that flow through the program during symbolic tracing and record all the operations (torch function calls, method calls, operators) that they touch into the growing FX Graph.
If you’re doing graph transforms, you can wrap your own Proxy method around a raw Node so that you can use the overloaded operators to add additional things to a Graph.
相關結構
FX主要的結構就是Graph和GraphModule了，其中A Graph is a data structure that represents a method on a GraphModule。可以理解為Graph中存放著網路中最關鍵的Node，這些node就是網路中的一個個節點（比如卷積、relu、add、concat等等），這些node記錄了對應的method和輸入輸出資訊，從而可以串起來組成網路的邏輯。
通過print_tabular()可以將graph中的node資訊列印出來：
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)
# 這裡列印module中的node
gm.graph.print_tabular()
列印資訊如下：
可以看到，對於輸入x，對應的IR型別是placeholder；對於權重資訊，對應的IR型別是get_attr；對於具體的實際操作（add、linear、sum、relu、topk等），對應著call_function、call_module這倆IR，最後的輸出對應著output這個IR。
同時還列印了每個node的輸入資訊和額外的引數資訊，通過這些資訊就可以把node連起來。
不過光有graph是不夠的，還需要GraphModule。GraphModule繼承於torch.nn.Module，包含了前向forward函式和網路中模組需要的引數，這些引數會被graph中的node呼叫。
總結一下，那就是graph中的node包含了網路的邏輯資訊，然後這些node前後呼叫關係會被FX重新組合為GraphModule中FX生成的前向forward程式碼（可以通過traced.code列印出來），而這些生成的程式碼會需要GraphModule中的引數資訊來保證順利執行。
修改Graph
既然知道graph中包含了網路的順序執行資訊，那麼想要修改網路，直接修改node就可以：
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)
# 對於graph中的node，FX會以順序的形式來表示這個網路
# 所以我們可以直接for迴圈來遍歷：
for node in graph.nodes:
# 檢測該node的IR型別是否是call_function
if node.op == 'call_function':
# 修改node.target為torch.mul，網路也因此變了
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)
簡單提一句，node.target代表call_function中call的是哪個target，而torch.add也就是pytorch自帶的操作op，呼叫這個node的時候會實際呼叫到torch.add。
優雅地修改graph網路
上述直接修改簡單粗暴，FX也貼心地為我們提供了Graph rewrites工具，我們可以藉助這些工具方便地增加或者刪除某一個node：
# 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)
藉助replace_pattern來修改網路
Graph rewrites工具都有了（相關概念是來源於編譯器），那麼match pattern肯定也有了，我們可以通過 replace_pattern()來對整個graph進行修改。pattern的話可以用fx自帶的也可以自己新增自己的規則:
# Sample module
class M(torch.nn.Module):
def __init__(self):
super().__init__()
def forward(self, x, w1, w2):
val1 = torch.neg(w1)
m1 = torch.cat([val1, w2]).sum()
val2 = torch.neg(w1)
m2 = torch.cat([val2, w2]).sum()
return x + torch.max(m1) + torch.max(m2)
# Symbolically trace an instance of `M`
traced = symbolic_trace(M())
# Define the pattern.
def pattern(a1, a2):
val1 = torch.neg(a1)
return torch.cat([val1, a2]).sum()
# Define the replacement (same rules as the pattern)
def replacement(w1, w2):
return torch.stack([w1, w2])
# Replace `pattern` with `replacement` in `traced`
replace_pattern(traced, pattern, replacement)
# After calling `replace_pattern`, the generated code is:
'''
def forward(self, x, w1, w2):
stack = torch.stack([w1, w2])
max_1 = torch.max(stack); stack = None
add = x + max_1; x = max_1 = None
stack_1 = torch.stack([w1, w2]); w1 = w2 = None
max_2 = torch.max(stack_1); stack_1 = None
add_1 = add + max_2; add = max_2 = None
return add_1
'''
Interpreter
Interpreter，即直譯器，這個名字用的好。其實就是以一個比較優雅的方式迴圈一個Graph的node並且執行它們，並同時順帶完成一些任務。比如我們想看下模型在執行期間每一層的shape變化：
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)
上述的propagate函式很簡單，遍歷一遍node然後記錄資訊到node.shape和node.dtype中。FX也提供了interpreter類，存放了一些util的function，我們直接繼承就可以使用（類似於上面這個ShapeProp）。
Transformer
Transformer就是對torch.nn.Module做一些變換，這些變換我們可以封裝成一個函式或者寫到類裡頭，其實Transformer也可以叫做PASS，總之就是對網路進行一些修改。比如這樣：
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`
# trace nn.Module
graph : torch.fx.Graph = tracer_class().trace(m)
# Step 2: 這裡對Graph進行修改
graph = ...
# Step 3: Construct a Module to return
return torch.fx.GraphModule(m, graph)
Your transform will take in an torch.nn.Module, acquire a Graph from it, do some modifications, and return a new torch.nn.Module. You should think of the torch.nn.Module that your FX transform returns as identical to a regular 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 torch.nn.Module will allow for composability.
當然也可以直接修改GraphModule，沒必要非要返回一個新的：
import torch
import torch.fx
def transform(m : nn.Module) -> nn.Module:
gm : torch.fx.GraphModule = torch.fx.symbolic_trace(m)
# 這裡修改 gm.graph
# <...>
# Recompile the forward() method of `gm` from its Graph
gm.recompile()
return gm
需要注意，gm.recompile()這句必須是要加上的，我們修改graph後，需要recompile來重新生成forward程式碼。
舉個FX的栗子
鋪墊了那麼多，簡單舉一個FX的實際例子吧，這裡我們用FX去量化一個基於CenterNet框架的目標檢測模型，backbone使用的是Resnet50，限於篇幅，本篇只介紹trace完模型和fuse的部分，量化和匯出trt之後的文章再說。
首先搭建CenterNet模型，然後進行trace：
model = FXCenterNet()
tracer = Tracer()
graph_module = GraphModule(model, tracer.trace(model))
其中trace的函式如下，大概就是遍歷model中的操作，按照規則轉換為node存放到graph中，包含attr和op、輸入輸出等資訊，最終返回graph這個IR結構：
@compatibility(is_backward_compatible=True)
def trace(self, root: Union[torch.nn.Module, Callable[..., Any]], concrete_args: Optional[Dict[str, Any]] = None) -> Graph:
# root FXCenterNet
if isinstance(root, torch.nn.Module):
self.root = root
fn = type(root).forward
self.submodule_paths = {mod: name for name, mod in root.named_modules()}
else:
self.root = torch.nn.Module()
fn = root
tracer_cls: Optional[Type['Tracer']] = getattr(self, '__class__', None)
self.graph = Graph(tracer_cls=tracer_cls)
# 這裡大概就是遍歷root中的操作，按照規則轉換為node存放到graph中，
# 包含attr和op、輸入輸出等資訊，最終返回graph這個IR結構
...
return self.graph
生成的self.graph型別是torch.fx.graph.Graph。
self.graph
<torch.fx.graph.Graph object at 0x7f57f59efdf0>
呼叫self.graph.print_tabular()列印graph的node資訊，可以看到熟悉的resnet-50-backbone的結構，以IR的形式組織起來：
生成graph後，開始組裝GraphModule，GraphModule是由graph生成的，GraphModule會把graph的node中的引數和模組資訊複製一份到自己：
@compatibility(is_backward_compatible=True)
class GraphModule(torch.nn.Module):
def __new__(cls: 'Type[GraphModule]', *args, **kwargs):
for t in cls.__mro__:
c = t.__qualname__.split('.')[-1]
if c != 'GraphModuleImpl':
cls = t
break
class GraphModuleImpl(cls): # type: ignore[misc, valid-type]
pass
return super().__new__(GraphModuleImpl)
@compatibility(is_backward_compatible=True)
def __init__(self,
root: Union[torch.nn.Module, Dict[str, Any]],
graph: Graph,
class_name: str = 'GraphModule'):
super().__init__()
self.__class__.__name__ = class_name
if isinstance(root, torch.nn.Module):
if hasattr(root, 'training'):
self.training = root.training
# 這裡拷貝graph中的引數資訊和模組資訊到self也就是GraphModule中
for node in graph.nodes:
if node.op in ['get_attr', 'call_module']:
assert isinstance(node.target, str)
_copy_attr(root, self, node.target)
elif isinstance(root, dict):
targets_to_copy = []
for node in graph.nodes:
if node.op in ['get_attr', 'call_module']:
assert isinstance(node.target, str)
if node.target not in root:
raise RuntimeError('Node ' + str(node) + ' referenced target ' + node.target +
' but that target was not provided in ``root``!')
targets_to_copy.append(node.target)
targets_to_copy.sort(key=lambda t: t.count('.'))
for target_to_copy in targets_to_copy:
_assign_attr(root[target_to_copy], self, target_to_copy)
else:
raise RuntimeError('Unsupported type ' + str(root) + ' passed for root!')
self.graph = graph
self._tracer_cls = None
if self.graph._tracer_cls and '<locals>' not in self.graph._tracer_cls.__qualname__:
self._tracer_cls = self.graph._tracer_cls
__jit_unused_properties__ = ['graph']
最終graph_module中包含了生成的程式碼，通過print(graph_module.code)列印出來：
def forward(self, input):
input_1 = input
upsampler_deconv_layers_0_bias = getattr(self.upsampler.deconv_layers, "0").bias
...
head_angle_0 = getattr(self.head.angle, "0")(upsampler_deconv_layers_11); upsampler_deconv_layers_11 = None
head_angle_1 = getattr(self.head.angle, "1")(head_angle_0); head_angle_0 = None
head_angle_2 = getattr(self.head.angle, "2")(head_angle_1); head_angle_1 = None
return {'hm': head_hm_2, 'wh': head_wh_2, 'reg': head_reg_2, 'angle': head_angle_2}
這個時候我們就有了trace後的Module，這個Module和原始模型並沒有區別，forward函式也是按照原始模型的forward生成的。因為我們只是簡單地trace了一遍，所以相同輸入結果也是一樣的：graph_module(input) == original_model(input)，畢竟沒幹啥特殊的。
OP融合
接下來就是fuse，這裡直接呼叫FX提供的fuse函式，其實裡頭也就是呼叫了Fuser：
def _fuse_fx(
graph_module: GraphModule,
is_qat: bool,
fuse_custom_config_dict: Optional[Dict[str, Any]] = None,
backend_config_dict: Optional[Dict[str, Any]] = None,
) -> GraphModule:
r""" Internal helper function to fuse modules in preparation for quantization
Args:
graph_module: GraphModule object from symbolic tracing (torch.fx.symbolic_trace)
"""
_check_is_graph_module(graph_module)
fuser = Fuser()
return fuser.fuse(
graph_module, is_qat, fuse_custom_config_dict, backend_config_dict)
來看看Fuser都幹了啥，其實很簡單，就是遍歷一遍input_graph = model.graph中的node，然後根據指定好的fuse規則進行融合，融合會涉及到修改graph結構：
class Fuser:
def fuse(
self,
model: GraphModule,
is_qat: bool,
fuse_custom_config_dict: Optional[Dict[str, Any]] = None,
backend_config_dict: Optional[Dict[str, Any]] = None,
) -> GraphModule:
if fuse_custom_config_dict is None:
fuse_custom_config_dict = {}
input_root = model
input_graph = model.graph
# 這裡首先copy 原始模型中的named_modules中，之後會根據fuse情況進行修改
self.modules = dict(input_root.named_modules())
...
# 這裡查詢匹配的fuse pattern
fusion_pairs = self._find_matches(
input_root, input_graph, fusion_pattern_to_fuse_handler_cls)
self.fused_graph = Graph()
env: Dict[Any, Any] = {}
def load_arg(a):
return map_arg(a, lambda node: env[node.name])
def get_root_node(node_pattern):
while not isinstance(node_pattern[-1], Node):
node_pattern = node_pattern[-1]
return node_pattern[-1]
for node in input_graph.nodes:
maybe_last_node, pattern, matched_node_pattern, obj = \
fusion_pairs.get(node.name, (None, None, None, None))
if maybe_last_node is node:
assert obj is not None
# TODO: currently we hard code the root node, which only works for
# a sequence of ops and assume the root node is the last node,
# we want to make this more general to support more complex patterns
root_node = get_root_node(matched_node_pattern) # 尋找fuse的根node
env[node.name] = obj.fuse( # 這裡將self傳入，對self進行修改
self, load_arg, root_node, matched_node_pattern, # type: ignore[arg-type]
fuse_custom_config_dict, fuser_method_mapping, is_qat)
elif maybe_last_node is None:
env[node.name] = self.fused_graph.node_copy(node, load_arg)
# node matched in patterns and is not root is removed here
preserved_attributes = set(fuse_custom_config_dict.get("preserved_attributes", []))
model = FusedGraphModule(input_root, self.fused_graph, preserved_attributes)
return model
def _find_matches(
self, root: GraphModule, graph: Graph,
patterns: Dict[Pattern, Callable]
) -> Dict[str, Tuple[Node, Pattern, NodePattern, FuseHandler]]:
modules = dict(root.named_modules())
match_map : Dict[str, Tuple[Node, Pattern, NodePattern, FuseHandler]] = {} # node name -> (root_node, match_value)
def apply_match(pattern, node, match, matched_node_pattern):
if isinstance(pattern, tuple):
s, *args = pattern
current_node_pattern: List[Node] = []
apply_match(s, node, match, current_node_pattern)
for subpattern, arg in zip(args, node.args):
apply_match(subpattern, arg, match, current_node_pattern)
matched_node_pattern.append(tuple(current_node_pattern))
else:
# the first pattern matches will take precedence
if node.name not in match_map:
matched_node_pattern.append(node)
root_node, pattern, handler = match
match_map[node.name] = (root_node, pattern, matched_node_pattern, handler)
# 這裡就是match過程
for node in reversed(graph.nodes):
if node.name not in match_map:
for pattern, value in patterns.items():
matched_node_pattern: List[Node] = []
if is_match(modules, node, pattern):
apply_match(pattern, node, (node, pattern, value(self, node)), matched_node_pattern)
return match_map
至於定義了哪些fuse的規則，可以在pytorch/torch/ao/quantization/fx/fusion_patterns.py這裡頭找到：
# /ao/quantization/fx/fusion_patterns.py
@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv1d))
@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv2d))
@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv3d))
@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv1d))
@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv2d))
@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv3d))
@register_fusion_pattern((torch.nn.BatchNorm1d, torch.nn.Conv1d))
@register_fusion_pattern((torch.nn.BatchNorm2d, torch.nn.Conv2d))
@register_fusion_pattern((torch.nn.BatchNorm3d, torch.nn.Conv3d))
@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm1d, torch.nn.Conv1d)))
@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm2d, torch.nn.Conv2d)))
@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm3d, torch.nn.Conv3d)))
@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm1d, torch.nn.Conv1d)))
@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm2d, torch.nn.Conv2d)))
@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm3d, torch.nn.Conv3d)))
@register_fusion_pattern((torch.nn.BatchNorm1d, torch.nn.Linear))
class DefaultFuseHandler(FuseHandler):
def __init__(
self,
quantizer: QuantizerCls,
node: Node):
super().__init__(quantizer, node)
def fuse(...):
# 這裡執行實際的融合操作
具體的融合操作在DefaultFuseHandler類中的fuse方法內執行，找到對應的fuser_method，然後呼叫，返回融合後的fused_module使用setattr來修改網路的modules，同樣也會通過node_copy修改graph中的node：
matched_module_types = get_matched_types(matched_modules)
module_parent_name, module_name = _parent_name(root_node.target)
fuser_method = get_fuser_method_new(matched_module_types, fuser_method_mapping)
# TODO: change the signature for fuser_method to take matched module patterns
# as input
fused_module = fuser_method(is_qat, *matched_modules)
# TODO: maybe add a pass to cleanup bn modules?
setattr(quantizer.modules[module_parent_name], module_name, fused_module) # 往fuse控制的新模型中加入 新的modules
return quantizer.fused_graph.node_copy(root_node, load_arg) # 往fuse控制的新graph中加入forward引數
其中，Conv+bn+relu的融合細節會呼叫pytorch/torch/ao/quantization/fuser_method_mappings.py中的fuse_conv_bn_relu函式：
def fuse_conv_bn_relu(is_qat, conv, bn, relu):
assert(conv.training == bn.training == relu.training),\
"Conv and BN both must be in the same mode (train or eval)."
fused_module : Optional[Type[nn.Sequential]] = None
map_to_fused_module_eval = {
nn.Conv1d: nni.ConvReLU1d,
nn.Conv2d: nni.ConvReLU2d,
nn.Conv3d: nni.ConvReLU3d,
}
fused_module = map_to_fused_module_eval.get(type(conv), None)
if fused_module is not None:
fused_conv = nn.utils.fusion.fuse_conv_bn_eval(conv, bn)
return fused_module(fused_conv, relu)
else:
raise NotImplementedError("Cannot fuse eval modules: {}".format((conv, bn, relu)))
上述fused_module為torch.nn.intrinsic.modules.fused.ConvReLU2d類，會呼叫fuse_conv_bn_eval來實際吸bn到conv：
def fuse_conv_bn_eval(conv, bn, transpose=False):
assert(not (conv.training or bn.training)), "Fusion only for eval!"
fused_conv = copy.deepcopy(conv)
fused_conv.weight, fused_conv.bias = \
fuse_conv_bn_weights(fused_conv.weight, fused_conv.bias,
bn.running_mean, bn.running_var, bn.eps, bn.weight, bn.bias, transpose)
return fused_conv
def fuse_conv_bn_weights(conv_w, conv_b, bn_rm, bn_rv, bn_eps, bn_w, bn_b, transpose=False):
if conv_b is None:
conv_b = torch.zeros_like(bn_rm)
if bn_w is None:
bn_w = torch.ones_like(bn_rm)
if bn_b is None:
bn_b = torch.zeros_like(bn_rm)
bn_var_rsqrt = torch.rsqrt(bn_rv + bn_eps)
if transpose:
shape = [1, -1] + [1] * (len(conv_w.shape) - 2)
else:
shape = [-1, 1] + [1] * (len(conv_w.shape) - 2)
conv_w = conv_w * (bn_w * bn_var_rsqrt).reshape(shape)
conv_b = (conv_b - bn_rm) * bn_var_rsqrt * bn_w + bn_b
return torch.nn.Parameter(conv_w), torch.nn.Parameter(conv_b)
吸完後，得到新的conv，然後帶入ConvReLU2d類。
class ConvReLU2d(_FusedModule):
r"""This is a sequential container which calls the Conv2d and ReLU modules.
During quantization this will be replaced with the corresponding fused module."""
def __init__(self, conv, relu):
assert type(conv) == Conv2d and type(relu) == ReLU, \
'Incorrect types for input modules{}{}'.format(
type(conv), type(relu))
super().__init__(conv, relu)
整體流程就是conv + bn->conv然後conv + relu -> ConvReLU2d。
fuse後的code就清爽很多了，bn和relu都被融合進去了（當然還有其他融合）:
def forward(self, input):
input_1 = input
backbone_conv1 = self.backbone.conv1(input_1)
backbone_maxpool = self.backbone.maxpool(backbone_relu)
backbone_layer1_0_conv1 = getattr(self.backbone.layer1, "0").conv1(backbone_maxpool)
backbone_layer1_0_conv2 = getattr(self.backbone.layer1, "0").conv2(backbone_layer1_0_relu)
backbone_layer1_0_conv3 = getattr(self.backbone.layer1, "0").conv3(backbone_layer1_0_relu_1)
...
head_reg_0 = getattr(self.head.reg, "0")(upsampler_deconv_layers_11)
head_reg_2 = getattr(self.head.reg, "2")(head_reg_1)
head_angle_0 = getattr(self.head.angle, "0")(upsampler_deconv_layers_11)
head_angle_2 = getattr(self.head.angle, "2")(head_angle_1)
return {'hm': head_hm_2, 'wh': head_wh_2, 'reg': head_reg_2, 'angle': head_angle_2}
至此，就得到了trace後和fuse後的模型，可以看到融合後的ConvReLU2d模組。
這個GraphModule和torch.nn.module的使用方式一模一樣，可以簡單輸入一個image驗證一下。
下一篇中我們會對這個GraphModule進行量化操作。
如何debug
那麼我們得到了最終的GraphModule，該如何debug呢，也就是一步一步單獨除錯。這也是有辦法的，除錯fx生成model的方式有三種：
直接通過pdb進行debug
我們是可以進入FX的Generated Code中的，也可以設定斷點：
列印生成的程式碼，並且和Module組合
因為graph中的node包含了指定邏輯，GraphModule中包含了模型權重等資訊，而這些權重資訊是通過原始的Module獲取的，那麼我們可以直接將生成的code放到原始Module子類的forward中，組成一個新的Module來呼叫。
# 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
"""
# 這裡繼承原始的Module
class SubclassM(M):
def __init__(self):
super().__init__()
# 把生成的程式碼粘到這裡
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()
是不是很符合常識！
使用to_folder函式
就像之前例子裡說到的，GraphModule.to_folder()是一個神奇的函式，可以直接將FX生成的module匯出為一個資料夾，資料夾中包含了模型需要的引數(.pt格式)和模型的定義。
而module.py的程式碼也幫你生成好了：
# 匯出的module.py中
import torch
from torch.nn import *
class FusedModel(torch.nn.Module):
def __init__(self):
super().__init__()
# 這裡載入權重引數資訊
self.backbone = torch.load(r'fx_debug/backbone.pt')
self.load_state_dict(torch.load(r'fx_debug/state_dict.pt'))
...
def forward(self, input):
# 這裡就是生成的code部分，也幫你寫到forward中了
input_1 = input
backbone_conv1 = self.backbone.conv1(input_1)
backbone_maxpool = self.backbone.maxpool(backbone_relu)
backbone_layer1_0_conv1 = getattr(self.backbone.layer1, "0").conv1(backbone_maxpool)
...
head_angle_0 = getattr(self.head.angle, "0")(upsampler_deconv_layers_11)
head_angle_2 = getattr(self.head.angle, "2")(head_angle_1)
return {'hm': head_hm_2, 'wh': head_wh_2, 'reg': head_reg_2, 'angle': head_angle_2}
是不是很強大？！
我們也可以修改這個生成的程式碼來做其他方面的實驗（不過這個匯出有一些bug，不知道是不是我使用姿勢不對）。
一些限制
torch.fx也是有一些限制的（畢竟不可能十全十美）。
因為Symbolic execution的限制。
Proxy object cannot be iterated. This can be attempted when the Proxy is used in a loop or as a *args or **kwargs function argument. See the torch.fx docs on pytorch.org for a more detailed explanation of what types of control flow can be traced, and check out the Proxy docstring for help troubleshooting Proxy iteration errors
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.
更詳細的限制可以看官方的介紹：
就先寫到這裡吧，關於FX的功能使用更多是在量化過程中體現了，下一篇的量化實操中，會結合量化過程來理解FX，同時也會總結下PTQ量化的流程和注意點，我是老潘，我們下一篇再見~
撩我吧
- 如果你與我志同道合於此，老潘很願意與你交流！
- 如果你喜歡老潘的內容，歡迎關注和支援~
- 如果有問題想要聯絡我，可加公眾號直接私信，點這裡！