使用自定义 C++ 类扩展 PyTorch#
本教程介绍了一个将 C++ 类绑定到 PyTorch 的 API。该 API 与 pybind11 非常相似，如果您熟悉该系统，大多数概念都将适用。
在 C++ 中实现和绑定类#
在本教程中，我们将定义一个简单的 C++ 类，该类在一个成员变量中维护持久状态。
// This header is all you need to do the C++ portions of this
// tutorial
#include <torch/script.h>
// This header is what defines the custom class registration
// behavior specifically. script.h already includes this, but
// we include it here so you know it exists in case you want
// to look at the API or implementation.
#include <torch/custom_class.h>
#include <string>
#include <vector>
template <class T>
struct MyStackClass : torch::CustomClassHolder {
std::vector<T> stack_;
MyStackClass(std::vector<T> init) : stack_(init.begin(), init.end()) {}
void push(T x) {
stack_.push_back(x);
}
T pop() {
auto val = stack_.back();
stack_.pop_back();
return val;
}
c10::intrusive_ptr<MyStackClass> clone() const {
return c10::make_intrusive<MyStackClass>(stack_);
}
void merge(const c10::intrusive_ptr<MyStackClass>& c) {
for (auto& elem : c->stack_) {
push(elem);
}
}
};
有几点需要注意
要通过自定义类扩展 PyTorch，您需要包含
torch/custom_class.h头文件。请注意，每当我们处理自定义类的实例时，我们都是通过
c10::intrusive_ptr<>的实例来操作的。您可以将intrusive_ptr视为智能指针，类似于std::shared_ptr，但引用计数直接存储在对象本身中，而不是像std::shared_ptr那样存储在单独的元数据块中。torch::Tensor内部使用相同的指针类型；自定义类也必须使用此指针类型，以便我们能够一致地管理不同的对象类型。第二点需要注意是，用户定义的类必须继承自
torch::CustomClassHolder。这确保了自定义类有空间来存储引用计数。
现在让我们看看如何使这个类对 PyTorch 可见，这个过程称为绑定类。
// Notice a few things:
// - We pass the class to be registered as a template parameter to
// `torch::class_`. In this instance, we've passed the
// specialization of the MyStackClass class ``MyStackClass<std::string>``.
// In general, you cannot register a non-specialized template
// class. For non-templated classes, you can just pass the
// class name directly as the template parameter.
// - The arguments passed to the constructor make up the "qualified name"
// of the class. In this case, the registered class will appear in
// Python and C++ as `torch.classes.my_classes.MyStackClass`. We call
// the first argument the "namespace" and the second argument the
// actual class name.
TORCH_LIBRARY(my_classes, m) {
m.class_<MyStackClass<std::string>>("MyStackClass")
// The following line registers the contructor of our MyStackClass
// class that takes a single `std::vector<std::string>` argument,
// i.e. it exposes the C++ method `MyStackClass(std::vector<T> init)`.
// Currently, we do not support registering overloaded
// constructors, so for now you can only `def()` one instance of
// `torch::init`.
.def(torch::init<std::vector<std::string>>())
// The next line registers a stateless (i.e. no captures) C++ lambda
// function as a method. Note that a lambda function must take a
// `c10::intrusive_ptr<YourClass>` (or some const/ref version of that)
// as the first argument. Other arguments can be whatever you want.
.def("top", [](const c10::intrusive_ptr<MyStackClass<std::string>>& self) {
return self->stack_.back();
})
// The following four lines expose methods of the MyStackClass<std::string>
// class as-is. `torch::class_` will automatically examine the
// argument and return types of the passed-in method pointers and
// expose these to Python and TorchScript accordingly. Finally, notice
// that we must take the *address* of the fully-qualified method name,
// i.e. use the unary `&` operator, due to C++ typing rules.
.def("push", &MyStackClass<std::string>::push)
.def("pop", &MyStackClass<std::string>::pop)
.def("clone", &MyStackClass<std::string>::clone)
.def("merge", &MyStackClass<std::string>::merge)
;
}
使用 CMake 将示例构建为 C++ 项目#
现在，我们将使用 CMake 构建系统来构建上述 C++ 代码。首先，将我们到目前为止介绍的所有 C++ 代码放入一个名为 class.cpp 的文件中。然后，编写一个简单的 CMakeLists.txt 文件并将其放在同一目录下。以下是 CMakeLists.txt 的样子：
cmake_minimum_required(VERSION 3.1 FATAL_ERROR)
project(custom_class)
find_package(Torch REQUIRED)
# Define our library target
add_library(custom_class SHARED class.cpp)
set(CMAKE_CXX_STANDARD 14)
# Link against LibTorch
target_link_libraries(custom_class "${TORCH_LIBRARIES}")
另外，创建一个 build 目录。您的文件树应该如下所示：
custom_class_project/
class.cpp
CMakeLists.txt
build/
继续调用 cmake 然后 make 来构建项目。
$ cd build
$ cmake -DCMAKE_PREFIX_PATH="$(python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)')" ..
-- The C compiler identification is GNU 7.3.1
-- The CXX compiler identification is GNU 7.3.1
-- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc
-- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++
-- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Found torch: /torchbind_tutorial/libtorch/lib/libtorch.so
-- Configuring done
-- Generating done
-- Build files have been written to: /torchbind_tutorial/build
$ make -j
Scanning dependencies of target custom_class
[ 50%] Building CXX object CMakeFiles/custom_class.dir/class.cpp.o
[100%] Linking CXX shared library libcustom_class.so
[100%] Built target custom_class
您会发现在 build 目录中有一个（除其他内容外）动态库文件。在 Linux 上，它可能名为 libcustom_class.so。因此，文件树应该看起来像：
custom_class_project/
class.cpp
CMakeLists.txt
build/
libcustom_class.so
从 Python 使用 C++ 类#
现在我们已经将类及其注册编译成一个 .so 文件，我们可以将该 .so 文件加载到 Python 中并进行尝试。这是一个演示脚本：
import torch
# `torch.classes.load_library()` allows you to pass the path to your .so file
# to load it in and make the custom C++ classes available to both Python and
# TorchScript
torch.classes.load_library("build/libcustom_class.so")
# You can query the loaded libraries like this:
print(torch.classes.loaded_libraries)
# prints {'/custom_class_project/build/libcustom_class.so'}
# We can find and instantiate our custom C++ class in python by using the
# `torch.classes` namespace:
#
# This instantiation will invoke the MyStackClass(std::vector<T> init)
# constructor we registered earlier
s = torch.classes.my_classes.MyStackClass(["foo", "bar"])
# We can call methods in Python
s.push("pushed")
assert s.pop() == "pushed"
# Test custom operator
s.push("pushed")
torch.ops.my_classes.manipulate_instance(s) # acting as s.pop()
assert s.top() == "bar"
# Returning and passing instances of custom classes works as you'd expect
s2 = s.clone()
s.merge(s2)
for expected in ["bar", "foo", "bar", "foo"]:
assert s.pop() == expected
# We can also use the class in TorchScript
# For now, we need to assign the class's type to a local in order to
# annotate the type on the TorchScript function. This may change
# in the future.
MyStackClass = torch.classes.my_classes.MyStackClass
@torch.jit.script
def do_stacks(s: MyStackClass): # We can pass a custom class instance
# We can instantiate the class
s2 = torch.classes.my_classes.MyStackClass(["hi", "mom"])
s2.merge(s) # We can call a method on the class
# We can also return instances of the class
# from TorchScript function/methods
return s2.clone(), s2.top()
stack, top = do_stacks(torch.classes.my_classes.MyStackClass(["wow"]))
assert top == "wow"
for expected in ["wow", "mom", "hi"]:
assert stack.pop() == expected
为自定义 C++ 类定义序列化/反序列化方法#
如果您尝试保存一个 ScriptModule，其中一个自定义绑定的 C++ 类作为属性，您将收到以下错误：
# export_attr.py
import torch
torch.classes.load_library('build/libcustom_class.so')
class Foo(torch.nn.Module):
def __init__(self):
super().__init__()
self.stack = torch.classes.my_classes.MyStackClass(["just", "testing"])
def forward(self, s: str) -> str:
return self.stack.pop() + s
scripted_foo = torch.jit.script(Foo())
scripted_foo.save('foo.pt')
loaded = torch.jit.load('foo.pt')
print(loaded.stack.pop())
$ python export_attr.py
RuntimeError: Cannot serialize custom bound C++ class __torch__.torch.classes.my_classes.MyStackClass. Please define serialization methods via def_pickle for this class. (pushIValueImpl at ../torch/csrc/jit/pickler.cpp:128)
这是因为 PyTorch 无法自动确定您要从 C++ 类中保存哪些信息。您必须手动指定。方法是使用 class_ 上的特殊 def_pickle 方法在类上定义 __getstate__ 和 __setstate__ 方法。
注意
__getstate__ 和 __setstate__ 的语义等同于 Python pickle 模块的语义。您可以 阅读更多关于我们如何使用这些方法的信息。
以下是我们可以在 MyStackClass 的注册中添加的 def_pickle 调用的示例，以包含序列化方法：
// class_<>::def_pickle allows you to define the serialization
// and deserialization methods for your C++ class.
// Currently, we only support passing stateless lambda functions
// as arguments to def_pickle
.def_pickle(
// __getstate__
// This function defines what data structure should be produced
// when we serialize an instance of this class. The function
// must take a single `self` argument, which is an intrusive_ptr
// to the instance of the object. The function can return
// any type that is supported as a return value of the TorchScript
// custom operator API. In this instance, we've chosen to return
// a std::vector<std::string> as the salient data to preserve
// from the class.
[](const c10::intrusive_ptr<MyStackClass<std::string>>& self)
-> std::vector<std::string> {
return self->stack_;
},
// __setstate__
// This function defines how to create a new instance of the C++
// class when we are deserializing. The function must take a
// single argument of the same type as the return value of
// `__getstate__`. The function must return an intrusive_ptr
// to a new instance of the C++ class, initialized however
// you would like given the serialized state.
[](std::vector<std::string> state)
-> c10::intrusive_ptr<MyStackClass<std::string>> {
// A convenient way to instantiate an object and get an
// intrusive_ptr to it is via `make_intrusive`. We use
// that here to allocate an instance of MyStackClass<std::string>
// and call the single-argument std::vector<std::string>
// constructor with the serialized state.
return c10::make_intrusive<MyStackClass<std::string>>(std::move(state));
});
注意
我们在 pickle API 中采取了与 pybind11 不同的方法。pybind11 有一个特殊的函数 pybind11::pickle()，您将其传递给 class_::def()，而我们有一个单独的 def_pickle 方法用于此目的。这是因为 torch::jit::pickle 这个名字已经被占用了，我们不想引起混淆。
一旦我们这样定义了（反）序列化行为，我们的脚本现在就可以成功运行了。
$ python ../export_attr.py
testing
定义接受或返回绑定 C++ 类的自定义运算符#
一旦您定义了自定义 C++ 类，您还可以将该类用作自定义运算符（即自由函数）的参数或返回值。假设您有以下自由函数：
c10::intrusive_ptr<MyStackClass<std::string>> manipulate_instance(const c10::intrusive_ptr<MyStackClass<std::string>>& instance) {
instance->pop();
return instance;
}
您可以通过在 TORCH_LIBRARY 块中运行以下代码来注册它：
m.def(
"manipulate_instance(__torch__.torch.classes.my_classes.MyStackClass x) -> __torch__.torch.classes.my_classes.MyStackClass Y",
manipulate_instance
);
完成此操作后，您可以像下面的示例一样使用该运算符：
class TryCustomOp(torch.nn.Module):
def __init__(self):
super(TryCustomOp, self).__init__()
self.f = torch.classes.my_classes.MyStackClass(["foo", "bar"])
def forward(self):
return torch.ops.my_classes.manipulate_instance(self.f)
注意
接受 C++ 类作为参数的运算符的注册要求该自定义类已经注册。您可以通过确保自定义类注册和您的自由函数定义在同一个 TORCH_LIBRARY 块中，并且自定义类注册在前来强制执行此操作。将来，我们可能会放宽此要求，以便可以按任何顺序注册这些。
结论#
本教程介绍了如何将 C++ 类公开给 PyTorch，如何注册其方法，如何从 Python 使用该类，以及如何使用该类保存和加载代码并在独立的 C++ 进程中运行该代码。您现在可以扩展您的 PyTorch 模型，使用与第三方 C++ 库交互的 C++ 类，或实现任何其他需要平滑融合 Python 和 C++ 之间界限的用例。
一如既往，如果您遇到任何问题或有疑问，可以使用我们的 论坛 或 GitHub issues 进行联系。此外，我们的 常见问题解答 (FAQ) 页面 可能包含有用的信息。
