目录

使用自定义 C++ 类扩展 TorchScript

创建时间:2020年1月23日 | 最后更新时间:2024年12月2日 | 最后验证时间:2024年11月5日

警告

TorchScript 已经不再处于活跃开发阶段。

本教程是自定义操作教程的后续内容,并介绍了我们将C++类绑定到TorchScript和Python同时使用的API。该API与pybind11非常相似,如果您熟悉该系统,大部分概念都可以迁移。

在C++中实现和绑定类

对于本教程,我们将定义一个简单的 C++ 类,该类通过成员变量维护持久化状态。

// 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);
    }
  }
};

有几个需要注意的地方:

  • torch/custom_class.h 是你需要包含的头部,以使用自定义类扩展 TorchScript。

  • 请注意,当我们处理自定义类的实例时,我们始终通过 c10::intrusive_ptr<> 的实例来操作。将 intrusive_ptr 理解为一个智能指针,类似于 std::shared_ptr,但引用计数直接存储在对象内部,而不是像 std::shared_ptr 那样存储在一个单独的元数据块中。torch::Tensor 内部也使用相同的指针类型;自定义类也必须使用这种指针类型,以便我们可以一致地管理不同类型的对象。

  • 需要注意的第二点是,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TorchScript可见,这个过程称为 绑定 该类:

// 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 构建系统构建上述 C++ 代码。首先,将我们迄今为止所涵盖的所有 C++ 代码放入一个名为 class.cpp 的文件中。 然后,在同一目录下编写一个简单的 CMakeLists.txt 文件。以下是 CMakeLists.txt 应该的样子:

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}")

此外,创建一个 build 目录。你的文件树应该如下所示:

custom_class_project/
  class.cpp
  CMakeLists.txt
  build/

我们假设您已经按照前面教程中描述的方式设置了环境。 继续并调用 cmake,然后运行 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

你会发现,在构建目录中现在(以及其他内容)存在一个动态库文件。在 Linux 上,它可能被命名为 libcustom_class.so。因此,文件树应该如下所示:

custom_class_project/
  class.cpp
  CMakeLists.txt
  build/
    libcustom_class.so

从 Python 和 TorchScript 使用 C++ 类

现在我们已经将我们的类及其注册编译到一个 .so 文件中, 我们可以将该 .so 文件加载到 Python 中并进行尝试。以下是一个脚本, 演示了这一点:

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

使用自定义类保存、加载和运行TorchScript代码

我们还可以在使用 libtorch 的 C++ 进程中使用自定义注册的 C++ 类。例如,让我们定义一个简单的 nn.Module,它实例化并调用我们的 MyStackClass 类的方法:

import torch

torch.classes.load_library('build/libcustom_class.so')


class Foo(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, s: str) -> str:
        stack = torch.classes.my_classes.MyStackClass(["hi", "mom"])
        return stack.pop() + s


scripted_foo = torch.jit.script(Foo())
print(scripted_foo.graph)

scripted_foo.save('foo.pt')

foo.pt 在我们的文件系统中现在包含我们刚刚定义的序列化 TorchScript 程序。

现在,我们将定义一个新的CMake项目,以展示如何加载此模型及其所需的.so文件。有关如何执行此操作的完整说明,请查看 在C++中加载TorchScript模型教程

同样如之前所述,让我们创建一个包含以下内容的文件结构:

cpp_inference_example/
  infer.cpp
  CMakeLists.txt
  foo.pt
  build/
  custom_class_project/
    class.cpp
    CMakeLists.txt
    build/

注意到我们已经复制了序列化的 foo.pt 文件,以及来自上方 custom_class_project 的源代码树。我们将把 custom_class_project 添加为这个 C++ 项目的依赖项,以便我们可以将自定义类编译到二进制文件中。

让我们用以下内容填充 infer.cpp

#include <torch/script.h>

#include <iostream>
#include <memory>

int main(int argc, const char* argv[]) {
  torch::jit::Module module;
  try {
    // Deserialize the ScriptModule from a file using torch::jit::load().
    module = torch::jit::load("foo.pt");
  }
  catch (const c10::Error& e) {
    std::cerr << "error loading the model\n";
    return -1;
  }

  std::vector<c10::IValue> inputs = {"foobarbaz"};
  auto output = module.forward(inputs).toString();
  std::cout << output->string() << std::endl;
}

同样,让我们定义我们的 CMakeLists.txt 文件:

cmake_minimum_required(VERSION 3.1 FATAL_ERROR)
project(infer)

find_package(Torch REQUIRED)

add_subdirectory(custom_class_project)

# Define our library target
add_executable(infer infer.cpp)
set(CMAKE_CXX_STANDARD 14)
# Link against LibTorch
target_link_libraries(infer "${TORCH_LIBRARIES}")
# This is where we link in our libcustom_class code, making our
# custom class available in our binary.
target_link_libraries(infer -Wl,--no-as-needed custom_class)

你知道的: cd buildcmakemake

$ 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: /local/miniconda3/lib/python3.7/site-packages/torch/lib/libtorch.so
  -- Configuring done
  -- Generating done
  -- Build files have been written to: /cpp_inference_example/build
$ make -j
  Scanning dependencies of target custom_class
  [ 25%] Building CXX object custom_class_project/CMakeFiles/custom_class.dir/class.cpp.o
  [ 50%] Linking CXX shared library libcustom_class.so
  [ 50%] Built target custom_class
  Scanning dependencies of target infer
  [ 75%] Building CXX object CMakeFiles/infer.dir/infer.cpp.o
  [100%] Linking CXX executable infer
  [100%] Built target infer

现在我们可以运行我们的激动人心的 C++ 二进制文件了:

$ ./infer
  momfoobarbaz

不可思议!

将自定义类移动到/从 IValues

也有可能你需要将自定义类从或移入 IValue``s, such as when you take or return ``IValue``s from TorchScript methods or you want to instantiate a custom class attribute in C++. For creating an ``IValue 的自定义 C++ 类实例:

  • torch::make_custom_class<T>() 提供了一个类似于 c10::intrusive_ptr<T> 的 API,它会接收你提供的任何一组参数,调用与这些参数匹配的 T 类型构造函数,并将实例包装起来返回。然而,它不仅仅返回一个指向自定义类对象的指针,而是返回一个 IValue 包装的对象。你可以直接将这个 IValue 传递给 TorchScript。

  • 如果已经有一个指向您类的 intrusive_ptr,您可以直接使用构造函数 IValue(intrusive_ptr<T>) 从它构建一个 IValue。

IValue 转换回自定义类:

  • IValue::toCustomClass<T>() 将返回一个指向 intrusive_ptr<T> 的引用,该引用指向 自定义类,而 IValue 包含此自定义类。内部,此函数会检查 T 是否已注册为自定义类,并且 IValue 是否确实包含 一个自定义类。您可以手动通过调用 isCustomClass() 来检查 IValue 是否包含自定义类。

定义自定义 C++ 类的序列化/反序列化方法

如果你尝试将一个带有自定义绑定的 C++ 类作为属性保存为 ScriptModule,你会得到以下错误:

# 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 TorchScript 无法自动确定从你的 C++ 类中保存哪些信息。您必须手动指定这些信息。实现这一点的方法是在类上定义 __getstate____setstate__ 方法,并使用 def_pickle 方法上的特殊方法在 class_ 上进行操作。

注意

TorchScript 中 __getstate____setstate__ 的语义与 Python 的 pickle 模块等效。你可以 了解更多 关于我们如何使用这些方法的信息。

这是一个示例,我们可以将def_pickle调用添加到注册 MyStackClass中以包含序列化方法:

    // 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

我们对pickle API采取了与pybind11不同的方法。与pybind11作为特殊函数pybind11::pickle()并将其传递给class_::def()不同,我们为此目的提供了一个单独的方法def_pickle。这是因为名称torch::jit::pickle已经被占用,我们不想引起混淆。

一旦我们以这种方式定义了(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

定义自定义操作符,这些操作符可以接受或返回绑定的C++类

一旦你定义了一个自定义的 C++ 类,你也可以将该类用作自定义操作符(即自由函数)的参数或返回值。假设你有以下这个自由函数:

c10::intrusive_ptr<MyStackClass<std::string>> manipulate_instance(const c10::intrusive_ptr<MyStackClass<std::string>>& instance) {
  instance->pop();
  return instance;
}

你可以在以下代码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
    );

请参阅 自定义操作教程 以获取有关注册 API 的更多详细信息。

完成这一步后,您可以像以下示例一样使用该操作:

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

注册一个以C++类作为参数的操作符需要确保该自定义类已经被注册。你可以通过确保自定义类的注册和你的自由函数定义位于同一个TORCH_LIBRARY代码块中,并且自定义类的注册优先于自由函数定义来强制实现这一点。在未来,我们可能会放宽这一要求,使得这些可以按任意顺序注册。

结论

本教程引导你了解如何将一个 C++ 类暴露给 TorchScript(以及由此扩展到 Python),如何注册其方法,如何从 Python 和 TorchScript 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。你现在已准备好通过与第三方 C++ 库交互或实现任何需要 Python、TorchScript 和 C++ 之间无缝融合的用例的 C++ 类来扩展你的 TorchScript 模型。

一如既往,如果您遇到任何问题或有任何疑问,您可以使用我们的 论坛GitHub 问题 与我们联系。此外,我们的 常见问题解答 (FAQ) 页面 可能包含有用的信息。

文档

访问 PyTorch 的全面开发人员文档

查看文档

教程

获取面向初学者和高级开发人员的深入教程

查看教程

资源

查找开发资源并解答您的问题

查看资源