使用自定义 C ++类扩展 TorchScript

原文： https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html

本教程是自定义运算符教程的后续教程，并介绍了我们为将 C ++类同时绑定到 TorchScript 和 Python 而构建的 API。 该 API 与 pybind11 非常相似，如果您熟悉该系统，则大多数概念都将转移过来。

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，该类在成员变量中保持持久状态。

  1. // This header is all you need to do the C++ portions of this
  2. // tutorial
  3. #include <torch/script.h>
  4. // This header is what defines the custom class registration
  5. // behavior specifically. script.h already includes this, but
  6. // we include it here so you know it exists in case you want
  7. // to look at the API or implementation.
  8. #include <torch/custom_class.h>
  10. #include <string>
  11. #include <vector>
  13. template <class T>
  14. struct Stack : torch::jit::CustomClassHolder {
  15.   std::vector<T> stack_;
  16.   Stack(std::vector<T> init) : stack_(init.begin(), init.end()) {}
  18.   void push(T x) {
  19.     stack_.push_back(x);
  20.   }
  21.   T pop() {
  22.     auto val = stack_.back();
  23.     stack_.pop_back();
  24.     return val;
  25.   }
  27.   c10::intrusive_ptr<Stack> clone() const {
  28.     return c10::make_intrusive<Stack>(stack_);
  29.   }
  31.   void merge(const c10::intrusive_ptr<Stack>& c) {
  32.     for (auto& elem : c->stack_) {
  33.       push(elem);
  34.     }
  35.   }
  36. };

有几件事要注意：

  • torch/custom_class.h是您需要使用自定义类扩展 TorchScript 的标头。
  • 注意，无论何时使用自定义类的实例，我们都通过c10::intrusive_ptr&lt;&gt;的实例来实现。 将intrusive_ptr视为类似于std::shared_ptr的智能指针。 使用此智能指针的原因是为了确保在语言(C ++，Python 和 TorchScript）之间对对象实例进行一致的生命周期管理。
  • 注意的第二件事是用户定义的类必须继承自torch::jit::CustomClassHolder。 这确保了所有设置都可以处理前面提到的生命周期管理系统。

现在让我们看一下如何使该类对 TorchScript 可见，该过程称为绑定该类：

  1. // Notice a few things:
  2. // - We pass the class to be registered as a template parameter to
  3. //   `torch::jit::class_`. In this instance, we've passed the
  4. //   specialization of the Stack class ``Stack<std::string>``.
  5. //   In general, you cannot register a non-specialized template
  6. //   class. For non-templated classes, you can just pass the
  7. //   class name directly as the template parameter.
  8. // - The single parameter to ``torch::jit::class_()`` is a
  9. //   string indicating the name of the class. This is the name
  10. //   the class will appear as in both Python and TorchScript.
  11. //   For example, our Stack class would appear as ``torch.classes.Stack``.
  12. static auto testStack =
  13.   torch::jit::class_<Stack<std::string>>("Stack")
  14.       // The following line registers the contructor of our Stack
  15.       // class that takes a single `std::vector<std::string>` argument,
  16.       // i.e. it exposes the C++ method `Stack(std::vector<T> init)`.
  17.       // Currently, we do not support registering overloaded
  18.       // constructors, so for now you can only `def()` one instance of
  19.       // `torch::jit::init`.
  20.       .def(torch::jit::init<std::vector<std::string>>())
  21.       // The next line registers a stateless (i.e. no captures) C++ lambda
  22.       // function as a method. Note that a lambda function must take a
  23.       // `c10::intrusive_ptr<YourClass>` (or some const/ref version of that)
  24.       // as the first argument. Other arguments can be whatever you want.
  25.       .def("top", [](const c10::intrusive_ptr<Stack<std::string>>& self) {
  26.         return self->stack_.back();
  27.       })
  28.       // The following four lines expose methods of the Stack<std::string>
  29.       // class as-is. `torch::jit::class_` will automatically examine the
  30.       // argument and return types of the passed-in method pointers and
  31.       // expose these to Python and TorchScript accordingly. Finally, notice
  32.       // that we must take the *address* of the fully-qualified method name,
  33.       // i.e. use the unary `&` operator, due to C++ typing rules.
  34.       .def("push", &Stack<std::string>::push)
  35.       .def("pop", &Stack<std::string>::pop)
  36.       .def("clone", &Stack<std::string>::clone)
  37.       .def("merge", &Stack<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，将到目前为止介绍的所有 C ++代码放入class.cpp文件中。 然后，编写一个简单的CMakeLists.txt文件并将其放置在同一目录中。 CMakeLists.txt的外观如下：

  1. cmake_minimum_required(VERSION 3.1 FATAL_ERROR)
  2. project(custom_class)
  4. find_package(Torch REQUIRED)
  6. # Define our library target
  7. add_library(custom_class SHARED class.cpp)
  8. set(CMAKE_CXX_STANDARD 14)
  9. # Link against LibTorch
  10. target_link_libraries(custom_class "${TORCH_LIBRARIES}")

另外，创建一个build目录。 您的文件树应如下所示：

  1. custom_class_project/
  2.   class.cpp
  3.   CMakeLists.txt
  4.   build/

现在，要构建项目，请继续从 PyTorch 网站下载适当的 libtorch 二进制文件。 将 zip 存档解压缩到某个位置(在项目目录中可能很方便），并记下将其解压缩到的路径。 接下来，继续调用 cmake，然后进行构建项目：

  1. $ cd build
  2. $ cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
  3.   -- The C compiler identification is GNU 7.3.1
  4.   -- The CXX compiler identification is GNU 7.3.1
  5.   -- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc
  6.   -- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc -- works
  7.   -- Detecting C compiler ABI info
  8.   -- Detecting C compiler ABI info - done
  9.   -- Detecting C compile features
  10.   -- Detecting C compile features - done
  11.   -- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++
  12.   -- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++ -- works
  13.   -- Detecting CXX compiler ABI info
  14.   -- Detecting CXX compiler ABI info - done
  15.   -- Detecting CXX compile features
  16.   -- Detecting CXX compile features - done
  17.   -- Looking for pthread.h
  18.   -- Looking for pthread.h - found
  19.   -- Looking for pthread_create
  20.   -- Looking for pthread_create - not found
  21.   -- Looking for pthread_create in pthreads
  22.   -- Looking for pthread_create in pthreads - not found
  23.   -- Looking for pthread_create in pthread
  24.   -- Looking for pthread_create in pthread - found
  25.   -- Found Threads: TRUE
  26.   -- Found torch: /torchbind_tutorial/libtorch/lib/libtorch.so
  27.   -- Configuring done
  28.   -- Generating done
  29.   -- Build files have been written to: /torchbind_tutorial/build
  30. $ make -j
  31.   Scanning dependencies of target custom_class
  32.   [ 50%] Building CXX object CMakeFiles/custom_class.dir/class.cpp.o
  33.   [100%] Linking CXX shared library libcustom_class.so
  34.   [100%] Built target custom_class

您会发现，构建目录中现在有一个动态库文件。 在 Linux 上，它可能名为libcustom_class.so。 因此，文件树应如下所示：

  1. custom_class_project/
  2.   class.cpp
  3.   CMakeLists.txt
  4.   build/
  5.     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为.so文件，我们可以将 <cite>.so</cite> 加载到 Python 中并进行尝试。 这是一个演示脚本的脚本：

  1. import torch
  3. # `torch.classes.load_library()` allows you to pass the path to your .so file
  4. # to load it in and make the custom C++ classes available to both Python and
  5. # TorchScript
  6. torch.classes.load_library("libcustom_class.so")
  7. # You can query the loaded libraries like this:
  8. print(torch.classes.loaded_libraries)
  9. # prints {'/custom_class_project/build/libcustom_class.so'}
  11. # We can find and instantiate our custom C++ class in python by using the
  12. # `torch.classes` namespace:
  13. #
  14. # This instantiation will invoke the Stack(std::vector<T> init) constructor
  15. # we registered earlier
  16. s = torch.classes.Stack(["foo", "bar"])
  18. # We can call methods in Python
  19. s.push("pushed")
  20. assert s.pop() == "pushed"
  22. # Returning and passing instances of custom classes works as you'd expect
  23. s2 = s.clone()
  24. s.merge(s2)
  25. for expected in ["bar", "foo", "bar", "foo"]:
  26.     assert s.pop() == expected
  28. # We can also use the class in TorchScript
  29. # For now, we need to assign the class's type to a local in order to
  30. # annotate the type on the TorchScript function. This may change
  31. # in the future.
  32. Stack = torch.classes.Stack
  34. @torch.jit.script
  35. def do_stacks(s : Stack): # We can pass a custom class instance to TorchScript
  36.     s2 = torch.classes.Stack(["hi", "mom"]) # We can instantiate the class
  37.     s2.merge(s) # We can call a method on the class
  38.     return s2.clone(), s2.top()  # We can also return instances of the class
  39.                                  # from TorchScript function/methods
  41. stack, top = do_stacks(torch.classes.Stack(["wow"]))
  42. assert top == "wow"
  43. for expected in ["wow", "mom", "hi"]:
  44.     assert stack.pop() == expected

使用自定义类保存，加载和运行 TorchScript 代码

我们也可以在使用 libtorch 的 C ++进程中使用自定义注册的 C ++类。 举例来说，让我们定义一个简单的nn.Module，该实例在我们的 Stack 类上实例化并调用一个方法：

  1. import torch
  3. torch.classes.load_library('libcustom_class.so')
  5. class Foo(torch.nn.Module):
  6.     def __init__(self):
  7.         super().__init__()
  9.     def forward(self, s : str) -> str:
  10.         stack = torch.classes.Stack(["hi", "mom"])
  11.         return stack.pop() + s
  13. scripted_foo = torch.jit.script(Foo())
  14. print(scripted_foo.graph)
  16. scripted_foo.save('foo.pt')

我们文件系统中的foo.pt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 ++教程中加载 TorchScript 模型。

与之前类似，让我们创建一个包含以下内容的文件结构：

  1. cpp_inference_example/
  2.   infer.cpp
  3.   CMakeLists.txt
  4.   foo.pt
  5.   build/
  6.   custom_class_project/
  7.     class.cpp
  8.     CMakeLists.txt
  9.     build/

请注意，我们已经复制了序列化的foo.pt文件以及上面custom_class_project的源代码树。 我们将添加custom_class_project作为对此 C ++项目的依赖项，以便我们可以将自定义类构建到二进制文件中。

让我们用以下内容填充infer.cpp：

  1. #include <torch/script.h>
  3. #include <iostream>
  4. #include <memory>
  6. int main(int argc, const char* argv[]) {
  7.   torch::jit::script::Module module;
  8.   try {
  9.     // Deserialize the ScriptModule from a file using torch::jit::load().
  10.     module = torch::jit::load("foo.pt");
  11.   }
  12.   catch (const c10::Error& e) {
  13.     std::cerr << "error loading the model\n";
  14.     return -1;
  15.   }
  17.   std::vector<c10::IValue> inputs = {"foobarbaz"};
  18.   auto output = module.forward(inputs).toString();
  19.   std::cout << output->string() << std::endl;
  20. }

同样，让我们​​定义我们的 CMakeLists.txt 文件：

  1. cmake_minimum_required(VERSION 3.1 FATAL_ERROR)
  2. project(infer)
  4. find_package(Torch REQUIRED)
  6. add_subdirectory(custom_class_project)
  8. # Define our library target
  9. add_executable(infer infer.cpp)
  10. set(CMAKE_CXX_STANDARD 14)
  11. # Link against LibTorch
  12. target_link_libraries(infer "${TORCH_LIBRARIES}")
  13. # This is where we link in our libcustom_class code, making our
  14. # custom class available in our binary.
  15. target_link_libraries(infer -Wl,--no-as-needed custom_class)

您知道练习：cd build，cmake和make：

  1. $ cd build
  2. $ cmake -DCMAKE_PREFIX_PATH=/path/to/libtorch ..
  3.   -- The C compiler identification is GNU 7.3.1
  4.   -- The CXX compiler identification is GNU 7.3.1
  5.   -- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc
  6.   -- Check for working C compiler: /opt/rh/devtoolset-7/root/usr/bin/cc -- works
  7.   -- Detecting C compiler ABI info
  8.   -- Detecting C compiler ABI info - done
  9.   -- Detecting C compile features
  10.   -- Detecting C compile features - done
  11.   -- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++
  12.   -- Check for working CXX compiler: /opt/rh/devtoolset-7/root/usr/bin/c++ -- works
  13.   -- Detecting CXX compiler ABI info
  14.   -- Detecting CXX compiler ABI info - done
  15.   -- Detecting CXX compile features
  16.   -- Detecting CXX compile features - done
  17.   -- Looking for pthread.h
  18.   -- Looking for pthread.h - found
  19.   -- Looking for pthread_create
  20.   -- Looking for pthread_create - not found
  21.   -- Looking for pthread_create in pthreads
  22.   -- Looking for pthread_create in pthreads - not found
  23.   -- Looking for pthread_create in pthread
  24.   -- Looking for pthread_create in pthread - found
  25.   -- Found Threads: TRUE
  26.   -- Found torch: /local/miniconda3/lib/python3.7/site-packages/torch/lib/libtorch.so
  27.   -- Configuring done
  28.   -- Generating done
  29.   -- Build files have been written to: /cpp_inference_example/build
  30. $ make -j
  31.   Scanning dependencies of target custom_class
  32.   [ 25%] Building CXX object custom_class_project/CMakeFiles/custom_class.dir/class.cpp.o
  33.   [ 50%] Linking CXX shared library libcustom_class.so
  34.   [ 50%] Built target custom_class
  35.   Scanning dependencies of target infer
  36.   [ 75%] Building CXX object CMakeFiles/infer.dir/infer.cpp.o
  37.   [100%] Linking CXX executable infer
  38.   [100%] Built target infer

现在我们可以运行令人兴奋的 C ++二进制文件：

  1. $ ./infer
  2.   momfoobarbaz

难以置信！

定义自定义 C ++类的序列化/反序列化方法

如果您尝试将具有自定义绑定 C ++类的ScriptModule保存为属性，则会出现以下错误：

  1. # export_attr.py
  2. import torch
  4. torch.classes.load_library('libcustom_class.so')
  6. class Foo(torch.nn.Module):
  7.   def __init__(self):
  8.       super().__init__()
  9.       self.stack = torch.classes.Stack(["just", "testing"])
  11.   def forward(self, s : str) -> str:
  12.       return self.stack.pop() + s
  14. scripted_foo = torch.jit.script(Foo())
  16. scripted_foo.save('foo.pt')
  1. $ python export_attr.py
  2. RuntimeError: Cannot serialize custom bound C++ class __torch__.torch.classes.Stack. Please define serialization methods via torch::jit::pickle_ for this class. (pushIValueImpl at ../torch/csrc/jit/pickler.cpp:128)

这是因为 TorchScript 无法自动找出 C ++类中保存的信息。 您必须手动指定。 这样做的方法是使用class_上的特殊def_pickle方法在类上定义__getstate__和__setstate__方法。

注意

TorchScript 中__getstate__和__setstate__的语义与 Python pickle 模块的语义相同。 您可以阅读更多有关如何使用这些方法的信息。

这是一个如何更新Stack类的注册码以包含序列化方法的示例：

  1. static auto testStack =
  2.   torch::jit::class_<Stack<std::string>>("Stack")
  3.       .def(torch::jit::init<std::vector<std::string>>())
  4.       .def("top", [](const c10::intrusive_ptr<Stack<std::string>>& self) {
  5.         return self->stack_.back();
  6.       })
  7.       .def("push", &Stack<std::string>::push)
  8.       .def("pop", &Stack<std::string>::pop)
  9.       .def("clone", &Stack<std::string>::clone)
  10.       .def("merge", &Stack<std::string>::merge)
  11.       // class_<>::def_pickle allows you to define the serialization
  12.       // and deserialization methods for your C++ class.
  13.       // Currently, we only support passing stateless lambda functions
  14.       // as arguments to def_pickle
  15.       .def_pickle(
  16.             // __getstate__
  17.             // This function defines what data structure should be produced
  18.             // when we serialize an instance of this class. The function
  19.             // must take a single `self` argument, which is an intrusive_ptr
  20.             // to the instance of the object. The function can return
  21.             // any type that is supported as a return value of the TorchScript
  22.             // custom operator API. In this instance, we've chosen to return
  23.             // a std::vector<std::string> as the salient data to preserve
  24.             // from the class.
  25.             [](const c10::intrusive_ptr<Stack<std::string>>& self)
  26.                 -> std::vector<std::string> {
  27.               return self->stack_;
  28.             },
  29.             // __setstate__
  30.             // This function defines how to create a new instance of the C++
  31.             // class when we are deserializing. The function must take a
  32.             // single argument of the same type as the return value of
  33.             // `__getstate__`. The function must return an intrusive_ptr
  34.             // to a new instance of the C++ class, initialized however
  35.             // you would like given the serialized state.
  36.             [](std::vector<std::string> state)
  37.                 -> c10::intrusive_ptr<Stack<std::string>> {
  38.               // A convenient way to instantiate an object and get an
  39.               // intrusive_ptr to it is via `make_intrusive`. We use
  40.               // that here to allocate an instance of Stack<std::string>
  41.               // and call the single-argument std::vector<std::string>
  42.               // constructor with the serialized state.
  43.               return c10::make_intrusive<Stack<std::string>>(std::move(state));
  44.             });

Note

我们采用与 pickle API 中的 pybind11 不同的方法。 pybind11 作为传递给class_::def()的特殊功能pybind11::pickle()，为此我们有一个单独的方法def_pickle。 这是因为名称torch::jit::pickle已经被使用，我们不想引起混淆。

以这种方式定义(反）序列化行为后，脚本现在可以成功运行：

  1. import torch
  3. torch.classes.load_library('libcustom_class.so')
  5. class Foo(torch.nn.Module):
  6.     def __init__(self):
  7.         super().__init__()
  8.         self.stack = torch.classes.Stack(["just", "testing"])
  10.     def forward(self, s : str) -> str:
  11.         return self.stack.pop() + s
  13. scripted_foo = torch.jit.script(Foo())
  15. scripted_foo.save('foo.pt')
  16. loaded = torch.jit.load('foo.pt')
  18. print(loaded.stack.pop())
  1. $ python ../export_attr.py
  2. testing

结论

本教程向您介绍了如何向 TorchScript(以及扩展为 Python）公开 C ++类，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以及运行该代码。 在独立的 C ++过程中。 现在，您可以使用与第三方 C ++库接口的 C ++类扩展 TorchScript 模型，或实现需要 Python，TorchScript 和 C ++之间的界线才能平滑融合的任何其他用例。

与往常一样，如果您遇到任何问题或疑问，可以使用我们的论坛或 GitHub 问题进行联系。 另外，我们的常见问题解答(FAQ）页面可能包含有用的信息。