pytorch/torch/custom_class.h

#pragma once

#include <ATen/core/function_schema.h>
#include <ATen/core/ivalue.h>
#include <ATen/core/jit_type.h>
#include <ATen/core/op_registration/op_registration.h>
#include <ATen/core/stack.h>
#include <c10/util/C++17.h>
#include <c10/util/Metaprogramming.h>
#include <c10/util/TypeList.h>
#include <c10/util/TypeTraits.h>
#include <torch/csrc/jit/api/custom_class.h>
#include <torch/csrc/jit/runtime/operator.h>
#include <torch/csrc/jit/api/compilation_unit.h>
#include <torch/csrc/jit/frontend/tracer.h>
#include <torch/csrc/utils/variadic.h>
#include <torch/custom_class_detail.h>
#include <iostream>
#include <sstream>


namespace torch {
namespace jit {

template <class... Types>
detail::types<void, Types...> init() {
  return detail::types<void, Types...>{};
}

// To bind custom classes into Torchscript, use an API very similar to Pybind's.
// Currently exposes one class `torch::jit::class_<T>` and 2 methods.
// - Constructing `torch::jit::class_<Foo>` registers `Foo` in Python and
// Torchscript, and puts it under `torch.classes.Foo` in Python.
// - torch::jit::class_<Foo>.def("method1", &Foo::method1) does some template
// metaprogramming to introspect the function types and register the operator
// for use in Torchscript.
// - torch::jit::class_<Foo>.def(torch::jit::init<int64_t, int64_t>()) registers
// the Foo(int, int) constructor.
// see test/custom_operator/classes.cpp and
// test/custom_operator/test_custom_classes.py for example usages

template <class CurClass>
class class_ {
  static_assert(std::is_base_of<CustomClassHolder, CurClass>::value,
    "torch::jit::class_<T> requires T to inherit from CustomClassHolder");

  std::string className;
  std::string qualClassName;
  ClassTypePtr classTypePtr;

  const std::string parentModule = "classes";
  const std::string topModule = "__torch__.torch";

 public:
  class_(std::string className_) : className(std::move(className_)) {
    qualClassName = topModule + "." + parentModule + "." + className;

    // We currently represent custom classes as torchscript classes with a
    // capsule attribute
    classTypePtr =
        ClassType::create(c10::QualifiedName(qualClassName), classCU());
    classTypePtr->addAttribute("capsule", CapsuleType::get());

    c10::getCustomClassTypeMap().insert({typeid(c10::intrusive_ptr<CurClass>).name(),
                              c10::StrongTypePtr(classCU(), classTypePtr)});
    c10::getCustomClassTypeMap().insert({typeid(c10::tagged_capsule<CurClass>).name(),
                              c10::StrongTypePtr(classCU(), classTypePtr)});

    classCU()->register_type(classTypePtr);
  }

  template <typename... Types>
  class_& def(detail::types<void, Types...>) { // Used in combination with
                                               // torch::jit::init<...>()
    auto func = [](c10::tagged_capsule<CurClass> self, Types... args) {
      auto classObj = c10::make_intrusive<CurClass>(args...);
      auto genericPtr = c10::static_intrusive_pointer_cast<torch::jit::CustomClassHolder>(std::move(classObj));
      auto capsule = IValue(std::move(genericPtr));
      auto object = std::move(self.ivalue).toObject();
      object->setSlot(0, std::move(capsule));
    };

    defineMethod("__init__", std::move(func));
    return *this;
  }
  template <typename Func>
  class_& def(std::string name, Func f) {
    auto wrapped_f = detail::wrap_func<CurClass, Func>(std::move(f));
    defineMethod(std::move(name), std::move(wrapped_f));
    return *this;
  }

  // Pickle
  template <typename GetStateFn, typename SetStateFn>
  class_& def_pickle(GetStateFn&& get_state, SetStateFn&& set_state) {
    static_assert(
        c10::guts::is_stateless_lambda<std::decay_t<GetStateFn>>::value &&
            c10::guts::is_stateless_lambda<std::decay_t<SetStateFn>>::value,
        "torch::jit::pickle_ currently only supports lambdas as "
        "__getstate__ and __setstate__ arguments.");
    def("__getstate__", std::forward<GetStateFn>(get_state));

    // __setstate__ needs to be registered with some custom handling:
    // We need to wrap the invocation of of the user-provided function
    // such that we take the return value (i.e. c10::intrusive_ptr<CurrClass>)
    // and assign it to the `capsule` attribute.
    using SetStateTraits =
        c10::guts::infer_function_traits_t<std::decay_t<SetStateFn>>;
    using SetStateArg = typename c10::guts::typelist::head_t<
        typename SetStateTraits::parameter_types>;
    auto setstate_wrapper = [set_state = std::move(set_state)](
                                c10::tagged_capsule<CurClass> self,
                                SetStateArg&& arg) {
      c10::intrusive_ptr<CurClass> classObj =
          at::guts::invoke(set_state, std::forward<SetStateArg>(arg));
      auto genericPtr =
          c10::static_intrusive_pointer_cast<torch::jit::CustomClassHolder>(
              classObj);
      auto capsule = IValue(genericPtr);
      auto object = self.ivalue.toObject();
      object->setSlot(0, capsule);
    };
    defineMethod(
        "__setstate__",
        detail::wrap_func<CurClass, decltype(setstate_wrapper)>(
            std::move(setstate_wrapper)));

    // type validation
    auto getstate_schema = classTypePtr->getMethod("__getstate__")->getSchema();
    auto format_getstate_schema = [&getstate_schema]() {
      std::stringstream ss;
      ss << getstate_schema;
      return ss.str();
    };
    TORCH_CHECK(
        getstate_schema.arguments().size() == 1,
        "__getstate__ should take exactly one argument: self. Got: ",
        format_getstate_schema());
    auto first_arg_type = getstate_schema.arguments().at(0).type();
    TORCH_CHECK(
        *first_arg_type == *classTypePtr,
        "self argument of __getstate__ must be the custom class type. Got ",
        first_arg_type->python_str());
    TORCH_CHECK(
        getstate_schema.returns().size() == 1,
        "__getstate__ should return exactly one value for serialization. Got: ",
        format_getstate_schema());
    auto ser_type = getstate_schema.returns().at(0).type();
    auto setstate_schema = classTypePtr->getMethod("__setstate__")->getSchema();
    auto arg_type = setstate_schema.arguments().at(1).type();
    TORCH_CHECK(
        (*arg_type == *ser_type),
        "__setstate__'s argument should be the same type as the "
        "return value of __getstate__. Got ",
        arg_type->python_str(),
        " but expected ",
        ser_type->python_str());

    return *this;
  }

 private:
  template <typename Func>
  void defineMethod(std::string name, Func func) {
    auto graph = std::make_shared<Graph>();
    auto qualFuncName = className + "::" + name;
    ensure_c10_registerer_defined();
    registeredOps().push_back(
        torch::RegisterOperators().op(qualFuncName, std::move(func)));
    auto func_symbol = c10::Symbol::fromQualString(qualFuncName);
    auto ops = torch::jit::getAllOperatorsFor(func_symbol);
    TORCH_CHECK(ops.size() == 1);
    auto &schema = ops[0]->schema();

    for (const auto& arg : schema.arguments()) {
      graph->addInput()->setType(arg.type());
    }

    auto opCall = graph->insertNode(graph->create(
        func_symbol, graph->inputs(), schema.returns().size()));
    Value* res;
    if (schema.returns().size() > 1) {
      const auto& returns = schema.returns();
      size_t op_invocation_idx = 0;
      for (const auto& ret : returns) {
        opCall->output(op_invocation_idx++)->setType(ret.type());
      }
      res = graph->insertNode(graph->createTuple(opCall->outputs()))->output();
    } else if (schema.returns().size() == 1) {
      const auto& returns = schema.returns();
      res = opCall->output()->setType(returns[0].type());
    } else {
      res = graph->insertConstant(IValue())->setType(NoneType::get());
    }
    graph->registerOutput(res);

    auto method = classCU()->create_function(qualClassName + "." + name, graph);
    classTypePtr->addMethod(method);
  }
};

} // namespace jit

} // namespace torch