pytorch/test/cpp/jit/test_module_api.cpp

#include <gtest/gtest.h>

#include <test/cpp/jit/test_utils.h>

#include <ATen/core/qualified_name.h>
#include <torch/csrc/jit/api/module.h>
#include <torch/csrc/jit/frontend/resolver.h>
#include <torch/csrc/jit/serialization/import.h>
#include <torch/csrc/jit/serialization/import_source.h>
#include <torch/csrc/jit/testing/file_check.h>
#include <torch/torch.h>

namespace torch {
namespace jit {

static constexpr std::string_view moduleInterfaceSrc = R"JIT(
class OneInterface(ModuleInterface):
    def one(self, x: Tensor, y: Tensor) -> Tensor:
        pass
)JIT";

static const std::vector<std::string> subModuleMethodsSrc = {R"JIT(
def one(self, x: Tensor, y: Tensor) -> Tensor:
    return self.attr * x + y + 1

def forward(self, x: Tensor) -> Tensor:
    return self.attr + x
)JIT"};

static const std::string parentForward = R"JIT(
def forward(self, x: Tensor) -> Tensor:
    return self.subMod1.one(x, x) + self.subMod2.one(x, x)
)JIT";

static void import_libs(
    std::shared_ptr<CompilationUnit> cu,
    const std::string& class_name,
    const std::shared_ptr<Source>& src,
    const std::vector<at::IValue>& tensor_table) {
  SourceImporter si(
      cu,
      &tensor_table,
      [&](const std::string& name) -> std::shared_ptr<Source> { return src; },
      /*version=*/2);
  si.loadType(QualifiedName(class_name));
}

TEST(ModuleAPITest, MethodRunAsync) {
  // Module m("m");
  // m.define(R"(
  //   def forward(self):
  //     r1 = torch.jit.fork(torch.mm, torch.rand(100,100),torch.rand(100,100))
  //     r2 = torch.jit.fork(torch.mm, torch.rand(100,100),torch.rand(100,100))
  //     return r1.wait() + r2.wait()
  // )");
  std::string filePath(__FILE__);
  auto testModelFile = filePath.substr(0, filePath.find_last_of("/\\") + 1);
  // borrow model file from TEST(GraphExecutorTest, runAsync_executor)
  testModelFile.append("test_interpreter_async.pt");
  auto m = load(testModelFile);

  auto counter = 0;
  std::mutex mtx;

  auto launcher = [&](std::function<void()> f) {
    mtx.lock();
    ++counter;
    mtx.unlock();
    at::launch(std::move(f));
  };

  auto method = m.get_method("forward");

  std::vector<IValue> stack;
  auto kwargs = std::unordered_map<std::string, at::IValue>();
  auto future = method.run_async(stack, kwargs, launcher);

  future->wait();

  // expect 2 forks and 2 wait callbacks being executed on provided taskLauncher
  // but ivalue::Future would be marked completed and release wait before
  // finishing all callbacks
  ASSERT_GE(counter, 2);
}

TEST(ModuleAPITest, Clone) {
  auto cu = std::make_shared<CompilationUnit>();
  // creating child module
  auto child = ClassType::create("child", cu, true);
  auto attr_name = "attr";
  child->addAttribute(attr_name, IntType::get());
  Module c1(cu, child);
  auto v1 = IValue(2);
  c1.register_attribute(attr_name, IntType::get(), v1, false);
  Module c2(cu, child);
  auto v2 = IValue(3);
  c2.register_attribute(attr_name, IntType::get(), v2, false);

  // attach two child module instance to parent that shares
  // ClassType
  auto parent = ClassType::create("parent", cu, true);
  Module p(cu, parent);
  p.register_attribute("c1", c1.type(), c1._ivalue(), false);
  p.register_attribute("c2", c2.type(), c2._ivalue(), false);

  // clone parent
  Module p2 = p.clone();
  // check the two child module has the same ClassType
  ASSERT_EQ(p2.attr("c1").type(), p2.attr("c2").type());
  // but different instances
  ASSERT_EQ(Module(p2.attr("c1").toObject()).attr(attr_name).toInt(), 2);
  ASSERT_EQ(Module(p2.attr("c2").toObject()).attr(attr_name).toInt(), 3);
}

TEST(ModuleAPITest, CloneWithModuleInterface) {
  auto cu = std::make_shared<CompilationUnit>();

  // define a initial module with two submods share same interface
  Module parentMod("parentMod", cu);
  Module subMod1("subMod1", cu);
  Module subMod2("subMod2", cu);

  std::vector<at::IValue> constantTable;
  import_libs(
      cu,
      "__torch__.OneInterface",
      std::make_shared<Source>(moduleInterfaceSrc),
      constantTable);

  auto v1 = IValue(2);
  subMod1.register_attribute("attr", IntType::get(), v1, false);

  auto v2 = IValue(4);
  subMod2.register_attribute("attr", IntType::get(), v2, false);

  for (const std::string& method : subModuleMethodsSrc) {
    subMod1.define(method, nativeResolver());
    subMod2.define(method, nativeResolver());
  }

  parentMod.register_attribute(
      "subMod1",
      cu->get_interface("__torch__.OneInterface"),
      subMod1._ivalue());
  parentMod.register_attribute(
      "subMod2",
      cu->get_interface("__torch__.OneInterface"),
      subMod2._ivalue());

  parentMod.define(parentForward, nativeResolver());

  Module clonedMod = parentMod.clone();

  // clone will copy both type and data, therefore we'll have a
  // different type
  ASSERT_NE(clonedMod.type(), parentMod.type());
}

TEST(ModuleAPITest, Copy) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto attr_name = "attr";
  cls->addAttribute(attr_name, IntType::get());
  Module m(cu, cls);
  auto v = IValue(2);
  m.register_attribute(attr_name, IntType::get(), v, false);

  Module m2 = m.clone();
  Module m3 = m.copy();

  // Make sure copy works
  ASSERT_EQ(m2.attr(attr_name).toInt(), 2);
  ASSERT_EQ(m3.attr(attr_name).toInt(), 2);

  // clone will copy both type and data, therefore we'll have a
  // different type
  ASSERT_NE(m.type(), m2.type());
  // copy only copies data, type is shared
  ASSERT_EQ(m.type(), m3.type());

  // change value of copied instance
  m3.register_attribute(attr_name, IntType::get(), IValue(3), false);
  // Verify value of original instance doesn't change
  ASSERT_EQ(m2.attr(attr_name).toInt(), 2);
  ASSERT_EQ(m3.attr(attr_name).toInt(), 3);
}

TEST(ModuleAPITest, DeepCopy) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto str_attr = "str_attr";
  auto int_attr = "int_attr";
  auto tensor_attr = "tensor_attr";
  auto tensor_list_attr = "tensor_list_attr";
  cls->addAttribute(int_attr, IntType::get());
  cls->addAttribute(str_attr, StringType::get());
  cls->addAttribute(tensor_attr, TensorType::get());
  cls->addAttribute(tensor_list_attr, ListType::ofTensors());
  Module m(cu, cls);
  c10::List<at::Tensor> list({at::rand(5), at::rand(5)});
  m.setattr(int_attr, IValue(2));
  m.setattr(str_attr, IValue("str"));
  m.setattr(tensor_attr, at::randn(5));
  m.setattr(tensor_list_attr, list);

  Module m2 = m.deepcopy();
  Module m3 = m.copy();
  // Make sure copy works
  ASSERT_EQ(m2.attr(int_attr).toInt(), 2);
  ASSERT_EQ(m3.attr(int_attr).toInt(), 2);

  // Test overlaps
  ASSERT_TRUE(!IValue(m2._ivalue()).overlaps(IValue(m._ivalue())));
  ASSERT_TRUE(IValue(m3._ivalue()).overlaps(IValue(m._ivalue())));

  // Both deepcopy and copy will preserve the type
  ASSERT_EQ(m.type(), m2.type());
  ASSERT_EQ(m.type(), m3.type());

  // change int value of copied instances
  m2.setattr(int_attr, IValue(3));
  m3.setattr(int_attr, IValue(4));

  // Verify value of original instance doesn't change
  ASSERT_EQ(m.attr(int_attr).toInt(), 2);
  ASSERT_EQ(m2.attr(int_attr).toInt(), 3);
  ASSERT_EQ(m3.attr(int_attr).toInt(), 4);

  // change Tensor value of copied instances
  at::Tensor t1 = m.attr(tensor_attr).toTensor();
  at::Tensor t2 =
      m2.attr(tensor_attr).toTensor(); // deepcopy will copy the Tensor
  at::Tensor t3 =
      m3.attr(tensor_attr).toTensor(); // copy will not copy the Tensor
  // check copy works
  ASSERT_TRUE(t1.equal(t2));
  ASSERT_TRUE(t1.equal(t3));

  // zero out t1
  t1.zero_();
  // check that t2 is not affected because it is a deep copy
  ASSERT_TRUE(!t1.equal(t2));
  // check that t3 is the same as t1 since it is a shallow copy
  ASSERT_TRUE(t1.equal(t3));
}

TEST(ModuleAPITest, DeepCopyString) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto attr1 = "attr1";
  cls->addAttribute(attr1, StringType::get());
  std::string str = "str";
  Module m(cu, cls);
  m.setattr(attr1, str);
  auto copied = m.deepcopy();
  auto original_str = str;
  ASSERT_EQ(copied.attr(attr1).toStringRef(), original_str);
  // check string mutation is not reflected in the copied module
  str += "str";
  ASSERT_EQ(copied.attr(attr1).toStringRef(), original_str);
}

TEST(ModuleAPITest, DeepCopyEnum) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto enum_attr = "enum_attr";
  auto int_enum_type = EnumType::create(
      "enum_class",
      IntType::get(),
      {{"enum_name_1", 1}, {"enum_name_2", 2}},
      cu);
  cls->addAttribute(enum_attr, int_enum_type);
  Module m(cu, cls);
  m.setattr(
      enum_attr,
      IValue(c10::make_intrusive<ivalue::EnumHolder>(
          int_enum_type, "enum_name_1", 1)));
  Module m2 = m.deepcopy();

  // Make sure deepcopy works
  c10::ivalue::EnumHolder* m2_holder = m2.attr(enum_attr).toEnumHolder().get();
  ASSERT_EQ(m2_holder->value().toInt(), 1);
  ASSERT_EQ(m2_holder->name(), "enum_name_1");
  ASSERT_EQ(m2_holder->type(), int_enum_type);

  // Test overlaps
  ASSERT_TRUE(!IValue(m2._ivalue()).overlaps(IValue(m._ivalue())));

  // Deepcopy will preserve the type
  ASSERT_EQ(m.type(), m2.type());

  // Change original, should not affect deepcopy
  m.setattr(
      enum_attr,
      IValue(c10::make_intrusive<ivalue::EnumHolder>(
          int_enum_type, "enum_name_2", 2)));
  ASSERT_NE(
      m.attr(enum_attr).toEnumHolder().get()->value().toInt(),
      m2.attr(enum_attr).toEnumHolder().get()->value().toInt());
}

TEST(ModuleAPITest, DeepCopyPreservesAliasing) {
  // check deepcopy preserves aliasing
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto attr1 = "attr1";
  auto attr2 = "attr2";
  auto attr3 = "attr3";
  auto attr4 = "attr4";
  cls->addAttribute(attr1, ListType::ofTensors());
  cls->addAttribute(attr2, ListType::ofTensors());
  cls->addAttribute(attr3, TensorType::get());
  cls->addAttribute(attr4, TensorType::get());
  Module m(cu, cls);
  auto t1 = at::rand(5);
  auto t2 = at::rand(5);
  auto t3 = at::rand(5);
  auto t4 = at::rand({5, 2});
  c10::List<at::Tensor> list1({t1, t2});
  c10::List<at::Tensor> list2({t1, t3});
  // first element of attr1 and attr2 are aliased
  m.setattr(attr1, list1);
  m.setattr(attr2, list2);
  m.setattr(attr3, t4);
  m.setattr(attr4, t4.view(-1));

  auto copied = m.deepcopy();
  // test tensor aliasing
  auto copied_attr1_t1 = copied.attr(attr1).toList().get(0);
  auto copied_attr2_t1 = copied.attr(attr2).toList().get(0);
  ASSERT_TRUE(copied_attr1_t1.isAliasOf(copied_attr2_t1));

  // test aliasing from view
  auto copied_attr3 = copied.attr(attr3);
  auto copied_attr4 = copied.attr(attr3);
  ASSERT_TRUE(copied_attr3.isAliasOf(copied_attr4));
}

TEST(ModuleAPITest, Constants) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  auto attr_name = "attr";
  auto const_name = "const";
  cls->addAttribute(attr_name, IntType::get());
  cls->addConstant(const_name, IValue(3));
  Module m(cu, cls);
  auto v = IValue(2);
  m.register_attribute(attr_name, IntType::get(), v, false);
  ASSERT_TRUE(m.hasattr(attr_name));
  ASSERT_TRUE(m.hasattr(const_name));
  ASSERT_EQ(m.attr(attr_name).toInt(), 2);
  ASSERT_EQ(m.attr(const_name).toInt(), 3);
}

TEST(ModuleAPITest, Parameters) {
  auto cu = std::make_shared<CompilationUnit>();
  auto cls = ClassType::create("foo.bar", cu, true);
  Module m(cu, cls);
  // Tensor parameter
  m.register_parameter(
      "tensor_param", at::empty({3}, at::kFloat), /* is_buffer */ false);
  // None parameter
  m.register_attribute(
      "none_param", NoneType::get(), IValue(), /* is_param */ true);
  m.register_attribute(
      "none_param2", NoneType::get(), IValue(), /* is_param */ true);
  auto param_list = m.parameters();
  ASSERT_EQ(param_list.size(), 1);
  ASSERT_TRUE(m.hasattr("tensor_param"));
  ASSERT_TRUE(m.hasattr("none_param"));
  ASSERT_TRUE(m.hasattr("none_param2"));
}

TEST(ModuleAPITest, Define) {
  Module m("m");
  m.register_parameter("foo", torch::ones({}), false);
  m.define(R"(
    def add_it(self, x, b : int = 4):
      return self.foo + x + b
  )");
  auto result = m.run_method("add_it", torch::ones({}));
  AT_ASSERT(result.toTensor().item<float>() == 6);
}

TEST(ModuleAPITest, Freezing) {
  Module m("m");
  m.register_parameter("foo", torch::ones({}), false);
  m.define(R"(
    def forward(self, x, b : int = 4):
      return self.foo + x + b
  )");
  m.eval();
  auto forward_g = m.get_method("forward").graph();
  testing::FileCheck().check("GetAttr")->run(*forward_g);

  // Removal of GetAttr is done by freezing
  auto frozen_mod = torch::jit::freeze(m);
  forward_g = frozen_mod.get_method("forward").graph();
  testing::FileCheck().check_not("GetAttr")->run(*forward_g);

  // If no training mode is set, the module is NOT frozen by OFI
  auto frozen_mod2 = torch::jit::optimize_for_inference(m);
  forward_g = frozen_mod2.get_method("forward").graph();
  testing::FileCheck().check("GetAttr")->run(*forward_g);
}

TEST(ModuleAPITest, OfiFreezesTraining) {
  Module m("m");
  m.register_parameter("foo", torch::ones({}), false);
  m.define(R"(
    def forward(self, x, b : int = 4):
      return self.foo + x + b
  )");
  m.register_attribute("training", BoolType::get(), true);
  m.eval();

  // Before freezing, we have a GetAttr check
  auto forward_g = m.get_method("forward").graph();
  testing::FileCheck().check("GetAttr")->run(*forward_g);

  // Demonstrate that freezing happens when OFI is called
  // Removal of GetAttr is done by freezing, but only when training
  // attribute is set
  auto frozen_mod = torch::jit::optimize_for_inference(m);
  forward_g = frozen_mod.get_method("forward").graph();
  testing::FileCheck().check_not("GetAttr")->run(*forward_g);
}

TEST(ModuleAPITest, OfiFreezesNoForward) {
  Module m("m");
  m.register_parameter("foo", torch::ones({}), false);
  m.define(R"(
    def bar(self, x, b : int = 4):
      return self.foo + x + b
  )");
  m.eval();

  // OFI is called without the presence of forward methods
  auto frozen_mod =
      torch::jit::optimize_for_inference(m, std::vector<std::string>{"bar"});
  ASSERT_EQ(
      m.run_method("bar", torch::ones({})).toTensor().item<float>(),
      frozen_mod.run_method("bar", torch::ones({})).toTensor().item<float>());
}

TEST(ModuleAPITest, To_CUDA) {
  Module m("test");
  {
    // test cuda to cpu for params and buffers
    m.register_parameter("foo", torch::ones({}, at::kCUDA), false);
    m.register_buffer("bar", torch::ones({}, at::kCUDA));

    m.to(at::kCUDA);
    m.to(at::kCPU);
    AT_ASSERT(m.attr("foo").toTensor().device().is_cpu());
    AT_ASSERT(m.attr("bar").toTensor().device().is_cpu());
  }
  {
    // test cpu to cuda for params and buffers
    m.register_parameter("foo", torch::ones({}), false);
    m.register_buffer("bar", torch::ones({}));

    m.to(at::kCUDA);
    AT_ASSERT(m.attr("foo").toTensor().device().is_cuda());
    AT_ASSERT(m.attr("bar").toTensor().device().is_cuda());
  }
}

} // namespace jit
} // namespace torch