Rename autograd namespace to torch and change torch.h into python.h (#7267)

* Rename autograd namespace to torch and change torch.h into python.h

* Include torch.h instead of python.h in test/cpp/api

* Change some mentions of torch.h to python.h in C++ extensions

* Set paths directly, without find_path

setup.py

@@ -312,7 +312,8 @@ class build_deps(Command):
         self.copy_tree('torch/csrc', 'torch/lib/include/torch/csrc/')
         self.copy_tree('third_party/pybind11/include/pybind11/',
                        'torch/lib/include/pybind11')
-        self.copy_file('torch/torch.h', 'torch/lib/include/torch/torch.h')
+        self.copy_file('torch/csrc/api/include/torch/python.h',
+                       'torch/lib/include/torch/python.h')
 
 
 build_dep_cmds = {}
@@ -600,7 +601,7 @@ main_sources = [
     "torch/csrc/DynamicTypes.cpp",
     "torch/csrc/assertions.cpp",
     "torch/csrc/byte_order.cpp",
-    "torch/csrc/torch.cpp",
+    "torch/csrc/api/src/python.cpp",
     "torch/csrc/utils.cpp",
     "torch/csrc/utils/cuda_lazy_init.cpp",
     "torch/csrc/utils/device.cpp",
@@ -901,6 +902,6 @@ if __name__ == '__main__':
                 'lib/include/torch/csrc/jit/*.h',
                 'lib/include/torch/csrc/utils/*.h',
                 'lib/include/torch/csrc/cuda/*.h',
-                'lib/include/torch/torch.h',
+                'lib/include/torch/python.h',
             ]
         })

test/cpp/api/container.cpp

@@ -1,8 +1,8 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
-using namespace autograd;
+using namespace torch;
 
 AUTOGRAD_CONTAINER_CLASS(TestModel) {
  public:

test/cpp/api/integration.cpp

@@ -1,8 +1,8 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
-using namespace autograd;
+using namespace torch;
 
 #include <iostream>
 

test/cpp/api/main.cpp

@@ -1,7 +1,7 @@
 #define CATCH_CONFIG_RUNNER
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
 #include <iostream>
 
@@ -16,7 +16,7 @@ int main(int argc, char* argv[]) {
     return return_code;
   }
 
-  if (!autograd::hasCuda()) {
+  if (!torch::hasCuda()) {
     std::cerr << "CUDA not available. Disabling CUDA tests" << std::endl;
     // ~ disables the [cuda] tag.
     session.configData().testsOrTags.push_back("~[cuda]");

test/cpp/api/misc.cpp

@@ -1,8 +1,8 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
-using namespace autograd;
+using namespace torch;
 
 TEST_CASE("misc") {
   SECTION("no_grad") {

test/cpp/api/optim.cpp

@@ -1,8 +1,8 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
-using namespace autograd;
+using namespace torch;
 
 bool test_optimizer_xor(Optimizer optim, std::shared_ptr<ContainerList> model) {
   float running_loss = 1;

test/cpp/api/rnn.cpp

@@ -1,8 +1,8 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
-using namespace autograd;
+using namespace torch;
 
 template <typename R, typename Func>
 bool test_RNN_xor(Func&& model_maker, bool cuda = false) {

test/cpp/api/serialization.cpp

@@ -1,10 +1,10 @@
 #include <catch.hpp>
 
-#include <torch/autograd.h>
+#include <torch/torch.h>
 
 #include "cereal/archives/portable_binary.hpp"
 
-using namespace autograd;
+using namespace torch;
 
 TEST_CASE("serialization") {
   SECTION("undefined") {

test/cpp_extensions/cuda_extension.cpp

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/python.h>
 
 // Declare the function from cuda_extension.cu. It will be compiled
 // separately with nvcc and linked with the object file of cuda_extension.cpp

test/cpp_extensions/doubler.h

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/python.h>
 
 struct Doubler {
   Doubler(int A, int B) {

test/cpp_extensions/extension.cpp

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/python.h>
 
 at::Tensor sigmoid_add(at::Tensor x, at::Tensor y) {
   return x.sigmoid() + y.sigmoid();

test/cpp_extensions/jit_extension.cpp

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/python.h>
 
 #include "doubler.h"
 

test/cpp_extensions/jit_extension2.cpp

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/python.h>
 
 using namespace at;
 

test/test_cpp_extensions.py

@@ -151,7 +151,7 @@ class TestCppExtension(common.TestCase):
         '''
 
         cpp_source2 = '''
-        #include <torch/torch.h>
+        #include <torch/python.h>
         at::Tensor sin_add(at::Tensor x, at::Tensor y);
         PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
           m.def("sin_add", &sin_add, "sin(x) + sin(y)");

tools/cpp_build/libtorch/CMakeLists.txt

@@ -10,14 +10,13 @@ if (VERBOSE)
   message(STATUS "ATEN_BUILD_PATH is ${ATEN_BUILD_PATH}")
 endif()
 
-find_path(ATEN_INCLUDE_DIR ATen/ATen.h PATHS "${ATEN_PATH}/src/" NO_DEFAULT_PATH)
-find_path(ATEN_BUILD_INCLUDE_DIR ATen/Type.h PATHS "${ATEN_BUILD_PATH}/src/ATen" NO_DEFAULT_PATH)
+set(ATEN_INCLUDE_DIR "${ATEN_PATH}/src/")
+set(ATEN_BUILD_INCLUDE_DIR "${ATEN_BUILD_PATH}/src/ATen")
+set(TORCH_SRC_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../../torch")
+
 find_library(ATEN_LIBRARY ATen PATHS "${ATEN_BUILD_PATH}/src/ATen" NO_DEFAULT_PATH)
-
 find_library(NANOPB_LIBRARY protobuf-nanopb PATHS "${NANOPB_BUILD_PATH}" NO_DEFAULT_PATH)
 
-find_path(TORCH_SRC_DIR torch.h PATHS "${CMAKE_CURRENT_SOURCE_DIR}/../../../torch" NO_DEFAULT_PATH)
-
 set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
@@ -256,7 +255,7 @@ target_include_directories(torch
   "${ATEN_INCLUDE_DIR}/TH"
   "${ATEN_BUILD_INCLUDE_DIR}"
   "${ATEN_BUILD_PATH}/src/TH"
-  "${TORCH_SRC_DIR}/../"
+  "${TORCH_SRC_DIR}/.."
   "${CMAKE_CURRENT_SOURCE_DIR}")
 
 if (NOT NO_API)

torch/csrc/api/include/torch/containers.h

@@ -5,9 +5,9 @@
 #include "torch/csrc/autograd/variable.h"
 
 #define AUTOGRAD_CONTAINER_CLASS(Type) \
-  class Type : public autograd::Container_CRTP<Type>
+  class Type : public torch::Container_CRTP<Type>
 
-namespace autograd {
+namespace torch {
 class ContainerImpl {
  public:
   virtual ~ContainerImpl() = default;
@@ -438,4 +438,4 @@ class RNN : public RNNBase<RNN> {
   }
 };
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/include/torch/detail.h

@@ -12,7 +12,7 @@
 #include "torch/csrc/utils/auto_gpu.h"
 
 #define AUTOGRAD_OPTIMIZER_CLASS(Type) \
-  class Type : public autograd::Optimizer_CRTP<Type>
+  class Type : public torch::Optimizer_CRTP<Type>
 #define AUTOGRAD_KWARG(CLS, TYP, NAME, DEFAULT, OPTION) \
   TYP NAME##_ = DEFAULT;                                \
   CLS& NAME(TYP x = OPTION) {                           \
@@ -25,7 +25,7 @@ namespace tag = torch::autograd;
 using IntVec = decltype(std::declval<at::IntList>().vec());
 } // namespace
 
-namespace autograd {
+namespace torch {
 namespace detail {
 extern tag::Engine engine;
 }
@@ -67,4 +67,4 @@ int getNumGPUs();
 bool hasCuda();
 bool hasCudnn();
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/include/torch/optimizers.h

@@ -6,7 +6,7 @@
 #include "cereal/access.hpp"
 #include "cereal/cereal.hpp"
 
-namespace autograd {
+namespace torch {
 class OptimizerImpl {
  public:
   OptimizerImpl(Container model) : model_(model) {}
@@ -137,4 +137,4 @@ AUTOGRAD_OPTIMIZER_CLASS(Adam) {
   std::unordered_map<std::string, at::Tensor> max_exp_avg_sq_buffer_;
 };
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/include/torch/serialization.h

@@ -9,7 +9,7 @@
 #include "cereal/types/unordered_map.hpp"
 #include "cereal/types/vector.hpp"
 
-namespace autograd {
+namespace torch {
 
 // Some convenience functions for saving and loading
 template <typename T>
@@ -35,12 +35,12 @@ void load(std::istream& stream, T* obj) {
 template <typename T>
 void save(std::string const& path, T const& obj) {
   std::ofstream os(path, std::ios::binary);
-  autograd::save(os, obj);
+  torch::save(os, obj);
 }
 template <typename T>
 void load(std::string const& path, T& obj) {
   std::ifstream is(path, std::ios::binary);
-  autograd::load(is, obj);
+  torch::load(is, obj);
 }
 
 namespace detail {
@@ -108,21 +108,21 @@ inline at::Backend backendFromId(int32_t id) {
 }
 
 } // namespace detail
-} // namespace autograd
+} // namespace torch
 
 // This is super ugly and I don't know how to simplify it
-CEREAL_REGISTER_TYPE(autograd::SGD);
-CEREAL_REGISTER_POLYMORPHIC_RELATION(autograd::OptimizerImpl, autograd::SGD);
-CEREAL_REGISTER_TYPE(autograd::Adagrad);
+CEREAL_REGISTER_TYPE(torch::SGD);
+CEREAL_REGISTER_POLYMORPHIC_RELATION(torch::OptimizerImpl, torch::SGD);
+CEREAL_REGISTER_TYPE(torch::Adagrad);
 CEREAL_REGISTER_POLYMORPHIC_RELATION(
-    autograd::OptimizerImpl,
-    autograd::Adagrad);
-CEREAL_REGISTER_TYPE(autograd::RMSprop);
+    torch::OptimizerImpl,
+    torch::Adagrad);
+CEREAL_REGISTER_TYPE(torch::RMSprop);
 CEREAL_REGISTER_POLYMORPHIC_RELATION(
-    autograd::OptimizerImpl,
-    autograd::RMSprop);
-CEREAL_REGISTER_TYPE(autograd::Adam);
-CEREAL_REGISTER_POLYMORPHIC_RELATION(autograd::OptimizerImpl, autograd::Adam);
+    torch::OptimizerImpl,
+    torch::RMSprop);
+CEREAL_REGISTER_TYPE(torch::Adam);
+CEREAL_REGISTER_POLYMORPHIC_RELATION(torch::OptimizerImpl, torch::Adam);
 
 namespace cereal {
 
@@ -162,11 +162,11 @@ loadBinary(BinaryInputArchive& archive, void* data, std::size_t size) {
 template <class Archive>
 void save(Archive& archive, at::Tensor const& tensor) {
   if (!tensor.defined()) {
-    int32_t typeId = ::autograd::detail::scalarTypeId(at::ScalarType::Undefined);
+    int32_t typeId = ::torch::detail::scalarTypeId(at::ScalarType::Undefined);
     archive(CEREAL_NVP(typeId));
     return;
   } else {
-    int32_t typeId = ::autograd::detail::scalarTypeId(tensor.type().scalarType());
+    int32_t typeId = ::torch::detail::scalarTypeId(tensor.type().scalarType());
     archive(CEREAL_NVP(typeId));
   }
   auto sizes = std::vector<int64_t>();
@@ -175,7 +175,7 @@ void save(Archive& archive, at::Tensor const& tensor) {
     sizes.push_back(s);
   }
   auto contig = tensor.toBackend(at::kCPU).contiguous();
-  int32_t backend = ::autograd::detail::backendId(tensor.type().backend());
+  int32_t backend = ::torch::detail::backendId(tensor.type().backend());
 
   archive(CEREAL_NVP(backend), CEREAL_NVP(sizes));
   agimpl::saveBinary(
@@ -195,7 +195,7 @@ void load(Archive& archive, at::Tensor& tensor) {
   at::ScalarType type;
   int32_t typeId;
   archive(CEREAL_NVP(typeId));
-  type = ::autograd::detail::scalarTypeFromId(typeId);
+  type = ::torch::detail::scalarTypeFromId(typeId);
   if (type == at::ScalarType::Undefined) {
     tensor = at::Tensor();
     return;
@@ -206,7 +206,7 @@ void load(Archive& archive, at::Tensor& tensor) {
   auto buf = std::vector<uint8_t>();
   archive(CEREAL_NVP(backendId), CEREAL_NVP(sizes));
 
-  at::Backend backend = ::autograd::detail::backendFromId(backendId);
+  at::Backend backend = ::torch::detail::backendFromId(backendId);
   if (!tensor.defined() || tensor.type().scalarType() != type) {
     tensor = at::getType(backend, type).tensor();
   }

torch/csrc/api/src/containers.cpp

@@ -1,6 +1,6 @@
 #include "torch/containers.h"
 
-namespace autograd {
+namespace torch {
 std::map<std::string, Variable> ContainerImpl::parameters() const {
   std::map<std::string, Variable> ret;
   for (auto pair : children_) {
@@ -634,4 +634,4 @@ variable_list Dropout2d::forward(variable_list inputs) {
   return lst;
 }
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/src/detail.cpp

@@ -13,7 +13,7 @@
 
 #include "torch/detail.h"
 
-namespace autograd {
+namespace torch {
 namespace detail {
 tag::Engine engine;
 }
@@ -68,4 +68,4 @@ bool hasCudnn() {
   return hasCuda() && AT_CUDNN_ENABLED();
 }
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/src/optimizers.cpp

@@ -1,6 +1,6 @@
 #include "torch/optimizers.h"
 
-namespace autograd {
+namespace torch {
 
 void OptimizerImpl::zero_grad() {
   for (auto p : model_->parameters()) {
@@ -196,4 +196,4 @@ void Adam::init_state() {
   exp_avg_sq_buffer_.clear();
 }
 
-} // namespace autograd
+} // namespace torch

torch/csrc/api/src/python.cpp

@@ -1,4 +1,4 @@
-#include <torch/torch.h>
+#include <torch/csrc/api/include/torch/python.h>
 #include <torch/csrc/autograd/generated/VariableType.h>
 #include <torch/csrc/autograd/variable.h>
 

torch/csrc/tensor/python_tensor.cpp

@@ -2,9 +2,8 @@
 
 #include <structmember.h>
 #include <pybind11/pybind11.h>
-#include <sstream>
 
-#include "torch/torch.h"
+#include "torch/csrc/api/include/torch/python.h"
 #include "torch/csrc/assertions.h"
 #include "torch/csrc/Dtype.h"
 #include "torch/csrc/DynamicTypes.h"
@@ -20,6 +19,9 @@
 #include "torch/csrc/utils/tensor_new.h"
 #include "torch/csrc/utils/tensor_types.h"
 
+#include <sstream>
+#include <vector>
+
 namespace torch { namespace tensor {
 
 using namespace at;

torch/utils/cpp_extension.py

@@ -504,7 +504,7 @@ def load_inline(name,
     the necessary header includes, as well as the (pybind11) binding code. More
     precisely, strings passed to ``cpp_sources`` are first concatenated into a
     single ``.cpp`` file. This file is then prepended with ``#include
-    <torch/torch.h>``.
+    <torch/python.h>``.
 
     Furthermore, if the ``functions`` argument is supplied, bindings will be
     automatically generated for each function specified. ``functions`` can
@@ -551,7 +551,7 @@ def load_inline(name,
     if isinstance(cuda_sources, str):
         cuda_sources = [cuda_sources]
 
-    cpp_sources.insert(0, '#include <torch/torch.h>')
+    cpp_sources.insert(0, '#include <torch/python.h>')
 
     # If `functions` is supplied, we create the pybind11 bindings for the user.
     # Here, `functions` is (or becomes, after some processing) a map from
@@ -747,7 +747,7 @@ def _write_ninja_file(path,
     sources = [os.path.abspath(file) for file in sources]
     includes = [os.path.abspath(file) for file in extra_include_paths]
 
-    # include_paths() gives us the location of torch/torch.h
+    # include_paths() gives us the location of torch/python.h
     includes += include_paths(with_cuda)
     # sysconfig.get_paths()['include'] gives us the location of Python.h
     includes.append(sysconfig.get_paths()['include'])