pybind.cc 98.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

7
http://www.apache.org/licenses/LICENSE-2.0
8 9 10 11 12 13

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
L
lgone2000 已提交
14
#include <Python.h>
C
chengduoZH 已提交
15
#include <algorithm>
16
#include <cstdlib>
C
chengduoZH 已提交
17
#include <map>
S
sneaxiy 已提交
18
#include <memory>
C
chengduoZH 已提交
19 20 21
#include <mutex>  // NOLINT // for call_once
#include <string>
#include <unordered_map>
22
#include <unordered_set>
C
chengduoZH 已提交
23 24
#include <utility>
#include <vector>
Y
Yi Wang 已提交
25 26
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
Z
Zhen Wang 已提交
27
#include "paddle/fluid/framework/feed_fetch_type.h"
Y
Yi Wang 已提交
28
#include "paddle/fluid/framework/framework.pb.h"
S
sneaxiy 已提交
29
#include "paddle/fluid/framework/garbage_collector.h"
H
hutuxian 已提交
30
#include "paddle/fluid/framework/io/fs.h"
31
#include "paddle/fluid/framework/ir/coalesce_grad_tensor_pass.h"
32
#include "paddle/fluid/framework/ir/pass_builder.h"
33
#include "paddle/fluid/framework/load_op_lib.h"
Y
Yi Wang 已提交
34 35 36
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
37
#include "paddle/fluid/framework/op_compatible_info.h"
S
sneaxiy 已提交
38
#include "paddle/fluid/framework/op_info.h"
39
#include "paddle/fluid/framework/op_registry.h"
Y
Yu Yang 已提交
40
#include "paddle/fluid/framework/parallel_executor.h"
Y
Yi Wang 已提交
41
#include "paddle/fluid/framework/prune.h"
Y
Refine  
Yu Yang 已提交
42
#include "paddle/fluid/framework/reader.h"
H
hong 已提交
43
#include "paddle/fluid/framework/save_load_util.h"
S
sneaxiy 已提交
44
#include "paddle/fluid/framework/scope_pool.h"
Y
Yi Wang 已提交
45
#include "paddle/fluid/framework/selected_rows.h"
46
#include "paddle/fluid/framework/trainer.h"
47
#include "paddle/fluid/framework/type_defs.h"
X
Xin Pan 已提交
48
#include "paddle/fluid/framework/version.h"
H
hong 已提交
49
#include "paddle/fluid/imperative/layer.h"
Y
Refine  
Yu Yang 已提交
50
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
51
#include "paddle/fluid/memory/allocation/mmap_allocator.h"
D
dzhwinter 已提交
52
#include "paddle/fluid/operators/activation_op.h"
S
sneaxiy 已提交
53
#include "paddle/fluid/operators/py_func_op.h"
54
#include "paddle/fluid/platform/cpu_helper.h"
Y
Yu Yang 已提交
55
#include "paddle/fluid/platform/cpu_info.h"
56
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
Y
Yi Wang 已提交
57
#include "paddle/fluid/platform/enforce.h"
58
#include "paddle/fluid/platform/init.h"
Y
Yi Wang 已提交
59 60
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
H
hutuxian 已提交
61
#include "paddle/fluid/pybind/box_helper_py.h"
Y
Yi Wang 已提交
62
#include "paddle/fluid/pybind/const_value.h"
D
dongdaxiang 已提交
63
#include "paddle/fluid/pybind/data_set_py.h"
Y
Yi Wang 已提交
64
#include "paddle/fluid/pybind/exception.h"
D
dongdaxiang 已提交
65
#include "paddle/fluid/pybind/fleet_wrapper_py.h"
66
#include "paddle/fluid/pybind/global_value_getter_setter.h"
67
#include "paddle/fluid/pybind/gloo_wrapper_py.h"
68
#include "paddle/fluid/pybind/imperative.h"
F
flame 已提交
69
#include "paddle/fluid/pybind/inference_api.h"
F
flame 已提交
70
#include "paddle/fluid/pybind/ir.h"
71
#include "paddle/fluid/pybind/pybind_boost_headers.h"
72

73
#ifdef PADDLE_WITH_NCCL
D
dongdaxiang 已提交
74
#include "paddle/fluid/pybind/nccl_wrapper_py.h"
W
wopeizl 已提交
75
#endif
76
#include "paddle/fluid/framework/data_type.h"
77 78
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
S
sneaxiy 已提交
79
#include "paddle/fluid/pybind/reader_py.h"
Y
Yi Wang 已提交
80
#include "paddle/fluid/pybind/tensor_py.h"
81
#include "paddle/fluid/string/to_string.h"
D
Dong Zhihong 已提交
82
#ifdef PADDLE_WITH_CUDA
83
#ifdef PADDLE_WITH_NCCL
Y
Yi Wang 已提交
84
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
P
peizhilin 已提交
85
#endif
Y
Yi Wang 已提交
86 87
#include "paddle/fluid/platform/cuda_profiler.h"
#include "paddle/fluid/platform/gpu_info.h"
D
Dong Zhihong 已提交
88 89
#endif

90 91 92 93
#ifdef PADDLE_WITH_DISTRIBUTE
#include "paddle/fluid/pybind/communicator_py.h"
#endif

M
minqiyang 已提交
94 95
#include "pybind11/stl.h"

96
DECLARE_bool(use_mkldnn);
97

Q
Qiao Longfei 已提交
98 99
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
100 101 102
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchUnmergedList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchType);
Q
Qiao Longfei 已提交
103

104
namespace paddle {
105
namespace pybind {
106
bool IsCompiledWithCUDA() {
107
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
108 109 110 111 112 113
  return false;
#else
  return true;
#endif
}

114 115 116 117 118 119 120 121
bool IsCompiledWithMKLDNN() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return true;
#endif
}

122
bool IsCompiledWithBrpc() {
123
#ifndef PADDLE_WITH_DISTRIBUTE
124 125
  return false;
#endif
126 127 128 129 130 131

#ifdef PADDLE_WITH_GRPC
  return false;
#endif

  return true;
132 133
}

Y
update  
Yancey1989 已提交
134
bool IsCompiledWithDIST() {
Y
Yancey1989 已提交
135
#ifdef PADDLE_WITH_DISTRIBUTE
Y
update  
Yancey1989 已提交
136 137 138 139 140 141
  return true;
#else
  return false;
#endif
}

S
sneaxiy 已提交
142 143 144 145 146 147 148 149 150 151
template <typename PlaceType1, typename PlaceType2>
static inline bool IsSamePlace(const PlaceType1 &p1, const PlaceType2 &p2) {
  return paddle::platform::Place(p1) == paddle::platform::Place(p2);
}

template <typename PlaceType>
static inline int PlaceIndex(const PlaceType &p) {
  return static_cast<int>(paddle::platform::Place(p).which());
}

H
hong 已提交
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
static PyObject *GetPythonAttribute(PyObject *obj, const char *attr_name) {
  // NOTE(zjl): PyObject_GetAttrString would return nullptr when attr_name
  // is not inside obj, but it would also set the error flag of Python.
  // If the error flag is set in C++, C++ code would not raise Exception,
  // but Python would raise Exception once C++ call ends.
  // To avoid unexpected Exception raised in Python, we check whether
  // attribute exists before calling PyObject_GetAttrString.
  //
  // Caution: PyObject_GetAttrString would increase reference count of PyObject.
  // Developer should call Py_DECREF manually after the attribute is not used.
  if (PyObject_HasAttrString(obj, attr_name)) {
    return PyObject_GetAttrString(obj, attr_name);
  } else {
    return nullptr;
  }
}

template <typename T>
static T PyObjectCast(PyObject *obj) {
  try {
    return py::cast<T>(py::handle(obj));
  } catch (py::cast_error &) {
    PADDLE_THROW("Python object is not type of %s", typeid(T).name());
  }
}

using PyNameVarBaseMap = std::unordered_map<std::string, py::handle>;

static std::vector<std::shared_ptr<imperative::VarBase>> GetVarBaseList(
    const PyNameVarBaseMap &state_dict) {
  std::vector<std::shared_ptr<imperative::VarBase>> vec_res;
  vec_res.reserve(state_dict.size());

  for (auto &para : state_dict) {
    PyObject *py_obj = para.second.ptr();
    if (!py_obj || py_obj == Py_None) {
      PADDLE_THROW("Save parameter [%s] is None", para.first);
    }
    vec_res.emplace_back(
191
        PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_obj));
H
hong 已提交
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227
  }

  return vec_res;
}

static std::vector<std::string> inline GetNameList(
    const py::handle &py_handle) {
  std::vector<std::string> vec_res;

  PyObject *py_obj = py_handle.ptr();  // get underlying PyObject
  // Python None is not nullptr in C++!
  if (!py_obj || py_obj == Py_None) {
    PADDLE_THROW("Save parameter list is None");
  }

  if (PyList_Check(py_obj)) {
    size_t len = PyList_GET_SIZE(py_obj);

    vec_res.reserve(len);

    const char *kNameField = "name";

    for (size_t i = 0; i < len; ++i) {
      PyObject *py_name =
          PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kNameField);
      PADDLE_ENFORCE_NOT_NULL(py_name);
      vec_res.emplace_back(PyObjectCast<std::string>(py_name));
      Py_DECREF(py_name);
    }
  } else {
    PADDLE_THROW("Set parameter should be a list");
  }

  return vec_res;
}

228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277
static void inline CreateVariableIfNotExit(
    const py::handle &py_handle, const framework::Scope &scope,
    const framework::Executor *exe = nullptr) {
  std::vector<std::string> vec_res;

  PyObject *py_obj = py_handle.ptr();  // get underlying PyObject
  // Python None is not nullptr in C++!
  if (!py_obj || py_obj == Py_None) {
    PADDLE_THROW("Save parameter list is None");
  }

  if (PyList_Check(py_obj)) {
    size_t len = PyList_GET_SIZE(py_obj);

    vec_res.reserve(len);

    const char *kNameField = "name";
    const char *kVarDescField = "desc";

    for (size_t i = 0; i < len; ++i) {
      PyObject *py_name =
          PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kNameField);
      PADDLE_ENFORCE_NOT_NULL(py_name);
      auto para_name = PyObjectCast<std::string>(py_name);
      Py_DECREF(py_name);

      auto var = scope.FindVar(para_name);
      if (var == nullptr) {
        PADDLE_ENFORCE_NE(exe, nullptr,
                          "Parameter not Initialized, "
                          "Please set argument [executor] not None "
                          "or run startup program first");
        PyObject *py_var_desc =
            PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kVarDescField);
        PADDLE_ENFORCE_NOT_NULL(py_var_desc);
        auto var_desc = PyObjectCast<framework::VarDesc>(py_var_desc);
        Py_DECREF(py_var_desc);
        var = const_cast<framework::Scope *>(&scope)->Var(para_name);
        auto *tensor_temp = var->GetMutable<framework::LoDTensor>();
        tensor_temp->Resize(framework::make_ddim(var_desc.GetShape()));
        tensor_temp->mutable_data(exe->GetPlace(), var_desc.GetDataType());
      }
    }
  } else {
    PADDLE_THROW("Set parameter should be a list");
  }

  return;
}

278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
static void AssertStaticGraphAndDygraphGradMakerNoDiff() {
  std::set<std::string> ops;
  for (auto &pair : framework::OpInfoMap::Instance().map()) {
    bool has_static_grad_maker = (pair.second.grad_op_maker_ != nullptr);
    bool has_dygraph_grad_maker =
        (pair.second.dygraph_grad_op_maker_ != nullptr);
    if (has_static_grad_maker ^ has_dygraph_grad_maker) {
      bool has_kernel =
          (framework::OperatorWithKernel::AllOpKernels().count(pair.first) > 0);
      if (has_kernel) {
        ops.insert(pair.first);
      } else {
        VLOG(5) << pair.first << " has no kernels, skip";
      }
    }
  }
  PADDLE_ENFORCE_EQ(ops.empty(), true,
                    platform::errors::Unimplemented(
                        "OperatorWithKernel [%s] have only static graph grad "
                        "maker or have only dygraph grad maker, which is not "
                        "allowed",
                        string::join_strings(ops, ',')));
}

302 303 304 305 306 307
#ifdef PADDLE_WITH_AVX
PYBIND11_MODULE(core_avx, m) {
#else
PYBIND11_MODULE(core_noavx, m) {
#endif

Y
Yu Yang 已提交
308 309 310
  // Not used, just make sure cpu_info.cc is linked.
  paddle::platform::CpuTotalPhysicalMemory();

Y
Refine  
Yu Yang 已提交
311
  paddle::memory::allocation::UseAllocatorStrategyGFlag();
S
sneaxiy 已提交
312

313 314
  AssertStaticGraphAndDygraphGradMakerNoDiff();

315
  m.doc() = "C++ core of PaddlePaddle";
316

317 318 319 320
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

321
  BindException(&m);
Y
Yu Yang 已提交
322

323 324
  m.def("set_num_threads", &platform::SetNumThreads);

6
633WHU 已提交
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
  m.def("from_dlpack", [](py::capsule *dltensor) {
    DLManagedTensor *dmt = reinterpret_cast<DLManagedTensor *>(
        PyCapsule_GetPointer(dltensor->ptr(), "dltensor"));
    PyCapsule_SetName(dltensor->ptr(), "used_dltensor");
    DLTensor dl = dmt->dl_tensor;
    Tensor tensor;

    if (dl.ctx.device_type == kDLCPU) {
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
#ifdef PADDLE_WITH_CUDA
    if (dl.ctx.device_type == kDLGPU) {
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
#endif
    return tensor;
  });

H
hong 已提交
343 344 345 346 347 348 349 350 351
  m.def("_save_static_dict",
        [](const std::string &str_file_name, const py::handle &vec_var_list,
           const Scope &scope) {
          std::vector<std::string> vec_name_list = GetNameList(vec_var_list);
          SaveStaticNameListToDisk(str_file_name, vec_name_list, scope);
        });

  m.def("_load_static_dict",
        [](const std::string &str_file_name, const py::handle &vec_var_list,
352
           const Scope &scope, const Executor *executor) {
H
hong 已提交
353
          std::vector<std::string> vec_name_list = GetNameList(vec_var_list);
354
          CreateVariableIfNotExit(vec_var_list, scope, executor);
H
hong 已提交
355 356 357
          LoadStaticNameListFromDisk(str_file_name, vec_name_list, scope);
        });

358 359 360 361 362 363
  m.def("_create_loaded_parameter",
        [](const py::handle &vec_var_list, const Scope &scope,
           const Executor *executor) {
          CreateVariableIfNotExit(vec_var_list, scope, executor);
        });

H
hong 已提交
364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
  m.def("_save_dygraph_dict", [](const std::string &str_file_name,
                                 const PyNameVarBaseMap &state_dict) {
    auto vec_var_base_list = GetVarBaseList(state_dict);

    SaveDygraphVarBaseListToDisk(str_file_name, vec_var_base_list);
  });

  m.def("_load_dygraph_dict", [](const std::string &str_file_name) {
    auto load_tensor = LoadDygraphVarBaseListFromDisk(str_file_name);

    std::unordered_map<std::string, std::shared_ptr<imperative::VarBase>>
        map_output;

    for (size_t i = 0; i < load_tensor.size(); ++i) {
      map_output.emplace(load_tensor[i]->Name(), load_tensor[i]);
    }

    return map_output;
  });
6
633WHU 已提交
383

384 385 386 387 388 389
  m.def("save_op_compatible_info", [](framework::ProgramDesc &desc) {
    framework::OpCompatibleMap op_compatible_map;
    op_compatible_map.InitOpCompatibleMap();
    return op_compatible_map.ConvertToProto(desc.OpCompatibleMap());
  });

S
sneaxiy 已提交
390
  m.def(
S
sneaxiy 已提交
391
      "_append_python_callable_object_and_return_id",
S
sneaxiy 已提交
392 393 394 395
      [](py::object py_obj) -> size_t {
        return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
      });

S
sneaxiy 已提交
396 397 398
  m.def("_get_use_default_grad_op_desc_maker_ops",
        [] { return OpInfoMap::Instance().GetUseDefaultGradOpDescMakerOps(); });

399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414
  m.def("_get_all_register_op_kernels", [] {
    auto &all_kernels = paddle::framework::OperatorWithKernel::AllOpKernels();
    std::unordered_map<std::string, std::vector<std::string>> all_kernels_info;
    for (auto &kernel_pair : all_kernels) {
      auto op_type = kernel_pair.first;
      std::vector<std::string> kernel_types;
      for (auto &info_pair : kernel_pair.second) {
        paddle::framework::OpKernelType kernel_type = info_pair.first;
        kernel_types.push_back(
            paddle::framework::KernelTypeToString(kernel_type));
      }
      all_kernels_info.emplace(op_type, kernel_types);
    }
    return all_kernels_info;
  });

S
sneaxiy 已提交
415 416 417
  // NOTE(zjl): ctest would load environment variables at the beginning even
  // though we have not `import paddle.fluid as fluid`. So we add this API
  // to enable eager deletion mode in unittest.
S
sneaxiy 已提交
418
  m.def("_set_eager_deletion_mode", &paddle::framework::SetEagerDeletionMode);
S
sneaxiy 已提交
419

420
  m.def("_set_fuse_parameter_group_size",
421
        &paddle::framework::ir::SetFuseParameterGroupsSize);
422
  m.def("_set_fuse_parameter_memory_size",
423
        &paddle::framework::ir::SetFuseParameterMemorySize);
424

S
sneaxiy 已提交
425 426 427
  m.add_object("_cleanup",
               py::capsule([]() { ScopePool::Instance().Clear(); }));

428 429
  m.def("_set_paddle_lib_path", &paddle::platform::dynload::SetPaddleLibPath);

430
  BindImperative(&m);
431

432
  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
433
      .def("__array__", [](Tensor &self) { return TensorToPyArray(self); })
S
sneaxiy 已提交
434 435
      .def("_is_initialized",
           [](const Tensor &self) { return self.IsInitialized(); })
Y
yuyang18 已提交
436
      .def("_get_dims",
437
           [](const Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
438
      .def("_set_dims",
Q
qijun 已提交
439
           [](Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
440
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
441
           })
Y
yuyang18 已提交
442
      .def("_set_layout",
D
dzhwinter 已提交
443 444 445
           [](Tensor &self, const std::string &layout) {
             self.set_layout(StringToDataLayout(layout));
           })
Y
yuyang18 已提交
446
      .def("_alloc_float",
D
dzhwinter 已提交
447
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
448
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
449
           })
Y
yuyang18 已提交
450
      .def("_alloc_float",
Y
Yu Yang 已提交
451
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
452
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
453
           })
454 455 456 457
      .def("_alloc_double",
           [](Tensor &self, paddle::platform::CPUPlace &place) {
             self.mutable_data<double>(place);
           })
Y
yuyang18 已提交
458
      .def("_alloc_int",
Y
Yu Yang 已提交
459
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
460
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
461
           })
Y
yuyang18 已提交
462
      .def("_alloc_int",
D
dzhwinter 已提交
463
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
464
             self.mutable_data<int>(place);
Q
qijun 已提交
465
           })
Y
yuyang18 已提交
466
      .def("_alloc_int",
C
chengduoZH 已提交
467 468 469
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
470
      .def("_alloc_float",
C
chengduoZH 已提交
471 472 473
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<float>(place);
           })
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488
      .def("_mutable_data",
           [](Tensor &self, paddle::platform::CPUPlace &place,
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
      .def("_mutable_data",
           [](Tensor &self, paddle::platform::CUDAPlace &place,
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
      .def("_mutable_data",
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place,
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
Z
Zeng Jinle 已提交
489
      .def("_clear", &Tensor::clear)
490
      .def("set", SetTensorFromPyArray<paddle::platform::CPUPlace>,
491
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
492
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPlace>,
493
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
494
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPinnedPlace>,
495 496
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false,
           R"DOC(
L
Leo Chen 已提交
497 498 499 500 501 502
        Set the data of LoDTensor on place with given numpy array.
        
        Args:
          lod (numpy.ndarray): The data to set.
          place (CPUPlace|CUDAPlace|CUDAPinnedPlace): The place where the 
          LoDTensor is to be set.
503 504
          zero_copy (bool, optional): Whether to share memory with the input numpy array.
          This parameter only works with CPUPlace. Default: False.
L
Leo Chen 已提交
505 506 507 508 509 510 511 512 513 514 515 516 517

        Returns:
            None.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                t = fluid.LoDTensor()
                t.set(np.ndarray([5, 30]), fluid.CPUPlace())
          )DOC")
518

L
Leo Chen 已提交
519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535
      .def("shape", [](Tensor &self) { return vectorize(self.dims()); }, R"DOC(
           Return the shape of LoDTensor.

           Returns:
               list[int]: The shape of LoDTensor.


           Examples:
               .. code-block:: python

                  import paddle.fluid as fluid
                  import numpy as np

                  t = fluid.LoDTensor()
                  t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                  print(t.shape())  # [5, 30]
           )DOC")
6
633WHU 已提交
536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
      .def("_to_dlpack",
           [](Tensor &self) {
             DLPackTensor dlpack_tensor(self, 1);
             DLManagedTensor *dmt =
                 dlpack_tensor.ToCudfCompatibleDLManagedTensor();
             auto capsule = py::capsule(
                 static_cast<void *>(dmt), "dltensor", [](PyObject *ptr) {
                   if (ptr) {
                     auto dltensor = new DLManagedTensor;
                     try {
                       dltensor = reinterpret_cast<DLManagedTensor *>(
                           PyCapsule_GetPointer(ptr, "used_dltensor"));
                       return;
                     } catch (...) {
                       dltensor = reinterpret_cast<DLManagedTensor *>(
                           PyCapsule_GetPointer(ptr, "dltensor"));
                     }
                     dltensor->deleter(dltensor);
                   }
                 });
             return capsule;
           })
Y
yuyang18 已提交
558 559 560 561
      .def("_set_float_element", TensorSetElement<float>)
      .def("_get_float_element", TensorGetElement<float>)
      .def("_set_double_element", TensorSetElement<double>)
      .def("_get_double_element", TensorGetElement<double>)
X
xuezhong 已提交
562
      .def("_place", [](Tensor &self) { return self.place(); })
W
wopeizl 已提交
563
      .def("_dtype", [](Tensor &self) { return self.type(); })
564
      .def("_share_data_with", &Tensor::ShareDataWith)
565 566 567 568 569 570
      .def("__getitem__", PySliceTensor, py::return_value_policy::reference)
      .def("__str__", [](const Tensor &self) {
        std::stringstream ostr;
        ostr << self;
        return ostr.str();
      });
Y
Yu Yang 已提交
571

L
Leo Chen 已提交
572
  // TODO(cql): add reference: en_user_guide_lod_tensor
X
Xin Pan 已提交
573
  py::class_<LoDTensor, Tensor>(m, "LoDTensor", R"DOC(
L
Leo Chen 已提交
574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
    LoDTensor is a Tensor with optional LoD (Level of Details) information, 
    it can be used for variable-length sequences, 
    see :ref:`user_guide_lod_tensor` for details.

    LoDTensor can be converted to numpy array using :code:`numpy.array(lod_tensor)`.

    You can skip the following explanation if you don't need to know details 
    of LoDTensor.

    The following two examples show how to use LODtensor to represent 
    variable-length sequences.
    
    Example 1:
    
    Suppose x is a LoDTensor representing a variable-length sequence. 
    It contains two logical subsequences, the length of first logical sequence 
    is 2 (e.g., number of samples is 2), the length of second logical sequence 
    is 3, and the total length is 5. The data of the first logical sequence is 
    [1, 2], [3, 4], and the data of the second logical sequence is [5, 6], 
    [7, 8], [9, 10]. The data dimension of each sample is 2. So, the final 
    shape of the LoDTensor is [5, 2], of which 5 is the total length and 2 is 
    the dimension of each sample.
    
    Logically, we can represent the variable-length sequence in two ways: one 
    is in the form of recursive sequence lengths, that is, 
    x.recursive_sequence_lengths=[[2, 3]]; the other is in the form of offsets, 
    that is, x.lod=[[0, 2, 2+3]]. These two representations are equivalent, and 
    you can set and retrieve recursive_sequence_lengths or LoD through the 
    corresponding interfaces of LoDTensor introduced later.

    Actually, in order to access sequence faster, Paddle uses offset to store 
    different lengths of sequences. 
    Therefore, the operations on recursive_sequence_lengths will be converted 
    to the operations on LoD eventually.
    
    .. code-block:: python

      y.data = [[1, 2], [3, 4],
                [5, 6], [7, 8],
                [9, 10], [11, 12], [13, 14]]

      y.shape = [2+2+3, 2]

      y.recursive_sequence_lengths = [[2, 1], [2, 2, 3]]

      y.lod = [[0, 2, 3], [0, 2, 4, 7]]

    Example 2:

    LoD may have more than one level (for example, a paragraph may have more 
    than one sentence and a sentence may have more than one word). Suppose y 
    is a LoDTensor and its lod_level is 2. 
    From level = 0, there are two logical sequences, the length of which is 
    2 and 1, respectively, indicating that the first logical sequence contains 
    two sub-sequences and the second logical sequence contains one sub-sequence. 
    From level = 1, the lengths of two sub-sequences contained by the first 
    logical sequence is 2 and 2, and the length of sub-sequence contained by 
    the second logical sequence is 3.
      
    Therefore, the LoDTensor is represented in the form of recursive sequence 
    lengths as y.recursive_sequence_lengths=[[2,1], [2,2,3]]; and equally, in 
    the form of offset, it is represented as y.lod=[[0,2,3], [0,2,4,7]].

    .. code-block:: python

      y.data = [[1, 2], [3, 4],
                [5, 6], [7, 8],
                [9, 10], [11, 12], [13, 14]]

      y.shape = [2+2+3, 2]

      y.recursive_sequence_lengths = [[2, 1], [2, 2, 3]]

      y.lod = [[0, 2, 3], [0, 2, 4, 7]]
Z
Zeng Jinle 已提交
648 649 650 651 652 653 654

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid

          t = fluid.LoDTensor()
X
Xin Pan 已提交
655 656

        )DOC")
657
      .def("__array__", [](Tensor &self) { return TensorToPyArray(self); })
658 659 660 661 662 663 664 665 666
      .def("__init__",
           [](LoDTensor &instance, const std::vector<std::vector<size_t>>
                                       &recursive_sequence_lengths) {
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
C
chengduo 已提交
667 668
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, -1), true,
669 670 671
                 "the provided recursive_sequence_lengths info is invalid");
             new (&instance) LoDTensor(new_offset_lod);
           })
Y
Yu Yang 已提交
672
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
G
gongweibao 已提交
673
      // We implement offset based LOD in C++ while we use length based with
H
hong 已提交
674 675
      // Python API. So we changed set_lod to set_recursive_sequence_lengths
      // to
G
gongweibao 已提交
676 677 678
      // avoid misuse.
      // The discussion is here:
      // https://github.com/PaddlePaddle/Paddle/issues/10855
D
dangqingqing 已提交
679
      .def("set_lod",
680
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
681
             // the input lod is offset-based level-of-detail info
Y
Yu Yang 已提交
682
             LoD new_lod;
683 684
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
C
chengduo 已提交
685 686 687
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_lod, vectorize(self.dims()).front()), true,
                 "the provided lod info is invalid");
688
             self.set_lod(new_lod);
S
sneaxiy 已提交
689 690 691 692 693
           },
           py::arg("lod"), R"DOC(
           Set LoD of the LoDTensor.

           Args:
L
Leo Chen 已提交
694 695 696 697
               lod (list[list[int]]): The lod to set.

           Returns:
                None.
Z
Zeng Jinle 已提交
698 699 700 701 702 703 704 705 706 707

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_lod([[0, 2, 5]])
L
Leo Chen 已提交
708
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
709
           )DOC")
710 711 712 713 714 715 716 717 718 719 720
      .def("set_recursive_sequence_lengths",
           [](LoDTensor &self, const std::vector<std::vector<size_t>>
                                   &recursive_sequence_lengths) {
             // the input recursive_sequence_lengths is length-based
             // level-of-detail info
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
C
chengduo 已提交
721 722
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, vectorize(self.dims()).front()), true,
723 724
                 "the provided recursive_sequence_lengths info is invalid");
             self.set_lod(new_offset_lod);
S
sneaxiy 已提交
725 726
           },
           py::arg("recursive_sequence_lengths"), R"DOC(
L
Leo Chen 已提交
727
           Set LoD of the LoDTensor according to recursive sequence lengths.
S
sneaxiy 已提交
728

L
Leo Chen 已提交
729
           For example, if recursive_sequence_lengths=[[2, 3]], which means
730
           there are two sequences with length 2 and 3 respectively, the
L
Leo Chen 已提交
731
           corresponding lod would be [[0, 2, 2+3]], i.e., [[0, 2, 5]].
S
sneaxiy 已提交
732 733

           Args:
L
Leo Chen 已提交
734 735 736 737
                recursive_sequence_lengths (list[list[int]]): The recursive sequence lengths.
           
           Returns:
                None.
Z
Zeng Jinle 已提交
738 739 740 741 742 743 744 745 746 747

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
L
Leo Chen 已提交
748 749
                 print(t.recursive_sequence_length())  # [[2, 3]]
                 print(t.lod())  # [[0, 2, 5]]
S
sneaxiy 已提交
750
           )DOC")
751 752 753 754 755 756 757 758
      .def("lod",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the offset-based lod info
             LoD lod = self.lod();
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
S
sneaxiy 已提交
759 760 761 762 763
           },
           R"DOC(
           Return the LoD of the LoDTensor.

           Returns:
L
Leo Chen 已提交
764 765
               list[list[int]]: The lod of the LoDTensor.
           
Z
Zeng Jinle 已提交
766 767 768 769 770 771 772 773 774 775
           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_lod([[0, 2, 5]])
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
776
           )DOC")
G
gongweibao 已提交
777
      // Set above comments of set_lod.
778 779 780 781 782 783 784 785
      .def("recursive_sequence_lengths",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the length-based lod info
             LoD lod = ConvertToLengthBasedLoD(self.lod());
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
S
sneaxiy 已提交
786 787
           },
           R"DOC(
L
Leo Chen 已提交
788 789
           Return the recursive sequence lengths corresponding to of the LodD 
           of the LoDTensor.
S
sneaxiy 已提交
790 791

           Returns:
L
Leo Chen 已提交
792
                list[list[int]]: The recursive sequence lengths.
Z
Zeng Jinle 已提交
793 794 795 796 797 798 799 800 801 802 803

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
                 print(t.recursive_sequence_lengths()) # [[2, 3]]
S
sneaxiy 已提交
804 805 806 807 808 809 810 811
           )DOC")
      .def("has_valid_recursive_sequence_lengths",
           [](LoDTensor &self) -> bool {
             // Check that the lod info is valid and match the outermost
             // dimension of the LoDTensor data
             return CheckLoD(self.lod(), vectorize(self.dims()).front());
           },
           R"DOC(
L
Leo Chen 已提交
812
           Check whether the LoD of the LoDTensor is valid.
S
sneaxiy 已提交
813 814

           Returns:
L
Leo Chen 已提交
815
               bool: Whether the LoD is valid.
Z
Zeng Jinle 已提交
816 817 818 819 820 821 822 823 824 825 826

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
                 print(t.has_valid_recursive_sequence_lengths()) # True
W
wopeizl 已提交
827 828 829 830 831 832 833
           )DOC")
      .def("__getitem__", PySliceTensor, py::return_value_policy::reference,
           R"DOC(
           Slice the original Tensor, and remove the LoD information.

           Returns:
               out (Tensor): new Tensor(NOT LoDTensor).
834
           )DOC")
835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852
      .def("__str__",
           [](const LoDTensor &self) {
             std::stringstream ostr;
             ostr << self;
             return ostr.str();
           })
      .def("_copy", [](const LoDTensor &self, const platform::Place &place) {
        // follow fetch_op's inplementation
        LoDTensor dst;
        if (self.IsInitialized() && self.numel() > 0) {
          TensorCopySync(self, place, &dst);
        } else {
          // Not copy, if the src tensor is empty.
          dst.clear();
          dst.Resize({0});
        }
        dst.set_lod(self.lod());
        return dst;
853
#ifdef _WIN32
854
      });
855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904
#else
           })
      .def(py::pickle(
          [](const LoDTensor &t) {  // __getstate__
            auto holder = t.Holder();
            PADDLE_ENFORCE_EQ(
              platform::is_cpu_place(holder->place()), true,
              platform::errors::PreconditionNotMet(
                  "LoDTensor is not on CPU."
                  "Now only LoDTensor on CPU can be serialized."));
            auto* mmap_writer_allocation =
              dynamic_cast<memory::allocation::MemoryMapWriterAllocation *>(
                holder.get());
            PADDLE_ENFORCE_NOT_NULL(mmap_writer_allocation,
              platform::errors::PreconditionNotMet(
                "LoDTensor is not in shared memory."
                "Now only LoDTensor on shared memory can be serialized."));
            int type_idx = static_cast<int>(t.type());

            return py::make_tuple(mmap_writer_allocation->ipc_name(),
                                  mmap_writer_allocation->size(),
                                  type_idx, vectorize(t.dims()), t.lod());
          },
          [](py::tuple t) {  // __setstate__
            if (t.size() != 5)
              throw std::runtime_error("Invalid LoDTensor state!");

            // 1. Create a new C++ instance
            LoDTensor tensor;

            // 2. Rebuild Allocation
            const std::string &ipc_name = t[0].cast<std::string>();
            size_t size = t[1].cast<size_t>();
            auto shared_reader_holder =
              memory::allocation::RebuildMemoryMapReaderAllocation(
                ipc_name, size);

            // 3. Maintain global fd set
            VLOG(3) << "LoDTensor ipc name: " << ipc_name;
            memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);

            // 4. Rebuild LoDTensor
            tensor.ResetHolderWithType(shared_reader_holder,
              static_cast<proto::VarType::Type>(t[2].cast<int>()));
            tensor.Resize(make_ddim(t[3].cast<std::vector<int>>()));
            tensor.set_lod(t[4].cast<framework::LoD>());

            return tensor;
          }));
#endif
D
dangqingqing 已提交
905

Q
qijun 已提交
906 907 908 909 910 911 912 913 914 915 916
  py::class_<SelectedRows>(m, "SelectedRows")
      .def("__init__",
           [](SelectedRows &instance) { new (&instance) SelectedRows(); })
      .def("__init__",
           [](SelectedRows &instance, const std::vector<int64_t> rows,
              const int64_t &height) {
             new (&instance) SelectedRows(rows, height);
           })
      .def("get_tensor",
           [](SelectedRows &self) { return self.mutable_value(); },
           py::return_value_policy::reference)
917 918
      .def("numel",
           [](SelectedRows &self) -> int64_t { return self.value().numel(); })
Q
qijun 已提交
919 920
      .def("set_height", &SelectedRows::set_height)
      .def("height", &SelectedRows::height)
Q
qijun 已提交
921 922 923 924 925 926 927 928 929
      .def("set_rows",
           [](SelectedRows &self, std::vector<int64_t> rows) {
#ifndef PADDLE_WITH_CUDA
             self.set_rows(rows);
#else
        Vector<int64_t> new_rows(rows);
        self.set_rows(new_rows);
#endif
           })
930
      .def("sync_index", [](SelectedRows &instance) { instance.SyncIndex(); })
931
      .def("rows", [](SelectedRows &self) {
932 933 934 935 936
        auto rows = self.rows();
        std::vector<int64_t> new_rows;
        new_rows.reserve(rows.size());
        std::copy(rows.begin(), rows.end(), std::back_inserter(new_rows));
        return new_rows;
937
      });
Q
qijun 已提交
938

939
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
940 941 942

All parameter, weight, gradient are variables in Paddle.
)DOC")
S
sneaxiy 已提交
943
      .def(py::init<>())
944
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
945
      .def("set_int",
946 947
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
948 949 950 951 952 953 954
      .def("is_float", [](const Variable &var) { return var.IsType<float>(); })
      .def("set_float",
           [](Variable &var, float val) -> void {
             *var.GetMutable<float>() = val;
           })
      .def("get_float",
           [](const Variable &var) -> float { return var.Get<float>(); })
Y
Yu Yang 已提交
955
      .def("get_tensor",
956 957
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
958 959
           },
           py::return_value_policy::reference)
960 961 962 963
      .def("get_bytes",
           [](Variable &self) {
             return py::bytes(*self.GetMutable<std::string>());
           })
Y
Yu Yang 已提交
964 965 966
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
967 968 969 970 971
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
972 973 974
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
975 976 977
      .def("get_fetch_list",
           [](Variable &self) { return self.GetMutable<FetchList>(); },
           py::return_value_policy::reference)
978
#if (defined(PADDLE_WITH_NCCL))
D
Dong Zhihong 已提交
979 980 981 982 983
      .def("get_communicator",
           [](Variable &self) -> platform::Communicator * {
             return self.GetMutable<platform::Communicator>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
984
#endif
Y
Refine  
Yu Yang 已提交
985 986
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
C
chengduo 已提交
987
             PADDLE_ENFORCE_EQ(self.IsType<framework::ReaderHolder>(), true);
Y
Refine  
Yu Yang 已提交
988 989
             return self.GetMutable<framework::ReaderHolder>();
           },
990 991 992 993 994
           py::return_value_policy::reference)
      .def("set_scope", [](Variable &self, Scope &scope) {
        auto scope_vec = self.GetMutable<std::vector<framework::Scope *>>();
        scope_vec->emplace_back(&scope);
      });
995

S
sneaxiy 已提交
996
  BindReader(&m);
Y
Refine  
Yu Yang 已提交
997

S
sneaxiy 已提交
998
  py::class_<Scope>(m, "_Scope", R"DOC(
Q
Qiao Longfei 已提交
999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011
    Scope is an association of a name to Variable. All variables belong to Scope.

    Variables in a parent scope can be retrieved from local scope.

    You need to specify a scope to run a Net, i.e., `exe.Run(&scope)`.
    One net can run in different scopes and update different variable in the
    scope.

    You can create var in a scope and get it from the scope.

    Examples:
        .. code-block:: python

1012
          import paddle.fluid as fluid
Q
Qiao Longfei 已提交
1013 1014 1015 1016 1017 1018
          # create tensor from a scope and set value to it.
          param = scope.var('Param').get_tensor()
          param_array = np.full((height, row_numel), 5.0).astype("float32")
          param.set(param_array, place)

        )DOC")
S
sneaxiy 已提交
1019 1020
      .def("_remove_from_pool",
           [](Scope &self) { ScopePool::Instance().Remove(&self); })
D
dongzhihong 已提交
1021
      .def("var",
1022
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
1023
             return self.Var(name);
Y
Yu Yang 已提交
1024
           },
S
sneaxiy 已提交
1025 1026
           py::arg("name"),
           R"DOC(
1027
           Find or create variable named :code:`name` in the current scope.
S
sneaxiy 已提交
1028

1029
           If the variable named :code:`name` does not exist in the
S
sneaxiy 已提交
1030
           current scope, the variable would be created. Otherwise,
1031
           return the existing variable.
S
sneaxiy 已提交
1032 1033

           Args:
1034 1035
               name (str): the variable name.

S
sneaxiy 已提交
1036
           Returns:
1037
               out (core.Variable): the found or created variable.
S
sneaxiy 已提交
1038 1039 1040 1041
           )DOC",
           py::return_value_policy::reference)
      .def("find_var", &Scope::FindVar, py::arg("name"),
           R"DOC(
1042
           Find variable named :code:`name` in the current scope or
S
sneaxiy 已提交
1043
           its parent scope. Return None if not found.
1044

S
sneaxiy 已提交
1045 1046
           Args:
               name (str): the variable name.
1047

S
sneaxiy 已提交
1048
           Returns:
1049
               out (core.Variable|None): the found variable or None.
S
sneaxiy 已提交
1050
           )DOC",
1051
           py::return_value_policy::reference)
1052
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
S
sneaxiy 已提交
1053 1054 1055 1056 1057 1058
           R"DOC(
           Create a new sub-scope of the current scope.

           Returns:
               out (core._Scope): the created sub-scope.
           )DOC",
1059
           py::return_value_policy::reference)
S
sneaxiy 已提交
1060 1061 1062
      .def("drop_kids", &Scope::DropKids,
           R"DOC(
           Delete all sub-scopes of the current scope.
S
sneaxiy 已提交
1063 1064
           )DOC")
      .def("_kids", &Scope::kids);
1065

S
sneaxiy 已提交
1066 1067 1068 1069 1070 1071
  m.def("Scope",
        []() -> Scope * {
          auto *s = new Scope();
          ScopePool::Instance().Insert(std::unique_ptr<Scope>(s));
          return s;
        },
S
sneaxiy 已提交
1072 1073
        R"DOC(
        Create a new scope.
1074

S
sneaxiy 已提交
1075 1076 1077
        Returns:
            out (core._Scope): the created scope.
        )DOC",
S
sneaxiy 已提交
1078 1079
        py::return_value_policy::reference);

Y
Yu Yang 已提交
1080 1081
  //! @note: Be careful! PyBind will return std::string as an unicode, not
  //! Python str. If you want a str object, you should cast them in Python.
Y
Yu Yang 已提交
1082 1083
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
1084 1085 1086 1087
    for (auto &iter : OpInfoMap::Instance().map()) {
      auto &info = iter.second;
      if (info.HasOpProtoAndChecker()) {
        std::string str;
C
chengduo 已提交
1088 1089
        PADDLE_ENFORCE_EQ(
            info.Proto().SerializeToString(&str), true,
1090 1091 1092 1093
            "Serialize OpProto Error. This could be a bug of Paddle.");
        ret_values.emplace_back(str);
      }
    }
Y
Yu Yang 已提交
1094 1095
    return ret_values;
  });
1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108
  m.def("get_op_attrs_default_value",
        [](py::bytes byte_name) -> paddle::framework::AttributeMap {
          std::string op_type = byte_name;
          paddle::framework::AttributeMap res;
          auto info = OpInfoMap::Instance().GetNullable(op_type);
          if (info != nullptr) {
            if (info->HasOpProtoAndChecker()) {
              auto op_checker = info->Checker();
              res = op_checker->GetAttrsDefaultValuesMap();
            }
          }
          return res;
        });
1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
  m.def(
      "get_grad_op_desc", [](const OpDesc &op_desc,
                             const std::unordered_set<std::string> &no_grad_set,
                             const std::vector<BlockDesc *> &grad_sub_block) {
        std::unordered_map<std::string, std::string> grad_to_var;
        std::vector<std::unique_ptr<OpDesc>> grad_op_descs =
            framework::OpInfoMap::Instance()
                .Get(op_desc.Type())
                .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
                               grad_sub_block);
        std::vector<OpDesc *> grad_op_desc_ptrs(grad_op_descs.size());
        std::transform(grad_op_descs.begin(), grad_op_descs.end(),
                       grad_op_desc_ptrs.begin(),
                       [](std::unique_ptr<OpDesc> &p) { return p.release(); });
        return std::make_pair(grad_op_desc_ptrs, grad_to_var);
      });
1125 1126 1127
  m.def("has_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasGradOpMaker();
  });
1128 1129 1130 1131 1132
  m.def("has_non_empty_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance()
        .Get(op_type)
        .HasNonEmptyGradOpMaker();
  });
1133 1134 1135
  m.def("has_infer_inplace", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasInferInplace();
  });
1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149
  m.def("infer_no_need_buffer_slots",
        [](const std::string op_type, const framework::VariableNameMap &inputs,
           const framework::VariableNameMap &outputs,
           const framework::AttributeMap &attrs) {
          auto infer_func = framework::OpInfoMap::Instance()
                                .Get(op_type)
                                .NoNeedBufferVarsInferer();
          if (infer_func) {
            return infer_func(inputs, outputs, attrs);
          } else {
            std::unordered_set<std::string> empty = {};
            return empty;
          }
        });
Y
Yu Yang 已提交
1150
  m.def("prune", [](const ProgramDesc &origin,
1151
                    const std::set<std::string> &feeded_var_names,
1152
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
1153
    ProgramDesc prog_with_targets(origin);
1154

1155
    for (const auto &t : targets) {
1156
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
1157
    }
1158
    proto::ProgramDesc pruned_desc;
1159 1160 1161 1162
    auto pruned_origin_block_id_map =
        Prune(*prog_with_targets.Proto(), feeded_var_names, &pruned_desc);
    return std::make_tuple(ProgramDesc(pruned_desc),
                           pruned_origin_block_id_map);
1163
  });
1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180
  m.def("prune_backward",
        [](const framework::ProgramDesc &program) {
          return PruneBackward(program);
        },
        R"DOC(
             Prune the backward part of a program, mostly called in
             program.clone(for_test=True).
              
             Args:
                   program (ProgramDesc): The original program.

             Returns:
                   tuple(ProgramDesc, map<int, int>): The first part is 
                   the pruned program desc, and the second part is a map
                   which contains the id pair of pruned block and corresponding
                   origin block.
           )DOC");
1181 1182 1183 1184
  m.def("empty_var_name",
        []() { return std::string(framework::kEmptyVarName); });
  m.def("grad_var_suffix",
        []() { return std::string(framework::kGradVarSuffix); });
1185 1186
  m.def("loaded_var_suffix",
        []() { return std::string(framework::kLoadedVarSuffix); });
1187 1188 1189
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
1190 1191
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
Q
qijun 已提交
1192
  // clang-format off
Y
Yu Yang 已提交
1193
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
1194 1195
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
1196
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
1197 1198 1199
                    return new paddle::platform::CPUDeviceContext();
                  })
      .def_static("create",
D
dzhwinter 已提交
1200
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
1201
                      -> paddle::platform::DeviceContext* {
1202
#ifndef PADDLE_WITH_CUDA
D
dzhwinter 已提交
1203
                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
Q
qijun 已提交
1204
#else
Q
qijun 已提交
1205
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
1206
#endif
C
chengduoZH 已提交
1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217
                  })
          .def_static("create",
                [](paddle::platform::CUDAPinnedPlace& place)
                        -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_CUDA
                  PADDLE_THROW(
                        "CUDAPinnedPlace is not supported in CPU device.");
#else
                  return new paddle::platform::CUDAPinnedDeviceContext(place);
#endif
                });;
D
Dong Zhihong 已提交
1218
// clang-format on
1219
#if defined(PADDLE_WITH_NCCL)
D
Dong Zhihong 已提交
1220 1221
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
1222
  py::class_<platform::CUDAPlace>(m, "CUDAPlace", R"DOC(
1223 1224 1225 1226 1227 1228 1229 1230
    **Note**:
        For multi-card tasks, please use `FLAGS_selected_gpus` environment variable to set the visible GPU device.
        The next version will fix the problem with `CUDA_VISIBLE_DEVICES` environment variable.

    CUDAPlace is a descriptor of a device.
    It represents a GPU device allocated or to be allocated with Tensor or LoDTensor.
    Each CUDAPlace has a dev_id to indicate the graphics card ID represented by the current CUDAPlace,
    staring from 0.
1231
    The memory of CUDAPlace with different dev_id is not accessible.
1232 1233 1234 1235 1236 1237 1238 1239
    Numbering here refers to the logical ID of the visible graphics card, not the actual ID of the graphics card.
    You can set visible GPU devices by setting the `CUDA_VISIBLE_DEVICES` environment variable.
    When the program starts, visible GPU devices will be numbered from 0.
    If `CUDA_VISIBLE_DEVICES` is not set, all devices are visible by default,
    and the logical ID is the same as the actual ID.

    Parameters:
        id (int): GPU device ID.
L
lujun 已提交
1240 1241 1242 1243

    Examples:
        .. code-block:: python

1244
          import paddle.fluid as fluid
L
lujun 已提交
1245 1246
          gpu_place = fluid.CUDAPlace(0)

1247
        )DOC")
S
sneaxiy 已提交
1248 1249 1250
      .def("__init__",
           [](platform::CUDAPlace &self, int dev_id) {
#ifdef PADDLE_WITH_CUDA
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid CUDAPlace(%d), device id must be 0 or "
                   "positive integer",
                   dev_id);
               std::exit(-1);
             }

             if (UNLIKELY(dev_id >= platform::GetCUDADeviceCount())) {
               if (platform::GetCUDADeviceCount() == 0) {
                 LOG(ERROR) << "Cannot use GPU because there is no GPU "
                               "detected on your "
                               "machine.";
                 std::exit(-1);
               } else {
                 LOG(ERROR) << string::Sprintf(
                     "Invalid CUDAPlace(%d), must inside [0, %d), because GPU "
                     "number on your machine is %d",
                     dev_id, platform::GetCUDADeviceCount(),
                     platform::GetCUDADeviceCount());
                 std::exit(-1);
               }
             }

S
sneaxiy 已提交
1275 1276
             new (&self) platform::CUDAPlace(dev_id);
#else
1277 1278 1279 1280 1281 1282 1283 1284 1285
             LOG(ERROR) << string::Sprintf(
                 "Cannot use GPU because you have installed CPU version "
                 "PaddlePaddle.\n"
                 "If you want to use GPU, please try to install GPU version "
                 "PaddlePaddle by: pip install paddlepaddle-gpu\n"
                 "If you only have CPU, please change CUDAPlace(%d) to be "
                 "CPUPlace().\n",
                 dev_id);
             std::exit(-1);
S
sneaxiy 已提交
1286 1287
#endif
           })
S
sneaxiy 已提交
1288 1289 1290 1291 1292 1293
      .def("_type", &PlaceIndex<platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPlace, platform::CUDAPinnedPlace>)
D
dzhwinter 已提交
1294
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
1295

1296
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace", R"DOC(
1297 1298
    CPUPlace is a descriptor of a device.
    It represents a CPU device allocated or to be allocated with Tensor or LoDTensor.
L
lujun 已提交
1299 1300 1301 1302

    Examples:
        .. code-block:: python

1303
          import paddle.fluid as fluid
1304
          cpu_place = fluid.CPUPlace()to be allocated
L
lujun 已提交
1305

1306
        )DOC")
1307
      .def(py::init<>())
S
sneaxiy 已提交
1308 1309 1310 1311 1312 1313
      .def("_type", &PlaceIndex<platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CPUPlace, platform::CUDAPinnedPlace>)
1314
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
1315

1316
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace", R"DOC(
1317 1318 1319 1320 1321 1322
    CUDAPinnedPlace is a descriptor of a device.
    It refers to the page locked memory allocated by the CUDA function `cudaHostAlloc()` in the host memory.
    The host operating system will not paging and exchanging the memory.
    It can be accessed through direct memory access technology to speed up the copy of data between the host and GPU.
    For more information on CUDA data transfer and `pinned memory`,
    please refer to `official document <https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#pinned-memory>`_ .
L
lujun 已提交
1323 1324 1325 1326

    Examples:
        .. code-block:: python

1327
          import paddle.fluid as fluid
L
lujun 已提交
1328 1329
          place = fluid.CUDAPinnedPlace()

1330
        )DOC")
S
sneaxiy 已提交
1331
      .def("__init__",
S
sneaxiy 已提交
1332
           [](platform::CUDAPinnedPlace &self) {
S
sneaxiy 已提交
1333 1334 1335
#ifndef PADDLE_WITH_CUDA
             PADDLE_THROW("Cannot use CUDAPinnedPlace in CPU only version");
#endif
S
sneaxiy 已提交
1336
             new (&self) platform::CUDAPinnedPlace();
S
sneaxiy 已提交
1337
           })
S
sneaxiy 已提交
1338 1339 1340 1341 1342 1343 1344 1345
      .def("_type", &PlaceIndex<platform::CUDAPinnedPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPinnedPlace, platform::Place>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPlace>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPinnedPlace>)
C
chengduoZH 已提交
1346 1347
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
1348 1349
  py::class_<platform::Place>(m, "Place")
      .def(py::init<>())
S
sneaxiy 已提交
1350 1351 1352 1353 1354
      .def("_type", &PlaceIndex<platform::Place>)
      .def("_equals", &IsSamePlace<platform::Place, platform::Place>)
      .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::Place, platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPinnedPlace>)
X
xuezhong 已提交
1355 1356
      .def("is_gpu_place",
           [](platform::Place &self) { return platform::is_gpu_place(self); })
S
sneaxiy 已提交
1357 1358 1359 1360 1361 1362
      .def("is_cpu_place",
           [](platform::Place &self) { return platform::is_cpu_place(self); })
      .def("is_cuda_pinned_place",
           [](platform::Place &self) {
             return platform::is_cuda_pinned_place(self);
           })
X
xuezhong 已提交
1363 1364 1365 1366
      .def("gpu_device_id",
           [](platform::Place &self) {
             return boost::get<platform::CUDAPlace>(self).device;
           })
S
sneaxiy 已提交
1367 1368
      .def("set_place", [](platform::Place &self,
                           const platform::Place &other) { self = other; })
Y
Yu Yang 已提交
1369 1370 1371 1372 1373
      .def("set_place",
           [](platform::Place &self, const platform::CPUPlace &cpu_place) {
             self = cpu_place;
           })
      .def("set_place",
D
dzhwinter 已提交
1374
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
1375
             self = gpu_place;
C
chengduoZH 已提交
1376 1377
           })
      .def("set_place", [](platform::Place &self,
C
chengduoZH 已提交
1378 1379
                           const platform::CUDAPinnedPlace &cuda_pinned_place) {
        self = cuda_pinned_place;
C
chengduoZH 已提交
1380
      });
Y
Yu Yang 已提交
1381

Y
Yu Yang 已提交
1382
  py::class_<OperatorBase>(m, "Operator")
C
chengduo 已提交
1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393
      .def_static(
          "create",
          [](py::bytes protobin) {
            proto::OpDesc desc;
            PADDLE_ENFORCE_EQ(desc.ParsePartialFromString(protobin), true,
                              "Cannot parse user input to OpDesc");
            PADDLE_ENFORCE_EQ(desc.IsInitialized(), true,
                              "User OpDesc is not initialized, reason %s",
                              desc.InitializationErrorString());
            return OpRegistry::CreateOp(desc);
          })
1394
      .def("run",
1395
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1396 1397 1398
              const platform::CPUPlace &place) { self.Run(scope, place); })
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1399
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
1400 1401 1402 1403 1404
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
1405 1406 1407 1408 1409 1410 1411
      .def("type",
           [](const OperatorBase &op) -> std::string { return op.Type(); })
      .def("outputs",
           [](const OperatorBase &op)
               -> std::map<std::string, std::vector<std::string>> {
                 return op.Outputs();
               })
Q
qijun 已提交
1412 1413
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
1414
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
1415
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
1416 1417 1418 1419
      .def("__str__", &OperatorBase::DebugString)
      .def("no_intermediate_outputs",
           [](const OperatorBase &op) { return op.OutputVars(false); })
      .def("support_gpu", &OperatorBase::SupportGPU);
Y
Yu Yang 已提交
1420

1421 1422 1423
  py::class_<framework::ExecutorPrepareContext>(m, "ExecutorPrepareContext")
      .def(py::init<const ProgramDesc &, size_t>());

1424 1425 1426 1427 1428 1429 1430 1431 1432
  py::class_<framework::TrainerBase, std::shared_ptr<framework::TrainerBase>>(
      m, "TrainerBase")
      .def("get_worker_scope",
           [](TrainerBase &self, int thread_id) -> Scope * {
             return self.GetWorkerScope(thread_id);
           },
           py::return_value_policy::reference)
      .def("finalize", &TrainerBase::Finalize);

F
fengjiayi 已提交
1433
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
1434
      .def(py::init<const platform::Place &>())
Y
Yancey1989 已提交
1435
      .def("close", &Executor::Close)
1436 1437
      .def("run_from_dataset", &Executor::RunFromDataset,
           py::call_guard<py::gil_scoped_release>())
D
Dong Daxiang 已提交
1438 1439
      .def("release_trainer", &Executor::ReleaseTrainer,
           py::call_guard<py::gil_scoped_release>())
1440 1441 1442 1443
      .def("init_for_dataset",
           [](Executor &self, const ProgramDesc &prog,
              const std::string &trainer_desc, Scope *scope,
              Dataset *dataset) -> std::shared_ptr<TrainerBase> {
D
Dong Daxiang 已提交
1444
             pybind11::gil_scoped_release release;
1445 1446 1447 1448 1449 1450 1451
             return self.InitForDataset(prog, trainer_desc, scope, dataset);
           })
      .def("run_from_dataset",
           [](Executor &self, std::shared_ptr<TrainerBase> trainer) {
             pybind11::gil_scoped_release release;
             self.RunFromDataset(trainer);
           })
1452 1453 1454
      .def("run_prepared_ctx",
           [](Executor &self, ExecutorPrepareContext *ctx, Scope *scope,
              std::map<std::string, const LoDTensor *> *feed_targets,
1455
              std::map<std::string, FetchType *> *fetch_targets,
1456 1457 1458 1459 1460 1461 1462 1463
              bool create_local_scope = true, bool create_vars = true,
              const std::string &feed_holder_name = "feed",
              const std::string &fetch_holder_name = "fetch") {
             pybind11::gil_scoped_release release;
             self.RunPreparedContext(ctx, scope, feed_targets, fetch_targets,
                                     create_local_scope, create_vars,
                                     feed_holder_name, fetch_holder_name);
           })
1464
      .def("run_prepared_ctx",
G
guru4elephant 已提交
1465 1466 1467 1468 1469 1470 1471
           [](Executor &self, ExecutorPrepareContext *ctx, Scope *scope,
              bool create_local_scope = true, bool create_vars = true,
              bool keep_kids = false) {
             pybind11::gil_scoped_release release;
             self.RunPreparedContext(ctx, scope, create_local_scope,
                                     create_vars, keep_kids);
           })
1472 1473 1474 1475 1476 1477 1478 1479 1480 1481
      .def("prepare",
           [](Executor &self, const ProgramDesc &program, int block_id,
              const std::vector<std::string> &skip_ref_cnt_vars =
                  std::vector<std::string>(),
              bool force_disable_gc = false) {
             pybind11::gil_scoped_release release;
             return self.Prepare(program, block_id, skip_ref_cnt_vars,
                                 force_disable_gc);
           })
      .def("create_variables", &Executor::CreateVariables)
S
sneaxiy 已提交
1482
      .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope,
S
sneaxiy 已提交
1483 1484
                     int block_id, bool create_local_scope, bool create_vars,
                     const std::vector<std::string> &fetch_vars) {
S
sneaxiy 已提交
1485
        pybind11::gil_scoped_release release;
S
sneaxiy 已提交
1486 1487
        self.Run(prog, scope, block_id, create_local_scope, create_vars,
                 fetch_vars);
S
sneaxiy 已提交
1488
      });
S
sneaxiy 已提交
1489

D
dzhwinter 已提交
1490
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
1491
  m.def("init_glog", framework::InitGLOG);
1492
  m.def("load_op_library", framework::LoadOpLib);
X
Xin Pan 已提交
1493 1494
  m.def("init_devices",
        [](bool init_p2p) { framework::InitDevices(init_p2p); });
1495

1496
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
1497
  m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
1498
  m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
Y
update  
Yancey1989 已提交
1499
  m.def("is_compiled_with_dist", IsCompiledWithDIST);
H
hutuxian 已提交
1500 1501 1502 1503 1504 1505 1506
  m.def("run_cmd",
        [](const std::string &cmd, int time_out = -1,
           int sleep_inter = -1) -> const std::string {
          return paddle::framework::shell_get_command_output(cmd, time_out,
                                                             sleep_inter);
        },
        py::arg("cmd"), py::arg("time_out") = -1, py::arg("sleep_inter") = -1);
1507 1508 1509 1510 1511 1512
#ifdef PADDLE_WITH_CUDA
  m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool {
    // Only GPUs with Compute Capability >= 53 support float16
    return platform::GetCUDAComputeCapability(place.device) >= 53;
  });
#endif
1513

1514
  m.def("set_feed_variable", framework::SetFeedVariable);
1515 1516 1517 1518 1519 1520 1521 1522 1523 1524
  m.def("get_fetch_variable",
        [](const Scope &scope, const std::string &var_name,
           size_t index) -> py::object {
          auto &var = framework::GetFetchVariable(scope, var_name, index);
          if (data_is_lod_tensor(var)) {
            return py::cast(boost::get<LoDTensor>(var));
          } else {
            return py::cast(boost::get<LoDTensorArray>(var));
          }
        });
1525
  m.def("get_variable_tensor", framework::GetVariableTensor);
Q
qijun 已提交
1526

X
Xin Pan 已提交
1527 1528
  m.def("_is_program_version_supported", IsProgramVersionSupported);

1529 1530 1531 1532 1533
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
1534
  BindGlobalValueGetterSetter(&m);
Y
Yu Yang 已提交
1535

Y
Yu Yang 已提交
1536 1537 1538 1539 1540 1541 1542 1543 1544
  py::class_<framework::LoDRankTable>(m, "LodRankTable")
      .def("items", [](framework::LoDRankTable &table) {
        std::vector<std::pair<size_t, size_t>> res;
        for (auto &item : table.items()) {
          res.push_back({item.index, item.length});
        }
        return res;
      });

Z
Zeng Jinle 已提交
1545
  py::class_<LoDTensorArray>(m, "LoDTensorArray", R"DOC(
1546
    LoDTensorArray is array of LoDTensor, it supports operator[], len() and for-loop iteration.
Z
Zeng Jinle 已提交
1547 1548 1549

    Examples:
        .. code-block:: python
1550

Z
Zeng Jinle 已提交
1551 1552 1553 1554
          import paddle.fluid as fluid

          arr = fluid.LoDTensorArray()
)DOC")
S
sneaxiy 已提交
1555 1556
      .def("__init__",
           [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); })
Y
Yu Yang 已提交
1557 1558 1559 1560 1561 1562 1563 1564 1565 1566
      .def("__getitem__",
           [](LoDTensorArray &self, size_t i) { return &self.at(i); },
           py::return_value_policy::reference)
      .def("__len__", [](LoDTensorArray &self) { return self.size(); })
      .def("__setitem__",
           [](LoDTensorArray &self, size_t i, const LoDTensor &t) {
             PADDLE_ENFORCE_LT(i, self.size());
             self[i].ShareDataWith(t);
             self[i].set_lod(t.lod());
           })
S
sneaxiy 已提交
1567 1568 1569 1570 1571 1572
      .def("append",
           [](LoDTensorArray &self, const LoDTensor &t) {
             self.emplace_back();
             self.back().ShareDataWith(t);
             self.back().set_lod(t.lod());
           },
Z
Zeng Jinle 已提交
1573 1574
           py::arg("tensor"), R"DOC(
             Append a LoDensor to LoDTensorArray.
1575 1576 1577 1578 1579 1580
              
             Args:
                   tensor (LoDTensor): The LoDTensor to be appended.

             Returns:
                   None.
Z
Zeng Jinle 已提交
1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591

             Examples:
                 .. code-block:: python

                   import paddle.fluid as fluid
                   import numpy as np

                   arr = fluid.LoDTensorArray()
                   t = fluid.LoDTensor()
                   t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                   arr.append(t)
1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602
           )DOC")
      .def("_move_to_list",
           [](LoDTensorArray &self) -> py::list {
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               res[i] = py::cast(std::move(self[i]));
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership);
Y
Yu Yang 已提交
1603

1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649
  py::class_<FetchList>(m, "FetchList", R"DOC( FetchList is a
        vector of boost::variant<LoDTensor, LoDTensorArray>.
        )DOC")
      .def("_move_to_list",
           [](FetchList &self) -> py::list {
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               if (data_is_lod_tensor(self[i])) {
                 auto &data = boost::get<LoDTensor>(self[i]);
                 res[i] = py::cast(std::move(data));
               } else {
                 auto &data = boost::get<LoDTensorArray>(self[i]);
                 py::list tmp(data.size());
                 for (size_t j = 0; j < data.size(); ++j) {
                   tmp[j] = py::cast(std::move(data[j]));
                 }
                 res[i] = std::move(tmp);
               }
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership)

      .def("append",
           [](FetchList &self, const LoDTensor &t) {
             self.emplace_back();
             auto &lod_tensor = boost::get<LoDTensor>(self.back());
             lod_tensor.ShareDataWith(t);
             lod_tensor.set_lod(t.lod());
           },
           py::arg("var"))

      .def("append",
           [](FetchList &self, const LoDTensorArray &t) {
             self.emplace_back();
             auto &lod_tensor_array = boost::get<LoDTensorArray>(self.back());
             for (size_t i = 0; i < t.size(); ++i) {
               lod_tensor_array[i].ShareDataWith(t[i]);
               lod_tensor_array[i].set_lod(t[i].lod());
             }
           },
           py::arg("var"));

  py::class_<FetchUnmergedList>(m, "FetchUnmergedList", R"DOC(
        FetchUnmergedList is 2-D array of FetchType(boost::variant(LoDTensor, LoDTensorArray)).
Z
Zhen Wang 已提交
1650 1651
        )DOC")
      .def("_move_to_list",
1652
           [](FetchUnmergedList &self) -> py::list {
Z
Zhen Wang 已提交
1653 1654 1655 1656
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               py::list tmp(self[i].size());
               for (size_t j = 0; j < self[i].size(); ++j) {
1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667
                 if (data_is_lod_tensor(self[i][j])) {
                   auto &var = boost::get<LoDTensor>(self[i][j]);
                   tmp[j] = py::cast(std::move(var));
                 } else {
                   auto &var = boost::get<LoDTensorArray>(self[i][j]);
                   py::list tmp_array(var.size());
                   for (size_t k = 0; k < var.size(); ++k) {
                     tmp_array[k] = std::move(var[k]);
                   }
                   tmp[j] = std::move(tmp_array);
                 }
Z
Zhen Wang 已提交
1668 1669 1670 1671 1672 1673 1674 1675 1676
               }
               res[i] = std::move(tmp);
               self[i].clear();
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership);

Y
Yu Yang 已提交
1677
  m.def("op_support_gpu", OpSupportGPU);
D
Dong Zhihong 已提交
1678
#ifdef PADDLE_WITH_CUDA
D
Dong Zhihong 已提交
1679
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
1680

P
peizhilin 已提交
1681
#ifndef _WIN32
D
dangqingqing 已提交
1682 1683 1684
  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
D
Dong Zhihong 已提交
1685
#endif
P
peizhilin 已提交
1686
#endif
Y
Yu Yang 已提交
1687

1688 1689 1690 1691 1692 1693
  py::enum_<platform::TracerOption>(m, "TracerOption", py::arithmetic())
      .value("kDefault", platform::TracerOption::kDefault)
      .value("kOpDetail", platform::TracerOption::kOpDetail)
      .value("kAllOpDetail", platform::TracerOption::kAllOpDetail)
      .export_values();

1694 1695 1696 1697
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
1698
      .value("kAll", platform::ProfilerState::kAll)
1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709
      .export_values();

  py::enum_<platform::EventSortingKey>(m, "EventSortingKey", py::arithmetic())
      .value("kDefault", platform::EventSortingKey::kDefault)
      .value("kCalls", platform::EventSortingKey::kCalls)
      .value("kTotal", platform::EventSortingKey::kTotal)
      .value("kMin", platform::EventSortingKey::kMin)
      .value("kMax", platform::EventSortingKey::kMax)
      .value("kAve", platform::EventSortingKey::kAve)
      .export_values();

1710
  m.def("set_tracer_option", platform::SetTracerOption);
1711 1712
  m.def("enable_profiler", platform::EnableProfiler);
  m.def("disable_profiler", platform::DisableProfiler);
X
Xin Pan 已提交
1713
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
1714
  m.def("reset_profiler", platform::ResetProfiler);
1715
  m.def("get_pass", [](const std::string &pass_type) {
W
WangZhen 已提交
1716 1717 1718
    auto pass = framework::ir::PassRegistry::Instance().Get(pass_type);
    return std::shared_ptr<framework::ir::Pass>(std::move(pass));
  });
Y
Yu Yang 已提交
1719

1720 1721
  m.def("size_of_dtype", framework::SizeOfType);

1722 1723 1724
  using VarQuantScale =
      std::unordered_map<std::string, std::pair<bool, LoDTensor>>;

1725 1726
  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
W
WangZhen 已提交
1727
      .def("has", &ir::Pass::Has)
1728 1729 1730
      .def("set_not_owned",
           [](ir::Pass &self, const std::string &attr_name, ProgramDesc &attr) {
             self.SetNotOwned<ProgramDesc>(attr_name, &attr);
W
WangZhen 已提交
1731
           })
1732
      .def(
1733
          "set",
1734 1735 1736
          [](ir::Pass &self, const std::string &name, const std::string &attr) {
            self.Set<std::string>(name, new std::string(attr));
          })
1737 1738
      .def("set", [](ir::Pass &self, const std::string &name,
                     bool val) { self.Set<bool>(name, new bool(val)); })
1739 1740
      .def("set", [](ir::Pass &self, const std::string &name,
                     int val) { self.Set<const int>(name, new int(val)); })
1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754
      .def("set",
           [](ir::Pass &self, const std::string &name,
              std::unordered_set<std::string> set) {
             self.Set(name, new std::unordered_set<std::string>(set));
           })
      .def("set",
           [](ir::Pass &self, const std::string &name,
              std::unordered_set<int> set) {
             self.Set(name, new std::unordered_set<int>(set));
           })
      .def("set",
           [](ir::Pass &self, const std::string &name, VarQuantScale scales) {
             self.Set(name, new VarQuantScale(scales));
           })
F
flame 已提交
1755 1756
      .def("type", &ir::Pass::Type)
      .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
1757
        self.Apply(graph.get());
F
flame 已提交
1758
      });
1759

X
fix  
Xin Pan 已提交
1760 1761
  py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
      m, "PassBuilder");
1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775
  pb.def(py::init())
      .def("append_pass",
           [](ir::PassBuilder &self,
              const std::string &pass_type) -> std::shared_ptr<ir::Pass> {
             return self.AppendPass(pass_type);
           })
      .def("all_passes", [](ir::PassBuilder &self) { return self.AllPasses(); })
      .def("insert_pass",
           [](ir::PassBuilder &self, size_t idx, const std::string &pass_type) {
             return self.InsertPass(idx, pass_type);
           })
      .def("remove_pass",
           [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });

Y
yuyang18 已提交
1776
  // -- python binds for parallel executor.
X
Xin Pan 已提交
1777

Y
yuyang18 已提交
1778
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
C
chengduo 已提交
1779 1780 1781 1782
  py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
    ExecutionStrategy allows the user to more preciously control how to run
    the program in ParallelExecutor by setting the property.

C
chengduo 已提交
1783 1784 1785
    Examples:
        .. code-block:: python

1786
          import paddle.fluid as fluid
1787 1788 1789 1790 1791 1792 1793 1794 1795 1796
          x = fluid.layers.data(name='x', shape=[13], dtype='float32')
          y = fluid.layers.data(name='y', shape=[1], dtype='float32')
          y_predict = fluid.layers.fc(input=x, size=1, act=None)

          cost = fluid.layers.square_error_cost(input=y_predict, label=y)
          avg_loss = fluid.layers.mean(cost)

          sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
          sgd_optimizer.minimize(avg_loss)

C
chengduo 已提交
1797 1798 1799
          exec_strategy = fluid.ExecutionStrategy()
          exec_strategy.num_threads = 4

1800 1801
          train_exe = fluid.ParallelExecutor(use_cuda=False,
                                             loss_name=avg_loss.name,
C
chengduo 已提交
1802 1803
                                             exec_strategy=exec_strategy)

C
chengduo 已提交
1804 1805
        )DOC");

Y
yuyang18 已提交
1806
  exec_strategy.def(py::init())
Y
yuyang18 已提交
1807 1808 1809 1810 1811
      .def_property(
          "num_threads",
          [](const ExecutionStrategy &self) { return self.num_threads_; },
          [](ExecutionStrategy &self, size_t num_threads) {
            self.num_threads_ = num_threads;
C
chengduo 已提交
1812 1813 1814 1815 1816 1817 1818 1819 1820 1821
          },
          R"DOC(The type is INT, num_threads represents the size of thread pool that
            used to run the operators of the current program in ParallelExecutor.
            If :math:`num\_threads=1`, all the operators will execute one by one,
            but the order maybe difference between iterations.
            If it is not set, it will be set in ParallelExecutor according to the
            device type and device count, for GPU, :math:`num\_threads=device\_count*4`, for CPU,
            :math:`num\_threads=CPU\_NUM*4`, the explanation of:math:`CPU\_NUM` is in ParallelExecutor.
            if it is not set, ParallelExecutor will get the cpu count by calling
            `multiprocessing.cpu_count()`. Default 0.)DOC")
Y
yuyang18 已提交
1822
      .def_property(
1823 1824 1825 1826
          "use_cuda",
          [](const ExecutionStrategy &self) { return self.use_cuda_; },
          [](ExecutionStrategy &self, bool use_cuda) {
            self.use_cuda_ = use_cuda;
C
chengduo 已提交
1827 1828 1829 1830
          })  // FIXME(chengduo): Doesn't add doc for 'use_cuda', use_cuda may
      // make user confuse, because ParallelExecutor has a parameter named
      // 'use_cuda' too, in current implementation, ParallelExecutor's
      // 'use_cuda' will rewrite ExecutionStrategy's 'use_cuda'.
Y
yuyang18 已提交
1831 1832 1833 1834 1835
      .def_property(
          "allow_op_delay",
          [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
          [](ExecutionStrategy &self, bool allow_op_delay) {
            self.allow_op_delay_ = allow_op_delay;
C
chengduo 已提交
1836 1837 1838
          },
          R"DOC(The type is BOOL, allow_op_delay represents whether to delay the
                communication operators to run, it may make the execution faster.
1839 1840
                Note that this option is invalid now, and it will be removed in
                next version. Default False.)DOC")
Y
yuyang18 已提交
1841 1842 1843 1844 1845 1846 1847
      .def_property(
          "num_iteration_per_drop_scope",
          [](const ExecutionStrategy &self) {
            return self.num_iteration_per_drop_scope_;
          },
          [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
            self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
C
chengduo 已提交
1848 1849 1850 1851
          },
          R"DOC(The type is INT, num_iteration_per_drop_scope indicates how
                many iterations to clean up the temp variables which
                is generated during execution. It may make the execution faster,
1852 1853
                because the temp variable's shape maybe the same between two iterations.
                Default 1.
C
chengduo 已提交
1854 1855 1856 1857 1858 1859

                NOTES:
                    1. If you fetch data when calling the 'run', the ParallelExecutor
                       will clean up the temp variables at the end of the current iteration.
                    2. In some NLP model, it may cause the GPU memory is insufficient,
                       in this case, you should reduce `num_iteration_per_drop_scope`.
1860
              )DOC")
Q
Qiao Longfei 已提交
1861 1862 1863 1864 1865 1866 1867 1868 1869
      .def_property(
          "num_iteration_per_run",
          [](const ExecutionStrategy &self) {
            return self.num_iteration_per_run_;
          },
          [](ExecutionStrategy &self, size_t num_iteration_per_run) {
            self.num_iteration_per_run_ = num_iteration_per_run;
          },
          R"DOC(This config that how many iteration the executor will run when
1870
                user call exe.run() in python
Q
Qiao Longfei 已提交
1871
              )DOC")
1872 1873 1874 1875 1876 1877 1878 1879
      .def_property(
          "use_thread_barrier",
          [](const ExecutionStrategy &self) { return self.thread_barrier_; },
          [](ExecutionStrategy &self, bool use_thread_barrier) {
            self.thread_barrier_ = use_thread_barrier;
          },
          R"DOC(This config that the this is distributed training with parameter server
              )DOC")
1880 1881 1882 1883 1884
      .def_property("_dry_run",
                    [](const ExecutionStrategy &self) { return self.dry_run_; },
                    [](ExecutionStrategy &self, bool dry_run) {
                      self.dry_run_ = dry_run;
                    });
C
chengduo 已提交
1885

Y
yuyang18 已提交
1886
  exec_strategy.def_property(
Y
yuyang18 已提交
1887 1888 1889 1890 1891 1892 1893
      "use_experimental_executor",
      [](const ExecutionStrategy &self) {
        return self.type_ == ExecutionStrategy::kExperimental;
      },
      [](ExecutionStrategy &self, bool experimental) {
        self.type_ = experimental ? ExecutionStrategy::kExperimental
                                  : ExecutionStrategy::kDefault;
Y
yuyang18 已提交
1894 1895
      });

C
chengduo 已提交
1896 1897 1898 1899
  py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy", R"DOC(
    BuildStrategy allows the user to more preciously control how to
    build the SSA Graph in ParallelExecutor by setting the property.

C
chengduo 已提交
1900 1901 1902
    Examples:
        .. code-block:: python

1903 1904
            import os
            import numpy as np
F
flame 已提交
1905
            import paddle.fluid as fluid
1906 1907 1908 1909 1910 1911 1912 1913 1914

            os.environ["CPU_NUM"] = '2'
            places = fluid.cpu_places()

            data = fluid.layers.data(name="x", shape=[1], dtype="float32")
            hidden = fluid.layers.fc(input=data, size=10)
            loss = fluid.layers.mean(hidden)
            fluid.optimizer.SGD(learning_rate=0.01).minimize(loss)

F
flame 已提交
1915
            build_strategy = fluid.BuildStrategy()
1916 1917
            build_strategy.enable_inplace = True
            build_strategy.memory_optimize = True
F
flame 已提交
1918
            build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
1919 1920 1921 1922
            program = fluid.compiler.CompiledProgram(fluid.default_main_program())
            program = program.with_data_parallel(loss_name=loss.name,
                                                 build_strategy=build_strategy,
                                                 places=places)
C
chengduo 已提交
1923
)DOC");
Y
yuyang18 已提交
1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939

  py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
      .value("Reduce", BuildStrategy::ReduceStrategy::kReduce)
      .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce);
  py::enum_<BuildStrategy::GradientScaleStrategy>(build_strategy,
                                                  "GradientScaleStrategy")
      .value("CoeffNumDevice",
             BuildStrategy::GradientScaleStrategy::kCoeffNumDevice)
      .value("One", BuildStrategy::GradientScaleStrategy::kOne)
      .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized);

  build_strategy.def(py::init())
      .def_property(
          "reduce_strategy",
          [](const BuildStrategy &self) { return self.reduce_; },
          [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
1940 1941 1942 1943
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
1944
            self.reduce_ = strategy;
C
chengduo 已提交
1945
          },
1946
          R"DOC((fluid.BuildStrategy.ReduceStrategy, optional): there are two reduce
C
chengduo 已提交
1947 1948
                strategies in ParallelExecutor, AllReduce and Reduce. If you want
                that all the parameters' optimization are done on all devices independently,
1949
                you should choose AllReduce; otherwise, if you choose Reduce, all the parameters'
C
chengduo 已提交
1950 1951
                optimization will be evenly distributed to different devices, and then
                broadcast the optimized parameter to other devices.
1952
                Default is 'AllReduce'.
F
flame 已提交
1953 1954 1955 1956 1957 1958 1959 1960

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
                  )DOC")
Y
yuyang18 已提交
1961 1962 1963 1964 1965
      .def_property(
          "gradient_scale_strategy",
          [](const BuildStrategy &self) { return self.gradient_scale_; },
          [](BuildStrategy &self,
             BuildStrategy::GradientScaleStrategy strategy) {
1966 1967 1968 1969
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
1970
            self.gradient_scale_ = strategy;
C
chengduo 已提交
1971
          },
1972 1973
          R"DOC((fluid.BuildStrategy.GradientScaleStrategy, optional): there are three
                ways of defining :math:`loss@grad` in ParallelExecutor, that is, CoeffNumDevice,
C
chengduo 已提交
1974 1975
                One and Customized. By default, ParallelExecutor sets the :math:`loss@grad`
                according to the number of devices. If you want to customize :math:`loss@grad`,
1976
                you can choose Customized. Default is 'CoeffNumDevice'.
F
flame 已提交
1977 1978 1979 1980 1981

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
C
chengduo 已提交
1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
                        import paddle.fluid.compiler as compiler
                        import numpy
                        import os

                        use_cuda = True
                        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
                        exe = fluid.Executor(place)

                        # NOTE: If you use CPU to run the program, you need
                        # to specify the CPU_NUM, otherwise, fluid will use
                        # all the number of the logic core as the CPU_NUM,
                        # in that case, the batch size of the input should be
                        # greater than CPU_NUM, if not, the process will be
                        # failed by an exception.
                        if not use_cuda:
                            os.environ['CPU_NUM'] = str(2)
                            places = fluid.cpu_places()
                        else:
                            places = places = fluid.cuda_places()

                        data = fluid.layers.data(name='X', shape=[1], dtype='float32')
                        hidden = fluid.layers.fc(input=data, size=10)
                        loss = fluid.layers.mean(hidden)
                        fluid.optimizer.SGD(learning_rate=0.01).minimize(loss)

                        fluid.default_startup_program().random_seed=1
                        exe.run(fluid.default_startup_program())

F
flame 已提交
2010
                        build_strategy = fluid.BuildStrategy()
C
chengduo 已提交
2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024
                        build_strategy.gradient_scale_strategy = \
                                 fluid.BuildStrategy.GradientScaleStrategy.Customized
                        compiled_prog = compiler.CompiledProgram(
                                 fluid.default_main_program()).with_data_parallel(
                                          loss_name=loss.name, build_strategy=build_strategy,
                                          places = places)

                        dev_count =  len(places)
                        x = numpy.random.random(size=(10, 1)).astype('float32')
                        loss_grad = numpy.ones((dev_count)).astype("float32") * 0.01
                        loss_grad_name = loss.name+"@GRAD"
                        loss_data = exe.run(compiled_prog,
                                             feed={"X": x, loss_grad_name : loss_grad},
                                             fetch_list=[loss.name, loss_grad_name])
F
flame 已提交
2025
                   )DOC")
Y
yuyang18 已提交
2026 2027 2028 2029
      .def_property(
          "debug_graphviz_path",
          [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
          [](BuildStrategy &self, const std::string &path) {
2030 2031 2032 2033
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2034
            self.debug_graphviz_path_ = path;
C
chengduo 已提交
2035
          },
2036
          R"DOC((str, optional): debug_graphviz_path indicates the path that
F
flame 已提交
2037
                writing the SSA Graph to file in the form of graphviz.
2038
                It is useful for debugging. Default is empty string, that is, ""
F
flame 已提交
2039 2040 2041 2042 2043 2044

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
C
chengduo 已提交
2045 2046
                        build_strategy.debug_graphviz_path = "./graph"

F
flame 已提交
2047
                    )DOC")
S
sneaxiy 已提交
2048 2049 2050 2051 2052 2053
      .def_property(
          "enable_sequential_execution",
          [](const BuildStrategy &self) {
            return self.enable_sequential_execution_;
          },
          [](BuildStrategy &self, bool b) {
2054 2055 2056 2057
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2058 2059
            self.enable_sequential_execution_ = b;
          },
2060 2061
          R"DOC((bool, optional): If set True, the execution order of ops would
                be the same as what is in the program. Default is False.
F
flame 已提交
2062 2063 2064 2065 2066 2067 2068 2069

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.enable_sequential_execution = True
          )DOC")
S
sneaxiy 已提交
2070 2071 2072 2073 2074 2075
      .def_property(
          "remove_unnecessary_lock",
          [](const BuildStrategy &self) {
            return self.remove_unnecessary_lock_;
          },
          [](BuildStrategy &self, bool b) {
2076 2077 2078 2079
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2080 2081
            self.remove_unnecessary_lock_ = b;
          },
2082 2083
          R"DOC((bool, optional): If set True, some locks in GPU ops would be
                released and ParallelExecutor would run faster. Default is True.
F
flame 已提交
2084 2085 2086 2087 2088 2089 2090 2091

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.remove_unnecessary_lock = True
          )DOC")
2092 2093 2094 2095
      .def_property(
          "num_trainers",
          [](const BuildStrategy &self) { return self.num_trainers_; },
          [](BuildStrategy &self, int num_trainers) {
2096
#ifdef WIN32
2097 2098
            PADDLE_THROW(platform::errors::Unavailable(
                "Windows has NO support to distribute mode."));
2099
#endif
2100 2101
            self.num_trainers_ = num_trainers;
          })
2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113
      .def_property(
          "trainers_endpoints",
          [](const BuildStrategy &self) { return self.trainers_endpoints_; },
          [](BuildStrategy &self,
             const std::vector<std::string> &trainers_endpoints) {
            self.trainers_endpoints_ = trainers_endpoints;
          })
      .def_property("trainer_id",
                    [](const BuildStrategy &self) { return self.trainer_id_; },
                    [](BuildStrategy &self, int trainer_id) {
                      self.trainer_id_ = trainer_id;
                    })
2114 2115 2116 2117 2118 2119
      .def_property(
          "nccl_comm_num",
          [](const BuildStrategy &self) { return self.nccl_comm_num_; },
          [](BuildStrategy &self, int nccl_comm_num) {
            self.nccl_comm_num_ = nccl_comm_num;
          })
2120
      .def_property("use_hierarchical_allreduce",
2121 2122 2123 2124 2125 2126
                    [](const BuildStrategy &self) {
                      return self.use_hierarchical_allreduce_;
                    },
                    [](BuildStrategy &self, bool use) {
                      self.use_hierarchical_allreduce_ = use;
                    })
2127
      .def_property("hierarchical_allreduce_inter_nranks",
2128 2129 2130 2131 2132 2133 2134
                    [](const BuildStrategy &self) {
                      return self.hierarchical_allreduce_inter_nranks_;
                    },
                    [](BuildStrategy &self, int nranks) {
                      self.hierarchical_allreduce_inter_nranks_ = nranks;
                    })

C
chengduo 已提交
2135 2136 2137 2138 2139 2140
      .def_property(
          "fuse_elewise_add_act_ops",
          [](const BuildStrategy &self) {
            return self.fuse_elewise_add_act_ops_;
          },
          [](BuildStrategy &self, bool b) {
2141 2142 2143 2144
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
C
chengduo 已提交
2145 2146
            self.fuse_elewise_add_act_ops_ = b;
          },
2147
          R"DOC((bool, optional): fuse_elewise_add_act_ops indicate whether
F
flame 已提交
2148
                to fuse elementwise_add_op and activation_op,
2149
                it may make the execution faster. Default is False.
F
flame 已提交
2150 2151 2152 2153 2154 2155 2156 2157

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.fuse_elewise_add_act_ops = True
                     )DOC")
Z
Zhen Wang 已提交
2158 2159 2160 2161
      .def_property(
          "fuse_bn_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_act_ops_; },
          [](BuildStrategy &self, bool b) {
2162
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
Z
Zhen Wang 已提交
2163
                              platform::errors::PreconditionNotMet(
2164 2165
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Z
Zhen Wang 已提交
2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178
            self.fuse_bn_act_ops_ = b;
          },
          R"DOC((bool, optional): fuse_bn_act_ops indicate whether
                to fuse batch_norm and activation_op,
                it may make the execution faster. Default is False.

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.fuse_bn_act_ops = True
                     )DOC")
2179 2180 2181 2182
      .def_property(
          "enable_auto_fusion",
          [](const BuildStrategy &self) { return self.enable_auto_fusion_; },
          [](BuildStrategy &self, bool b) {
2183
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
2184
                              platform::errors::PreconditionNotMet(
2185 2186
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200
            self.enable_auto_fusion_ = b;
          },
          R"DOC((bool, optional): Whether to enable fusing subgraph to a
                fusion_group. Now we only support fusing subgraph that composed
                of elementwise-like operators, such as elementwise_add/mul
                without broadcast and activations.

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.enable_auto_fusion = True
                    )DOC")
2201 2202 2203 2204 2205 2206
      .def_property(
          "fuse_relu_depthwise_conv",
          [](const BuildStrategy &self) {
            return self.fuse_relu_depthwise_conv_;
          },
          [](BuildStrategy &self, bool b) {
2207 2208 2209 2210
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2211 2212
            self.fuse_relu_depthwise_conv_ = b;
          },
2213
          R"DOC((bool, optional): fuse_relu_depthwise_conv indicate whether
F
flame 已提交
2214 2215 2216
                to fuse relu and depthwise_conv2d,
                it will save GPU memory and may make the execution faster.
                This options is only available in GPU devices.
2217
                Default is False.
F
flame 已提交
2218 2219 2220 2221 2222 2223 2224 2225

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.fuse_relu_depthwise_conv = True
          )DOC")
C
chengduo 已提交
2226 2227 2228 2229 2230 2231
      .def_property("fuse_broadcast_ops",
                    [](const BuildStrategy &self) {
                      return self.fuse_broadcast_ops_ == true ||
                             self.fuse_broadcast_ops_ == boost::none;
                    },
                    [](BuildStrategy &self, bool b) {
2232 2233 2234 2235
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2236 2237
                      self.fuse_broadcast_ops_ = b;
                    },
2238
                    R"DOC((bool, optional): fuse_broadcast_op indicates whether
2239 2240 2241 2242
                      to fuse the broadcast ops. Note that, in Reduce mode,
                      fusing broadcast ops may make the program faster. Because
                      fusing broadcast OP equals delaying the execution of all
                      broadcast Ops, in this case, all nccl streams are used only
2243 2244 2245 2246 2247 2248 2249 2250 2251
                      for NCCLReduce operations for a period of time. Default False.

                      Examples:
                          .. code-block:: python

                              import paddle.fluid as fluid
                              build_strategy = fluid.BuildStrategy()
                              build_strategy.fuse_broadcast_ops = True
                    )DOC")
C
chengduo 已提交
2252 2253
      .def_property("fuse_all_optimizer_ops",
                    [](const BuildStrategy &self) {
C
chengduo 已提交
2254 2255
                      return self.fuse_all_optimizer_ops_ == true ||
                             self.fuse_all_optimizer_ops_ == boost::none;
C
chengduo 已提交
2256 2257
                    },
                    [](BuildStrategy &self, bool b) {
2258 2259 2260 2261
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2262 2263
                      self.fuse_all_optimizer_ops_ = b;
                    })
Q
qingqing01 已提交
2264 2265 2266 2267
      .def_property(
          "sync_batch_norm",
          [](const BuildStrategy &self) { return self.sync_batch_norm_; },
          [](BuildStrategy &self, bool b) {
2268 2269 2270 2271
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Q
qingqing01 已提交
2272 2273
            self.sync_batch_norm_ = b;
          },
2274
          R"DOC((bool, optional): sync_batch_norm indicates whether to use
Q
qingqing01 已提交
2275 2276 2277
                synchronous batch normalization which synchronizes the mean
                and variance through multi-devices in training phase.
                Current implementation doesn't support FP16 training and CPU.
2278 2279
                And only synchronous on one machine, not all machines. 
                Default is False.
F
flame 已提交
2280 2281 2282 2283 2284 2285 2286 2287

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.sync_batch_norm = True
                )DOC")
D
dzhwinter 已提交
2288 2289
      .def_property(
          "memory_optimize",
2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303
          [](const BuildStrategy &self) -> py::object {
            if (self.memory_optimize_) {
              return py::cast(self.memory_optimize_.get());
            } else {
              return py::cast(nullptr);
            }
          },
          [](BuildStrategy &self, const py::handle &value) {
            auto *py_obj = value.ptr();
            if (py_obj == nullptr || py_obj == Py_None) {
              self.memory_optimize_ = boost::none;
            } else if (PyBool_Check(py_obj)) {
              self.memory_optimize_ = (py_obj == Py_True);
            } else {
2304 2305 2306
              PADDLE_THROW(platform::errors::InvalidArgument(
                  "BuildStrategy.memory_optimize must be set to None, False or "
                  "True"));
2307 2308
            }
          },
2309
          R"DOC((bool, optional): memory opitimize aims to save total memory
2310
                consumption, set to True to enable it.
2311

2312 2313 2314
                Default None. None means framework would choose to use or not use 
                this strategy automatically. Currently, None means that it is 
                enabled when GC is disabled, and disabled when GC is enabled. 
2315
                True means enabling and False means disabling. Default is None.)DOC")
2316 2317 2318
      .def_property(
          "is_distribution",
          [](const BuildStrategy &self) { return self.is_distribution_; },
2319 2320 2321
          [](BuildStrategy &self, bool b) {
#ifdef WIN32
            if (b) {
2322 2323
              PADDLE_THROW(platform::errors::Unavailable(
                  "Windows has NO support to distribute mode."));
2324 2325 2326 2327 2328
            }
#else
            self.is_distribution_ = b;
#endif
          })
Q
can run  
Qiao Longfei 已提交
2329 2330 2331
      .def_property("async_mode",
                    [](const BuildStrategy &self) { return self.async_mode_; },
                    [](BuildStrategy &self, bool b) { self.async_mode_ = b; })
D
dzhwinter 已提交
2332
      .def_property(
D
dzhwinter 已提交
2333 2334 2335
          "enable_inplace",
          [](const BuildStrategy &self) { return self.enable_inplace_; },
          [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; })
C
chengduo 已提交
2336 2337
      .def_property(
          "fuse_all_reduce_ops",
C
chengduo 已提交
2338 2339 2340 2341
          [](const BuildStrategy &self) {
            return self.fuse_all_reduce_ops_ == true ||
                   self.fuse_all_reduce_ops_ == boost::none;
          },
C
chengduo 已提交
2342
          [](BuildStrategy &self, bool b) { self.fuse_all_reduce_ops_ = b; })
2343 2344 2345 2346 2347 2348 2349
      .def_property("enable_backward_optimizer_op_deps",
                    [](const BuildStrategy &self) {
                      return self.enable_backward_optimizer_op_deps_;
                    },
                    [](BuildStrategy &self, bool b) {
                      self.enable_backward_optimizer_op_deps_ = b;
                    })
2350 2351 2352 2353
      .def_property(
          "cache_runtime_context",
          [](const BuildStrategy &self) { return self.cache_runtime_context_; },
          [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
2354 2355 2356 2357 2358 2359 2360 2361 2362
      .def_property(
          "mkldnn_enabled_op_types",
          [](const BuildStrategy &self) {
            return self.mkldnn_enabled_op_types_;
          },
          [](BuildStrategy &self,
             const std::unordered_set<std::string> &mkldnn_enabled_op_types) {
            self.mkldnn_enabled_op_types_ = mkldnn_enabled_op_types;
          })
2363
      .def("_finalize_strategy_and_create_passes",
X
fix  
Xin Pan 已提交
2364
           [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
2365 2366 2367 2368 2369
             return self.CreatePassesFromStrategy(true);
           },
           R"DOC(Allow user to customized passes. Normally model-specific
                optimization passes should be defined in this way. BuildStrategy
                cannot be updated after being finalized.)DOC");
Y
yuyang18 已提交
2370 2371

  pe.def(py::init<const std::vector<platform::Place> &,
Y
Yan Xu 已提交
2372
                  const std::vector<std::string> &, const std::string &,
X
Xin Pan 已提交
2373
                  Scope *, std::vector<Scope *> &, const ExecutionStrategy &,
X
Xin Pan 已提交
2374
                  const BuildStrategy &, ir::Graph *>())
Y
Yu Yang 已提交
2375 2376 2377 2378
      // NOTE: even we return a vec<Scope*>* to Python use reference policy.
      // We still cannot get local_scope from this vector, since the element
      // of vec<Scope*> will be freed by Python GC. We can only return Scope*
      // one by one and mark them as reference.
2379 2380 2381 2382 2383
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
2384 2385 2386
      .def("drop_local_exe_scopes", &ParallelExecutor::DropLocalExeScopes)
      .def("_need_create_local_exe_scopes",
           &ParallelExecutor::NeedCreateLocalExeScope)
Y
Yu Yang 已提交
2387 2388 2389 2390
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
2391 2392
      .def("run",
           [](ParallelExecutor &self,
Z
Zhen Wang 已提交
2393 2394 2395 2396 2397 2398 2399 2400
              const std::vector<std::string> &fetch_tensors,
              bool return_merged) -> py::object {
             paddle::framework::FetchResultType ret;
             {
               pybind11::gil_scoped_release release;
               ret = self.Run(fetch_tensors, return_merged);
             }
             if (return_merged) {
2401 2402
               return py::cast(
                   std::move(boost::get<paddle::framework::FetchList>(ret)));
Z
Zhen Wang 已提交
2403 2404 2405 2406
             } else {
               return py::cast(std::move(
                   boost::get<paddle::framework::FetchUnmergedList>(ret)));
             }
2407 2408
           })
      .def("device_count", &ParallelExecutor::DeviceCount);
Y
Yu Yang 已提交
2409

D
dongdaxiang 已提交
2410
  BindFleetWrapper(&m);
2411
  BindGlooWrapper(&m);
H
hutuxian 已提交
2412
  BindBoxHelper(&m);
H
hutuxian 已提交
2413 2414 2415
#ifdef PADDLE_WITH_BOX_PS
  BindBoxWrapper(&m);
#endif
2416
#ifdef PADDLE_WITH_NCCL
D
dongdaxiang 已提交
2417
  BindNCCLWrapper(&m);
W
wopeizl 已提交
2418
#endif
F
flame 已提交
2419 2420
  BindGraph(&m);
  BindNode(&m);
F
flame 已提交
2421
  BindInferenceApi(&m);
2422
  BindDataset(&m);
2423 2424 2425
#ifdef PADDLE_WITH_DISTRIBUTE
  BindCommunicator(&m);
#endif
L
Luo Tao 已提交
2426
}
2427
}  // namespace pybind
2428
}  // namespace paddle