pybind.cc 101.6 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"
H
hutuxian 已提交
59
#include "paddle/fluid/platform/monitor.h"
Y
Yi Wang 已提交
60 61
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
H
hutuxian 已提交
62
#include "paddle/fluid/pybind/box_helper_py.h"
Y
Yi Wang 已提交
63
#include "paddle/fluid/pybind/const_value.h"
D
dongdaxiang 已提交
64
#include "paddle/fluid/pybind/data_set_py.h"
Y
Yi Wang 已提交
65
#include "paddle/fluid/pybind/exception.h"
D
dongdaxiang 已提交
66
#include "paddle/fluid/pybind/fleet_wrapper_py.h"
67
#include "paddle/fluid/pybind/global_value_getter_setter.h"
68
#include "paddle/fluid/pybind/gloo_wrapper_py.h"
69
#include "paddle/fluid/pybind/imperative.h"
F
flame 已提交
70
#include "paddle/fluid/pybind/inference_api.h"
F
flame 已提交
71
#include "paddle/fluid/pybind/ir.h"
72
#include "paddle/fluid/pybind/pybind_boost_headers.h"
73

74
#ifdef PADDLE_WITH_NCCL
D
dongdaxiang 已提交
75
#include "paddle/fluid/pybind/nccl_wrapper_py.h"
W
wopeizl 已提交
76
#endif
77
#include "paddle/fluid/framework/data_type.h"
78 79
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
S
sneaxiy 已提交
80
#include "paddle/fluid/pybind/reader_py.h"
Y
Yi Wang 已提交
81
#include "paddle/fluid/pybind/tensor_py.h"
82
#include "paddle/fluid/string/to_string.h"
D
Dong Zhihong 已提交
83
#ifdef PADDLE_WITH_CUDA
84
#ifdef PADDLE_WITH_NCCL
Y
Yi Wang 已提交
85
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
P
peizhilin 已提交
86
#endif
Y
Yi Wang 已提交
87 88
#include "paddle/fluid/platform/cuda_profiler.h"
#include "paddle/fluid/platform/gpu_info.h"
D
Dong Zhihong 已提交
89 90
#endif

91 92 93 94
#ifdef PADDLE_WITH_DISTRIBUTE
#include "paddle/fluid/pybind/communicator_py.h"
#endif

Y
Yanghello 已提交
95 96 97 98
#ifdef PADDLE_WITH_CRYPTO
#include "paddle/fluid/pybind/crypto.h"
#endif

M
minqiyang 已提交
99 100
#include "pybind11/stl.h"

101
DECLARE_bool(use_mkldnn);
102

Q
Qiao Longfei 已提交
103 104
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
105 106 107
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchUnmergedList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchType);
Q
Qiao Longfei 已提交
108

109
namespace paddle {
110
namespace pybind {
111
bool IsCompiledWithCUDA() {
112
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
113 114 115 116 117 118
  return false;
#else
  return true;
#endif
}

119 120 121 122 123 124 125 126
bool IsCompiledWithMKLDNN() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return true;
#endif
}

127
bool IsCompiledWithBrpc() {
128
#ifndef PADDLE_WITH_DISTRIBUTE
129 130
  return false;
#endif
131 132 133 134 135 136

#ifdef PADDLE_WITH_GRPC
  return false;
#endif

  return true;
137 138
}

Y
update  
Yancey1989 已提交
139
bool IsCompiledWithDIST() {
Y
Yancey1989 已提交
140
#ifdef PADDLE_WITH_DISTRIBUTE
Y
update  
Yancey1989 已提交
141 142 143 144 145 146
  return true;
#else
  return false;
#endif
}

S
sneaxiy 已提交
147 148 149 150 151 152 153 154 155 156
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
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
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 &) {
179 180 181
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Python object is not type of %s, the real type is %s",
        typeid(T).name(), obj->ob_type->tp_name));
H
hong 已提交
182 183 184 185 186 187 188 189 190 191 192 193 194
  }
}

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) {
195 196
      PADDLE_THROW(platform::errors::InvalidArgument(
          "The parameter [%s] to save is None", para.first));
H
hong 已提交
197 198
    }
    vec_res.emplace_back(
199
        PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_obj));
H
hong 已提交
200 201 202 203 204 205 206 207 208 209 210 211
  }

  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) {
212 213
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameter list to save is None"));
H
hong 已提交
214 215 216 217 218 219 220 221 222 223 224 225
  }

  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);
226 227 228
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to save is None"));
H
hong 已提交
229 230 231 232
      vec_res.emplace_back(PyObjectCast<std::string>(py_name));
      Py_DECREF(py_name);
    }
  } else {
233 234
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to save is not a list"));
H
hong 已提交
235 236 237 238
  }
  return vec_res;
}

239 240 241 242 243 244 245 246
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) {
247 248
    PADDLE_THROW(
        platform::errors::InvalidArgument("The parameter list to set is None"));
249 250 251 252 253 254 255 256 257 258 259 260 261
  }

  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);
262 263 264
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to set is None"));
265 266 267 268 269
      auto para_name = PyObjectCast<std::string>(py_name);
      Py_DECREF(py_name);

      auto var = scope.FindVar(para_name);
      if (var == nullptr) {
270 271 272 273 274
        PADDLE_ENFORCE_NOT_NULL(exe,
                                platform::errors::InvalidArgument(
                                    "Parameter not Initialized, "
                                    "Please set argument [executor] not None "
                                    "or run startup program first"));
275 276
        PyObject *py_var_desc =
            PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kVarDescField);
277 278 279
        PADDLE_ENFORCE_NOT_NULL(
            py_var_desc, platform::errors::InvalidArgument(
                             "The var_desc of parameter to set is None"));
280 281 282 283 284 285 286 287 288
        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 {
289 290
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to set is not a list"));
291 292 293 294 295
  }

  return;
}

296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
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, ',')));
}

320 321 322 323 324 325
#ifdef PADDLE_WITH_AVX
PYBIND11_MODULE(core_avx, m) {
#else
PYBIND11_MODULE(core_noavx, m) {
#endif

Y
Yu Yang 已提交
326 327 328
  // Not used, just make sure cpu_info.cc is linked.
  paddle::platform::CpuTotalPhysicalMemory();

Y
Refine  
Yu Yang 已提交
329
  paddle::memory::allocation::UseAllocatorStrategyGFlag();
S
sneaxiy 已提交
330

331 332
  AssertStaticGraphAndDygraphGradMakerNoDiff();

333
  m.doc() = "C++ core of PaddlePaddle";
334

335 336 337 338
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

339
  BindException(&m);
Y
Yu Yang 已提交
340

341 342
  m.def("set_num_threads", &platform::SetNumThreads);

6
633WHU 已提交
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
  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
361 362 363 364 365 366 367 368 369
  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,
370
           const Scope &scope, const Executor *executor) {
H
hong 已提交
371
          std::vector<std::string> vec_name_list = GetNameList(vec_var_list);
372
          CreateVariableIfNotExit(vec_var_list, scope, executor);
H
hong 已提交
373 374 375
          LoadStaticNameListFromDisk(str_file_name, vec_name_list, scope);
        });

376 377 378 379 380 381
  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
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400
  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
401

402 403 404 405 406 407
  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
408
  m.def(
S
sneaxiy 已提交
409
      "_append_python_callable_object_and_return_id",
S
sneaxiy 已提交
410 411 412 413
      [](py::object py_obj) -> size_t {
        return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
      });

S
sneaxiy 已提交
414 415 416
  m.def("_get_use_default_grad_op_desc_maker_ops",
        [] { return OpInfoMap::Instance().GetUseDefaultGradOpDescMakerOps(); });

417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
  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
433 434 435
  // 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
436
  m.def("_set_eager_deletion_mode", &paddle::framework::SetEagerDeletionMode);
S
sneaxiy 已提交
437

438
  m.def("_set_fuse_parameter_group_size",
439
        &paddle::framework::ir::SetFuseParameterGroupsSize);
440
  m.def("_set_fuse_parameter_memory_size",
441
        &paddle::framework::ir::SetFuseParameterMemorySize);
442

S
sneaxiy 已提交
443 444 445
  m.add_object("_cleanup",
               py::capsule([]() { ScopePool::Instance().Clear(); }));

446 447
  m.def("_set_paddle_lib_path", &paddle::platform::dynload::SetPaddleLibPath);

448
  BindImperative(&m);
449

450
  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
451
      .def("__array__", [](Tensor &self) { return TensorToPyArray(self); })
S
sneaxiy 已提交
452 453
      .def("_is_initialized",
           [](const Tensor &self) { return self.IsInitialized(); })
Y
yuyang18 已提交
454
      .def("_get_dims",
455
           [](const Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
456
      .def("_set_dims",
Q
qijun 已提交
457
           [](Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
458
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
459
           })
Y
yuyang18 已提交
460
      .def("_set_layout",
D
dzhwinter 已提交
461 462 463
           [](Tensor &self, const std::string &layout) {
             self.set_layout(StringToDataLayout(layout));
           })
Y
yuyang18 已提交
464
      .def("_alloc_float",
D
dzhwinter 已提交
465
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
466
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
467
           })
Y
yuyang18 已提交
468
      .def("_alloc_float",
Y
Yu Yang 已提交
469
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
470
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
471
           })
472 473 474 475
      .def("_alloc_double",
           [](Tensor &self, paddle::platform::CPUPlace &place) {
             self.mutable_data<double>(place);
           })
Y
yuyang18 已提交
476
      .def("_alloc_int",
Y
Yu Yang 已提交
477
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
478
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
479
           })
Y
yuyang18 已提交
480
      .def("_alloc_int",
D
dzhwinter 已提交
481
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
482
             self.mutable_data<int>(place);
Q
qijun 已提交
483
           })
Y
yuyang18 已提交
484
      .def("_alloc_int",
C
chengduoZH 已提交
485 486 487
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
488
      .def("_alloc_float",
C
chengduoZH 已提交
489 490 491
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<float>(place);
           })
492 493 494 495 496 497 498 499 500 501 502 503 504 505 506
      .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
507
      .def("_clear", &Tensor::clear)
508
      .def("set", SetTensorFromPyArray<paddle::platform::CPUPlace>,
509
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
510
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPlace>,
511
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
512
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPinnedPlace>,
513 514
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false,
           R"DOC(
L
Leo Chen 已提交
515 516 517 518 519 520
        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.
521 522
          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
523 524 525 526 527 528 529 530 531 532 533 534 535

        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")
536

L
Leo Chen 已提交
537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
      .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
554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575
      .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
576 577 578 579
      .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
580
      .def("_place", [](Tensor &self) { return self.place(); })
W
wopeizl 已提交
581
      .def("_dtype", [](Tensor &self) { return self.type(); })
582
      .def("_share_data_with", &Tensor::ShareDataWith)
583 584 585 586 587 588
      .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
589

L
Leo Chen 已提交
590
  // TODO(cql): add reference: en_user_guide_lod_tensor
X
Xin Pan 已提交
591
  py::class_<LoDTensor, Tensor>(m, "LoDTensor", R"DOC(
L
Leo Chen 已提交
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665
    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
666 667 668 669 670 671 672

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid

          t = fluid.LoDTensor()
X
Xin Pan 已提交
673 674

        )DOC")
675
      .def("__array__", [](Tensor &self) { return TensorToPyArray(self); })
676 677 678 679 680 681 682 683 684
      .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
685 686
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, -1), true,
687 688 689 690
                 platform::errors::InvalidArgument(
                     "The provided recursive_sequence_lengths info is invalid, "
                     "the LoD converted by recursive_sequence_lengths is %s",
                     new_lod));
691 692
             new (&instance) LoDTensor(new_offset_lod);
           })
Y
Yu Yang 已提交
693
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
G
gongweibao 已提交
694
      // We implement offset based LOD in C++ while we use length based with
H
hong 已提交
695 696
      // Python API. So we changed set_lod to set_recursive_sequence_lengths
      // to
G
gongweibao 已提交
697 698 699
      // avoid misuse.
      // The discussion is here:
      // https://github.com/PaddlePaddle/Paddle/issues/10855
D
dangqingqing 已提交
700
      .def("set_lod",
701
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
702
             // the input lod is offset-based level-of-detail info
Y
Yu Yang 已提交
703
             LoD new_lod;
704 705
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
C
chengduo 已提交
706 707
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_lod, vectorize(self.dims()).front()), true,
708 709
                 platform::errors::InvalidArgument(
                     "The provided LoD is invalid, the LoD is %s", new_lod));
710
             self.set_lod(new_lod);
S
sneaxiy 已提交
711 712 713 714 715
           },
           py::arg("lod"), R"DOC(
           Set LoD of the LoDTensor.

           Args:
L
Leo Chen 已提交
716 717 718 719
               lod (list[list[int]]): The lod to set.

           Returns:
                None.
Z
Zeng Jinle 已提交
720 721 722 723 724 725 726 727 728 729

           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
730
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
731
           )DOC")
732 733 734 735 736 737 738 739 740 741 742
      .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
743 744
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, vectorize(self.dims()).front()), true,
745 746 747 748 749
                 platform::errors::InvalidArgument(
                     "The provided recursive_sequence_lengths info is invalid, "
                     "the LoD converted by recursive_sequence_lengths is "
                     "%s",
                     new_lod));
750
             self.set_lod(new_offset_lod);
S
sneaxiy 已提交
751 752
           },
           py::arg("recursive_sequence_lengths"), R"DOC(
L
Leo Chen 已提交
753
           Set LoD of the LoDTensor according to recursive sequence lengths.
S
sneaxiy 已提交
754

L
Leo Chen 已提交
755
           For example, if recursive_sequence_lengths=[[2, 3]], which means
756
           there are two sequences with length 2 and 3 respectively, the
L
Leo Chen 已提交
757
           corresponding lod would be [[0, 2, 2+3]], i.e., [[0, 2, 5]].
S
sneaxiy 已提交
758 759

           Args:
L
Leo Chen 已提交
760 761 762 763
                recursive_sequence_lengths (list[list[int]]): The recursive sequence lengths.
           
           Returns:
                None.
Z
Zeng Jinle 已提交
764 765 766 767 768 769 770 771 772 773

           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
774 775
                 print(t.recursive_sequence_length())  # [[2, 3]]
                 print(t.lod())  # [[0, 2, 5]]
S
sneaxiy 已提交
776
           )DOC")
777 778 779 780 781 782 783 784
      .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
785 786 787 788 789
           },
           R"DOC(
           Return the LoD of the LoDTensor.

           Returns:
L
Leo Chen 已提交
790 791
               list[list[int]]: The lod of the LoDTensor.
           
Z
Zeng Jinle 已提交
792 793 794 795 796 797 798 799 800 801
           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
802
           )DOC")
G
gongweibao 已提交
803
      // Set above comments of set_lod.
804 805 806 807 808 809 810 811
      .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
812 813
           },
           R"DOC(
L
Leo Chen 已提交
814 815
           Return the recursive sequence lengths corresponding to of the LodD 
           of the LoDTensor.
S
sneaxiy 已提交
816 817

           Returns:
L
Leo Chen 已提交
818
                list[list[int]]: The recursive sequence lengths.
Z
Zeng Jinle 已提交
819 820 821 822 823 824 825 826 827 828 829

           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
830 831 832 833 834 835 836 837
           )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
838
           Check whether the LoD of the LoDTensor is valid.
S
sneaxiy 已提交
839 840

           Returns:
L
Leo Chen 已提交
841
               bool: Whether the LoD is valid.
Z
Zeng Jinle 已提交
842 843 844 845 846 847 848 849 850 851 852

           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
853 854 855 856 857 858 859
           )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).
860
           )DOC")
861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878
      .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;
879
#ifdef _WIN32
880
      });
881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930
#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
931

Q
qijun 已提交
932 933 934 935 936 937 938 939 940 941 942
  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)
943 944
      .def("numel",
           [](SelectedRows &self) -> int64_t { return self.value().numel(); })
Q
qijun 已提交
945 946
      .def("set_height", &SelectedRows::set_height)
      .def("height", &SelectedRows::height)
Q
qijun 已提交
947 948 949 950 951 952 953 954 955
      .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
           })
956
      .def("sync_index", [](SelectedRows &instance) { instance.SyncIndex(); })
957
      .def("rows", [](SelectedRows &self) {
958 959 960 961 962
        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;
963
      });
Q
qijun 已提交
964

965
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
966 967 968

All parameter, weight, gradient are variables in Paddle.
)DOC")
S
sneaxiy 已提交
969
      .def(py::init<>())
970
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
971
      .def("set_int",
972 973
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
974 975 976 977 978 979 980
      .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
981
      .def("get_tensor",
982 983
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
984 985
           },
           py::return_value_policy::reference)
986 987 988 989
      .def("get_bytes",
           [](Variable &self) {
             return py::bytes(*self.GetMutable<std::string>());
           })
Y
Yu Yang 已提交
990 991 992
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
993 994 995 996 997
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
998 999 1000
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
1001 1002 1003
      .def("get_fetch_list",
           [](Variable &self) { return self.GetMutable<FetchList>(); },
           py::return_value_policy::reference)
1004
#if (defined(PADDLE_WITH_NCCL))
D
Dong Zhihong 已提交
1005 1006 1007 1008 1009
      .def("get_communicator",
           [](Variable &self) -> platform::Communicator * {
             return self.GetMutable<platform::Communicator>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
1010
#endif
Y
Refine  
Yu Yang 已提交
1011 1012
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
1013 1014 1015 1016
             PADDLE_ENFORCE_EQ(
                 self.IsType<framework::ReaderHolder>(), true,
                 platform::errors::InvalidArgument(
                     "The variable is not type of ReaderHolder."));
Y
Refine  
Yu Yang 已提交
1017 1018
             return self.GetMutable<framework::ReaderHolder>();
           },
1019 1020 1021 1022 1023
           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);
      });
1024

S
sneaxiy 已提交
1025
  BindReader(&m);
Y
Refine  
Yu Yang 已提交
1026

S
sneaxiy 已提交
1027
  py::class_<Scope>(m, "_Scope", R"DOC(
Q
Qiao Longfei 已提交
1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
    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

1041
          import paddle.fluid as fluid
Q
Qiao Longfei 已提交
1042 1043 1044 1045 1046 1047
          # 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
1048 1049
      .def("_remove_from_pool",
           [](Scope &self) { ScopePool::Instance().Remove(&self); })
D
dongzhihong 已提交
1050
      .def("var",
1051
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
1052
             return self.Var(name);
Y
Yu Yang 已提交
1053
           },
S
sneaxiy 已提交
1054 1055
           py::arg("name"),
           R"DOC(
1056
           Find or create variable named :code:`name` in the current scope.
S
sneaxiy 已提交
1057

1058
           If the variable named :code:`name` does not exist in the
S
sneaxiy 已提交
1059
           current scope, the variable would be created. Otherwise,
1060
           return the existing variable.
S
sneaxiy 已提交
1061 1062

           Args:
1063 1064
               name (str): the variable name.

S
sneaxiy 已提交
1065
           Returns:
1066
               out (core.Variable): the found or created variable.
S
sneaxiy 已提交
1067 1068 1069 1070
           )DOC",
           py::return_value_policy::reference)
      .def("find_var", &Scope::FindVar, py::arg("name"),
           R"DOC(
1071
           Find variable named :code:`name` in the current scope or
S
sneaxiy 已提交
1072
           its parent scope. Return None if not found.
1073

S
sneaxiy 已提交
1074 1075
           Args:
               name (str): the variable name.
1076

S
sneaxiy 已提交
1077
           Returns:
1078
               out (core.Variable|None): the found variable or None.
S
sneaxiy 已提交
1079
           )DOC",
1080
           py::return_value_policy::reference)
1081
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
S
sneaxiy 已提交
1082 1083 1084 1085 1086 1087
           R"DOC(
           Create a new sub-scope of the current scope.

           Returns:
               out (core._Scope): the created sub-scope.
           )DOC",
1088
           py::return_value_policy::reference)
S
sneaxiy 已提交
1089 1090 1091
      .def("drop_kids", &Scope::DropKids,
           R"DOC(
           Delete all sub-scopes of the current scope.
S
sneaxiy 已提交
1092 1093
           )DOC")
      .def("_kids", &Scope::kids);
1094

S
sneaxiy 已提交
1095 1096 1097 1098 1099 1100
  m.def("Scope",
        []() -> Scope * {
          auto *s = new Scope();
          ScopePool::Instance().Insert(std::unique_ptr<Scope>(s));
          return s;
        },
S
sneaxiy 已提交
1101 1102
        R"DOC(
        Create a new scope.
1103

S
sneaxiy 已提交
1104 1105 1106
        Returns:
            out (core._Scope): the created scope.
        )DOC",
S
sneaxiy 已提交
1107 1108
        py::return_value_policy::reference);

Y
Yu Yang 已提交
1109 1110
  //! @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
1111 1112
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
1113 1114 1115 1116
    for (auto &iter : OpInfoMap::Instance().map()) {
      auto &info = iter.second;
      if (info.HasOpProtoAndChecker()) {
        std::string str;
C
chengduo 已提交
1117 1118
        PADDLE_ENFORCE_EQ(
            info.Proto().SerializeToString(&str), true,
1119 1120
            platform::errors::Fatal(
                "Serialize OpProto Error. This could be a bug of Paddle."));
1121 1122 1123
        ret_values.emplace_back(str);
      }
    }
Y
Yu Yang 已提交
1124 1125
    return ret_values;
  });
1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
  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;
        });
1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154
  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);
      });
1155 1156 1157
  m.def("has_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasGradOpMaker();
  });
1158 1159 1160 1161 1162
  m.def("has_non_empty_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance()
        .Get(op_type)
        .HasNonEmptyGradOpMaker();
  });
1163 1164 1165
  m.def("has_infer_inplace", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasInferInplace();
  });
1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179
  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
1180
  m.def("prune", [](const ProgramDesc &origin,
1181
                    const std::set<std::string> &feeded_var_names,
1182
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
1183
    ProgramDesc prog_with_targets(origin);
1184

1185
    for (const auto &t : targets) {
1186
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
1187
    }
1188
    proto::ProgramDesc pruned_desc;
1189 1190 1191 1192
    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);
1193
  });
1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210
  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");
1211 1212 1213 1214
  m.def("empty_var_name",
        []() { return std::string(framework::kEmptyVarName); });
  m.def("grad_var_suffix",
        []() { return std::string(framework::kGradVarSuffix); });
1215 1216
  m.def("loaded_var_suffix",
        []() { return std::string(framework::kLoadedVarSuffix); });
1217 1218 1219
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
1220 1221
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
Q
qijun 已提交
1222
  // clang-format off
Y
Yu Yang 已提交
1223
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
1224 1225
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
1226
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
1227 1228 1229
                    return new paddle::platform::CPUDeviceContext();
                  })
      .def_static("create",
D
dzhwinter 已提交
1230
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
1231
                      -> paddle::platform::DeviceContext* {
1232
#ifndef PADDLE_WITH_CUDA
1233 1234 1235 1236
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use CUDAPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
Q
qijun 已提交
1237
#else
Q
qijun 已提交
1238
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
1239
#endif
C
chengduoZH 已提交
1240 1241 1242 1243 1244
                  })
          .def_static("create",
                [](paddle::platform::CUDAPinnedPlace& place)
                        -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_CUDA
1245 1246 1247 1248
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use CUDAPinnedPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
C
chengduoZH 已提交
1249 1250 1251 1252
#else
                  return new paddle::platform::CUDAPinnedDeviceContext(place);
#endif
                });;
D
Dong Zhihong 已提交
1253
// clang-format on
1254
#if defined(PADDLE_WITH_NCCL)
D
Dong Zhihong 已提交
1255 1256
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
1257
  py::class_<platform::CUDAPlace>(m, "CUDAPlace", R"DOC(
1258 1259 1260 1261 1262 1263 1264 1265
    **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.
1266
    The memory of CUDAPlace with different dev_id is not accessible.
1267 1268 1269 1270 1271 1272 1273 1274
    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
1275 1276 1277 1278

    Examples:
        .. code-block:: python

1279
          import paddle.fluid as fluid
L
lujun 已提交
1280 1281
          gpu_place = fluid.CUDAPlace(0)

1282
        )DOC")
S
sneaxiy 已提交
1283 1284 1285
      .def("__init__",
           [](platform::CUDAPlace &self, int dev_id) {
#ifdef PADDLE_WITH_CUDA
1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
             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
1310 1311
             new (&self) platform::CUDAPlace(dev_id);
#else
1312 1313 1314 1315 1316 1317 1318 1319 1320
             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
1321 1322
#endif
           })
S
sneaxiy 已提交
1323 1324 1325 1326 1327 1328
      .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
1329
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
1330

1331
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace", R"DOC(
1332 1333
    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
1334 1335 1336 1337

    Examples:
        .. code-block:: python

1338
          import paddle.fluid as fluid
1339
          cpu_place = fluid.CPUPlace()
L
lujun 已提交
1340

1341
        )DOC")
1342
      .def(py::init<>())
S
sneaxiy 已提交
1343 1344 1345 1346 1347 1348
      .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>)
1349
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
1350

1351
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace", R"DOC(
1352 1353 1354 1355 1356 1357
    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
1358 1359 1360 1361

    Examples:
        .. code-block:: python

1362
          import paddle.fluid as fluid
L
lujun 已提交
1363 1364
          place = fluid.CUDAPinnedPlace()

1365
        )DOC")
S
sneaxiy 已提交
1366
      .def("__init__",
S
sneaxiy 已提交
1367
           [](platform::CUDAPinnedPlace &self) {
S
sneaxiy 已提交
1368
#ifndef PADDLE_WITH_CUDA
1369 1370 1371
             PADDLE_THROW(platform::errors::PermissionDenied(
                 "Cannot use CUDAPinnedPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
S
sneaxiy 已提交
1372
#endif
S
sneaxiy 已提交
1373
             new (&self) platform::CUDAPinnedPlace();
S
sneaxiy 已提交
1374
           })
S
sneaxiy 已提交
1375 1376 1377 1378 1379 1380 1381 1382
      .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
1383 1384
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
1385 1386
  py::class_<platform::Place>(m, "Place")
      .def(py::init<>())
S
sneaxiy 已提交
1387 1388 1389 1390 1391
      .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
1392 1393
      .def("is_gpu_place",
           [](platform::Place &self) { return platform::is_gpu_place(self); })
S
sneaxiy 已提交
1394 1395 1396 1397 1398 1399
      .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
1400 1401
      .def("gpu_device_id",
           [](platform::Place &self) {
1402
             return BOOST_GET_CONST(platform::CUDAPlace, self).device;
X
xuezhong 已提交
1403
           })
S
sneaxiy 已提交
1404 1405
      .def("set_place", [](platform::Place &self,
                           const platform::Place &other) { self = other; })
Y
Yu Yang 已提交
1406 1407 1408 1409 1410
      .def("set_place",
           [](platform::Place &self, const platform::CPUPlace &cpu_place) {
             self = cpu_place;
           })
      .def("set_place",
D
dzhwinter 已提交
1411
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
1412
             self = gpu_place;
C
chengduoZH 已提交
1413 1414
           })
      .def("set_place", [](platform::Place &self,
C
chengduoZH 已提交
1415 1416
                           const platform::CUDAPinnedPlace &cuda_pinned_place) {
        self = cuda_pinned_place;
C
chengduoZH 已提交
1417
      });
Y
Yu Yang 已提交
1418

Y
Yu Yang 已提交
1419
  py::class_<OperatorBase>(m, "Operator")
C
chengduo 已提交
1420 1421 1422 1423 1424
      .def_static(
          "create",
          [](py::bytes protobin) {
            proto::OpDesc desc;
            PADDLE_ENFORCE_EQ(desc.ParsePartialFromString(protobin), true,
1425 1426 1427 1428 1429 1430 1431
                              platform::errors::InvalidArgument(
                                  "Cannot parse user input to OpDesc"));
            PADDLE_ENFORCE_EQ(
                desc.IsInitialized(), true,
                platform::errors::InvalidArgument(
                    "The provided OpDesc is not initialized, the reason is: %s",
                    desc.InitializationErrorString()));
C
chengduo 已提交
1432 1433
            return OpRegistry::CreateOp(desc);
          })
1434
      .def("run",
1435
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1436 1437 1438
              const platform::CPUPlace &place) { self.Run(scope, place); })
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1439
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
1440 1441 1442 1443 1444
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
1445 1446 1447 1448 1449 1450 1451
      .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
1452 1453
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
1454
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
1455
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
1456 1457 1458 1459
      .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
1460

1461 1462 1463
  py::class_<framework::ExecutorPrepareContext>(m, "ExecutorPrepareContext")
      .def(py::init<const ProgramDesc &, size_t>());

1464 1465 1466 1467 1468 1469 1470 1471 1472
  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
1473
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
1474
      .def(py::init<const platform::Place &>())
Y
Yancey1989 已提交
1475
      .def("close", &Executor::Close)
1476 1477
      .def("run_from_dataset", &Executor::RunFromDataset,
           py::call_guard<py::gil_scoped_release>())
D
Dong Daxiang 已提交
1478 1479
      .def("release_trainer", &Executor::ReleaseTrainer,
           py::call_guard<py::gil_scoped_release>())
1480 1481 1482 1483
      .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
1484
             pybind11::gil_scoped_release release;
1485 1486 1487 1488 1489 1490 1491
             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);
           })
1492 1493 1494
      .def("run_prepared_ctx",
           [](Executor &self, ExecutorPrepareContext *ctx, Scope *scope,
              std::map<std::string, const LoDTensor *> *feed_targets,
1495
              std::map<std::string, FetchType *> *fetch_targets,
1496 1497 1498 1499 1500 1501 1502 1503
              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);
           })
1504
      .def("run_prepared_ctx",
G
guru4elephant 已提交
1505 1506 1507 1508 1509 1510 1511
           [](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);
           })
1512 1513 1514 1515 1516 1517 1518 1519 1520 1521
      .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
1522
      .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope,
S
sneaxiy 已提交
1523 1524
                     int block_id, bool create_local_scope, bool create_vars,
                     const std::vector<std::string> &fetch_vars) {
S
sneaxiy 已提交
1525
        pybind11::gil_scoped_release release;
S
sneaxiy 已提交
1526 1527
        self.Run(prog, scope, block_id, create_local_scope, create_vars,
                 fetch_vars);
S
sneaxiy 已提交
1528
      });
S
sneaxiy 已提交
1529

D
dzhwinter 已提交
1530
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
1531
  m.def("init_glog", framework::InitGLOG);
1532
  m.def("load_op_library", framework::LoadOpLib);
X
Xin Pan 已提交
1533 1534
  m.def("init_devices",
        [](bool init_p2p) { framework::InitDevices(init_p2p); });
1535

1536
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
1537
  m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
1538
  m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
Y
update  
Yancey1989 已提交
1539
  m.def("is_compiled_with_dist", IsCompiledWithDIST);
H
hutuxian 已提交
1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558

  m.def("get_float_stats", []() {
    std::vector<paddle::platform::ExportedStatValue<float>> float_stats;
    paddle::platform::StatRegistry<float>::Instance().publish(float_stats);
    std::unordered_map<std::string, float> stats_map;
    for (const auto &stat : float_stats) {
      stats_map[stat.key] = stat.value;
    }
    return stats_map;
  });
  m.def("get_int_stats", []() {
    std::vector<paddle::platform::ExportedStatValue<int64_t>> int_stats;
    paddle::platform::StatRegistry<int64_t>::Instance().publish(int_stats);
    std::unordered_map<std::string, int64_t> stats_map;
    for (const auto &stat : int_stats) {
      stats_map[stat.key] = stat.value;
    }
    return stats_map;
  });
H
hutuxian 已提交
1559 1560 1561 1562 1563 1564 1565
  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);
G
gongweibao 已提交
1566 1567 1568 1569 1570 1571 1572 1573 1574
  m.def("shell_execute_cmd",
        [](const std::string &cmd, int time_out = 0, int sleep_inter = 0,
           bool redirect_stderr = false) -> std::vector<std::string> {
          return paddle::framework::shell_execute_cmd(
              cmd, time_out, sleep_inter, redirect_stderr);
        },
        py::arg("cmd"), py::arg("time_out") = 0, py::arg("sleep_inter") = 0,
        py::arg("redirect_stderr") = false);

1575 1576 1577 1578 1579 1580
#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
1581

1582
  m.def("set_feed_variable", framework::SetFeedVariable);
1583 1584 1585 1586 1587
  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)) {
1588
            return py::cast(BOOST_GET(LoDTensor, var));
1589
          } else {
1590
            return py::cast(BOOST_GET(LoDTensorArray, var));
1591 1592
          }
        });
1593
  m.def("get_variable_tensor", framework::GetVariableTensor);
Q
qijun 已提交
1594

X
Xin Pan 已提交
1595 1596
  m.def("_is_program_version_supported", IsProgramVersionSupported);

1597 1598 1599 1600 1601
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
1602
  BindGlobalValueGetterSetter(&m);
Y
Yu Yang 已提交
1603

Y
Yu Yang 已提交
1604 1605 1606 1607 1608 1609 1610 1611 1612
  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
1613
  py::class_<LoDTensorArray>(m, "LoDTensorArray", R"DOC(
1614
    LoDTensorArray is array of LoDTensor, it supports operator[], len() and for-loop iteration.
Z
Zeng Jinle 已提交
1615 1616 1617

    Examples:
        .. code-block:: python
1618

Z
Zeng Jinle 已提交
1619 1620 1621 1622
          import paddle.fluid as fluid

          arr = fluid.LoDTensorArray()
)DOC")
S
sneaxiy 已提交
1623 1624
      .def("__init__",
           [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); })
Y
Yu Yang 已提交
1625 1626 1627 1628 1629 1630
      .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) {
1631 1632 1633 1634
             PADDLE_ENFORCE_LT(i, self.size(),
                               platform::errors::InvalidArgument(
                                   "The index to set is larger than the size "
                                   "of LoDTensorArray."));
Y
Yu Yang 已提交
1635 1636 1637
             self[i].ShareDataWith(t);
             self[i].set_lod(t.lod());
           })
S
sneaxiy 已提交
1638 1639 1640 1641 1642 1643
      .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
1644 1645
           py::arg("tensor"), R"DOC(
             Append a LoDensor to LoDTensorArray.
1646 1647 1648 1649 1650 1651
              
             Args:
                   tensor (LoDTensor): The LoDTensor to be appended.

             Returns:
                   None.
Z
Zeng Jinle 已提交
1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662

             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)
1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673
           )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
1674

1675 1676 1677 1678 1679 1680 1681 1682
  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])) {
1683
                 auto &data = BOOST_GET(LoDTensor, self[i]);
1684 1685
                 res[i] = py::cast(std::move(data));
               } else {
1686
                 auto &data = BOOST_GET(LoDTensorArray, self[i]);
1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701
                 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();
1702
             auto &lod_tensor = BOOST_GET(LoDTensor, self.back());
1703 1704 1705 1706 1707 1708 1709 1710
             lod_tensor.ShareDataWith(t);
             lod_tensor.set_lod(t.lod());
           },
           py::arg("var"))

      .def("append",
           [](FetchList &self, const LoDTensorArray &t) {
             self.emplace_back();
1711
             auto &lod_tensor_array = BOOST_GET(LoDTensorArray, self.back());
1712 1713 1714 1715 1716 1717 1718 1719 1720
             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
1721 1722
        )DOC")
      .def("_move_to_list",
1723
           [](FetchUnmergedList &self) -> py::list {
Z
Zhen Wang 已提交
1724 1725 1726 1727
             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) {
1728
                 if (data_is_lod_tensor(self[i][j])) {
1729
                   auto &var = BOOST_GET(LoDTensor, self[i][j]);
1730 1731
                   tmp[j] = py::cast(std::move(var));
                 } else {
1732
                   auto &var = BOOST_GET(LoDTensorArray, self[i][j]);
1733 1734 1735 1736 1737 1738
                   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
1739 1740 1741 1742 1743 1744 1745 1746 1747
               }
               res[i] = std::move(tmp);
               self[i].clear();
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership);

Y
Yu Yang 已提交
1748
  m.def("op_support_gpu", OpSupportGPU);
D
Dong Zhihong 已提交
1749
#ifdef PADDLE_WITH_CUDA
D
Dong Zhihong 已提交
1750
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
1751

P
peizhilin 已提交
1752
#ifndef _WIN32
D
dangqingqing 已提交
1753 1754 1755
  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
D
Dong Zhihong 已提交
1756
#endif
P
peizhilin 已提交
1757
#endif
Y
Yu Yang 已提交
1758

1759 1760 1761 1762 1763 1764
  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();

1765 1766 1767 1768
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
1769
      .value("kAll", platform::ProfilerState::kAll)
1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780
      .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();

1781
  m.def("set_tracer_option", platform::SetTracerOption);
1782 1783
  m.def("enable_profiler", platform::EnableProfiler);
  m.def("disable_profiler", platform::DisableProfiler);
X
Xin Pan 已提交
1784
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
1785
  m.def("reset_profiler", platform::ResetProfiler);
1786
  m.def("get_pass", [](const std::string &pass_type) {
W
WangZhen 已提交
1787 1788 1789
    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
1790

1791 1792
  m.def("size_of_dtype", framework::SizeOfType);

1793 1794 1795
  using VarQuantScale =
      std::unordered_map<std::string, std::pair<bool, LoDTensor>>;

1796 1797
  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
W
WangZhen 已提交
1798
      .def("has", &ir::Pass::Has)
1799 1800 1801
      .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
1802
           })
1803
      .def(
1804
          "set",
1805 1806 1807
          [](ir::Pass &self, const std::string &name, const std::string &attr) {
            self.Set<std::string>(name, new std::string(attr));
          })
1808 1809
      .def("set", [](ir::Pass &self, const std::string &name,
                     bool val) { self.Set<bool>(name, new bool(val)); })
1810 1811
      .def("set", [](ir::Pass &self, const std::string &name,
                     int val) { self.Set<const int>(name, new int(val)); })
1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825
      .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
1826 1827
      .def("type", &ir::Pass::Type)
      .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
1828
        self.Apply(graph.get());
F
flame 已提交
1829
      });
1830

X
fix  
Xin Pan 已提交
1831 1832
  py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
      m, "PassBuilder");
1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846
  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
1847
  // -- python binds for parallel executor.
X
Xin Pan 已提交
1848

Y
yuyang18 已提交
1849
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
C
chengduo 已提交
1850 1851 1852 1853
  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
1854 1855 1856
    Examples:
        .. code-block:: python

1857
          import paddle.fluid as fluid
1858 1859 1860 1861 1862 1863 1864 1865 1866 1867
          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
1868 1869 1870
          exec_strategy = fluid.ExecutionStrategy()
          exec_strategy.num_threads = 4

1871 1872
          train_exe = fluid.ParallelExecutor(use_cuda=False,
                                             loss_name=avg_loss.name,
C
chengduo 已提交
1873 1874
                                             exec_strategy=exec_strategy)

C
chengduo 已提交
1875 1876
        )DOC");

Y
yuyang18 已提交
1877
  exec_strategy.def(py::init())
Y
yuyang18 已提交
1878 1879 1880 1881 1882
      .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
1883 1884 1885 1886 1887 1888 1889 1890 1891 1892
          },
          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
1893
      .def_property(
1894 1895 1896 1897
          "use_cuda",
          [](const ExecutionStrategy &self) { return self.use_cuda_; },
          [](ExecutionStrategy &self, bool use_cuda) {
            self.use_cuda_ = use_cuda;
C
chengduo 已提交
1898 1899 1900 1901
          })  // 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
1902 1903 1904 1905 1906
      .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
1907 1908 1909
          },
          R"DOC(The type is BOOL, allow_op_delay represents whether to delay the
                communication operators to run, it may make the execution faster.
1910 1911
                Note that this option is invalid now, and it will be removed in
                next version. Default False.)DOC")
Y
yuyang18 已提交
1912 1913 1914 1915 1916 1917 1918
      .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
1919 1920 1921 1922
          },
          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,
1923 1924
                because the temp variable's shape maybe the same between two iterations.
                Default 1.
C
chengduo 已提交
1925 1926 1927 1928 1929 1930

                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`.
1931
              )DOC")
Q
Qiao Longfei 已提交
1932 1933 1934 1935 1936 1937 1938 1939 1940
      .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
1941
                user call exe.run() in python
Q
Qiao Longfei 已提交
1942
              )DOC")
1943 1944 1945 1946 1947 1948 1949 1950
      .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")
1951 1952 1953 1954 1955
      .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
1956

Y
yuyang18 已提交
1957
  exec_strategy.def_property(
Y
yuyang18 已提交
1958 1959 1960 1961 1962 1963 1964
      "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
1965 1966
      });

C
chengduo 已提交
1967 1968 1969 1970
  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
1971 1972 1973
    Examples:
        .. code-block:: python

1974 1975
            import os
            import numpy as np
F
flame 已提交
1976
            import paddle.fluid as fluid
1977 1978 1979 1980 1981 1982 1983 1984 1985

            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
1986
            build_strategy = fluid.BuildStrategy()
1987 1988
            build_strategy.enable_inplace = True
            build_strategy.memory_optimize = True
F
flame 已提交
1989
            build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
1990 1991 1992 1993
            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
1994
)DOC");
Y
yuyang18 已提交
1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010

  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) {
2011 2012 2013 2014
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2015
            self.reduce_ = strategy;
C
chengduo 已提交
2016
          },
2017
          R"DOC((fluid.BuildStrategy.ReduceStrategy, optional): there are two reduce
C
chengduo 已提交
2018 2019
                strategies in ParallelExecutor, AllReduce and Reduce. If you want
                that all the parameters' optimization are done on all devices independently,
2020
                you should choose AllReduce; otherwise, if you choose Reduce, all the parameters'
C
chengduo 已提交
2021 2022
                optimization will be evenly distributed to different devices, and then
                broadcast the optimized parameter to other devices.
2023
                Default is 'AllReduce'.
F
flame 已提交
2024 2025 2026 2027 2028 2029 2030 2031

                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
2032 2033 2034 2035 2036
      .def_property(
          "gradient_scale_strategy",
          [](const BuildStrategy &self) { return self.gradient_scale_; },
          [](BuildStrategy &self,
             BuildStrategy::GradientScaleStrategy strategy) {
2037 2038 2039 2040
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2041
            self.gradient_scale_ = strategy;
C
chengduo 已提交
2042
          },
2043 2044
          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
2045 2046
                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`,
2047
                you can choose Customized. Default is 'CoeffNumDevice'.
F
flame 已提交
2048 2049 2050 2051 2052

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
C
chengduo 已提交
2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080
                        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
2081
                        build_strategy = fluid.BuildStrategy()
C
chengduo 已提交
2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095
                        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
2096
                   )DOC")
Y
yuyang18 已提交
2097 2098 2099 2100
      .def_property(
          "debug_graphviz_path",
          [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
          [](BuildStrategy &self, const std::string &path) {
2101 2102 2103 2104
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2105
            self.debug_graphviz_path_ = path;
C
chengduo 已提交
2106
          },
2107
          R"DOC((str, optional): debug_graphviz_path indicates the path that
F
flame 已提交
2108
                writing the SSA Graph to file in the form of graphviz.
2109
                It is useful for debugging. Default is empty string, that is, ""
F
flame 已提交
2110 2111 2112 2113 2114 2115

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
C
chengduo 已提交
2116 2117
                        build_strategy.debug_graphviz_path = "./graph"

F
flame 已提交
2118
                    )DOC")
S
sneaxiy 已提交
2119 2120 2121 2122 2123 2124
      .def_property(
          "enable_sequential_execution",
          [](const BuildStrategy &self) {
            return self.enable_sequential_execution_;
          },
          [](BuildStrategy &self, bool b) {
2125 2126 2127 2128
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2129 2130
            self.enable_sequential_execution_ = b;
          },
2131 2132
          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
2133 2134 2135 2136 2137 2138 2139 2140

                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
2141 2142 2143 2144 2145 2146
      .def_property(
          "remove_unnecessary_lock",
          [](const BuildStrategy &self) {
            return self.remove_unnecessary_lock_;
          },
          [](BuildStrategy &self, bool b) {
2147 2148 2149 2150
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2151 2152
            self.remove_unnecessary_lock_ = b;
          },
2153 2154
          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
2155 2156 2157 2158 2159 2160 2161 2162

                Examples:
                    .. code-block:: python

                        import paddle.fluid as fluid
                        build_strategy = fluid.BuildStrategy()
                        build_strategy.remove_unnecessary_lock = True
          )DOC")
2163 2164 2165 2166
      .def_property(
          "num_trainers",
          [](const BuildStrategy &self) { return self.num_trainers_; },
          [](BuildStrategy &self, int num_trainers) {
2167
#ifdef WIN32
2168
            PADDLE_THROW(platform::errors::Unavailable(
2169
                "Distribution mode is not supported on Windows platform."));
2170
#endif
2171 2172
            self.num_trainers_ = num_trainers;
          })
2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184
      .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;
                    })
2185 2186 2187 2188 2189 2190
      .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;
          })
2191
      .def_property("use_hierarchical_allreduce",
2192 2193 2194 2195 2196 2197
                    [](const BuildStrategy &self) {
                      return self.use_hierarchical_allreduce_;
                    },
                    [](BuildStrategy &self, bool use) {
                      self.use_hierarchical_allreduce_ = use;
                    })
2198
      .def_property("hierarchical_allreduce_inter_nranks",
2199 2200 2201 2202 2203 2204 2205
                    [](const BuildStrategy &self) {
                      return self.hierarchical_allreduce_inter_nranks_;
                    },
                    [](BuildStrategy &self, int nranks) {
                      self.hierarchical_allreduce_inter_nranks_ = nranks;
                    })

C
chengduo 已提交
2206 2207 2208 2209 2210 2211
      .def_property(
          "fuse_elewise_add_act_ops",
          [](const BuildStrategy &self) {
            return self.fuse_elewise_add_act_ops_;
          },
          [](BuildStrategy &self, bool b) {
2212 2213 2214 2215
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
C
chengduo 已提交
2216 2217
            self.fuse_elewise_add_act_ops_ = b;
          },
2218
          R"DOC((bool, optional): fuse_elewise_add_act_ops indicate whether
F
flame 已提交
2219
                to fuse elementwise_add_op and activation_op,
2220
                it may make the execution faster. Default is False.
F
flame 已提交
2221 2222 2223 2224 2225 2226 2227 2228

                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
2229 2230 2231 2232
      .def_property(
          "fuse_bn_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_act_ops_; },
          [](BuildStrategy &self, bool b) {
2233
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
Z
Zhen Wang 已提交
2234
                              platform::errors::PreconditionNotMet(
2235 2236
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Z
Zhen Wang 已提交
2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249
            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")
2250 2251 2252 2253
      .def_property(
          "enable_auto_fusion",
          [](const BuildStrategy &self) { return self.enable_auto_fusion_; },
          [](BuildStrategy &self, bool b) {
2254
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
2255
                              platform::errors::PreconditionNotMet(
2256 2257
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271
            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")
2272 2273 2274 2275 2276 2277
      .def_property(
          "fuse_relu_depthwise_conv",
          [](const BuildStrategy &self) {
            return self.fuse_relu_depthwise_conv_;
          },
          [](BuildStrategy &self, bool b) {
2278 2279 2280 2281
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2282 2283
            self.fuse_relu_depthwise_conv_ = b;
          },
2284
          R"DOC((bool, optional): fuse_relu_depthwise_conv indicate whether
F
flame 已提交
2285 2286 2287
                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.
2288
                Default is False.
F
flame 已提交
2289 2290 2291 2292 2293 2294 2295 2296

                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
2297 2298 2299 2300 2301 2302
      .def_property("fuse_broadcast_ops",
                    [](const BuildStrategy &self) {
                      return self.fuse_broadcast_ops_ == true ||
                             self.fuse_broadcast_ops_ == boost::none;
                    },
                    [](BuildStrategy &self, bool b) {
2303 2304 2305 2306
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2307 2308
                      self.fuse_broadcast_ops_ = b;
                    },
2309
                    R"DOC((bool, optional): fuse_broadcast_op indicates whether
2310 2311 2312 2313
                      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
2314 2315 2316 2317 2318 2319 2320 2321 2322
                      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
2323 2324
      .def_property("fuse_all_optimizer_ops",
                    [](const BuildStrategy &self) {
C
chengduo 已提交
2325 2326
                      return self.fuse_all_optimizer_ops_ == true ||
                             self.fuse_all_optimizer_ops_ == boost::none;
C
chengduo 已提交
2327 2328
                    },
                    [](BuildStrategy &self, bool b) {
2329 2330 2331 2332
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2333 2334
                      self.fuse_all_optimizer_ops_ = b;
                    })
Q
qingqing01 已提交
2335 2336 2337 2338
      .def_property(
          "sync_batch_norm",
          [](const BuildStrategy &self) { return self.sync_batch_norm_; },
          [](BuildStrategy &self, bool b) {
2339 2340 2341 2342
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Q
qingqing01 已提交
2343 2344
            self.sync_batch_norm_ = b;
          },
2345
          R"DOC((bool, optional): sync_batch_norm indicates whether to use
Q
qingqing01 已提交
2346 2347 2348
                synchronous batch normalization which synchronizes the mean
                and variance through multi-devices in training phase.
                Current implementation doesn't support FP16 training and CPU.
2349 2350
                And only synchronous on one machine, not all machines. 
                Default is False.
F
flame 已提交
2351 2352 2353 2354 2355 2356 2357 2358

                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
2359 2360
      .def_property(
          "memory_optimize",
2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374
          [](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 {
2375 2376 2377
              PADDLE_THROW(platform::errors::InvalidArgument(
                  "BuildStrategy.memory_optimize must be set to None, False or "
                  "True"));
2378 2379
            }
          },
2380
          R"DOC((bool, optional): memory opitimize aims to save total memory
2381
                consumption, set to True to enable it.
2382

2383 2384 2385
                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. 
2386
                True means enabling and False means disabling. Default is None.)DOC")
2387 2388 2389
      .def_property(
          "is_distribution",
          [](const BuildStrategy &self) { return self.is_distribution_; },
2390 2391 2392
          [](BuildStrategy &self, bool b) {
#ifdef WIN32
            if (b) {
2393
              PADDLE_THROW(platform::errors::Unavailable(
2394
                  "Distribution mode is not supported on Windows platform."));
2395 2396 2397 2398 2399
            }
#else
            self.is_distribution_ = b;
#endif
          })
Q
can run  
Qiao Longfei 已提交
2400 2401 2402
      .def_property("async_mode",
                    [](const BuildStrategy &self) { return self.async_mode_; },
                    [](BuildStrategy &self, bool b) { self.async_mode_ = b; })
D
dzhwinter 已提交
2403
      .def_property(
D
dzhwinter 已提交
2404 2405 2406
          "enable_inplace",
          [](const BuildStrategy &self) { return self.enable_inplace_; },
          [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; })
C
chengduo 已提交
2407 2408
      .def_property(
          "fuse_all_reduce_ops",
C
chengduo 已提交
2409 2410 2411 2412
          [](const BuildStrategy &self) {
            return self.fuse_all_reduce_ops_ == true ||
                   self.fuse_all_reduce_ops_ == boost::none;
          },
C
chengduo 已提交
2413
          [](BuildStrategy &self, bool b) { self.fuse_all_reduce_ops_ = b; })
2414 2415 2416 2417 2418 2419 2420
      .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;
                    })
2421 2422 2423 2424
      .def_property(
          "cache_runtime_context",
          [](const BuildStrategy &self) { return self.cache_runtime_context_; },
          [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
2425 2426 2427 2428 2429 2430 2431 2432 2433
      .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;
          })
2434
      .def("_finalize_strategy_and_create_passes",
X
fix  
Xin Pan 已提交
2435
           [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
2436 2437 2438 2439 2440
             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
2441 2442

  pe.def(py::init<const std::vector<platform::Place> &,
Y
Yan Xu 已提交
2443
                  const std::vector<std::string> &, const std::string &,
X
Xin Pan 已提交
2444
                  Scope *, std::vector<Scope *> &, const ExecutionStrategy &,
X
Xin Pan 已提交
2445
                  const BuildStrategy &, ir::Graph *>())
Y
Yu Yang 已提交
2446 2447 2448 2449
      // 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.
2450 2451 2452 2453 2454
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
2455 2456 2457
      .def("drop_local_exe_scopes", &ParallelExecutor::DropLocalExeScopes)
      .def("_need_create_local_exe_scopes",
           &ParallelExecutor::NeedCreateLocalExeScope)
Y
Yu Yang 已提交
2458 2459 2460 2461
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
2462 2463
      .def("run",
           [](ParallelExecutor &self,
Z
Zhen Wang 已提交
2464 2465 2466 2467 2468 2469 2470 2471
              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) {
2472
               return py::cast(
2473
                   std::move(BOOST_GET(paddle::framework::FetchList, ret)));
Z
Zhen Wang 已提交
2474 2475
             } else {
               return py::cast(std::move(
2476
                   BOOST_GET(paddle::framework::FetchUnmergedList, ret)));
Z
Zhen Wang 已提交
2477
             }
2478 2479
           })
      .def("device_count", &ParallelExecutor::DeviceCount);
Y
Yu Yang 已提交
2480

D
dongdaxiang 已提交
2481
  BindFleetWrapper(&m);
2482
  BindGlooWrapper(&m);
H
hutuxian 已提交
2483
  BindBoxHelper(&m);
H
hutuxian 已提交
2484 2485 2486
#ifdef PADDLE_WITH_BOX_PS
  BindBoxWrapper(&m);
#endif
2487
#ifdef PADDLE_WITH_NCCL
D
dongdaxiang 已提交
2488
  BindNCCLWrapper(&m);
W
wopeizl 已提交
2489
#endif
F
flame 已提交
2490 2491
  BindGraph(&m);
  BindNode(&m);
F
flame 已提交
2492
  BindInferenceApi(&m);
2493
  BindDataset(&m);
Y
Yanghello 已提交
2494 2495 2496
#ifdef PADDLE_WITH_CRYPTO
  BindCrypto(&m);
#endif
2497 2498
#ifdef PADDLE_WITH_DISTRIBUTE
  BindCommunicator(&m);
2499 2500
  BindCommunicatorContext(&m);
  BindLargeScaleKV(&m);
2501
#endif
L
Luo Tao 已提交
2502
}
2503
}  // namespace pybind
2504
}  // namespace paddle