imperative.cc 52.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

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

http://www.apache.org/licenses/LICENSE-2.0

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. */

#include "paddle/fluid/pybind/imperative.h"
16

17
#include <Python.h>
18 19 20 21
#include <pybind11/chrono.h>
#include <pybind11/complex.h>
#include <pybind11/functional.h>
#include <pybind11/stl.h>
22

23
#include <memory>
24
#include <set>
J
Jiabin Yang 已提交
25
#include <string>
26
#include <unordered_map>
27
#include <unordered_set>
28
#include <utility>
J
Jiabin Yang 已提交
29
#include <vector>
30

31
#include "paddle/fluid/imperative/all_reduce.h"
32
#include "paddle/fluid/imperative/amp_auto_cast.h"
33
#include "paddle/fluid/imperative/basic_engine.h"
34
#include "paddle/fluid/imperative/data_loader.h"
35
#include "paddle/fluid/imperative/layer.h"
J
Jiabin Yang 已提交
36
#include "paddle/fluid/imperative/nccl_context.h"
37
#include "paddle/fluid/imperative/partial_grad_engine.h"
38
#include "paddle/fluid/imperative/profiler.h"
39
#include "paddle/fluid/imperative/tracer.h"
M
minqiyang 已提交
40
#include "paddle/fluid/imperative/type_defs.h"
41
#include "paddle/fluid/memory/allocation/mmap_allocator.h"
42
#include "paddle/fluid/pybind/op_function.h"
43
#include "paddle/fluid/pybind/pybind_boost_headers.h"
L
Leo Chen 已提交
44
#include "paddle/fluid/pybind/tensor_py.h"
45

46 47 48
namespace paddle {
namespace pybind {

49 50
namespace py = ::pybind11;

51 52 53 54
class Layer : public imperative::Layer {
 public:
  using imperative::Layer::Layer;  // Inherit constructors

55 56 57 58
  std::vector<std::shared_ptr<imperative::VarBase>> Forward(
      const std::vector<std::shared_ptr<imperative::VarBase>> &inputs)
      override {
    PYBIND11_OVERLOAD(std::vector<std::shared_ptr<imperative::VarBase>>, Layer,
J
Jiabin Yang 已提交
59
                      Forward, inputs);  // NOLINT
60 61 62
  }
};

L
Leo Chen 已提交
63 64 65 66 67
static const platform::Place PyObjectToPlace(const py::object &place_obj) {
  if (py::isinstance<platform::CPUPlace>(place_obj)) {
    return place_obj.cast<platform::CPUPlace>();
  } else if (py::isinstance<platform::CUDAPlace>(place_obj)) {
    return place_obj.cast<platform::CUDAPlace>();
68 69
  } else if (py::isinstance<platform::XPUPlace>(place_obj)) {
    return place_obj.cast<platform::XPUPlace>();
L
Leo Chen 已提交
70 71
  } else if (py::isinstance<platform::CUDAPinnedPlace>(place_obj)) {
    return place_obj.cast<platform::CUDAPinnedPlace>();
72 73
  } else if (py::isinstance<platform::Place>(place_obj)) {
    return place_obj.cast<platform::Place>();
L
Leo Chen 已提交
74 75
  } else {
    PADDLE_THROW(platform::errors::InvalidArgument(
76 77
        "Place should be one of "
        "Place/CPUPlace/XPUPlace/CUDAPlace/CUDAPinnedPlace"));
L
Leo Chen 已提交
78 79 80 81 82 83 84
  }
}

static void InitTensorForVarBase(imperative::VarBase *self,
                                 const py::array &array,
                                 const platform::Place place,
                                 bool persistable = false,
85 86
                                 bool zero_copy = false, std::string name = "",
                                 int stop_gradient = -1) {
L
Leo Chen 已提交
87
  if (name == "") {
88 89
    name =
        imperative::GetCurrentTracer()->GenerateUniqueName("generated_tensor");
L
Leo Chen 已提交
90
  }
91 92 93
  VLOG(5) << "Init Tensor as: / name: " << name
          << " / persistable: " << persistable << " / zero_copy: " << zero_copy
          << " / stop_gradient: " << stop_gradient;
L
Leo Chen 已提交
94
  new (self) imperative::VarBase(name);
95
  auto *tensor = self->MutableVar()->GetMutable<framework::LoDTensor>();
L
Leo Chen 已提交
96 97
  if (platform::is_cpu_place(place)) {
    SetTensorFromPyArray<platform::CPUPlace>(
98
        tensor, array, BOOST_GET_CONST(platform::CPUPlace, place), zero_copy);
99 100 101
  } else if (platform::is_xpu_place(place)) {
    SetTensorFromPyArray<platform::XPUPlace>(
        tensor, array, BOOST_GET_CONST(platform::XPUPlace, place), zero_copy);
L
Leo Chen 已提交
102 103
  } else if (platform::is_gpu_place(place)) {
    SetTensorFromPyArray<platform::CUDAPlace>(
104
        tensor, array, BOOST_GET_CONST(platform::CUDAPlace, place), zero_copy);
L
Leo Chen 已提交
105 106
  } else if (platform::is_cuda_pinned_place(place)) {
    SetTensorFromPyArray<platform::CUDAPinnedPlace>(
107 108
        tensor, array, BOOST_GET_CONST(platform::CUDAPinnedPlace, place),
        zero_copy);
109
  } else {
L
Leo Chen 已提交
110
    PADDLE_THROW(platform::errors::InvalidArgument(
111
        "Place should be one of CPUPlace/XPUPlace/CUDAPlace/CUDAPinnedPlace"));
J
Jiabin Yang 已提交
112
  }
113 114 115
  if (stop_gradient != -1) {
    self->SetOverridedStopGradient(stop_gradient);
  }
L
Leo Chen 已提交
116
  self->SetPersistable(persistable);
117 118 119 120 121 122
  self->SetType(framework::proto::VarType::LOD_TENSOR);
  self->SetDataType(tensor->type());
}

static void InitVarBaseFromNumpyWithKwargs(imperative::VarBase *self,
                                           const py::kwargs &kwargs) {
123
  VLOG(4) << "Init VarBase from kwargs: ";
124 125
  PADDLE_ENFORCE_EQ(
      kwargs.contains("value"), true,
126 127
      platform::errors::NotFound(
          "The kwargs used to create Varbase misses argument: value"));
L
Leo Chen 已提交
128 129 130 131 132 133 134 135
  auto persistable = kwargs.contains("persistable")
                         ? kwargs["persistable"].cast<bool>()
                         : false;
  auto array = kwargs.contains("value") ? kwargs["value"].cast<py::array>()
                                        : py::array();
  auto zero_copy =
      kwargs.contains("zero_copy") ? kwargs["zero_copy"].cast<bool>() : false;
  auto name = kwargs.contains("name") ? kwargs["name"].cast<std::string>() : "";
136 137 138
  auto stop_gradient = kwargs.contains("stop_gradient")
                           ? kwargs["stop_gradient"].cast<int>()
                           : -1;
L
Leo Chen 已提交
139 140 141
  auto default_place = imperative::GetCurrentTracer()->ExpectedPlace();
  auto place = kwargs.contains("place") ? PyObjectToPlace(kwargs["place"])
                                        : default_place;
142 143
  InitTensorForVarBase(self, array, place, persistable, zero_copy, name,
                       stop_gradient);
144
}
145

146 147 148
template <typename P>
static void InitVarBaseFromNumpyWithArg(imperative::VarBase *self,
                                        const py::array &array, const P &place,
L
Leo Chen 已提交
149 150
                                        bool persistable = false,
                                        bool zero_copy = false,
151 152 153 154 155
                                        std::string name = "",
                                        int stop_gradient = -1) {
  VLOG(4) << "Init VarBase from Arg: ";
  // 0: self, 1: value, 2: place, 3: persistable, 4: zero_copy, 5: name , 6:
  // stop_gradient
L
Leo Chen 已提交
156
  if (name == "") {
157 158
    name =
        imperative::GetCurrentTracer()->GenerateUniqueName("generated_tensor");
L
Leo Chen 已提交
159
  }
160 161 162
  VLOG(5) << "Init Tensor as: / name: " << name
          << " / persistable: " << persistable << " / zero_copy: " << zero_copy
          << " / stop_gradient: " << stop_gradient;
L
Leo Chen 已提交
163
  new (self) imperative::VarBase(name);
164 165
  self->SetPersistable(persistable);
  auto *tensor = self->MutableVar()->GetMutable<framework::LoDTensor>();
166 167 168
  if (stop_gradient != -1) {
    self->SetOverridedStopGradient(stop_gradient);
  }
169 170 171 172 173 174
  SetTensorFromPyArray<P>(tensor, array, place, zero_copy);
  self->SetType(framework::proto::VarType::LOD_TENSOR);
  self->SetDataType(tensor->type());
}

static void InitVarBaseFromNumpyWithArgDefault(imperative::VarBase *self,
L
Leo Chen 已提交
175
                                               const py::array &array) {
176
  VLOG(4) << "Init VarBase from numpy: ";
L
Leo Chen 已提交
177 178
  auto place = imperative::GetCurrentTracer()->ExpectedPlace();
  InitTensorForVarBase(self, array, place);
179
}
180

181 182 183 184 185
static void InitVarBaseFromTensorWithArgDefault(
    imperative::VarBase *self, const framework::LoDTensor &tensor) {
  VLOG(4) << "Init VarBase";
  auto place = imperative::GetCurrentTracer()->ExpectedPlace();
  new (self) imperative::VarBase(
186
      imperative::GetCurrentTracer()->GenerateUniqueName("generated_tensor"));
187 188 189 190 191 192 193 194 195 196 197 198 199 200
  self->SetPersistable(false);
  self->SetType(framework::proto::VarType::LOD_TENSOR);
  self->SetDataType(tensor.type());
  auto *new_tensor = self->MutableVar()->GetMutable<framework::LoDTensor>();
  // Same place,share data directly
  if (place == tensor.place()) {
    new_tensor->ShareDataWith(tensor);
    VLOG(4) << "Same place, do ShareDataWith";
  } else {
    framework::TensorCopy(tensor, place, new_tensor);
    VLOG(4) << "Different place, do TensorCopy";
  }
}

201 202 203 204 205
static std::string GetTypeName(const imperative::VarBase &var) {
  if (var.Type() == framework::proto::VarType::RAW) {
    return "RAW";
  } else if (!var.Var().IsInitialized()) {
    return "nullptr";
206
  } else {
207
    return framework::ToTypeName(var.Var().Type());
208 209
  }
}
L
Leo Chen 已提交
210

211
using PyNameVarBaseMap = std::unordered_map<std::string, py::handle>;
212 213 214 215 216 217

template <typename T>
static T PyObjectCast(PyObject *obj) {
  try {
    return py::cast<T>(py::handle(obj));
  } catch (py::cast_error &) {
218 219
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Python object is not type of %s", typeid(T).name()));
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234
  }
}

// NOTE(zjl): py::handle is a very light wrapper of PyObject *.
// Unlike py::object, py::handle does not change reference count of PyObject *.
static std::vector<std::shared_ptr<imperative::VarBase>>
GetVarBaseListFromPyHandle(const py::handle &handle) {
  PyObject *py_obj = handle.ptr();  // get underlying PyObject
  // Python None is not nullptr in C++!
  if (!py_obj || py_obj == Py_None) {
    return {};
  }

  std::vector<std::shared_ptr<imperative::VarBase>> result;

235
  if (PyList_Check(py_obj)) {  // List of VarBase
236 237 238
    size_t len = PyList_GET_SIZE(py_obj);
    result.reserve(len);
    for (size_t i = 0; i < len; ++i) {
239 240 241
      PyObject *py_ivar = PyList_GET_ITEM(py_obj, i);
      PADDLE_ENFORCE_NOT_NULL(
          py_ivar, platform::errors::InvalidArgument("Python Object is NULL"));
242 243 244
      result.emplace_back(
          PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_ivar));
    }
245
  } else if (PyTuple_Check(py_obj)) {  // Tuple of VarBase
246 247 248
    size_t len = PyTuple_GET_SIZE(py_obj);
    result.reserve(len);
    for (size_t i = 0; i < len; ++i) {
249 250 251
      PyObject *py_ivar = PyTuple_GET_ITEM(py_obj, i);
      PADDLE_ENFORCE_NOT_NULL(
          py_ivar, platform::errors::InvalidArgument("Python Object is NULL"));
252 253 254
      result.emplace_back(
          PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_ivar));
    }
255 256 257
  } else {  // VarBase
    result.emplace_back(
        PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_obj));
258 259 260 261 262
  }

  return result;
}

J
Jiabin Yang 已提交
263 264 265
static imperative::NameVarBaseMap ConvertToNameVarBaseMap(
    const PyNameVarBaseMap &map) {
  imperative::NameVarBaseMap result;
266 267 268 269 270 271
  for (auto &pair : map) {
    auto var_vec = GetVarBaseListFromPyHandle(pair.second);
    if (!var_vec.empty()) {
      result.emplace(pair.first, std::move(var_vec));
    }
  }
J
Jiabin Yang 已提交
272

273 274 275
  PADDLE_ENFORCE_EQ(
      PyErr_Occurred(), nullptr,
      platform::errors::InvalidArgument(py::str(py::handle(PyErr_Occurred()))));
276 277 278
  return result;
}

279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
static bool PyCheckInteger(PyObject *obj) {
#if PY_VERSION_HEX < 0x03000000
  return (PyLong_Check(obj) || PyInt_Check(obj)) && !PyBool_Check(obj);
#else
  return PyLong_Check(obj) && !PyBool_Check(obj);
#endif
}

// NOTE(zhiqiu): Revised version of PySlice_GetIndices. From:
// https://github.com/python/cpython/blob/8d21aa21f2cbc6d50aab3f420bb23be1d081dac4/Objects/sliceobject.c#L103
// Original PySlice_GetIndices return wrong result when
// slice_item contains long int, such as arr[:180L].
// NOT sure why this happens !!!
// Besides, PySlice_GetIndices cannot raise error when float in slice item.
// So, I make a revised version of PySlice_GetIndices, named to
// _PySlice_GetIndices. Try to use _PySlice_Unpack which is more robust than
// PySlice_GetIndices in the future.
static int _PySlice_GetIndices(PySliceObject *r, Py_ssize_t length,
                               Py_ssize_t *start, Py_ssize_t *stop,
                               Py_ssize_t *step) {
  /* XXX support long ints */
  if (r->step == Py_None) {
    *step = 1;
  } else {
    if (PyCheckInteger(r->step)) {
      *step = PyLong_AsLong(r->step);
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Currently, VarBase.__getitem__() only allows None or integers in "
          "slice item, but received %s.",
          std::string(Py_TYPE(r->step)->tp_name)));
    }
  }
  if (r->start == Py_None) {
    *start = *step < 0 ? length - 1 : 0;
  } else {
    if (PyCheckInteger(r->start)) {
      *start = PyLong_AsLong(r->start);
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Currently, VarBase.__getitem__() only allows None or integers in "
          "slice item, but received %s.",
          std::string(Py_TYPE(r->start)->tp_name)));
    }
    if (*start < 0) *start += length;
  }
  if (r->stop == Py_None) {
    *stop = *step < 0 ? -1 : length;
  } else {
    if (PyCheckInteger(r->stop)) {
      *stop = PyLong_AsLong(r->stop);
    } else {
      PADDLE_THROW(platform::errors::InvalidArgument(
          "Currently, VarBase.__getitem__() only allows None or integers in "
          "slice item, but received %s.",
          std::string(Py_TYPE(r->stop)->tp_name)));
    }
    if (*stop < 0) *stop += length;
  }
  if (*stop > length) return -1;
  if (*start >= length) return -1;
  if (*step == 0) return -1;
  return 0;
}

S
songyouwei 已提交
344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367
static void ParseIndexingSlice(framework::LoDTensor *tensor, PyObject *_index,
                               std::vector<int> *slice_axes,
                               std::vector<int> *slice_starts,
                               std::vector<int> *slice_ends,
                               std::vector<int> *slice_strides,
                               std::vector<int> *decrease_axis,
                               std::vector<int> *infer_flags) {
  // We allow indexing by Integers, Slices, and tuples of those
  // types.
  // Ellipsis and None are not supported yet.
  // wrap to tuple
  PyObject *index = !PyTuple_Check(_index) ? PyTuple_Pack(1, _index) : _index;
  PADDLE_ENFORCE_EQ(
      tensor->IsInitialized(), true,
      platform::errors::InvalidArgument("tensor has not been initialized"));
  const auto &shape = tensor->dims();
  const int rank = shape.size();
  const int size = PyTuple_GET_SIZE(index);
  PADDLE_ENFORCE_EQ(
      size <= rank, true,
      platform::errors::InvalidArgument(
          "too many indices (%d) for tensor of dimension %d", size, rank));
  for (int dim = 0; dim < size; ++dim) {
    PyObject *slice_item = PyTuple_GetItem(index, dim);
368 369 370 371 372 373 374
    PADDLE_ENFORCE_EQ(PyCheckInteger(slice_item) || PySlice_Check(slice_item),
                      true,
                      platform::errors::InvalidArgument(
                          "Currently, VarBase.__getitem__() only allows "
                          "indexing by Integers, Slices, and tuples of "
                          "these types, but received %s in %dth slice item",
                          std::string(Py_TYPE(slice_item)->tp_name), dim + 1));
S
songyouwei 已提交
375 376
    infer_flags->push_back(1);
    int dim_len = shape[dim];
377 378
    if (PyCheckInteger(slice_item)) {
      // integer, PyLong_AsLong supports both int and long
S
songyouwei 已提交
379
      int start = static_cast<int>(PyLong_AsLong(slice_item));
H
hong 已提交
380
      auto s_t = start;
S
songyouwei 已提交
381
      start = start < 0 ? start + dim_len : start;
H
hong 已提交
382 383 384 385 386 387 388 389 390 391
      if (start >= dim_len) {
        std::string str_error_message =
            "The starting index " + std::to_string(s_t) +
            " of slice is out of bounds in tensor " + std::to_string(dim) +
            "-th axis, it shound be in the range of [" +
            std::to_string(-dim_len) + ", " + std::to_string(dim_len) + ")";
        // py::index_error is corresponding to IndexError in Python
        // Used to indicate out of bounds access in __getitem__, __setitem__
        throw py::index_error(str_error_message);
      }
S
songyouwei 已提交
392 393 394 395 396 397
      slice_axes->push_back(dim);
      slice_starts->push_back(start);
      slice_ends->push_back(start + 1);
      slice_strides->push_back(1);
      decrease_axis->push_back(dim);
    } else {
398
      // slice item
S
songyouwei 已提交
399
      Py_ssize_t start, end, step;
400 401 402
      PySliceObject *p = reinterpret_cast<PySliceObject *>(slice_item);
      _PySlice_GetIndices(p, dim_len, &start, &end, &step);

S
songyouwei 已提交
403
      // :: or : or 0:dim_len:1
404 405 406
      if (start == 0 && end == dim_len && step == 1) {
        continue;
      }
S
songyouwei 已提交
407 408 409 410 411 412 413 414 415
      slice_axes->push_back(dim);
      slice_starts->push_back(start);
      slice_ends->push_back(end);
      slice_strides->push_back(step);
    }
  }
  if (!PyTuple_Check(_index)) Py_DecRef(index);
}

416
// Bind Methods
J
Jiabin Yang 已提交
417
void BindImperative(py::module *m_ptr) {
418 419
  auto &m = *m_ptr;

420 421
  BindOpFunctions(&m);

422 423
#ifndef _WIN32
  // Dygraph DataLoader signal handler
424 425 426 427 428 429 430 431 432 433 434 435 436
  m.def("_set_process_pids", [](int64_t key, py::object &obj) {
    PADDLE_ENFORCE_EQ(
        py::isinstance<py::tuple>(obj) || py::isinstance<py::list>(obj), true,
        platform::errors::InvalidArgument(
            "The subprocess ids set in DataLoader is illegal."
            "Expected data type is tuple or list, but received %s",
            obj.get_type()));
    py::list pids = py::cast<py::list>(obj);
    std::set<pid_t> pids_set = {};
    for (size_t i = 0; i < pids.size(); i++) {
      pids_set.insert(pids[i].cast<pid_t>());
    }
    imperative::SetLoadProcessPIDs(key, pids_set);
437
  });
438 439
  m.def("_erase_process_pids",
        [](int64_t key) { imperative::EraseLoadProcessPIDs(key); });
440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511
  m.def("_set_process_signal_handler",
        []() { imperative::SetLoadProcessSignalHandler(); });
  m.def("_throw_error_if_process_failed",
        []() { imperative::ThrowErrorIfLoadProcessFailed(); });

  // Dygraph DataLoader reader process & thread related functions
  m.def(
      "_convert_to_tensor_list",
      [](py::object &obj) -> py::list {
        // 0. input data check
        PADDLE_ENFORCE(
            py::isinstance<py::tuple>(obj) || py::isinstance<py::list>(obj),
            platform::errors::InvalidArgument(
                "The batch data read into DataLoader is illegal."
                "Expected data type is tuple or list, but received %s",
                obj.get_type()));
        py::list batch = py::cast<py::list>(obj);
        py::list tensors;
        for (size_t i = 0; i < batch.size(); ++i) {
          // 1. cast to python array
          auto array = batch[i].cast<py::array>();
          PADDLE_ENFORCE_NE(
              string::Sprintf("%s", array.dtype()).compare("object"), 0,
              platform::errors::InvalidArgument(
                  "Faild to convert input data to a regular ndarray.\n  * "
                  "Usually this means the input data contains nested "
                  "lists with different lengths.\n  * Check the reader "
                  "function passed to 'set_(sample/sample_list/batch)"
                  "_generator' to locate the data causes this issue."));
          // 2. construcct LoDTensor
          framework::LoDTensor t;
          SetTensorFromPyArray<platform::CPUPlace>(&t, array,
                                                   platform::CPUPlace(), true);
          // 3. allocate shared memory
          void *data_ptr = t.data<void>();
          size_t data_size = t.numel() * framework::SizeOfType(t.type());
          auto shared_writer_holder =
              memory::allocation::AllocateMemoryMapWriterAllocation(data_size);
          // 4. maintain mmap fd set & backup ipc_name
          const std::string &ipc_name = shared_writer_holder->ipc_name();
          memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);
          // 5. copy data & reset holder
          memory::Copy(platform::CPUPlace(), shared_writer_holder->ptr(),
                       platform::CPUPlace(), data_ptr, data_size);
          t.ResetHolder(shared_writer_holder);
          // 6. append to result list
          tensors.append(t);
        }
        return tensors;
      },
      py::return_value_policy::take_ownership);

  m.def("_remove_tensor_list_mmap_fds", [](py::list &tensor_list) {
    for (size_t i = 0; i < tensor_list.size(); ++i) {
      auto t = tensor_list[i].cast<framework::LoDTensor>();
      auto *mmap_writer_allocation =
          dynamic_cast<memory::allocation::MemoryMapWriterAllocation *>(
              t.Holder().get());
      PADDLE_ENFORCE_NOT_NULL(
          mmap_writer_allocation,
          platform::errors::NotFound("The shared memory of LoDTensor in "
                                     "DataLoader's child process has been "
                                     "released."));
      memory::allocation::MemoryMapFdSet::Instance().Remove(
          mmap_writer_allocation->ipc_name());
    }
  });

  m.def("_cleanup_mmap_fds",
        []() { memory::allocation::MemoryMapFdSet::Instance().Clear(); });
#endif

512 513 514 515 516
  m.def("start_imperative_gperf_profiler",
        []() { imperative::StartProfile(); });

  m.def("stop_imperative_gperf_profiler", []() { imperative::StopProfile(); });

Z
Zeng Jinle 已提交
517 518 519
  m.def("_is_dygraph_debug_enabled",
        []() { return imperative::IsDebugEnabled(); });
  m.def("_dygraph_debug_level", []() { return imperative::GetDebugLevel(); });
520 521 522 523
  m.def("_switch_tracer",
        [](const std::shared_ptr<imperative::Tracer> &tracer) {
          imperative::SetCurrentTracer(tracer);
        });
Z
Zeng Jinle 已提交
524

525
  py::class_<imperative::VarBase, std::shared_ptr<imperative::VarBase>>(
526
      m, "VarBase", R"DOC()DOC")
Z
Zeng Jinle 已提交
527
      .def_static("_alive_vars", &imperative::VarBase::AliveVarNames)
J
Jiabin Yang 已提交
528
      .def("__init__",
529 530 531
           [](imperative::VarBase &self, framework::proto::VarType::Type dtype,
              const std::vector<int> &dims, const py::handle &name,
              framework::proto::VarType::Type type, bool persistable) {
532
             VLOG(4) << "Init VarBase";
533 534 535
             std::string act_name = "";
             if (!name.ptr() || name.ptr() == Py_None) {
               act_name = imperative::GetCurrentTracer()->GenerateUniqueName(
536
                   "generated_tensor");
537 538 539 540
             } else {
               act_name = name.cast<std::string>();
             }
             new (&self) imperative::VarBase(act_name);
J
Jiabin Yang 已提交
541 542 543 544 545 546 547 548 549
             self.SetPersistable(persistable);
             self.SetType(type);
             self.SetDataType(dtype);
             if (type == framework::proto::VarType::LOD_TENSOR) {
               auto *tensor =
                   self.MutableVar()->GetMutable<framework::LoDTensor>();
               tensor->Resize(framework::make_ddim(dims));
             }
           })
550 551
      .def("__init__", &InitVarBaseFromNumpyWithArg<platform::CPUPlace>,
           py::arg("value"), py::arg("place"), py::arg("persistable") = false,
552 553
           py::arg("zero_copy") = false, py::arg("name") = "",
           py::arg("stop_gradient") = -1)
554 555 556 557
      .def("__init__", &InitVarBaseFromNumpyWithArg<platform::XPUPlace>,
           py::arg("value"), py::arg("place"), py::arg("persistable") = false,
           py::arg("zero_copy") = false, py::arg("name") = "",
           py::arg("stop_gradient") = -1)
558 559
      .def("__init__", &InitVarBaseFromNumpyWithArg<platform::CUDAPlace>,
           py::arg("value"), py::arg("place"), py::arg("persistable") = false,
560 561
           py::arg("zero_copy") = false, py::arg("name") = "",
           py::arg("stop_gradient") = -1)
562 563
      .def("__init__", &InitVarBaseFromNumpyWithArg<platform::CUDAPinnedPlace>,
           py::arg("value"), py::arg("place"), py::arg("persistable") = false,
564 565
           py::arg("zero_copy") = false, py::arg("name") = "",
           py::arg("stop_gradient") = -1)
L
Leo Chen 已提交
566
      .def("__init__", &InitVarBaseFromNumpyWithArgDefault, py::arg("value"))
567
      .def("__init__", &InitVarBaseFromTensorWithArgDefault, py::arg("tensor"))
568
      .def("__init__", &InitVarBaseFromNumpyWithKwargs)
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595
      .def("__setitem__",
           [](std::shared_ptr<imperative::VarBase> &self, py::handle _index,
              py::object &value_obj) {
             auto self_tensor =
                 self->MutableVar()->GetMutable<framework::LoDTensor>();
             auto self_numpy = TensorToPyArray(*self_tensor);

             if (py::isinstance<py::array>(value_obj) ||
                 py::isinstance<py::int_>(value_obj) ||
                 py::isinstance<py::float_>(value_obj)) {
               auto value_numpy = value_obj;
               self_numpy[_index] = value_numpy;
               SetTensorFromPyArray(self_tensor, self_numpy,
                                    self_tensor->place(), true);

             } else {
               auto value =
                   value_obj.cast<std::shared_ptr<imperative::VarBase>>();
               auto value_tensor =
                   value->MutableVar()->GetMutable<framework::LoDTensor>();
               auto value_numpy = TensorToPyArray(*value_tensor);

               self_numpy[_index] = value_numpy;
               SetTensorFromPyArray(self_tensor, self_numpy,
                                    self_tensor->place(), true);
             }
           })
596
      .def("__getitem__",
S
songyouwei 已提交
597
           [](std::shared_ptr<imperative::VarBase> &self, py::handle _index) {
598
             std::vector<int> slice_axes, slice_starts, slice_ends,
S
songyouwei 已提交
599 600 601 602 603 604
                 slice_strides, decrease_axis, infer_flags;
             auto tensor =
                 self->MutableVar()->GetMutable<framework::LoDTensor>();
             ParseIndexingSlice(tensor, _index.ptr(), &slice_axes,
                                &slice_starts, &slice_ends, &slice_strides,
                                &decrease_axis, &infer_flags);
605 606 607 608
             // release gil and do tracing
             py::gil_scoped_release release;
             const auto &tracer = imperative::GetCurrentTracer();
             if (slice_axes.empty()) {
S
songyouwei 已提交
609
               return self;
610
             } else {
S
songyouwei 已提交
611
               imperative::NameVarBaseMap ins = {{"Input", {self}}};
612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633
               framework::AttributeMap attrs = {
                   {"axes", slice_axes},
                   {"starts", slice_starts},
                   {"ends", slice_ends},
                   {"infer_flags", infer_flags},
                   {"decrease_axis", decrease_axis}};
               auto out = std::shared_ptr<imperative::VarBase>(
                   new imperative::VarBase(tracer->GenerateUniqueName()));
               imperative::NameVarBaseMap outs = {{"Out", {out}}};
               std::string op_type = "slice";
               for (auto stride : slice_strides) {
                 if (stride != 1) {
                   op_type = "strided_slice";
                   attrs.insert({"strides", slice_strides});
                   attrs.erase("decrease_axis");
                   break;
                 }
               }
               tracer->TraceOp(op_type, ins, outs, std::move(attrs));
               return out;
             }
           })
634 635 636 637 638 639 640
      .def("numpy",
           [](imperative::VarBase &self) -> py::array {
             const auto &tensor =
                 self.MutableVar()->Get<framework::LoDTensor>();
             PADDLE_ENFORCE_EQ(
                 tensor.IsInitialized(), true,
                 platform::errors::InvalidArgument(
641
                     "Tensor of %s is Empty, please check if it has no data.",
642 643 644 645 646
                     self.Name()));
             return TensorToPyArray(tensor, true);
           },
           R"DOC(

Z
Zhou Wei 已提交
647 648
        Returns a numpy array shows the value of current Tensor.
        
649
        Returns:
Z
Zhou Wei 已提交
650
            ndarray: The numpy value of current Tensor.
651 652

        Returns type:
Z
Zhou Wei 已提交
653
            ndarray: dtype is same as current Tensor
654 655 656 657

        Examples:
            .. code-block:: python

Z
Zhou Wei 已提交
658
                import paddle
659 660
                import numpy as np
                data = np.random.uniform(-1, 1, [30, 10, 32]).astype('float32')
Z
Zhou Wei 已提交
661 662 663 664
                linear = paddle.nn.Linear(32, 64)
                data = paddle.to_tensor(data)
                x = linear(data)
                print(x.numpy())
665 666 667

       )DOC")
      .def("detach",
Z
Zhou Wei 已提交
668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715
           [](const imperative::VarBase
                  &self) -> std::shared_ptr<imperative::VarBase> {
             PADDLE_ENFORCE_EQ(
                 self.Var().IsInitialized(), true,
                 platform::errors::InvalidArgument(
                     "Tensor %s has not been initialized!", self.Name()));
             PADDLE_ENFORCE_EQ(
                 self.Var().IsType<framework::LoDTensor>() ||
                     self.Var().IsType<framework::SelectedRows>(),
                 true,
                 platform::errors::InvalidArgument(
                     "Type of Tensor[%s] must be LoDTensor or SelectedRows!",
                     self.Name()));
             auto detach_var = std::make_shared<imperative::VarBase>(
                 true, "detach_" + self.Name());
             detach_var->SetPersistable(self.Persistable());
             detach_var->SetType(self.Type());
             detach_var->SetDataType(self.DataType());
             if (self.Var().IsType<framework::LoDTensor>()) {
               const auto &origin_tensor =
                   self.Var().Get<framework::LoDTensor>();
               PADDLE_ENFORCE_EQ(
                   origin_tensor.IsInitialized(), true,
                   platform::errors::InvalidArgument(
                       "Tensor %s has not been initialized!", self.Name()));

               auto *detach_tensor =
                   detach_var->MutableVar()->GetMutable<framework::LoDTensor>();
               detach_tensor->ShareDataWith(origin_tensor);
             } else {
               const auto &origin_selected_rows =
                   self.Var().Get<framework::SelectedRows>();
               PADDLE_ENFORCE_EQ(
                   origin_selected_rows.value().IsInitialized(), true,
                   platform::errors::InvalidArgument(
                       "Tensor %s has not been initialized!", self.Name()));

               auto *detach_selected_rows =
                   detach_var->MutableVar()
                       ->GetMutable<framework::SelectedRows>();
               detach_selected_rows->set_height(origin_selected_rows.height());
               detach_selected_rows->set_rows(origin_selected_rows.rows());
               detach_selected_rows->mutable_value()->ShareDataWith(
                   origin_selected_rows.value());
             }
             VLOG(3) << "The detached Tensor(" << detach_var->Name()
                     << ") share data with " << self.Name();
             return detach_var;
716
           },
Z
Zhou Wei 已提交
717
           py::return_value_policy::take_ownership, R"DOC(
718

719
        Returns a new Tensor, detached from the current graph.
Z
Zhou Wei 已提交
720 721
        It will share data with origin Tensor and always doesn't have a Tensor copy.
        In addition, the detached Tensor doesn't provide gradient propagation.
722

723
        Returns: The detached Tensor.
724 725 726 727

        Examples:
            .. code-block:: python

728
                import paddle
Z
Zhou Wei 已提交
729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753

                x = paddle.to_tensor(1.0, stop_gradient=False)
                detach_x = x.detach()
                detach_x[:] = 10.0
                print(x)  # Tensor(shape=[1], dtype=float32, place=CPUPlace, stop_gradient=False,
                          #        [10.])
                y = x**2
                y.backward()
                print(x.grad)         # [20.0]
                print(detach_x.grad)  # None, 'stop_gradient=True' by default

                detach_x.stop_gradient = False # Set stop_gradient to be False, supported auto-grad
                z = detach_x**3
                z.backward()

                print(x.grad)         # [20.0], detach_x is detached from x's graph, not affect each other
                print(detach_x.grad)  # [300.0], detach_x has its own graph

                # Due to sharing of data with origin Tensor, There are some unsafe operations:
                y = 2 * x
                detach_x[:] = 5.0
                y.backward() 
                # It will raise Error:
                #   one of the variables needed for gradient computation has been modified by an inplace operation.
             
754 755 756
       )DOC")
      .def("clear_gradient", &imperative::VarBase::ClearGradient, R"DOC(

757
        Only for Tensor that has gradient, normally we use this for Parameters since other temporary Tensor doesen't has gradient.
758

759
        The Gradient of current Tensor will be set to ``0`` .
760 761 762 763 764 765

        Returns:  None

        Examples:
             .. code-block:: python

766
                import paddle
Z
Zhou Wei 已提交
767 768 769 770 771 772 773
                input = paddle.uniform([10, 2])
                linear = paddle.nn.Linear(2, 3)
                out = linear(input)
                out.backward()
                print("Before clear_gradient, linear.weight.grad: {}".format(linear.weight.grad))
                linear.weight.clear_gradient()
                print("After clear_gradient, linear.weight.grad: {}".format(linear.weight.grad))
774
      )DOC")
Z
Zhou Wei 已提交
775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822
      .def("clone",
           [](std::shared_ptr<imperative::VarBase> &self) {
             const auto &tensor = self->Var().Get<framework::LoDTensor>();
             PADDLE_ENFORCE_EQ(
                 tensor.IsInitialized(), true,
                 platform::errors::InvalidArgument(
                     "%s has not been initialized", self->Name()));
             auto tracer = imperative::GetCurrentTracer();
             auto new_var = std::make_shared<imperative::VarBase>(
                 true, tracer->GenerateUniqueName(self->Name() + "_clone"));
             framework::AttributeMap attrs;
             imperative::NameVarBaseMap ins = {{"X", {self}}};
             imperative::NameVarBaseMap outs = {{"Out", {new_var}}};
             tracer->TraceOp("assign", ins, outs, attrs);
             return new_var;
           },
           py::return_value_policy::copy, R"DOC(

        Returns a new Tensor, which is clone of origin Tensor, and it remains in the current graph.
        It will always have a Tensor copy.
        Tn addition, the cloned Tensor provides gradient propagation.

        Returns: The cloned Tensor.

        Examples:
            .. code-block:: python

              import paddle

              x = paddle.to_tensor(1.0, stop_gradient=False)
              clone_x = x.clone()
              y = clone_x**2
              y.backward()
              print(clone_x.stop_gradient) # False
              print(clone_x.grad)          # [2.0], support gradient propagation
              print(x.stop_gradient)       # False
              print(x.grad)                # [2.0], clone_x support gradient propagation for x

              x = paddle.to_tensor(1.0)
              clone_x = x.clone()
              clone_x.stop_gradient = False
              z = clone_x**3
              z.backward()
              print(clone_x.stop_gradient) # False
              print(clone_x.grad)          # [3.0], support gradient propagation
              print(x.stop_gradient) # True
              print(x.grad)          # None
       )DOC")
L
Leo Chen 已提交
823
      .def("_run_backward",
824 825
           [](imperative::VarBase &self, const imperative::Tracer &tracer,
              bool retain_graph) {
826 827
             // TODO(jiabin): when we impl more backward execution we can
             // select them
828
             auto *engine = tracer.GetEngine();
829
             engine->Init(&self, retain_graph);
830
             VLOG(3) << "Start backward";
L
Leo Chen 已提交
831 832 833 834 835 836 837 838 839 840
             engine->Execute();
             VLOG(3) << "Finish backward";
           },
           py::call_guard<py::gil_scoped_release>())
      .def("_grad_name", &imperative::VarBase::GradVarName)
      .def("_grad_value",
           [](imperative::VarBase &self) {
             return self.MutableGradVar()->Get<framework::LoDTensor>();
           },
           py::return_value_policy::reference)
841 842 843 844
      .def("_set_grad_type",
           [](imperative::VarBase &self, framework::proto::VarType::Type type) {
             self.MutableGradVarBase()->SetType(type);
           })
845
      .def("_grad_ivar",
J
Jiabin Yang 已提交
846 847
           [](const imperative::VarBase &self) {
             auto &grad_var = self.GradVarBase();
848 849 850 851 852 853 854 855 856 857 858
             if (grad_var && grad_var->Var().IsInitialized()) {
               auto *tensor =
                   grad_var->MutableVar()->IsType<framework::LoDTensor>()
                       ? grad_var->MutableVar()
                             ->GetMutable<framework::LoDTensor>()
                       : grad_var->MutableVar()
                             ->GetMutable<framework::SelectedRows>()
                             ->mutable_value();
               if (tensor->IsInitialized()) {
                 return grad_var;
               }
J
Jiabin Yang 已提交
859
             }
860
             return std::shared_ptr<imperative::VarBase>(nullptr);
J
Jiabin Yang 已提交
861 862
           },
           py::return_value_policy::copy)
863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
      .def("_is_sparse",
           [](imperative::VarBase &self) {
             return self.Var().IsType<framework::SelectedRows>();
           })
      .def("_allreduce",
           [](imperative::VarBase &self,
              const imperative::ParallelStrategy &strategy) {
             if (strategy.nranks_ > 1) {
#ifdef PADDLE_WITH_NCCL
#if NCCL_VERSION_CODE >= 2212
               imperative::AllReduce(self.Var(), self.MutableVar(), strategy);
#else
               if (!self.Var().IsType<framework::SelectedRows>()) {
                 imperative::AllReduce(self.Var(), self.MutableVar(), strategy);
               } else {
                 PADDLE_THROW(platform::errors::Unimplemented(
                     "Imperative SelectedRows allreduce is not supported when "
                     "paddle is compiled with NCCL verison lower than v2.2.12. "
                     "You can set is_sparse=False for the Layer containing "
                     "this argument, such as Embedding(is_sparse=False)."));
               }
#endif  // NCCL_VERSION_CODE
#else
               PADDLE_THROW(platform::errors::Unimplemented(
                   "Imperative allreduce is not supported when paddle is "
                   "not compiled with NCCL."));
#endif  // PADDLE_WITH_NCCL
             }
           },
           py::call_guard<py::gil_scoped_release>())
893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013
      .def("cpu",
           [](const std::shared_ptr<imperative::VarBase> &self) {
             if (platform::is_cpu_place(self->Place())) {
               return self;
             } else {
               auto new_var = self->NewVarBase(platform::CPUPlace(), true);
               new_var->SetOverridedStopGradient(self->OverridedStopGradient());
               return new_var;
             }
           },
           R"DOC(
        Returns a copy of this Tensor in CPU memory.

        If this Tensor is already in CPU memory, then no copy is performed and the original Tensor is returned.

        Examples:
            .. code-block:: python

              import paddle
              x = paddle.to_tensor(1.0, place=paddle.CUDAPlace(0))
              print(x.place)    # CUDAPlace(0)
              
              y = x.cpu()
              print(y.place)    # CPUPlace

              )DOC")
      .def("pin_memory",
           [](const std::shared_ptr<imperative::VarBase> &self) {
#ifndef PADDLE_WITH_CUDA
             PADDLE_THROW(platform::errors::PermissionDenied(
                 "Cannot copy this Tensor to pinned memory in CPU version "
                 "Paddle, "
                 "Please recompile or reinstall Paddle with CUDA support."));
#endif
             if (platform::is_cuda_pinned_place(self->Place())) {
               return self;
             } else {
               auto new_var =
                   self->NewVarBase(platform::CUDAPinnedPlace(), true);
               new_var->SetOverridedStopGradient(self->OverridedStopGradient());
               return new_var;
             }
           },
           R"DOC(
        Returns a copy of this Tensor in pin memory.

        If this Tensor is already in pin memory, then no copy is performed and the original Tensor is returned.

        Examples:
            .. code-block:: python

              import paddle
              x = paddle.to_tensor(1.0, place=paddle.CUDAPlace(0))
              print(x.place)      # CUDAPlace(0)

              y = x.pin_memory()
              print(y.place)      # CUDAPinnedPlace

      )DOC")
      .def("cuda",
           [](const std::shared_ptr<imperative::VarBase> &self, int device_id,
              bool blocking) {
#ifndef PADDLE_WITH_CUDA
             PADDLE_THROW(platform::errors::PermissionDenied(
                 "Cannot copy this Tensor to GPU in CPU version Paddle, "
                 "Please recompile or reinstall Paddle with CUDA support."));
#else
             int device_count = platform::GetCUDADeviceCount();
             if (device_id == -1) {
               if (platform::is_gpu_place(self->Place())) {
                 return self;
               } else {
                 device_id = 0;
               }
             }
             PADDLE_ENFORCE_GE(
                 device_id, 0,
                 platform::errors::InvalidArgument(
                     "Can not copy Tensor to Invalid CUDAPlace(%d), device id "
                     "must inside [0, %d)",
                     device_id, device_count));
             PADDLE_ENFORCE_LT(
                 device_id, device_count,
                 platform::errors::InvalidArgument(
                     "Can not copy Tensor to Invalid CUDAPlace(%d), device id "
                     "must inside [0, %d)",
                     device_id, device_count));
             platform::CUDAPlace place = platform::CUDAPlace(device_id);
             if (platform::is_same_place(self->Place(), place)) {
               return self;
             } else {
               auto new_var = self->NewVarBase(place, blocking);
               new_var->SetOverridedStopGradient(self->OverridedStopGradient());
               return new_var;
             }
#endif
           },
           py::arg("device_id") = -1, py::arg("blocking") = true, R"DOC(
        Returns a copy of this Tensor in GPU memory.

        If this Tensor is already in GPU memory and device_id is default, 
        then no copy is performed and the original Tensor is returned.
        
        Args:
            device_id(int, optional): The destination GPU device id. Defaults to the current device.
            blocking(bool, optional): If False and the source is in pinned memory, the copy will be 
              asynchronous with respect to the host. Otherwise, the argument has no effect. Default: False.

        Examples:
            .. code-block:: python

              import paddle
              x = paddle.to_tensor(1.0, place=paddle.CPUPlace())
              print(x.place)        # CPUPlace

              y = x.cuda()
              print(y.place)        # CUDAPlace(0)

              y = x.cuda(1)
              print(y.place)        # CUDAPlace(1)
       )DOC")
1014 1015
      .def("_copy_to",
           [](const imperative::VarBase &self, const platform::CPUPlace &place,
J
Jiabin Yang 已提交
1016 1017
              bool blocking) { return self.NewVarBase(place, blocking); },
           py::return_value_policy::copy)
1018 1019 1020 1021 1022
      .def("_copy_to",
           [](const imperative::VarBase &self,
              const platform::CUDAPinnedPlace &place,
              bool blocking) { return self.NewVarBase(place, blocking); },
           py::return_value_policy::copy)
1023 1024 1025 1026
      .def("_copy_to",
           [](const imperative::VarBase &self, const platform::XPUPlace &place,
              bool blocking) { return self.NewVarBase(place, blocking); },
           py::return_value_policy::copy)
1027 1028
      .def("_copy_to",
           [](const imperative::VarBase &self, const platform::CUDAPlace &place,
J
Jiabin Yang 已提交
1029 1030 1031
              bool blocking) { return self.NewVarBase(place, blocking); },
           py::return_value_policy::copy)
      .def("value", [](imperative::VarBase &self) { return self.MutableVar(); },
1032 1033 1034
           py::return_value_policy::reference)
      .def_property("name", &imperative::VarBase::Name,
                    &imperative::VarBase::SetName)
L
Leo Chen 已提交
1035 1036 1037 1038 1039
      .def_property("stop_gradient",
                    &imperative::VarBase::OverridedStopGradient,
                    &imperative::VarBase::SetOverridedStopGradient)
      .def_property("persistable", &imperative::VarBase::Persistable,
                    &imperative::VarBase::SetPersistable)
J
Jiabin Yang 已提交
1040 1041 1042 1043
      .def_property_readonly(
          "shape",
          [](imperative::VarBase &self) {
            if (self.Var().IsType<framework::LoDTensor>()) {
1044
              return framework::vectorize<int>(
J
Jiabin Yang 已提交
1045
                  self.Var().Get<framework::LoDTensor>().dims());
1046 1047 1048
            } else if (self.Var().IsType<framework::SelectedRows>()) {
              return framework::vectorize<int>(
                  self.Var().Get<framework::SelectedRows>().value().dims());
J
Jiabin Yang 已提交
1049
            } else {
1050 1051
              VLOG(2) << "It is meaningless to get shape of "
                         "variable type "
J
Jiabin Yang 已提交
1052 1053 1054 1055
                      << GetTypeName(self);
              return std::vector<int>();
            }
          })
1056 1057 1058
      .def_property_readonly(
          "place", [](imperative::VarBase &self) { return self.Place(); },
          py::return_value_policy::copy)
1059 1060 1061 1062 1063 1064
      .def_property_readonly("_place_str",
                             [](imperative::VarBase &self) {
                               std::stringstream ostr;
                               ostr << self.Place();
                               return ostr.str();
                             })
J
Jiabin Yang 已提交
1065
      .def_property_readonly("type", &imperative::VarBase::Type)
L
Leo Chen 已提交
1066
      .def_property_readonly("dtype", &imperative::VarBase::DataType);
1067 1068 1069

  py::class_<imperative::Layer, Layer /* <--- trampoline*/> layer(m, "Layer");
  layer.def(py::init<>())
1070 1071 1072 1073 1074
      .def("forward",
           [](imperative::Layer &self,
              const std::vector<std::shared_ptr<imperative::VarBase>> &inputs) {
             return self.Forward(inputs);
           });
1075

1076 1077 1078 1079 1080
  py::class_<imperative::jit::ProgramDescTracer>(m, "ProgramDescTracer", "")
      .def("create_program_desc",
           &imperative::jit::ProgramDescTracer::CreateProgramDesc)
      .def("reset", &imperative::jit::ProgramDescTracer::Reset);

1081
  py::class_<imperative::Tracer, std::shared_ptr<imperative::Tracer>>(
1082
      m, "Tracer", R"DOC()DOC")
1083
      .def("__init__",
J
Jiabin Yang 已提交
1084
           [](imperative::Tracer &self) { new (&self) imperative::Tracer(); })
1085 1086 1087
      .def_property("_enable_program_desc_tracing",
                    &imperative::Tracer::IsProgramDescTracingEnabled,
                    &imperative::Tracer::SetEnableProgramDescTracing)
1088 1089
      .def_property("_enable_autocast", &imperative::Tracer::IsAutoCastEnabled,
                    &imperative::Tracer::SetEnableAutoCast)
1090
      .def_property("_has_grad", &imperative::Tracer::HasGrad,
1091
                    &imperative::Tracer::SetHasGrad)
1092 1093 1094 1095 1096 1097 1098 1099
      .def_property(
          "_expected_place",
          [](const imperative::Tracer &self) -> py::object {
            return py::cast(self.ExpectedPlace());
          },
          [](imperative::Tracer &self, const py::object &obj) {
            if (py::isinstance<platform::CUDAPlace>(obj)) {
              auto p = obj.cast<platform::CUDAPlace *>();
L
Leo Chen 已提交
1100
              self.SetExpectedPlace(*p);
1101 1102 1103
            } else if (py::isinstance<platform::XPUPlace>(obj)) {
              auto p = obj.cast<platform::XPUPlace *>();
              self.SetExpectedPlace(*p);
1104 1105
            } else if (py::isinstance<platform::CPUPlace>(obj)) {
              auto p = obj.cast<platform::CPUPlace *>();
L
Leo Chen 已提交
1106
              self.SetExpectedPlace(*p);
1107 1108
            } else if (py::isinstance<platform::CUDAPinnedPlace>(obj)) {
              auto p = obj.cast<platform::CUDAPinnedPlace *>();
L
Leo Chen 已提交
1109
              self.SetExpectedPlace(*p);
1110
            } else {
L
Leo Chen 已提交
1111
              PADDLE_THROW(platform::errors::InvalidArgument(
1112 1113
                  "Incompatible Place Type: supports XPUPlace, CUDAPlace, "
                  "CPUPlace, "
L
Leo Chen 已提交
1114 1115
                  "and CUDAPinnedPlace, "
                  "but got Unknown Type!"));
1116 1117
            }
          })
1118 1119 1120
      .def("_get_program_desc_tracer",
           &imperative::Tracer::GetProgramDescTracer,
           py::return_value_policy::reference)
1121
      .def("_generate_unique_name", &imperative::Tracer::GenerateUniqueName,
1122
           py::arg("key") = "dygraph_tmp")
1123 1124 1125 1126 1127
      .def(
          "_set_amp_op_list",
          [](imperative::Tracer &self,
             std::unordered_set<std::string> &allow_ops,
             std::unordered_set<std::string> &block_ops) {
1128 1129
            // NOTE(zhiqiu): The automatic conversion in pybind11 between
            // c++
1130 1131 1132 1133
            // STL and python set/list/dict involve a copy operation that
            // prevents pass-by-reference semantics, so it is ok to swap.
            // The reaseon why not directly pass
            // std::shared_ptr<std::unordered_set<std::string>>
1134 1135
            // is that pybind11 forbid shared_ptr<T> where T is not custom
            // type.
1136 1137 1138 1139 1140 1141 1142 1143 1144
            imperative::AmpOperators::Instance().GetAllowOps()->swap(allow_ops);
            imperative::AmpOperators::Instance().GetBlockOps()->swap(block_ops);
          })
      .def("_get_amp_op_list",
           [](imperative::Tracer &self) {
             return std::make_tuple(
                 *(imperative::AmpOperators::Instance().GetAllowOps()),
                 *(imperative::AmpOperators::Instance().GetBlockOps()));
           })
1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157
      .def("trace",
           [](imperative::Tracer &self, const std::string &type,
              const PyNameVarBaseMap &ins, const PyNameVarBaseMap &outs,
              framework::AttributeMap attrs, const platform::XPUPlace &place,
              bool trace_backward) {
             auto ins_map = ConvertToNameVarBaseMap(ins);
             auto outs_map = ConvertToNameVarBaseMap(outs);
             {
               py::gil_scoped_release release;
               self.TraceOp(type, std::move(ins_map), std::move(outs_map),
                            std::move(attrs), place, trace_backward);
             }
           })
M
minqiyang 已提交
1158
      .def("trace",
J
Jiabin Yang 已提交
1159 1160 1161 1162 1163 1164
           [](imperative::Tracer &self, const std::string &type,
              const PyNameVarBaseMap &ins, const PyNameVarBaseMap &outs,
              framework::AttributeMap attrs, const platform::CUDAPlace &place,
              bool trace_backward) {
             auto ins_map = ConvertToNameVarBaseMap(ins);
             auto outs_map = ConvertToNameVarBaseMap(outs);
1165 1166
             {
               py::gil_scoped_release release;
J
Jiabin Yang 已提交
1167 1168
               self.TraceOp(type, std::move(ins_map), std::move(outs_map),
                            std::move(attrs), place, trace_backward);
1169
             }
M
minqiyang 已提交
1170
           })
J
Jiabin Yang 已提交
1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183
      .def("trace",
           [](imperative::Tracer &self, const std::string &type,
              const PyNameVarBaseMap &ins, const PyNameVarBaseMap &outs,
              framework::AttributeMap attrs, const platform::CPUPlace &place,
              bool trace_backward) {
             auto ins_map = ConvertToNameVarBaseMap(ins);
             auto outs_map = ConvertToNameVarBaseMap(outs);
             {
               py::gil_scoped_release release;
               self.TraceOp(type, std::move(ins_map), std::move(outs_map),
                            std::move(attrs), place, trace_backward);
             }
           });
1184 1185

  // define parallel context
1186 1187 1188
  py::class_<imperative::ParallelStrategy> parallel_strategy(
      m, "ParallelStrategy", "");
  parallel_strategy.def(py::init())
1189 1190
      .def_property(
          "nranks",
1191 1192
          [](const imperative::ParallelStrategy &self) { return self.nranks_; },
          [](imperative::ParallelStrategy &self, int nranks) {
1193 1194 1195
            self.nranks_ = nranks;
          })
      .def_property("local_rank",
1196
                    [](const imperative::ParallelStrategy &self) {
1197 1198
                      return self.local_rank_;
                    },
1199
                    [](imperative::ParallelStrategy &self, int local_rank) {
1200 1201 1202 1203
                      self.local_rank_ = local_rank;
                    })
      .def_property(
          "trainer_endpoints",
1204
          [](const imperative::ParallelStrategy &self) {
1205 1206
            return self.trainer_endpoints_;
          },
1207
          [](imperative::ParallelStrategy &self, std::vector<std::string> eps) {
1208 1209 1210
            self.trainer_endpoints_ = eps;
          })
      .def_property("current_endpoint",
1211
                    [](const imperative::ParallelStrategy &self) {
1212 1213
                      return self.current_endpoint_;
                    },
1214 1215
                    [](imperative::ParallelStrategy &self,
                       const std::string &ep) { self.current_endpoint_ = ep; });
1216 1217 1218 1219 1220 1221 1222 1223

  m.def(
      "dygraph_partial_grad",
      [](const std::vector<std::shared_ptr<imperative::VarBase>> &input_targets,
         const std::vector<std::shared_ptr<imperative::VarBase>>
             &output_targets,
         const std::vector<std::shared_ptr<imperative::VarBase>> &output_grads,
         const std::vector<std::shared_ptr<imperative::VarBase>> &no_grad_vars,
1224 1225
         const platform::Place &place, bool create_graph, bool retain_graph,
         bool allow_unused, bool only_inputs) {
Z
Zeng Jinle 已提交
1226 1227
        imperative::PartialGradEngine engine(
            input_targets, output_targets, output_grads, no_grad_vars, place,
1228
            create_graph, retain_graph, allow_unused, only_inputs);
1229 1230 1231 1232 1233
        engine.Execute();
        return engine.GetResult();
      },
      py::call_guard<py::gil_scoped_release>());

1234
#if defined(PADDLE_WITH_NCCL)
1235 1236
  py::class_<imperative::NCCLParallelContext> nccl_ctx(m,
                                                       "NCCLParallelContext");
1237 1238

  nccl_ctx
1239 1240 1241
      .def(py::init<const imperative::ParallelStrategy &,
                    const platform::CUDAPlace &>())
      .def("init", [](imperative::NCCLParallelContext &self) { self.Init(); });
1242
#endif
1243 1244 1245 1246
}

}  // namespace pybind
}  // namespace paddle