inference_api.cc 23.1 KB
Newer Older
F
flame 已提交
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/inference_api.h"
16
#include <pybind11/numpy.h>
F
flame 已提交
17 18
#include <pybind11/stl.h>
#include <cstring>
19
#include <functional>
F
flame 已提交
20
#include <iostream>
21
#include <iterator>
22
#include <map>
23
#include <memory>
F
flame 已提交
24
#include <string>
25
#include <type_traits>
26
#include <unordered_set>
27
#include <utility>
F
flame 已提交
28 29
#include <vector>
#include "paddle/fluid/inference/api/analysis_predictor.h"
30
#include "paddle/fluid/inference/api/helper.h"
F
flame 已提交
31
#include "paddle/fluid/inference/api/paddle_inference_api.h"
32
#include "paddle/fluid/inference/api/paddle_pass_builder.h"
33
#include "paddle/fluid/inference/utils/io_utils.h"
F
flame 已提交
34 35 36 37 38

namespace py = pybind11;

namespace paddle {
namespace pybind {
39 40 41
using paddle::AnalysisPredictor;
using paddle::NativeConfig;
using paddle::NativePaddlePredictor;
F
flame 已提交
42
using paddle::PaddleBuf;
43 44
using paddle::PaddleDType;
using paddle::PaddlePassBuilder;
F
flame 已提交
45 46
using paddle::PaddlePlace;
using paddle::PaddlePredictor;
47 48 49
using paddle::PaddleTensor;
using paddle::PassStrategy;
using paddle::ZeroCopyTensor;
F
flame 已提交
50

51 52 53 54 55 56 57 58 59 60
namespace {
void BindPaddleDType(py::module *m);
void BindPaddleBuf(py::module *m);
void BindPaddleTensor(py::module *m);
void BindPaddlePlace(py::module *m);
void BindPaddlePredictor(py::module *m);
void BindNativeConfig(py::module *m);
void BindNativePredictor(py::module *m);
void BindAnalysisConfig(py::module *m);
void BindAnalysisPredictor(py::module *m);
61 62
void BindZeroCopyTensor(py::module *m);
void BindPaddlePassBuilder(py::module *m);
F
flame 已提交
63

64
#ifdef PADDLE_WITH_MKLDNN
65
void BindMkldnnQuantizerConfig(py::module *m);
66
#endif
67 68 69 70

template <typename T>
PaddleBuf PaddleBufCreate(py::array_t<T> data) {
  PaddleBuf buf(data.size() * sizeof(T));
71
  std::copy_n(static_cast<const T *>(data.data()), data.size(),
72 73 74 75 76 77 78
              static_cast<T *>(buf.data()));
  return buf;
}

template <typename T>
void PaddleBufReset(PaddleBuf &buf, py::array_t<T> data) {  // NOLINT
  buf.Resize(data.size() * sizeof(T));
79
  std::copy_n(static_cast<const T *>(data.data()), data.size(),
80 81 82 83 84 85 86 87 88 89 90
              static_cast<T *>(buf.data()));
}

template <typename T>
PaddleTensor PaddleTensorCreate(
    py::array_t<T> data, const std::string name = "",
    const std::vector<std::vector<size_t>> &lod = {}, bool copy = true) {
  PaddleTensor tensor;

  if (copy) {
    PaddleBuf buf(data.size() * sizeof(T));
91
    std::copy_n(static_cast<const T *>(data.data()), data.size(),
92 93 94 95 96 97
                static_cast<T *>(buf.data()));
    tensor.data = std::move(buf);
  } else {
    tensor.data = PaddleBuf(data.mutable_data(), data.size() * sizeof(T));
  }

98
  tensor.dtype = inference::PaddleTensorGetDType<T>();
99 100 101 102 103 104 105 106
  tensor.name = name;
  tensor.lod = lod;
  tensor.shape.resize(data.ndim());
  std::copy_n(data.shape(), data.ndim(), tensor.shape.begin());

  return tensor;
}

107
py::dtype PaddleDTypeToNumpyDType(PaddleDType dtype) {
108
  py::dtype dt;
109
  switch (dtype) {
110 111 112 113 114 115 116 117 118 119
    case PaddleDType::INT32:
      dt = py::dtype::of<int32_t>();
      break;
    case PaddleDType::INT64:
      dt = py::dtype::of<int64_t>();
      break;
    case PaddleDType::FLOAT32:
      dt = py::dtype::of<float>();
      break;
    default:
120 121 122
      PADDLE_THROW(platform::errors::Unimplemented(
          "Unsupported data type. Now only supports INT32, INT64 and "
          "FLOAT32."));
123
  }
124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154

  return dt;
}

py::array PaddleTensorGetData(PaddleTensor &tensor) {  // NOLINT
  py::dtype dt = PaddleDTypeToNumpyDType(tensor.dtype);
  return py::array(std::move(dt), {tensor.shape}, tensor.data.data());
}

template <typename T>
void ZeroCopyTensorCreate(ZeroCopyTensor &tensor,  // NOLINT
                          py::array_t<T> data) {
  std::vector<int> shape;
  std::copy_n(data.shape(), data.ndim(), std::back_inserter(shape));
  tensor.Reshape(std::move(shape));
  tensor.copy_from_cpu(static_cast<const T *>(data.data()));
}

size_t PaddleGetDTypeSize(PaddleDType dt) {
  size_t size{0};
  switch (dt) {
    case PaddleDType::INT32:
      size = sizeof(int32_t);
      break;
    case PaddleDType::INT64:
      size = sizeof(int64_t);
      break;
    case PaddleDType::FLOAT32:
      size = sizeof(float);
      break;
    default:
155 156 157
      PADDLE_THROW(platform::errors::Unimplemented(
          "Unsupported data type. Now only supports INT32, INT64 and "
          "FLOAT32."));
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178
  }
  return size;
}

py::array ZeroCopyTensorToNumpy(ZeroCopyTensor &tensor) {  // NOLINT
  py::dtype dt = PaddleDTypeToNumpyDType(tensor.type());
  auto tensor_shape = tensor.shape();
  py::array::ShapeContainer shape(tensor_shape.begin(), tensor_shape.end());
  py::array array(dt, std::move(shape));

  switch (tensor.type()) {
    case PaddleDType::INT32:
      tensor.copy_to_cpu(static_cast<int32_t *>(array.mutable_data()));
      break;
    case PaddleDType::INT64:
      tensor.copy_to_cpu(static_cast<int64_t *>(array.mutable_data()));
      break;
    case PaddleDType::FLOAT32:
      tensor.copy_to_cpu<float>(static_cast<float *>(array.mutable_data()));
      break;
    default:
179 180 181
      PADDLE_THROW(platform::errors::Unimplemented(
          "Unsupported data type. Now only supports INT32, INT64 and "
          "FLOAT32."));
182 183
  }
  return array;
184
}
185 186 187 188 189 190

py::bytes SerializePDTensorToBytes(PaddleTensor &tensor) {  // NOLINT
  std::stringstream ss;
  paddle::inference::SerializePDTensorToStream(&ss, tensor);
  return static_cast<py::bytes>(ss.str());
}
191
}  // namespace
192

F
flame 已提交
193 194 195 196 197 198 199 200 201 202
void BindInferenceApi(py::module *m) {
  BindPaddleDType(m);
  BindPaddleBuf(m);
  BindPaddleTensor(m);
  BindPaddlePlace(m);
  BindPaddlePredictor(m);
  BindNativeConfig(m);
  BindNativePredictor(m);
  BindAnalysisConfig(m);
  BindAnalysisPredictor(m);
203 204
  BindZeroCopyTensor(m);
  BindPaddlePassBuilder(m);
205 206 207
#ifdef PADDLE_WITH_MKLDNN
  BindMkldnnQuantizerConfig(m);
#endif
F
flame 已提交
208 209 210 211 212
  m->def("create_paddle_predictor",
         &paddle::CreatePaddlePredictor<AnalysisConfig>);
  m->def("create_paddle_predictor",
         &paddle::CreatePaddlePredictor<NativeConfig>);
  m->def("paddle_dtype_size", &paddle::PaddleDtypeSize);
213
  m->def("paddle_tensor_to_bytes", &SerializePDTensorToBytes);
F
flame 已提交
214 215
}

216
namespace {
F
flame 已提交
217 218 219
void BindPaddleDType(py::module *m) {
  py::enum_<PaddleDType>(*m, "PaddleDType")
      .value("FLOAT32", PaddleDType::FLOAT32)
220 221
      .value("INT64", PaddleDType::INT64)
      .value("INT32", PaddleDType::INT32);
F
flame 已提交
222 223 224 225 226 227 228 229
}

void BindPaddleBuf(py::module *m) {
  py::class_<PaddleBuf>(*m, "PaddleBuf")
      .def(py::init<size_t>())
      .def(py::init([](std::vector<float> &data) {
        auto buf = PaddleBuf(data.size() * sizeof(float));
        std::memcpy(buf.data(), static_cast<void *>(data.data()), buf.length());
G
Gabor Buella 已提交
230
        return buf;
F
flame 已提交
231
      }))
232 233 234
      .def(py::init(&PaddleBufCreate<int32_t>))
      .def(py::init(&PaddleBufCreate<int64_t>))
      .def(py::init(&PaddleBufCreate<float>))
F
flame 已提交
235 236 237 238 239 240
      .def("resize", &PaddleBuf::Resize)
      .def("reset",
           [](PaddleBuf &self, std::vector<float> &data) {
             self.Resize(data.size() * sizeof(float));
             std::memcpy(self.data(), data.data(), self.length());
           })
241 242 243
      .def("reset", &PaddleBufReset<int32_t>)
      .def("reset", &PaddleBufReset<int64_t>)
      .def("reset", &PaddleBufReset<float>)
244
      .def("empty", &PaddleBuf::empty)
245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
      .def("tolist",
           [](PaddleBuf &self, const std::string &dtype) -> py::list {
             py::list l;
             if (dtype == "int32") {
               auto *data = static_cast<int32_t *>(self.data());
               auto size = self.length() / sizeof(int32_t);
               l = py::cast(std::vector<int32_t>(data, data + size));
             } else if (dtype == "int64") {
               auto *data = static_cast<int64_t *>(self.data());
               auto size = self.length() / sizeof(int64_t);
               l = py::cast(std::vector<int64_t>(data, data + size));
             } else if (dtype == "float32") {
               auto *data = static_cast<float *>(self.data());
               auto size = self.length() / sizeof(float);
               l = py::cast(std::vector<float>(data, data + size));
             } else {
261 262 263
               PADDLE_THROW(platform::errors::Unimplemented(
                   "Unsupported data type. Now only supports INT32, INT64 and "
                   "FLOAT32."));
264 265 266
             }
             return l;
           })
F
flame 已提交
267 268 269 270 271 272 273 274 275 276
      .def("float_data",
           [](PaddleBuf &self) -> std::vector<float> {
             auto *data = static_cast<float *>(self.data());
             return {data, data + self.length() / sizeof(*data)};
           })
      .def("int64_data",
           [](PaddleBuf &self) -> std::vector<int64_t> {
             int64_t *data = static_cast<int64_t *>(self.data());
             return {data, data + self.length() / sizeof(*data)};
           })
277 278 279 280
      .def("int32_data",
           [](PaddleBuf &self) -> std::vector<int32_t> {
             int32_t *data = static_cast<int32_t *>(self.data());
             return {data, data + self.length() / sizeof(*data)};
F
flame 已提交
281 282 283 284 285 286 287
           })
      .def("length", &PaddleBuf::length);
}

void BindPaddleTensor(py::module *m) {
  py::class_<PaddleTensor>(*m, "PaddleTensor")
      .def(py::init<>())
288 289 290 291 292 293 294 295 296 297 298 299 300
      .def(py::init(&PaddleTensorCreate<int32_t>), py::arg("data"),
           py::arg("name") = "",
           py::arg("lod") = std::vector<std::vector<size_t>>(),
           py::arg("copy") = true)
      .def(py::init(&PaddleTensorCreate<int64_t>), py::arg("data"),
           py::arg("name") = "",
           py::arg("lod") = std::vector<std::vector<size_t>>(),
           py::arg("copy") = true)
      .def(py::init(&PaddleTensorCreate<float>), py::arg("data"),
           py::arg("name") = "",
           py::arg("lod") = std::vector<std::vector<size_t>>(),
           py::arg("copy") = true)
      .def("as_ndarray", &PaddleTensorGetData)
F
flame 已提交
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
      .def_readwrite("name", &PaddleTensor::name)
      .def_readwrite("shape", &PaddleTensor::shape)
      .def_readwrite("data", &PaddleTensor::data)
      .def_readwrite("dtype", &PaddleTensor::dtype)
      .def_readwrite("lod", &PaddleTensor::lod);
}

void BindPaddlePlace(py::module *m) {
  py::enum_<PaddlePlace>(*m, "PaddlePlace")
      .value("UNK", PaddlePlace::kUNK)
      .value("CPU", PaddlePlace::kCPU)
      .value("GPU", PaddlePlace::kGPU);
}

void BindPaddlePredictor(py::module *m) {
  auto paddle_predictor = py::class_<PaddlePredictor>(*m, "PaddlePredictor");
  paddle_predictor
      .def("run",
           [](PaddlePredictor &self, const std::vector<PaddleTensor> &inputs) {
             std::vector<PaddleTensor> outputs;
             self.Run(inputs, &outputs);
             return outputs;
           })
      .def("get_input_tensor", &PaddlePredictor::GetInputTensor)
      .def("get_output_tensor", &PaddlePredictor::GetOutputTensor)
326 327
      .def("get_input_names", &PaddlePredictor::GetInputNames)
      .def("get_output_names", &PaddlePredictor::GetOutputNames)
F
flame 已提交
328
      .def("zero_copy_run", &PaddlePredictor::ZeroCopyRun)
329 330
      .def("clone", &PaddlePredictor::Clone)
      .def("get_serialized_program", &PaddlePredictor::GetSerializedProgram);
F
flame 已提交
331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373

  auto config = py::class_<PaddlePredictor::Config>(paddle_predictor, "Config");
  config.def(py::init<>())
      .def_readwrite("model_dir", &PaddlePredictor::Config::model_dir);
}

void BindNativeConfig(py::module *m) {
  py::class_<NativeConfig, PaddlePredictor::Config>(*m, "NativeConfig")
      .def(py::init<>())
      .def_readwrite("use_gpu", &NativeConfig::use_gpu)
      .def_readwrite("device", &NativeConfig::device)
      .def_readwrite("fraction_of_gpu_memory",
                     &NativeConfig::fraction_of_gpu_memory)
      .def_readwrite("prog_file", &NativeConfig::prog_file)
      .def_readwrite("param_file", &NativeConfig::param_file)
      .def_readwrite("specify_input_name", &NativeConfig::specify_input_name)
      .def("set_cpu_math_library_num_threads",
           &NativeConfig::SetCpuMathLibraryNumThreads)
      .def("cpu_math_library_num_threads",
           &NativeConfig::cpu_math_library_num_threads);
}

void BindNativePredictor(py::module *m) {
  py::class_<NativePaddlePredictor, PaddlePredictor>(*m,
                                                     "NativePaddlePredictor")
      .def(py::init<const NativeConfig &>())
      .def("init", &NativePaddlePredictor::Init)
      .def("run",
           [](NativePaddlePredictor &self,
              const std::vector<PaddleTensor> &inputs) {
             std::vector<PaddleTensor> outputs;
             self.Run(inputs, &outputs);
             return outputs;
           })
      .def("get_input_tensor", &NativePaddlePredictor::GetInputTensor)
      .def("get_output_tensor", &NativePaddlePredictor::GetOutputTensor)
      .def("zero_copy_run", &NativePaddlePredictor::ZeroCopyRun)
      .def("clone", &NativePaddlePredictor::Clone)
      .def("scope", &NativePaddlePredictor::scope,
           py::return_value_policy::reference);
}

void BindAnalysisConfig(py::module *m) {
374 375 376 377 378
  py::class_<AnalysisConfig> analysis_config(*m, "AnalysisConfig");

  py::enum_<AnalysisConfig::Precision>(analysis_config, "Precision")
      .value("Float32", AnalysisConfig::Precision::kFloat32)
      .value("Int8", AnalysisConfig::Precision::kInt8)
Z
Zhaolong Xing 已提交
379
      .value("Half", AnalysisConfig::Precision::kHalf)
380 381
      .export_values();

382 383
  analysis_config.def(py::init<>())
      .def(py::init<const AnalysisConfig &>())
F
flame 已提交
384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
      .def(py::init<const std::string &>())
      .def(py::init<const std::string &, const std::string &>())
      .def("set_model", (void (AnalysisConfig::*)(const std::string &)) &
                            AnalysisConfig::SetModel)
      .def("set_model", (void (AnalysisConfig::*)(const std::string &,
                                                  const std::string &)) &
                            AnalysisConfig::SetModel)
      .def("set_prog_file", &AnalysisConfig::SetProgFile)
      .def("set_params_file", &AnalysisConfig::SetParamsFile)
      .def("model_dir", &AnalysisConfig::model_dir)
      .def("prog_file", &AnalysisConfig::prog_file)
      .def("params_file", &AnalysisConfig::params_file)
      .def("enable_use_gpu", &AnalysisConfig::EnableUseGpu,
           py::arg("memory_pool_init_size_mb"), py::arg("device_id") = 0)
      .def("disable_gpu", &AnalysisConfig::DisableGpu)
      .def("use_gpu", &AnalysisConfig::use_gpu)
      .def("gpu_device_id", &AnalysisConfig::gpu_device_id)
      .def("memory_pool_init_size_mb",
           &AnalysisConfig::memory_pool_init_size_mb)
      .def("fraction_of_gpu_memory_for_pool",
           &AnalysisConfig::fraction_of_gpu_memory_for_pool)
      .def("switch_ir_optim", &AnalysisConfig::SwitchIrOptim,
           py::arg("x") = true)
      .def("ir_optim", &AnalysisConfig::ir_optim)
408
      .def("enable_memory_optim", &AnalysisConfig::EnableMemoryOptim)
409
      .def("enable_profile", &AnalysisConfig::EnableProfile)
410
      .def("disable_glog_info", &AnalysisConfig::DisableGlogInfo)
411
      .def("glog_info_disabled", &AnalysisConfig::glog_info_disabled)
412
      .def("set_optim_cache_dir", &AnalysisConfig::SetOptimCacheDir)
F
flame 已提交
413 414 415 416 417 418 419 420 421
      .def("switch_use_feed_fetch_ops", &AnalysisConfig::SwitchUseFeedFetchOps,
           py::arg("x") = true)
      .def("use_feed_fetch_ops_enabled",
           &AnalysisConfig::use_feed_fetch_ops_enabled)
      .def("switch_specify_input_names",
           &AnalysisConfig::SwitchSpecifyInputNames, py::arg("x") = true)
      .def("specify_input_name", &AnalysisConfig::specify_input_name)
      .def("enable_tensorrt_engine", &AnalysisConfig::EnableTensorRtEngine,
           py::arg("workspace_size") = 1 << 20, py::arg("max_batch_size") = 1,
422
           py::arg("min_subgraph_size") = 3,
N
nhzlx 已提交
423
           py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32,
424 425 426
           py::arg("use_static") = false, py::arg("use_calib_mode") = true)
      .def("set_trt_dynamic_shape_info",
           &AnalysisConfig::SetTRTDynamicShapeInfo,
427 428 429 430 431
           py::arg("min_input_shape") =
               std::map<std::string, std::vector<int>>({}),
           py::arg("max_input_shape") =
               std::map<std::string, std::vector<int>>({}),
           py::arg("optim_input_shape") =
432 433
               std::map<std::string, std::vector<int>>({}),
           py::arg("disable_trt_plugin_fp16") = false)
F
flame 已提交
434
      .def("tensorrt_engine_enabled", &AnalysisConfig::tensorrt_engine_enabled)
435 436 437 438 439
      .def("enable_lite_engine", &AnalysisConfig::EnableLiteEngine,
           py::arg("precision_mode") = AnalysisConfig::Precision::kFloat32,
           py::arg("passes_filter") = std::vector<std::string>(),
           py::arg("ops_filter") = std::vector<std::string>())
      .def("lite_engine_enabled", &AnalysisConfig::lite_engine_enabled)
F
flame 已提交
440 441 442 443 444 445 446 447 448
      .def("switch_ir_debug", &AnalysisConfig::SwitchIrDebug,
           py::arg("x") = true)
      .def("enable_mkldnn", &AnalysisConfig::EnableMKLDNN)
      .def("mkldnn_enabled", &AnalysisConfig::mkldnn_enabled)
      .def("set_cpu_math_library_num_threads",
           &AnalysisConfig::SetCpuMathLibraryNumThreads)
      .def("cpu_math_library_num_threads",
           &AnalysisConfig::cpu_math_library_num_threads)
      .def("to_native_config", &AnalysisConfig::ToNativeConfig)
449 450 451 452 453
      .def("enable_quantizer", &AnalysisConfig::EnableMkldnnQuantizer)
#ifdef PADDLE_WITH_MKLDNN
      .def("quantizer_config", &AnalysisConfig::mkldnn_quantizer_config,
           py::return_value_policy::reference)
#endif
F
flame 已提交
454 455 456
      .def("set_mkldnn_op", &AnalysisConfig::SetMKLDNNOp)
      .def("set_model_buffer", &AnalysisConfig::SetModelBuffer)
      .def("model_from_memory", &AnalysisConfig::model_from_memory)
457 458 459 460
      .def("delete_pass",
           [](AnalysisConfig &self, const std::string &pass) {
             self.pass_builder()->DeletePass(pass);
           })
F
flame 已提交
461 462 463 464
      .def("pass_builder", &AnalysisConfig::pass_builder,
           py::return_value_policy::reference);
}

465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486
#ifdef PADDLE_WITH_MKLDNN
void BindMkldnnQuantizerConfig(py::module *m) {
  py::class_<MkldnnQuantizerConfig> quantizer_config(*m,
                                                     "MkldnnQuantizerConfig");
  quantizer_config.def(py::init<const MkldnnQuantizerConfig &>())
      .def(py::init<>())
      .def("set_quant_data",
           [](MkldnnQuantizerConfig &self,
              const std::vector<PaddleTensor> &data) {
             auto warmup_data =
                 std::make_shared<std::vector<PaddleTensor>>(data);
             self.SetWarmupData(warmup_data);
             return;
           })
      .def("set_quant_batch_size", &MkldnnQuantizerConfig::SetWarmupBatchSize)
      .def(
          "set_enabled_op_types",
          (void (MkldnnQuantizerConfig::*)(std::unordered_set<std::string> &)) &
              MkldnnQuantizerConfig::SetEnabledOpTypes);
}
#endif

F
flame 已提交
487 488 489 490 491 492 493 494 495 496 497 498 499
void BindAnalysisPredictor(py::module *m) {
  py::class_<AnalysisPredictor, PaddlePredictor>(*m, "AnalysisPredictor")
      .def(py::init<const AnalysisConfig &>())
      .def("init", &AnalysisPredictor::Init)
      .def(
          "run",
          [](AnalysisPredictor &self, const std::vector<PaddleTensor> &inputs) {
            std::vector<PaddleTensor> outputs;
            self.Run(inputs, &outputs);
            return outputs;
          })
      .def("get_input_tensor", &AnalysisPredictor::GetInputTensor)
      .def("get_output_tensor", &AnalysisPredictor::GetOutputTensor)
500 501 502
      .def("get_input_names", &AnalysisPredictor::GetInputNames)
      .def("get_output_names", &AnalysisPredictor::GetOutputNames)
      .def("get_input_tensor_shape", &AnalysisPredictor::GetInputTensorShape)
F
flame 已提交
503
      .def("zero_copy_run", &AnalysisPredictor::ZeroCopyRun)
504 505 506 507 508 509 510
      .def("create_feed_fetch_var", &AnalysisPredictor::CreateFeedFetchVar)
      .def("prepare_feed_fetch", &AnalysisPredictor::PrepareFeedFetch)
      .def("prepare_argument", &AnalysisPredictor::PrepareArgument)
      .def("optimize_inference_program",
           &AnalysisPredictor::OptimizeInferenceProgram)
      .def("analysis_argument", &AnalysisPredictor::analysis_argument,
           py::return_value_policy::reference)
F
flame 已提交
511 512
      .def("clone", &AnalysisPredictor::Clone)
      .def("scope", &AnalysisPredictor::scope,
513
           py::return_value_policy::reference)
514 515 516 517
      .def("program", &AnalysisPredictor::program,
           py::return_value_policy::reference)
      .def("get_serialized_program", &AnalysisPredictor::GetSerializedProgram)
      .def("mkldnn_quantize", &AnalysisPredictor::MkldnnQuantize)
518 519
      .def("SaveOptimModel", &AnalysisPredictor::SaveOptimModel,
           py::arg("dir"));
F
flame 已提交
520
}
521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578

void BindZeroCopyTensor(py::module *m) {
  py::class_<ZeroCopyTensor>(*m, "ZeroCopyTensor")
      .def("reshape", &ZeroCopyTensor::Reshape)
      .def("copy_from_cpu", &ZeroCopyTensorCreate<int32_t>)
      .def("copy_from_cpu", &ZeroCopyTensorCreate<int64_t>)
      .def("copy_from_cpu", &ZeroCopyTensorCreate<float>)
      .def("copy_to_cpu", &ZeroCopyTensorToNumpy)
      .def("shape", &ZeroCopyTensor::shape)
      .def("set_lod", &ZeroCopyTensor::SetLoD)
      .def("lod", &ZeroCopyTensor::lod)
      .def("type", &ZeroCopyTensor::type);
}

void BindPaddlePassBuilder(py::module *m) {
  py::class_<PaddlePassBuilder>(*m, "PaddlePassBuilder")
      .def(py::init<const std::vector<std::string> &>())
      .def("set_passes",
           [](PaddlePassBuilder &self, const std::vector<std::string> &passes) {
             self.ClearPasses();
             for (auto pass : passes) {
               self.AppendPass(std::move(pass));
             }
           })
      .def("append_pass", &PaddlePassBuilder::AppendPass)
      .def("insert_pass", &PaddlePassBuilder::InsertPass)
      .def("delete_pass",
           [](PaddlePassBuilder &self, const std::string &pass_type) {
             self.DeletePass(pass_type);
           })
      .def("append_analysis_pass", &PaddlePassBuilder::AppendAnalysisPass)
      .def("turn_on_debug", &PaddlePassBuilder::TurnOnDebug)
      .def("debug_string", &PaddlePassBuilder::DebugString)
      .def("all_passes", &PaddlePassBuilder::AllPasses,
           py::return_value_policy::reference)
      .def("analysis_passes", &PaddlePassBuilder::AnalysisPasses);

  py::class_<PassStrategy, PaddlePassBuilder>(*m, "PassStrategy")
      .def(py::init<const std::vector<std::string> &>())
      .def("enable_cudnn", &PassStrategy::EnableCUDNN)
      .def("enable_mkldnn", &PassStrategy::EnableMKLDNN)
      .def("enable_mkldnn_quantizer", &PassStrategy::EnableMkldnnQuantizer)
      .def("use_gpu", &PassStrategy::use_gpu);

  py::class_<CpuPassStrategy, PassStrategy>(*m, "CpuPassStrategy")
      .def(py::init<>())
      .def(py::init<const CpuPassStrategy &>())
      .def("enable_cudnn", &CpuPassStrategy::EnableCUDNN)
      .def("enable_mkldnn", &CpuPassStrategy::EnableMKLDNN)
      .def("enable_mkldnn_quantizer", &CpuPassStrategy::EnableMkldnnQuantizer);

  py::class_<GpuPassStrategy, PassStrategy>(*m, "GpuPassStrategy")
      .def(py::init<>())
      .def(py::init<const GpuPassStrategy &>())
      .def("enable_cudnn", &GpuPassStrategy::EnableCUDNN)
      .def("enable_mkldnn", &GpuPassStrategy::EnableMKLDNN)
      .def("enable_mkldnn_quantizer", &GpuPassStrategy::EnableMkldnnQuantizer);
}
579
}  // namespace
F
flame 已提交
580 581
}  // namespace pybind
}  // namespace paddle