io.cpp 13.8 KB
Newer Older
W
wangliu 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
/* 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 "io.h"
#include <fstream>
#include <vector>
#include "common/log.h"
L
liuruilong 已提交
19 20 21

#include "common/enforce.h"
#include "framework/framework.pb-c.h"
L
liuruilong 已提交
22 23
#include "framework/lod_tensor.h"
#include "framework/operator.h"
L
liuruilong 已提交
24
#include "framework/program/program-optimize/program_optimize.h"
L
liuruilong 已提交
25 26 27 28
#include "framework/program/program_desc.h"
#include "framework/program/var_desc.h"
#include "framework/scope.h"
#include "framework/tensor.h"
W
wangliu 已提交
29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170

namespace paddle_mobile {
using framework::Variable;

void ReadBinaryFile(const std::string &filename, std::string *contents) {
  std::ifstream fin(filename, std::ios::in | std::ios::binary);
  PADDLE_MOBILE_ENFORCE(fin.is_open(), "open file: %s failed",
                        filename.c_str());
  fin.seekg(0, std::ios::end);
  contents->clear();
  contents->resize(fin.tellg());
  fin.seekg(0, std::ios::beg);
  fin.read(&(contents->at(0)), contents->size());
  fin.close();
}

static size_t ReadBuffer(const char *file_name, uint8_t **out) {
  printf("%s \n", file_name);
  FILE *fp;
  fp = fopen(file_name, "rb");
  PADDLE_MOBILE_ENFORCE(fp != NULL, " %s open failed !", file_name);

  fseek(fp, 0, SEEK_END);
  size_t size = ftell(fp);
  rewind(fp);

  DLOG << "model size: " << size;

  *out = reinterpret_cast<uint8_t *>(malloc(size));

  size_t cur_len = 0;
  size_t nread;
  while ((nread = fread(*out + cur_len, 1, size - cur_len, fp)) != 0) {
    cur_len += nread;
  }
  fclose(fp);
  return cur_len;
}

template <typename Dtype, Precision P>
void Loader<Dtype, P>::LoadVar(framework::Variable *variable,
                               const framework::VarDesc &var_desc,
                               const std::string &file_path) {
  auto tensor = variable->GetMutable<framework::LoDTensor>();
  std::ifstream is(file_path);
  PADDLE_MOBILE_ENFORCE(is.is_open(), "open file: %s failed",
                        file_path.c_str());

  std::fpos<mbstate_t> pos;
  pos = is.tellg();  // save   current   position
  is.seekg(0, std::ios::end);
  is.seekg(pos);  // restore   saved   position

  // 1. version
  uint32_t version;
  is.read(reinterpret_cast<char *>(&version), sizeof(version));

  // 2 Lod information
  uint64_t lod_level;
  is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
  auto &lod = *tensor->mutable_lod();
  lod.resize(lod_level);
  for (uint64_t i = 0; i < lod_level; ++i) {
    uint64_t size;
    is.read(reinterpret_cast<char *>(&size), sizeof(size));
    std::vector<size_t> tmp(size / sizeof(size_t));
    is.read(reinterpret_cast<char *>(tmp.data()),
            static_cast<std::streamsize>(size));
    for (auto j : tmp) {
      LOG(kLOG_DEBUG1) << "    lod - " << j;
    }
    lod[i] = tmp;
  }

  // 3. tensor version
  uint32_t tensor_version;
  is.read(reinterpret_cast<char *>(&tensor_version), sizeof(tensor_version));

  // 4. tensor desc
  int32_t size;
  is.read(reinterpret_cast<char *>(&size), sizeof(size));
  std::unique_ptr<char[]> buf(new char[size]);
  is.read(reinterpret_cast<char *>(buf.get()), size);

  const framework::TensorDesc &desc = var_desc.Tensor_desc();

  PaddleMobile__Framework__Proto__VarType__TensorDesc *tensor_desc = NULL;
  //  void *v;
  //  PaddleMobile__Framework__Proto__VarType__TensorDesc_Closure()(tensor_desc,
  //  buf.get());

  //  DLOG << "PaddleMobile__Framework__Proto__VarType__TensorDesc_Closure- " <<
  //  tensor_desc;

  //  framework::TensorDesc &tensor_desc = variable->
  //  PaddleMobile__Framework__Proto__ProgramDesc *c_program;
  //  uint8_t *proto_buf = NULL;
  //  size_t read_size = ReadBuffer(file_path.c_str(), &proto_buf);
  //  c_program = paddle_mobile__framework__proto__program_desc__unpack(NULL,
  //  read_size, buf);

  //  paddle_mobile__framework__proto__var_type__tensor_desc__init()

  int memory_size = 1;
  for (auto l : desc.Dims()) {
    memory_size *= l;
  }

  tensor->Resize(framework::make_ddim(desc.Dims()));

  void *memory = tensor;
  int type_size = 0;
  switch (desc.DataType()) {
    case framework::VARTYPE_TYPE_FP16:
      type_size = 2;
      break;
    case framework::VARTYPE_TYPE_FP32:
      type_size = 4;
      memory = tensor->mutable_data<float>();
      break;
    case framework::VARTYPE_TYPE_FP64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_INT32:
      type_size = 4;
      break;
    case framework::VARTYPE_TYPE_INT64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_BOOL:
      type_size = 1;
      break;
    default:
      break;
  }

  is.read(static_cast<char *>(memory), memory_size * type_size);
  is.close();
}

template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
L
liuruilong 已提交
171
    const std::string &dirname, bool optimize) {
W
wangliu 已提交
172 173 174 175 176 177 178 179 180
  std::string model_filename = dirname + "/__model__";
  PaddleMobile__Framework__Proto__ProgramDesc *c_program;
  uint8_t *buf = NULL;
  size_t read_size = ReadBuffer(model_filename.c_str(), &buf);

  PADDLE_MOBILE_ENFORCE(buf != NULL, "read from __model__ is null");

  c_program = paddle_mobile__framework__proto__program_desc__unpack(
      NULL, read_size, buf);
W
wangliu 已提交
181
  //
W
wangliu 已提交
182
  PADDLE_MOBILE_ENFORCE(c_program != NULL, "program is null");
W
wangliu 已提交
183
  //
W
wangliu 已提交
184
  DLOG << "n_ops: " << (*c_program->blocks)->n_ops;
W
wangliu 已提交
185
  //
W
wangliu 已提交
186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
  std::shared_ptr<framework::ProgramDesc> originProgramDesc =
      std::make_shared<framework::ProgramDesc>(c_program);

  framework::Program<Dtype, P> program;
  program.model_path = dirname;
  program.originProgram = originProgramDesc;

  std::shared_ptr<framework::Scope> scope =
      std::make_shared<framework::Scope>();
  program.scope = scope;
  originProgramDesc->Block(0);

  for (const auto &block : originProgramDesc->Blocks()) {
    for (int i = 0; i < block->Vars().size(); ++i) {
      std::shared_ptr<framework::VarDesc> var_desc = block->Vars()[i];
      //      DLOG << "var name-- " << var_desc->Name();
      auto var = scope->Var(var_desc->Name());

      if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
        if (var_desc->Persistable() &&
            var_desc->Type() != framework::VARTYPE_TYPE_FEED_MINIBATCH &&
            var_desc->Type() != framework::VARTYPE_TYPE_FETCH_LIST) {
          auto dim = var_desc->Tensor_desc().Dims();
          auto tensor = var->GetMutable<framework::LoDTensor>();
          tensor->Resize(framework::make_ddim(dim));
        } else {
          auto dim = var_desc->Tensor_desc().Dims();
          PADDLE_MOBILE_ENFORCE(dim.size() > 0, "dim size is 0");
          dim[0] = 1;
          auto tensor = var->GetMutable<framework::LoDTensor>();
          tensor->Resize(framework::make_ddim(dim));
        }
      } else {
        // TODO(codeWorm): some.
      }
    }
  }

L
liuruilong 已提交
224 225
  //  originProgramDesc->Description("program: ");

L
liuruilong 已提交
226 227
  if (optimize) {
    framework::ProgramOptimize program_optimize;
L
liuruilong 已提交
228 229
    program.optimizeProgram =
        program_optimize.FushionOptimize(originProgramDesc);
L
liuruilong 已提交
230
  }
L
liuruilong 已提交
231 232 233 234 235 236
  if (optimize) {
    program.optimizeProgram->Description("optimize: ");
  } else {
    originProgramDesc->Description("program: ");
  }

W
wangliu 已提交
237 238 239 240 241 242 243 244 245
  paddle_mobile__framework__proto__program_desc__free_unpacked(c_program, NULL);
  return program;
}

template class Loader<CPU, Precision::FP32>;

#pragma mark - executor

template <typename Dtype, Precision P>
L
liuruilong 已提交
246 247
Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
                             bool use_optimize)
L
liuruilong 已提交
248
    : program_(p), batch_size_(batch_size), use_optimize_(use_optimize) {
W
wangliu 已提交
249 250 251 252 253 254 255 256 257 258 259 260 261 262
  if (use_optimize_) {
    to_predict_program_ = program_.optimizeProgram;
  } else {
    to_predict_program_ = program_.originProgram;
  }
  Variable *variable_ptr = program_.scope->Var("batch_size");
  variable_ptr[0].SetValue<int>(batch_size);
  const std::vector<std::shared_ptr<framework::BlockDesc>> blocks =
      to_predict_program_->Blocks();
  for (int i = 0; i < blocks.size(); ++i) {
    std::shared_ptr<framework::BlockDesc> block_desc = blocks[i];
    std::vector<std::shared_ptr<framework::OpDesc>> ops = block_desc->Ops();
    for (int j = 0; j < ops.size(); ++j) {
      std::shared_ptr<framework::OpDesc> op = ops[j];
L
liuruilong 已提交
263
      DLOG << "create op: " << op->Type();
W
wangliu 已提交
264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
      auto op_base = framework::OpRegistry<Dtype>::CreateOp(
          op->Type(), op->GetInputs(), op->GetOutputs(), op->GetAttrMap(),
          program_.scope);
      op_base->InferShape();

      ops_of_block_[*block_desc.get()].push_back(op_base);
    }
  }
  InitMemory();
}

template <typename Dtype, Precision P>
void Executor<Dtype, P>::LoadMemory(const framework::VarDesc var_desc,
                                    framework::LoDTensor *tensor,
                                    const std::string &file_path) {
  std::ifstream is(file_path);
  PADDLE_MOBILE_ENFORCE(is.is_open(), "open file: %s failed",
                        file_path.c_str());
  std::fpos<mbstate_t> pos;
  pos = is.tellg();  // save   current   position
  is.seekg(0, std::ios::end);
  is.seekg(pos);  // restore   saved   position

  // 1. version
  uint32_t version;
  is.read(reinterpret_cast<char *>(&version), sizeof(version));

  // 2 Lod information
  uint64_t lod_level;
  is.read(reinterpret_cast<char *>(&lod_level), sizeof(lod_level));
  auto &lod = *tensor->mutable_lod();
  lod.resize(lod_level);
  for (uint64_t i = 0; i < lod_level; ++i) {
    uint64_t size;
    is.read(reinterpret_cast<char *>(&size), sizeof(size));
    std::vector<size_t> tmp(size / sizeof(size_t));
    is.read(reinterpret_cast<char *>(tmp.data()),
            static_cast<std::streamsize>(size));
    for (auto j : tmp) {
      LOG(kLOG_DEBUG1) << "    lod - " << j;
    }
    lod[i] = tmp;
  }

  // 3. tensor version
  uint32_t tensor_version;
  is.read(reinterpret_cast<char *>(&tensor_version), sizeof(tensor_version));

  // 4. tensor desc
  int32_t size;
  is.read(reinterpret_cast<char *>(&size), sizeof(size));
  std::unique_ptr<char[]> buf(new char[size]);
  is.read(reinterpret_cast<char *>(buf.get()), size);

  const framework::TensorDesc &desc = var_desc.Tensor_desc();

  int memory_size = 1;
  for (auto l : desc.Dims()) {
    memory_size *= l;
  }

  tensor->Resize(framework::make_ddim(desc.Dims()));

  void *memory = tensor;
  int type_size = 0;
  switch (desc.DataType()) {
    case framework::VARTYPE_TYPE_FP16:
      type_size = 2;
      break;
    case framework::VARTYPE_TYPE_FP32:
      type_size = 4;
      memory = tensor->mutable_data<float>();
      break;
    case framework::VARTYPE_TYPE_FP64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_INT32:
      type_size = 4;
      break;
    case framework::VARTYPE_TYPE_INT64:
      type_size = 8;
      break;
    case framework::VARTYPE_TYPE_BOOL:
      type_size = 1;
      break;
    default:
      break;
  }

  is.read(static_cast<char *>(memory), memory_size * type_size);
  is.close();
}

template <typename Dtype, Precision P>
void Executor<Dtype, P>::InitMemory() {
  for (const auto &block : to_predict_program_->Blocks()) {
    for (const auto &var_desc : block->Vars()) {
      auto var = program_.scope->Var(var_desc->Name());
      if (var_desc->Persistable()) {
        auto tensor = var->template GetMutable<framework::LoDTensor>();
        if (var_desc->Name() == "feed" || var_desc->Name() == "fetch") {
          continue;
        }
        LoadMemory(*var_desc, tensor,
                   program_.model_path + "/" + var_desc->Name());
      } else {
        if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
          auto tensor = var->template GetMutable<framework::LoDTensor>();

          tensor->template mutable_data<Ptype>();
        }
      }
    }
  }
}

template <typename Dtype, Precision P>
W
wangliu 已提交
381 382
std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
    const framework::Tensor &t) {
W
wangliu 已提交
383 384 385 386 387 388
  framework::Variable *g_feed_value = program_.scope->Var("feed");
  framework::Tensor *feed_tensor =
      g_feed_value->GetMutable<framework::LoDTensor>();
  feed_tensor->Resize(t.dims());
  feed_tensor->ShareDataWith(t);
  std::shared_ptr<framework::BlockDesc> to_predict_block =
W
wangliu 已提交
389
      to_predict_program_->Block(0);
W
wangliu 已提交
390 391 392 393
  for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
    auto op = ops_of_block_[*to_predict_block.get()][j];
    op->Run();
  }
W
wangliu 已提交
394 395 396 397 398 399 400 401 402 403 404 405 406 407
  auto ops = ops_of_block_[*to_predict_program_->Block(0)];
  auto last_op = ops.rbegin();
  auto output_map = (*last_op)->Outputs();
  std::vector<std::string> out_keys = (*last_op)->GetOutKeys();
  PADDLE_MOBILE_ENFORCE(out_keys.size() > 0, "the last op contains no output");
  framework::LoDTensor *output_tensor =
      framework::GetVarValue<framework::LoDTensor>(out_keys[0], output_map,
                                                   *(program_.scope));
  return std::shared_ptr<framework::Tensor>(output_tensor);
}
template <typename Dtype, Precision P>
std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
    const framework::Tensor &t, int block_id) {
  return Predict(t);
W
wangliu 已提交
408 409 410
}

template <typename Dtype, Precision P>
L
liuruilong 已提交
411
std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
W
wangliu 已提交
412 413
    const std::vector<Ptype> &input, const std::vector<int64_t> &dims) {
  framework::Tensor tensor(input, framework::make_ddim(dims));
W
wangliu 已提交
414 415 416 417 418 419 420 421
  std::shared_ptr<framework::Tensor> output_tensor = Predict(tensor, 0);
  Executor<Dtype, P>::Ptype *output_ptr =
      output_tensor->data<typename Executor<Dtype, P>::Ptype>();
  std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
  for (int j = 0; j < output_tensor->numel(); ++j) {
    result_vector.push_back(output_ptr[j]);
  }
  return result_vector;
W
wangliu 已提交
422 423 424 425 426
}

template class Executor<CPU, Precision::FP32>;

}  // namespace paddle_mobile