analysis_predictor.cc 18.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

Y
Yan Chunwei 已提交
15
#include "paddle/fluid/inference/api/analysis_predictor.h"
16 17
#include <glog/logging.h>
#include <algorithm>
18
#include <memory>
19 20
#include <string>
#include <vector>
21
#include "paddle/fluid/framework/feed_fetch_method.h"
22
#include "paddle/fluid/framework/feed_fetch_type.h"
Y
Yan Chunwei 已提交
23
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
24
#include "paddle/fluid/framework/ir/pass.h"
25
#include "paddle/fluid/framework/naive_executor.h"
26
#include "paddle/fluid/framework/scope.h"
27
#include "paddle/fluid/inference/api/helper.h"
28
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
29
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
30 31 32
#if PADDLE_WITH_TENSORRT
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#endif
33
#include "paddle/fluid/inference/utils/singleton.h"
34
#include "paddle/fluid/platform/cpu_helper.h"
T
tensor-tang 已提交
35 36 37
#include "paddle/fluid/platform/profiler.h"

DECLARE_bool(profile);
38
DECLARE_int32(paddle_num_threads);
39 40 41

namespace paddle {

42 43
using contrib::AnalysisConfig;

44 45 46 47 48 49 50 51 52 53 54
namespace {
bool IsPersistable(const framework::VarDesc *var) {
  if (var->Persistable() &&
      var->GetType() != framework::proto::VarType::FEED_MINIBATCH &&
      var->GetType() != framework::proto::VarType::FETCH_LIST) {
    return true;
  }
  return false;
}
}  // namespace

Y
Yan Chunwei 已提交
55
bool AnalysisPredictor::Init(
56 57
    const std::shared_ptr<framework::Scope> &parent_scope,
    const std::shared_ptr<framework::ProgramDesc> &program) {
58
  VLOG(30) << "Predictor::init()";
T
tensor-tang 已提交
59 60 61 62 63 64 65 66 67 68
#if !defined(_WIN32)
  if (FLAGS_profile) {
    LOG(WARNING) << "Profiler is actived, might affect the performance";
    LOG(INFO) << "You can turn off by set gflags '-profile false'";
    auto tracking_device = config_.use_gpu ? platform::ProfilerState::kAll
                                           : platform::ProfilerState::kCPU;
    platform::EnableProfiler(tracking_device);
  }
#endif

69 70 71
  // no matter with or without MKLDNN
  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);

72 73 74 75 76 77 78 79 80 81 82 83 84
  if (!PrepareScope(parent_scope)) {
    return false;
  }
  if (!CreateExecutor()) {
    return false;
  }
  if (!PrepareProgram(program)) {
    return false;
  }

  // Prepare executor, create local variables.
  if (!PrepareExecutor()) {
    return true;
Y
Yan Chunwei 已提交
85
  }
86 87 88 89 90 91 92 93 94

  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();

  return true;
}

bool AnalysisPredictor::PrepareScope(
    const std::shared_ptr<framework::Scope> &parent_scope) {
Y
Yan Chunwei 已提交
95
  if (parent_scope) {
96 97 98
    PADDLE_ENFORCE_NOT_NULL(
        parent_scope,
        "Both program and parent_scope should be set in Clone mode.");
Y
Yan Chunwei 已提交
99
    scope_ = parent_scope;
100
    status_is_cloned_ = true;
Y
Yan Chunwei 已提交
101 102 103
  } else {
    paddle::framework::InitDevices(false);
    scope_.reset(new paddle::framework::Scope());
104
    status_is_cloned_ = false;
Y
Yan Chunwei 已提交
105
  }
106 107 108 109 110
  sub_scope_ = &scope_->NewScope();
  return true;
}
bool AnalysisPredictor::PrepareProgram(
    const std::shared_ptr<framework::ProgramDesc> &program) {
111 112
  if (!program) {
    if (!LoadProgramDesc()) return false;
113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133

    // Optimize the program, and load parameters and modify them in the
    // scope_.
    // This will change the scope_ address.
    if (config_.enable_ir_optim) {
      status_ir_optim_enabled_ = true;
      OptimizeInferenceProgram();
    } else {
      // If the parent_scope is passed, we assert that the persistable variables
      // are already created, so just create the no persistable variables.

      // If not cloned, the parameters should be loaded
      // OptimizeInferenceProgram.
      // So in both cases, just the local variables are needed to load, not the
      // parematers.
      executor_->CreateVariables(*inference_program_, 0, true, sub_scope_);

      // Load parameters
      LOG(INFO) << "load parameters ";
      LoadParameters();
    }
Y
Yan Chunwei 已提交
134
  } else {
135 136
    // If the program is passed from external, no need to optimize it, this
    // logic is used in the clone scenario.
137 138
    inference_program_ = program;
  }
M
Michal Gallus 已提交
139

140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
  executor_->CreateVariables(*inference_program_, 0, false, sub_scope_);

  return true;
}
bool AnalysisPredictor::CreateExecutor() {
  if (config_.use_gpu) {
    status_use_gpu_ = true;
    place_ = paddle::platform::CUDAPlace(config_.device);
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  executor_.reset(new paddle::framework::NaiveExecutor(place_));
  return true;
}
bool AnalysisPredictor::PrepareExecutor() {
  executor_->Prepare(sub_scope_, *inference_program_, 0,
156 157
                     config_.use_feed_fetch_ops);

158
  PADDLE_ENFORCE_NOT_NULL(sub_scope_);
Y
Yan Chunwei 已提交
159

160 161 162 163 164 165
  return true;
}

bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                            std::vector<PaddleTensor> *output_data,
                            int batch_size) {
166
  VLOG(30) << "Predictor::predict";
167 168 169 170 171 172 173
  inference::Timer timer;
  timer.tic();
  // set feed variable
  std::vector<framework::LoDTensor> feeds;
  framework::Scope *scope = sub_scope_ ? sub_scope_ : scope_.get();
  if (!SetFeed(inputs, scope)) {
    LOG(ERROR) << "fail to set feed";
Y
Yan Chunwei 已提交
174
    return false;
175
  }
M
Michal Gallus 已提交
176

177 178 179
  // Run the inference program
  // if share variables, we need not create variables
  executor_->Run();
180

181 182 183 184
  // get fetch variable
  if (!GetFetch(output_data, scope)) {
    LOG(ERROR) << "fail to get fetches";
    return false;
T
tensor-tang 已提交
185
  }
186
  VLOG(30) << "predict cost: " << timer.toc() << "ms";
Y
Yan Chunwei 已提交
187 188 189 190

  // Fix TensorArray reuse not cleaned bug.
  tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
  tensor_array_batch_cleaner_.ResetTensorArray();
191 192
  return true;
}
193

194 195
bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                framework::Scope *scope) {
196
  VLOG(30) << "Predictor::set_feed";
197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
  if (inputs.size() != feeds_.size()) {
    LOG(ERROR) << "wrong feed input size, need " << feeds_.size() << " but get "
               << inputs.size();
    return false;
  }

  // Cache the inputs memory for better concurrency performance.
  feed_tensors_.resize(inputs.size());

  for (size_t i = 0; i < inputs.size(); ++i) {
    auto &input = feed_tensors_[i];
    framework::DDim ddim = framework::make_ddim(inputs[i].shape);
    void *input_ptr;
    if (inputs[i].dtype == PaddleDType::INT64) {
      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
    } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
    } else {
      LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
      return false;
    }

    // TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
    std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
                inputs[i].data.length());
    // TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
    framework::LoD lod;
    for (auto &level : inputs[i].lod) {
      lod.emplace_back(level);
    }
    input.set_lod(lod);
    int idx = -1;
    if (config_.specify_input_name) {
      idx = feed_names_[inputs[i].name];
    } else {
      idx = boost::get<int>(feeds_[i]->GetAttr("col"));
    }
    framework::SetFeedVariable(scope, input, "feed", idx);
  }
  return true;
}

template <typename T>
void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch,
                                    PaddleTensor *output) {
  // set shape.
  auto shape = framework::vectorize(fetch.dims());
  output->shape.assign(shape.begin(), shape.end());
  // set data.
  const T *data = fetch.data<T>();
  int num_elems = inference::VecReduceToInt(shape);
  output->data.Resize(num_elems * sizeof(T));
  // The fetched tensor output by fetch op, should always in CPU memory, so just
  // copy.
  memcpy(output->data.data(), data, num_elems * sizeof(T));
  // set lod
  output->lod.clear();
  for (auto &level : fetch.lod()) {
    output->lod.emplace_back(level.begin(), level.end());
  }
}

bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
                                 framework::Scope *scope) {
261
  VLOG(30) << "Predictor::get_fetch";
262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
  outputs->resize(fetchs_.size());
  for (size_t i = 0; i < fetchs_.size(); ++i) {
    int idx = boost::get<int>(fetchs_[i]->GetAttr("col"));
    PADDLE_ENFORCE((size_t)idx == i);
    framework::LoDTensor &fetch =
        framework::GetFetchVariable(*scope, "fetch", idx);
    auto type = fetch.type();
    auto output = &(outputs->at(i));
    if (type == typeid(float)) {
      GetFetchOne<float>(fetch, output);
      output->dtype = PaddleDType::FLOAT32;
    } else if (type == typeid(int64_t)) {
      GetFetchOne<int64_t>(fetch, output);
      output->dtype = PaddleDType::INT64;
    } else {
      LOG(ERROR) << "unknown type, only support float32 and int64 now.";
    }
  }
Y
Yan Chunwei 已提交
280 281
  return true;
}
282

283
// NOTE All the members in AnalysisConfig should be copied to Argument.
Y
Yan Chunwei 已提交
284
void AnalysisPredictor::OptimizeInferenceProgram() {
285 286 287
  status_program_optimized_ = true;

  argument_.SetUseGPU(config_.use_gpu);
Y
Yan Chunwei 已提交
288 289
  // Analyze inference_program
  if (!config_.model_dir.empty()) {
290
    argument_.SetModelDir(config_.model_dir);
Y
Yan Chunwei 已提交
291 292 293 294 295
  } else {
    PADDLE_ENFORCE(
        !config_.param_file.empty(),
        "Either model_dir or (param_file, prog_file) should be set.");
    PADDLE_ENFORCE(!config_.prog_file.empty());
296 297
    argument_.SetModelProgramPath(config_.prog_file);
    argument_.SetModelParamsPath(config_.param_file);
Y
Yan Chunwei 已提交
298
  }
299

300 301 302 303
  if (config_.use_gpu && config_.use_tensorrt_) {
    argument_.SetUseTensorRT(true);
    argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
    argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
W
Wojciech Uss 已提交
304
  }
305

306 307 308 309 310 311 312 313 314
  auto passes = config_.pass_builder()->AllPasses();
  if (!config_.enable_ir_optim) passes.clear();
  argument_.SetIrAnalysisPasses(passes);
  argument_.SetScopeNotOwned(const_cast<framework::Scope *>(scope_.get()));
  Analyzer().Run(&argument_);

  PADDLE_ENFORCE(argument_.scope_valid());
  VLOG(5) << "to prepare executor";
  ARGUMENT_CHECK_FIELD((&argument_), ir_analyzed_program);
Y
Yan Chunwei 已提交
315
  inference_program_.reset(
316
      new framework::ProgramDesc(argument_.ir_analyzed_program()));
317
  LOG(INFO) << "== optimize end ==";
Y
Yan Chunwei 已提交
318
}
319 320

template <>
321 322
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) {
323
  VLOG(30) << "create AnalysisConfig";
324 325 326 327 328 329 330 331 332 333 334 335 336
  if (config.use_gpu) {
    // 1. GPU memeroy
    PADDLE_ENFORCE_GT(
        config.fraction_of_gpu_memory, 0.f,
        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
    PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device);
    std::vector<std::string> flags;
    if (config.fraction_of_gpu_memory >= 0.0f ||
        config.fraction_of_gpu_memory <= 0.95f) {
      flags.push_back("dummpy");
      std::string flag = "--fraction_of_gpu_memory_to_use=" +
                         std::to_string(config.fraction_of_gpu_memory);
      flags.push_back(flag);
337
      VLOG(30) << "set flag: " << flag;
338 339 340 341 342
      framework::InitGflags(flags);
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
343
  if (!dynamic_cast<AnalysisPredictor *>(predictor.get())->Init(nullptr)) {
344 345
    return nullptr;
  }
346
  return std::move(predictor);
347 348
}

349
void AnalysisPredictor::PrepareFeedFetch() {
350 351
  PADDLE_ENFORCE_NOT_NULL(sub_scope_);
  CreateFeedFetchVar(sub_scope_);
352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
  for (auto *op : inference_program_->Block(0).AllOps()) {
    if (op->Type() == "feed") {
      int idx = boost::get<int>(op->GetAttr("col"));
      if (feeds_.size() <= static_cast<size_t>(idx)) {
        feeds_.resize(idx + 1);
      }
      feeds_[idx] = op;
      feed_names_[op->Output("Out")[0]] = idx;
    } else if (op->Type() == "fetch") {
      int idx = boost::get<int>(op->GetAttr("col"));
      if (fetchs_.size() <= static_cast<size_t>(idx)) {
        fetchs_.resize(idx + 1);
      }
      fetchs_[idx] = op;
    }
  }
}

370 371 372 373 374 375 376 377
void AnalysisPredictor::CreateFeedFetchVar(framework::Scope *scope) {
  PADDLE_ENFORCE_NOT_NULL(scope);
  auto *var = scope->Var("feed");
  var->GetMutable<framework::FeedFetchList>();
  var = scope->Var("fetch");
  var->GetMutable<framework::FeedFetchList>();
}

378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetInputTensor(
    const std::string &name) {
  PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = true;
  res->SetName(name);
  return res;
}

std::unique_ptr<ZeroCopyTensor> AnalysisPredictor::GetOutputTensor(
    const std::string &name) {
  PADDLE_ENFORCE(executor_->scope()->FindVar(name), "no name called %s", name);
  std::unique_ptr<ZeroCopyTensor> res(
      new ZeroCopyTensor(static_cast<void *>(executor_->scope())));
  res->input_or_output_ = false;
  res->SetName(name);
  return res;
}

bool AnalysisPredictor::ZeroCopyRun() {
  executor_->Run();
Y
Yan Chunwei 已提交
400 401 402
  // Fix TensorArray reuse not cleaned bug.
  tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get());
  tensor_array_batch_cleaner_.ResetTensorArray();
403 404 405 406 407
  return true;
}

bool AnalysisPredictor::LoadProgramDesc() {
  // Initialize the inference program
408
  std::string filename;
409
  if (!config_.model_dir.empty()) {
410
    filename = config_.model_dir + "/__model__";
411 412 413 414
  } else if (!config_.prog_file.empty() && !config_.param_file.empty()) {
    // All parameters are saved in a single file.
    // The file names should be consistent with that used
    // in Python API `fluid.io.save_inference_model`.
415
    filename = config_.prog_file;
416
  } else {
417 418 419 420 421
    if (config_.model_dir.empty() && config_.prog_file.empty()) {
      LOG(ERROR)
          << "Either model_dir or (prog_file, param_file) should be set.";
      return false;
    }
422 423 424 425 426
    LOG(ERROR) << string::Sprintf(
        "not valid model path '%s' or program path '%s'.", config_.model_dir,
        config_.param_file);
    return false;
  }
427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499

  std::string pb_content;
  // Read binary
  std::ifstream fin(filename, std::ios::in | std::ios::binary);
  PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
  fin.seekg(0, std::ios::end);

  pb_content.resize(fin.tellg());
  fin.seekg(0, std::ios::beg);
  fin.read(&(pb_content.at(0)), pb_content.size());
  fin.close();

  // Create ProgramDesc
  framework::proto::ProgramDesc proto;
  proto.ParseFromString(pb_content);
  inference_program_.reset(new framework::ProgramDesc(proto));
  return true;
}

bool AnalysisPredictor::LoadParameters() {
  PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
                          "The inference program should be loaded first.");
  const auto &global_block = inference_program_->MutableBlock(0);

  // create a temporary program to load parameters.

  std::unique_ptr<framework::ProgramDesc> load_program(
      new framework::ProgramDesc());
  framework::BlockDesc *load_block = load_program->MutableBlock(0);
  std::vector<std::string> params;

  for (auto *var : global_block->AllVars()) {
    if (IsPersistable(var)) {
      VLOG(3) << "persistable variable's name: " << var->Name();

      framework::VarDesc *new_var = load_block->Var(var->Name());
      new_var->SetShape(var->GetShape());
      new_var->SetDataType(var->GetDataType());
      new_var->SetType(var->GetType());
      new_var->SetLoDLevel(var->GetLoDLevel());
      new_var->SetPersistable(true);

      if (!config_.param_file.empty()) {
        params.push_back(new_var->Name());
      } else {
        // append_op
        framework::OpDesc *op = load_block->AppendOp();
        op->SetType("load");
        op->SetOutput("Out", {new_var->Name()});
        op->SetAttr("file_path", {config_.model_dir + "/" + new_var->Name()});
        op->CheckAttrs();
      }
    }
  }

  if (!config_.param_file.empty()) {
    // sort paramlist to have consistent ordering
    std::sort(params.begin(), params.end());
    // append just the load_combine op
    framework::OpDesc *op = load_block->AppendOp();
    op->SetType("load_combine");
    op->SetOutput("Out", params);
    op->SetAttr("file_path", {config_.param_file});
    op->CheckAttrs();
  }

  // Use NaiveExecutor to Load parameters.
  platform::CPUPlace place;
  framework::NaiveExecutor e(place);
  e.Prepare(scope_.get(), *load_program, 0, false);
  e.Run();
  VLOG(3) << "get " << scope_->LocalVarNames().size() << " vars after load";

500 501
  return true;
}
502 503 504 505 506 507 508 509 510 511 512 513 514

AnalysisPredictor::~AnalysisPredictor() {
#if !defined(_WIN32)
  if (FLAGS_profile) {
    platform::DisableProfiler(platform::EventSortingKey::kTotal,
                              "./profile.log");
  }
#endif
  if (sub_scope_) {
    scope_->DeleteScope(sub_scope_);
  }
}

515 516 517 518 519 520
std::unique_ptr<PaddlePredictor> AnalysisPredictor::Clone() {
  auto *x = new AnalysisPredictor(config_);
  x->Init(scope_, inference_program_);
  return std::unique_ptr<PaddlePredictor>(x);
}

Y
Yan Chunwei 已提交
521 522
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
523
    const contrib::AnalysisConfig &config) {
Y
Yan Chunwei 已提交
524 525 526 527
  return CreatePaddlePredictor<contrib::AnalysisConfig,
                               PaddleEngineKind::kAnalysis>(config);
}

528
}  // namespace paddle
529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550

#if PADDLE_WITH_TENSORRT
USE_TRT_CONVERTER(elementwise_add_weight);
USE_TRT_CONVERTER(elementwise_add_tensor);
USE_TRT_CONVERTER(elementwise_sub_tensor);
USE_TRT_CONVERTER(elementwise_div_tensor);
USE_TRT_CONVERTER(elementwise_mul_tensor);
USE_TRT_CONVERTER(elementwise_max_tensor);
USE_TRT_CONVERTER(elementwise_min_tensor);
USE_TRT_CONVERTER(elementwise_pow_tensor);
USE_TRT_CONVERTER(mul);
USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu);
USE_TRT_CONVERTER(sigmoid);
USE_TRT_CONVERTER(tanh);
USE_TRT_CONVERTER(fc);
USE_TRT_CONVERTER(pool2d);
USE_TRT_CONVERTER(softmax);
USE_TRT_CONVERTER(batch_norm);
USE_TRT_CONVERTER(concat);
USE_TRT_CONVERTER(dropout);
USE_TRT_CONVERTER(pad);
551
USE_TRT_CONVERTER(split);
552 553
USE_TRT_CONVERTER(prelu);
USE_TRT_CONVERTER(conv2d_transpose);
554
#endif