analysis_predictor.cc 12.2 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
#include <memory>
17 18
#include <string>
#include <vector>
19
#include "paddle/fluid/framework/feed_fetch_method.h"
Y
Yan Chunwei 已提交
20
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
21
#include "paddle/fluid/framework/ir/pass.h"
22
#include "paddle/fluid/framework/naive_executor.h"
23
#include "paddle/fluid/framework/scope.h"
24
#include "paddle/fluid/inference/api/helper.h"
25
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
26
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
27
#include "paddle/fluid/inference/utils/singleton.h"
28
#include "paddle/fluid/platform/cpu_helper.h"
T
tensor-tang 已提交
29 30 31
#include "paddle/fluid/platform/profiler.h"

DECLARE_bool(profile);
32
DECLARE_int32(paddle_num_threads);
33 34 35

namespace paddle {

36 37
using contrib::AnalysisConfig;

Y
Yan Chunwei 已提交
38
bool AnalysisPredictor::Init(
39 40
    const std::shared_ptr<framework::Scope> &parent_scope,
    const std::shared_ptr<framework::ProgramDesc> &program) {
Y
Yan Chunwei 已提交
41
  VLOG(3) << "Predictor::init()";
T
tensor-tang 已提交
42 43 44 45 46 47 48 49 50 51
#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

52 53 54
  // no matter with or without MKLDNN
  paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);

Y
Yan Chunwei 已提交
55 56
  if (config_.use_gpu) {
    place_ = paddle::platform::CUDAPlace(config_.device);
57 58
    LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim "
                    "is turned false.";
59
    config_.enable_ir_optim = false;
Y
Yan Chunwei 已提交
60 61 62 63 64 65 66 67 68 69
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  if (parent_scope) {
    scope_ = parent_scope;
    sub_scope_ = &(parent_scope->NewScope());
  } else {
    paddle::framework::InitDevices(false);
    scope_.reset(new paddle::framework::Scope());
  }
70

71
  executor_.reset(new paddle::framework::NaiveExecutor(place_));
72

73 74 75
  if (!program) {
    if (!LoadProgramDesc()) return false;
    OptimizeInferenceProgram();
Y
Yan Chunwei 已提交
76
  } else {
77 78
    inference_program_ = program;
  }
M
Michal Gallus 已提交
79 80 81 82 83

  if (config_._use_mkldnn) {
    executor_->EnableMKLDNN(*inference_program_);
  }

84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102
  executor_->Prepare(scope_.get(), *inference_program_, 0,
                     config_.use_feed_fetch_ops);

  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();
  return true;
}

bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                            std::vector<PaddleTensor> *output_data,
                            int batch_size) {
  VLOG(3) << "Predictor::predict";
  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
103
    return false;
104
  }
M
Michal Gallus 已提交
105

106 107 108
  // Run the inference program
  // if share variables, we need not create variables
  executor_->Run();
109

110 111 112 113
  // get fetch variable
  if (!GetFetch(output_data, scope)) {
    LOG(ERROR) << "fail to get fetches";
    return false;
T
tensor-tang 已提交
114
  }
115 116 117
  VLOG(3) << "predict cost: " << timer.toc() << "ms";
  return true;
}
118

119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204
bool AnalysisPredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                framework::Scope *scope) {
  VLOG(3) << "Predictor::set_feed";
  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) {
  VLOG(3) << "Predictor::get_fetch";
  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
205 206
  return true;
}
207

Y
Yan Chunwei 已提交
208 209
void AnalysisPredictor::OptimizeInferenceProgram() {
  LOG(INFO) << "optimize begin";
210
  FLAGS_IA_enable_ir = config_.enable_ir_optim;
Y
Yan Chunwei 已提交
211 212 213 214 215 216 217 218 219 220 221 222 223 224
  FLAGS_IA_enable_tensorrt_subgraph_engine = false;
  FLAGS_IA_output_storage_path = "";  // Don't output the model.
  // Analyze inference_program
  if (!config_.model_dir.empty()) {
    argument_.fluid_model_dir.reset(new std::string(config_.model_dir));
  } else {
    PADDLE_ENFORCE(
        !config_.param_file.empty(),
        "Either model_dir or (param_file, prog_file) should be set.");
    PADDLE_ENFORCE(!config_.prog_file.empty());
    argument_.fluid_model_program_path.reset(
        new std::string(config_.prog_file));
    argument_.fluid_model_param_path.reset(new std::string(config_.param_file));
  }
225

Y
Yan Chunwei 已提交
226 227
  argument_.origin_program_desc.reset(
      new ProgramDesc(*inference_program_->Proto()));
Y
Yan Chunwei 已提交
228 229 230
  PADDLE_ENFORCE(
      config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude,
      "Only kExclude is supported yet.");
231 232
  Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);

Y
Yan Chunwei 已提交
233 234 235 236
  CHECK(argument_.transformed_program_desc);
  VLOG(5) << "to prepare executor";
  inference_program_.reset(
      new framework::ProgramDesc(*argument_.transformed_program_desc));
237 238 239 240 241 242
  if (argument_.Has(framework::ir::kParamScopeAttr)) {
    // Update scope.
    scope_.reset(
        argument_.Release<framework::Scope>(framework::ir::kParamScopeAttr));
  }
  LOG(INFO) << "== optimize end ==";
Y
Yan Chunwei 已提交
243
}
244 245

template <>
246 247
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) {
248
  VLOG(3) << "create AnalysisConfig";
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
  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);
      VLOG(3) << "set flag: " << flag;
      framework::InitGflags(flags);
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
268
  if (!dynamic_cast<AnalysisPredictor *>(predictor.get())->Init(nullptr)) {
269 270 271 272 273
    return nullptr;
  }
  return predictor;
}

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
void AnalysisPredictor::PrepareFeedFetch() {
  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;
    }
  }
}

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();
  return true;
}

bool AnalysisPredictor::LoadProgramDesc() {
  // Initialize the inference program
  std::unique_ptr<framework::Executor> tmp_exe(
      new framework::Executor(platform::CPUPlace()));
  if (!config_.model_dir.empty()) {
    // Parameters are saved in separate files sited in
    // the specified `dirname`.
    inference_program_ = paddle::inference::Load(
        static_cast<framework::Executor *>(tmp_exe.get()), scope_.get(),
        config_.model_dir);
  } 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`.
    inference_program_ = paddle::inference::Load(
        static_cast<framework::Executor *>(tmp_exe.get()), scope_.get(),
        config_.prog_file, config_.param_file);
  } else {
    LOG(ERROR) << string::Sprintf(
        "not valid model path '%s' or program path '%s'.", config_.model_dir,
        config_.param_file);
    return false;
  }
  return true;
}
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
349 350
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
351
    const contrib::AnalysisConfig &config) {
Y
Yan Chunwei 已提交
352 353 354 355
  return CreatePaddlePredictor<contrib::AnalysisConfig,
                               PaddleEngineKind::kAnalysis>(config);
}

356
}  // namespace paddle