paddle_inference_api_impl.cc 9.1 KB
Newer Older
X
Xin Pan 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
/* 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 <sys/time.h>
#include <algorithm>
#include <map>
#include <set>
#include <sstream>
#include <string>
#include <utility>
#include <vector>

#include "paddle/contrib/inference/paddle_inference_api_impl.h"

namespace paddle {
namespace {

// Timer for timer
class Timer {
W
Wu Yi 已提交
31
 public:
X
Xin Pan 已提交
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
  double start;
  double startu;
  void tic() {
    struct timeval tp;
    gettimeofday(&tp, NULL);
    start = tp.tv_sec;
    startu = tp.tv_usec;
  }
  double toc() {
    struct timeval tp;
    gettimeofday(&tp, NULL);
    double used_time_ms =
        (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0;
    return used_time_ms;
  }
};

template <class T>
std::string num2str(T a) {
  std::stringstream istr;
  istr << a;
  return istr.str();
}
}  // namespace

Y
Yan Chunwei 已提交
57
bool NativePaddlePredictor::Init() {
X
Xin Pan 已提交
58 59
  VLOG(3) << "Predictor::init()";

Y
Yan Chunwei 已提交
60
  if (config_.use_gpu) {
X
Xin Pan 已提交
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
    place_ = paddle::platform::CUDAPlace(config_.device);
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  paddle::framework::InitDevices(false);
  executor_.reset(new paddle::framework::Executor(place_));
  scope_.reset(new paddle::framework::Scope());

  // Initialize the inference program
  if (!config_.model_dir.empty()) {
    // Parameters are saved in separate files sited in
    // the specified `dirname`.
    inference_program_ = paddle::inference::Load(
        executor_.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(
        executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
  } else {
    LOG(ERROR) << "fail to load inference model.";
    return false;
  }
  ctx_ = executor_->Prepare(*inference_program_, 0);

Y
Yan Chunwei 已提交
87 88 89 90 91 92 93
  // Create temporary variables first, so that the first batch do not need to
  // create variables in the runtime. This is the logics of the old inference
  // API.
  // TODO(Superjomn) this should be modified when `Clone` is valid for
  // multi-thread application.
  executor_->CreateVariables(*inference_program_, scope_.get(), 0);

X
Xin Pan 已提交
94 95 96 97 98 99
  // Get the feed_target_names and fetch_target_names
  feed_target_names_ = inference_program_->GetFeedTargetNames();
  fetch_target_names_ = inference_program_->GetFetchTargetNames();
  return true;
}

Y
Yan Chunwei 已提交
100 101
bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                std::vector<PaddleTensor> *output_data) {
X
Xin Pan 已提交
102 103 104 105
  VLOG(3) << "Predictor::predict";
  Timer timer;
  timer.tic();
  // set feed variable
106 107
  std::map<std::string, const framework::LoDTensor *> feed_targets;
  std::vector<framework::LoDTensor> feeds;
X
Xin Pan 已提交
108 109 110 111 112 113 114 115
  if (!SetFeed(inputs, &feeds)) {
    LOG(ERROR) << "fail to set feed";
    return false;
  }
  for (size_t i = 0; i < feed_target_names_.size(); ++i) {
    feed_targets[feed_target_names_[i]] = &feeds[i];
  }
  // get fetch variable
116 117
  std::map<std::string, framework::LoDTensor *> fetch_targets;
  std::vector<framework::LoDTensor> fetchs;
X
Xin Pan 已提交
118 119 120 121 122 123 124 125 126 127
  fetchs.resize(fetch_target_names_.size());
  for (size_t i = 0; i < fetch_target_names_.size(); ++i) {
    fetch_targets[fetch_target_names_[i]] = &fetchs[i];
  }
  // Run the inference program
  // if share variables, we need not create variables
  executor_->RunPreparedContext(ctx_.get(),
                                scope_.get(),
                                &feed_targets,
                                &fetch_targets,
Y
Yan Chunwei 已提交
128
                                false /* don't create variable eatch time */);
X
Xin Pan 已提交
129 130 131 132 133 134 135 136
  if (!GetFetch(fetchs, output_data)) {
    LOG(ERROR) << "fail to get fetchs";
    return false;
  }
  VLOG(3) << "predict cost: " << timer.toc() << "ms";
  return true;
}

Y
Yan Chunwei 已提交
137
std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
X
Xin Pan 已提交
138
  VLOG(3) << "Predictor::clone";
Y
Yan Chunwei 已提交
139 140 141 142
  std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));

  if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init()) {
    LOG(ERROR) << "fail to call Init";
X
Xin Pan 已提交
143 144
    return nullptr;
  }
145 146
  // fix manylinux compile error.
  return std::move(cls);
X
Xin Pan 已提交
147 148
}

Y
Yan Chunwei 已提交
149 150
bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
                                    std::vector<framework::LoDTensor> *feeds) {
X
Xin Pan 已提交
151 152 153 154 155 156
  VLOG(3) << "Predictor::set_feed";
  if (inputs.size() != feed_target_names_.size()) {
    LOG(ERROR) << "wrong feed input size.";
    return false;
  }
  for (size_t i = 0; i < feed_target_names_.size(); ++i) {
157 158
    framework::LoDTensor input;
    framework::DDim ddim = framework::make_ddim(inputs[i].shape);
X
Xin Pan 已提交
159 160
    void *input_ptr;
    if (inputs[i].dtype == PaddleDType::INT64) {
161
      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
X
Xin Pan 已提交
162
    } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
163
      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
X
Xin Pan 已提交
164 165 166 167 168 169 170 171 172 173 174 175 176 177
    } 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);
    feeds->push_back(input);
  }
  return true;
}

Y
Yan Chunwei 已提交
178
bool NativePaddlePredictor::GetFetch(
179
    const std::vector<framework::LoDTensor> &fetchs,
X
Xin Pan 已提交
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243
    std::vector<PaddleTensor> *outputs) {
  VLOG(3) << "Predictor::get_fetch";
  outputs->resize(fetchs.size());
  for (size_t i = 0; i < fetchs.size(); ++i) {
    // TODO(panyx0718): Support fetch of other types.
    if (fetchs[i].type() != typeid(float)) {
      LOG(ERROR) << "only support fetching float now.";
      return false;
    }
    std::vector<int> shape;
    auto dims_i = fetchs[i].dims();
    auto lod = fetchs[i].lod();
    const float *output_ptr = fetchs[i].data<float>();
    // const int64_t* output_ptr = fetchs[i].data<int64_t>();
    auto num = fetchs[i].numel();
    std::vector<float> data;
    if (0 == lod.size()) {
      std::copy(output_ptr, output_ptr + num, std::back_inserter(data));
      for (int j = 0; j < dims_i.size(); ++j) {
        shape.push_back(dims_i[j]);
      }
    } else {
      // for batch detection
      // image[0] -> output[0] shape {145, 6}
      // image[1] -> output[1] shape {176, 6}
      // then,
      // the batch output shape {321, 6}
      // the lod {{0, 145, 321}}
      // so we should append output[0] to {176, 6}
      size_t max_dim = 0;
      for (size_t j = 1; j < lod[0].size(); j++) {
        max_dim = std::max(max_dim, lod[0][j] - lod[0][j - 1]);
      }
      size_t common_dim = lod[0].back() == 0 ? 0 : num / lod[0].back();
      if (max_dim > 0) {
        data.resize((lod[0].size() - 1) * max_dim * common_dim, 0);
      }
      for (size_t j = 1; j < lod[0].size(); j++) {
        size_t start = lod[0][j - 1] * common_dim;
        size_t end = lod[0][j] * common_dim;
        if (end > start) {
          std::copy(output_ptr + start,
                    output_ptr + end,
                    data.begin() + (j - 1) * max_dim * common_dim);
        }
      }
      shape.push_back(lod[0].size() - 1);
      shape.push_back(max_dim);
      for (int j = 1; j < dims_i.size(); ++j) {
        shape.push_back(dims_i[j]);
      }
    }

    outputs->at(i).shape = shape;
    outputs->at(i).data.length = sizeof(float) * data.size();
    outputs->at(i).data.data = malloc(outputs->at(i).data.length);
    std::memcpy(
        outputs->at(i).data.data, data.data(), outputs->at(i).data.length);
    outputs->at(i).dtype = PaddleDType::FLOAT32;
    // TODO(panyx0718): support other types? fill tensor name? avoid a copy.
  }
  return true;
}

244
template <>
Y
Yan Chunwei 已提交
245
std::unique_ptr<PaddlePredictor>
Y
Yan Chunwei 已提交
246
CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
Y
Yan Chunwei 已提交
247 248 249 250
    const NativeConfig &config) {
  VLOG(3) << "create NativePaddlePredictor";
  if (config.use_gpu) {
    // 1. GPU memeroy
Y
Yan Chunwei 已提交
251 252 253
    PADDLE_ENFORCE(
        config.fraction_of_gpu_memory > 0.f,
        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
Y
Yan Chunwei 已提交
254 255 256 257 258 259 260 261 262 263
    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=" +
                         num2str<float>(config.fraction_of_gpu_memory);
      flags.push_back(flag);
      VLOG(3) << "set flag: " << flag;
      framework::InitGflags(flags);
    }
X
Xin Pan 已提交
264 265
  }

Y
Yan Chunwei 已提交
266 267
  std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
  if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init()) {
X
Xin Pan 已提交
268 269
    return nullptr;
  }
270
  return std::move(predictor);
X
Xin Pan 已提交
271 272 273
}

}  // namespace paddle