test_NetworkCompare.cpp 9.5 KB
Newer Older
Z
zhangjinchao01 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 31 32 33 34 35 36 37 38 39 40 41 42 43
/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.

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. */

#undef PADDLE_DISABLE_TIMER
#include <paddle/utils/PythonUtil.h>
#include <cstdlib>
#include <algorithm>
#include <gtest/gtest.h>

#include "paddle/trainer/Trainer.h"
#include "paddle/utils/Stat.h"
#include "TestUtil.h"

using namespace paddle;  // NOLINT
using namespace std;     // NOLINT

P_DECLARE_int32(gpu_id);
P_DECLARE_double(checkgrad_eps);
P_DEFINE_bool(use_label, true, "input label or sequence label");
P_DEFINE_bool(static_para, false, "static parameter");

struct DataIn {
  std::vector<Argument> inArgs;
  std::vector<MatrixPtr> outGrads;
  std::vector<VectorPtr> paraValues;
};

struct DataOut {
  std::vector<MatrixPtr> outValues;
  std::vector<VectorPtr> paraGrads;
};

44 45
void initArgument(DataIn& data,
                  const std::string& configPath,
Z
zhangjinchao01 已提交
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
                  bool useGpu = FLAGS_use_gpu) {
  TrainerConfigHelper config(configPath);
  size_t batchSize = config.getOptConfig().batch_size();

  for (const auto& layer_name : config.getModelConfig().input_layer_names()) {
    auto layer_config = std::find_if(config.getModelConfig().layers().begin(),
                                     config.getModelConfig().layers().end(),
                                     [=](const LayerConfig& layer_config) {
                                       return layer_config.name() == layer_name;
                                     });
    CHECK(layer_config != config.getModelConfig().layers().end());

    size_t layerSize = layer_config->size();
    Argument arg;
    arg.value = Matrix::create(batchSize, layerSize, false, useGpu);
    arg.grad = Matrix::create(batchSize, layerSize, false, useGpu);
    arg.value->randomizeUniform();
    arg.value->add(-0.5);
    arg.value->sigmoid(*arg.value);
    arg.grad->zeroMem();
    if (FLAGS_use_label) {
      arg.ids = VectorT<int>::create(batchSize, useGpu);
      arg.ids->rand(layerSize);
    }
    generateSequenceStartPositions(batchSize, arg.sequenceStartPositions);
    data.inArgs.push_back(arg);
  }

  for (const auto& layer_name : config.getModelConfig().output_layer_names()) {
    auto layer_config = std::find_if(config.getModelConfig().layers().begin(),
                                     config.getModelConfig().layers().end(),
                                     [=](const LayerConfig& layer_config) {
                                       return layer_config.name() == layer_name;
                                     });
    CHECK(layer_config != config.getModelConfig().layers().end());

    size_t layerSize = layer_config->size();
    MatrixPtr grad = Matrix::create(batchSize, layerSize, false, useGpu);
    grad->randomizeUniform();
    data.outGrads.push_back(grad);
  }

  for (const auto& para_config : config.getModelConfig().parameters()) {
    VectorPtr value = Vector::create(para_config.size(), useGpu);
    value->randnorm(0, 2);
    data.paraValues.push_back(value);
  }
}

void calcGradient(DataIn& in, DataOut& out, const std::string& configPath) {
  *ThreadLocalRand::getSeed() = 0;
  srand(0);

  Trainer trainer;
  auto config = std::make_shared<TrainerConfigHelper>(configPath);
  trainer.init(config, false);

  std::vector<ParameterPtr> parameters;
  vector<Argument> outArgs;

  auto gradientMachine = trainer.getGradientMachine();
  parameters = gradientMachine->getParameters();
  if (FLAGS_static_para) {
    for (size_t i = 0; i < parameters.size(); i++) {
      parameters[i]->getBuf(PARAMETER_VALUE)->one();
    }
  } else {
    for (size_t i = 0; i < in.paraValues.size(); i++) {
      parameters[i]->getBuf(PARAMETER_VALUE)->copyFrom(*in.paraValues[i]);
    }
  }
  gradientMachine->start(trainer.getConfig(), nullptr);
  gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN);
  for (size_t i = 0; i < in.outGrads.size(); i++) {
120 121
    // If the all the layers in the config have no parameters, also
    // not set NeedGradient(), the outArgs[i] will be nullptr.
Z
zhangjinchao01 已提交
122 123 124 125
    outArgs[i].grad->copyFrom(*in.outGrads[i]);
  }
  gradientMachine->backward();
  for (size_t i = 0; i < in.outGrads.size(); i++) {
126 127 128 129
    MatrixPtr value = Matrix::create(outArgs[i].value->getHeight(),
                                     outArgs[i].value->getWidth(),
                                     false,
                                     false);
Z
zhangjinchao01 已提交
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
    value->copyFrom(*outArgs[i].value);
    out.outValues.push_back(value);
  }
  for (size_t i = 0; i < in.paraValues.size(); i++) {
    VectorPtr grad = Vector::create(
        parameters[i]->getBuf(PARAMETER_GRADIENT)->getSize(), false);
    grad->copyFrom(*parameters[i]->getBuf(PARAMETER_GRADIENT));
    out.paraGrads.push_back(grad);
  }

  for (int i = 0; i < 20; i++) {
    REGISTER_TIMER("forward");
    gradientMachine->forward(in.inArgs, &outArgs, PASS_TRAIN);
  }
  for (int i = 0; i < 20; i++) {
    REGISTER_TIMER("backward");
    gradientMachine->backward();
  }

  gradientMachine->finish();
}

152 153 154 155 156 157
void checkBuffer(real* A,
                 const char* desA,
                 real* B,
                 const char* desB,
                 size_t len,
                 size_t width = 1) {
Z
zhangjinchao01 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176
  int nNum = 0;
  for (size_t i = 0; i < len; ++i) {
    real diff = fabs(A[i] - B[i]);
    if (diff > 0.0f &&
        diff / std::max(fabs(A[i]), fabs(B[i])) > FLAGS_checkgrad_eps) {
      nNum++;
      LOG(INFO) << "Row: " << i / width << ", " << desA << " : " << A[i]
                << "    " << desB << " : " << B[i];
    }
  }
  EXPECT_EQ(0, nNum);
}

void compareGradient(DataOut& outA, DataOut& outB) {
  LOG(INFO) << "------------------------------"
            << " Check Network Output "
            << "------------------------------";
  for (size_t i = 0; i < outA.outValues.size(); ++i) {
    LOG(INFO) << "OUTPUT VALUE: " << i;
177 178 179 180
    checkBuffer(outA.outValues[i]->getData(),
                "network A output",
                outB.outValues[i]->getData(),
                "network B output",
Z
zhangjinchao01 已提交
181 182 183 184 185 186 187 188 189 190
                outA.outValues[i]->getElementCnt(),
                outA.outValues[i]->getWidth());
  }

  if (!FLAGS_static_para) {
    LOG(INFO) << "------------------------------"
              << " Check Parameters "
              << "------------------------------";
    for (size_t i = 0; i < outA.paraGrads.size(); ++i) {
      LOG(INFO) << "PARAMETER GRADIENT: " << i;
191 192 193 194
      checkBuffer(outA.paraGrads[i]->getData(),
                  "Network A",
                  outB.paraGrads[i]->getData(),
                  "Network B",
Z
zhangjinchao01 已提交
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
                  outA.paraGrads[i]->getSize());
    }
  }
}

void compareNetwork(const std::string& config_file_a,
                    const std::string& config_file_b) {
  DataIn in;
  initArgument(in, config_file_a);

  DataOut dataA;
  calcGradient(in, dataA, config_file_a);
  LOG(INFO) << "forwardBackward of Network A is finished";
  globalStat.printSegTimerStatus();
  globalStat.reset();
  LOG(INFO) << "\n\n";

  DataOut dataB;
  calcGradient(in, dataB, config_file_b);
  LOG(INFO) << "forwardBackward of the Network B is finished";
  globalStat.printSegTimerStatus();
  globalStat.reset();
  LOG(INFO) << "\n\n";

  compareGradient(dataA, dataB);
}

TEST(Compare, concat_dotmul) {
  std::string config_file_a = "./gserver/tests/concat_dotmul_a.conf";
  std::string config_file_b = "./gserver/tests/concat_dotmul_b.conf";
  compareNetwork(config_file_a, config_file_b);
}

TEST(Compare, concat_fullmatrix) {
  std::string config_file_a = "./gserver/tests/concat_fullmatrix_a.conf";
  std::string config_file_b = "./gserver/tests/concat_fullmatrix_b.conf";
  compareNetwork(config_file_a, config_file_b);
}

TEST(Compare, concat_table) {
  std::string config_file_a = "./gserver/tests/concat_table_a.conf";
  std::string config_file_b = "./gserver/tests/concat_table_b.conf";
  compareNetwork(config_file_a, config_file_b);
}

240 241 242 243 244 245 246 247 248
#ifndef PADDLE_ONLY_CPU
TEST(Compare, img_pool) {
  std::string config_file_a = "./gserver/tests/img_pool_a.conf";
  std::string config_file_b = "./gserver/tests/img_pool_b.conf";
  bool useGpu = FLAGS_use_gpu;
  FLAGS_use_gpu = true;
  compareNetwork(config_file_a, config_file_b);
  FLAGS_use_gpu = useGpu;
}
249 250 251 252 253 254 255 256 257

TEST(Compare, img_conv) {
  std::string config_file_a = "./gserver/tests/img_conv_a.conf";
  std::string config_file_b = "./gserver/tests/img_conv_b.conf";
  bool useGpu = FLAGS_use_gpu;
  FLAGS_use_gpu = true;
  compareNetwork(config_file_a, config_file_b);
  FLAGS_use_gpu = useGpu;
}
258

259
// Test cudnn_conv and exconv give the same result
260
TEST(Compare, img_conv2) {
261 262
  std::string config_file_a = "./gserver/tests/img_conv_a.conf";
  std::string config_file_b = "./gserver/tests/img_conv_c.conf";
263 264 265 266 267
  bool useGpu = FLAGS_use_gpu;
  FLAGS_use_gpu = true;
  compareNetwork(config_file_a, config_file_b);
  FLAGS_use_gpu = useGpu;
}
268 269
#endif

Z
zhangjinchao01 已提交
270 271 272 273 274 275 276 277 278 279 280 281 282 283 284
P_DEFINE_string(config_file_a, "", "config of one network to compare");
P_DEFINE_string(config_file_b, "", "config of another network to compare");
TEST(Compare, network) {
  if (FLAGS_config_file_a != "" && FLAGS_config_file_b != "") {
    compareNetwork(FLAGS_config_file_a, FLAGS_config_file_b);
  }
}

int main(int argc, char** argv) {
  testing::InitGoogleTest(&argc, argv);
  paddle::initMain(argc, argv);
  initPython(argc, argv);
  int ret = RUN_ALL_TESTS();
  return ret;
}