test_RecurrentLayer.cpp 13.4 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 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
/* 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. */

#include <gtest/gtest.h>
#include <vector>
#include <paddle/utils/Version.h>
#include "paddle/gserver/layers/DataLayer.h"
#include "paddle/gserver/layers/Layer.h"
#include "ModelConfig.pb.h"

#include "TestUtil.h"

using namespace paddle;  // NOLINT
using namespace std;     // NOLINT
P_DECLARE_bool(use_gpu);
P_DECLARE_bool(rnn_use_batch);
P_DECLARE_int32(fixed_seq_length);

void checkError(const Matrix& matrix1, const Matrix& matrix2) {
  CHECK(matrix1.getHeight() == matrix2.getHeight());
  CHECK(matrix1.getWidth() == matrix2.getWidth());
#ifndef PADDLE_TYPE_DOUBLE
  real err = 1e-3;
#else
  real err = 1e-10;
#endif

  int height = matrix1.getHeight();
  int width = matrix1.getWidth();
  const real* data1 = matrix1.getData();
  const real* data2 = matrix2.getData();
  int count = 0;
  for (int i = 0; i < height; i++) {
    for (int j = 0; j < width; j++) {
      if (fabs(data1[i * width + j] - data2[i * width + j]) > err) {
        count++;
      }
    }
  }
  EXPECT_EQ(count, 0) << "There are " << count << " different element.";
}

void checkError(const CpuVector& vector1, const CpuVector& vector2) {
  CHECK(vector1.getSize() == vector2.getSize());
#ifndef PADDLE_TYPE_DOUBLE
  real err = 1e-3;
#else
  real err = 1e-10;
#endif

  int size = vector1.getSize();
  const real* data1 = vector1.getData();
  const real* data2 = vector2.getData();
  int count = 0;
  for (int i = 0; i < size; i++) {
    if (fabs(data1[i] - data2[i]) > err) {
      count++;
    }
  }
  EXPECT_EQ(count, 0) << "There are " << count << " different element.";
}

LayerPtr creatDataLayer(string name, size_t batchSize, int layerSize,
                        bool useGpu) {
  LayerConfig dataConfig;
  dataConfig.set_name(name);
  dataConfig.set_type("data");
  dataConfig.set_size(layerSize);
  LayerPtr layer = LayerPtr(new DataLayer(dataConfig));

  Argument data;
  data.value = Matrix::create(batchSize, layer->getSize(), false, useGpu);
  data.grad = Matrix::create(batchSize, layer->getSize(), false, useGpu);
  data.value->randomizeUniform();
  data.value->add(-0.5);
  data.value->sigmoid(*data.value);
  data.grad->zeroMem();

  generateSequenceStartPositions(batchSize, data.sequenceStartPositions);

  DataLayerPtr dataLayer = std::dynamic_pointer_cast<DataLayer>(layer);
  dataLayer->setData(data);
  dataLayer->forward(PASS_GC);

  return layer;
}

ParameterPtr creatParameter(string name, int pid, size_t paraSize,
                            bool useGpu) {
  ParameterConfig paraConfig;
  paraConfig.set_name(name);
  paraConfig.set_size(paraSize);

  ParameterPtr parameter =
      std::make_shared<Parameter>(paraConfig, useGpu, /*initialize */ false);
  parameter->enableType(PARAMETER_VALUE);
  parameter->enableType(PARAMETER_GRADIENT);
  parameter->randomize();
  parameter->setID(pid);

  return parameter;
}

ParameterPtr creatParameterBias(string name, int pid, size_t paraSize,
                                bool useGpu) {
  ParameterConfig paraConfig;
  paraConfig.set_name(name);
  paraConfig.set_size(paraSize);
  paraConfig.set_initial_std(1);

  ParameterPtr parameter =
      std::make_shared<Parameter>(paraConfig, useGpu, /*initialize */ true);
  parameter->randomize();
  parameter->setID(pid);

  return parameter;
}

LayerPtr initRecurrentLayer(LayerConfig layerConfig, size_t batchSize,
                            int layerSize, bool useGpu) {
  FLAGS_use_gpu = useGpu;
  LayerMap layerMap;
  ParameterMap parameterMap;
  LayerPtr dataLayer = creatDataLayer("layer_0", batchSize, layerSize, useGpu);
  layerMap[dataLayer->getName()] = dataLayer;

  ParameterPtr para =
      creatParameter("para_0", 0, layerSize * layerSize, useGpu);
  parameterMap[para->getName()] = para;

  layerConfig.add_inputs();
  LayerInputConfig& input = *(layerConfig.mutable_inputs(0));
  input.set_input_layer_name("layer_0");
  input.set_input_parameter_name("para_0");
  LayerPtr testLayer = Layer::create(layerConfig);
  layerMap[testLayer->getName()] = testLayer;

  testLayer->init(layerMap, parameterMap);
  testLayer->setNeedGradient(true);

  return testLayer;
}

void checkRecurrentLayer(LayerPtr testLayer) {
  const VectorPtr& weightGrad =
      (testLayer->getParameters()[0])->getBuf(PARAMETER_GRADIENT);
  const MatrixPtr& inputGrad = testLayer->getPrev(0)->getOutputGrad();
  CpuVector seqPara(weightGrad->getSize());
  CpuVector batPara(weightGrad->getSize());
  CpuMatrix seqInputGrad(inputGrad->getHeight(), inputGrad->getWidth());
  CpuMatrix batInputGrad(inputGrad->getHeight(), inputGrad->getWidth());

  CpuMatrix outputGrad(inputGrad->getHeight(), inputGrad->getWidth());
  outputGrad.randomizeUniform();

  /* use sequence calculate */
  FLAGS_rnn_use_batch = false;
  weightGrad->zero();
  inputGrad->zero();
  testLayer->forward(PASS_GC);
  testLayer->getOutputGrad()->copyFrom(outputGrad);
  testLayer->backward();
  seqPara.copyFrom(*weightGrad);
  seqInputGrad.copyFrom(*inputGrad);

  /* use batch calculate */
  FLAGS_rnn_use_batch = true;
  weightGrad->zero();
  inputGrad->zero();
  testLayer->forward(PASS_GC);
  testLayer->getOutputGrad()->copyFrom(outputGrad);
  testLayer->backward();
  batPara.copyFrom(*weightGrad);
  batInputGrad.copyFrom(*inputGrad);

  /* check */
  checkError(seqInputGrad, batInputGrad);
  checkError(seqPara, batPara);
}

TEST(Layer, RecurrentLayer) {
  LayerConfig layerConfig;
  layerConfig.set_name("rnn");
  layerConfig.set_type("recurrent");
  layerConfig.set_active_type("tanh");
  for (auto layerSize : {1, 10, 64, 128, 256, 512}) {
    for (auto batchSize : {1, 5, 20, 100, 128}) {
      for (auto useGpu : {false, true}) {
        for (auto reversed : {false, true}) {
          LOG(INFO) << " layerSize=" << layerSize << " batchSize=" << batchSize
                    << " useGpu=" << useGpu << " reversed=" << reversed;
          layerConfig.set_size(layerSize);
          layerConfig.set_reversed(reversed);
          LayerPtr testLayer =
              initRecurrentLayer(layerConfig, batchSize, layerSize, useGpu);
          checkRecurrentLayer(testLayer);
        }
      }
    }
  }
}

#define protected public
#include "paddle/gserver/layers/LstmLayer.h"
#include "paddle/gserver/layers/GatedRecurrentLayer.h"
template<class T>
class TestRecurrentLayer {
public:
  LayerConfig config_;
  bool useGpu_;
  bool useBatch_;
  LayerPtr testLayer_;
  LayerPtr dataLayer_;
  ParameterPtr para_;
  ParameterPtr bias_;
  LayerMap layerMap_;
  ParameterMap parameterMap_;
  TestRecurrentLayer(const LayerConfig& config,
    bool useGpu, bool useBatch = false)
    : config_(config), useGpu_(useGpu), useBatch_(useBatch) {}
  void init(size_t batchSize) {
    FLAGS_use_gpu = useGpu_;
    testLayer_ = Layer::create(config_);
    if (typeid(T) == typeid(GatedRecurrentLayer)) {
      dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(),
                                  batchSize, config_.size() * 3, useGpu_);
      para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(),
                             0, config_.size() * config_.size() * 3, useGpu_);
      bias_ = creatParameterBias(config_.bias_parameter_name(),
                                 1, config_.size() * 3, useGpu_);
    } else if (typeid(T) == typeid(LstmLayer)) {
      dataLayer_ = creatDataLayer(config_.mutable_inputs(0)->input_layer_name(),
                                  batchSize, config_.size() * 4, useGpu_);
      para_ = creatParameter(config_.mutable_inputs(0)->input_parameter_name(),
                             0, config_.size() * config_.size() * 4, useGpu_);
      bias_ = creatParameterBias(config_.bias_parameter_name(),
                                 1, config_.size() * 7, useGpu_);
    }
    layerMap_[dataLayer_->getName()] = dataLayer_;
    parameterMap_[para_->getName()] = para_;
    parameterMap_[bias_->getName()] = bias_;

    layerMap_[testLayer_->getName()] = testLayer_;
    testLayer_->init(layerMap_, parameterMap_);
    testLayer_->setNeedGradient(true);
    (dynamic_cast<T*>(testLayer_.get()))->useBatch_ = useBatch_;
  }
  void forward() {
    FLAGS_use_gpu = useGpu_;
    testLayer_->forward(PASS_GC);
  }
  void backward() {
    FLAGS_use_gpu = useGpu_;
    testLayer_->backward(nullptr);
  }
};

template<class T>
void checkRecurrentLayer(LayerConfig layerConfig, size_t batchSize,
                         bool cpuBatch, bool gpuBatch) {
  TestRecurrentLayer<T> testCpu(layerConfig, false, cpuBatch);
  TestRecurrentLayer<T> testGpu(layerConfig, true, gpuBatch);
  testCpu.init(batchSize);
  testGpu.init(batchSize);
  auto checkError = [](MatrixPtr cpu, MatrixPtr gpu,
                       int numSequences, const char* str) {
    CpuMatrix check(gpu->getHeight(), gpu->getWidth());
    check.copyFrom(*gpu);
    int height = cpu->getHeight();
    int width = cpu->getWidth();
    const real* data1 = cpu->getData();
    const real* data2 = check.getData();
    int count = 0;
    for (int i = 0; i < height; i++) {
      for (int j = 0; j < width; j++) {
        if (fabs(data1[i * width + j] - data2[i * width + j]) / numSequences >
            1e-4) {
          count++;
        }
      }
    }
    EXPECT_EQ(count, 0) << "[" << str << "]" <<
      "There are " << count << " different element.";
  };
  T* cpuLayer = dynamic_cast<T*>(testCpu.testLayer_.get());
  T* gpuLayer = dynamic_cast<T*>(testGpu.testLayer_.get());

  Argument& cpuInput = testCpu.dataLayer_->getOutput();
  Argument& gpuInput = testGpu.dataLayer_->getOutput();
  gpuInput.resizeAndCopyFrom(cpuInput, true);

  const VectorPtr& cpuVec = testCpu.para_->getBuf(PARAMETER_VALUE);
  const VectorPtr& gpuVec = testGpu.para_->getBuf(PARAMETER_VALUE);
  gpuVec->copyFrom(*cpuVec);

  const VectorPtr& cpuBiasVec = testCpu.bias_->getBuf(PARAMETER_VALUE);
  const VectorPtr& gpuBiasVec = testGpu.bias_->getBuf(PARAMETER_VALUE);
  gpuBiasVec->copyFrom(*cpuBiasVec);

  /* check forward */
  testCpu.forward();
  testGpu.forward();

  checkError(cpuLayer->getOutputValue(),
             gpuLayer->getOutputValue(), 1, "outputValue");

  /* check backward */
  cpuLayer->getOutputGrad()->randomizeUniform();
  gpuLayer->getOutputGrad()->copyFrom(*cpuLayer->getOutputGrad());
  hl_stream_synchronize(HPPL_STREAM_DEFAULT);

  testCpu.backward();
  testGpu.backward();

  // check input grad
  checkError(cpuInput.grad, gpuInput.grad, 1, "inputGrad");
  // check weight grad
  int numSequences = cpuInput.getNumSequences();
  checkError(cpuLayer->weight_->getWGrad(), gpuLayer->weight_->getWGrad(),
             numSequences, "weightGrad");
  // check bias grad
  checkError(cpuLayer->bias_->getWGrad(), gpuLayer->bias_->getWGrad(),
             numSequences, "biasGrad");
}

TEST(Layer, GatedRecurrentLayer) {
  LayerConfig layerConfig;
  layerConfig.set_type("gated_recurrent");
  layerConfig.set_active_type("sigmoid");
  layerConfig.set_active_gate_type("sigmoid");

  layerConfig.add_inputs();
  LayerInputConfig& input = *(layerConfig.mutable_inputs(0));
  input.set_input_layer_name("layer_0");
  input.set_input_parameter_name("para_0");
  layerConfig.set_bias_parameter_name("bias");

  for (auto frameSize : {32, 64, 128, 256, 512}) {
    for (auto batchSize : {1, 5, 100, 500}) {
      for (auto reversed : {false, true}) {
        for (auto cpuBatch : {false, true}) {
          for (auto gpuBatch : {false, true}) {
            LOG(INFO) << " batchSize=" << batchSize
                      << " frameSize=" << frameSize << " reversed=" << reversed
                      << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch;
            layerConfig.set_size(frameSize);
            layerConfig.set_reversed(reversed);
            checkRecurrentLayer<GatedRecurrentLayer>(
              layerConfig, batchSize, cpuBatch, gpuBatch);
          }
        }
      }
    }
  }
}

TEST(Layer, LstmLayer) {
  LayerConfig layerConfig;
  layerConfig.set_type("lstmemory");
  layerConfig.set_active_type("relu");
372
  layerConfig.set_active_state_type("tanh");
Z
zhangjinchao01 已提交
373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
  layerConfig.set_active_gate_type("sigmoid");

  layerConfig.add_inputs();
  LayerInputConfig& input = *(layerConfig.mutable_inputs(0));
  input.set_input_layer_name("layer_0");
  input.set_input_parameter_name("para_0");
  layerConfig.set_bias_parameter_name("bias");

  for (auto frameSize : {32, 64, 128, 256, 512}) {
    for (auto batchSize : {1, 5, 100, 500}) {
      for (auto reversed : {false, true}) {
        for (auto cpuBatch : {false, true}) {
          for (auto gpuBatch : {false, true}) {
            LOG(INFO) << " batchSize=" << batchSize
                      << " frameSize=" << frameSize << " reversed=" << reversed
                      << " cpuBatch=" << cpuBatch << " gpuBatch=" << gpuBatch;
            layerConfig.set_size(frameSize);
            layerConfig.set_reversed(reversed);
            checkRecurrentLayer<LstmLayer>
              (layerConfig, batchSize, cpuBatch, gpuBatch);
          }
        }
      }
    }
  }
}

int main(int argc, char** argv) {
  if (version::isWithGpu()) {
    testing::InitGoogleTest(&argc, argv);
    initMain(argc, argv);
    return RUN_ALL_TESTS();
  } else {
    return 0;
  }
}