recurrent_op_test.cc 8.1 KB
Newer Older
Y
Yan Chunwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/*
  Copyright (c) 2016 PaddlePaddle Authors. 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.
*/

Y
Yan Chunwei 已提交
14 15
#include "paddle/operators/recurrent_op.h"

Y
Yan Chunwei 已提交
16 17 18
#include <glog/logging.h>
#include <gtest/gtest.h>

Y
Yi Wang 已提交
19
#include "paddle/framework/ddim.h"
Y
Yan Chunwei 已提交
20 21 22
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/tensor.h"
Y
Yan Chunwei 已提交
23
#include "paddle/operators/net_op.h"
Y
Yan Chunwei 已提交
24

25 26
namespace paddle {
namespace operators {
Y
Yu Yang 已提交
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
using namespace paddle::framework;

class RecurrentGradientAlgorithmTest : public ::testing::Test {
 protected:
  virtual void SetUp() override {
    CreateGlobalVariables();
    CreateStepScopes();
    CreateStepNet();
    CreateRNNGradientAlgorithm();

    // segment inputs
    SegmentInputs();
    // link forward memories
    LinkeMemories();
  }

  virtual void TearDown() override {}

  void CreateGlobalVariables() {
    // inputs: x
    LOG(INFO) << "create global variable x";
    Variable* x = scope_.NewVar("x");
    DDim dims =
        make_ddim({10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/});
    x->GetMutable<Tensor>()->mutable_data<float>(dims, platform::CPUPlace());
    // inputs: h_boot
    LOG(INFO) << "create global variable h_boot";
    Variable* h_boot = scope_.NewVar("h_boot");
    h_boot->GetMutable<Tensor>()->mutable_data<float>(
        make_ddim({20 /*batch size*/, 30 /*input dim*/}), platform::CPUPlace());
    // inputs: w
    LOG(INFO) << "create global variable w";
    Variable* w = scope_.NewVar("rnn/w");
    w->GetMutable<Tensor>()->mutable_data<float>(make_ddim({30, 30}),
                                                 platform::CPUPlace());
    // inputs: h_grad
    LOG(INFO) << "create variable h_grad";
    Variable* dh = scope_.NewVar("h_grad");
    dh->GetMutable<Tensor>()->mutable_data<float>(make_ddim({10, 20, 30}),
                                                  platform::CPUPlace());
    // inputs: step_scopes
    LOG(INFO) << "create variable step_scopes";
    scope_.NewVar("step_scopes");
    // inputs: step_net
    LOG(INFO) << "create variable step_net";
    scope_.NewVar("step_net");
    // outputs: w_grad
    LOG(INFO) << "create global variable w_grad";
    scope_.NewVar("rnn/w_grad");
    // outputs: x_grad
    LOG(INFO) << "create global variable x_grad";
    scope_.NewVar("x_grad");
    // outputs: h_boot_grad
    LOG(INFO) << "create global variable h_boot_grad";
    scope_.NewVar("h_boot_grad");
  }

  void CreateStepScopes() {
    auto step_scopes =
        scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
    for (int i = 0; i < 10; ++i) {
      auto& scope = scope_.NewScope();
      auto pre_t = scope.NewVar("rnn/pre_h")->GetMutable<Tensor>();
      pre_t->mutable_data<float>({20, 30}, platform::CPUPlace());
      auto tensor = scope.NewVar("rnn/h")->GetMutable<Tensor>();
      tensor->mutable_data<float>({20, 30}, platform::CPUPlace());

      // for unit test of ConcatOutputs
      auto xg = scope.NewVar("rnn/x_grad")->GetMutable<Tensor>();
      xg->mutable_data<float>({20, 30}, platform::CPUPlace());

      step_scopes->emplace_back(&scope);
    }

    // last time step
    auto g = (*step_scopes)[9]->NewVar("rnn/h_pre_grad")->GetMutable<Tensor>();
    g->mutable_data<float>({20, 30}, platform::CPUPlace());
  }

  void CreateRNNGradientAlgorithm() {
    std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
    arg->step_net = "step_net";
    arg->step_scopes = "step_scopes";
    rnn::Link inlink;
    inlink.external = "h_grad";
    inlink.internal = "rnn/h_grad";
    arg->inlinks = std::vector<rnn::Link>{inlink};

    rnn::Link outlink;
    outlink.external = "x_grad";
    outlink.internal = "rnn/x_grad";
    arg->outlinks = std::vector<rnn::Link>{outlink};

    rnn::MemoryAttr mem_attr;
    mem_attr.pre_var = "rnn/h_pre_grad";
    mem_attr.var = "rnn/h_grad";
    mem_attr.boot_var = "h_boot_grad";
    arg->memories = std::vector<rnn::MemoryAttr>{mem_attr};

    rnn_grad_algo_.Init(std::move(arg));
  }

  void CreateStepNet() {
    LOG(INFO) << "create variable step_net";
    Variable* var = scope_.NewVar("step_net");
    auto net = var->GetMutable<NetOp>();
    // TODO(qingqing) modify backward op create for RNNOp unit test
    // and the unit test will be removed to Python.
    // net->AddOp(OpRegistry::CreateOp("mul", {"X", {"rnn/h_pre", "rnn/w",
    //     "rnn/s_grad"}}, {"Y", {"rnn/h_pre_grad", "rnn/w_grad"}}, {}));

    // net->AddOp(OpRegistry::CreateOp("add_two", {"X", {"rnn/h_grad"}},
    //			    {"Y", {"rnn/x_grad"}}, {"Out", "rnn/s_grad"}}, {}));
    net->CompleteAddOp();
  }

  void SegmentInputs() {
    LOG(INFO) << "segment inputs";
    std::vector<std::string> inlinks = {"x"};
    std::vector<std::string> inlinks_alias = {"rnn/x"};

    rnn::Link inlink;
    inlink.external = "x";
    inlink.internal = "rnn/x";
    auto step_scopes =
        scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
    rnn::SegmentInputs(*step_scopes, std::vector<rnn::Link>{inlink}, 10,
                       true /*infer_shape_mode*/);
  }

  void LinkeMemories() {
    LOG(INFO) << "link memories";
    rnn::MemoryAttr mem_attr;
    mem_attr.pre_var = "rnn/h_pre";
    mem_attr.var = "rnn/h";
    mem_attr.boot_var = "boot_h";
    std::vector<rnn::MemoryAttr> memories;
    memories.push_back(mem_attr);
    auto step_scopes =
        scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
    for (int i = 1; i < 10; ++i) {
      rnn::LinkMemories(*step_scopes, memories, i, -1,
                        true /*infer_shape_mode*/);
    }
  }

  Scope scope_;
  RecurrentGradientAlgorithm rnn_grad_algo_;
};

// TEST_F(RecurrentGradientAlgorithmTest, Run) {
//   platform::CPUDeviceContext ctx;
//   rnn_grad_algo_.Run(scope_, ctx);
// }

}  // namespace operators
}  // namespace paddle

TEST(RecurrentOp, LinkMemories) {
  using namespace paddle::framework;
  using namespace paddle::platform;
  using namespace paddle::operators;

  // create and init step scopes
  size_t len = 10;
  std::vector<Scope*> step_scopes;
  for (size_t i = 0; i < len; ++i) {
    auto scope = new Scope();
    scope->NewVar("pre_h");
    auto tensor = scope->NewVar("h")->GetMutable<Tensor>();
    float* data = tensor->mutable_data<float>({15, 20}, CPUPlace());
    for (size_t j = 0; j < 15 * 20; ++j) {
      data[j] = rand() * (1. / (double)RAND_MAX);
    }
    step_scopes.push_back(scope);
  }

  // create MemoryAttr
  rnn::MemoryAttr mem_attr;
  mem_attr.pre_var = "pre_h";
  mem_attr.var = "h";
  mem_attr.boot_var = "boot_h";
  std::vector<rnn::MemoryAttr> memories;
  memories.push_back(mem_attr);

  for (size_t i = 1; i < len; ++i) {
    rnn::LinkMemories(step_scopes, memories, i, -1, false
                      /*infer_shape_mode*/);
  }
  // check
  for (size_t i = 0; i < len - 1; ++i) {
    const float* a =
        step_scopes[i]->FindVar("h")->GetMutable<Tensor>()->data<float>();
    const float* b = step_scopes[i + 1]
                         ->FindVar("pre_h")
                         ->GetMutable<Tensor>()
                         ->data<float>();
    for (size_t j = 0; j < 15 * 20; ++j) {
      ASSERT_FLOAT_EQ(a[j], b[j]);
    }
  }

  for (int i = len - 2; i >= 0; --i) {
    rnn::LinkMemories(step_scopes, memories, i, 1, false
                      /*infer_shape_mode*/);
  }
  // check
  for (int i = len - 2; i >= 0; --i) {
    const float* a =
        step_scopes[i]->FindVar("pre_h")->GetMutable<Tensor>()->data<float>();
    const float* b =
        step_scopes[i + 1]->FindVar("h")->GetMutable<Tensor>()->data<float>();
    for (size_t j = 0; j < 15 * 20; ++j) {
      ASSERT_FLOAT_EQ(a[j], b[j]);
    }
  }

  for (auto s : step_scopes) {
    delete s;
  }
}

USE_OP(add_two);
USE_OP(mul);
F
Fix bug  
fengjiayi 已提交
252
USE_OP_ITSELF(recurrent_op);