backward_test.cc 14.3 KB
Newer Older
Y
Yu Yang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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
Yu Yang 已提交
15
#include "paddle/framework/backward.h"
D
dongzhihong 已提交
16

Y
Yu Yang 已提交
17 18
#include <gtest/gtest.h>
#include "paddle/framework/op_registry.h"
Y
Yan Chunwei 已提交
19
#include "paddle/operators/net_op.h"
Y
Yu Yang 已提交
20

Y
Yu Yang 已提交
21 22 23
namespace paddle {
namespace framework {

D
dongzhihong 已提交
24 25 26 27 28 29 30
using OperatorBase = framework::OperatorBase;
using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
using OpProto = framework::OpProto;
using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;

Y
Yu Yang 已提交
31
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
Y
Yu Yang 已提交
32
 public:
Y
Yu Yang 已提交
33
  RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
Y
Yu Yang 已提交
34
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
35 36 37
    AddInput("X", "Input X of Add").AsNoGradient();
    AddInput("b", "Bias of Add").AsNoGradient();
    AddOutput("Out", "Out of Add").AsNoGradient();
Y
Yu Yang 已提交
38 39 40 41
    AddComment("Add Op");
  }
};

Y
Yu Yang 已提交
42 43 44 45
class MulOpMaker : public OpProtoAndCheckerMaker {
 public:
  MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
46 47
    AddInput("X", "A");
    AddInput("Y", "B");
Y
Yu Yang 已提交
48 49 50 51 52 53 54 55 56 57
    AddOutput("Out", "Out");
    AddComment("Mul");
  }
};

class SigmoidOpMaker : public OpProtoAndCheckerMaker {
 public:
  SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("X", "X");
Y
Yu Yang 已提交
58
    AddOutput("Out", "Y");
Y
Yu Yang 已提交
59 60 61 62
    AddComment("Sigmoid");
  }
};

D
dongzhihong 已提交
63 64 65 66 67
class NoGradOpMaker : public OpProtoAndCheckerMaker {
 public:
  NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("X", "X input");
Y
Yu Yang 已提交
68
    AddOutput("Out", "Y output");
D
dongzhihong 已提交
69 70 71 72
    AddComment("NoGradOp, same input output. no Grad");
  }
};

D
dongzhihong 已提交
73
class FcOp : public operators::NetOp {
Y
Yu Yang 已提交
74
 public:
Y
Yu Yang 已提交
75 76 77
  FcOp(const std::string &type, const VarNameMap &inputs,
       const VarNameMap &outputs, const AttributeMap &attrs)
      : NetOp(type, inputs, outputs, attrs) {
Y
Yu Yang 已提交
78 79 80
    AddOp(OpRegistry::CreateOp("mul",
                               {{"X", {Input("X")}}, {"Y", {Input("W")}}},
                               {{"Out", {Output("mul_result")}}}, {}));
81
    auto input_b = Inputs("b");
Y
Yu Yang 已提交
82
    std::string before_act = "mul_result";
83
    if (input_b.size() != 0) {
Y
Yu Yang 已提交
84
      AddOp(OpRegistry::CreateOp(
85
          "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
Y
Yu Yang 已提交
86
          {{"Out", {Output("add_result")}}}, {}));
Y
Yu Yang 已提交
87 88 89
      before_act = "add_result";
    } else {
      auto out_varname = Output("add_result");
90 91
      if (out_varname != kEmptyVarName) {
        this->Rename(out_varname, kEmptyVarName);
Y
Yu Yang 已提交
92
      }
Y
Yu Yang 已提交
93
    }
Y
Yu Yang 已提交
94

Y
Yu Yang 已提交
95 96
    AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
                               {{"Out", {Output("Out")}}}, {}));
Y
Yu Yang 已提交
97 98 99 100 101 102 103 104 105 106 107
    CompleteAddOp(false);
  }
};

class FcOpMaker : public OpProtoAndCheckerMaker {
 public:
  FcOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("X", "x");
    AddInput("W", "w");
    AddInput("b", "b");
Y
Yu Yang 已提交
108 109
    AddOutput("mul_result", "").AsIntermediate();
    AddOutput("add_result", "").AsIntermediate();
Y
Yu Yang 已提交
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
    AddOutput("Out", "");
    AddComment("");
  }
};

class ManyOutputOpMaker : public OpProtoAndCheckerMaker {
 public:
  ManyOutputOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("x", "x");
    AddOutput("y", "y");
    AddOutput("z", "z");
    AddComment("");
  }
};

class FillZeroOpMaker : public OpProtoAndCheckerMaker {
 public:
  FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("x", "x");
    AddOutput("out", "out");
    AddComment("");
  }
};
Y
Yu Yang 已提交
135 136 137 138 139

class AddOpMaker : public OpProtoAndCheckerMaker {
 public:
  AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
140
    AddInput("X", "x").AsDuplicable();
Y
Yu Yang 已提交
141 142 143 144
    AddOutput("Y", "y");
    AddComment("");
  }
};
Y
Yu Yang 已提交
145 146 147 148
}  // namespace framework
}  // namespace paddle

namespace f = paddle::framework;
D
dongzhihong 已提交
149
namespace ops = paddle::operators;
Y
Yu Yang 已提交
150
using EnforceNotMet = paddle::platform::EnforceNotMet;
F
fengjiayi 已提交
151 152 153 154 155 156 157
REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
            f::NOP);
REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
F
fengjiayi 已提交
158
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
F
fengjiayi 已提交
159 160
REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
            f::NOP);
Y
Yu Yang 已提交
161

162 163 164 165 166
TEST(Backward, simple_op_grad) {
  auto fwd = f::OpRegistry::CreateOp(
      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
  ASSERT_NE(fwd, nullptr);
  auto gop = f::OpRegistry::CreateGradOp(*fwd);
Q
qiaolongfei 已提交
167 168
  ASSERT_EQ(1UL, gop->Inputs().size());
  ASSERT_EQ("rowwise_add_grad", gop->Type());
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
  ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
  ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
}

TEST(Backward, simple_op_not_need_grad) {
  auto fwd = f::OpRegistry::CreateOp(
      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
  ASSERT_NE(fwd, nullptr);
  auto gop = f::Backward(*fwd, {"x"});
  ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName);

  auto no_input_gop = f::Backward(*fwd, {"x", "b"});
  ASSERT_NE(no_input_gop, nullptr);
  ASSERT_TRUE(no_input_gop->IsNetOp());
  ASSERT_EQ(0UL,
            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
}

TEST(Backward, net_fc_backward_normal) {
  std::shared_ptr<f::OperatorBase> fwd =
      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
                              {{"mul_result", {"mul_res"}},
                               {"add_result", {"add_re"}},
                               {"Out", {"out"}}},
                              {});
  ASSERT_NE(fwd, nullptr);
  std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
  ASSERT_TRUE(gop->IsNetOp());
  auto net = static_cast<ops::NetOp *>(gop.get());

  ASSERT_NO_THROW(net->DebugString());

  ASSERT_EQ(3UL, net->ops_.size());

  f::OperatorBase &d_sigmoid = *net->ops_[0];
Q
qiaolongfei 已提交
204
  ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
205 206

  f::OperatorBase &d_add = *net->ops_[1];
Q
qiaolongfei 已提交
207
  ASSERT_EQ("rowwise_add_grad", d_add.Type());
208 209

  f::OperatorBase &d_mul = *net->ops_[2];
Q
qiaolongfei 已提交
210
  ASSERT_EQ("mul_grad", d_mul.Type());
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229
}

TEST(Backward, net_fc_backward_not_have_b) {
  std::shared_ptr<f::OperatorBase> fwd =
      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}},
                              {{"mul_result", {"mul_res"}},
                               {"add_result", {"add_res"}},
                               {"Out", {"tmp"}}},
                              {});
  ASSERT_NE(fwd, nullptr);
  std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
  ASSERT_TRUE(gop->IsNetOp());
  auto net = static_cast<ops::NetOp *>(gop.get());

  ASSERT_NO_THROW(net->DebugString());

  ASSERT_EQ(2UL, net->ops_.size());

  f::OperatorBase &d_sigmoid = *net->ops_[0];
Q
qiaolongfei 已提交
230
  ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
231 232

  f::OperatorBase &d_mul = *net->ops_[1];
Q
qiaolongfei 已提交
233
  ASSERT_EQ("mul_grad", d_mul.Type());
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
}

TEST(Backward, net_input_of_network_not_need_grad) {
  ops::NetOp net;
  net.AddOp(f::OpRegistry::CreateOp(
      "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
      {{"mul_result", {"mul_tmp_0"}},
       {"add_result", {"add_tmp_0"}},
       {"Out", {"hidden0"}}},
      {}));
  net.AddOp(f::OpRegistry::CreateOp(
      "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
      {{"mul_result", {"mul_tmp_1"}},
       {"add_result", {"add_tmp_1"}},
       {"Out", {"hidden1"}}},
      {}));
  net.CompleteAddOp();
  auto bwd = Backward(net, {"x"});  // x@GRAD is not need.
  ASSERT_TRUE(bwd->IsNetOp());
  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());

  auto output_vars = bwd_net->OutputVars(true);
  std::unordered_set<std::string> all_outputs =
      std::unordered_set<std::string>(output_vars.begin(), output_vars.end());
  all_outputs.erase(f::kEmptyVarName);

  for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
    ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end());
  }

  // Not Generated X
  ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end());

  ASSERT_EQ(2UL, bwd_net->ops_.size());
  ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
  auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
  ASSERT_EQ(3UL, first_fc_grad->ops_.size());
  ASSERT_EQ(f::kEmptyVarName,
            first_fc_grad->ops_[2]->Output(f::GradVarName("X")));
}

TEST(Backward, net_shared_weight) {
  ops::NetOp net;
  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
                                    {{"Out", {"out"}}}, {}));
  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
                                    {{"Out", {"FinalOut"}}}, {}));
  net.CompleteAddOp();

  auto bwd = f::Backward(net, {});
  ASSERT_TRUE(bwd->IsNetOp());
  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
  ASSERT_EQ(3UL, bwd_net->ops_.size());
Q
qiaolongfei 已提交
287
  ASSERT_EQ("add", bwd_net->ops_[2]->Type());
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
}

TEST(Backward, op_register_grad_not_for_network) {
  auto fwd =
      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
                              {{"mul_result", {"mul_out"}},
                               {"add_result", {"add_out"}},
                               {"Out", {"out1"}}},
                              {{"temporary_index", std::vector<int>{0, 1}}});

  ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
}

TEST(Backward, op_all_input_are_not_need) {
  auto fwd = f::OpRegistry::CreateOp(
      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
  auto backward = f::Backward(*fwd, {"x", "b"});
  ASSERT_TRUE(backward->IsNetOp());
  auto net = static_cast<ops::NetOp *>(backward.get());
  ASSERT_TRUE(net->ops_.empty());
}

TEST(Backward, op_all_output_are_not_need) {
  auto fwd = f::OpRegistry::CreateOp(
      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
  auto backward = f::Backward(*fwd, {"out"});
  ASSERT_TRUE(backward->IsNetOp());
  auto net = static_cast<ops::NetOp *>(backward.get());
  ASSERT_TRUE(net->ops_.empty());
}

TEST(Backward, op_part_of_output_are_not_need) {
  auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
                                     {{"y", {"Y"}}, {"z", {"Z"}}}, {});
  auto backward = f::Backward(*fwd, {"Z"});
  ASSERT_TRUE(backward->IsNetOp());
  auto net = static_cast<ops::NetOp *>(backward.get());
  ASSERT_EQ(net->ops_.size(), 2UL);

  auto &fill_zero = *net->ops_[0];
Q
qiaolongfei 已提交
328
  ASSERT_EQ("fill_zeros_like", fill_zero.Type());
329 330 331 332 333 334
  ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
  ASSERT_EQ("Z", fill_zero.Input("Src"));
  ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));

  auto &d_many_out = *net->ops_[1];
Q
qiaolongfei 已提交
335 336
  ASSERT_EQ("many_output_op_grad", d_many_out.Type());
  ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size());  // I/O/OG
337 338 339 340 341 342 343 344 345 346 347
  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
            d_many_out.Input(f::GradVarName("z")));
  ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
  ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
}

TEST(Backward, op_part_of_input_are_not_need) {
  auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}},
                                     {{"Out", {"out"}}}, {});
  auto backward = f::Backward(*fwd, {"a"});
  auto &grad_mul = *backward;
Q
qiaolongfei 已提交
348 349 350
  ASSERT_EQ(grad_mul.Type(), "mul_grad");
  ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
  ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385
  ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
  ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
  ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
  ASSERT_EQ(grad_mul.Input("X"), "a");
  ASSERT_EQ(grad_mul.Input("Y"), "b");
  ASSERT_EQ(grad_mul.Input("Out"), "out");
}

TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
  ops::NetOp net;
  net.AddOp(f::OpRegistry::CreateOp(
      "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
      {{"mul_result", {"mul_out1"}},
       {"add_result", {"add_out1"}},
       {"Out", {"out1"}}},
      {}));
  net.AddOp(f::OpRegistry::CreateOp(
      "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
      {{"mul_result", {"mul_out2"}},
       {"add_result", {"tmp_out2"}},
       {"Out", {"out2"}}},
      {}));
  net.AddOp(f::OpRegistry::CreateOp(
      "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
      {{"mul_result", {"mul_out3"}},
       {"add_result", {"tmp_out3"}},
       {"Out", {"out3"}}},
      {}));
  net.CompleteAddOp();

  auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
  ASSERT_TRUE(backward->IsNetOp());
  auto bwd_net = static_cast<ops::NetOp *>(backward.get());
  ASSERT_EQ(bwd_net->ops_.size(), 3UL);
  auto &grad_fc = *bwd_net->ops_[0];
Y
Yu Yang 已提交
386 387

  const char *all = paddle::operators::NetOp::kAll;
Q
qiaolongfei 已提交
388
  EXPECT_EQ(grad_fc.Inputs(all).size(),
389 390 391 392
            2UL       /* external input number */
                + 1UL /* external output number*/
                + 1UL /* number of gradient of external output*/
                + 2U /* internal variable number*/);
Q
qiaolongfei 已提交
393
  EXPECT_EQ(grad_fc.Outputs(all).size(),
394 395 396 397
            2UL       /* input number of mul*/
                + 2UL /* input number of rowwise_add
                       */
                + 1UL /* input number of sigmod */);
Q
qiaolongfei 已提交
398 399 400 401
  EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
  EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
  EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
  EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
402
}