backward_test.cc 14.5 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 32
class EmptyOp : public OperatorBase {
 public:
33
  DEFINE_OPERATOR_CTOR(EmptyOp, OperatorBase);
Y
Yu Yang 已提交
34
  void InferShape(const Scope &scope) const override {}
D
dongzhihong 已提交
35
  void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
Y
Yu Yang 已提交
36 37
};

Y
Yu Yang 已提交
38
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
Y
Yu Yang 已提交
39
 public:
Y
Yu Yang 已提交
40
  RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
Y
Yu Yang 已提交
41
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
42 43 44
    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
45 46 47 48
    AddComment("Add Op");
  }
};

Y
Yu Yang 已提交
49 50 51 52
class MulOpMaker : public OpProtoAndCheckerMaker {
 public:
  MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
53 54
    AddInput("X", "A");
    AddInput("Y", "B");
Y
Yu Yang 已提交
55 56 57 58 59 60 61 62 63 64
    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
65
    AddOutput("Out", "Y");
Y
Yu Yang 已提交
66 67 68 69
    AddComment("Sigmoid");
  }
};

D
dongzhihong 已提交
70 71 72 73 74
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
75
    AddOutput("Out", "Y output");
D
dongzhihong 已提交
76 77 78 79
    AddComment("NoGradOp, same input output. no Grad");
  }
};

D
dongzhihong 已提交
80
class FcOp : public operators::NetOp {
Y
Yu Yang 已提交
81
 public:
82
  DEFINE_OPERATOR_CTOR(FcOp, operators::NetOp)
Y
Yu Yang 已提交
83
  void Init() override {
Y
Yu Yang 已提交
84 85 86
    AddOp(OpRegistry::CreateOp("mul",
                               {{"X", {Input("X")}}, {"Y", {Input("W")}}},
                               {{"Out", {Output("mul_result")}}}, {}));
87
    auto input_b = Inputs("b");
Y
Yu Yang 已提交
88
    std::string before_act = "mul_result";
89
    if (input_b.size() != 0) {
Y
Yu Yang 已提交
90
      AddOp(OpRegistry::CreateOp(
91
          "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
Y
Yu Yang 已提交
92
          {{"Out", {Output("add_result")}}}, {}));
Y
Yu Yang 已提交
93 94 95
      before_act = "add_result";
    } else {
      auto out_varname = Output("add_result");
96 97
      if (out_varname != kEmptyVarName) {
        this->Rename(out_varname, kEmptyVarName);
Y
Yu Yang 已提交
98
      }
Y
Yu Yang 已提交
99
    }
Y
Yu Yang 已提交
100

Y
Yu Yang 已提交
101 102
    AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
                               {{"Out", {Output("Out")}}}, {}));
Y
Yu Yang 已提交
103 104 105 106 107 108 109 110 111 112 113
    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
114 115
    AddOutput("mul_result", "").AsIntermediate();
    AddOutput("add_result", "").AsIntermediate();
Y
Yu Yang 已提交
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
    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
141 142 143 144 145

class AddOpMaker : public OpProtoAndCheckerMaker {
 public:
  AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
146
    AddInput("X", "x").AsDuplicable();
Y
Yu Yang 已提交
147 148 149 150
    AddOutput("Y", "y");
    AddComment("");
  }
};
Y
Yu Yang 已提交
151 152 153 154
}  // namespace framework
}  // namespace paddle

namespace f = paddle::framework;
D
dongzhihong 已提交
155
namespace ops = paddle::operators;
Y
Yu Yang 已提交
156 157 158 159 160 161 162
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
D
dongzhihong 已提交
163
REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
Y
Yu Yang 已提交
164
REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
Y
Yu Yang 已提交
165 166
REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
D
dongzhihong 已提交
167 168 169
REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
Y
Yu Yang 已提交
170

171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394
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);
  ASSERT_EQ(1UL, gop->inputs_.size());
  ASSERT_EQ("rowwise_add_grad", gop->type_);
  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];
  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);

  f::OperatorBase &d_add = *net->ops_[1];
  ASSERT_EQ("rowwise_add_grad", d_add.type_);

  f::OperatorBase &d_mul = *net->ops_[2];
  ASSERT_EQ("mul_grad", d_mul.type_);
}

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];
  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);

  f::OperatorBase &d_mul = *net->ops_[1];
  ASSERT_EQ("mul_grad", d_mul.type_);
}

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());
  ASSERT_EQ("add", bwd_net->ops_[2]->type_);
}

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];
  ASSERT_EQ("fill_zeros_like", fill_zero.type_);
  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];
  ASSERT_EQ("many_output_op_grad", d_many_out.type_);
  ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
  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;
  ASSERT_EQ(grad_mul.type_, "mul_grad");
  ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
  ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
  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
395 396 397

  const char *all = paddle::operators::NetOp::kAll;
  EXPECT_EQ(grad_fc.inputs_[all].size(),
398 399 400 401
            2UL       /* external input number */
                + 1UL /* external output number*/
                + 1UL /* number of gradient of external output*/
                + 2U /* internal variable number*/);
Y
Yu Yang 已提交
402
  EXPECT_EQ(grad_fc.outputs_[all].size(),
403 404 405 406
            2UL       /* input number of mul*/
                + 2UL /* input number of rowwise_add
                       */
                + 1UL /* input number of sigmod */);
Y
Yu Yang 已提交
407 408 409 410
  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);
411
}