recurrent_op.cc 24.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Y
Yan Chunwei 已提交
2

L
Luo Tao 已提交
3 4 5
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
Y
Yan Chunwei 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
Y
Yan Chunwei 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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

Y
Yu Yang 已提交
15
#include <vector>
Y
Yi Wang 已提交
16 17
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/op_registry.h"
Y
Yan Chunwei 已提交
18 19 20

namespace paddle {
namespace operators {
Y
Yu Yang 已提交
21 22 23 24 25 26 27
constexpr char kInputs[] = "inputs";
constexpr char kInitialStates[] = "initial_states";
constexpr char kParameters[] = "parameters";
constexpr char kOutputs[] = "outputs";
constexpr char kStepScopes[] = "step_scopes";
constexpr char kExStates[] = "ex_states";
constexpr char kStates[] = "states";
28
constexpr char kStepBlock[] = "sub_block";
Y
Yu Yang 已提交
29 30 31 32 33 34 35
constexpr char kReverse[] = "reverse";
constexpr char kIsTrain[] = "is_train";
#define GRAD_SUFFIX "@GRAD"
constexpr char kInputGrads[] = "inputs" GRAD_SUFFIX;
constexpr char kOutputGrads[] = "outputs" GRAD_SUFFIX;
constexpr char kParamGrads[] = "parameters" GRAD_SUFFIX;
constexpr char kInitStateGrads[] = "initial_states" GRAD_SUFFIX;
Y
Yan Chunwei 已提交
36

Y
Yu Yang 已提交
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
using StepScopeVar = std::vector<framework::Scope *>;

// StepScopes manages scopes inside RNN.
//    StepScopes::CurScope() get the current scope
//    StepScopes::ExScope() get the ex-scope, or scope in previous time step.
//    StepScopes::Next() move to next time step.
//
// if is_train = False, then
//   there are two scopes for the RNN and just support forward.
// else
//   the len(scopes) == seq_len
//
// if is_backward = True, then
//   reversely access scopes
// else
//   access scopes from begin to end.
class StepScopes {
 public:
  StepScopes(const framework::Scope &parent, StepScopeVar *scopes,
             bool is_train, size_t seq_len, bool is_backward = false)
      : counter_(is_backward ? seq_len - 1 : 0UL),
        scopes_(scopes),
        is_train_(is_train),
        is_backward_(is_backward) {
    size_t num_step_scopes = is_train ? seq_len : 2;
    PADDLE_ENFORCE(is_train || !is_backward,
                   "Cannot backward when is not training");
    if (!is_backward_) {
      PADDLE_ENFORCE(scopes->empty());
      scopes->reserve(static_cast<size_t>(num_step_scopes));
      for (size_t i = 0; i < num_step_scopes; ++i) {
        scopes->emplace_back(&parent.NewScope());
      }
Y
Yan Chunwei 已提交
70
    }
Y
Yu Yang 已提交
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
  }

  framework::Scope &CurScope() { return GetScope(counter_); }

  framework::Scope &ExScope() {
    auto &scope = GetScope(is_backward_ ? counter_ + 1 : counter_ - 1);
    return scope;
  }

  void Next() {
    if (is_backward_) {
      --counter_;
    } else {
      ++counter_;
    }
  }

 private:
  framework::Scope &GetScope(size_t scope_id) const {
    if (!is_train_) {
      scope_id %= 2;
    }
    PADDLE_ENFORCE_LT(scope_id, scopes_->size());
    return *(*scopes_)[scope_id];
  }

  size_t counter_;
  StepScopeVar *scopes_;
  bool is_train_;
  bool is_backward_;
};

// Base class for RecurrentOp/RecurrentGradOp
//    Some common protected functions for RecurrentOp/RecurrentGradOp
class RecurrentBase : public framework::OperatorBase {
 public:
  RecurrentBase(const std::string &type,
                const framework::VariableNameMap &inputs,
                const framework::VariableNameMap &outputs,
                const framework::AttributeMap &attrs)
      : OperatorBase(type, inputs, outputs, attrs) {}

 protected:
  // Get SequenceLength from Scope
  //   The sequence length is got from input tensor. The input tensor's
  //   dimension should be [SEQ_LEN, ..., ...]. The first of the tensor's shape
  //   is SEQ_LEN. The second of the tensor's shape could be the batch size or
  //   nested sequence length.
  int64_t GetSequenceLength(const framework::Scope &scope) const {
    // Dim format SEQ_LEN, BATCH_SIZE, ...
    int64_t seq_len = -1;
    auto &all_inputs = Inputs(kInputs);
    PADDLE_ENFORCE(!all_inputs.empty());
    for (auto &iname : all_inputs) {
      auto *var = scope.FindVar(iname);
      PADDLE_ENFORCE(var != nullptr);
      PADDLE_ENFORCE(var->IsType<framework::LoDTensor>());
      auto &dim = var->Get<framework::LoDTensor>().dims();
      if (seq_len == -1) {
        seq_len = dim[0];
      } else {
        PADDLE_ENFORCE_EQ(seq_len, dim[0]);
      }
    }
    return seq_len;
  }

  // for src_tensor, dst_tensor in zip(map(src_scope.FindVar, src_vars),
  //                                   map(dst_scope.Var, dst_vars)):
  //   dst_tensor.ShareDataWith(src_tensor)
  static void LinkTensor(const framework::Scope &src_scope,
                         const std::vector<std::string> &src_vars,
                         framework::Scope *dst_scope,
                         const std::vector<std::string> &dst_vars) {
    LinkTensorWithCallback(
        src_scope, src_vars, dst_scope, dst_vars,
        [&](const framework::Tensor &src, framework::Tensor *dst) {
          dst->ShareDataWith(src);
        });
  }

  // for src_tensor, dst_tensor in zip(map(src_scope.FindVar, src_vars),
  //                                   map(dst_scope.Var, dst_vars)):
  //   callback(src_tensor, &dst_tensor)
  template <typename Callback>
  static void LinkTensorWithCallback(const framework::Scope &src_scope,
                                     const std::vector<std::string> &src_vars,
                                     framework::Scope *dst_scope,
                                     const std::vector<std::string> &dst_vars,
                                     Callback callback) {
    PADDLE_ENFORCE_EQ(src_vars.size(), dst_vars.size());
    for (size_t i = 0; i < dst_vars.size(); ++i) {
      VLOG(10) << "Link " << src_vars[i] << " to " << dst_vars[i];
      AccessTensor(src_scope, src_vars[i], dst_scope, dst_vars[i], callback);
    }
  }

  // for src_tensor, dst_tensor in zip(map(src_scope.FindVar, src_vars),
  //                                   map(dst_scope.FindVar, dst_vars)):
  //   callback(src_tensor, &dst_tensor)
  template <typename Callback>
  static void LinkTensorWithCallback(const framework::Scope &src_scope,
                                     const std::vector<std::string> &src_vars,
                                     const framework::Scope &dst_scope,
                                     const std::vector<std::string> &dst_vars,
                                     Callback callback) {
    PADDLE_ENFORCE_EQ(src_vars.size(), dst_vars.size());
    for (size_t i = 0; i < dst_vars.size(); ++i) {
      VLOG(10) << "Link " << src_vars[i] << " to " << dst_vars[i];
      AccessTensor(src_scope, src_vars[i], dst_scope, dst_vars[i], callback);
    }
  }

  // (seq_len, shape) -> return [seq_len] + list(shape)
  static framework::DDim PrependDims(size_t seq_len,
                                     const framework::DDim &src) {
    auto dims = framework::vectorize(src);
    dims.insert(dims.begin(), static_cast<int64_t>(seq_len));
    return framework::make_ddim(dims);
  }

 private:
  template <typename Callback>
  static void AccessTensor(const framework::Scope &src_scope,
                           const std::string &src_var_name,
                           framework::Scope *dst_scope,
                           const std::string &dst_var_name, Callback callback) {
    auto *src_var = src_scope.FindVar(src_var_name);
    PADDLE_ENFORCE(src_var != nullptr);
    auto &src_tensor = src_var->Get<framework::LoDTensor>();

    auto *dst_var = dst_scope->Var(dst_var_name);
    auto *dst_tensor = dst_var->GetMutable<framework::LoDTensor>();
    callback(src_tensor, dst_tensor);
  }

  template <typename Callback>
  static void AccessTensor(const framework::Scope &src_scope,
                           const std::string &src_var_name,
                           const framework::Scope &dst_scope,
                           const std::string &dst_var_name, Callback callback) {
    auto *src_var = src_scope.FindVar(src_var_name);
    PADDLE_ENFORCE(src_var != nullptr);
    auto &src_tensor = src_var->Get<framework::LoDTensor>();
    auto *dst_var = dst_scope.FindVar(dst_var_name);
    PADDLE_ENFORCE(dst_var != nullptr);
    auto *dst_tensor = dst_var->GetMutable<framework::LoDTensor>();
    callback(src_tensor, dst_tensor);
  }
};

class RecurrentOp : public RecurrentBase {
 public:
  RecurrentOp(const std::string &type, const framework::VariableNameMap &inputs,
              const framework::VariableNameMap &outputs,
              const framework::AttributeMap &attrs)
      : RecurrentBase(type, inputs, outputs, attrs) {}

229 230 231
 private:
  void RunImpl(const framework::Scope &scope,
               const platform::Place &place) const override {
Y
Yu Yang 已提交
232 233 234 235 236
    auto seq_len = static_cast<size_t>(this->GetSequenceLength(scope));
    VLOG(3) << "Static RNN input sequence length = " << seq_len;
    StepScopes scopes = CreateStepScopes(scope, seq_len);
    auto reverse = Attr<bool>(kReverse);

D
dzhwinter 已提交
237
    framework::Executor executor(place);
Y
Yu Yang 已提交
238
    auto *block = Attr<framework::BlockDesc *>(kStepBlock);
D
dzhwinter 已提交
239

Y
Yu Yang 已提交
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
    auto *program = block->Program();

    for (size_t i = 0; i < seq_len; ++i) {
      size_t seq_offset = reverse ? seq_len - i - 1 : i;
      VLOG(3) << "Recurrent operate at the time step " << seq_offset;

      auto &cur_scope = scopes.CurScope();

      // Link outside::input --> inside::input
      //   inside::input = outside::input[seq_offset: seq_offset+1]
      LinkTensorWithCallback(
          scope, Inputs(kInputs), &cur_scope, Inputs(kInputs),
          [&seq_offset](const framework::Tensor &outside,
                        framework::Tensor *inside) {
            inside->ShareDataWith(outside.Slice(seq_offset, seq_offset + 1));
            auto dims = framework::vectorize(inside->dims());
            dims.erase(dims.begin());
            inside->Resize(framework::make_ddim(dims));
          });

      if (i == 0) {
        // Link initial states  --> ex_states
        LinkTensor(scope, Inputs(kInitialStates), &cur_scope,
                   Attr<std::vector<std::string>>(kExStates));
      } else {
        auto &ex_scope = scopes.ExScope();
        // Link ex_scope::state --> cur_scope::ex_state
        LinkTensor(ex_scope, Attr<std::vector<std::string>>(kStates),
                   &cur_scope, Attr<std::vector<std::string>>(kExStates));
      }

      // Every inputs are linked now, execute!
      executor.Run(*program, &cur_scope, block->ID(),
                   false /*create_local_scope*/);

D
dzhwinter 已提交
275
      // get device context from pool
Y
Yu Yang 已提交
276 277 278
      platform::DeviceContextPool &pool =
          platform::DeviceContextPool::Instance();
      auto &dev_ctx = *pool.Get(place);
D
dzhwinter 已提交
279

Y
Yu Yang 已提交
280 281 282 283 284 285 286 287
      // Copy inside::output -> outside::output
      //    outside::output[seq_offset: seq_offset + 1] = inside::output
      this->LinkTensorWithCallback(
          cur_scope, Outputs(kOutputs), scope, Outputs(kOutputs),
          [&](const framework::LoDTensor &src_tensor,
              framework::LoDTensor *dst_tensor) {
            if (i == 0) {  // create output tensor at begin
              dst_tensor->Resize(PrependDims(seq_len, src_tensor.dims()));
D
dzhwinter 已提交
288
              dst_tensor->mutable_data(place, src_tensor.type());
Y
Yu Yang 已提交
289 290 291 292 293
            }

            auto dst_out = dst_tensor->Slice(seq_offset, seq_offset + 1);
            // Explicit copy output since the local RNN scope can be destroyed
            // early.
294
            framework::Copy(src_tensor, place, dev_ctx, &dst_out);
Y
Yu Yang 已提交
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
          });

      scopes.Next();
    }
  }

 private:
  StepScopes CreateStepScopes(const framework::Scope &scope,
                              size_t seq_len) const {
    auto *var = scope.FindVar(Output(kStepScopes));
    PADDLE_ENFORCE(var != nullptr);
    return StepScopes(scope, var->GetMutable<StepScopeVar>(),
                      Attr<bool>(kIsTrain), seq_len);
  }
};

class RecurrentGradOp : public RecurrentBase {
 public:
  RecurrentGradOp(const std::string &type,
                  const framework::VariableNameMap &inputs,
                  const framework::VariableNameMap &outputs,
                  const framework::AttributeMap &attrs)
      : RecurrentBase(type, inputs, outputs, attrs) {}

319 320 321
 private:
  void RunImpl(const framework::Scope &scope,
               const platform::Place &place) const override {
Y
Yu Yang 已提交
322 323 324 325
    auto seq_len = static_cast<size_t>(GetSequenceLength(scope));
    StepScopes scopes = CreateStepScopes(scope, seq_len);
    auto reverse = Attr<bool>(kReverse);

D
dzhwinter 已提交
326
    framework::Executor executor(place);
Y
Yu Yang 已提交
327
    auto *block = Attr<framework::BlockDesc *>(kStepBlock);
D
dzhwinter 已提交
328

Y
Yu Yang 已提交
329 330
    auto *program = block->Program();

D
dzhwinter 已提交
331
    // get device context from pool
Y
Yu Yang 已提交
332 333
    platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
    auto &dev_ctx = *pool.Get(place);
D
dzhwinter 已提交
334

Y
Yu Yang 已提交
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
    for (size_t step_id = 0; step_id < seq_len; ++step_id) {
      size_t seq_offset = reverse ? step_id : seq_len - step_id - 1;
      VLOG(3) << "Recurrent backward operate at the time step " << seq_offset;
      auto &cur_scope = scopes.CurScope();
      // Link outside::output_grads --> inside::output_grads
      //   inside::output_grad = outside::output_grad[seq_offset:seq_offset+1]
      LinkTensorWithCallback(
          scope, Inputs(kOutputGrads), &cur_scope, Inputs(kOutputGrads),
          [&](const framework::Tensor &outside, framework::Tensor *inside) {
            inside->ShareDataWith(outside.Slice(seq_offset, seq_offset + 1));
            auto dims = framework::vectorize(inside->dims());
            dims.erase(dims.begin());
            inside->Resize(framework::make_ddim(dims));
          });
      auto og_set = List2Set(Inputs(kOutputGrads));

      if (VLOG_IS_ON(10)) {
        std::ostringstream sout;
        std::copy(og_set.begin(), og_set.end(),
                  std::ostream_iterator<std::string>(sout, ","));
        VLOG(10) << " RNN output gradients = [" << sout.str() << "]";
      }

      // Link states
      //   if cur_scope::cur_state_grad in out_grads:
      //     cur_scope::cur_state_grad += ex_scope::ex_state_grad
      //   else:
      //     ex_scope::ex_state_grad --> cur_scope::cur_state_grad
      if (step_id != 0) {  // not at beginning
        auto &ex_scope = scopes.ExScope();
        auto ex_state_grads =
            GradVarLists(Attr<std::vector<std::string>>(kExStates));
        auto cur_state_grads =
            GradVarLists(Attr<std::vector<std::string>>(kStates));

        PADDLE_ENFORCE_EQ(ex_state_grads.size(), cur_state_grads.size());
        for (size_t i = 0; i < ex_state_grads.size(); ++i) {
          auto &cur_grad = cur_state_grads[i];
          auto &ex_grad = ex_state_grads[i];
          auto &ex_tensor =
              ex_scope.FindVar(ex_grad)->Get<framework::LoDTensor>();

          VLOG(10) << " RNN link " << cur_grad << " from " << ex_grad;
          auto *cur_grad_var = cur_scope.Var(cur_grad);
          auto cur_grad_tensor =
              cur_grad_var->GetMutable<framework::LoDTensor>();
381
          framework::Copy(ex_tensor, place, dev_ctx, cur_grad_tensor);
Y
Yu Yang 已提交
382
        }
Y
Yan Chunwei 已提交
383
      }
Y
Yu Yang 已提交
384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402

      VLOG(5) << "Recurrent memory linking finished ";
      // Run step block with cur_scope
      executor.Run(*program, &cur_scope, block->ID(),
                   false /*create_local_scope*/);

      VLOG(5) << "executor.Run finished ";

      auto local_var_names = LocalVarNames(cur_scope);

      // Accumulate params
      //   if (step == 0):
      //      outside::param_grad = 0.0
      //   outside::param_grad += inside::param_grad
      {
        auto &pg_names = Outputs(kParamGrads);
        auto &p_names = Inputs(kParameters);
        PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size());

Y
Yu Yang 已提交
403 404
        for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) {
          auto inside_grad_name = framework::GradVarName(p_names[param_id]);
Y
Yu Yang 已提交
405 406 407 408 409 410 411 412 413 414 415 416

          // If does not compute gradient of that variable inside rnn, just
          // continue
          if (local_var_names.find(inside_grad_name) == local_var_names.end()) {
            continue;
          }

          // zero gradient variable in step 0
          if (step_id == 0) {
            auto &inside_tensor = cur_scope.FindVar(inside_grad_name)
                                      ->Get<framework::LoDTensor>();
            framework::AttributeMap attrs;
F
fengjiayi 已提交
417
            attrs["dtype"] = framework::ToDataType(inside_tensor.type());
Y
Yu Yang 已提交
418 419 420 421
            attrs["shape"] = framework::vectorize2int(inside_tensor.dims());
            attrs["value"] = 0.0f;

            auto zero_op = framework::OpRegistry::CreateOp(
Y
Yiqun Liu 已提交
422 423
                "fill_constant", framework::VariableNameMap{},
                {{"Out", {pg_names[param_id]}}}, attrs);
D
dzhwinter 已提交
424
            zero_op->Run(scope, place);
Y
Yu Yang 已提交
425 426
          }

Y
Yu Yang 已提交
427
          auto new_inside_name = cur_scope.Rename(inside_grad_name);
Y
Yu Yang 已提交
428 429 430
          // sum gradient

          auto sum_op = framework::OpRegistry::CreateOp(
Y
Yu Yang 已提交
431
              "sum", {{"X", {pg_names[param_id], new_inside_name}}},
Y
Yiqun Liu 已提交
432
              {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
D
dzhwinter 已提交
433
          sum_op->Run(cur_scope, place);
Y
Yu Yang 已提交
434 435

          cur_scope.Rename(new_inside_name, inside_grad_name);
Y
Yu Yang 已提交
436
        }
Y
Yan Chunwei 已提交
437
      }
Y
Yu Yang 已提交
438 439 440 441 442 443 444 445 446 447 448 449 450
      VLOG(5) << "Accumulate Parameter finished ";

      // Copy input gradient from inside to outside
      //   outside::input_grad[seq_offset: seq_offset + 1] = inside::input_grad
      LinkTensorWithCallback(
          cur_scope, GradVarLists(Inputs(kInputs)), scope, Outputs(kInputGrads),
          [&](const framework::LoDTensor &inside,
              framework::LoDTensor *outside) {
            if (inside.memory_size() == 0) {  // IG is not created.
              return;
            }
            if (step_id == 0) {  // alloc memory
              outside->Resize(PrependDims(seq_len, inside.dims()));
D
dzhwinter 已提交
451
              outside->mutable_data(place, inside.type());
Y
Yu Yang 已提交
452 453 454
            }

            auto dst = outside->Slice(seq_offset, seq_offset + 1);
455
            framework::Copy(inside, place, dev_ctx, &dst);
Y
Yu Yang 已提交
456 457 458 459 460 461 462 463 464 465 466
          });
      VLOG(5) << "Link outside gradient finished ";

      if (step_id + 1 == seq_len) {  // at_end
        // copy initialize states gradient from inside to outside
        LinkTensorWithCallback(
            cur_scope, GradVarLists(Attr<std::vector<std::string>>(kExStates)),
            scope, Outputs(kInitStateGrads),
            [&](const framework::LoDTensor &inside,
                framework::LoDTensor *outside) {
              outside->Resize(inside.dims());
D
dzhwinter 已提交
467
              outside->mutable_data(place, inside.type());
468
              framework::Copy(inside, place, dev_ctx, outside);
Y
Yu Yang 已提交
469 470 471 472
            });
        VLOG(5) << "Link initialize state gradient finished ";
      }
      scopes.Next();
Y
Yan Chunwei 已提交
473 474
    }
  }
Y
Yu Yang 已提交
475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496

 private:
  StepScopes CreateStepScopes(const framework::Scope &scope,
                              size_t seq_len) const {
    auto *var = scope.FindVar(Input(kStepScopes));
    PADDLE_ENFORCE(var != nullptr);
    return StepScopes(scope, var->GetMutable<StepScopeVar>(),
                      Attr<bool>(kIsTrain), seq_len, true /*is_backward*/);
  }

  std::unordered_set<std::string> List2Set(
      const std::vector<std::string> &list) const {
    std::unordered_set<std::string> local_var_name_set;
    local_var_name_set.reserve(list.size());
    for (auto &each : list) {
      local_var_name_set.insert(each);
    }
    return local_var_name_set;
  }

  std::unordered_set<std::string> LocalVarNames(
      const framework::Scope &scope) const {
Y
Yang Yu 已提交
497
    return this->List2Set(scope.LocalVarNames());
Y
Yu Yang 已提交
498 499 500 501 502 503 504 505 506 507 508 509
  }
  static std::vector<std::string> GradVarLists(
      const std::vector<std::string> &var_names) {
    std::vector<std::string> retv;
    retv.reserve(var_names.size());
    std::transform(var_names.begin(), var_names.end(), std::back_inserter(retv),
                   framework::GradVarName);
    return retv;
  }
};

class RecurrentOpProtoMaker : public framework::OpProtoAndCheckerMaker {
510
 public:
511
  RecurrentOpProtoMaker(OpProto *proto, OpAttrChecker *op_checker)
Y
Yan Chunwei 已提交
512
      : OpProtoAndCheckerMaker(proto, op_checker) {
Y
Yu Yang 已提交
513 514 515 516
    AddInput(kInputs, "rnn inputs").AsDuplicable();
    AddInput(kInitialStates, "rnn initial states").AsDuplicable();
    AddInput(kParameters,
             "Parameters are used by step block as its input. However, the "
K
kexinzhao 已提交
517 518
             "input is not a sequence tensor. Every time step, each operator "
             "in step block just use the parameter directly.")
Y
Yu Yang 已提交
519
        .AsDuplicable();
Y
Yu Yang 已提交
520
    AddOutput(kOutputs,
K
kexinzhao 已提交
521
              "The output sequence of RNN. The sequence length must be same.")
Y
Yu Yang 已提交
522
        .AsDuplicable();
Y
Yu Yang 已提交
523
    AddOutput(kStepScopes,
K
kexinzhao 已提交
524
              "StepScopes contain all local variables in each time step.");
Y
Yu Yang 已提交
525 526 527 528 529 530 531 532 533 534 535
    AddAttr<std::vector<std::string>>(kExStates,
                                      string::Sprintf(
                                          R"DOC(The ex-state variable names.
The ex-state means the state value in the ex-timestep or the previous time step
[%s, %s, %s] must be the same order)DOC",
                                          kExStates, kStates, kInitStateGrads));
    AddAttr<std::vector<std::string>>(
        kStates,
        string::Sprintf(
            "The state variable names. [%s, %s, %s] must be the same order",
            kExStates, kStates, kInitStateGrads));
Y
Yu Yang 已提交
536
    AddAttr<framework::BlockDesc *>(kStepBlock, "The step block inside RNN");
Y
Yu Yang 已提交
537 538
    AddAttr<bool>(kReverse, R"DOC(Calculate RNN reversely or not.
By default reverse=False
Y
Yan Chunwei 已提交
539

Y
Yu Yang 已提交
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562
Assume the input data is [A, B, C, D]

if reverse is False:
  the computation of RNN is like
      A          B          C         D
      |          |          |         |
      v          v          v         v
     rnn -----> rnn -----> rnn ----> rnn
      |          |          |         |
      v          v          v         v
      o          o          o         o

if reverse is True
  the computation of RNN is like
      A          B          C         D
      |          |          |         |
      v          v          v         v
     rnn <----- rnn <----- rnn <---- rnn
      |          |          |         |
      v          v          v         v
      o          o          o         o
)DOC").SetDefault(false);
    AddAttr<bool>(kIsTrain, "").SetDefault(true);
K
kexinzhao 已提交
563 564 565 566 567
    AddComment(R"DOC(
Static Length Recurrent Operator.

The static length recurrent operator can only operate on fixed size sequence
data, i.e. in each mini-batch, the sequence length of all inputs are the same.
Y
Yu Yang 已提交
568 569 570 571 572 573 574 575

)DOC");
  }
};

class RecurrentGradOpDescMaker : public framework::SingleGradOpDescMaker {
 public:
  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
Y
Yan Chunwei 已提交
576

Y
Yu Yang 已提交
577
 protected:
Y
Yu Yang 已提交
578 579
  virtual std::unique_ptr<framework::OpDesc> Apply() const {
    auto *grad = new framework::OpDesc();
Y
Yu Yang 已提交
580 581 582 583
    grad->SetType("recurrent_grad");
    for (auto &input_param : this->InputNames()) {
      grad->SetInput(input_param, this->Input(input_param));
      grad->SetOutput(framework::GradVarName(input_param),
584
                      this->InputGrad(input_param, false));
Y
Yu Yang 已提交
585 586 587 588 589 590 591 592 593 594 595 596 597 598 599
    }

    for (auto &output_param : this->OutputNames()) {
      if (output_param == kStepScopes) {
        grad->SetInput(output_param, this->Output(output_param));
        grad->SetInput(framework::GradVarName(output_param),
                       this->Output(output_param));
      } else {
        grad->SetInput(output_param, this->Output(output_param));
        grad->SetInput(framework::GradVarName(output_param),
                       this->OutputGrad(output_param));
      }
    }
    grad->SetAttrMap(this->Attrs());
    grad->SetBlockAttr(kStepBlock, *grad_block_[0]);
Y
Yan Chunwei 已提交
600

Y
Yu Yang 已提交
601
    return std::unique_ptr<framework::OpDesc>(grad);
Y
Yan Chunwei 已提交
602 603 604
  }
};

Y
Yu Yang 已提交
605 606 607 608 609 610 611
class RecurrentGradOpShapeInference : public framework::InferShapeBase {
 public:
  void operator()(framework::InferShapeContext *ctx) const override {
    std::vector<std::string> input{kInputs, kInitialStates};
    std::vector<std::string> output{kOutputs};
    for (auto &s : input) {
      PADDLE_ENFORCE(ctx->HasInputs(s));
612 613 614
      PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(s)),
                     "Cannot find the gradient variable %s",
                     framework::GradVarName(s));
Y
Yu Yang 已提交
615 616 617 618 619 620
    }
    for (auto &s : output) {
      PADDLE_ENFORCE(ctx->HasInputs(s));
    }
    for (auto &s : input) {
      ctx->SetOutputsDim(framework::GradVarName(s), ctx->GetInputsDim(s));
Y
Yan Chunwei 已提交
621
    }
Y
Yu Yang 已提交
622 623 624 625 626 627 628
    if (ctx->HasInputs(kParameters)) {
      PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName(kParameters)));
      ctx->SetOutputsDim(framework::GradVarName(kParameters),
                         ctx->GetInputsDim(kParameters));
    }
  }
};
Y
Yan Chunwei 已提交
629 630 631 632

}  // namespace operators
}  // namespace paddle

Y
Yu Yang 已提交
633 634 635 636 637
REGISTER_OPERATOR(recurrent, paddle::operators::RecurrentOp,
                  paddle::operators::RecurrentOpProtoMaker,
                  paddle::operators::RecurrentGradOpDescMaker);
REGISTER_OPERATOR(recurrent_grad, paddle::operators::RecurrentGradOp,
                  paddle::operators::RecurrentGradOpShapeInference);