recurrent_network_op.cc 16.5 KB
Newer Older
Y
Yan Chunwei 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
/* 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. */

#include "paddle/operators/recurrent_network_op.h"

#include <glog/logging.h>
#include <cstring>
#include <sstream>

#include "paddle/framework/net.h"
#include "paddle/framework/op_registry.h"
#include "paddle/platform/enforce.h"

namespace paddle {
namespace operators {

namespace rnn {

void SegmentInputs(std::vector<std::shared_ptr<Scope>>& step_scopes,
                   const std::vector<Link>& inlinks,
                   const size_t seq_len) {
  PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided.");
  for (size_t i = 0; i < inlinks.size(); ++i) {
    Tensor* input =
        step_scopes[0]->GetVariable(inlinks[i].external)->GetMutable<Tensor>();
    DDim dims = input->dims();
    PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
                   "all the inlinks must have same length");
    DDim step_dims = slice_ddim(dims, 1, dims.size());
    for (size_t j = 0; j < seq_len; j++) {
      Tensor* step_input = step_scopes[j]
                               ->CreateVariable(inlinks[i].internal)
                               ->GetMutable<Tensor>();
      *step_input = input->Slice<float>(j, j + 1);
      step_input->Resize(step_dims);
    }
  }
}

void ConcatOutputs(std::vector<std::shared_ptr<Scope>>& step_scopes,
                   const std::vector<Link>& outlinks,
                   const size_t seq_len) {
  for (size_t i = 0; i < outlinks.size(); i++) {
    Tensor* output =
        step_scopes[0]->GetVariable(outlinks[i].external)->GetMutable<Tensor>();

    // TODO(qingiqng) remove following code after adding
    // InferShape in RecurrentGradientOp
    DDim step_dims = step_scopes[0]
                         ->GetVariable(outlinks[i].internal)
                         ->GetMutable<Tensor>()
                         ->dims();
    std::vector<int> dims_vec = vectorize(step_dims);
    dims_vec.insert(dims_vec.begin(), seq_len);
    output->mutable_data<float>(make_ddim(dims_vec), platform::CPUPlace());

    for (size_t j = 0; j < seq_len; j++) {
      Tensor* step_output = step_scopes[j]
                                ->GetVariable(outlinks[i].internal)
                                ->GetMutable<Tensor>();
72 73
      // TODO(luotao02) data type and platform::DeviceContext() should set
      // correctly
Y
Yan Chunwei 已提交
74
      (output->Slice<float>(j, j + 1))
75
          .CopyFrom<float>(*step_output, platform::CPUPlace());
Y
Yan Chunwei 已提交
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
    }
  }
}

void LinkMemories(std::vector<std::shared_ptr<Scope>>& scopes,
                  const std::vector<rnn::MemoryAttr>& memories,
                  size_t step_id,
                  int offset) {
  PADDLE_ENFORCE(step_id < scopes.size(),
                 "step [%d] is out of range of step scopes' size [%d]",
                 step_id,
                 scopes.size());
  PADDLE_ENFORCE(static_cast<int>(step_id) + offset >= 0,
                 "offset [%d] must be large than -[%d]",
                 offset,
                 step_id);
  PADDLE_ENFORCE(step_id + offset < scopes.size(),
                 "offset [%d] is out of range, it must be less than (%d - %d)",
                 offset,
                 scopes.size(),
                 step_id);
  std::shared_ptr<Scope> scope = scopes[step_id];
  std::shared_ptr<Scope> linked_scope = scopes[step_id + offset];
  for (auto& attr : memories) {
    auto mem = scope->CreateVariable(attr.pre_var)->GetMutable<Tensor>();
    // maybe share variable is better?
    auto linked_mem = linked_scope->GetVariable(attr.var)->GetMutable<Tensor>();
    mem->ShareDataWith<float>(*linked_mem);

    // TODO(qingqing) remove following code
    // the memory of current step should be allocated in step net
    auto m = scope->CreateVariable(attr.var)->GetMutable<Tensor>();
    // for unit test, as addOp and mulOp are null currently, if not
    // mutable_data, mem.data() in output will be error. We will
    // remove this line after merge the correct addOp and mulOp.
    m->mutable_data<float>(mem->dims(), platform::CPUPlace());
  }
}

void InitArgument(const ArgumentName& name,
                  Argument* arg,
                  const OperatorBase& op) {
  arg->step_net = op.Input(name.step_net);
  arg->step_scopes = op.Output(name.step_scopes);

  auto inlinks = op.Inputs(name.inlinks);
  auto inlink_alias = op.GetAttr<std::vector<std::string>>(name.inlink_alias);
  PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(),
                 "the size of inlinks and inlink_alias don't match:%d,%d",
                 inlinks.size(),
                 inlink_alias.size());
  for (size_t i = 0; i < inlinks.size(); ++i) {
    rnn::Link link;
    link.external = inlinks[i];
    link.internal = inlink_alias[i];
    (arg->inlinks).push_back(link);
  }

  auto outlinks = op.Outputs(name.outlinks);
  auto outlink_alias = op.GetAttr<std::vector<std::string>>(name.outlink_alias);
  PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(),
                 "the size of outlinks and outlink_alias don't match:%d,%d",
                 outlinks.size(),
                 outlink_alias.size());
  for (size_t i = 0; i < outlinks.size(); ++i) {
    rnn::Link link;
    link.external = outlinks[i];
    link.internal = outlink_alias[i];
    (arg->outlinks).push_back(link);
  }

  auto boot_memories = op.Inputs(name.boot_memories);

  // attributes
  auto memories = op.GetAttr<std::vector<std::string>>(name.memories);
  auto pre_memories = op.GetAttr<std::vector<std::string>>(name.pre_memories);

  PADDLE_ENFORCE(memories.size() == boot_memories.size(),
                 "the size of memories, boot_memories don't match:%d,%d",
                 memories.size(),
                 boot_memories.size());
  PADDLE_ENFORCE(pre_memories.size() == boot_memories.size(),
                 "the size of pre_memories, boot_memories don't match:%d,%d",
                 pre_memories.size(),
                 boot_memories.size());
  PADDLE_ENFORCE(memories.size() > 0, "more than 1 memories should be set");

  for (size_t i = 0; i < memories.size(); ++i) {
    rnn::MemoryAttr mem_attr;
    mem_attr.var = memories[i];
    mem_attr.pre_var = pre_memories[i];
    mem_attr.boot_var = boot_memories[i];
    (arg->memories).push_back(mem_attr);
  }
}

}  // namespace rnn

void RecurrentAlgorithm::InferShape(const std::shared_ptr<Scope>& scope) const {
  seq_len_ = scope->GetVariable((arg_->inlinks[0]).external)
                 ->GetMutable<Tensor>()
                 ->dims()[0];
  CreateScopes(scope);
  auto step_scopes = GetStepScopes(scope);

  // SegmentInputs is called in InferShape. The input must hold memory in
  // SegmentInputs. But the other op only set dimension for the output in
  // InferShape. That's a problem. Wether the RNN op needs InferShape or not?
  // Wether the following functions (SegmentInputs, InitMemories, ...) need
  // to rewrite for RNN op?
  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);

  InitMemories(step_scopes[0]);

  PADDLE_ENFORCE(scope->HasVariable(arg_->step_net),
                 "stepnet [%s] is not in scope.",
                 arg_->step_net);
  Variable* net = scope->GetVariable(arg_->step_net);
  PADDLE_ENFORCE(net != nullptr, "failed to get step net");
  // If the InferShape is called in OperatorBase's run function,
  // the rnn op only needs to do InferShape for the first time step
  for (size_t i = 0; i < seq_len_; i++) {
    if (i > 0) {
      rnn::LinkMemories(step_scopes, arg_->memories, i, -1);
    }
    net->GetMutable<NetOp>()->InferShape(step_scopes[i]);
  }

  auto outlinks = arg_->outlinks;
  for (size_t i = 0; i < outlinks.size(); i++) {
    DDim step_dims = step_scopes[0]
                         ->GetVariable(outlinks[i].internal)
                         ->GetMutable<Tensor>()
                         ->dims();
    std::vector<int> dims_vec = vectorize(step_dims);
    // now only support fixed length
    dims_vec.insert(dims_vec.begin(), seq_len_);
    Tensor* output =
        step_scopes[0]->GetVariable(outlinks[i].external)->GetMutable<Tensor>();
    output->Resize(make_ddim(dims_vec));
  }
}

void RecurrentAlgorithm::Run(const std::shared_ptr<Scope>& scope,
                             const platform::DeviceContext& dev_ctx) const {
  auto step_scopes = GetStepScopes(scope);

  Variable* net = scope->GetVariable(arg_->step_net);
  for (size_t step_id = 0; step_id < seq_len_; step_id++) {
    // the link memory is done in InferShape
    // maybe remove following code after testing
    if (step_id > 0) {
      rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1);
    }
    net->GetMutable<NetOp>()->Run(step_scopes[step_id], dev_ctx);
  }

  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_);
}

void RecurrentAlgorithm::CreateScopes(std::shared_ptr<Scope> scope) const {
  // TODO(xxx) Only two scopes are needed for inference, this case will be
  // supported later.
  auto step_scopes = scope->GetVariable(arg_->step_scopes)
                         ->GetMutable<std::vector<std::shared_ptr<Scope>>>();

  if (seq_len_ > step_scopes->size()) {
    for (size_t i = step_scopes->size(); i < seq_len_; ++i) {
      std::shared_ptr<Scope> step_scope = std::make_shared<Scope>(scope);

      // Now all variables in scope must be created outside of op.
      auto net_op = scope->GetVariable(arg_->step_net)->GetMutable<NetOp>();
      for (auto& input : net_op->inputs_) {
        step_scope->CreateVariable(input);
      }
      for (auto& output : net_op->outputs_) {
        step_scope->CreateVariable(output);
      }

      step_scopes->push_back(std::make_shared<Scope>(step_scope));
    }
  }
}

void RecurrentAlgorithm::InitMemories(std::shared_ptr<Scope> step_scope) const {
  for (auto& attr : arg_->memories) {
    Tensor* pre_mem =
        step_scope->CreateVariable(attr.pre_var)->GetMutable<Tensor>();
    PADDLE_ENFORCE(step_scope->HasVariable(attr.boot_var),
                   "memory [%s]'s boot variable [%s] not exists",
                   attr.var,
                   attr.boot_var);
    Tensor* boot_mem =
        step_scope->GetVariable(attr.boot_var)->GetMutable<Tensor>();
    pre_mem->ShareDataWith<float>(*boot_mem);

    // TODO(qingqing) remove following code
    // the memory of current step should be allocated in step net
    // here for unit test
    auto cur_step_mem =
        step_scope->CreateVariable(attr.var)->GetMutable<Tensor>();
    cur_step_mem->mutable_data<float>(boot_mem->dims(), platform::CPUPlace());
  }
}

const rnn::ArgumentName RecurrentOp::kArgName{"step_net",
                                              "step_scopes",
                                              "inlinks",
                                              "outlinks",
                                              "inlink_alias",
                                              "outlink_alias",
                                              "memories",
                                              "pre_memories",
                                              "boot_memories"};

const rnn::ArgumentName RecurrentGradientOp::kArgName{"step_net",
                                                      "step_scopes",
                                                      "outlink@grad",
                                                      "inlink@grad",
                                                      "inlink_alias",
                                                      "outlink_alias",
                                                      "memories",
                                                      "pre_memories",
                                                      "boot_memories@grad"};

void RecurrentOp::Init() {
  OperatorBase::Init();
  std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
  rnn::InitArgument(kArgName, arg.get(), *this);
  alg_.Init(std::move(arg));
}

class RecurrentAlgorithmProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
public:
  RecurrentAlgorithmProtoAndCheckerMaker(OpProto* proto,
                                         OpAttrChecker* op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    const auto& name = RecurrentOp::kArgName;
    // inputs and outputs stored in proto
Y
Yu Yang 已提交
315 316 317 318
    AddInput(name.inlinks, "the input that need to be segmented for each step.")
        .SetMultiple();
    AddInput(name.boot_memories, "variables to initialize memories.")
        .SetMultiple();
Y
Yan Chunwei 已提交
319 320
    AddInput(name.step_net, "network shared by all steps.");

Y
Yu Yang 已提交
321 322
    AddOutput(name.outlinks, "the output that need to concated for all steps.")
        .SetMultiple();
Y
Yan Chunwei 已提交
323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420
    AddOutput(name.step_scopes, "step scopes");

    // Attributes stored in AttributeMap
    AddAttr<std::vector<std::string>>(name.inlink_alias, "alias of inlinks");
    AddAttr<std::vector<std::string>>(name.outlink_alias, "alias of outlinks");
    AddAttr<std::vector<std::string>>(name.pre_memories,
                                      "names of pre-memories");
    AddAttr<std::vector<std::string>>(name.memories, "names of memories");

    AddComment("This is a recurrent group operator.");
  }
};

void RecurrentGradientAlgorithm::Run(
    const std::shared_ptr<Scope>& scope,
    const platform::DeviceContext& dev_ctx) const {
  auto step_scopes = GetStepScopes(scope);
  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);
  PADDLE_ENFORCE(scope->HasVariable(arg_->step_net),
                 "step net is not in scope.");
  Variable* net = scope->GetVariable(arg_->step_net);
  PADDLE_ENFORCE(net != nullptr, "failed to get step net");
  for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
    if (static_cast<size_t>(step_id) != seq_len_ - 1) {
      rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
    }
    net->GetMutable<NetOp>()->Run(step_scopes[step_id], dev_ctx);
  }
  LinkBootMemoryGradients(step_scopes[0]);
  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_);
}

void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
    std::shared_ptr<Scope> step_scope) const {
  for (auto& attr : arg_->memories) {
    Tensor* mem_grad =
        step_scope->CreateVariable(attr.var)->GetMutable<Tensor>();
    PADDLE_ENFORCE(mem_grad != nullptr,
                   "boot_tensor should be retrieved before");
    PADDLE_ENFORCE(step_scope->HasVariable(attr.boot_var),
                   "memory [%s]'s boot variable [%s] not exists",
                   attr.var,
                   attr.boot_var);
    Tensor* boot_mem_grad =
        step_scope->CreateVariable(attr.boot_var)->GetMutable<Tensor>();
    boot_mem_grad->ShareDataWith<float>(*mem_grad);
  }
}

void RecurrentGradientAlgorithm::InferShape(
    const std::shared_ptr<Scope>& scope) const {
  seq_len_ = scope->GetVariable((arg_->inlinks[0]).external)
                 ->GetMutable<Tensor>()
                 ->dims()[0];
  auto step_scopes = GetStepScopes(scope);
  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);

  PADDLE_ENFORCE(scope->HasVariable(arg_->step_net),
                 "step net is not in scope.");
  Variable* net = scope->GetVariable(arg_->step_net);
  PADDLE_ENFORCE(net != nullptr, "failed to get step net");

  for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
    if (static_cast<size_t>(step_id) != seq_len_ - 1) {
      rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
    }
    net->GetMutable<NetOp>()->InferShape(step_scopes[step_id]);
  }

  auto outlinks = arg_->outlinks;
  for (size_t i = 0; i < outlinks.size(); i++) {
    DDim step_dims = step_scopes[0]
                         ->GetVariable(outlinks[i].internal)
                         ->GetMutable<Tensor>()
                         ->dims();
    std::vector<int> dims_vec = vectorize(step_dims);
    // now only support fixed length
    dims_vec.insert(dims_vec.begin(), seq_len_);
    Tensor* output =
        step_scopes[0]->GetVariable(outlinks[i].external)->GetMutable<Tensor>();
    output->Resize(make_ddim(dims_vec));
  }
  LinkBootMemoryGradients(step_scopes[0]);
}

void RecurrentGradientOp::Init() {
  OperatorBase::Init();
  std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
  rnn::InitArgument(kArgName, arg.get(), *this);
  alg_.Init(std::move(arg));
}

}  // namespace operators
}  // namespace paddle

REGISTER_OP(recurrent_op,
            paddle::operators::RecurrentOp,
            paddle::operators::RecurrentAlgorithmProtoAndCheckerMaker);