while_op.cc 13.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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 <vector>
#include "paddle/framework/executor.h"
17
#include "paddle/framework/lod_tensor_array.h"
18 19
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
20
#include "paddle/operators/detail/safe_ref.h"
21 22 23 24 25 26 27

namespace paddle {
namespace operators {

using StepScopeVar = std::vector<framework::Scope *>;
using LoDTensor = framework::LoDTensor;

28
constexpr char kStepBlock[] = "sub_block";
29 30 31
constexpr char kCondition[] = "Condition";
constexpr char kStepScopes[] = "StepScopes";
constexpr char kParameters[] = "X";
32 33
constexpr char kParamGrads[] = "X@GRAD";
constexpr char kOutputs[] = "Out";
34 35 36 37 38 39 40 41 42

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

  void Run(const framework::Scope &scope,
43
           const platform::Place &dev_place) const override {
44 45 46 47
    PADDLE_ENFORCE_NOT_NULL(scope.FindVar(Input(kCondition)));
    auto &cond = scope.FindVar(Input(kCondition))->Get<LoDTensor>();
    PADDLE_ENFORCE_EQ(cond.dims(), paddle::framework::make_ddim({1}));

48
    framework::Executor executor(dev_place);
49
    auto *block = Attr<framework::BlockDesc *>(kStepBlock);
50

51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
    auto *program = block->Program();

    auto step_scopes =
        scope.FindVar(Output(kStepScopes))->GetMutable<StepScopeVar>();

    while (cond.data<bool>()[0]) {
      auto &current_scope = scope.NewScope();
      step_scopes->push_back(&current_scope);

      executor.Run(*program, &current_scope, block->ID(),
                   false /*create_local_scope*/);
    }
  }
};

class WhileOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
68
  WhileOpMaker(OpProto *proto, OpAttrChecker *op_checker)
69 70 71 72 73 74 75 76 77
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput(kParameters,
             "A set of variables, which are required by operators inside the "
             "block of While Op.")
        .AsDuplicable();
    AddInput(
        kCondition,
        "(Bool) An scalar. When it's False, the While Op will be terminated.")
        .AsDuplicable();
78
    AddOutput(kOutputs,
79
              "A set of variables, which will be assigned with values "
80
              "generated by the operators inside the block of While Op.")
81 82 83 84 85
        .AsDuplicable();
    AddOutput(kStepScopes,
              "(StepScopeVar) A vector of local scope, which size equals the "
              "step number of While Op. The i'th scope storages temporary "
              "variables generated in the i'th step.");
86 87
    AddAttr<framework::BlockDesc *>(kStepBlock,
                                    "The step block inside WhileOp");
88 89 90 91 92 93 94 95 96 97 98 99 100
    AddComment(R"DOC(
)DOC");
  }
};

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

  void Run(const framework::Scope &scope,
101 102
           const platform::Place &dev_place) const override {
    framework::Executor executor(dev_place);
103
    auto *block = Attr<framework::BlockDesc *>(kStepBlock);
104 105 106 107 108
    auto *program = block->Program();

    auto *step_scopes =
        scope.FindVar(Input(kStepScopes))->GetMutable<StepScopeVar>();

109 110 111 112 113 114
    auto outside_og_names = Inputs(framework::GradVarName(kOutputs));
    auto inside_og_names =
        Attr<std::vector<std::string>>("original_output_grad");

    PADDLE_ENFORCE_EQ(outside_og_names.size(), inside_og_names.size());

115 116
    for (auto cur_scope_iter = step_scopes->rbegin();
         cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) {
117 118 119 120 121 122 123 124 125
      VLOG(3) << "Start backward at time_step "
              << cur_scope_iter - step_scopes->rbegin();
      framework::Scope &cur_scope = **cur_scope_iter;
      // Link OG from outside to inside
      for (size_t i = 0; i < outside_og_names.size(); ++i) {
        auto outside_og_name = outside_og_names[i];
        auto inside_og_name = inside_og_names[i];
        VLOG(10) << "Linking outside " << outside_og_name << " --> inside "
                 << inside_og_name;
126 127 128 129 130 131
        auto &og_outside =
            detail::Ref(scope.FindVar(outside_og_name),
                        "Cannot find Outside Gradient %s", outside_og_name);
        auto &og_inside =
            detail::Ref(cur_scope.Var(inside_og_name),
                        "Cannot find inside gradient %s", inside_og_name);
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
        if (og_outside.Type().hash_code() ==
            typeid(framework::LoDTensor).hash_code()) {
          auto &outside_tensor = og_outside.Get<framework::LoDTensor>();
          auto &inside_tensor =
              detail::Ref(og_inside.GetMutable<framework::LoDTensor>());
          inside_tensor.set_lod(outside_tensor.lod());
          inside_tensor.ShareDataWith(outside_tensor);
        } else if (og_outside.Type().hash_code() ==
                   typeid(framework::LoDTensorArray).hash_code()) {
          auto &outside_array = og_outside.Get<framework::LoDTensorArray>();
          auto &inside_array =
              detail::Ref(og_inside.GetMutable<framework::LoDTensorArray>());
          VLOG(10) << outside_og_name << " size = " << outside_array.size();
          inside_array.resize(outside_array.size());

          for (size_t j = 0; j < inside_array.size(); ++j) {
            VLOG(10) << j << " " << outside_array[j].numel();
            if (outside_array[j].numel() != 0) {
              inside_array[j].set_lod(outside_array[j].lod());
              inside_array[j].ShareDataWith(outside_array[j]);
            } else {
              PADDLE_ENFORCE_EQ(inside_array[j].numel(), 0);
            }
          }
        }
      }

159 160 161 162 163
      executor.Run(*program, *cur_scope_iter, block->ID(), false);

      auto &pg_names = Outputs(kParamGrads);
      auto &p_names = Inputs(kParameters);
      PADDLE_ENFORCE_EQ(pg_names.size(), p_names.size());
164 165
      for (size_t param_id = 0; param_id < pg_names.size(); ++param_id) {
        if (pg_names[param_id] == framework::kEmptyVarName) {
166
          continue;  // parameter doesn't have gradient
167 168
        }
        auto inside_grad_name = framework::GradVarName(p_names[param_id]);
169

170
        //  // TODO(tonyyang-svail): Not sure we need the following
171 172 173 174 175 176 177 178 179 180 181
        //  // 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 (cur_scope_iter == step_scopes->rbegin()) {
          auto *var = (*cur_scope_iter)->FindVar(inside_grad_name);
182
          PADDLE_ENFORCE_NOT_NULL(var, "Can not find var %s", inside_grad_name);
183 184 185
          if (var->IsType<LoDTensor>()) {
            auto &inside_tensor = var->Get<framework::LoDTensor>();
            framework::AttributeMap attrs;
186
            attrs["dtype"] = framework::ToDataType(inside_tensor.type());
187 188 189 190
            attrs["shape"] = framework::vectorize2int(inside_tensor.dims());
            attrs["value"] = 0.0f;

            auto zero_op = framework::OpRegistry::CreateOp(
Y
Yiqun Liu 已提交
191 192
                "fill_constant", framework::VariableNameMap{},
                {{"Out", {pg_names[param_id]}}}, attrs);
193
            zero_op->Run(scope, dev_place);
194 195 196
          }
        }

197
        auto new_inside_name = cur_scope.Rename(inside_grad_name);
198
        auto sum_op = framework::OpRegistry::CreateOp(
199
            "sum", {{"X", {pg_names[param_id], new_inside_name}}},
Y
Yiqun Liu 已提交
200
            {{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
201
        sum_op->Run(cur_scope, dev_place);
202
        cur_scope.Rename(new_inside_name, inside_grad_name);
203 204 205 206 207 208 209 210 211 212
      }
    }
  }
};

class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker {
 public:
  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;

 protected:
213 214
  std::unique_ptr<framework::OpDesc> Apply() const override {
    auto *grad = new framework::OpDesc();
215
    grad->SetType("while_grad");
216
    grad->SetInput(kParameters, Input(kParameters));
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235

    // Not all of IGs will be generated by inner gradient operators of while op.
    // Ignore IGs that is not generated by the inside block.
    auto igs = InputGrad(kParameters, /*do not drop empty gradient*/ false);
    std::unordered_set<std::string> all_outs;
    for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) {
      for (auto &oname : grad_block_[0]->Op(i)->OutputArgumentNames()) {
        all_outs.insert(oname);
      }
    }
    for (auto &each_ig : igs) {
      if (all_outs.find(each_ig) == all_outs.end()) {
        VLOG(10) << "Ignore " << each_ig;
        each_ig = framework::kEmptyVarName;
      }
    }

    grad->SetOutput(framework::GradVarName(kParameters), igs);

236 237 238 239 240
    grad->SetInput(kOutputs, Output(kOutputs));

    // OG should be re-calculated by step blocks, since many outputs of while op
    // do not need to calculate gradients.
    std::unordered_set<std::string> block_ins;
241
    auto *fwd_block = this->grad_block_[0]->ParentBlock();
242 243 244 245 246 247 248
    {
      for (auto &p : Input(kParameters)) {
        block_ins.insert(p);
      }
      for (auto &o : Output(kOutputs)) {
        block_ins.insert(o);
      }
249
    }
250 251 252 253 254 255
    std::unordered_set<std::string> extra_inputs;
    for (size_t i = 0; i < grad_block_[0]->OpSize(); ++i) {
      for (auto &input_name : grad_block_[0]->Op(i)->InputArgumentNames()) {
        if (block_ins.find(input_name) != block_ins.end()) {
          continue;
        }
256 257 258 259 260 261 262

        // If the input of Op is generated by the forward block, do not make it
        // as input again.
        if (fwd_block->FindVar(input_name) != nullptr) {
          continue;
        }

263 264
        extra_inputs.insert(input_name);
      }
265

266 267
      for (auto &output_name : grad_block_[0]->Op(i)->OutputArgumentNames()) {
        block_ins.insert(output_name);
268 269
      }
    }
270 271 272 273 274 275 276

    std::vector<std::string> extra_inputs_list;
    extra_inputs_list.resize(extra_inputs.size());
    std::copy(extra_inputs.begin(), extra_inputs.end(),
              extra_inputs_list.begin());
    grad->SetInput(framework::GradVarName(kOutputs), extra_inputs_list);
    grad->SetInput(kStepScopes, Output(kStepScopes));
277 278
    grad->SetAttrMap(this->Attrs());
    grad->SetBlockAttr(kStepBlock, *grad_block_[0]);
279 280 281
    // record the original output gradient names, since the gradient name of
    // while operator could be renamed.
    grad->SetAttr("original_output_grad", extra_inputs_list);
282

283
    return std::unique_ptr<framework::OpDesc>(grad);
284 285 286
  }
};

287 288
class WhileGradOpVarTypeInference : public framework::VarTypeInference {
 public:
289 290
  void operator()(const framework::OpDesc &op_desc,
                  framework::BlockDesc *block) const override {
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
    auto p_names = op_desc.Input(kParameters);
    auto pg_names = op_desc.Output(framework::GradVarName(kParameters));

    for (size_t i = 0; i < p_names.size(); ++i) {
      auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i]));
      auto *g_var = block->FindVarRecursive(pg_names[i]);
      if (g_var != nullptr) {  // Gradient could be @EMPTY@
        VLOG(5) << "Setting " << pg_names[i] << " following " << p_names[i]
                << " type: " << p_var.GetType();
        g_var->SetType(p_var.GetType());
        g_var->SetDataType(p_var.GetDataType());
      }
    }
  }
};

class WhileGradOpShapeInference : public framework::InferShapeBase {
 public:
  void operator()(framework::InferShapeContext *ctx) const override {
    ctx->HasInputs(kParameters);
    ctx->HasOutputs(framework::GradVarName(kParameters));
    ctx->HasInputs(kOutputs);
    ctx->HasInputs(framework::GradVarName(kOutputs));

    auto p_names = ctx->Inputs(kParameters);
    auto pg_names = ctx->Outputs(kParamGrads);
    auto var_types = ctx->GetInputsVarType(kParameters);
    std::vector<std::string> names_to_set;
    std::vector<framework::DDim> dims_to_set;
    for (size_t i = 0; i < p_names.size(); ++i) {
      if (pg_names[i] == framework::kEmptyVarName) {
        continue;
      }
324
      auto dims = ctx->GetInputsElementDim(kParameters, i);
325
      if (var_types[i] == framework::proto::VarDesc::LOD_TENSOR) {
326
        names_to_set.push_back(pg_names[i]);
327
        dims_to_set.push_back(dims);
328
      } else if (var_types[i] == framework::proto::VarDesc::LOD_TENSOR_ARRAY) {
329 330
        // not sure how to set the dim of LOD_TENSOR_ARRAY
        names_to_set.push_back(pg_names[i]);
331
        dims_to_set.push_back(dims);
332 333 334 335 336 337
      }
    }
    ctx->SetDims(names_to_set, dims_to_set);
  }
};

338 339 340 341 342 343
}  // namespace operators
}  // namespace paddle

REGISTER_OPERATOR(while, paddle::operators::WhileOp,
                  paddle::operators::WhileOpMaker,
                  paddle::operators::WhileGradOpDescMaker);
344 345 346
REGISTER_OPERATOR(while_grad, paddle::operators::WhileGradOp,
                  paddle::operators::WhileGradOpShapeInference,
                  paddle::operators::WhileGradOpVarTypeInference);
新手
引导
客服 返回
顶部