tracer.cc 10.3 KB
Newer Older
J
Jiabin Yang 已提交
1
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14
//
// 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/fluid/imperative/tracer.h"
M
minqiyang 已提交
15
#include <unordered_set>
16
#include <utility>
C
chengduo 已提交
17
#include "paddle/fluid/platform/profiler.h"
M
minqiyang 已提交
18

19
namespace paddle {
M
minqiyang 已提交
20 21
namespace imperative {

J
Jiabin Yang 已提交
22 23 24 25 26 27 28 29 30 31
static std::vector<std::unique_ptr<framework::OpDesc>> CreateGradOpDescs(
    const framework::OpInfo& op_info, const framework::OpDesc& op_desc,
    const std::unordered_set<std::string>& no_grad_set,
    const std::vector<framework::BlockDesc*>& grad_sub_block,
    std::unordered_map<std::string, std::string>* grad_to_var) {
  if (op_info.grad_op_maker_) {
    return op_info.grad_op_maker_(op_desc, no_grad_set, grad_to_var,
                                  grad_sub_block);
  } else {
    return {};
X
Xin Pan 已提交
32
  }
M
minqiyang 已提交
33 34
}

35 36 37 38 39 40 41 42 43 44
static void PassStopGradient(const NameVarBaseMap& outs, bool generate_grad) {
  for (const auto& name_pair : outs) {
    for (const auto& vb : name_pair.second) {
      VLOG(6) << "Set output: " << vb->Name() << "'s OverridedStopGradient as "
              << generate_grad;
      vb->InnerSetOverridedStopGradient(generate_grad);
    }
  }
}

J
Jiabin Yang 已提交
45 46 47 48 49 50 51 52 53 54 55 56 57
void Tracer::TraceOp(const std::string& type, const NameVarBaseMap& ins,
                     const NameVarBaseMap& outs, framework::AttributeMap attrs,
                     const platform::Place& place, bool trace_backward) {
  platform::RecordEvent event(type);
  VLOG(1) << "Trace Op: " << type;
  size_t op_id = GenerateUniqueId();
  auto op = OpBase::Create(op_id, type, ins, outs, std::move(attrs), place);
  op->Run(ins, outs);

  if (ComputeRequiredGrad(ins, outs, trace_backward)) {
    TraceBackward(op, framework::OpDesc(op->Type(), op->InputNameMap(),
                                        op->OutputNameMap(), op->Attrs()),
                  ins, outs);
58 59
  } else {
    VLOG(3) << "No Grad to track for Op: " << type;
60
  }
M
minqiyang 已提交
61 62
}

J
Jiabin Yang 已提交
63
bool Tracer::ComputeRequiredGrad(const NameVarBaseMap& ins,
64
                                 const NameVarBaseMap& outs,
J
Jiabin Yang 已提交
65
                                 bool trace_backward) {
66 67 68 69 70 71 72 73 74 75 76 77 78
  if (!trace_backward) return false;

  for (const auto& name_pair : ins) {
    for (const auto& var_base : name_pair.second) {
      if (!var_base->OverridedStopGradient()) {
        VLOG(6) << "Find out input: " << var_base->Name()
                << "'s GeneratedGrad is True";
        PassStopGradient(outs, var_base->OverridedStopGradient());
        return true;
      }
    }
  }
  return false;
M
minqiyang 已提交
79 80
}

J
Jiabin Yang 已提交
81 82 83 84 85 86 87
void Tracer::TraceBackward(const std::shared_ptr<OpBase>& fwd_op,
                           const framework::OpDesc& fwd_op_desc,
                           const NameVarBaseMap& ins,
                           const NameVarBaseMap& outs) {
  // grad_to_var is a map of framework::GradVarName(in_var_name/out_var_name) ->
  // in_var_name/out_var_name
  std::unordered_map<std::string, std::string> grad_to_var;
88

J
Jiabin Yang 已提交
89 90 91
  // Get grad_op_desc using fwd_op_desc
  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs_ =
      CreateGradOpDescs(fwd_op->Info(), fwd_op_desc, {}, {}, &grad_to_var);
92

J
Jiabin Yang 已提交
93
  // Create grad_ops using grad_op_descs
94

J
Jiabin Yang 已提交
95
  size_t grad_op_num = grad_op_descs_.size();
96

J
Jiabin Yang 已提交
97 98
  VLOG(3) << "Create " << grad_op_num << " grad op desc(s) to op "
          << fwd_op->Type();
99

J
Jiabin Yang 已提交
100 101
  if (grad_op_num == 0) {
    return;
102
  }
J
Jiabin Yang 已提交
103 104 105 106 107 108 109 110 111 112
  // Build a map to record var_name -> std::shared_ptr<VarBase>*,
  // so that we can find suitable var in grad op descs
  std::unordered_map<std::string, const std::shared_ptr<VarBase>*> name_to_var;
  for (auto& pair : ins) {
    for (auto& var : pair.second) {
      auto& var_ptr = name_to_var[var->Name()];
      PADDLE_ENFORCE_EQ(var_ptr == nullptr || var_ptr->get() == var.get(), true,
                        "There are different variables with same name %s",
                        var->Name());
      var_ptr = &var;
M
minqiyang 已提交
113 114 115
    }
  }

J
Jiabin Yang 已提交
116 117 118 119 120 121 122
  for (auto& pair : outs) {
    for (auto& var : pair.second) {
      auto& var_ptr = name_to_var[var->Name()];
      PADDLE_ENFORCE_EQ(var_ptr == nullptr || var_ptr->get() == var.get(), true,
                        "There are different variables with same name %s",
                        var->Name());
      var_ptr = &var;
M
minqiyang 已提交
123 124 125
    }
  }

J
Jiabin Yang 已提交
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
  // Build backward ins and outs

  for (size_t i = 0; i < grad_op_num; i++) {
    // Step1: build grad op and add them to engine

    // Use trace id to decide the order of gradient sum in sorted sum mode
    size_t trace_id = fwd_op->id();
    std::shared_ptr<OpBase> grad_op =
        OpBase::Create(trace_id, (*(grad_op_descs_[i].get())), fwd_op->place());

    // this OpBase* is just used to manage op's life time
    engine_->InsertOp(grad_op.get(), grad_op);

    std::unordered_set<OpBase*> visited_preceding_ops;
    // Step2 : prepare grad_in vars and bind them with grad_op,
    // set inputs' grad_op as current grad_op
    for (const auto& grad_ins : grad_op_descs_[i]->Inputs()) {
      if (grad_ins.second.empty()) continue;
      auto& bwd_in = (*grad_op->GetMutableInsMap())[grad_ins.first];
      bwd_in.reserve(grad_ins.second.size());

      for (auto& grad_in_var_name : grad_ins.second) {
        auto iter = grad_to_var.find(grad_in_var_name);

        if (iter != grad_to_var.end()) {
          // If it is a grad var, find its coresponding forward var
          auto& fwd_var_name = iter->second;
          auto fwd_var_iter = name_to_var.find(fwd_var_name);
          PADDLE_ENFORCE_EQ(fwd_var_iter != name_to_var.end(), true,
                            "Cannot find forward variable named %s",
                            fwd_var_name);
157
          const auto& tmp = (*(fwd_var_iter->second))->GradVarBase();
J
Jiabin Yang 已提交
158
          PADDLE_ENFORCE_NOT_NULL(
159
              tmp.get(),
J
Jiabin Yang 已提交
160 161 162
              "Grad of %s should "
              "not be NULL when we Track_Backward Input of %s",
              (*(fwd_var_iter->second))->Name(), grad_op->Type());
163 164 165 166 167 168 169 170
          // Create grad_in's dim in tensor for Grad Dependency compute
          auto* tensor = tmp->MutableVar()->GetMutable<framework::LoDTensor>();
          tensor->Resize((*(fwd_var_iter->second))
                             ->Var()
                             .Get<framework::LoDTensor>()
                             .dims());
          // Add Grad Op for grad_in
          tmp->AddGradOps(grad_op);
J
Jiabin Yang 已提交
171 172
          VLOG(3) << "Add Grad Op " << grad_op->Type() << " for :"
                  << (*(fwd_var_iter->second))->GradVarBase()->Name();
173 174 175
          // Add Grad var input to engine set
          engine_->InsertGradVar(tmp.get());
          VLOG(3) << "Add Grad: " << tmp->Name() << " in to Engine";
J
Jiabin Yang 已提交
176 177 178 179 180 181 182 183 184
          bwd_in.emplace_back((*(fwd_var_iter->second))->GradVarBase());
        } else {
          // If it is a forward var, just add it
          auto fwd_var_iter = name_to_var.find(grad_in_var_name);
          PADDLE_ENFORCE_EQ(fwd_var_iter != name_to_var.end(), true,
                            "Cannot find forward variable named %s",
                            grad_in_var_name);
          bwd_in.emplace_back(*(fwd_var_iter->second));
        }
185
        VLOG(3) << "Set backward input from fwd var" << grad_ins.first << " of "
J
Jiabin Yang 已提交
186 187 188 189
                << grad_op->Type() << " to be "
                << (bwd_in.back() ? bwd_in.back()->Name() : "nullptr");
      }
    }
190

J
Jiabin Yang 已提交
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
    // Step3: prepare grad_out vars and using their grad_ops to set current
    // grad_op's preceding op
    for (auto& grad_outs : grad_op_descs_[i]->Outputs()) {
      if (grad_outs.second.empty()) continue;
      auto& bwd_out = (*grad_op->GetMutableOutsMap())[grad_outs.first];
      bwd_out.reserve(grad_outs.second.size());

      for (auto& grad_out_var_name : grad_outs.second) {
        auto iter = grad_to_var.find(grad_out_var_name);
        PADDLE_ENFORCE_EQ(iter != grad_to_var.end(), true,
                          "Cannot find output of input grad %s in op %s",
                          grad_out_var_name, fwd_op->Type());
        auto fwd_var_iter = name_to_var.find(iter->second);
        PADDLE_ENFORCE_EQ(fwd_var_iter != name_to_var.end(), true,
                          "Cannot find forward variable named %s",
                          iter->second);
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
        const auto& tmp = (*(fwd_var_iter->second))->GradVarBase();

        PADDLE_ENFORCE_NOT_NULL(tmp.get(),
                                "Grad output: %s of op: %s should not be NULL",
                                (tmp->Name(), grad_op->Type()));

        if ((!tmp->OverridedStopGradient()) || (grad_outs.second.size() > 1)) {
          VLOG(3) << "Set backward output " << grad_outs.first << " of "
                  << grad_op->Type() << " to be " << tmp->Name()
                  << ". Its Overrided Stop_Gradient is: False";
          bwd_out.emplace_back(tmp);
          auto grad_pending_ops =
              (*(fwd_var_iter->second))->GradVarBase()->GradOps();
          if (VLOG_IS_ON(3) && !grad_pending_ops.empty()) {
            VLOG(3) << "Add grad_pending Op of :"
                    << (*(fwd_var_iter->second))->GradVarBase()->Name()
                    << " It's grad_pending Op are: ";
            for (const auto& op : grad_pending_ops) {
              VLOG(3) << op->Type();
            }
X
Xin Pan 已提交
227
          }
228 229 230 231 232 233 234 235
          if (!grad_pending_ops.empty()) {
            for (const auto& op : grad_pending_ops) {
              PADDLE_ENFORCE_NOT_NULL(op,
                                      "No nullptr should be grad_pending op");
              if (visited_preceding_ops.count(op) == 0) {
                visited_preceding_ops.insert(op);
                grad_op->InsertGradPendingOps(op);
              }
J
Jiabin Yang 已提交
236
            }
237 238 239
          } else {
            VLOG(5) << "Hit leaf VarBase"
                    << (*(fwd_var_iter->second))->GradVarBase()->Name();
J
Jiabin Yang 已提交
240 241
          }
        } else {
242 243 244
          VLOG(3) << "Skip backward output " << grad_outs.first << " of "
                  << grad_op->Type() << " Named: " << tmp->Name()
                  << ", since its Overrided Stop_Gradient is: True";
M
minqiyang 已提交
245 246 247
        }
      }
    }
J
Jiabin Yang 已提交
248 249
    // To ensure numeric stability as static graph
    grad_op->SortGradPendingOps();
M
minqiyang 已提交
250 251
  }
}
J
Jiabin Yang 已提交
252

M
minqiyang 已提交
253
}  // namespace imperative
254
}  // namespace paddle