tracer.h 6.4 KB
Newer Older
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
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// 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.

#pragma once

#include <map>
#include <string>
#include <vector>

#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/imperative/engine.h"
#include "paddle/fluid/imperative/layer.h"

namespace paddle {
namespace imperative {

void CreateGradOp(const framework::OpDesc& op_desc,
                  const std::unordered_set<std::string>& no_grad_set,
                  const std::vector<framework::BlockDesc*>& grad_sub_block,
                  framework::OpDesc** grad_op_desc,
                  std::unordered_map<std::string, std::string>* grad_to_var) {
  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
      framework::OpInfoMap::Instance()
          .Get(op_desc.Type())
          .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
  // TODO(panyx0718): Leak?
  *grad_op_desc = grad_op_descs[0].release();
}

X
Xin Pan 已提交
43 44 45 46 47 48 49 50
void InitVar(framework::Variable* var, framework::Variable* grad_var) {
  auto& var_t = var->Get<framework::LoDTensor>();
  float* data =
      grad_var->GetMutable<framework::LoDTensor>()->mutable_data<float>(
          var_t.dims(), platform::CPUPlace());
  std::fill(data, data + var_t.numel(), 0.0);
}

51 52
class Tracer {
 public:
53
  explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {}
54

M
minqiyang 已提交
55
  virtual ~Tracer() {}
56

X
Xin Pan 已提交
57 58 59
  void Trace(OpBase* op,
             const std::map<std::string, std::vector<VarBase*>>& inputs,
             const std::map<std::string, std::vector<VarBase*>>& outputs,
60
             framework::BlockDesc* block) {
X
Xin Pan 已提交
61 62
    std::map<std::string, VarBase*> vars;

63
    framework::OpDesc* op_desc = op->op_desc_;
M
minqiyang 已提交
64
    VLOG(3) << "tracer tracing " << op_desc->Type();
65 66 67 68 69
    op_desc->InferShape(*block);
    op_desc->InferVarType(block);
    std::unique_ptr<framework::OperatorBase> op_base =
        framework::OpRegistry::CreateOp(*op_desc);

X
Xin Pan 已提交
70 71 72 73 74 75 76 77 78 79 80 81 82
    framework::VariableValueMap invars_map;
    framework::VariableValueMap outvars_map;

    op->input_vars_ = inputs;
    for (auto it : op->input_vars_) {
      auto& invars = invars_map[it.first];
      for (VarBase* inp : it.second) {
        PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
                                op->op_desc_->Type(), inp->var_desc_->Name());

        invars.push_back(inp->var_);
        vars[inp->var_desc_->Name()] = inp;
        if (inp->pre_op_) {
X
Xin Pan 已提交
83 84
          op->pre_ops_[it.first].push_back(inp->pre_op_);
          op->pre_ops_out_idx_[it.first].push_back(inp->pre_op_out_idx_);
85
        } else {
X
Xin Pan 已提交
86
          op->pre_ops_[it.first].push_back(nullptr);
87
        }
X
Xin Pan 已提交
88
        VLOG(3) << "input vname " << inp->var_desc_->Name() << " "
X
clean  
Xin Pan 已提交
89
                << inp->var_->IsInitialized();
90 91 92
      }
    }

X
Xin Pan 已提交
93 94 95 96 97 98 99 100 101 102
    op->output_vars_ = outputs;
    for (auto it : op->output_vars_) {
      auto& outvars = outvars_map[it.first];
      const std::vector<VarBase*>& outputs = it.second;
      for (size_t i = 0; i < outputs.size(); ++i) {
        VarBase* out = outputs[i];
        outvars.push_back(out->var_);
        vars[out->var_desc_->Name()] = out;

        framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
103
        if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
X
Xin Pan 已提交
104
          out->var_->GetMutable<framework::LoDTensor>();
105 106 107
        } else {
          LOG(ERROR) << "tracer doesn't support yet";
        }
108
        out->stop_gradient_ = stop_gradient;
X
Xin Pan 已提交
109 110 111
        out->pre_op_ = op;
        out->pre_op_out_name_ = it.first;
        out->pre_op_out_idx_ = i;
M
minqiyang 已提交
112

X
Xin Pan 已提交
113 114
        VLOG(3) << "output vname " << out->var_desc_->Name() << " "
                << out->var_->IsInitialized();
115 116
      }
    }
M
minqiyang 已提交
117

M
minqiyang 已提交
118
    VLOG(3) << "tracer running " << op_desc->Type();
X
Xin Pan 已提交
119
    framework::RuntimeContext ctx(invars_map, outvars_map);
X
Xin Pan 已提交
120 121 122 123 124 125 126 127 128 129 130

    // TODO(panyx0718): Cache p.
    framework::OperatorWithKernel* op_kernel =
        dynamic_cast<framework::OperatorWithKernel*>(op_base.get());
    PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");

    framework::Scope scope;
    platform::CPUPlace place;
    PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
    p.op.RuntimeInferShape(scope, place, ctx);
    p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
X
Xin Pan 已提交
131

M
minqiyang 已提交
132
    if (!stop_gradient) {
M
minqiyang 已提交
133 134 135 136
      framework::OpDesc* grad_op_desc;
      auto grad_to_var = new std::unordered_map<std::string, std::string>();
      CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
      op->grad_op_desc_ = grad_op_desc;
X
Xin Pan 已提交
137 138 139 140 141

      for (auto it : grad_op_desc->Inputs()) {
        auto& grad_in_vars = op->grad_input_vars_[it.first];
        for (const std::string& grad_invar : it.second) {
          block->FindRecursiveOrCreateVar(grad_invar);
X
Xin Pan 已提交
142 143
          auto var_it = grad_to_var->find(grad_invar);
          if (var_it == grad_to_var->end()) {
X
Xin Pan 已提交
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
            auto fwd_var_it = vars.find(grad_invar);
            PADDLE_ENFORCE(fwd_var_it != vars.end());
            grad_in_vars.push_back(fwd_var_it->second->var_);
          } else {
            VarBase* var = vars[var_it->second];
            if (!var->grads_->IsInitialized()) {
              InitVar(var->var_, var->grads_);
            }
            grad_in_vars.push_back(var->grads_);
          }
        }
      }
      for (auto it : grad_op_desc->Outputs()) {
        auto& grad_out_vars = op->grad_output_vars_[it.first];
        for (const std::string& grad_outvar : it.second) {
          block->FindRecursiveOrCreateVar(grad_outvar);
X
Xin Pan 已提交
160 161
          auto var_it = grad_to_var->find(grad_outvar);
          PADDLE_ENFORCE(var_it != grad_to_var->end());
X
Xin Pan 已提交
162 163 164 165 166 167 168
          VarBase* var = vars[var_it->second];
          if (!var->grads_->IsInitialized()) {
            InitVar(var->var_, var->grads_);
          }
          grad_out_vars.push_back(var->grads_);
        }
      }
M
minqiyang 已提交
169
    }
170

171 172 173 174 175 176 177 178 179
    op->block_ = block;
  }

 private:
  framework::BlockDesc* root_block_;
};

}  // namespace imperative
}  // namespace paddle