tracer.cc 7.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
// 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.

#include "paddle/fluid/imperative/tracer.h"

namespace paddle {
M
minqiyang 已提交
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
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();
}

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);
}

void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
                   const VarBasePtrMap& outputs, framework::BlockDesc* block,
                   const bool stop_gradient) {
  std::map<std::string, VarBase*> vars;

  framework::OpDesc* op_desc = op->op_desc_;
  VLOG(3) << "tracer tracing " << op_desc->Type();
  op_desc->InferShape(*block);
  op_desc->InferVarType(block);
  std::unique_ptr<framework::OperatorBase> op_base =
      framework::OpRegistry::CreateOp(*op_desc);

  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) {
M
minqiyang 已提交
61
      PADDLE_ENFORCE_NOT_NULL(inp->var_, "op %s input %s nullptr",
M
minqiyang 已提交
62 63
                              op->op_desc_->Type(), inp->var_desc_->Name());

M
minqiyang 已提交
64
      invars.push_back(inp->var_);
M
minqiyang 已提交
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82
      vars[inp->var_desc_->Name()] = inp;
      if (inp->pre_op_) {
        op->pre_ops_[it.first].push_back(inp->pre_op_);
        op->pre_ops_out_idx_[it.first].push_back(inp->pre_op_out_idx_);
      } else {
        op->pre_ops_[it.first].push_back(nullptr);
      }
      VLOG(3) << "input vname " << inp->var_desc_->Name() << " "
              << inp->var_->IsInitialized();
    }
  }

  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];
M
minqiyang 已提交
83
      outvars.push_back(out->var_);
M
minqiyang 已提交
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
      vars[out->var_desc_->Name()] = out;

      framework::VarDesc* var_desc = block->FindVar(out->var_desc_->Name());
      if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
        out->var_->GetMutable<framework::LoDTensor>();
      } else {
        LOG(ERROR) << "tracer doesn't support yet";
      }
      out->stop_gradient_ = stop_gradient;
      out->pre_op_ = op;
      out->pre_op_out_name_ = it.first;
      out->pre_op_out_idx_ = i;

      VLOG(3) << "output vname " << out->var_desc_->Name() << " "
              << out->var_->IsInitialized();
    }
  }

  VLOG(3) << "tracer running " << op_desc->Type();
  framework::RuntimeContext ctx(invars_map, outvars_map);

  // 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));

  if (!stop_gradient) {
    framework::OpDesc* grad_op_desc;
118 119 120 121
    // TODO(panyx): Is this leaked?
    std::unique_ptr<std::unordered_map<std::string, std::string>> grad_to_var(
        new std::unordered_map<std::string, std::string>());
    CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var.get());
M
minqiyang 已提交
122 123 124 125 126 127 128 129 130 131
    op->grad_op_desc_ = grad_op_desc;

    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);
        auto var_it = grad_to_var->find(grad_invar);
        if (var_it == grad_to_var->end()) {
          auto fwd_var_it = vars.find(grad_invar);
          PADDLE_ENFORCE(fwd_var_it != vars.end());
132
          // Forward inputs or outputs.
M
minqiyang 已提交
133
          grad_in_vars.push_back(fwd_var_it->second->var_);
M
minqiyang 已提交
134 135
        } else {
          VarBase* var = vars[var_it->second];
M
minqiyang 已提交
136 137
          if (!var->grads_->var_->IsInitialized()) {
            InitVar(var->var_, var->grads_->var_);
M
minqiyang 已提交
138
          }
139
          // Douts.
M
minqiyang 已提交
140
          grad_in_vars.push_back(var->grads_->var_);
M
minqiyang 已提交
141 142 143 144 145 146 147 148 149 150 151
        }
      }
    }

    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);
        auto var_it = grad_to_var->find(grad_outvar);
        PADDLE_ENFORCE(var_it != grad_to_var->end());
        VarBase* var = vars[var_it->second];
M
minqiyang 已提交
152 153
        if (!var->grads_->var_->IsInitialized()) {
          InitVar(var->var_, var->grads_->var_);
M
minqiyang 已提交
154
        }
M
minqiyang 已提交
155
        grad_out_vars.push_back(var->grads_->var_);
M
minqiyang 已提交
156 157 158 159 160 161 162
      }
    }
  }

  op->block_ = block;
}

163 164 165 166
std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
                                      const std::vector<VarBase*>& inputs,
                                      bool stop_gradient) {
  VLOG(3) << "py_trace";
X
Xin Pan 已提交
167 168
  op->input_vars_[PyLayer::kFwdInp] = inputs;
  op->output_vars_[PyLayer::kFwdOut] = PyLayer::Apply(op->forward_id_, inputs);
169 170
  for (VarBase* inp : inputs) {
    if (inp->pre_op_) {
X
Xin Pan 已提交
171 172
      op->pre_ops_[PyLayer::kFwdInp].push_back(inp->pre_op_);
      op->pre_ops_out_idx_[PyLayer::kFwdInp].push_back(inp->pre_op_out_idx_);
173
    } else {
X
Xin Pan 已提交
174
      op->pre_ops_[PyLayer::kFwdInp].push_back(nullptr);
175 176 177
    }
  }

X
Xin Pan 已提交
178
  auto& outputs = op->output_vars_[PyLayer::kFwdOut];
179 180 181 182
  for (size_t i = 0; i < outputs.size(); ++i) {
    VarBase* out = outputs[i];
    out->stop_gradient_ = stop_gradient;
    out->pre_op_ = op;
X
Xin Pan 已提交
183
    out->pre_op_out_name_ = PyLayer::kFwdOut;
184 185 186
    out->pre_op_out_idx_ = i;
  }
  if (!stop_gradient) {
X
Xin Pan 已提交
187 188 189 190
    auto& grad_input_vars =
        op->grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)];
    auto& grad_output_vars =
        op->grad_output_vars_[framework::GradVarName(PyLayer::kFwdOut)];
191 192 193 194 195 196 197 198

    for (const VarBase* inp : inputs) {
      grad_input_vars.push_back(inp->var_);
    }
    for (VarBase* out : outputs) {
      grad_input_vars.push_back(out->var_);
    }
    for (VarBase* out : outputs) {
M
minqiyang 已提交
199
      grad_input_vars.push_back(out->grads_->var_);
200 201 202 203 204
      if (!grad_input_vars.back()->IsInitialized()) {
        InitVar(out->var_, grad_input_vars.back());
      }
    }
    for (const VarBase* inp : inputs) {
M
minqiyang 已提交
205
      grad_output_vars.push_back(inp->grads_->var_);
206 207 208 209 210 211 212 213
      if (!grad_output_vars.back()->IsInitialized()) {
        InitVar(inp->var_, grad_output_vars.back());
      }
    }
  }
  return outputs;
}

M
minqiyang 已提交
214
}  // namespace imperative
215
}  // namespace paddle