sum_op.cc 6.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
/* 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/sum_op.h"
#include <vector>
14
#include "paddle/framework/var_type_inference.h"
15
#include "paddle/operators/detail/safe_ref.h"
16 17 18 19 20 21 22 23 24

namespace paddle {
namespace operators {
using framework::Tensor;

class SumOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

25
  void InferShape(framework::InferShapeContext* ctx) const override {
26
    PADDLE_ENFORCE(ctx->HasInputs("X"), "Inputs(X) should not be null");
27

28 29
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of SumOp should not be null.");
30 31 32 33 34
    if (ctx->IsRuntime() &&
        ctx->GetOutputsVarType("Out")[0] ==
            framework::VarDesc::LOD_TENSOR_ARRAY) {
      return;  // skip runtime infershape when is tensor array;
    }
35

36
    auto x_dims = ctx->GetInputsDim("X");
37
    size_t N = x_dims.size();
Q
qijun 已提交
38
    PADDLE_ENFORCE_GT(N, 1, "Input tensors count should > 1.");
39 40

    auto in_dim = x_dims[0];
41 42
    for (size_t i = 1; i < N; i++) {
      auto dim = x_dims[i];
43
      PADDLE_ENFORCE_EQ(in_dim, dim, "Input tensors must have same shape");
Q
qijun 已提交
44
    }
45 46
    ctx->SetOutputDim("Out", in_dim);
    ctx->ShareLoD("X", /*->*/ "Out");
47
  }
48 49

 protected:
Y
Yu Yang 已提交
50
  framework::OpKernelType GetKernelType(
51 52 53
      const framework::ExecutionContext& ctx) const override {
    auto x_vars = ctx.MultiInputVar("X");
    if (x_vars[0]->IsType<framework::LoDTensor>()) {
Y
Yu Yang 已提交
54 55 56
      return framework::OpKernelType(
          framework::ToDataType(x_vars[0]->Get<framework::LoDTensor>().type()),
          ctx.device_context());
57
    } else if (x_vars[0]->IsType<framework::SelectedRows>()) {
Y
Yu Yang 已提交
58 59 60 61
      return framework::OpKernelType(
          framework::ToDataType(
              x_vars[0]->Get<framework::SelectedRows>().value().type()),
          ctx.device_context());
62
    } else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
63 64 65 66 67 68 69
      for (auto& x_var : x_vars) {
        auto& array = x_var->Get<framework::LoDTensorArray>();
        for (auto& each : array) {
          if (each.numel() != 0) {
            return framework::OpKernelType(framework::ToDataType(each.type()),
                                           ctx.device_context());
          }
70 71
        }
      }
72
      PADDLE_THROW("Cannot find the input data type by all input data");
73 74 75 76
    }
    PADDLE_THROW("Unexpected branch. Input type is %s",
                 x_vars[0]->Type().name());
  }
77 78 79 80
};

class SumOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
81
  SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
82
      : OpProtoAndCheckerMaker(proto, op_checker) {
83 84 85
    AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
        .AsDuplicable();
    AddOutput("Out", "(Tensor) The output tensor of sum operator.");
86
    AddComment(R"DOC(
87
Sum operator.
88

89 90 91
This operators sums the input tensors. All the inputs can carry the 
LoD (Level of Details) information. However, the output only shares 
the LoD information with the first input.
92
)DOC");
93 94 95
  }
};

96 97 98 99 100
class SumOpVarTypeInference : public framework::VarTypeInference {
 public:
  void operator()(const framework::OpDescBind& op_desc,
                  framework::BlockDescBind* block) const override {
    auto& inputs = op_desc.Input("X");
101
    auto var_type = framework::VarDesc::SELECTED_ROWS;
102

103 104 105 106 107
    for (auto& name : op_desc.Input("X")) {
      VLOG(10) << name << " "
               << block->FindRecursiveOrCreateVar(name)->GetType();
    }

108 109
    bool any_input_is_lod_tensor = std::any_of(
        inputs.begin(), inputs.end(), [block](const std::string& name) {
110 111
          return block->FindRecursiveOrCreateVar(name)->GetType() ==
                 framework::VarDesc::LOD_TENSOR;
112
        });
113 114

    auto is_tensor_array = [block](const std::string& name) {
115
      return detail::Ref(block->FindRecursiveOrCreateVar(name)).GetType() ==
116 117 118 119 120 121 122 123 124
             framework::VarDesc::LOD_TENSOR_ARRAY;
    };

    bool any_input_is_tensor_array =
        std::any_of(inputs.begin(), inputs.end(), is_tensor_array);
    bool all_inputs_are_tensor_array =
        std::all_of(inputs.begin(), inputs.end(), is_tensor_array);

    if (any_input_is_tensor_array) {
125 126 127 128 129 130 131 132 133 134
      if (!all_inputs_are_tensor_array) {
        std::ostringstream os;
        for (auto& each : inputs) {
          os << "    " << each << " type is "
             << detail::Ref(block->FindRecursiveOrCreateVar(each)).GetType()
             << "\n";
        }
        PADDLE_ENFORCE(all_inputs_are_tensor_array,
                       "Not all inputs are tensor array:\n%s", os.str());
      }
135 136 137
      var_type = framework::VarDesc::LOD_TENSOR_ARRAY;
    } else if (any_input_is_lod_tensor) {
      var_type = framework::VarDesc::LOD_TENSOR;
138 139 140
    }

    auto out_var_name = op_desc.Output("Out").front();
141 142 143 144
    auto& out_var = detail::Ref(block->FindRecursiveOrCreateVar(out_var_name));
    out_var.SetType(var_type);
    auto& in_var = detail::Ref(block->FindVarRecursive(inputs.front()));
    out_var.SetDataType(in_var.GetDataType());
145 146 147
  }
};

148
class SumGradMaker : public framework::GradOpDescMakerBase {
149
 public:
150
  using framework::GradOpDescMakerBase::GradOpDescMakerBase;
151

Y
Yu Yang 已提交
152 153
  std::vector<std::unique_ptr<framework::OpDescBind>> operator()()
      const override {
154
    auto x_grads = InputGrad("X");
Y
Yu Yang 已提交
155
    std::vector<std::unique_ptr<framework::OpDescBind>> grad_ops;
156 157 158 159
    grad_ops.reserve(x_grads.size());
    auto og = OutputGrad("Out");
    std::transform(x_grads.begin(), x_grads.end(), std::back_inserter(grad_ops),
                   [&og](const std::string& x_grad) {
Y
Yu Yang 已提交
160 161 162 163 164 165
                     auto* grad_op = new framework::OpDescBind();
                     grad_op->SetType("scale");
                     grad_op->SetInput("X", og);
                     grad_op->SetOutput("Out", {x_grad});
                     grad_op->SetAttr("scale", 1.0f);
                     return std::unique_ptr<framework::OpDescBind>(grad_op);
166 167
                   });
    return grad_ops;
168 169 170 171 172 173 174
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
175

176 177
REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
                  ops::SumOpVarTypeInference);
178 179
REGISTER_OP_CPU_KERNEL(sum, ops::SumKernel<paddle::platform::CPUPlace, float>,
                       ops::SumKernel<paddle::platform::CPUPlace, double>);
新手
引导
客服 返回
顶部