sgd_op.cc 3.4 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
/* 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/sgd_op.h"

namespace paddle {
namespace operators {

D
dongzhihong 已提交
20
class SGDOp : public framework::OperatorWithKernel {
Y
Yu Yang 已提交
21 22 23
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

Q
qijun 已提交
24
  void InferShape(framework::InferShapeContext* ctx) const override {
25 26 27 28 29 30 31 32
    PADDLE_ENFORCE(ctx->HasInput("Param"),
                   "Input(Param) of SGDOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Grad"),
                   "Input(Grad) of SGDOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                   "Input(LearningRate) of SGDOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                   "Output(ParamOut) of SGDOp should not be null.");
Q
Qiao Longfei 已提交
33

34
    auto lr_dims = ctx->GetInputDim("LearningRate");
35 36
    PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
                      "Learning rate should have 1 element");
37
    auto param_dim = ctx->GetInputDim("Param");
Q
qijun 已提交
38 39
    // TODO(qijun): check dimensions of Param and Grad at complie
    // and run time.
40
    ctx->SetOutputDim("ParamOut", param_dim);
Q
Qiao Longfei 已提交
41 42 43
  }
};

D
dongzhihong 已提交
44
class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
45
 public:
Q
qijun 已提交
46
  SGDOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
47
      : OpProtoAndCheckerMaker(proto, op_checker) {
48 49 50 51
    AddInput("Param", "Input parameter");
    AddInput("LearningRate", "Learning rate of SGD");
    AddInput("Grad", "Input gradient");
    AddOutput("ParamOut", "output parameter");
Q
Qiao Longfei 已提交
52 53 54 55 56 57 58 59 60
    AddComment(R"DOC(

Simplest sgd algorithm.

param_out = param - learning_rate * grad;

)DOC");
  }
};
Q
qijun 已提交
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92

template <typename T>
struct SparseSGDFunctor<platform::CPUPlace, T> {
  void operator()(const platform::DeviceContext& ctx,
                  const framework::SelectedRows& input,
                  const framework::Tensor& learning_rate,
                  framework::Tensor* output) {
    auto in_height = input.height();
    auto out_dims = output->dims();
    PADDLE_ENFORCE_EQ(in_height, out_dims[0]);

    auto& in_value = input.value();
    auto& in_rows = input.rows();

    int64_t in_row_numel = in_value.numel() / in_rows.size();
    PADDLE_ENFORCE_EQ(in_row_numel, output->numel() / in_height);

    auto* in_data = in_value.data<T>();
    auto* out_data = output->data<T>();
    auto* lr = learning_rate.data<T>();

    for (size_t i = 0; i < in_rows.size(); i++) {
      for (int64_t j = 0; j < in_row_numel; j++) {
        out_data[in_rows[i] * in_row_numel + j] -=
            lr[0] * in_data[i * in_row_numel + j];
      }
    }
  }
};

template struct SparseSGDFunctor<platform::CPUPlace, float>;

Q
Qiao Longfei 已提交
93 94 95
}  // namespace operators
}  // namespace paddle

D
dongzhihong 已提交
96
namespace ops = paddle::operators;
F
fengjiayi 已提交
97
REGISTER_OP_WITHOUT_GRADIENT(sgd, ops::SGDOp, ops::SGDOpMaker);
D
dongzhihong 已提交
98 99
REGISTER_OP_CPU_KERNEL(sgd,
                       ops::SGDOpKernel<paddle::platform::CPUPlace, float>);