rmsprop_op.cc 4.7 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
/* 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/rmsprop_op.h"

namespace paddle {
namespace operators {

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

24
  void InferShape(framework::InferShapeContext *ctx) const override {
25
    PADDLE_ENFORCE(ctx->HasInput("Param"),
K
Kavya Srinet 已提交
26
                   "Input(Param) of RmspropOp should not be null.");
27 28 29 30
    PADDLE_ENFORCE(ctx->HasInput("MeanSquare"),
                   "Input(MeanSquare) of RmspropOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                   "Input(LearningRate) of RmspropOp should not be null.");
31
    PADDLE_ENFORCE(ctx->HasInput("Grad"),
K
Kavya Srinet 已提交
32
                   "Input(Grad) of RmspropOp should not be null.");
33
    PADDLE_ENFORCE(ctx->HasInput("Moment"),
K
Kavya Srinet 已提交
34
                   "Input(Moment) of RmspropOp should not be null.");
35 36 37 38

    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                   "Output(param_out) of RmspropOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("MomentOut"),
39 40 41
                   "Output(Momentum_out) of RmspropOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("MeanSquareOut"),
                   "Output(MeanSquareOut) of RmspropOp should not be null.");
42 43 44 45 46

    auto param_dim = ctx->GetInputDim("Param");
    PADDLE_ENFORCE_EQ(
        param_dim, ctx->GetInputDim("Grad"),
        "Param and grad input of RmspropOp should have the same dimension.");
47 48 49 50 51 52
    PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("Moment"),
                      "Param and Momentum input of RmspropOp "
                      "should have the same dimension.");
    PADDLE_ENFORCE_EQ(param_dim, ctx->GetInputDim("MeanSquare"),
                      "Param and Momentum input of RmspropOp "
                      "should have the same dimension.");
53

K
Kavya Srinet 已提交
54 55 56 57
    auto lr_dim = ctx->GetInputDim("LearningRate");
    PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1,
                      "Learning Rate should be a scalar.");

58 59
    ctx->SetOutputDim("ParamOut", param_dim);
    ctx->SetOutputDim("MomentOut", param_dim);
60
    ctx->SetOutputDim("MeanSquareOut", param_dim);
61 62 63 64 65 66 67 68
  }
};

class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  RmspropOpMaker(framework::OpProto *proto,
                 framework::OpAttrChecker *op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
    AddInput("Param",
             "(Tensor, default Tensor<float>) "
             "Input parameter value that has to be updated");
    AddInput("MeanSquare",
             "(Tensor, default Tensor<float>)"
             " The mean square value that gets updated");
    AddInput("LearningRate",
             "(Tensor, default Tensor<float>) "
             "The learning rate should be a tensor of size 1");
    AddInput("Grad",
             "(Tensor, default Tensor<float>) "
             "Input gradient of the parameter");
    AddInput("Moment",
             "(Tensor, default Tensor<float>) The moment that gets updated");

    AddOutput("ParamOut", "(Tensor) Output updated parameter value");
    AddOutput("MomentOut", "(Tensor) Output updated moment");
    AddOutput("MeanSquareOut", "(Tensor) Output Mean squared updated value");

    AddAttr<float>("epsilon",
                   "(float, default 1e-10) Constant "
                   "for numerical stability.")
91
        .SetDefault(1.0e-10f);
92 93 94
    AddAttr<float>("decay",
                   "(float, default 0.9) "
                   "Discounting factor for coming gradient.")
95
        .SetDefault(0.9f);
96
    AddAttr<float>("momentum", "(float, default 0.0) Constant value")
97
        .SetDefault(0.0f);
98 99 100 101
    AddComment(R"DOC(

RMSprop

102 103 104 105
MeanSquareOut = decay * MeanSquare + (1 - decay) * Grad * Grad
MomentOut = momentum * Moment +
            LearningRate * Grad / sqrt(MeanSquareOut + epsilon)
ParamOut = Param -  MomentOut
106

107
The original slides that proposed RMSprop: Slide 29 of
108 109 110 111 112 113 114 115 116 117 118 119
http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)

)DOC");
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(rmsprop, ops::RmspropOp, ops::RmspropOpMaker);
REGISTER_OP_CPU_KERNEL(rmsprop,
                       ops::RmspropOpKernel<paddle::platform::CPUPlace, float>);