proximal_adagrad_op.cc 4.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/proximal_adagrad_op.h"
16 17 18 19

namespace paddle {
namespace operators {

D
dzhwinter 已提交
20
using Tensor = framework::Tensor;
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
class ProximalAdagradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext *ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Param"),
                   "Input(Param) of ProximalAdagradOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Moment"),
                   "Input(Moment) of ProximalAdagradOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Grad"),
                   "Input(Grad) of ProximalAdagradOp should not be null.");
    PADDLE_ENFORCE(
        ctx->HasInput("LearningRate"),
        "Input(LearningRate) of ProximalAdagradOp should not be null.");

    PADDLE_ENFORCE(ctx->HasOutput("ParamOut"),
                   "Output(ParamOut) of ProximalAdagradOp should not be null.");
    PADDLE_ENFORCE(
        ctx->HasOutput("MomentOut"),
        "Output(MomentOut) of ProximalAdagradOp should not be null.");

    auto param_dim = ctx->GetInputDim("Param");
    PADDLE_ENFORCE_EQ(
        param_dim, ctx->GetInputDim("Grad"),
        "Param and Grad of ProximalAdagrad Op must have same dimension.");

    PADDLE_ENFORCE_EQ(
        param_dim, ctx->GetInputDim("Moment"),
        "Param and Moment of ProximalAdagrad Op must have same dimension.");

    auto lr_dim = ctx->GetInputDim("LearningRate");
    PADDLE_ENFORCE_EQ(framework::product(lr_dim), 1,
                      "Learning Rate should be a scalar.");

    ctx->SetOutputDim("ParamOut", param_dim);
    ctx->SetOutputDim("MomentOut", param_dim);
  }
D
dzhwinter 已提交
59 60 61 62 63 64
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    auto input_data_type =
        framework::ToDataType(ctx.Input<Tensor>("Param")->type());
    return framework::OpKernelType(input_data_type, ctx.GetPlace());
  }
65 66 67 68
};

class ProximalAdagradOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
69
  void Make() override {
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 that has to be updated.");
    AddInput("Moment",
             "(Tensor, default Tensor<float>) "
             "Moment parameter that has to be updated.");
    AddInput("Grad",
             "(Tensor, default Tensor<float>) "
             "Input gradient of the parameter.");
    AddInput("LearningRate",
             "(Tensor, default Tensor<float>) "
             "The learning rate should be a tensor of size 1.");

    AddOutput("ParamOut", "(Tensor) Output updated parameter value.");
    AddOutput("MomentOut", "(Tensor) Output updated moment value.");

    AddAttr<float>("l1",
                   "(float, default 0.0) "
                   "L1 regularization strength.")
        .SetDefault(0.0f);
    AddAttr<float>("l2",
91
                   "(float, default 0.0) "
92 93 94
                   "L2 regularization strength.")
        .SetDefault(0.0f);
    AddComment(R"DOC(
95
Proximal Adagrad Optimizer.
96

97
Optimizer that implements the proximal adagrad algorithm:
98

99 100 101 102 103 104
$$
moment = moment + grad * grad \\
prox\_param = param - learning\_rate * grad * (1 / \sqrt{moment}) \\
param = sign(prox\_param) / (1 + learning\_rate * l2) *
        \max(|prox\_param| - learning\_rate * l1 , 0)
$$
105 106 107 108 109

The paper that proposed Proximal GD: 
(http://papers.nips.cc/paper/3793-efficient-learning-using-forward-backward-splitting.pdf)
Here, we use the adagrad learning rate as specified here: 
(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
110

111 112 113 114 115 116 117 118 119 120 121
)DOC");
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(proximal_adagrad, ops::ProximalAdagradOp,
                             ops::ProximalAdagradOpMaker);
REGISTER_OP_CPU_KERNEL(
    proximal_adagrad,
Q
QI JUN 已提交
122
    ops::ProximalAdagradOpKernel<paddle::platform::CPUDeviceContext, float>);
反馈
建议
客服 返回
顶部