adagrad_op.cc 5.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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/adagrad_op.h"

Q
QI JUN 已提交
17 18 19 20 21
#include <cmath>

#include "paddle/operators/math/math_function.h"
#include "paddle/operators/math/selected_rows_functor.h"

22 23 24 25 26 27 28
namespace paddle {
namespace operators {

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

Q
QI JUN 已提交
29
  void InferShape(framework::InferShapeContext* ctx) const override {
K
Kexin Zhao 已提交
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
    PADDLE_ENFORCE(ctx->HasInput("Param"),
                   "Input(Param) of AdagradOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Grad"),
                   "Input(Grad) of AdagradOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Moment"),
                   "Input(Moment) of AdagradOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("LearningRate"),
                   "Input(LearningRate) of AdagradOp should not be null.");

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

    auto lr_dims = ctx->GetInputDim("LearningRate");
45
    PADDLE_ENFORCE_EQ(framework::product(lr_dims), 1,
K
Kexin Zhao 已提交
46 47
                      "LearningRate should have one element");
    auto param_dims = ctx->GetInputDim("Param");
48
    PADDLE_ENFORCE_EQ(
K
Kexin Zhao 已提交
49 50
        param_dims, ctx->GetInputDim("Grad"),
        "Param and Grad input of AdagradOp should have the same dimension.");
51
    PADDLE_ENFORCE_EQ(
K
Kexin Zhao 已提交
52 53
        param_dims, ctx->GetInputDim("Moment"),
        "Param and Moment input of AdagradOp should have the same dimension.");
54

K
Kexin Zhao 已提交
55 56
    ctx->SetOutputDim("ParamOut", param_dims);
    ctx->SetOutputDim("MomentOut", param_dims);
57 58 59 60 61
  }
};

class AdagradOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
62
  AdagradOpMaker(OpProto* proto, OpAttrChecker* op_checker)
63
      : OpProtoAndCheckerMaker(proto, op_checker) {
K
Kexin Zhao 已提交
64 65 66 67 68 69 70 71 72 73 74 75
    AddInput("Param", "(Tensor) Input parameter");
    AddInput("Grad", "(Tensor) Input gradient");
    AddInput("Moment", "(Tensor) Second moment");
    AddInput("LearningRate", "(Tensor) Learning rate");

    AddOutput("ParamOut", "(Tensor) Output parameter");
    AddOutput("MomentOut", "(Tensor) Output second moment");

    AddAttr<float>("epsilon",
                   "(float, default 1.0e-6) "
                   "Constant for numerical stability")
        .SetDefault(1.0e-6f);
76 77 78 79
    AddComment(R"DOC(

Adaptive Gradient Algorithm (Adagrad).

80 81
The update is done as follows:

82 83
$$moment\_out = moment + grad * grad \\
param\_out = param - \frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
84
$$
85 86

The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
87 88 89
does not have the epsilon attribute. It is added here in our implementation
as also proposed here: http://cs231n.github.io/neural-networks-3/#ada
for numerical stability to avoid the division by zero error.
90 91 92 93

)DOC");
  }
};
Q
QI JUN 已提交
94 95 96 97 98 99 100 101

namespace {
size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
  return std::find(rows.begin(), rows.end(), value) - rows.begin();
}
}  // namespace

template <typename T>
Q
QI JUN 已提交
102 103
struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
  void operator()(const platform::CPUDeviceContext& context,
Q
QI JUN 已提交
104 105 106 107 108
                  const framework::SelectedRows& grad,
                  const framework::Tensor& learning_rate, T epsilon,
                  framework::Tensor* moment, framework::Tensor* param) {
    // 1. g_m.rows = set(g.rows)
    auto grad_width = grad.value().dims()[1];
T
wip  
typhoonzero 已提交
109 110 111 112
    math::scatter::MergeAdd<platform::CPUDeviceContext, T> merge_func;
    auto grad_merge = merge_func(context, grad);
    auto& merge_rows = grad_merge.rows();
    auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
Q
QI JUN 已提交
113 114

    // 2. m += g_m * g_m
T
wip  
typhoonzero 已提交
115 116
    math::scatter::Mul<platform::CPUDeviceContext, T> sqare_func;
    auto grad_square = sqare_func(context, grad_merge, grad_merge);
Q
QI JUN 已提交
117

Q
QI JUN 已提交
118
    math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
T
wip  
typhoonzero 已提交
119
    functor(context, grad_square, moment);
Q
QI JUN 已提交
120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135

    // 3. update parameter
    auto* lr = learning_rate.data<T>();
    auto* param_data = param->data<T>();
    auto* moment_data = moment->data<T>();

    for (size_t i = 0; i < merge_rows.size(); i++) {
      for (int64_t j = 0; j < grad_width; j++) {
        param_data[merge_rows[i] * grad_width + j] -=
            lr[0] * grad_merge_data[i * grad_width + j] /
            (std::sqrt(moment_data[merge_rows[i] * grad_width + j]) + epsilon);
      }
    }
  }
};

Q
QI JUN 已提交
136 137
template struct SparseAdagradFunctor<platform::CPUDeviceContext, float>;
template struct SparseAdagradFunctor<platform::CPUDeviceContext, double>;
138 139 140 141 142
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker);
Q
QI JUN 已提交
143
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
144 145
    adagrad, ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, double>);