momentum_op.h 2.0 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
S
sidgoyal78 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15

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. */

#pragma once
Y
Yi Wang 已提交
16 17
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
S
sidgoyal78 已提交
18 19 20 21

namespace paddle {
namespace operators {

22
template <typename T>
S
sidgoyal78 已提交
23 24 25
class MomentumOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
S
sidgoyal78 已提交
26 27 28 29 30 31
    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
    auto velocity_out = ctx.Output<framework::Tensor>("VelocityOut");
    auto param = ctx.Input<framework::Tensor>("Param");
    auto velocity = ctx.Input<framework::Tensor>("Velocity");
    auto grad = ctx.Input<framework::Tensor>("Grad");
    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
S
sidgoyal78 已提交
32 33 34 35

    param_out->mutable_data<T>(ctx.GetPlace());
    velocity_out->mutable_data<T>(ctx.GetPlace());

36
    T mu = static_cast<T>(ctx.Attr<float>("mu"));
37
    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
S
sidgoyal78 已提交
38

S
sidgoyal78 已提交
39 40 41 42 43 44
    auto p_out = framework::EigenVector<T>::Flatten(*param_out);
    auto v_out = framework::EigenVector<T>::Flatten(*velocity_out);

    auto p = framework::EigenVector<T>::Flatten(*param);
    auto v = framework::EigenVector<T>::Flatten(*velocity);
    auto g = framework::EigenVector<T>::Flatten(*grad);
45
    auto* lr = learning_rate->data<T>();
S
sidgoyal78 已提交
46

47
    v_out = v * mu + g;
K
kavyasrinet 已提交
48
    if (use_nesterov) {
49
      p_out = p - (g - v_out * mu) * lr[0];
K
kavyasrinet 已提交
50
    } else {
51
      p_out = p - lr[0] * v_out;
K
kavyasrinet 已提交
52
    }
S
sidgoyal78 已提交
53 54 55 56 57
  }
};

}  // namespace operators
}  // namespace paddle