momentum_op.h 2.0 KB
Newer Older
S
sidgoyal78 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
/* 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. */

#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"

namespace paddle {
namespace operators {

template <typename Place, typename T>
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 36 37

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

    float mu = ctx.Attr<float>("mu");

S
sidgoyal78 已提交
38 39 40 41 42 43 44 45
    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);
    auto lr = framework::EigenVector<T>::Flatten(*learning_rate);

S
sidgoyal78 已提交
46 47
    auto place = ctx.GetEigenDevice<Place>();

S
sidgoyal78 已提交
48 49 50
    Eigen::DSizes<int, 1> grad_dsize(grad->numel());
    v_out.device(place) = v * mu + g;
    p_out.device(place) = p - lr.broadcast(grad_dsize) * v_out;
S
sidgoyal78 已提交
51 52 53 54 55
  }
};

}  // namespace operators
}  // namespace paddle