maxout_op.h 2.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
W
wanghaox 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

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
17 18 19
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/maxouting.h"
W
wanghaox 已提交
20 21 22 23 24 25

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;

Q
QI JUN 已提交
26
template <typename DeviceContext, typename T>
W
wanghaox 已提交
27 28 29 30 31 32
class MaxOutKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    const Tensor* in_x = context.Input<Tensor>("X");
    Tensor* out = context.Output<Tensor>("Out");
    int groups = context.template Attr<int>("groups");
33
    int axis = context.template Attr<int>("axis");
34 35 36
    if (axis < 0) {
      axis += in_x->dims().size();
    }
W
wanghaox 已提交
37

Q
QI JUN 已提交
38 39
    math::MaxOutFunctor<DeviceContext, T> maxout_forward;
    maxout_forward(context.template device_context<DeviceContext>(), *in_x, out,
40
                   groups, axis);
W
wanghaox 已提交
41 42 43
  }
};

Q
QI JUN 已提交
44
template <typename DeviceContext, typename T>
W
wanghaox 已提交
45 46 47 48 49 50 51 52 53
class MaxOutGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    const Tensor* in_x = context.Input<Tensor>("X");
    const Tensor* out = context.Input<Tensor>("Out");
    const Tensor* out_grad =
        context.Input<Tensor>(framework::GradVarName("Out"));
    Tensor* in_x_grad = context.Output<Tensor>(framework::GradVarName("X"));
    int groups = context.template Attr<int>("groups");
54
    int axis = context.template Attr<int>("axis");
55 56 57 58
    if (axis < 0) {
      axis += in_x->dims().size();
    }

Q
QI JUN 已提交
59 60
    auto& device_ctx = context.template device_context<DeviceContext>();
    math::SetConstant<DeviceContext, T> zero;
W
wanghaox 已提交
61 62
    if (in_x_grad) {
      in_x_grad->mutable_data<T>(context.GetPlace());
W
wanghaox 已提交
63
      zero(device_ctx, in_x_grad, static_cast<T>(0.0));
Q
QI JUN 已提交
64
      math::MaxOutGradFunctor<DeviceContext, T> maxout_backward;
65 66
      maxout_backward(device_ctx, *in_x, in_x_grad, *out, *out_grad, groups,
                      axis);
W
wanghaox 已提交
67 68 69 70 71 72
    }
  }
};

}  // namespace operators
}  // namespace paddle