softmax_op.h 2.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/* 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 <glog/logging.h>
#include <paddle/framework/operator.h>

namespace paddle {
namespace operators {

Q
qijun 已提交
23
template <typename Place, typename T>
24 25
class SoftmaxKernel : public framework::OpKernel {
public:
Q
qijun 已提交
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
  void Compute(const framework::KernelContext& context) const override {
    auto input = context.Input(0)->Get<framework::Tensor>();
    auto* output = context.Output(0)->GetMutable<framework::Tensor>();

    auto logits = input.matrix<T>();
    auto softmax = output->matrix<T>();

    const int kBatchDim = 0;
    const int kClassDim = 1;

    const int batch_size = logits.dimension(kBatchDim);
    const int num_classes = logits.dimension(kClassDim);

    Eigen::DSizes<int, 1> along_class(kClassDim);
    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int, 2> one_by_class(1, num_classes);

    auto shifted_logits = (logits - logits.maximum(along_class)
                                        .eval()
                                        .reshape(batch_by_one)
                                        .broadcast(one_by_class));

    softmax.device(*(context.GetEigenDevice<Place>())) = shifted_logits.exp();

    softmax.device(*(context.GetEigenDevice<Place>())) =
        (softmax * softmax.sum(along_class)
                       .inverse()
                       .eval()
                       .reshape(batch_by_one)
                       .broadcast(one_by_class));
56 57 58 59
  }
};
}  // namespace operators
}  // namespace paddle