#include <algorithm>

#include <vector>

#include "caffe/layer.hpp"

#include "caffe/util/math_functions.hpp"

#include "caffe/vision_layers.hpp"

namespace caffe {

template <typename Dtype>

void SoftmaxLayer<Dtype>::Reshape(const vector<Blob<Dtype>*>& bottom,

vector<Blob<Dtype>*>* top) {

(*top)[0]->Reshape(bottom[0]->num(), bottom[0]->channels(),

bottom[0]->height(), bottom[0]->width());

sum_multiplier_.Reshape(1, bottom[0]->channels(), 1, 1);

Dtype* multiplier_data = sum_multiplier_.mutable_cpu_data();

for (int i = 0; i < sum_multiplier_.count(); ++i) {

multiplier_data[i] = 1.;

}

scale_.Reshape(bottom[0]->num(), 1, bottom[0]->height(), bottom[0]->width());

}

template <typename Dtype>

void SoftmaxLayer<Dtype>::Forward_cpu(const vector<Blob<Dtype>*>& bottom,

vector<Blob<Dtype>*>* top) {

const Dtype* bottom_data = bottom[0]->cpu_data();

Dtype* top_data = (*top)[0]->mutable_cpu_data();

Dtype* scale_data = scale_.mutable_cpu_data();

int num = bottom[0]->num();

int channels = bottom[0]->channels();

int dim = bottom[0]->count() / bottom[0]->num();

int spatial_dim = bottom[0]->height() * bottom[0]->width();

caffe_copy(bottom[0]->count(), bottom_data, top_data);

// We need to subtract the max to avoid numerical issues, compute the exp,

// and then normalize.

for (int i = 0; i < num; ++i) {

// initialize scale_data to the first plane

caffe_copy(spatial_dim, bottom_data + i * dim, scale_data);

for (int j = 0; j < channels; j++) {

for (int k = 0; k < spatial_dim; k++) {

scale_data[k] = std::max(scale_data[k],

bottom_data[i * dim + j * spatial_dim + k]);

}

}

// subtraction

caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels, spatial_dim,

1, -1., sum_multiplier_.cpu_data(), scale_data, 1., top_data + i * dim);

// exponentiation

caffe_exp<Dtype>(dim, top_data + i * dim, top_data + i * dim);

// sum after exp

caffe_cpu_gemv<Dtype>(CblasTrans, channels, spatial_dim, 1.,

top_data + i * dim, sum_multiplier_.cpu_data(), 0., scale_data);

// division

for (int j = 0; j < channels; j++) {

caffe_div(spatial_dim, top_data + (*top)[0]->offset(i, j), scale_data,

top_data + (*top)[0]->offset(i, j));

}

}

}

template <typename Dtype>

void SoftmaxLayer<Dtype>::Backward_cpu(const vector<Blob<Dtype>*>& top,

const vector<bool>& propagate_down,

vector<Blob<Dtype>*>* bottom) {

const Dtype* top_diff = top[0]->cpu_diff();

const Dtype* top_data = top[0]->cpu_data();

Dtype* bottom_diff = (*bottom)[0]->mutable_cpu_diff();

Dtype* scale_data = scale_.mutable_cpu_data();

int num = top[0]->num();

int channels = top[0]->channels();

int dim = top[0]->count() / top[0]->num();

int spatial_dim = top[0]->height() * top[0]->width();

caffe_copy(top[0]->count(), top_diff, bottom_diff);

for (int i = 0; i < num; ++i) {

// compute dot(top_diff, top_data) and subtract them from the bottom diff

for (int k = 0; k < spatial_dim; ++k) {

scale_data[k] = caffe_cpu_strided_dot<Dtype>(channels,

bottom_diff + i * dim + k, spatial_dim,

top_data + i * dim + k, spatial_dim);

}

// subtraction

caffe_cpu_gemm<Dtype>(CblasNoTrans, CblasNoTrans, channels, spatial_dim, 1,

-1., sum_multiplier_.cpu_data(), scale_data, 1., bottom_diff + i * dim);

}

// elementwise multiplication

caffe_mul(top[0]->count(), bottom_diff, top_data, bottom_diff);

}

#ifdef CPU_ONLY

STUB_GPU(SoftmaxLayer);

#endif

INSTANTIATE_CLASS(SoftmaxLayer);

} // namespace caffe