void LRNLayer<Dtype>::CrossChannelForward_cpu(
const vector<Blob<Dtype>*>& bottom, const 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();
Dtype alpha_over_size = alpha_ / size_;
caffe_sqr(num_ * channels_ * height_ * width_, bottom_data, top_data);
caffe_set(num_ * channels_ * height_ * width_, Dtype(k_), scale_data);
#ifdef _OPENMP
#pragma omp parallel for collapse(2)
#endif
for (int n = 0; n < num_; n++) {
for (int c = 0; c < channels_; c++) {
for (int i = c - pre_pad_; i <= c + pre_pad_; i++) {
if ((i >= 0) && (i < channels_)) {
caffe_axpy<Dtype>(height_ * width_, alpha_over_size,
top_data + scale_.offset(n, i),
scale_data + scale_.offset(n, c));
}
}
}
}
caffe_powx<Dtype>(scale_.count(), scale_data, -beta_, top_data);
caffe_mul<Dtype>(scale_.count(), top_data, bottom_data, top_data);
}
void LRNLayer<Dtype>::CrossChannelForward_cpu(
const vector<Blob<Dtype>*>& bottom, const 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();
// start with the constant value
for (int i = 0; i < scale_.count(); ++i) {
scale_data[i] = k_; // 初始化为1.0
}
// 调用Blob的构造函数，定义一个补零后的Blob对象，其channels维度比bottom[0]增加了size_-1
Blob<Dtype> padded_square(1, channels_ + size_ - 1, height_, width_);
Dtype* padded_square_data = padded_square.mutable_cpu_data();
caffe_set(padded_square.count(), Dtype(0), padded_square_data); // 清零
Dtype alpha_over_size = alpha_ / size_; // 公式中的alpha/n
// go through the images
for (int n = 0; n < num_; ++n) { // 循环遍历每一个样本
// compute the padded square y[i] = a[i]^2
//计算当前样本的bottom[0]的每个feature map中每个元素的平方，但前后pre_pad_个FM保持0
caffe_sqr(channels_ * height_ * width_,
bottom_data + bottom[0]->offset(n),
padded_square_data + padded_square.offset(0, pre_pad_));
// Create the first channel scale y[i] = alpha*x[i] + y[i]
// 计算第一批size_个相邻的feature map(即padded_square中相邻的第0个到第size_个FM)的对应每个元素乘以因子后的累加和
// 并加上scale_中对应的值，最后得到的值存储中scale_中
// 即完成公式中k_+sum_under_i(x_i^2)的计算
for (int c = 0; c < size_; ++c) {
caffe_axpy<Dtype>(height_ * width_, alpha_over_size,
padded_square_data + padded_square.offset(0, c),
scale_data + scale_.offset(n, 0));
}
// 这里使用了类似FIFO的形式计算其余scale_参数，
// 每次先复制上一次计算的padded_square的size_个相邻的feature map的结果，假设此时这size_个相邻的feature map的最后一个的标号是Index
// 然后再加上第index+1个feature map的相应处理后的值
// 最后减去上一次求和计算中的第一个feature map的值,即第index-size_+1个feature map的值
// 以上三步操作，完成计算的是LRN层按顺序输出的一个feature map，所以需要for循环channels_1次
for (int c = 1; c < channels_; ++c) {
// copy previous scale
caffe_copy<Dtype>(height_ * width_,
scale_data + scale_.offset(n, c - 1),
scale_data + scale_.offset(n, c));
// add head
caffe_axpy<Dtype>(height_ * width_, alpha_over_size,
padded_square_data + padded_square.offset(0, c + size_ - 1),
scale_data + scale_.offset(n, c));
// subtract tail
caffe_axpy<Dtype>(height_ * width_, -alpha_over_size,
padded_square_data + padded_square.offset(0, c - 1),
scale_data + scale_.offset(n, c));
}
}
// In the end, compute output 完成最后的开根和求墒的操作, 并存入top[0]中
caffe_powx<Dtype>(scale_.count(), scale_data, -beta_, top_data);
caffe_mul<Dtype>(scale_.count(), top_data, bottom_data, top_data); // 矩阵乘法
}
void LRNLayer<Dtype>::CrossChannelForward_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();
// start with the constant value
for (int i = 0; i < scale_.count(); ++i) {
scale_data[i] = 1.;
}
Blob<Dtype> padded_square(1, channels_ + size_ - 1, height_, width_);
Dtype* padded_square_data = padded_square.mutable_cpu_data();
caffe_set(padded_square.count(), Dtype(0), padded_square_data);
Dtype alpha_over_size = alpha_ / size_;
// go through the images
for (int n = 0; n < num_; ++n) {
// compute the padded square
caffe_sqr(channels_ * height_ * width_,
bottom_data + bottom[0]->offset(n),
padded_square_data + padded_square.offset(0, pre_pad_));
// Create the first channel scale
for (int c = 0; c < size_; ++c) {
caffe_axpy<Dtype>(height_ * width_, alpha_over_size,
padded_square_data + padded_square.offset(0, c),
scale_data + scale_.offset(n, 0));
}
for (int c = 1; c < channels_; ++c) {
// copy previous scale
caffe_copy<Dtype>(height_ * width_,
scale_data + scale_.offset(n, c - 1),
scale_data + scale_.offset(n, c));
// add head
caffe_axpy<Dtype>(height_ * width_, alpha_over_size,
padded_square_data + padded_square.offset(0, c + size_ - 1),
scale_data + scale_.offset(n, c));
// subtract tail
caffe_axpy<Dtype>(height_ * width_, -alpha_over_size,
padded_square_data + padded_square.offset(0, c - 1),
scale_data + scale_.offset(n, c));
}
// for (int i = 0; i < scale_.count(); ++i) {
// if (scale_data[i] < 0 )
// LOG(FATAL) << "found negative norm term " << scale_data[i] << " @ " << i;
// }
}
// In the end, compute output
caffe_powx<Dtype>(scale_.count(), scale_data, -beta_, top_data);
caffe_mul<Dtype>(scale_.count(), top_data, bottom_data, top_data);
}
