有的时候我们需要在Caffe中添加新的Layer,现在在做的项目中,需要有一个L2 Normalization Layer,Caffe中居然没有,所以要自己添加。
所以最重要的是如何实现forward_cpu(forward_gpu), backward_cpu(backward_gpu).
1. L2 Normalization Forward Pass(向前传导)
1.1 Formula Deduction(公式推导)
1.2 Implementation(实现)
template <typename Dtype>
void NormalizationLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
const vector<Blob<Dtype>*>& top) {
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* top_data = top[0]->mutable_gpu_data();
Dtype* squared_data = squared_.mutable_gpu_data();
Dtype normsqr;
int n = bottom[0]->num();
int d = bottom[0]->count() / n;
caffe_gpu_powx(n*d, bottom_data, Dtype(2), squared_data);
for (int i=0; i<n; ++i) {
caffe_gpu_asum<Dtype>(d, squared_data+i*d, &normsqr);
caffe_gpu_scale<Dtype>(d, pow(normsqr, -0.5), bottom_data+i*d, top_data+i*d);
}
}
2. L2 Normalization Backward Propagation
2.1 Formula Deduction(公式推导)
First is the gradient regardless of upper layer:
2.2 Implementation(实现)
template <typename Dtype>
void NormalizationLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
const vector<bool>& propagate_down, const vector<Blob<Dtype>*>& bottom) {
const Dtype* top_diff = top[0]->gpu_diff();
const Dtype* top_data = top[0]->gpu_data();
const Dtype* bottom_data = bottom[0]->gpu_data();
Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
int n = top[0]->num();
int d = top[0]->count() / n;
Dtype a;
for (int i=0; i<n; ++i) {
caffe_gpu_dot(d, top_data+i*d, top_diff+i*d, &a);
caffe_gpu_scale(d, a, top_data+i*d, bottom_diff+i*d);
caffe_gpu_sub(d, top_diff+i*d, bottom_diff+i*d, bottom_diff+i*d);
caffe_gpu_dot(d, bottom_data+i*d, bottom_data+i*d, &a);
caffe_gpu_scale(d, Dtype(pow(a, -0.5)), bottom_diff+i*d, bottom_diff+i*d);
}
}
全部修改可看一下 链接:
https://github.com/freesouls/caffe/tree/master/src/caffe/layers