caffe源码解读(2)－center_loss_layer.cpp

• center_loss公式定义

• center_loss_layer.cpp源码解读

• center loss,softmax loss在mnist数据集上的对比实验

定义

“Center Loss: simultaneously learning a center for deep features of each class and penalizing the distances between the deep features and their corresponding class centers. 参考论文：　A Discriminative Feature Learning Approach for　Deep Face　Recognition。　关于对center loss的理解，可参考知乎回答链接。

公式

(1) Forward Computation

$$L_{c}=\frac{1}{2N}\sum_{i=1}^{N}|x_{i}-c|_{2}^{2}$$
(2) Backward Computation
$$\frac{\partial L_{c}}{\partial x_{i}}=x_{i}-c$$
(3) Update Equation
$$\Delta c=-\frac{\alpha }{N}\sum_{i=1}^{N}(x_{i}-c)$$

代码

(１) LayerSetUp

namespace caffe{
    template<typename Dtype>
    void CenterLossLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
        const vector<Blob<Dtype>*>& top){
        LossLayer<Dtype>::LayerSetUp(bottom, top);
        CHECK_EQ(bottom[0]->num(), bottom[1]->num()); 
         //2个bottom,feature = bottom[0]->cpu_data();label = bottom[1]->cpu_data()
         //1个top,top[0]->mutable_cpu_data()[0] = loss;
        int channels = bottom[0]->channels();
        int num = bottom[0]->num();
         //获取center loss层的参数
         //loss weight 即参数lambda,用来调节center loss占比
        alpha = this->layer_param_.center_loss_param().alpha();
        lossWeight = this->layer_param_.center_loss_param().loss_weight();
        clusterNum = this->layer_param_.center_loss_param().cluster_num();

        center_info_.Reshape(clusterNum, channels, 1, 1);
        center_loss_.Reshape(num, channels, 1, 1);
        center_update_count_.resize(clusterNum);
        //caffe_set对center_info_.mutable_cpu_data()初始化
        caffe_set(clusterNum * channels, Dtype(0.0), center_info_.mutable_cpu_data());
    }

(2) Forward前向传播

template<typename Dtype>
    void CenterLossLayer<Dtype>::Forward_cpu(
        const vector<Blob<Dtype>*> &bottom,
        const vector<Blob<Dtype>*> &top){
        //２个bottom输入
        const Dtype *feature = bottom[0]->cpu_data();
        const Dtype *label = bottom[1]->cpu_data();
        int num = bottom[0]->num();
        int channels = bottom[0]->channels();
        //初始化loss
        Dtype loss = 0;
        caffe_set(clusterNum * channels, Dtype(0.0), center_info_.mutable_cpu_diff());
        for(int i = 0; i < clusterNum; ++i){
            center_update_count_[i] = 1;
        }
        for(int i = 0; i < num; ++i){
            int targetLabel = label[i];
            //caffe_sub做减法：center_loss.mutable_cpu_data=feature-center_info_.cpu_data()
            //即公式中xi-c
            caffe_sub(channels, feature + i * channels,
            center_info_.cpu_data() + targetLabel * channels,
            center_loss_.mutable_cpu_data() + i * channels);
            // store the update loss and number
            caffe_add(channels, center_loss_.cpu_data() + i * channels,
            center_info_.cpu_diff() + targetLabel * channels,
            center_info_.mutable_cpu_diff() + targetLabel * channels);
    　　　　　center_update_count_[targetLabel]++;
　　　　　　 //此处即按公式(1)计算center loss
　　　　　　 //并将loss作为top输出
            loss += caffe_cpu_dot(channels, center_loss_.cpu_data() + i * channels,
            center_loss_.cpu_data() + i * channels) * lossWeight / Dtype(2.0) / static_cast<Dtype>(num);
        }
        top[0]->mutable_cpu_data()[0] = loss;
        // update center loss.按公式(3)更新类中心：c
        for(int i = 0; i < clusterNum; ++i){
            Dtype scale = -alpha * lossWeight / Dtype(center_update_count_[i]);
            caffe_scal(channels, scale, center_info_.mutable_cpu_diff() + i * channels);
        }
        center_info_.Update();
    }

(3) Backward反向传播

    template<typename Dtype>
    void CenterLossLayer<Dtype>::Backward_cpu(
        const vector<Blob<Dtype>*> &top,
        const vector<bool> &propagate_down,
        const vector<Blob<Dtype>*> &bottom){
        int num = bottom[0]->num();
        int channels = bottom[0]->channels();
        //center_loss_.mutable_cpu_data()=feature-center_info_.cpu_data()
        //按公式（２）计算反向传播偏导
        caffe_scal(num * channels, lossWeight, center_loss_.mutable_cpu_data());
        Dtype *out = bottom[0]->mutable_cpu_diff();
        //center_loss_.cpu_data()拷贝到out中进行backward运算
        caffe_copy(num * channels, center_loss_.cpu_data(), out);

    }

实验

Github 上有开源的整个项目的代码[链接]（https://github.com/wangwen39/center-loss），新手可以用来练手。特征可视化可直接参考caffe主页：http://nbviewer.jupyter.org/github/BVLC/caffe/blob/master/examples/siamese/mnist_siamese.ipynb
mnist数据集共有10个类别的手写体数字0-9，通过对比实验可以看出，center loss能够很好的使类类之间的距离增大，同时使类内更加聚拢，从而达到更好的分类准确度。
① softmax loss

② center loss + softmax loss