参考文献依然是放前面：https://blog.csdn.net/caicaiatnbu/category_9096319.html

darknet版本： https://github.com/AlexeyAB/darknet，与原始的版本还是有一点区别的。

因为第一次读源码，我就直接按照参考文献的顺序来了，到时候再查漏补缺，加油！

今天看的是：cost_layer，主要完成损失函数的前向计算以及损失函数的求导，损失函数的误差反向传播。

直接放代码注解：比较简单

#include "cost_layer.h"
#include "utils.h"
#include "dark_cuda.h"
#include "blas.h"
#include <math.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>


/**
 * 根据输入的损失函数名称,返回定义的枚举类型的损失函数类别
 * @param s 损失函数的名称
 * @return 损失函数类别: 枚举类型
 * 说明: 如果不匹配,默认采用 SSE
 */
COST_TYPE get_cost_type(char *s)
{
    if (strcmp(s, "sse")==0) return SSE;
    if (strcmp(s, "masked")==0) return MASKED;
    if (strcmp(s, "smooth")==0) return SMOOTH;
    fprintf(stderr, "Couldn't find cost type %s, going with SSE\n", s);
    return SSE;
}

/**
 * 获得定义的枚举类型的损失函数字符串描述
 * @param a 损失函数类别: 枚举类型
 * @return 返回损失函数的字符串描述
 * 说明: 如果不匹配, 默认采用SSE
 */
char *get_cost_string(COST_TYPE a)
{
    switch(a){
        case SSE:
            return "sse";
        case MASKED:
            return "masked";
        case SMOOTH:
            return "smooth";
		default:
			return "sse";
    }
}

/**
 * 构建损失函数层
 * @param batch 该层输入中一个batch所含有图片的张数,等于net.batch
 * @param inputs 损失函数层每张输入图片的元素个数
 * @param cost_type 损失函数类型
 * @param scale
 * @return 损失函数层 l
 */

/*parser.c文件中有一个关于cost layer的使用
可以看到scale和type都是外面传进来的参数
cost_layer parse_cost(list *options, size_params params)
{
    char *type_s = option_find_str(options, "type", "sse");
    COST_TYPE type = get_cost_type(type_s); //h获得cost layer对应的类型
    float scale = option_find_float_quiet(options, "scale",1);
    cost_layer layer = make_cost_layer(params.batch, params.inputs, type, scale);
    layer.ratio =  option_find_float_quiet(options, "ratio",0);
    return layer;
}
*/
cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type, float scale)
{
    fprintf(stderr, "cost                                           %4d\n",  inputs);
    cost_layer l = { (LAYER_TYPE)0 };//初定义一个layer
    l.type = COST;//给定layer的类型，为损失层

    l.scale = scale;
    l.batch = batch;//batch数
    l.inputs = inputs;//该层的输入参数大小
    l.outputs = inputs;//该层的输出参数大小
    l.cost_type = cost_type;//损失函数类型
    l.delta = (float*)xcalloc(inputs * batch, sizeof(float));//构建layer的误差项空间
    l.output = (float*)xcalloc(inputs * batch, sizeof(float));//构建layer输出的特征空间
    l.cost = (float*)xcalloc(1, sizeof(float));//损失函数的值

    //前向和后向操作
    l.forward = forward_cost_layer;
    l.backward = backward_cost_layer;
    #ifdef GPU
    l.forward_gpu = forward_cost_layer_gpu;
    l.backward_gpu = backward_cost_layer_gpu;

    l.delta_gpu = cuda_make_array(l.delta, inputs*batch);
    l.output_gpu = cuda_make_array(l.output, inputs*batch);
    #endif
    return l;
}

//重构cost layer中要用到的存储空间
void resize_cost_layer(cost_layer *l, int inputs)
{
    l->inputs = inputs;
    l->outputs = inputs;
    l->delta = (float*)xrealloc(l->delta, inputs * l->batch * sizeof(float));//util.c定义的xrealloc，重新分配空间
    l->output = (float*)xrealloc(l->output, inputs * l->batch * sizeof(float));
#ifdef GPU
    cuda_free(l->delta_gpu);
    cuda_free(l->output_gpu);
    l->delta_gpu = cuda_make_array(l->delta, inputs*l->batch);
    l->output_gpu = cuda_make_array(l->output, inputs*l->batch);
#endif
}

/**
 * 损失函数层的前向传播函数
 * @param l 当前损失函数层
 * @param net 整个网络
 */
void forward_cost_layer(cost_layer l, network_state state)
{
    if (!state.truth) return;
    if(l.cost_type == MASKED){// MASKED只发现在darknet9000.cfg中使用
        int i;
        for(i = 0; i < l.batch*l.inputs; ++i){
            if(state.truth[i] == SECRET_NUM) state.input[i] = SECRET_NUM;
        }
    }
    if(l.cost_type == SMOOTH){// 如果损失函数是 smooth l1，调用smooth_l1_cpu
        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
        /*
        在blas.c中：
        明确:smooth(x) = 0.5* x^2 / beta    |x| < 1,
                       = |x| - 0.5 * beta    otherwise. 
        梯度:d(smooth_l1_loss(x))/d(x) = x           |x| < 1
                                       = +/- 1       otherwise.           
        void smooth_l1_cpu(int n, float *pred, float *truth, float *delta, float *error)
        {   
            int i;
            for(i = 0; i < n; ++i){
                float diff = truth[i] - pred[i];
                float abs_val = fabs(diff);//|x|
                if(abs_val < 1) { //做判断
                    error[i] = diff * diff;
                    delta[i] = diff;
                }
                else {
                    error[i] = 2*abs_val - 1;
                    delta[i] = (diff > 0) ? 1 : -1;
                }
            }
        }
        */

    } else {//其它，调用l2_cpu
        l2_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
        /*
        在blas.c中：
        SSE, 即L2, 误差平方和,可以发现这里并没有乘以1/2.
        明确:L2(x) = x**2
        梯度:dL2(x) /dx = 2x

        void l2_cpu(int n, float *pred, float *truth, float *delta, float *error)
        {
            int i;
            for(i = 0; i < n; ++i){
                float diff = truth[i] - pred[i];//x
                error[i] = diff * diff;
                delta[i] = diff;
            }
        }
        */
    }
    // 求loss总和，求cost layer的输出总和
    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
    /*
    在util.c中
    float sum_array(float *a, int n)
    {
        int i;
        float sum = 0;
        for(i = 0; i < n; ++i) sum += a[i];
        return sum;
    }
    */

}

/**
 * 损失函数层的反向传播函数
 * @param l 当前损失函数层
 * @param net 整个网络
 */
void backward_cost_layer(const cost_layer l, network_state state)
{
    // net.data += l.scale * l.delta
    //这里我没太懂，等我往后看
    axpy_cpu(l.batch*l.inputs, l.scale, l.delta, 1, state.delta, 1);
    /*
    在blas.c中：Y += alpha * X
    void axpy_cpu(int N, float ALPHA, float *X, int INCX, float *Y, int INCY)
    {
        int i;
        for(i = 0; i < N; ++i) Y[i*INCY] += ALPHA*X[i*INCX];
    }
    */
}

#ifdef GPU

void pull_cost_layer(cost_layer l)
{
    cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}

void push_cost_layer(cost_layer l)
{
    cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}

int float_abs_compare (const void * a, const void * b)
{
    float fa = *(const float*) a;
    if(fa < 0) fa = -fa;
    float fb = *(const float*) b;
    if(fb < 0) fb = -fb;
    return (fa > fb) - (fa < fb);
}

void forward_cost_layer_gpu(cost_layer l, network_state state)
{
    if (!state.truth) return;
    if (l.cost_type == MASKED) {
        mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
    }

    if(l.cost_type == SMOOTH){
        smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
    } else {
        l2_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
    }

    if(l.ratio){
        cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
        qsort(l.delta, l.batch*l.inputs, sizeof(float), float_abs_compare);
        int n = (1-l.ratio) * l.batch*l.inputs;
        float thresh = l.delta[n];
        thresh = 0;
        printf("%f\n", thresh);
        supp_ongpu(l.batch*l.inputs, thresh, l.delta_gpu, 1);
    }

    cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
}

void backward_cost_layer_gpu(const cost_layer l, network_state state)
{
    axpy_ongpu(l.batch*l.inputs, l.scale, l.delta_gpu, 1, state.delta, 1);
}
#endif