darknet 修改focal loss

https://github.com/AlexeyAB/darknet/blob/17520296c730c7d7e2683452b11bf50fc8959688/src/yolo_layer.c#L115-L119

由于不知道为什么focal loss要用if，因此直接将if置为1

在yolo_layer.c中

delta_yolo_class 计算类别误差
void delta_yolo_class(float *output, float *delta, int index, int class, int classes, int stride, float *avg_cat)
{
    int n;
    if (delta[index]){ //应该不会进入这个判断，因为 delta[index] 初值为0
        delta[index + stride*class] = 1 - output[index + stride*class];
        if(avg_cat) *avg_cat += output[index + stride*class];
        return;
    }
    for(n = 0; n < classes; ++n){ //对所有类别，如果预测正确，则误差为 1-predict，否则为 0-predict
        delta[index + stride*n] = ((n == class)?1 : 0) - output[index + stride*n];
        if(n == class && avg_cat) *avg_cat += output[index + stride*n];
    }
}
修改后如下：

void delta_yolo_class(float *output, float *delta, int index, int class, int classes, int stride, float *avg_cat)
{
    int n;
    if (delta[index]){
        delta[index + stride*class] = 1 - output[index + stride*class];
        if(avg_cat) *avg_cat += output[index + stride*class];
        return;
    }

    // Focal loss
    if (1) {
        // Focal Loss
        float alpha = 0.5;    // 0.25 or 0.5
        //float gamma = 2;    // hardcoded in many places of the grad-formula    

        int ti = index + stride*class;
        float pt = output[ti] + 0.000000000000001F;
        // http://fooplot.com/#W3sidHlwZSI6MCwiZXEiOiItKDEteCkqKDIqeCpsb2coeCkreC0xKSIsImNvbG9yIjoiIzAwMDAwMCJ9LHsidHlwZSI6MTAwMH1d
        float grad = -(1 - pt) * (2 * pt*logf(pt) + pt - 1);    // http://blog.csdn.net/linmingan/article/details/77885832    
        //float grad = (1 - pt) * (2 * pt*logf(pt) + pt - 1);    // https://github.com/unsky/focal-loss

        for (n = 0; n < classes; ++n) {
            delta[index + stride*n] = (((n == class) ? 1 : 0) - output[index + stride*n]);

            delta[index + stride*n] *= alpha*grad;

            if (n == class) *avg_cat += output[index + stride*n];
        }
    }

    else{
        for(n = 0; n < classes; ++n){
        delta[index + stride*n] = ((n == class)?1 : 0) - output[index + stride*n];
        if(n == class && avg_cat) *avg_cat += output[index + stride*n];
        }
    }
}