detection_layer层的实现

detection_layer对应Yolov1的实现,理解detection_layer的实现主要需要理解Yolov1中输出层的存储方式,以及论文中损失函数的构成。
论文深入理解,请参考YOLO v1深入理解。本文只对detection_layer的实现进行注释分析。

/*
** 构建detection层
** 输入: batch         一个batch中含有的图片张数(等于net.batch)
**       inputs        该层输入数据维度大小
**       n                一个grid cell预测bound box的数量
**       side            grid cell的尺寸大小
**       classes       目标类别数
**       coords       学习的位置参数个数
**       max_boxes       图像真实目标的最多个数
** 返回: detection_layer
*/
detection_layer make_detection_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
{
    detection_layer l = { (LAYER_TYPE)0 };
    l.type = DETECTION;
    l.n = n;
    l.batch = batch;
    l.inputs = inputs;
    l.classes = classes;
    l.coords = coords;
    l.rescore = rescore;
    l.side = side;
    l.w = side;
    l.h = side;
    assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
    l.cost = (float*)calloc(1, sizeof(float));
    l.outputs = l.inputs;
    //一个grid cell 只负责预测一个ground truth,因此此处不需要乘以l.n
    l.truths = l.side*l.side*(1+l.coords+l.classes);
    l.output = (float*)calloc(batch * l.outputs, sizeof(float));
    l.delta = (float*)calloc(batch * l.outputs, sizeof(float));

    l.forward = forward_detection_layer;
    l.backward = backward_detection_layer;
#ifdef GPU
    l.forward_gpu = forward_detection_layer_gpu;
    l.backward_gpu = backward_detection_layer_gpu;
    l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
    l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
#endif
    fprintf(stderr, "Detection Layer\n");
    srand(time(0));
    return l;
}

Yolov1中的损失函数如下:
图1 yolov1的损失函数
void forward_detection_layer(const detection_layer l, network_state state)
{
    //7*7=49个grid cell
    int locations = l.side*l.side;
    int i,j;
    memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));//copy state.input to l.output
    //if(l.reorg) reorg(l.output, l.w*l.h, size*l.n, l.batch, 1);
    int b;
    if (l.softmax){
        for(b = 0; b < l.batch; ++b){
            int index = b*l.inputs;
            for (i = 0; i < locations; ++i) {
                int offset = i*l.classes;
                //注意输出数据的存储方式(跟yolov2,v3的区别)
                softmax(l.output + index + offset, l.classes, 1,
                        l.output + index + offset, 1);
            }
        }
    }
    if(state.train){
        float avg_iou = 0;
        float avg_cat = 0;
        float avg_allcat = 0;
        float avg_obj = 0;
        float avg_anyobj = 0;
        int count = 0;
        *(l.cost) = 0;
        int size = l.inputs * l.batch;
        //初始化当前层的sensitivity
        memset(l.delta, 0, size * sizeof(float));
        for (b = 0; b < l.batch; ++b){
            int index = b*l.inputs;
            for (i = 0; i < locations; ++i) {
                int truth_index = (b*locations + i)*(1+l.coords+l.classes);//ground truth索引
                int is_obj = state.truth[truth_index];
                for (j = 0; j < l.n; ++j) {
                    //预测边框的索引(第i个grid cell的第j个bounding box)
                    int p_index = index + locations*l.classes + i*l.n + j;
                    //预测边框置信度损失(相当于对第四行对$\hat_{C_i}$求导)
                    l.delta[p_index] = l.noobject_scale*(0 - l.output[p_index]);
                    *(l.cost) += l.noobject_scale*pow(l.output[p_index], 2);//相当于第四行,当该box没有对象的时候,Ci为0
                    avg_anyobj += l.output[p_index];
                }

                int best_index = -1;
                float best_iou = 0;
                float best_rmse = 20;
                //该grid cell没有object,则只计算第四行的损失
                if (!is_obj){
                    continue;
                }
                //20个类别概率索引
                int class_index = index + i*l.classes;
                for(j = 0; j < l.classes; ++j) {
                    l.delta[class_index+j] = l.class_scale * (state.truth[truth_index+1+j] - l.output[class_index+j]);
                    //计算第5行
                    *(l.cost) += l.class_scale * pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2);
                    if(state.truth[truth_index + 1 + j]) avg_cat += l.output[class_index+j];
                    avg_allcat += l.output[class_index+j];
                }
                //注意state.truth的存储方式
                box truth = float_to_box(state.truth + truth_index + 1 + l.classes);
                truth.x /= l.side;
                truth.y /= l.side;

                for(j = 0; j < l.n; ++j){
                    //边框坐标位置索引
                    int box_index = index + locations*(l.classes + l.n) + (i*l.n + j) * l.coords;
                    box out = float_to_box(l.output + box_index);
                    out.x /= l.side;
                    out.y /= l.side;

                    if (l.sqrt){
                        out.w = out.w*out.w;
                        out.h = out.h*out.h;
                    }

                    float iou  = box_iou(out, truth);
                    //iou = 0;
                    float rmse = box_rmse(out, truth);
                    if(best_iou > 0 || iou > 0){
                        if(iou > best_iou){
                            best_iou = iou;
                            best_index = j;
                        }
                    }else{
                        if(rmse < best_rmse){
                            best_rmse = rmse;
                            best_index = j;
                        }
                    }
                }

                if(l.forced){
                    if(truth.w*truth.h < .1){
                        best_index = 1;
                    }else{
                        best_index = 0;
                    }
                }
                if(l.random && *(state.net.seen) < 64000){
                    best_index = rand()%l.n;
                }

                int box_index = index + locations*(l.classes + l.n) + (i*l.n + best_index) * l.coords;
                int tbox_index = truth_index + 1 + l.classes;

                box out = float_to_box(l.output + box_index);
                out.x /= l.side;
                out.y /= l.side;
                if (l.sqrt) {
                    out.w = out.w*out.w;
                    out.h = out.h*out.h;
                }
                float iou  = box_iou(out, truth);

                //printf("%d,", best_index);
                int p_index = index + locations*l.classes + i*l.n + best_index;//负责预测ground truth的bounding box
                *(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2);
                *(l.cost) += l.object_scale * pow(1-l.output[p_index], 2);
                avg_obj += l.output[p_index];
                l.delta[p_index] = l.object_scale * (1.-l.output[p_index]);

                if(l.rescore){
                    l.delta[p_index] = l.object_scale * (iou - l.output[p_index]);
                }
                //第一行求导
                l.delta[box_index+0] = l.coord_scale*(state.truth[tbox_index + 0] - l.output[box_index + 0]);
                l.delta[box_index+1] = l.coord_scale*(state.truth[tbox_index + 1] - l.output[box_index + 1]);
                 //第二行求导
                l.delta[box_index+2] = l.coord_scale*(state.truth[tbox_index + 2] - l.output[box_index + 2]);
                l.delta[box_index+3] = l.coord_scale*(state.truth[tbox_index + 3] - l.output[box_index + 3]);
                if(l.sqrt){
                    l.delta[box_index+2] = l.coord_scale*(sqrt(state.truth[tbox_index + 2]) - l.output[box_index + 2]);
                    l.delta[box_index+3] = l.coord_scale*(sqrt(state.truth[tbox_index + 3]) - l.output[box_index + 3]);
                }
                *(l.cost) += pow(1-iou, 2);
                avg_iou += iou;
                ++count;
            }
        }

        if(0){
            float* costs = (float*)calloc(l.batch * locations * l.n, sizeof(float));
            for (b = 0; b < l.batch; ++b) {
                int index = b*l.inputs;
                for (i = 0; i < locations; ++i) {
                    for (j = 0; j < l.n; ++j) {
                        int p_index = index + locations*l.classes + i*l.n + j;
                        costs[b*locations*l.n + i*l.n + j] = l.delta[p_index]*l.delta[p_index];
                    }
                }
            }
            int indexes[100];
            top_k(costs, l.batch*locations*l.n, 100, indexes);
            float cutoff = costs[indexes[99]];
            for (b = 0; b < l.batch; ++b) {
                int index = b*l.inputs;
                for (i = 0; i < locations; ++i) {
                    for (j = 0; j < l.n; ++j) {
                        int p_index = index + locations*l.classes + i*l.n + j;
                        if (l.delta[p_index]*l.delta[p_index] < cutoff) l.delta[p_index] = 0;
                    }
                }
            }
            free(costs);
        }
        *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
        printf("Detection Avg IOU: %f, Pos Cat: %f, All Cat: %f, Pos Obj: %f, Any Obj: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_allcat/(count*l.classes), avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
        //if(l.reorg) reorg(l.delta, l.w*l.h, size*l.n, l.batch, 0);
    }
}
/*
** 获取detection层的检测边框(跟get_detection_detections一样只是返回结果不一样)
** 输入: l         网络检测层
**       w, h       网络输入图像的宽高
**       thresh        阈值
**       probs        类别概率
**       boxes      检测到的边框
**       only_objectness       用于决定probs返回的是置信度还是对象类别概率
** 返回: detection_layer
*/
void get_detection_boxes(layer l, int w, int h, float thresh, float **probs, box *boxes, int only_objectness)
{
    int i,j,n;
    float *predictions = l.output;
    //int per_cell = 5*num+classes;
    for (i = 0; i < l.side*l.side; ++i){
        int row = i / l.side;
        int col = i % l.side;
        for(n = 0; n < l.n; ++n){
            int index = i*l.n + n;//预测候选框索引(最多能预测7*7*2=98个候选框)
            int p_index = l.side*l.side*l.classes + i*l.n + n;
            float scale = predictions[p_index];
            int box_index = l.side*l.side*(l.classes + l.n) + (i*l.n + n)*4;
            boxes[index].x = (predictions[box_index + 0] + col) / l.side * w;
            boxes[index].y = (predictions[box_index + 1] + row) / l.side * h;
            boxes[index].w = pow(predictions[box_index + 2], (l.sqrt?2:1)) * w;
            boxes[index].h = pow(predictions[box_index + 3], (l.sqrt?2:1)) * h;
            for(j = 0; j < l.classes; ++j){
                int class_index = i*l.classes;
                float prob = scale*predictions[class_index+j];
                probs[index][j] = (prob > thresh) ? prob : 0;
            }
            if(only_objectness){
                probs[index][0] = scale;
            }
        }
    }
}
/*
** 获取detection层的检测结果
** 输入: batch         一个batch中含有的图片张数(等于net.batch)
**       w, h       网络输入图像的宽高
**       thresh        阈值
**       dets       检测结果
** 返回: detection_layer
*/
void get_detection_detections(layer l, int w, int h, float thresh, detection *dets)
{
    int i, j, n;
    float *predictions = l.output;
    //int per_cell = 5*num+classes;
    for (i = 0; i < l.side*l.side; ++i) {
        int row = i / l.side;
        int col = i % l.side;
        for (n = 0; n < l.n; ++n) {
            int index = i*l.n + n;//预测候选框索引(最多能预测7*7*2=98个候选框)
            int p_index = l.side*l.side*l.classes + i*l.n + n;////预测候选框目标置信度存储索引
            float scale = predictions[p_index];
            int box_index = l.side*l.side*(l.classes + l.n) + (i*l.n + n) * 4;////预测候选框坐标位置存储索引
            box b;
            b.x = (predictions[box_index + 0] + col) / l.side * w;
            b.y = (predictions[box_index + 1] + row) / l.side * h;
            b.w = pow(predictions[box_index + 2], (l.sqrt ? 2 : 1)) * w;
            b.h = pow(predictions[box_index + 3], (l.sqrt ? 2 : 1)) * h;
            dets[index].bbox = b;
            dets[index].objectness = scale;
            for (j = 0; j < l.classes; ++j) {
                int class_index = i*l.classes;
                float prob = scale*predictions[class_index + j];//第i个grid cell中的第n个预测框是第j类的概率
                dets[index].prob[j] = (prob > thresh) ? prob : 0;
            }
        }
    }
}

你可能感兴趣的:(detection_layer层的实现)