darknet之region_layer源碼分析

void forward_region_layer(const region_layer l, network_state state)
{
    int i,j,b,t,n;
    // 25
    int size = l.coords + l.classes + 1;
    memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));
    #ifndef GPU
    flatten(l.output, l.w*l.h, size*l.n, l.batch, 1);
    #endif
    // 每個預測框的置信度
    for (b = 0; b < l.batch; ++b){
        for(i = 0; i < l.h*l.w*l.n; ++i){
            int index = size*i + b*l.outputs;
            l.output[index + 4] = logistic_activate(l.output[index + 4]);
        }
    }


#ifndef GPU
    if (l.softmax_tree){
        for (b = 0; b < l.batch; ++b){
            for(i = 0; i < l.h*l.w*l.n; ++i){
                int index = size*i + b*l.outputs;
                softmax_tree(l.output + index + 5, 1, 0, 1, l.softmax_tree, l.output + index + 5);
            }
        }
    } else if (l.softmax){
        for (b = 0; b < l.batch; ++b){
            for(i = 0; i < l.h*l.w*l.n; ++i){
                int index = size*i + b*l.outputs;
                softmax(l.output + index + 5, l.classes, 1, l.output + index + 5, 1);
            }
        }
    }
#endif
    if(!state.train) return;
    memset(l.delta, 0, l.outputs * l.batch * sizeof(float));
    float avg_iou = 0;
    float recall = 0;
    float avg_cat = 0;
    float avg_obj = 0;
    float avg_anyobj = 0;
    int count = 0;
    int class_count = 0;
    *(l.cost) = 0;
    for (b = 0; b < l.batch; ++b) {
        if(l.softmax_tree){
            int onlyclass_id = 0;
            // x,y,w,h,class_id
            for(t = 0; t < l.max_boxes; ++t){
                box truth = float_to_box(state.truth + t*5 + b*l.truths);
                if(!truth.x) break; // continue;
                int class_id = state.truth[t*5 + b*l.truths + 4];
                float maxp = 0;
                int maxi = 0;
                // ???
                if(truth.x > 100000 && truth.y > 100000){
                    for(n = 0; n < l.n*l.w*l.h; ++n){
                        int index = size*n + b*l.outputs + 5;
                        float scale =  l.output[index-1];
                        float p = scale*get_hierarchy_probability(l.output + index, l.softmax_tree, class_id);
                        if(p > maxp){
                            maxp = p;
                            maxi = n;
                        }
                    }
                    int index = size*maxi + b*l.outputs + 5;
                    // 類別的誤差t-o,證明輸出到誤差經歷了mse
                    delta_region_class(l.output, l.delta, index, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat, l.focal_loss);
                    ++class_count;
                    onlyclass_id = 1;
                    break;
                }
            }
            if(onlyclass_id) continue;
        }
        // output組織方式 : h(13)*w(13)*n(5)*size(25)
        for (j = 0; j < l.h; ++j) {
            for (i = 0; i < l.w; ++i) {
                for (n = 0; n < l.n; ++n) {
                    int index = size*(j*l.w*l.n + i*l.n + n) + b*l.outputs;
                    // 25中的0,1,2,3：x,y,w,h，l.biases就是anchor
                    box pred = get_region_box(l.output, l.biases, n, index, i, j, l.w, l.h);
                    float best_iou = 0;
                    int best_class_id = -1;
                    // 每個預測框與所有ground truth計算IOU
                    for(t = 0; t < l.max_boxes; ++t){
                        box truth = float_to_box(state.truth + t*5 + b*l.truths);
                        int class_id = state.truth[t * 5 + b*l.truths + 4];
                        if (class_id >= l.classes) continue; // if label contains class_id more than number of classes in the cfg-file
                        if(!truth.x) break; // continue;
                        float iou = box_iou(pred, truth);
                        if (iou > best_iou) {
                            best_class_id = state.truth[t*5 + b*l.truths + 4];
                            best_iou = iou;
                        }
                    }
                    // 25中的4 : confidence
                    avg_anyobj += l.output[index + 4];
                    // logistic_gradient是sigmoid的導數，置信度誤差
                    // (t-o)*o*(1-o) 證明輸出到誤差經歷了sigmoid和mse
                    l.delta[index + 4] = l.noobject_scale * ((0 - l.output[index + 4]) * logistic_gradient(l.output[index + 4]));
                    if(l.classfix == -1) l.delta[index + 4] = l.noobject_scale * ((best_iou - l.output[index + 4]) * logistic_gradient(l.output[index + 4]));
                    else{
                        // 大於閾值不參與反向傳播
                        if (best_iou > l.thresh) {
                            l.delta[index + 4] = 0;
                            if(l.classfix > 0){
                                delta_region_class(l.output, l.delta, index + 5, best_class_id, l.classes, l.softmax_tree, l.class_scale*(l.classfix == 2 ? l.output[index + 4] : 1), &avg_cat, l.focal_loss);
                                ++class_count;
                            }
                        }
                    }

                    if(*(state.net.seen) < 12800){
                        box truth = {0};
                        truth.x = (i + .5)/l.w;
                        truth.y = (j + .5)/l.h;
                        truth.w = l.biases[2*n];
                        truth.h = l.biases[2*n+1];
                        if(DOABS){
                            truth.w = l.biases[2*n]/l.w;
                            truth.h = l.biases[2*n+1]/l.h;
                        }
                        // 同時計算IOU和delta，座標的誤差
                        // x,y的誤差經歷了sigmoid和mse; w,h的誤差經歷了mse
                        delta_region_box(truth, l.output, l.biases, n, index, i, j, l.w, l.h, l.delta, .01);
                    }
                }
            }
        }
        for(t = 0; t < l.max_boxes; ++t){
            box truth = float_to_box(state.truth + t*5 + b*l.truths);
            int class_id = state.truth[t * 5 + b*l.truths + 4];
            if (class_id >= l.classes) {
                printf("\n Warning: in txt-labels class_id=%d >= classes=%d in cfg-file. In txt-labels class_id should be [from 0 to %d] \n", class_id, l.classes, l.classes-1);
                getchar();
                continue; // if label contains class_id more than number of classes in the cfg-file
            }

            if(!truth.x) break; // continue;
            float best_iou = 0;
            int best_index = 0;
            int best_n = 0;
            // cell的座標
            i = (truth.x * l.w);
            j = (truth.y * l.h);
            //printf("%d %f %d %f\n", i, truth.x*l.w, j, truth.y*l.h);
            box truth_shift = truth;
            // 移動到原點，爲了計算IOU，預測框同理
            truth_shift.x = 0;
            truth_shift.y = 0;
            //printf("index %d %d\n",i, j);
            // 遍歷ground truth所在cell的5個預測框或先驗框anchor，找出IOU最大的
            for(n = 0; n < l.n; ++n){
                int index = size*(j*l.w*l.n + i*l.n + n) + b*l.outputs;
                box pred = get_region_box(l.output, l.biases, n, index, i, j, l.w, l.h);
                if(l.bias_match){
                    pred.w = l.biases[2*n];
                    pred.h = l.biases[2*n+1];
                    if(DOABS){
                        pred.w = l.biases[2*n]/l.w;
                        pred.h = l.biases[2*n+1]/l.h;
                    }
                }
                //printf("pred: (%f, %f) %f x %f\n", pred.x, pred.y, pred.w, pred.h);
                // 移動到原點，爲了計算IOU，對應上面
                pred.x = 0;
                pred.y = 0;
                float iou = box_iou(pred, truth_shift);
                if (iou > best_iou){
                    best_index = index;
                    best_iou = iou;
                    best_n = n;
                }
            }
            //printf("%d %f (%f, %f) %f x %f\n", best_n, best_iou, truth.x, truth.y, truth.w, truth.h);

            // 同時計算IOU和delta
            float iou = delta_region_box(truth, l.output, l.biases, best_n, best_index, i, j, l.w, l.h, l.delta, l.coord_scale);
            if(iou > .5) recall += 1;
            avg_iou += iou;

            //l.delta[best_index + 4] = iou - l.output[best_index + 4];
            avg_obj += l.output[best_index + 4];
            // (t-o)*o*(1-o) 證明輸出到誤差經歷了sigmoid和mse
            l.delta[best_index + 4] = l.object_scale * (1 - l.output[best_index + 4]) * logistic_gradient(l.output[best_index + 4]);
            if (l.rescore) {
                l.delta[best_index + 4] = l.object_scale * (iou - l.output[best_index + 4]) * logistic_gradient(l.output[best_index + 4]);
            }

            if (l.map) class_id = l.map[class_id];
            // 一般只計算delta
            delta_region_class(l.output, l.delta, best_index + 5, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat, l.focal_loss);
            ++count;
            ++class_count;
        }
    }
    //printf("\n");
    #ifndef GPU
    flatten(l.delta, l.w*l.h, size*l.n, l.batch, 0);
    #endif
    *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
    printf("Region Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, Avg Recall: %f,  count: %d\n", avg_iou/count, avg_cat/class_count, avg_obj/count, avg_anyobj/(l.w*l.h*l.n*l.batch), recall/count, count);
}