1 void validate_classifier_multi(char *datacfg, char *filename, char *weightfile)
2 {
3 int i, j;
4 network net = parse_network_cfg(filename);
5 set_batch_network(&net, 1);
6 if(weightfile){
7 load_weights(&net, weightfile);
8 }
9 srand(time(0));
10
11 list *options = read_data_cfg(datacfg);//读.data文件到option列表中
12
13 char *label_list = option_find_str(options, "labels", "data/labels.list");
14 //从读到的.data生成的option列表去找对饮的字段如labels，将labels的配置路径放到label_list指针中，
15 //然后如果labels的配置路径是"data/labels.list"，打印“使用默认配置”字样
16 char *valid_list = option_find_str(options, "valid", "data/train.list");// l,key,def; return def
17 int classes = option_find_int(options, "classes", 2);
18 int topk = option_find_int(options, "top", 1);
19 if (topk > classes) topk = classes;//找的比类别还多
20
21 char **labels = get_labels(label_list);
22 //将labels.list标签名读到lables字符指针，可以通过labels[i]访问标签
23 list *plist = get_paths(valid_list);//得到验证集的数据路径
24 int scales[] = {224, 288, 320, 352, 384};
25 int nscales = sizeof(scales)/sizeof(scales[0]);
26
27 char **paths = (char **)list_to_array(plist);
28 int m = plist->size;
29 free_list(plist);
30
31 float avg_acc = 0;
32 float avg_topk = 0;
33 int* indexes = (int*)calloc(topk, sizeof(int));
34
35 for(i = 0; i < m; ++i){
36 int class_id = -1;//一般用负数初始化
37 char *path = paths[i];//这里的路径名包括文件名之外的路径吗？
38 for(j = 0; j < classes; ++j){
39 if(strstr(path, labels[j])){
40 //在path字符串中查找labels[j]字符串第一次出现的位置
41 class_id = j;
42 //这里实现了数据集在训练过程中的类别的确定。还是看匹配，只要标签在文件名中
43 break;
44 }
45 }
46 float* pred = (float*)calloc(classes, sizeof(float));
47 image im = load_image_color(paths[i], 0, 0);
48 for(j = 0; j < nscales; ++j){
49 image r = resize_min(im, scales[j]);
50 resize_network(&net, r.w, r.h);
51 float *p = network_predict(net, r.data);
52 if(net.hierarchy) hierarchy_predictions(p, net.outputs, net.hierarchy, 1);
53 axpy_cpu(classes, 1, p, 1, pred, 1);
54 flip_image(r);
55 p = network_predict(net, r.data);
56 axpy_cpu(classes, 1, p, 1, pred, 1);
57 if(r.data != im.data) free_image(r);
58 }
59 free_image(im);
60 top_k(pred, classes, topk, indexes);
61 free(pred);
62 if(indexes[0] == class_id) avg_acc += 1;
63 for(j = 0; j < topk; ++j){
64 if(indexes[j] == class_id) avg_topk += 1;
65 }
66
67 printf("%d: top 1: %f, top %d: %f\n", i, avg_acc/(i+1), topk, avg_topk/(i+1));
68 }
69 }
70
71
72 void predict_classifier(char *datacfg, char *cfgfile, char *weightfile, char *filename, int top)
73 {//反初始化主要是类对象的析构
74 network net = parse_network_cfg_custom(cfgfile, 1, 0);
75 if(weightfile){
76 load_weights(&net, weightfile);
77 }
78 set_batch_network(&net, 1);
79 srand(2222222);
80
81 fuse_conv_batchnorm(net);
82 calculate_binary_weights(net);
83
84 list *options = read_data_cfg(datacfg);
85
86 char *name_list = option_find_str(options, "names", 0);
87 if(!name_list) name_list = option_find_str(options, "labels", "data/labels.list");
88 int classes = option_find_int(options, "classes", 2);
89 if (top == 0) top = option_find_int(options, "top", 1);
90 if (top > classes) top = classes;
91
92 int i = 0;
93 char **names = get_labels(name_list);
94 clock_t time;
95 int* indexes = (int*)calloc(top, sizeof(int));
96 char buff[256];
97 char *input = buff;
98 //int size = net.w;
99 while(1){
100 if(filename){
101 strncpy(input, filename, 256);//将filename的前256个字符复制到input中。
102 }else{
103 printf("Enter Image Path: ");
104 fflush(stdout);
105 input = fgets(input, 256, stdin);
106 if(!input) return;
107 strtok(input, "\n");
108 }
109 image im = load_image_color(input, 0, 0);
110 image r = letterbox_image(im, net.w, net.h);
111 //image r = resize_min(im, size);
112 //resize_network(&net, r.w, r.h);
113 printf("%d %d\n", r.w, r.h);
114
115 float *X = r.data;
116 time=clock();
117 float *predictions = network_predict(net, X);
118 if(net.hierarchy) hierarchy_predictions(predictions, net.outputs, net.hierarchy, 0);
119 top_k(predictions, net.outputs, top, indexes);
120 //按得分来排top k,indexes是新的排序指针，按升序排列，prediction越大的在indexes里面的id越是靠后。
121 printf("%s: Predicted in %f seconds.\n", input, sec(clock()-time));
122 for(i = 0; i < top; ++i){
123 int index = indexes[i];
124 //hierarchy是一个树形结构体指针变量。应该是没有的。
125 if(net.hierarchy) printf("%d, %s: %f, parent: %s \n",index, names[index], predictions[index], (net.hierarchy->parent[index] >= 0) ? names[net.hierarchy->parent[index]] : "Root");
126 else printf("%s: %f\n",names[index], predictions[index]);
127 //names[index]是分类的对应的类别名称如yb,ye,yf
128 //predictions[index]是推理置信度
129 }
130 if(r.data != im.data) free_image(r);
131 free_image(im);
132 if (filename) break;//可以批量测试，如果filename是False，跳出
133 }
134 }
135
浙公网安备 33010602011771号