| import sys |
| import os |
| import numpy as np |
| import torch |
|
|
|
|
| def get_topn_same_recall(score, labels, N, threshold = 0.0): |
| gt_num = list(labels).count(0) |
| hit = 0.0 |
|
|
| for i in range(min(N, len(labels))): |
| if labels[i] == 0 and score[i] >= threshold: |
| hit += 1 |
| if gt_num == 0 or N == 0: |
| return 0.0 |
|
|
| return hit / min(N, gt_num) |
|
|
|
|
|
|
| def get_topn_hit_rate(labels, N): |
| """HR@N: Top-N中是否存在至少一个正样本(label==0)，是则为1，否则为0""" |
| for i in range(min(N, len(labels))): |
| if labels[i] == 0: |
| return 1.0 |
| return 0.0 |
|
|
|
|
| def TopNHitRate(query_dic, TopN): |
| """计算所有query的平均HR@N""" |
| hr_sum = 0.0 |
| query_num = 0.0 |
|
|
| for _, data in query_dic.items(): |
| query_score = np.array(data['score']) |
| query_label = np.array(data['label']) |
|
|
| index_score = np.argsort(-query_score) |
| query_label_sort = query_label[index_score] |
|
|
| if 0 in query_label: |
| hr_sum += get_topn_hit_rate(query_label_sort, TopN) |
| query_num += 1 |
|
|
| if query_num == 0: |
| return 0.0 |
| return hr_sum / query_num |
|
|
|
|
| def TopNAcc(query_dic, TopN): |
| same_ap = 0.0 |
| same_query_num = 0.0 |
| |
| for _, data in query_dic.items(): |
| query_score = np.array(data['score']) |
| query_label = np.array(data['label']) |
|
|
| index_score = np.argsort(-query_score) |
| query_score_sort = query_score[index_score] |
| query_label_sort = query_label[index_score] |
|
|
| if 0 in query_label: |
| same_ap += get_topn_same_recall(query_score_sort, query_label_sort, TopN) |
| same_query_num += 1 |
| |
| if same_query_num == 0: |
| return 0.0 |
| return same_ap / same_query_num |
|
|
|
|
| def tongkuan_score_print(query_dic_all, sim_weight): |
| results_all = [] |
| results_all.append(['same_Top1','same_Top5','same_Top10', 'same_Top20','same_Top30','sim', 'HR@1', 'HR@5', 'HR@10', 'HR@20']) |
|
|
| query_score_dic = {} |
| for query, value in query_dic_all.items(): |
| query_score_dic[query] = {} |
| query_score_dic[query]['score'] = value['sim'] |
| query_score_dic[query]['label'] = value['label'] |
|
|
| |
| same_rate_Top1 = TopNAcc(query_score_dic, 1) |
| same_rate_Top5 = TopNAcc(query_score_dic, 5) |
| same_rate_Top10 = TopNAcc(query_score_dic, 10) |
| same_rate_Top20 = TopNAcc(query_score_dic, 20) |
| same_rate_Top30 = TopNAcc(query_score_dic, 30) |
| sim = sim_weight[0] * same_rate_Top1 + sim_weight[1] * same_rate_Top5 + sim_weight[2] * same_rate_Top10 + sim_weight[3] * same_rate_Top20 + sim_weight[4] * same_rate_Top30 |
|
|
| hr_1 = TopNHitRate(query_score_dic, 1) |
| hr_5 = TopNHitRate(query_score_dic, 5) |
| hr_10 = TopNHitRate(query_score_dic, 10) |
| hr_20 = TopNHitRate(query_score_dic, 20) |
|
|
| print ('##### wo margin #####') |
| print ('same_rate_Top1: ', same_rate_Top1) |
| print ('same_rate_Top5: ', same_rate_Top5) |
| print ('same_rate_Top10: ', same_rate_Top10) |
| print ('same_rate_Top20: ', same_rate_Top20) |
| print ('same_rate_Top30: ', same_rate_Top30) |
| print ('sim: ', sim) |
| print ('HR@1: ', hr_1) |
| print ('HR@5: ', hr_5) |
| print ('HR@10: ', hr_10) |
| print ('HR@20: ', hr_20) |
|
|
| results_all.append([same_rate_Top1, same_rate_Top5, same_rate_Top10, same_rate_Top20, same_rate_Top30, sim, hr_1, hr_5, hr_10, hr_20]) |
|
|
| results_all_dict = {} |
| for i in range(len(results_all[0])): |
| results_all_dict[results_all[0][i]] = results_all[1][i] |
| return results_all_dict |
