AbnormalDetection/train.py at master · EachSheep/AbnormalDetection · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
import os
import torch
from torch.utils.tensorboard import SummaryWriter
import json
import time

from sklearn.metrics import average_precision_score, roc_auc_score, recall_score, precision_score
from sklearn.metrics import precision_recall_curve, roc_curve
import numpy as np

from Trainer import Trainer
from Tester import Tester
from argparser import args
from utils.set_logger import set_logger

if __name__ == '__main__':

    torch.manual_seed(args.random_seed)
    if args.cache_dir == '':
        runs_dir = os.path.join(args.experiment_dir, 'runs')
    else:
        runs_dir = os.path.join(args.cache_dir, 'runs')
    if not os.path.exists(runs_dir):
        os.makedirs(runs_dir)
    summarywriter_dir = os.path.join(runs_dir, 'train')
    if os.path.exists(summarywriter_dir):
        for root, dirs, files in os.walk(summarywriter_dir, topdown=False):
            for name in files:
                os.remove(os.path.join(root, name))
            for name in dirs:
                os.rmdir(os.path.join(root, name))
        os.rmdir(summarywriter_dir)
        os.makedirs(summarywriter_dir)
    else:
        os.makedirs(summarywriter_dir)

    logger = set_logger(summarywriter_dir)
    # writer = SummaryWriter(summarywriter_dir)
    trainer = Trainer(args)

    begin_time = time.time()
    for cur_epoch in range(0, trainer.args.epochs):
        cur_train_loss = trainer.train(cur_epoch)
        logger.info("TRAIN LOSS: %.4f" % (cur_train_loss))
        # writer.add_scalars("mini-valid/loss", {"train" : cur_train_loss}, cur_epoch)

        # if (cur_epoch + 1) % 5 == 0:
        #     cur_roc, cur_pr, precision, recall, F1, cur_test_loss, cur_label, cur_predict, total_uid  = trainer.eval(cur_epoch)
        #     logger.info("ROC-AUC: %.4f, PR-AUC: %.4f, PRECISION: %.4f, RECALL: %.4f, F1: %.4f, VALID LOSS: %.4f" % (cur_roc, cur_pr, precision, recall, F1, cur_test_loss))
        #     writer.add_scalars("mini-valid/loss", {"loss_valid" : cur_train_loss}, cur_epoch)
        #     writer.add_scalar("mini-valid/roc", cur_roc, cur_epoch)
        #     writer.add_scalar("mini-valid/pr", cur_pr, cur_epoch)
        #     writer.add_scalar("mini-valid/precision", precision, cur_epoch)
        #     writer.add_scalar("mini-valid/recall", recall, cur_epoch)
        #     writer.add_scalar("mini-valid/F1", F1, cur_epoch)
        #     writer.add_pr_curve(f"mini-valid/pr_curve-{cur_epoch}", cur_label, cur_predict, global_step=cur_epoch)

    # writer.flush()
    # writer.close()
    if args.cache_dir == '':
        model_path = os.path.join(args.experiment_dir, 'models', args.weight_name)
    else:
        model_path = os.path.join(args.cache_dir, 'models', args.weight_name)
    trainer.save_weights(model_path = model_path)
    end_time = time.time()
    print("Training time: {:.3f}".format((end_time - begin_time))) # 113.505 / 856408 = 0.00013
    args.train_time_per_epoch = "{:.3f}".format((end_time - begin_time) / args.epochs)
    # cur_roc, cur_pr, precision, recall, F1, cur_test_loss, cur_label, cur_predict, total_uid = trainer.test()
    # np.save(os.path.join(summarywriter_dir, 'valid_label.npy'), cur_label)
    # np.save(os.path.join(summarywriter_dir, 'valid_predict.npy'), cur_predict)
    # np.save(os.path.join(summarywriter_dir, 'valid_uid.npy'), total_uid)
    # logger.info("ROC-AUC: %.4f, PR-AUC: %.4f, PRECISION: %.4f, RECALL: %.4f, F1: %.4f, VALID LOSS: %.4f" % (cur_roc, cur_pr, precision, recall, F1, cur_test_loss))

    # args.ROCAUC = cur_roc
    # args.PRAUC = cur_pr
    # args.PRECISION = precision
    # args.RECALL = recall
    # args.F1 = F1
    # args.VALID_LOSS = cur_test_loss
    if args.cache_dir == '':
        preserve_dir = os.path.join(args.experiment_dir, f'{args.log_dir}-{args.log_label}')
    else:
        preserve_dir = os.path.join(args.cache_dir, f'{args.log_dir}-{args.log_label}')
    if not os.path.exists(preserve_dir):
        os.makedirs(preserve_dir)
    setting_path = os.path.join(preserve_dir, 'train_setting.json')
    json.dump(args.__dict__, open(setting_path, 'w'), indent=4)

    # valid
    # tester = Tester(args)
    # cur_roc, cur_pr, precision, recall, F1, cur_test_loss, cur_label, cur_predict, total_uid = tester.eval(state_dict_path = model_path)
    # logger.info("ROC-AUC: %.4f, PR-AUC: %.4f, PRECISION: %.4f, RECALL: %.4f, F1: %.4f, VALID LOSS: %.4f" % (cur_roc, cur_pr, precision, recall, F1, cur_test_loss))
    # np.save(os.path.join(summarywriter_dir, 'valid_label.npy'), cur_label)
    # np.save(os.path.join(summarywriter_dir, 'valid_predict.npy'), cur_predict)
    # np.save(os.path.join(summarywriter_dir, 'valid_uid.npy'), total_uid)
    # preserve_dir = os.path.join(args.experiment_dir, f'{args.log_dir}-{args.log_label}')
    # if not os.path.exists(preserve_dir):
    #     os.makedirs(preserve_dir)
    # args.ROCAUC = cur_roc
    # args.PRAUC = cur_pr
    # args.PRECISION = precision
    # args.RECALL = recall
    # args.F1 = F1
    # args.VALID_LOSS = cur_test_loss
    # setting_path = os.path.join(preserve_dir, 'train_setting.json')
    # json.dump(args.__dict__, open(setting_path, 'w'), indent=4)

    # 大概15个epoch收敛
    # 训练全部的大概:
    # 测试:444.0410s,(1165655 + 30855)

    # devnet 5个epoch收敛: 总共耗时:10851.901(12个epoch),因此耗时:10851.901/12*5 = 4521.625

    # 新数据集的结果

    # BCE, 随机采样
    # after_newtest_data :  88.53 39.23
    # after_newtest_data :  88.53 59.40
    # after_newtest_data :  88.53 55.48
    # after_newtest_data :  88.53 52.78

    # # BCE, 平衡采样
    # after_newtest_data :  89.01 39.33
    # after_newtest_data :  89.01 56.20
    # after_newtest_data :  89.01 54.10
    # after_newtest_data :  89.01 51.94

    # WeightedBCE, 随机采样
    # # 1781.497 2 epochs
    # 0.1, 0.9的权重
    # after_newtest_data :  87.50 37.97
    # after_newtest_data :  87.50 61.10
    # after_newtest_data :  87.50 56.84
    # after_newtest_data :  87.50 51.63

    # WeightedBCE, 随机采样
    # # 1781.497 2 epochs
    # 0.01, 0.99的权重
    # after_newtest_data :  87.40 37.17
    # after_newtest_data :  87.40 66.40
    # after_newtest_data :  87.40 56.62
    # after_newtest_data :  87.40 51.78


    # WeightedBCE, 平衡采样
    # # 1780.615 2 epochs
    # 0.1, 0.9的权重
    # after_newtest_data :  88.81 39.51
    # after_newtest_data :  88.81 63.30
    # after_newtest_data :  88.81 58.00
    # after_newtest_data :  88.81 53.35

    # # 0.01, 0.99的权重
    # after_newtest_data :  88.62 39.25
    # after_newtest_data :  88.62 74.70
    # after_newtest_data :  88.62 58.10
    # after_newtest_data :  88.62 53.20

    # 0.001, 0.999的权重
    # after_newtest_data :  88.29 39.08
    # after_newtest_data :  88.29 74.00
    # after_newtest_data :  88.29 59.24
    # after_newtest_data :  88.29 53.58

    # # # 0.01, 0.99的权重，部分训练
    # after_newtest_data :  87.16 36.56
    # after_newtest_data :  87.16 58.10
    # after_newtest_data :  87.16 54.62
    # after_newtest_data :  87.16 51.68

    # 对比学习+WeightedBCE, 平衡采样
    # 896.305, 1 epoch
    # after_newtest_data :  88.53 39.03
    # after_newtest_data :  88.53 53.60
    # after_newtest_data :  88.53 52.18
    # after_newtest_data :  88.53 51.15

    # DevNet, 平衡采样
    # after_newtest_data :  85.73 32.06
    # after_newtest_data :  85.73 41.00
    # after_newtest_data :  85.73 46.28
    # after_newtest_data :  85.73 47.92

    # # DevNet, 随机采样
    # after_newtest_data :  62.28 6.04
    # after_newtest_data :  62.28 10.40
    # after_newtest_data :  62.28 6.78
    # after_newtest_data :  62.28 6.33