프로젝트

일반

사용자정보

통계
| 개정판:

hytos / DTI_PID / WebServer / symbol_training / train.py @ bfb338d6

이력 | 보기 | 이력해설 | 다운로드 (12.9 KB)

1
"""
2
Training For Small Object
3
"""
4
import os
5
import argparse
6
import torch.nn as nn
7
from torch.utils.data import DataLoader
8
from src.doftech_dataset import DoftechDataset
9
from src.utils import *
10
from src.loss import YoloLoss
11
from src.yolo_net import Yolo
12
from src.yolo_doftech import YoloD
13
import shutil
14
import visdom
15
import cv2
16
import pickle
17
import numpy as np
18
from src.vis_utils import array_tool as at
19
from src.vis_utils.vis_tool import visdom_bbox
20

    
21
loss_data = {'X': [], 'Y': [], 'legend_U':['total', 'coord', 'conf', 'cls']}
22

    
23
# 형상 CLASS
24
DOFTECH_CLASSES= ['gate', 'globe', 'butterfly', 'check', 'ball', 'relief',
25
                  '3way_solenoid', 'gate_pressure', 'globe_pressure', 'butterfly_pressure', 'ball_shutoff', 'ball_pressure','ball_motor', 'plug_pressure',
26
                  'circle', 'inst_console', 'inst_console_dcs', 'inst_console_sih', 'logic_dcs', 'utility', 'specialty_items', 'logic', 'logic_local_console_dcs',
27
                  'reducer', 'blind_spectacle_open', 'blind_insertion_open', 'blind_spectacle_close', 'blind_insertion_close',
28
                  'strainer_basket', 'strainer_conical', 'fitting_capillary_tubing', 'meter_ultrasonic', 'strainer_y', 'tube_pitot'
29
                  ,'opc']
30

    
31
use_voc_model = True
32
use_visdom = False
33
if use_visdom:
34
    visdom = visdom.Visdom(port='8088')
35
    #visdom = visdom.Visdom()
36

    
37
def train(name=None, classes=None, bigs=None, root_path=None, pre_trained_model_path=None):
38
    global DOFTECH_CLASSES
39
    DOFTECH_CLASSES = classes
40

    
41
    parser = argparse.ArgumentParser("You Only Look Once: Unified, Real-Time Object Detection")
42
    parser.add_argument("--image_size", type=int, default=512, help="The common width and height for all images")
43
    parser.add_argument("--batch_size", type=int, default=10, help="The number of images per batch")
44

    
45
    # Training 기본 Setting
46
    parser.add_argument("--momentum", type=float, default=0.9)
47
    parser.add_argument("--decay", type=float, default=0.0005)
48
    parser.add_argument("--dropout", type=float, default=0.5)
49
    parser.add_argument("--num_epoches", type=int, default=205)
50
    parser.add_argument("--test_interval", type=int, default=20, help="Number of epoches between testing phases")
51
    parser.add_argument("--object_scale", type=float, default=1.0)
52
    parser.add_argument("--noobject_scale", type=float, default=0.5)
53
    parser.add_argument("--class_scale", type=float, default=1.0)
54
    parser.add_argument("--coord_scale", type=float, default=5.0)
55
    parser.add_argument("--reduction", type=int, default=32)
56
    parser.add_argument("--es_min_delta", type=float, default=0.0,
57
                        help="Early stopping's parameter: minimum change loss to qualify as an improvement")
58
    parser.add_argument("--es_patience", type=int, default=0,
59
                        help="Early stopping's parameter: number of epochs with no improvement after which training will be stopped. Set to 0 to disable this technique.")
60

    
61
    # 확인해야 하는 PATH
62
    parser.add_argument("--data_path", type=str, default=os.path.join(root_path, 'training'), help="the root folder of dataset") # 학습 데이터 경로 -> image와 xml의 상위 경로 입력
63
    parser.add_argument("--data_path_test", type=str, default=os.path.join(root_path, 'test'), help="the root folder of dataset") # 테스트 데이터 경로 -> test할 이미지만 넣으면 됨
64
    #parser.add_argument("--pre_trained_model_type", type=str, choices=["model", "params"], default="model")
65
    parser.add_argument("--pre_trained_model_path", type=str, default=pre_trained_model_path) # Pre-training 된 모델 경로
66

    
67
    parser.add_argument("--saved_path", type=str, default=os.path.join(root_path, 'checkpoint')) # training 된 모델 저장 경로
68
    parser.add_argument("--conf_threshold", type=float, default=0.35)
69
    parser.add_argument("--nms_threshold", type=float, default=0.5)
70
    opt = parser.parse_args()
71

    
72
    if not os.path.isdir(opt.saved_path):
73
        os.mkdir(opt.saved_path)
74

    
75
    # 학습할 클래스들을 저장하고 인식 시 불러와 사용합니다.
76
    with open(os.path.join(opt.saved_path, name + "_info.info"), 'w') as stream:
77
        con = str(len(DOFTECH_CLASSES))
78
        names = '\n'.join(DOFTECH_CLASSES)
79
        bigs = '\n'.join(bigs)
80
        con = con + '\n' + names + '\n' + '***bigs***' + '\n' + bigs
81
        stream.write(con)
82

    
83
    if torch.cuda.is_available():
84
        torch.cuda.manual_seed(123)
85
    else:
86
        torch.manual_seed(123)
87

    
88
    learning_rate_schedule = {"0": 1e-5, "5": 1e-4,
89
                              "80": 1e-5, "110": 1e-6}
90

    
91
    training_params = {"batch_size": 1,#opt.batch_size,
92
                       "shuffle": True,
93
                       "drop_last": True,
94
                       "collate_fn": custom_collate_fn}
95

    
96
    test_params = {"batch_size": opt.batch_size,
97
                   "shuffle": False,
98
                   "drop_last": False,
99
                   "collate_fn": custom_collate_fn}
100

    
101
    training_set = DoftechDataset(opt.data_path, opt.image_size, is_training=True, classes=DOFTECH_CLASSES)
102
    training_generator = DataLoader(training_set, **training_params)
103

    
104
    test_set = DoftechDataset(opt.data_path_test, opt.image_size, is_training=False, classes=DOFTECH_CLASSES)
105
    test_generator = DataLoader(test_set, **test_params)
106

    
107
    # BUILDING MODEL =======================================================================
108
    if use_voc_model :
109
        pre_model = Yolo(20).cuda()
110
        pre_model.load_state_dict(torch.load(opt.pre_trained_model_path), strict=False)
111
        model = YoloD(pre_model, training_set.num_classes).cuda()
112
    else :
113
        model = Yolo(training_set.num_classes).cuda()
114

    
115
    nn.init.normal_(list(model.modules())[-1].weight, 0, 0.01)
116

    
117
    criterion = YoloLoss(training_set.num_classes, model.anchors, opt.reduction)
118
    optimizer = torch.optim.SGD(model.parameters(), lr=1e-5, momentum=opt.momentum, weight_decay=opt.decay)
119

    
120
    best_loss = 1e10
121
    best_epoch = 0
122
    num_iter_per_epoch = len(training_generator)
123
    loss_step = 0
124
    # ======================================================================================
125

    
126
    # TRAINING =============================================================================
127
    save_count = 0
128

    
129
    model.train()
130
    for epoch in range(opt.num_epoches):
131
        if str(epoch) in learning_rate_schedule.keys():
132
            for param_group in optimizer.param_groups:
133
                param_group['lr'] = learning_rate_schedule[str(epoch)]
134

    
135
        for iter, batch in enumerate(training_generator):
136
            image, label, image2 = batch
137
            image = Variable(image.cuda(), requires_grad=False)
138
            if torch.cuda.is_available():
139
                image = Variable(image.cuda(), requires_grad=True)
140
                origin = Variable(image2.cuda(), requires_grad=False)
141
            else:
142
                image = Variable(image, requires_grad=True)
143

    
144
            optimizer.zero_grad()
145
            logits = model(image)
146
            loss, loss_coord, loss_conf, loss_cls = criterion(logits, label)
147
            loss.backward()
148
            optimizer.step()
149

    
150
            if iter % (opt.test_interval * 5) == 0:
151
                print("Epoch: {}/{}, Iteration: {}/{}, Lr: {}, Loss:{:.5f} (Coord:{:.5f} Conf:{:.5f} Cls:{:.5f})".format
152
                    (epoch + 1, opt.num_epoches, iter + 1, num_iter_per_epoch, optimizer.param_groups[0]['lr'], loss,
153
                    loss_coord,loss_conf,loss_cls))
154

    
155
                if use_visdom:
156
                    predictions = post_processing(logits, opt.image_size, DOFTECH_CLASSES, model.anchors, opt.conf_threshold,
157
                                                  opt.nms_threshold)
158

    
159
                    gt_image = at.tonumpy(image[0])
160
                    gt_image = visdom_bbox(gt_image, label[0])
161
                    visdom.image(gt_image, opts=dict(title='gt_box_image'), win=3)
162
                    #
163
                    origin_image = at.tonumpy(origin[0])
164
                    origin_image = visdom_bbox(origin_image, [])
165
                    visdom.image(origin_image, opts=dict(title='origin_box_image'), win=4)
166

    
167
                    image = at.tonumpy(image[0])
168

    
169
                    if len(predictions) != 0:
170
                        box_image = visdom_bbox(image, predictions[0])
171
                        visdom.image(box_image, opts=dict(title='box_image'), win=2)
172

    
173
                    elif len(predictions) == 0:
174
                        box_image = visdom_bbox(image, [])
175
                        visdom.image(box_image, opts=dict(title='box_image'), win=2)
176

    
177
                    loss_dict = {
178
                        'total' : loss.item(),
179
                        'coord' : loss_coord.item(),
180
                        'conf' : loss_conf.item(),
181
                        'cls' : loss_cls.item()
182
                    }
183

    
184
                    visdom_loss(visdom, loss_step, loss_dict)
185
                    loss_step = loss_step + 1
186

    
187
        if epoch % opt.test_interval == 0:
188
            model.eval()
189
            loss_ls = []
190
            loss_coord_ls = []
191
            loss_conf_ls = []
192
            loss_cls_ls = []
193
            for te_iter, te_batch in enumerate(test_generator):
194
                te_image, te_label, _ = te_batch
195
                num_sample = len(te_label)
196
                if torch.cuda.is_available():
197
                    te_image = te_image.cuda()
198
                with torch.no_grad():
199
                    te_logits = model(te_image)
200
                    batch_loss, batch_loss_coord, batch_loss_conf, batch_loss_cls = criterion(te_logits, te_label)
201
                loss_ls.append(batch_loss * num_sample)
202
                loss_coord_ls.append(batch_loss_coord * num_sample)
203
                loss_conf_ls.append(batch_loss_conf * num_sample)
204
                loss_cls_ls.append(batch_loss_cls * num_sample)
205

    
206
            te_loss = sum(loss_ls) / test_set.__len__()
207
            te_coord_loss = sum(loss_coord_ls) / test_set.__len__()
208
            te_conf_loss = sum(loss_conf_ls) / test_set.__len__()
209
            te_cls_loss = sum(loss_cls_ls) / test_set.__len__()
210
            print("Test>> Epoch: {}/{}, Lr: {}, Loss:{:.5f} (Coord:{:.5f} Conf:{:.5f} Cls:{:.5f})".format(
211
                epoch + 1, opt.num_epoches, optimizer.param_groups[0]['lr'], te_loss, te_coord_loss, te_conf_loss, te_cls_loss))
212

    
213
            model.train()
214
            if te_loss + opt.es_min_delta < best_loss:
215
                save_count += 1
216
                best_loss = te_loss
217
                best_epoch = epoch
218
                print("SAVE MODEL")
219
                # for debug for each loss
220
                #torch.save(model.state_dict(), os.path.join(opt.saved_path, name + "_only_params_" + str(save_count) + "_" + "{:.5f}".format(best_loss) + ".pth"))
221
                #torch.save(model, os.path.join(opt.saved_path, name + "_whole_model_" + str(save_count) + "_" + "{:.5f}".format(best_loss) + ".pth"))
222
                # save
223
                torch.save(model.state_dict(), os.path.join(opt.saved_path, name + "_only_params.pth"))
224
                torch.save(model, os.path.join(opt.saved_path, name + "_whole_model.pth"))
225
            else:
226
                save_count += 1
227
                # for debug for each loss
228
                #torch.save(model.state_dict(), os.path.join(opt.saved_path, name + "_only_params_" + str(save_count) + "_" + "{:.5f}".format(te_loss) + ".pth"))
229
                #torch.save(model, os.path.join(opt.saved_path, name + "_whole_model_" + str(save_count) + "_" + "{:.5f}".format(te_loss) + ".pth"))
230

    
231
            # Early stopping
232
            if epoch - best_epoch > opt.es_patience > 0:
233
                print("Stop training at epoch {}. The lowest loss achieved is {}".format(epoch, te_loss))
234
                break
235

    
236
def visdom_loss(visdom, loss_step, loss_dict):
237
    loss_data['X'].append(loss_step)
238
    loss_data['Y'].append([loss_dict[k] for k in loss_data['legend_U']])
239
    visdom.line(
240
        X=np.stack([np.array(loss_data['X'])] * len(loss_data['legend_U']), 1),
241
        Y=np.array(loss_data['Y']),
242
        win=30,
243
        opts=dict(xlabel='Step',
244
                  ylabel='Loss',
245
                  title='YOLO_V2',
246
                  legend=loss_data['legend_U']),
247
        update='append'
248
    )
249

    
250
if __name__ == "__main__":
251
    datas = ['gate', 'globe', 'butterfly', 'check', 'ball', 'relief',
252
                  '3way_solenoid', 'gate_pressure', 'globe_pressure', 'butterfly_pressure', 'ball_shutoff', 'ball_pressure','ball_motor', 'plug_pressure',
253
                  'inst', 'func_valve', 'inst_console', 'inst_console_dcs', 'inst_console_sih', 'logic_dcs', 'utility', 'specialty_items', 'logic', 'logic_local_console_dcs',
254
                  'reducer', 'blind_spectacle_open', 'blind_insertion_open', 'blind_spectacle_close', 'blind_insertion_close',
255
                  'strainer_basket', 'strainer_conical', 'fitting_capillary_tubing', 'meter_ultrasonic', 'strainer_y', 'tube_pitot',
256
                  'opc']
257
    data_path = os.path.join(os.path.dirname(os.path.realpath(__file__)) + '\\Data\\', 'VnV')
258
    train(name='VnV', classes=datas, root_path=data_path, pre_trained_model_path=os.path.dirname(os.path.realpath(
259
                                                       __file__)) + '\\pre_trained_model\\only_params_trained_yolo_voc')
클립보드 이미지 추가 (최대 크기: 500 MB)