-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathparty.py
executable file
·255 lines (210 loc) · 10.7 KB
/
party.py
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
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
import os, sys
sys.path.append(os.path.dirname(os.path.abspath(os.path.dirname(__file__))))
import torch
import time
from torch.utils.data import DataLoader
from pycocotools.cocoeval import COCOeval
import visdom
import tempfile
import json
from dataset.coco_dataset import COCO_Dataset
from dataset.voc_dataset import VOC_Dataset
from util.utils import detect
from util.utils import detect_retina
from util.voc_eval import voc_eval
from util.voc_eval_retina import voc_eval_retina
from option import opts, device
from util.evaluator import Evaluator
from tqdm import tqdm
class Party:
vis = None # 3 : Use Visdom
if opts.visdom:
vis = visdom.Visdom(port=opts.visdom_port)
num_classes = 0 # 4, 5 : Dataset & Dataloader ( coco:80 voc:20 )
train_loader = None
test_loader = None
if opts.model == 'yolov2_vgg_16':
image_resize = 416
elif opts.model == 'ssd_vgg_16':
image_resize = 300
elif opts.model == 'retinanet_resnet_50':
image_resize = 600
model = None # 6 : Model
coder = None
priors_cxcy = None
criterion = None # 7 : Criterion
optimizer = None # 8 : Optimizer
scheduler = None # 9 : Scheduler
epochs_start = opts.epochs_start # * : Training
epochs = opts.epochs
lr = opts.lr
def getDataLoader(self, dataset_type, dataset_root, batch_size, num_workers):
if dataset_type == 'coco':
self.num_classes = 81
train_set = COCO_Dataset(root=dataset_root, set_name='train2017', split='train', download=True, resize=self.image_resize)
test_set = COCO_Dataset(root=dataset_root, set_name='val2017', split='test', download=True, resize=self.image_resize)
elif dataset_type == 'voc':
self.num_classes = 21
train_set = VOC_Dataset(root=dataset_root, split='train', download=True, resize=self.image_resize)
test_set = VOC_Dataset(root=dataset_root, split='test', download=True, resize=self.image_resize)
if opts.model == 'yolov2_vgg_16':
self.num_classes = self.num_classes - 1
train_loader = DataLoader(dataset=train_set,
batch_size=batch_size,
collate_fn=train_set.collate_fn,
shuffle=True,
pin_memory=True,
num_workers=num_workers)
test_loader = DataLoader(dataset=test_set,
batch_size=1,
collate_fn=test_set.collate_fn,
shuffle=False)
return train_loader, test_loader
def train(self, epoch):
tic = time.time()
self.model.train()
# yolo v2 일 때, warm-up training
if opts.model == 'yolov2_vgg_16':
# warm up training
if epoch < 5:
for param_group in self.optimizer.param_groups:
param_group['lr'] = 1e-5
elif epoch == 5:
for param_group in self.optimizer.param_groups:
param_group['lr'] = 1e-4
for idx, datas in enumerate(self.train_loader):
images = datas[0]
boxes = datas[1]
labels = datas[2]
# assign to cuda
images = images.to(device)
boxes = [b.to(device) for b in boxes]
labels = [l.to(device) for l in labels]
preds = self.model(images) # Feed Forward
loss, losses = self.criterion(preds, boxes, labels) # Loss
# backward and update
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
toc = time.time()
for param_group in self.optimizer.param_groups:
lr = param_group['lr']
# for each steps (print for checking)
if idx % 100 == 0:
print('Epoch: [{0}]\t'
'Step: [{1}/{2}]\t'
'Loss: {loss:.4f}\t'
'Learning rate: {lr:.7f} s \t'
'Training Time : {time:.4f}\t'
.format(epoch, idx, len(self.train_loader),
loss=loss,
lr=lr,
time=toc - tic))
# VISDOM
if self.vis is not None:
self.vis.line(
X=torch.ones((1, 1)).cpu() * idx + epoch * self.train_loader.__len__(), # step
Y=torch.Tensor([loss]).unsqueeze(0).cpu(),
win='train_loss',
update='append',
opts=dict(xlabel='step',
ylabel='Loss',
title='training loss',
legend=['Total Loss']))
# Save Checkpoint
if not os.path.exists(opts.save_path):
os.mkdir(opts.save_path)
checkpoint = {'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict()}
if self.scheduler is not None:
checkpoint = {'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict()}
torch.save(checkpoint, os.path.join(opts.save_path, opts.save_filename) + '.{}.pth.tar'.format(epoch))
if self.scheduler is not None:
self.scheduler.step()
def test(self, epoch):
# @@@ Load Trained Model : 학습된 모델 불러오기 @@@
print('Testing Start fo <', opts.model, '> Using Dataset :<', opts.dataset_type, '>')
self.model.eval()
check_point = torch.load(os.path.join(opts.save_path, opts.save_filename) + '.{}.pth.tar'.format(epoch),
map_location=device)
self.model.load_state_dict(check_point['model_state_dict'], strict=True)
tic = time.time()
sum_loss = 0
evaluator = Evaluator(data_type=opts.dataset_type)
# @@@ Testing for test datasets @@@
with torch.no_grad():
for idx, datas in enumerate(self.test_loader):
images = datas[0]
boxes = datas[1]
labels = datas[2]
images = images.to(device)
boxes = [b.to(device) for b in boxes]
labels = [l.to(device) for l in labels]
# Get Loss
preds = self.model(images) # feed forward
loss, _ = self.criterion(preds, boxes, labels) # calculate loss
sum_loss += loss.item()
# @@@@@ Get EVALUATION Results @@@@@
if opts.test_eval:
# FIXME ) RetinaNet
if opts.model =='retinanet_resnet_50':
pred_boxes, pred_labels, pred_scores = detect(pred=preds,
coder=self.coder,
min_score=opts.conf_thres,
n_classes=self.num_classes)
# pred_boxes, pred_labels, pred_scores = detect_retina(self.priors_cxcy,
# pred=preds,
# min_score=0.1,
# max_overlap=0.45,
# top_k=200)
else:
pred_boxes, pred_labels, pred_scores = detect(pred=preds,
coder=self.coder,
min_score=opts.conf_thres,
n_classes=self.num_classes)
if opts.dataset_type == 'voc':
img_name = datas[3][0]
img_info = datas[4][0]
info = (pred_boxes, pred_labels, pred_scores, img_name, img_info)
elif opts.dataset_type == 'coco':
img_id = self.test_loader.dataset.img_id[idx]
img_info = self.test_loader.dataset.coco.loadImgs(ids=img_id)[0]
coco_ids = self.test_loader.dataset.coco_ids
info = (pred_boxes, pred_labels, pred_scores, img_id, img_info, coco_ids)
evaluator.get_info(info)
toc = time.time()
# @@@@@ Get Results @@@@@
if idx % 10 == 0 or idx == len(self.test_loader) - 1:
print('Epoch: [{0}]\t'
'Step: [{1}/{2}]\t'
'Loss: {loss:.4f}\t'
'Time : {time:.4f}\t'
.format(epoch,
idx, len(self.test_loader),
loss=loss,
time=toc-tic))
# @@@ Evaluation Start @@@
print('Start Evaluation...')
mAP = evaluator.evaluate(self.test_loader.dataset)
mean_loss = sum_loss / len(self.test_loader)
# @@@ VISDOM @@@
if self.vis is not None:
# loss plot
self.vis.line(X=torch.ones((1, 2)).cpu() * epoch, # step
Y=torch.Tensor([mean_loss, mAP]).unsqueeze(0).cpu(),
win='test_loss',
update='append',
opts=dict(xlabel='step',
ylabel='test',
title='test loss',
legend=['test Loss', 'mAP']))
def resume(self):
checkpoint = torch.load(os.path.join(opts.save_path, opts.save_filename)+'.{}.pth.tar'.format(opts.epochs_start-1))
self.model.load_state_dict(checkpoint['model_state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
self.scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
print('\n...Loaded checkpoint from epoch %d...\n' % (int(opts.epochs_start) - 1))