-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathmain.py
209 lines (165 loc) · 8.51 KB
/
main.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
import os
import argparse
import torch
import time
import pickle
import numpy as np
from torch import nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from dataset import VideoAnomalyDataset_C3D
from models import model
from tqdm import tqdm
from aggregate import remake_video_output, evaluate_auc, remake_video_3d_output
torch.backends.cudnn.benchmark = False
# Config
def get_configs():
parser = argparse.ArgumentParser(description="VAD-Jigsaw config")
parser.add_argument("--val_step", type=int, default=500)
parser.add_argument("--print_interval", type=int, default=100)
parser.add_argument("--epochs", type=int, default=100)
parser.add_argument("--gpu_id", type=str, default=0)
parser.add_argument("--log_date", type=str, default=None)
parser.add_argument("--batch_size", type=int, default=64)
parser.add_argument("--static_threshold", type=float, default=0.3)
parser.add_argument("--sample_num", type=int, default=5)
parser.add_argument("--filter_ratio", type=float, default=0.8)
parser.add_argument("--checkpoint", type=str, default=None)
parser.add_argument("--dataset", type=str, default="shanghaitech", choices=['shanghaitech', 'ped2', 'avenue'])
args = parser.parse_args()
args.device = torch.device("cuda:{}".format(args.gpu_id) if torch.cuda.is_available() else "cpu")
if args.dataset in ['shanghaitech', 'avenue']:
args.filter_ratio = 0.8
elif args.dataset == 'ped2':
args.filter_ratio = 0.5
return args
def train(args):
if not args.log_date:
running_date = time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime())
else:
running_date = args.log_date
print("The running_data : {}".format(running_date))
for k,v in vars(args).items():
print("-------------{} : {}".format(k, v))
# Load Data
data_dir = f"/irip/wangguodong_2020/projects/datasets/vad/{args.dataset}/training"
detect_pkl = f'detect/{args.dataset}_train_detect_result_yolov3.pkl'
vad_dataset = VideoAnomalyDataset_C3D(data_dir,
dataset=args.dataset,
detect_dir=detect_pkl,
fliter_ratio=args.filter_ratio,
frame_num=args.sample_num,
static_threshold=args.static_threshold)
vad_dataloader = DataLoader(vad_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4, pin_memory=True)
net = model.WideBranchNet(time_length=args.sample_num, num_classes=[args.sample_num ** 2, 81])
if args.checkpoint is not None:
state = torch.load(args.checkpoint)
print('load ' + args.checkpoint)
net.load_state_dict(state, strict=True)
net.cuda()
smoothed_auc, smoothed_auc_avg, _ = val(args, net)
exit(0)
net.cuda(args.device)
net = net.train()
criterion = nn.CrossEntropyLoss(reduction='mean')
optimizer = optim.Adam(params=net.parameters(), lr=1e-4)
# Train
log_dir = './log/{}/'.format(running_date)
writer = SummaryWriter(log_dir)
t0 = time.time()
global_step = 0
max_acc = -1
timestamp_in_max = None
for epoch in range(args.epochs):
for it, data in enumerate(vad_dataloader):
video, obj, temp_labels, spat_labels, t_flag = data['video'], data['obj'], data['label'], data["trans_label"], data["temporal"]
n_temp = t_flag.sum().item()
obj = obj.cuda(args.device, non_blocking=True)
temp_labels = temp_labels[t_flag].long().view(-1).cuda(args.device)
spat_labels = spat_labels[~t_flag].long().view(-1).cuda(args.device)
temp_logits, spat_logits = net(obj)
temp_logits = temp_logits[t_flag].view(-1, args.sample_num)
spat_logits = spat_logits[~t_flag].view(-1, 9)
temp_loss = criterion(temp_logits, temp_labels)
spat_loss = criterion(spat_logits, spat_labels)
loss = temp_loss + spat_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
writer.add_scalar('Train/Loss', loss.item(), global_step=global_step)
writer.add_scalar('Train/Temporal', temp_loss.item(), global_step=global_step)
writer.add_scalar('Train/Spatial', spat_loss.item(), global_step=global_step)
if (global_step + 1) % args.print_interval == 0:
print("[{}:{}/{}]\tloss: {:.6f} t_loss: {:.6f} s_loss: {:.6f} \ttime: {:.6f}".\
format(epoch, it + 1, len(vad_dataloader), loss.item(), temp_loss.item(), spat_loss.item(), time.time() - t0))
t0 = time.time()
global_step += 1
if global_step % args.val_step == 0 and epoch >= 5:
smoothed_auc, smoothed_auc_avg, temp_timestamp = val(args, net)
writer.add_scalar('Test/smoothed_auc', smoothed_auc, global_step=global_step)
writer.add_scalar('Test/smoothed_auc_avg', smoothed_auc_avg, global_step=global_step)
if smoothed_auc > max_acc:
max_acc = smoothed_auc
timestamp_in_max = temp_timestamp
save = './checkpoint/{}_{}.pth'.format('best', running_date)
torch.save(net.state_dict(), save)
print('cur max: ' + str(max_acc) + ' in ' + timestamp_in_max)
net = net.train()
def val(args, net=None):
if not args.log_date:
running_date = time.strftime("%Y-%m-%d-%H-%M-%S", time.localtime())
else:
running_date = args.log_date
print("The running_date : {}".format(running_date))
# Load Data
data_dir = f"/irip/wangguodong_2020/projects/datasets/vad/{args.dataset}/testing"
detect_pkl = f'detect/{args.dataset}_test_detect_result_yolov3.pkl'
testing_dataset = VideoAnomalyDataset_C3D(data_dir,
dataset=args.dataset,
detect_dir=detect_pkl,
fliter_ratio=args.filter_ratio,
frame_num=args.sample_num)
testing_data_loader = DataLoader(testing_dataset, batch_size=256, shuffle=False, num_workers=4, drop_last=False)
net.eval()
video_output = {}
for data in tqdm(testing_data_loader):
videos = data["video"]
frames = data["frame"].tolist()
obj = data["obj"].cuda(args.device)
with torch.no_grad():
temp_logits, spat_logits = net(obj)
temp_logits = temp_logits.view(-1, args.sample_num, args.sample_num)
spat_logits = spat_logits.view(-1, 9, 9)
spat_probs = F.softmax(spat_logits, -1)
diag = torch.diagonal(spat_probs, offset=0, dim1=-2, dim2=-1)
scores = diag.min(-1)[0].cpu().numpy()
temp_probs = F.softmax(temp_logits, -1)
diag2 = torch.diagonal(temp_probs, offset=0, dim1=-2, dim2=-1)
scores2 = diag2.min(-1)[0].cpu().numpy()
for video_, frame_, s_score_, t_score_ in zip(videos, frames, scores, scores2):
if video_ not in video_output:
video_output[video_] = {}
if frame_ not in video_output[video_]:
video_output[video_][frame_] = []
video_output[video_][frame_].append([s_score_, t_score_])
micro_auc, macro_auc = save_and_evaluate(video_output, running_date, dataset=args.dataset)
return micro_auc, macro_auc, running_date
def save_and_evaluate(video_output, running_date, dataset='shanghaitech'):
pickle_path = './log/video_output_ori_{}.pkl'.format(running_date)
with open(pickle_path, 'wb') as write:
pickle.dump(video_output, write, pickle.HIGHEST_PROTOCOL)
if dataset == 'shanghaitech':
video_output_spatial, video_output_temporal, video_output_complete = remake_video_output(video_output, dataset=dataset)
else:
video_output_spatial, video_output_temporal, video_output_complete = remake_video_3d_output(video_output, dataset=dataset)
evaluate_auc(video_output_spatial, dataset=dataset)
evaluate_auc(video_output_temporal, dataset=dataset)
smoothed_res, smoothed_auc_list = evaluate_auc(video_output_complete, dataset=dataset)
return smoothed_res.auc, np.mean(smoothed_auc_list)
if __name__ == '__main__':
if not os.path.exists('checkpoint'):
os.makedirs('checkpoint')
args = get_configs()
train(args)