-
Notifications
You must be signed in to change notification settings - Fork 0
/
run.py
164 lines (135 loc) · 6.76 KB
/
run.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
import numpy as np
import torch
import torch.optim
from torch.optim.lr_scheduler import *
from torch.autograd import Variable
from tensorboardX import SummaryWriter
import os, json
import configs
from utils.data_utils import SimpleDataManager
from methods.simple import Simple
from utils.io_utils import model_dict, parse_args, get_resume_file, get_best_file
class Experiment():
def __init__(self, params):
np.random.seed(10)
image_size = 224
real_base_file = configs.data_dir['real'] + 'base.json'
real_val_file = configs.data_dir['real'] + 'val.json'
rumor_base_file = configs.data_dir['rumor'] + 'base.json'
rumor_val_file = configs.data_dir['rumor'] + 'val.json'
real_base_datamgr = SimpleDataManager(image_size, configs.real_normalize_param, batch_size=32)
real_base_loader = real_base_datamgr.get_data_loader(real_base_file, aug=params.train_aug)
rumor_base_datamgr = SimpleDataManager(image_size, configs.rumor_normalize_param, batch_size=32)
rumor_base_loader = rumor_base_datamgr.get_data_loader(rumor_base_file, aug=params.train_aug)
real_val_datamgr = SimpleDataManager(image_size, configs.real_normalize_param, batch_size=32)
real_val_loader = real_val_datamgr.get_data_loader(real_val_file, aug=params.train_aug)
rumor_val_datamgr = SimpleDataManager(image_size, configs.rumor_normalize_param, batch_size=32)
rumor_val_loader = rumor_val_datamgr.get_data_loader(rumor_val_file, aug=params.train_aug)
novel_file = configs.data_dir['test'] + 'novel.json'
novel_datamgr = SimpleDataManager(image_size, batch_size=64)
novel_loader = novel_datamgr.get_data_loader(novel_file, aug=False, shuffle=False)
optimizer = params.optimizer
model = Simple(model_dict[params.model], 10)
model = model.cuda()
key = params.tag
writer = SummaryWriter(log_dir=os.path.join(params.vis_log, key))
params.checkpoint_dir = '%s/checkpoints/%s' % (configs.save_dir, params.checkpoint_dir)
if not os.path.isdir(params.vis_log):
os.makedirs(params.vis_log)
outfile_template = os.path.join(params.checkpoint_dir.replace("checkpoints", "features"), "%s.hdf5")
if params.mode == 'train' and not os.path.isdir(params.checkpoint_dir):
os.makedirs(params.checkpoint_dir)
if params.resume or params.mode == 'test':
if params.mode == 'test':
resume_file = get_best_file(params.checkpoint_dir)
else:
resume_file = get_resume_file(params.checkpoint_dir)
if resume_file is not None:
tmp = torch.load(resume_file)
params.start_epoch = tmp['epoch'] + 1
model.load_state_dict(tmp['state'])
print('Model Loaded!')
self.params = params
self.image_size = image_size
self.optimizer = optimizer
self.outfile_template = outfile_template
self.novel_file = novel_file
self.novel_loader = novel_loader
self.real_base_loader = real_base_loader
self.real_val_loader = real_val_loader
self.rumor_base_loader = rumor_base_loader
self.rumor_val_loader = rumor_val_loader
self.writer = writer
self.model = model
self.key = key
def train(self):
if self.optimizer == 'Adam':
train_optimizer = torch.optim.Adam(self.model.parameters())
train_scheduler = None
elif self.optimizer == 'SGD':
train_optimizer = torch.optim.SGD(self.model.parameters(), lr=self.params.train_lr, momentum=0.9,
weight_decay=0.001)
train_scheduler = StepLR(train_optimizer, step_size=50, gamma=0.1)
else:
raise ValueError('Unknown optimizer, please define by yourself')
max_acc = 0
start_epoch = self.params.start_epoch
stop_epoch = self.params.stop_epoch
for epoch in range(start_epoch, stop_epoch):
self.model.train()
self.model.train_loop(epoch, self.real_base_loader, self.rumor_base_loader, train_optimizer,
train_scheduler, self.writer)
self.model.eval()
acc = self.test('val', epoch)
if acc > max_acc: # for baseline and baseline++, we don't use validation here so we let acc = -1
print("best model! save...")
max_acc = acc
outfile = os.path.join(self.params.checkpoint_dir, 'best_model.tar')
torch.save({'epoch': epoch, 'state': self.model.state_dict()}, outfile)
if (epoch % self.params.save_freq == 0) or (epoch == stop_epoch - 1):
outfile = os.path.join(self.params.checkpoint_dir, '{:d}.tar'.format(epoch))
torch.save({'epoch': epoch, 'state': self.model.state_dict()}, outfile)
def test(self, split='novel', epoch=0):
self.outfile = self.outfile_template % split
if split == 'novel':
pred = self.model.test_loop(self.novel_loader)
self.produce_test_result(pred)
else:
acc = self.model.val_loop(epoch, self.real_val_loader, self.rumor_val_loader, self.writer)
print('Test Acc = %4.2f%%' % acc)
return acc
def produce_test_result(self, pred):
with open(self.novel_file, 'r') as f:
j = json.load(f)
with open('submit.csv', 'w') as f:
f.write('id,label\n')
for i, image in enumerate(j['image_names']):
f.write('%s,%s\n' % (image.split('/')[-1].split('.')[0], 1 - pred[i]))
def run(self):
if self.params.mode == 'train':
self.train()
elif self.params.mode == 'test':
self.test()
elif self.params.mode == 'save_feat':
self.save_feat()
def save_feat(self):
feats = np.array([[0] * 513])
for i, (x, y) in enumerate(self.real_base_loader):
feat = self.model.feature(Variable(x.cuda()))
feat = feat.detach().cpu().numpy()
feat = np.concatenate([feat, y[:, np.newaxis]], 1)
feats = np.concatenate([feats, feat], 0)
if i % 100 == 0:
print(f'{i * 32} real samples done!')
for i, (x, y) in enumerate(self.rumor_base_loader):
feat = self.model.feature(Variable(x.cuda()))
feat = feat.detach().cpu().numpy()
feat = np.concatenate([feat, y[:, np.newaxis]], 1)
feats = np.concatenate([feats, feat], 0)
if i % 100 == 0:
print(f'{i * 32} rumor samples done!')
np.savetxt("features.csv", feats[1:], delimiter=',')
if __name__ == '__main__':
params = parse_args()
exp = Experiment(params)
exp.run()