-
Notifications
You must be signed in to change notification settings - Fork 23
/
Copy pathmain.py
174 lines (153 loc) · 7.75 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
import pickle
import os
import json
import numpy as np
import torch
import torchvision
import torch.nn as nn
import torch.optim as optim
import sklearn.metrics
from dataset import MDataset
from methods import RandomMethod, SmallExact, Seafaring
from environments import get_env
import argparse
import random
def measure(method, test_dataset, n_classes, n_round=100, epoch_per_round=10, batchsize=16, device='cpu'):
train_dataset = method.get_dataset()
accs = []
aucs = []
for i_round in range(n_round + 1):
model = torchvision.models.resnet18(weights=torchvision.models.ResNet18_Weights.DEFAULT)
model.fc = nn.Linear(512, n_classes)
method.set_model(model)
device = torch.device(device)
model.to(device)
ns = train_dataset.labelcount()
print('ns', ns)
weight = ns.sum() / (ns + 1e-9)
weight = torch.FloatTensor(weight)
weight = weight.to(device)
criterion = nn.CrossEntropyLoss(weight)
optimizer = optim.SGD(model.parameters(), lr=0.0001, momentum=0.9)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batchsize, shuffle=True)
for epoch in range(epoch_per_round):
print(i_round, epoch, len(train_dataset), end=' ')
model.train()
losss = []
for batch_idx, (data, target, idx) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
losss.append(float(loss))
loss = np.mean(losss)
print(loss, end='\r')
model.eval()
acc = 0
res = []
preds = []
for batch_idx, (data, target, idx) in enumerate(test_dataset):
data = data.to(device)
output = model(data.unsqueeze(0))[0].cpu()
pred = output.argmax()
preds.append(int(pred))
acc += int(pred == target)
res.append((float(torch.softmax(output, dim=0)[1]), target))
fpr, tpr, thresholds = sklearn.metrics.roc_curve([c for a, c in res], [a for a, c in res], pos_label=1)
auc = sklearn.metrics.auc(fpr, tpr)
aucs.append(auc)
acc = acc / len(test_dataset)
accs.append(acc)
print()
print(acc, auc)
print(np.bincount(preds, minlength=n_classes))
if i_round < n_round:
train_dataset = method.increment_dataset()
return [accs, aucs, model]
def build_data(n, poslabels, tag_to_image_filepath):
with open(tag_to_image_filepath, 'rb') as f:
tag_to_item_dict = pickle.load(f)
existing_images = set(map(lambda x: x.split('.')[0], os.listdir('imgs')))
tag_to_item_dict = {
tag: list(filter(lambda x: x in existing_images, li)) for tag, li in tag_to_item_dict.items()
}
testlabels = np.random.randint(len(poslabels), size=n)
n_testlabels = np.bincount(testlabels, minlength=len(poslabels))
pos = sum([[('imgs/' + x + '.jpg', 1) for x in tag_to_item_dict[t][-n_testlabels[i]:]] for i, t in enumerate(poslabels)], [])
tags = [tag for tag in tag_to_item_dict.keys() if len(tag_to_item_dict[tag]) >= 1]
neg_tags = [tags[i] for i in np.random.randint(len(tags), size=n)]
neg = [('imgs/' + tag_to_item_dict[tag][0] + '.jpg', 0) for tag in neg_tags]
data = pos + neg
return data
def build_user_vitual_user(n, user, test, threshold):
with open('virtual_user_source.pickle', 'rb') as f:
li = pickle.load(f)[user]
pos = list(map(lambda x: (x[1], 1), filter(lambda x: x[0] > threshold, li)))
neg = list(map(lambda x: (x[1], 0), filter(lambda x: x[0] < threshold, li)))
print(len(pos), len(neg))
if test == -1:
data = pos[:n] + neg[:n]
else:
data = pos[test:][-n:] + neg[test:][-n:]
return data
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--method', choices=['Seafaring', 'Random', 'SmallExact'], default='MaxMax')
parser.add_argument('--env', choices=['OpenImage', 'Flickr'], default='OpenImage')
parser.add_argument('--apikey', type=str, default=None, help='API key of Flickr. Valid only for Flickr env.')
parser.add_argument('--tiara_budget', type=int, default=1000)
parser.add_argument('--budget_per_round', type=int, default=1)
parser.add_argument('--initdata', type=int, default=1, help='NumSizeber of the initial labelled data.')
parser.add_argument('--testdata', type=int, default=100, help='Size of the test dataset.')
parser.add_argument('--nround', type=int, default=100, help='Number of rounds of active learning.')
parser.add_argument('--nepoch', type=int, default=100, help='Number of epochs for training the target model.')
parser.add_argument('--alpha', type=float, default=1.0, help='The alpha parameter of Tiara.')
parser.add_argument('--threshold', type=float, default=0.6, help='Thoreshold of Positive data. Valid only for Flickr env.')
parser.add_argument('--batchsize', type=int, default=16)
parser.add_argument('--poolsize', type=int, default=1000, help='Size of the poolsize for SmallExact method')
parser.add_argument('--device', type=str, default='cpu')
parser.add_argument('--poslabels', type=str, nargs='+', default=['Cat'], help='List of positive labels. Valid only for OpenImage env.')
parser.add_argument('--user', type=int, default=0, help='Id of the target virtual user, i.e., category. Valid only for Flickr env. See also create_virtual_users.py.')
parser.add_argument('--initialtags', type=str, default=None, help='Path to the tag file.')
parser.add_argument('--resdir', type=str, default='results')
args = parser.parse_args()
if not os.path.exists(args.resdir):
os.makedirs(args.resdir)
with open(f'{args.resdir}/args.json', 'w') as f:
json.dump(vars(args), f, indent=4)
print(vars(args))
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
n_classes = 2
if args.env == 'OpenImage':
test_data = build_data(args.testdata, args.poslabels, 'openimage_tag_to_image_test.pickle')
test_data = MDataset(test_data)
init_data = build_data(args.initdata, args.poslabels, 'openimage_tag_to_image.pickle')
else:
test_data = build_user_vitual_user(args.testdata, args.user, -1, args.threshold)
test_data = MDataset(test_data)
init_data = build_user_vitual_user(args.initdata, args.user, args.testdata, args.threshold)
env = get_env(args.env, args.apikey, args.initialtags, args.poslabels, args.user, args.device, args.threshold)
if args.method == 'Random':
method = RandomMethod(init_data, budget_per_round=args.budget_per_round, poslabels=args.poslabels, env=env)
elif args.method == 'SmallExact':
method = SmallExact(init_data, budget_per_round=args.budget_per_round, poslabels=args.poslabels, poolsize=args.poolsize, device=args.device, env=env)
else:
method = Seafaring(init_data, budget_per_round=args.budget_per_round, tiara_budget=args.tiara_budget, alpha=args.alpha, poslabels=args.poslabels, device=args.device, env=env)
accs, aucs, model = measure(method, test_data, n_classes, args.nround, args.nepoch, args.batchsize, args.device)
res = [accs, aucs]
print(res)
if hasattr(env, 'save_cache'):
env.save_cache()
with open(f'{args.resdir}/res.pickle', 'wb') as f:
pickle.dump(res, f)
if hasattr(method, 'maxprob'):
with open(f'{args.resdir}/maxprob.pickle', 'wb') as f:
pickle.dump(method.maxprob, f)
torch.save(model.state_dict(), f'{args.resdir}/model.pth')
if __name__ == '__main__':
main()