-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbsearch_decouple_search_continue_channel_wise.py
198 lines (143 loc) · 8.08 KB
/
bsearch_decouple_search_continue_channel_wise.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
import argparse
import os
import random
import numpy as np
import torch.backends.cudnn
import torch.optim
import torch.utils.data
from config import PATH
from datasets.aux_dataset import FeatureDataset
from model.classifier import get_classifier
from utils import accuracy, AvgMetric, write, LabelSmoothingCrossEntropy
torch.backends.cudnn.benchmark = True
def train_loop(args, classifier, train_loader, lr, epsilon, M, epochs, test_loader=None, test_M=1, test_epsilon=0.0, flag='training'):
write('Starting {} from epsilon={:.4f} with M={}'.format(flag, epsilon, M), args.log_file)
optimizer = torch.optim.Adam(classifier.parameters(), lr=lr * args.batch_size_per_gpu / 256.)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, epochs, eta_min=0)
criterion = LabelSmoothingCrossEntropy()
for ep in range(1, epochs+1):
classifier.train()
loss_metric = AvgMetric()
for _it, (features, targets, _) in enumerate(train_loader):
features = features.float().cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
assert (features.size(1) == args.feat_dim) and (len(features.size()) == 2)
bs = features.size(0)
with torch.no_grad():
features = features.unsqueeze(1) + (torch.randn(size=[features.size(0), M, features.size(1)]).cuda() * epsilon * args.train_std.reshape(1,1,-1))
features = features.reshape(-1, args.feat_dim)
logits = classifier(features)
loss = criterion(logits, targets.reshape(-1, 1).repeat(1, M).reshape(-1))
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_metric.update(loss.item(), bs)
write('\rEpoch : {}/{} Iter : {}/{} Epsilon : {:.4f} Lr : {:.6f} Loss : {:.4f}'.format(ep, epochs, _it+1, len(train_loader), epsilon, scheduler.get_last_lr()[0], loss_metric.show()), end='\n' if _it+1 == len(train_loader) else '', log_file=args.log_file)
scheduler.step()
if test_loader is not None:
write('Start testing when training, epsilon={} with M={}'.format(test_epsilon, test_M), args.log_file)
classifier.eval()
acc_metric = AvgMetric()
with torch.no_grad():
for features, targets, _ in test_loader:
features = features.float().cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
bs = features.size(0)
features = features.unsqueeze(1) + (torch.randn(size=[features.size(0), test_M, features.size(1)]).cuda() * test_epsilon * args.train_std.reshape(1, 1, -1))
features = features.reshape(-1, args.feat_dim)
logits = classifier(features)
acc = accuracy(logits, targets.reshape(-1, 1).repeat(1, test_M).reshape(-1))
acc_metric.update(acc, bs)
write('samples in test loop is {}'.format(acc_metric.n), args.log_file)
test_acc = acc_metric.show()
write('Test_acc : {:.4f}'.format(test_acc), args.log_file)
return test_acc
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--shot', default=1, type=int)
parser.add_argument('--dataset', default='Imagenet', choices=['Imagenet', 'CUB'])
parser.add_argument('--arch', default='deit_small_p16', choices=['deit_small_p16', 'deit_large_p7', 'deit_base_p4', 'resnet50'])
parser.add_argument('--pretrain_method', default='DINO', choices=['DINO', 'MSN', 'MoCov3', 'SimCLR', 'BYOL', 'CLIP', 'DenseCL'])
parser.add_argument('--round', default=0.9, type=float)
parser.add_argument('--M', default=200, type=int)
parser.add_argument('--pre_epochs', default=100, type=int)
parser.add_argument('--cycle_epochs', default=20, type=int)
parser.add_argument('--final_epochs', default=100, type=int)
parser.add_argument('--search_lr', default=1.0, type=float)
parser.add_argument('--right', default=10.0, type=float)
parser.add_argument('--batch_size_per_gpu', default=256, type=int)
parser.add_argument('--batch_size_test_per_gpu', default=256, type=int)
parser.add_argument('--num_workers_per_gpu', default=0, type=int)
parser.add_argument('--seed', default=0, type=int)
args = parser.parse_args()
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
if args.dataset == 'Imagenet':
args.num_classes = 1000
elif args.dataset == 'CUB':
args.num_classes = 200
else:
raise NotImplementedError
if 'deit_small' in args.arch:
args.feat_dim = 384
elif 'deit_base' in args.arch:
args.feat_dim = 768
elif 'deit_large' in args.arch:
args.feat_dim = 1024
elif 'resnet50' in args.arch:
args.feat_dim = 2048
else:
raise NotImplementedError
exp_dir = os.path.join(PATH, 'checkpoint', args.pretrain_method, args.dataset, args.arch)
exp_dir_ft = os.path.join(exp_dir, 'IbM2', 'search_lr_{}_bs_{}_rd_{}_right_{}_M_{}_Pre_{}_C_{}_F_{}_seed_{}'.format(args.search_lr, args.batch_size_per_gpu, args.round, args.right, args.M, args.pre_epochs, args.cycle_epochs, args.final_epochs, args.seed), '{}shot'.format(args.shot))
args.exp_dir_ft = exp_dir_ft
args.log_file = os.path.join(exp_dir_ft, 'search_log.txt')
if not os.path.exists(exp_dir_ft):
os.makedirs(exp_dir_ft)
if os.path.isfile(args.log_file):
os.remove(args.log_file)
write(vars(args), args.log_file)
epsilons = []
for run in range(3):
write('\n\n', args.log_file)
args.train_path = os.path.join(PATH, 'checkpoint', args.pretrain_method, args.dataset, args.arch, 'features', 'train_{}shot_{}.pth'.format(args.shot, run))
train_set = FeatureDataset(args.train_path)
write('Run : {} the number of samples in train set is {}'.format(run, len(train_set)), args.log_file)
#####################################
args.train_std = torch.std(train_set.data, dim=0).cuda()
assert args.train_std.size(0) == args.feat_dim
write('\n\n', args.log_file)
#####################################
write('Run : {} Std in train(plain) set is {}'.format(run, args.train_std), args.log_file)
train_loader = torch.utils.data.DataLoader(train_set, shuffle=True, batch_size=args.batch_size_per_gpu, num_workers=args.num_workers_per_gpu, pin_memory=True)
init_classifier = get_classifier(args)
write(init_classifier, args.log_file)
write('Searching layers is feature layer', args.log_file)
write('********** Baseline **********', args.log_file)
th = train_loop(args, init_classifier, train_loader, lr=args.search_lr, epsilon=0.0, M=1, epochs=args.pre_epochs, test_loader=train_loader, test_M=1, test_epsilon=0.0, flag='training before search (to get th)')
write('Naive test_acc in train set with no epsilon is : {:.2f}'.format(th), args.log_file)
th = min(args.round, th)
write('\n', args.log_file)
write('\n', args.log_file)
write('********** Searching Stage **********', args.log_file)
write('Threshold : {:.2f}'.format(th), args.log_file)
left_bound = 0.0
right_bound = args.right
epsilon = right_bound / 2
search_classifier = get_classifier(args)
while True:
train_acc_now = train_loop(args, search_classifier, train_loader, lr=args.search_lr, epsilon=epsilon, M=args.M, epochs=args.cycle_epochs, test_loader=train_loader, test_M=args.M, test_epsilon=epsilon, flag='searching')
if train_acc_now > th:
left_bound = epsilon # harder
else:
right_bound = epsilon # easier
epsilon = (left_bound + right_bound) / 2.0
if right_bound - left_bound < 0.05:
break
epsilons.append(epsilon)
write('Epsilon after search is {:.4f}'.format(epsilon), args.log_file)
write('***Final Epsilon {:.4f} {:.4f} {:.4f} ***'.format(epsilons[0], epsilons[1], epsilons[2]), args.log_file)
if __name__ == '__main__':
main()