-
Notifications
You must be signed in to change notification settings - Fork 1
/
ScanMix.py
237 lines (207 loc) · 10.8 KB
/
ScanMix.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
from __future__ import print_function
import sys
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import random
import os
import argparse
import numpy as np
import copy
import pdb
from pathlib import Path
import matplotlib.pyplot as plt
from utils.config import create_config
from utils.common_config import get_train_transformations, get_val_transformations, get_scan_transformations,\
get_train_dataset, get_train_dataloader,\
get_val_dataset, get_val_dataloader,\
get_model, get_criterion
from utils.evaluate_utils import scanmix_test
from utils.train_utils import scanmix_train, scanmix_eval_train, scanmix_warmup, scanmix_scan
parser = argparse.ArgumentParser(description='DivideMix')
parser.add_argument('--noise_mode', default='sym')
parser.add_argument('--lambda_u', default=25, type=float, help='weight for unsupervised loss')
parser.add_argument('--r', default=0, type=float, help='noise ratio')
parser.add_argument('--seed', default=123)
parser.add_argument('--gpuid', default=0, type=int)
parser.add_argument('--inference', default=None, type=str)
parser.add_argument('--load_state_dict', default=None, type=str)
parser.add_argument('--cudaid', default=0)
parser.add_argument('--dividemix_only', action='store_true')
parser.add_argument('--lr_sl', type=float, default=None)
parser.add_argument('--config_env',
help='Config file for the environment')
parser.add_argument('--config_exp',
help='Config file for the experiment')
args = parser.parse_args()
device = device = torch.device('cuda:{}'.format(args.cudaid))
random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
#meta_info
meta_info = copy.deepcopy(args.__dict__)
p = create_config(args.config_env, args.config_exp, meta_info)
meta_info['dataset'] = p['dataset']
meta_info['noise_file'] = '{}/{:.2f}'.format(p['noise_dir'], args.r)
if args.noise_mode == 'asym':
meta_info['noise_file'] += '_asym'
elif 'semantic' in args.noise_mode:
meta_info['noise_file'] += '_{}'.format(args.noise_mode)
meta_info['noise_file'] += '.json'
meta_info['probability'] = None
meta_info['pred'] = None
meta_info['noise_rate'] = args.r
Path(os.path.join(p['scanmix_dir'], 'savedDicts')).mkdir(parents=True, exist_ok=True)
class NegEntropy(object):
def __call__(self,outputs):
probs = torch.softmax(outputs, dim=1)
return torch.mean(torch.sum(probs.log()*probs, dim=1))
def create_model():
model = get_model(p, p['scan_model'])
model = model.to(device)
return model
test_log = open(os.path.join(p['scanmix_dir'], 'acc.txt'), 'w')
stats_log = open(os.path.join(p['scanmix_dir'], 'stats.txt'), 'w')
def get_loader(p, mode, meta_info):
if mode == 'test':
meta_info['mode'] = 'test'
val_transformations = get_val_transformations(p)
val_dataset = get_val_dataset(p, val_transformations, meta_info=meta_info)
val_dataloader = get_val_dataloader(p, val_dataset)
return val_dataloader
elif mode == 'train':
meta_info['mode'] = 'labeled'
train_transformations = get_train_transformations(p)
labeled_dataset = get_train_dataset(p, train_transformations,
split='train', to_noisy_dataset=p['to_noisy_dataset'], meta_info=meta_info)
labeled_dataloader = get_train_dataloader(p, labeled_dataset)
meta_info['mode'] = 'unlabeled'
unlabeled_dataset = get_train_dataset(p, train_transformations,
split='train', to_noisy_dataset=p['to_noisy_dataset'], meta_info=meta_info)
unlabeled_dataloader = get_train_dataloader(p, unlabeled_dataset)
return labeled_dataloader, unlabeled_dataloader
elif mode == 'eval_train':
meta_info['mode'] = 'all'
eval_transformations = get_val_transformations(p)
eval_dataset = get_train_dataset(p, eval_transformations,
split='train', to_noisy_dataset=p['to_noisy_dataset'], meta_info=meta_info)
eval_dataloader = get_val_dataloader(p, eval_dataset)
return eval_dataloader
elif mode == 'warmup':
meta_info['mode'] = 'all'
warmup_transformations = get_train_transformations(p)
warmup_dataset = get_train_dataset(p, warmup_transformations,
split='train', to_noisy_dataset=p['to_noisy_dataset'], meta_info=meta_info)
warmup_dataloader = get_train_dataloader(p, warmup_dataset, explicit_batch_size=p['batch_size']*2)
return warmup_dataloader
elif mode == 'neighbors':
meta_info['mode'] = 'neighbor'
train_transformations = get_train_transformations(p)
neighbor_dataset = get_train_dataset(p, train_transformations,
split='train', to_neighbors_dataset=True, to_noisy_dataset=p['to_noisy_dataset'], meta_info=meta_info)
neighbor_dataloader = get_train_dataloader(p, neighbor_dataset, explicit_batch_size=p['batch_size_scan'])
return neighbor_dataloader
else:
raise NotImplementedError
CE = nn.CrossEntropyLoss(reduction='none')
CEloss = nn.CrossEntropyLoss()
criterion_dm, criterion_sl = get_criterion(p)
conf_penalty = NegEntropy()
def main():
print('| Building net')
net1 = create_model()
net2 = create_model()
cudnn.benchmark = True
optimizer1 = optim.SGD(net1.parameters(), lr=p['lr'], momentum=0.9, weight_decay=5e-4)
optimizer2 = optim.SGD(net2.parameters(), lr=p['lr'], momentum=0.9, weight_decay=5e-4)
if args.load_state_dict is not None:
print('Loading saved state dict from {}'.format(args.load_state_dict))
checkpoint = torch.load(args.load_state_dict)
net1.load_state_dict(checkpoint['net1_state_dict'])
net2.load_state_dict(checkpoint['net2_state_dict'])
optimizer1.load_state_dict(checkpoint['optimizer1'])
optimizer2.load_state_dict(checkpoint['optimizer2'])
start_epoch = checkpoint['epoch']+1
# test current state
test_loader = get_loader(p, 'test', meta_info)
acc = scanmix_test(start_epoch-1,net1,net2,test_loader, device=device)
print('\nEpoch:%d Accuracy:%.2f\n'%(start_epoch-1,acc))
test_log.write('Epoch:%d Accuracy:%.2f\n'%(start_epoch-1,acc))
test_log.flush()
else:
start_epoch = 0
all_loss = [[],[]] # save the history of losses from two networks
for epoch in range(start_epoch, p['num_epochs']+1):
lr=p['lr']
if epoch >= 150:
lr /= 10
for param_group in optimizer1.param_groups:
param_group['lr'] = lr
for param_group in optimizer2.param_groups:
param_group['lr'] = lr
test_loader = get_loader(p, 'test', meta_info)
eval_loader = get_loader(p, 'eval_train', meta_info)
if epoch<p['warmup']:
warmup_trainloader = get_loader(p, 'warmup', meta_info)
print('Warmup Net1')
scanmix_warmup(epoch,net1,optimizer1,warmup_trainloader, CEloss, conf_penalty, args.noise_mode, device=device)
print('\nWarmup Net2')
scanmix_warmup(epoch,net2,optimizer2,warmup_trainloader, CEloss, conf_penalty, args.noise_mode, device=device)
if epoch == p['warmup']-1:
prob1,_,_=scanmix_eval_train(args,net1,[], epoch, eval_loader, CE, device=device)
prob2,_,_=scanmix_eval_train(args,net2,[], epoch, eval_loader, CE, device=device)
pred1 = (prob1 > p['p_threshold'])
pred2 = (prob2 > p['p_threshold'])
noise1 = len((1-pred1).nonzero()[0])/len(pred1)
noise2 = len((1-pred2).nonzero()[0])/len(pred2)
predicted_noise = (noise1 + noise2) / 2
print('\nPREDICTED NOISE RATE: {}'.format(predicted_noise))
if predicted_noise <= 0.6:
args.lr_sl = 0.00001
p['augmentation_strategy'] = 'dividemix'
else:
args.lr_sl = 0.001
p['augmentation_strategy'] = 'ours'
else:
prob1,all_loss[0],pl_1=scanmix_eval_train(args,net1,all_loss[0], epoch, eval_loader, CE, device=device)
prob2,all_loss[1],pl_2=scanmix_eval_train(args,net2,all_loss[1], epoch, eval_loader, CE, device=device)
pred1 = (prob1 > p['p_threshold'])
pred2 = (prob2 > p['p_threshold'])
print('[DM] Train Net1')
meta_info['probability'] = prob2
meta_info['pred'] = pred2
labeled_trainloader, unlabeled_trainloader = get_loader(p, 'train', meta_info)
scanmix_train(p, epoch,net1,net2,optimizer1,labeled_trainloader, unlabeled_trainloader, criterion_dm, args.lambda_u, device=device) # train net1
print('\n[DM] Train Net2')
meta_info['probability'] = prob1
meta_info['pred'] = pred1
labeled_trainloader, unlabeled_trainloader = get_loader(p, 'train', meta_info)
scanmix_train(p, epoch,net2,net1,optimizer2,labeled_trainloader, unlabeled_trainloader, criterion_dm, args.lambda_u, device=device) # train net2
if not args.dividemix_only:
for param_group in optimizer1.param_groups:
param_group['lr'] = args.lr_sl
for param_group in optimizer2.param_groups:
param_group['lr'] = args.lr_sl
meta_info['predicted_labels'] = pl_2
neighbor_dataloader = get_loader(p, 'neighbors', meta_info)
print('\n[SL] Train Net1')
scanmix_scan(neighbor_dataloader, net1, criterion_sl, optimizer1, epoch, device)
meta_info['predicted_labels'] = pl_1
neighbor_dataloader = get_loader(p, 'neighbors', meta_info)
print('\n[SL] Train Net2')
scanmix_scan(neighbor_dataloader, net2, criterion_sl, optimizer2, epoch, device)
acc = scanmix_test(epoch,net1,net2,test_loader, device=device)
print('\nEpoch:%d Accuracy:%.2f\n'%(epoch,acc))
test_log.write('Epoch:%d Accuracy:%.2f\n'%(epoch,acc))
test_log.flush()
torch.save({
'net1_state_dict': net1.state_dict(),
'net2_state_dict': net2.state_dict(),
'epoch': epoch,
'optimizer1': optimizer1.state_dict(),
'optimizer2': optimizer2.state_dict(),
}, os.path.join(p['scanmix_dir'], 'savedDicts/checkpoint.json'))
if __name__ == "__main__":
main()