-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathtrain_cifar.py
143 lines (114 loc) · 6.25 KB
/
train_cifar.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
import os
import pickle
import numpy as np
import torch
from sklearn.mixture import GaussianMixture
from train import warmup, train
def save_losses(input_loss, exp):
name = './stats/cifar100/losses{}.pcl'
nm = name.format(exp)
if os.path.exists(nm):
loss_history = pickle.load(open(nm, "rb"))
else:
loss_history, clean_history = [], []
loss_history.append(input_loss)
pickle.dump(loss_history, open(nm, "wb"))
def eval_train(model, eval_loader, CE, all_loss, epoch, net, device, r, stats_log):
model.eval()
losses = torch.zeros(50000)
losses_clean = torch.zeros(50000)
with torch.no_grad():
for batch_idx, (inputs, _, targets, index, targets_clean) in enumerate(eval_loader):
inputs, targets, targets_clean = inputs.to(device), targets.to(device), targets_clean.to(device)
outputs = model(inputs)
loss = CE(outputs, targets)
clean_loss = CE(outputs, targets_clean)
for b in range(inputs.size(0)):
losses[index[b]] = loss[b]
losses_clean[index[b]] = clean_loss[b]
losses = (losses - losses.min()) / (losses.max() - losses.min())
all_loss.append(losses)
history = torch.stack(all_loss)
if r >= 0.9: # average loss over last 5 epochs to improve convergence stability
input_loss = history[-5:].mean(0)
input_loss = input_loss.reshape(-1, 1)
else:
input_loss = losses.reshape(-1, 1)
# exp = '_std_tpc_oracle'
# save_losses(input_loss, exp)
gmm = GaussianMixture(n_components=2, max_iter=200, tol=1e-2, reg_covar=5e-4)
gmm.fit(input_loss)
clean_idx, noisy_idx = gmm.means_.argmin(), gmm.means_.argmax()
stats_log.write('Epoch {} (net {}): GMM results: {} with weight {}\t'
'{} with weight {}\n'.format(epoch, net, gmm.means_[clean_idx], gmm.weights_[clean_idx],
gmm.means_[noisy_idx], gmm.weights_[noisy_idx]))
stats_log.flush()
prob = gmm.predict_proba(input_loss)
prob = prob[:, clean_idx]
return prob, all_loss, losses_clean
def run_test(epoch, net1, net2, test_loader, device, test_log):
net1.eval()
net2.eval()
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(test_loader):
inputs, targets = inputs.to(device), targets.to(device)
outputs1 = net1(inputs)
outputs2 = net2(inputs)
outputs = outputs1 + outputs2
_, predicted = torch.max(outputs, 1)
total += targets.size(0)
correct += predicted.eq(targets).cpu().sum().item()
acc = 100. * correct / total
print("\n| Test Epoch #%d\t Accuracy: %.2f%%\n" % (epoch, acc))
test_log.write('Epoch:%d Accuracy:%.2f\n' % (epoch, acc))
test_log.flush()
def run_train_loop(net1, optimizer1, sched1, net2, optimizer2, sched2, criterion, CEloss, CE, loader, p_threshold,
warm_up, num_epochs, all_loss, batch_size, num_class, device, lambda_u, T, alpha, noise_mode,
dataset, r, conf_penalty, stats_log, loss_log, test_log):
for epoch in range(num_epochs + 1):
test_loader = loader.run('test')
eval_loader = loader.run('eval_train')
if epoch < warm_up:
warmup_trainloader = loader.run('warmup')
print('Warmup Net1')
warmup(epoch, net1, optimizer1, warmup_trainloader, CEloss, conf_penalty, device, dataset, r, num_epochs,
noise_mode)
print('\nWarmup Net2')
warmup(epoch, net2, optimizer2, warmup_trainloader, CEloss, conf_penalty, device, dataset, r, num_epochs,
noise_mode)
prob1, all_loss[0], losses_clean1 = eval_train(net1, eval_loader, CE, all_loss[0], epoch, 1, device, r,
stats_log)
prob2, all_loss[1], losses_clean2 = eval_train(net2, eval_loader, CE, all_loss[1], epoch, 2, device, r,
stats_log)
p_thr2 = np.clip(p_threshold, prob2.min() + 1e-5, prob2.max() - 1e-5)
pred2 = prob2 > p_thr2
loss_log.write('{},{},{},{},{}\n'.format(epoch, losses_clean2[pred2].mean(), losses_clean2[pred2].std(),
losses_clean2[~pred2].mean(), losses_clean2[~pred2].std()))
loss_log.flush()
loader.run('train', pred2, prob2) # count metrics
else:
print('Train Net1')
prob2, all_loss[1], losses_clean2 = eval_train(net2, eval_loader, CE, all_loss[1], epoch, 2, device, r,
stats_log)
p_thr2 = np.clip(p_threshold, prob2.min() + 1e-5, prob2.max() - 1e-5)
pred2 = prob2 > p_thr2
loss_log.write('{},{},{},{},{}\n'.format(epoch, losses_clean2[pred2].mean(), losses_clean2[pred2].std(),
losses_clean2[~pred2].mean(), losses_clean2[~pred2].std()))
loss_log.flush()
labeled_trainloader, unlabeled_trainloader = loader.run('train', pred2, prob2) # co-divide
train(epoch, net1, net2, criterion, optimizer1, labeled_trainloader, unlabeled_trainloader, lambda_u,
batch_size, num_class, device, T, alpha, warm_up, dataset, r, noise_mode, num_epochs) # train net1
print('\nTrain Net2')
prob1, all_loss[0], losses_clean1 = eval_train(net1, eval_loader, CE, all_loss[0], epoch, 1, device, r,
stats_log)
p_thr1 = np.clip(p_threshold, prob1.min() + 1e-5, prob1.max() - 1e-5)
pred1 = prob1 > p_thr1
labeled_trainloader, unlabeled_trainloader = loader.run('train', pred1, prob1) # co-divide
train(epoch, net2, net1, criterion, optimizer2, labeled_trainloader, unlabeled_trainloader, lambda_u,
batch_size, num_class, device, T, alpha, warm_up, dataset, r, noise_mode, num_epochs) # train net2
run_test(epoch, net1, net2, test_loader, device, test_log)
sched1.step()
sched2.step()
torch.save(net1.state_dict(), './final_checkpoints/final_checkpoint.pth.tar')