-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_clothing1m_dividemix.py
304 lines (253 loc) · 12.3 KB
/
train_clothing1m_dividemix.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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
from __future__ import print_function
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 torchvision
import torchvision.models as models
import random
import os
import time
import argparse
import numpy as np
from dividemix_clothing1m_loader import clothing_dataloader
from sklearn.mixture import GaussianMixture
from networks.resnet import resnet50
parser = argparse.ArgumentParser(description='PyTorch Clothing1M Training')
parser.add_argument('--use_noisy_val', action='store_true', default=False, help='Using the noisy validation setting. By default, using the benchmark setting.')
parser.add_argument('--batch_size', default=32, type=int, help='train batchsize')
parser.add_argument('--lr', '--learning_rate', default=0.002, type=float, help='initial learning rate')
parser.add_argument('--alpha', default=0.5, type=float, help='parameter for Beta')
parser.add_argument('--lambda_u', default=0, type=float, help='weight for unsupervised loss')
parser.add_argument('--p_threshold', default=0.5, type=float, help='clean probability threshold')
parser.add_argument('--T', default=0.5, type=float, help='sharpening temperature')
parser.add_argument('--num_epochs', default=80, type=int)
parser.add_argument('--id', default='clothing1m')
parser.add_argument('--data_path', default='data/Clothing1M', type=str, help='path to dataset')
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--gpuid', default=0, type=int)
parser.add_argument('--num_class', default=14, type=int)
parser.add_argument('--num_batches', default=1000, type=int)
args = parser.parse_args()
torch.cuda.set_device(args.gpuid)
random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
def log(path, str):
print(str)
with open(path, 'a') as file:
file.write(str)
exp_name = 'dividemix_official_clothing1m_batch{}_seed{}'.format(args.batch_size, args.seed)
if args.use_noisy_val:
exp_name = 'nv_'+exp_name
logpath = '{}.txt'.format(exp_name)
log(logpath, 'Settings: {}\n'.format(args))
if not os.path.exists(exp_name):
os.makedirs(exp_name)
# Training
def train(epoch,net,net2,optimizer,labeled_trainloader,unlabeled_trainloader):
net.train()
net2.eval() #fix one network and train the other
unlabeled_train_iter = iter(unlabeled_trainloader)
num_iter = (len(labeled_trainloader.dataset)//args.batch_size)+1
for batch_idx, (inputs_x, inputs_x2, labels_x, w_x) in enumerate(labeled_trainloader):
try:
inputs_u, inputs_u2 = unlabeled_train_iter.next()
except:
unlabeled_train_iter = iter(unlabeled_trainloader)
inputs_u, inputs_u2 = unlabeled_train_iter.next()
batch_size = inputs_x.size(0)
# Transform label to one-hot
labels_x = torch.zeros(batch_size, args.num_class).scatter_(1, labels_x.view(-1,1), 1)
w_x = w_x.view(-1,1).type(torch.FloatTensor)
inputs_x, inputs_x2, labels_x, w_x = inputs_x.cuda(), inputs_x2.cuda(), labels_x.cuda(), w_x.cuda()
inputs_u, inputs_u2 = inputs_u.cuda(), inputs_u2.cuda()
with torch.no_grad():
# label co-guessing of unlabeled samples
outputs_u11 = net(inputs_u)
outputs_u12 = net(inputs_u2)
outputs_u21 = net2(inputs_u)
outputs_u22 = net2(inputs_u2)
pu = (torch.softmax(outputs_u11, dim=1) + torch.softmax(outputs_u12, dim=1) + torch.softmax(outputs_u21, dim=1) + torch.softmax(outputs_u22, dim=1)) / 4
ptu = pu**(1/args.T) # temparature sharpening
targets_u = ptu / ptu.sum(dim=1, keepdim=True) # normalize
targets_u = targets_u.detach()
# label refinement of labeled samples
outputs_x = net(inputs_x)
outputs_x2 = net(inputs_x2)
px = (torch.softmax(outputs_x, dim=1) + torch.softmax(outputs_x2, dim=1)) / 2
px = w_x*labels_x + (1-w_x)*px
ptx = px**(1/args.T) # temparature sharpening
targets_x = ptx / ptx.sum(dim=1, keepdim=True) # normalize
targets_x = targets_x.detach()
# mixmatch
l = np.random.beta(args.alpha, args.alpha)
l = max(l, 1-l)
all_inputs = torch.cat([inputs_x, inputs_x2, inputs_u, inputs_u2], dim=0)
all_targets = torch.cat([targets_x, targets_x, targets_u, targets_u], dim=0)
idx = torch.randperm(all_inputs.size(0))
input_a, input_b = all_inputs, all_inputs[idx]
target_a, target_b = all_targets, all_targets[idx]
mixed_input = l * input_a[:batch_size*2] + (1 - l) * input_b[:batch_size*2]
mixed_target = l * target_a[:batch_size*2] + (1 - l) * target_b[:batch_size*2]
logits = net(mixed_input)
Lx = -torch.mean(torch.sum(F.log_softmax(logits, dim=1) * mixed_target, dim=1))
# regularization
prior = torch.ones(args.num_class)/args.num_class
prior = prior.cuda()
pred_mean = torch.softmax(logits, dim=1).mean(0)
penalty = torch.sum(prior*torch.log(prior/pred_mean))
loss = Lx + penalty
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# sys.stdout.write('\r')
# sys.stdout.write('Clothing1M | Epoch [%3d/%3d] Iter[%3d/%3d]\t Labeled loss: %.4f '
# %(epoch, args.num_epochs, batch_idx+1, num_iter, Lx.item()))
# sys.stdout.flush()
def warmup(net,optimizer,dataloader):
net.train()
for batch_idx, (inputs, labels, path) in enumerate(dataloader):
inputs, labels = inputs.cuda(), labels.cuda()
optimizer.zero_grad()
outputs = net(inputs)
loss = CEloss(outputs, labels)
penalty = conf_penalty(outputs)
L = loss + penalty
L.backward()
optimizer.step()
# sys.stdout.write('\r')
# sys.stdout.write('|Warm-up: Iter[%3d/%3d]\t CE-loss: %.4f Conf-Penalty: %.4f'
# %(batch_idx+1, args.num_batches, loss.item(), penalty.item()))
# sys.stdout.flush()
def val(net,val_loader,epoch,k):
net.eval()
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(val_loader):
inputs, targets = inputs.cuda(), targets.cuda()
outputs = net(inputs)
_, predicted = torch.max(outputs, 1)
total += targets.size(0)
correct += predicted.eq(targets).cpu().sum().item()
acc = 100.*correct/total
#print("\n| Validation\t Net%d Acc: %.2f%%" %(k,acc))
update = False
if acc > best_acc[k-1]:
update = True
best_acc[k-1] = acc
best_epoch[k-1] = epoch
#save_point = './checkpoint/%s_net%d.pth.tar'%(args.id,k)
torch.save(net.state_dict(), '{}/net{}_best.pth'.format(exp_name, k))
torch.save(net.state_dict(), '{}/net{}_epoch{}.pth'.format(exp_name, k, epoch))
return acc, update
def test(net1,net2,test_loader):
net1.eval()
net2.eval()
correct1 = 0
correct2 = 0
correct = 0
total = 0
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(test_loader):
inputs, targets = inputs.cuda(), targets.cuda()
outputs1 = net1(inputs)
outputs2 = net2(inputs)
outputs = outputs1+outputs2 #""" use the ensemble of two models ??? """
_, predicted1 = torch.max(outputs1, 1)
_, predicted2 = torch.max(outputs2, 1)
_, predicted = torch.max(outputs, 1)
total += targets.size(0)
correct1 += predicted1.eq(targets).cpu().sum().item()
correct2 += predicted2.eq(targets).cpu().sum().item()
correct += predicted.eq(targets).cpu().sum().item()
return 100.*correct1/total, 100.*correct2/total, 100.*correct/total
def eval_train(epoch,model):
model.eval()
num_samples = args.num_batches*args.batch_size
losses = torch.zeros(num_samples)
paths = []
n=0
with torch.no_grad():
for batch_idx, (inputs, targets, path) in enumerate(eval_loader):
inputs, targets = inputs.cuda(), targets.cuda()
outputs = model(inputs)
loss = CE(outputs, targets)
for b in range(inputs.size(0)):
losses[n]=loss[b]
paths.append(path[b])
n+=1
# sys.stdout.write('\r')
# sys.stdout.write('| Evaluating loss Iter %3d\t' %(batch_idx))
# sys.stdout.flush()
losses = (losses-losses.min())/(losses.max()-losses.min())
losses = losses.reshape(-1,1)
gmm = GaussianMixture(n_components=2,max_iter=10,reg_covar=5e-4,tol=1e-2)
gmm.fit(losses)
prob = gmm.predict_proba(losses)
prob = prob[:,gmm.means_.argmin()]
return prob,paths
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 = resnet50(pretrained=True)
model.fc = nn.Linear(2048,args.num_class)
model = model.cuda()
return model
# log=open('./checkpoint/%s.txt'%args.id,'w')
# log.flush()
loader = clothing_dataloader(root=args.data_path,batch_size=args.batch_size,num_workers=5,num_batches=args.num_batches, use_noisy_val=args.use_noisy_val)
# print('| Building net')
net1 = create_model()
net2 = create_model()
test_net1 = create_model()
test_net2 = create_model()
cudnn.benchmark = True
optimizer1 = optim.SGD(net1.parameters(), lr=args.lr, momentum=0.9, weight_decay=1e-3)
optimizer2 = optim.SGD(net2.parameters(), lr=args.lr, momentum=0.9, weight_decay=1e-3)
CE = nn.CrossEntropyLoss(reduction='none')
CEloss = nn.CrossEntropyLoss()
conf_penalty = NegEntropy()
best_acc = [0,0]
best_epoch = [0,0]
for epoch in range(args.num_epochs+1):
t0 = time.time()
lr=args.lr
if epoch >= 40:
lr /= 10
for param_group in optimizer1.param_groups:
param_group['lr'] = lr
for param_group in optimizer2.param_groups:
param_group['lr'] = lr
if epoch<1: # warm up
train_loader = loader.run('warmup')
warmup(net1,optimizer1,train_loader)
train_loader = loader.run('warmup')
warmup(net2,optimizer2,train_loader)
else:
pred1 = (prob1 > args.p_threshold) # divide dataset
pred2 = (prob2 > args.p_threshold)
labeled_trainloader, unlabeled_trainloader = loader.run('train',pred2,prob2,paths=paths2) # co-divide
train(epoch,net1,net2,optimizer1,labeled_trainloader, unlabeled_trainloader) # train net1
labeled_trainloader, unlabeled_trainloader = loader.run('train',pred1,prob1,paths=paths1) # co-divide
train(epoch,net2,net1,optimizer2,labeled_trainloader, unlabeled_trainloader) # train net2
val_loader = loader.run('val') # validation
val_acc1, update1 = val(net1,val_loader,epoch,1)
val_acc2, update2 = val(net2,val_loader,epoch,2)
if update1 or update2:
test_loader = loader.run('test')
test_net1.load_state_dict(torch.load('{}/net1_best.pth'.format(exp_name)))
test_net2.load_state_dict(torch.load('{}/net2_best.pth'.format(exp_name)))
test_acc1, test_acc2, test_acc = test(test_net1,test_net2,test_loader)
eval_loader = loader.run('eval_train') # evaluate training data loss for next epoch
prob1,paths1 = eval_train(epoch,net1)
eval_loader = loader.run('eval_train')
prob2,paths2 = eval_train(epoch,net2)
log(logpath, 'Epoch: {}/{}, Time: {:.1f}s. '.format(epoch, args.num_epochs, time.time()-t0))
log(logpath, 'Val1: {:.2f}%, Val2: {:.2f}%, Epoch_best: net1 {} net2 {}; Test1: {:.2f}%, Test2: {:.2f}%; Test_ensemble: {:.2f}%.\n'.format(
val_acc1, val_acc2, best_epoch[0], best_epoch[1], test_acc1, test_acc2, test_acc))