forked from CoinCheung/mixMatch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
205 lines (175 loc) · 5.37 KB
/
train.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
import random
import time
import numpy as np
import torch
import torch.nn as nn
from model import WideResnet
from cifar import get_train_loader, get_val_loader, OneHot
from label_guessor import LabelGuessor
from mixup import MixUp
from loss import CrossEntropyLoss
from ema import EMA
## some hyper-parameters are borrowed from the official repository
wresnet_k = 2
wresnet_n = 28
n_classes = 10
n_workers = 0
lr = 0.002
n_epoches = 1024
batchsize = 64
n_imgs_per_epoch = 64 * 1024
n_guesses = 2
temperature = 0.5
mixup_alpha = 0.75
lam_u = 75
ema_alpha = 0.999
weight_decay = 0.02
## settings
torch.multiprocessing.set_sharing_strategy('file_system')
# torch.manual_seed(123)
# random.seed(123)
# np.random.seed(123)
# torch.backends.cudnn.deterministic = True
def set_model():
model = WideResnet(n_classes, k=wresnet_k, n=wresnet_n) # wide resnet-28
model.train()
model.cuda()
criteria_x = CrossEntropyLoss().cuda()
criteria_u = nn.MSELoss().cuda()
return model, criteria_x, criteria_u
def train_one_epoch(
model,
criteria_x,
criteria_u,
optim,
ema,
wd,
dltrain_x,
dltrain_u,
lb_guessor,
mixuper,
lambda_u,
lambda_u_once,
):
n_iters_per_epoch = n_imgs_per_epoch // batchsize
one_hot = OneHot(n_classes)
dl_x, dl_u = iter(dltrain_x), iter(dltrain_u)
loss_avg, loss_x_avg, loss_u_avg = [], [], []
st = time.time()
for it in range(n_iters_per_epoch):
ims_x, lbs_x = next(dl_x)
ims_u, _ = next(dl_u)
with torch.no_grad():
ims_x, lbs_x = ims_x[0].cuda(), one_hot(lbs_x).cuda()
ims_u = [im.cuda() for im in ims_u]
lbs_u = lb_guessor(model, ims_u).cuda()
ims = torch.cat([ims_x]+ims_u, dim=0)
lbs = torch.cat([lbs_x]+[lbs_u for _ in range(n_guesses)], dim=0)
ims, lbs = mixuper(ims, lbs)
optim.zero_grad()
logits = model(ims)
logits_x = logits[:batchsize]
lbs_x = lbs[:batchsize]
logits_u = logits[batchsize:]
preds_u = torch.softmax(logits_u, dim=1)
lbs_u = lbs[batchsize:]
loss_x = criteria_x(logits_x, lbs_x)
loss_u = criteria_u(preds_u, lbs_u)
lam_u = lambda_u + lambda_u_once * it
loss = loss_x + lam_u * loss_u
loss.backward()
optim.step()
do_weight_decay(model, wd)
ema.update_params()
loss_avg.append(loss.item())
loss_x_avg.append(loss_x.item())
loss_u_avg.append(loss_u.item())
if (it+1) % 512 == 0:
ed = time.time()
t = ed -st
loss_avg = sum(loss_avg) / len(loss_avg)
loss_x_avg = sum(loss_x_avg) / len(loss_x_avg)
loss_u_avg = sum(loss_u_avg) / len(loss_u_avg)
msg = ', '.join([
'iter: {}',
'loss_avg: {:.4f}',
'loss_u: {:.4f}',
'loss_x: {:.4f}',
'lam_u: {:.4f}',
'time: {:.2f}',
]).format(
it+1, loss_avg, loss_u, loss_x, lam_u, t
)
loss_avg, loss_x_avg, loss_u_avg = [], [], []
st = ed
print(msg)
ema.update_buffer()
def evaluate(ema):
ema.apply_shadow()
ema.model.eval()
ema.model.cuda()
dlval = get_val_loader(
batch_size=128, num_workers=n_workers, root='cifar10'
)
matches = []
for ims, lbs in dlval:
ims = ims[0].cuda()
lbs = lbs.cuda()
with torch.no_grad():
logits = ema.model(ims)
scores = torch.softmax(logits, dim=1)
_, preds = torch.max(scores, dim=1)
match = lbs == preds
matches.append(match)
matches = torch.cat(matches, dim=0).float()
acc = torch.mean(matches)
ema.restore()
return acc
@torch.no_grad()
def do_weight_decay(model, decay):
for param in model.parameters():
param.copy_(param * decay)
def train():
model, criteria_x, criteria_u = set_model()
n_iters_per_epoch = n_imgs_per_epoch // batchsize
dltrain_x, dltrain_u = get_train_loader(
batchsize, n_iters_per_epoch, L=250, K=n_guesses
)
lb_guessor = LabelGuessor(model, T=temperature)
mixuper = MixUp(mixup_alpha)
ema = EMA(model, ema_alpha)
optim = torch.optim.Adam(model.parameters(), lr=lr)
n_iters_per_epoch = n_imgs_per_epoch // batchsize
lam_u_epoch = float(lam_u) / n_epoches
lam_u_once = lam_u_epoch / n_iters_per_epoch
train_args = dict(
model=model,
criteria_x=criteria_x,
criteria_u=criteria_u,
optim=optim,
ema=ema,
wd = 1 - weight_decay * lr,
dltrain_x=dltrain_x,
dltrain_u=dltrain_u,
lb_guessor=lb_guessor,
mixuper=mixuper,
lambda_u=0,
lambda_u_once=lam_u_once,
)
best_acc = -1
print('start to train')
for e in range(n_epoches):
model.train()
print('epoch: {}'.format(e))
train_args['lambda_u'] = e * lam_u_epoch
train_one_epoch(**train_args)
torch.cuda.empty_cache()
acc = evaluate(ema)
best_acc = acc if best_acc < acc else best_acc
log_msg = [
'epoch: {}'.format(e),
'acc: {:.4f}'.format(acc),
'best_acc: {:.4f}'.format(best_acc)]
print(', '.join(log_msg))
if __name__ == '__main__':
train()