-
Notifications
You must be signed in to change notification settings - Fork 6
/
train.py
273 lines (222 loc) · 9.24 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
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
import os
import sys
import shutil
import numpy as np
import time, datetime
import torch
import random
import logging
import argparse
import torch.nn as nn
import torch.utils
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.utils.data.distributed
sys.path.append("../../")
from utils import *
from torchvision import datasets, transforms
from torch.autograd import Variable
from birealnet import birealnet18
import torchvision.models as models
from modules import Binarize
parser = argparse.ArgumentParser("birealnet18")
parser.add_argument('--batch_size', type=int, default=512, help='batch size')
parser.add_argument('--epochs', type=int, default=256, help='num of training epochs')
parser.add_argument('--bit-num', type=int, default=5, help='bits per kernel')
parser.add_argument('--learning_rate', type=float, default=0.001, help='init learning rate')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
parser.add_argument('--weight_decay', type=float, default=0, help='weight decay')
parser.add_argument('--save', type=str, default='./models', help='path for saving trained models')
parser.add_argument('--data', metavar='DIR', help='path to dataset')
parser.add_argument('--label_smooth', type=float, default=0.1, help='label smoothing')
parser.add_argument('--teacher', type=str, default='resnet34', help='path of ImageNet')
parser.add_argument('-j', '--workers', default=40, type=int, metavar='N',
help='number of data loading workers (default: 4)')
args = parser.parse_args()
CLASSES = 1000
if not os.path.exists('log'):
os.mkdir('log')
log_format = '%(asctime)s %(message)s'
logging.basicConfig(stream=sys.stdout, level=logging.INFO,
format=log_format, datefmt='%m/%d %I:%M:%S %p')
fh = logging.FileHandler(os.path.join('log/log.txt'))
fh.setFormatter(logging.Formatter(log_format))
logging.getLogger().addHandler(fh)
def main():
if not torch.cuda.is_available():
sys.exit(1)
start_t = time.time()
cudnn.benchmark = True
cudnn.enabled=True
logging.info("args = %s", args)
# load model
model_teacher = models.__dict__[args.teacher](pretrained=True)
model_teacher = nn.DataParallel(model_teacher).cuda()
for p in model_teacher.parameters():
p.requires_grad = False
model_teacher.eval()
model_student = birealnet18(bit_num=args.bit_num)
logging.info('student:')
logging.info(model_student)
model_student = nn.DataParallel(model_student).cuda()
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
criterion_smooth = CrossEntropyLabelSmooth(CLASSES, args.label_smooth)
criterion_smooth = criterion_smooth.cuda()
criterion_kd = DistributionLoss()
all_parameters = model_student.parameters()
weight_parameters = []
for pname, p in model_student.named_parameters():
if p.ndimension() == 4 or 'conv' in pname:
weight_parameters.append(p)
weight_parameters_id = list(map(id, weight_parameters))
other_parameters = list(filter(lambda p: id(p) not in weight_parameters_id, all_parameters))
optimizer = torch.optim.Adam(
[{'params' : other_parameters},
{'params' : weight_parameters, 'weight_decay' : args.weight_decay}],
lr=args.learning_rate,)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lambda step : (1.0-step/args.epochs), last_epoch=-1)
start_epoch = 0
best_top1_acc= 0
checkpoint_tar = os.path.join(args.save, 'checkpoint_ba.pth.tar')
checkpoint = torch.load(checkpoint_tar)
model_student.load_state_dict(checkpoint['state_dict'], strict=False)
# For resume
checkpoint_tar = os.path.join(args.save, 'checkpoint.pth.tar')
if os.path.exists(checkpoint_tar):
logging.info('loading checkpoint {} ..........'.format(checkpoint_tar))
checkpoint = torch.load(checkpoint_tar)
start_epoch = checkpoint['epoch']
best_top1_acc = checkpoint['best_top1_acc']
model_student.load_state_dict(checkpoint['state_dict'], strict=False)
logging.info("loaded checkpoint {} epoch = {}" .format(checkpoint_tar, checkpoint['epoch']))
# adjust the learning rate according to the checkpoint
for epoch in range(start_epoch):
scheduler.step()
# load training data
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# data augmentation
crop_scale = 0.08
lighting_param = 0.1
train_transforms = transforms.Compose([
transforms.RandomResizedCrop(224, scale=(crop_scale, 1.0)),
Lighting(lighting_param),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize])
train_dataset = datasets.ImageFolder(
traindir,
transform=train_transforms)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
# load validation data
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(valdir, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
# train the model
epoch = start_epoch
while epoch < args.epochs:
train_obj, train_top1_acc, train_top5_acc = train(epoch, train_loader, model_student, model_teacher, criterion_kd, optimizer, scheduler)
valid_obj, valid_top1_acc, valid_top5_acc = validate(epoch, val_loader, model_student, criterion, args)
is_best = False
if valid_top1_acc > best_top1_acc:
best_top1_acc = valid_top1_acc
is_best = True
save_checkpoint({
'epoch': epoch,
'state_dict': model_student.state_dict(),
'best_top1_acc': best_top1_acc,
'optimizer' : optimizer.state_dict(),
}, is_best, args.save)
epoch += 1
training_time = (time.time() - start_t) / 3600
print('total training time = {} hours'.format(training_time))
def train(epoch, train_loader, model_student, model_teacher, criterion, optimizer, scheduler):
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
progress = ProgressMeter(
len(train_loader),
[batch_time, data_time, losses, top1, top5],
prefix="Epoch: [{}]".format(epoch))
model_student.train()
model_teacher.eval()
end = time.time()
scheduler.step()
for param_group in optimizer.param_groups:
cur_lr = param_group['lr']
print('learning_rate:', cur_lr)
for i, (images, target) in enumerate(train_loader):
data_time.update(time.time() - end)
images = images.cuda()
target = target.cuda()
# compute outputy
logits_student = model_student(images)
logits_teacher = model_teacher(images)
loss = criterion(logits_student, logits_teacher)
# measure accuracy and record loss
prec1, prec5 = accuracy(logits_student, target, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n) # accumulated loss
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
progress.display(i)
#if i % 100 != 0:
# continue
#for name, module in model_student.named_modules():
# if isinstance(module, Binarize):
# module.update = True
return losses.avg, top1.avg, top5.avg
def validate(epoch, val_loader, model, criterion, args):
batch_time = AverageMeter('Time', ':6.3f')
losses = AverageMeter('Loss', ':.4e')
top1 = AverageMeter('Acc@1', ':6.2f')
top5 = AverageMeter('Acc@5', ':6.2f')
progress = ProgressMeter(
len(val_loader),
[batch_time, losses, top1, top5],
prefix='Test: ')
# switch to evaluation mode
model.eval()
with torch.no_grad():
end = time.time()
for i, (images, target) in enumerate(val_loader):
images = images.cuda()
target = target.cuda()
# compute output
logits = model(images)
loss = criterion(logits, target)
# measure accuracy and record loss
pred1, pred5 = accuracy(logits, target, topk=(1, 5))
n = images.size(0)
losses.update(loss.item(), n)
top1.update(pred1[0], n)
top5.update(pred5[0], n)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
progress.display(i)
print(' * acc@1 {top1.avg:.3f} acc@5 {top5.avg:.3f}'
.format(top1=top1, top5=top5))
return losses.avg, top1.avg, top5.avg
if __name__ == '__main__':
main()