-
Notifications
You must be signed in to change notification settings - Fork 8
/
train_cifar10.py
202 lines (166 loc) · 8.36 KB
/
train_cifar10.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
import matplotlib
matplotlib.use('Agg')
import argparse, time, logging
import numpy as np
import mxnet as mx
from mxnet import gluon, nd
from mxnet import autograd as ag
from mxnet.gluon import nn
from mxnet.gluon.data.vision import transforms
import gluoncv as gcv
gcv.utils.check_version('0.6.0')
from cifarresnet import get_model
from gluoncv.utils import makedirs, TrainingHistory
from gluoncv.data import transforms as gcv_transforms
# CLI
def parse_args():
parser = argparse.ArgumentParser(description='Train a model for image classification.')
parser.add_argument('--batch-size', type=int, default=32,
help='training batch size per device (CPU/GPU).')
parser.add_argument('--num-gpus', type=int, default=0,
help='number of gpus to use.')
parser.add_argument('--model', type=str, default='resnet',
help='model to use. options are resnet and wrn. default is resnet.')
parser.add_argument('--attention', type=str, choices=['eca', 'gct', 'se', 'se-v1', 'se-v2', 'se-v3', None],
default=None, help='channel attention type to use in residual blocks.')
parser.add_argument('-j', '--num-data-workers', dest='num_workers', default=4, type=int,
help='number of preprocessing workers')
parser.add_argument('--num-epochs', type=int, default=3,
help='number of training epochs.')
parser.add_argument('--random-crop', action='store_true',
help='whether use random crop augmentation during training. default is false.')
parser.add_argument('--lr', type=float, default=0.1,
help='learning rate. default is 0.1.')
parser.add_argument('--optimizer', type=str, default='nag',
help='optimizer type, default is nag.')
parser.add_argument('--momentum', type=float, default=0.9,
help='momentum value for optimizer, default is 0.9.')
parser.add_argument('--wd', type=float, default=0.0001,
help='weight decay rate. default is 0.0001.')
parser.add_argument('--lr-decay', type=float, default=0.1,
help='decay rate of learning rate. default is 0.1.')
parser.add_argument('--lr-decay-period', type=int, default=0,
help='period in epoch for learning rate decays. default is 0 (has no effect).')
parser.add_argument('--lr-decay-epoch', type=str, default='40,60',
help='epochs at which learning rate decays. default is 40,60.')
parser.add_argument('--drop-rate', type=float, default=0.0,
help='dropout rate for wide resnet. default is 0.')
parser.add_argument('--mode', type=str,
help='mode in which to train the model. options are imperative, hybrid')
parser.add_argument('--save-period', type=int, default=10,
help='period in epoch of model saving.')
parser.add_argument('--save-dir', type=str, default='params',
help='directory of saved models')
parser.add_argument('--resume-from', type=str,
help='resume training from the model')
parser.add_argument('--save-plot-dir', type=str, default='.',
help='the path to save the history plot')
opt = parser.parse_args()
return opt
def main():
opt = parse_args()
batch_size = opt.batch_size
classes = 10
num_gpus = opt.num_gpus
batch_size *= max(1, num_gpus)
context = [mx.gpu(i) for i in range(num_gpus)] if num_gpus > 0 else [mx.cpu()]
num_workers = opt.num_workers
lr_decay = opt.lr_decay
lr_decay_epoch = [int(i) for i in opt.lr_decay_epoch.split(',')] + [np.inf]
model_name = opt.model
kwargs = {'classes': classes, 'attention': opt.attention}
net = get_model(model_name, **kwargs)
if opt.resume_from:
net.load_parameters(opt.resume_from, ctx = context)
optimizer_args = {'learning_rate': opt.lr, 'wd': opt.wd}
if opt.optimizer == 'adam':
optimizer = mx.optimizer.Adam(**optimizer_args)
elif opt.optimizer == 'nag':
optimizer = mx.optimizer.NAG(momentum=opt.momentum, **optimizer_args)
else:
optimizer = mx.optimizer.SGD(momentum=opt.momentum, **optimizer_args)
save_period = opt.save_period
if opt.save_dir and save_period:
save_dir = opt.save_dir
makedirs(save_dir)
else:
save_dir = ''
save_period = 0
plot_path = opt.save_plot_dir
logging.basicConfig(level=logging.INFO)
logging.info(opt)
transforms_list_test = [
transforms.ToTensor(),
transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010])
]
transforms_list_train = [gcv_transforms.RandomCrop(32, pad=4)] if opt.random_crop else []
transforms_list_train += [transforms.RandomFlipLeftRight()] + transforms_list_test
transform_train = transforms.Compose(transforms_list_train)
transform_test = transforms.Compose(transforms_list_test)
def test(ctx, val_data):
metric = mx.metric.Accuracy()
for i, batch in enumerate(val_data):
data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0)
label = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0)
outputs = [net(X) for X in data]
metric.update(label, outputs)
return metric.get()
def train(epochs, ctx):
if isinstance(ctx, mx.Context):
ctx = [ctx]
net.initialize(mx.init.Xavier(), ctx=ctx)
train_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(train=True).transform_first(transform_train),
batch_size=batch_size, shuffle=True, last_batch='discard', num_workers=num_workers)
val_data = gluon.data.DataLoader(
gluon.data.vision.CIFAR10(train=False).transform_first(transform_test),
batch_size=batch_size, shuffle=False, num_workers=num_workers)
trainer = gluon.Trainer(net.collect_params(), optimizer)
metric = mx.metric.Accuracy()
train_metric = mx.metric.Accuracy()
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
train_history = TrainingHistory(['training-error', 'validation-error'])
iteration = 0
lr_decay_count = 0
best_val_score = 0
for epoch in range(epochs):
tic = time.time()
train_metric.reset()
metric.reset()
train_loss = 0
num_batch = len(train_data)
if epoch == lr_decay_epoch[lr_decay_count]:
trainer.set_learning_rate(trainer.learning_rate*lr_decay)
lr_decay_count += 1
for i, batch in enumerate(train_data):
data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0)
label = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0)
with ag.record():
output = [net(X) for X in data]
loss = [loss_fn(yhat, y) for yhat, y in zip(output, label)]
for l in loss:
l.backward()
trainer.step(batch_size)
train_loss += sum([l.sum().asscalar() for l in loss])
train_metric.update(label, output)
name, acc = train_metric.get()
iteration += 1
train_loss /= batch_size * num_batch
name, acc = train_metric.get()
name, val_acc = test(ctx, val_data)
train_history.update([1-acc, 1-val_acc])
train_history.plot(save_path='%s/%s_history.png'%(plot_path, model_name))
if val_acc > best_val_score:
best_val_score = val_acc
net.save_parameters('%s/%.4f-cifar-%s-%d-best.params'%(save_dir, best_val_score, model_name, epoch))
logging.info('[Epoch %d] train=%f val=%f loss=%f time: %f' %
(epoch, acc, val_acc, train_loss, time.time()-tic))
if save_period and save_dir and (epoch + 1) % save_period == 0:
net.save_parameters('%s/cifar10-%s-%d.params'%(save_dir, model_name, epoch))
if save_period and save_dir:
net.save_parameters('%s/cifar10-%s-%d.params'%(save_dir, model_name, epochs-1))
if opt.mode == 'hybrid':
net.hybridize()
train(opt.num_epochs, context)
if __name__ == '__main__':
main()