-
Notifications
You must be signed in to change notification settings - Fork 37
/
train_oneflow.py
154 lines (127 loc) · 4.42 KB
/
train_oneflow.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
import oneflow as flow
import argparse
import numpy as np
import os
import time
from models.vgg import vgg16, vgg19, vgg16_bn, vgg19_bn
from utils.ofrecord_data_utils import OFRecordDataLoader
model_dict = {
"vgg16": vgg16,
"vgg19": vgg19,
"vgg16_bn": vgg16_bn,
"vgg19_bn": vgg19_bn,
}
def _parse_args():
parser = argparse.ArgumentParser("flags for train vgg")
parser.add_argument(
"--save_checkpoint_path",
type=str,
default="./checkpoints",
help="save checkpoint root dir",
)
parser.add_argument(
"--load_checkpoint", type=str, default="", help="load checkpoint"
)
parser.add_argument(
"--ofrecord_path", type=str, default="./ofrecord", help="dataset path"
)
# training hyper-parameters
parser.add_argument(
"--learning_rate", type=float, default=0.001, help="learning rate"
)
parser.add_argument("--mom", type=float, default=0.9, help="momentum")
parser.add_argument("--epochs", type=int, default=1000, help="training epochs")
parser.add_argument(
"--train_batch_size", type=int, default=32, help="train batch size"
)
parser.add_argument("--val_batch_size", type=int, default=32, help="val batch size")
parser.add_argument(
"--model",
type=str,
default="vgg16",
help="choose a model from vgg16, vgg16_bn, vgg19, vgg19_bn",
)
return parser.parse_args()
def main(args):
train_data_loader = OFRecordDataLoader(
ofrecord_root=args.ofrecord_path,
mode="train",
dataset_size=9469, # NOTE(Liang Depeng): needs to explictly set the dataset size
batch_size=args.train_batch_size,
)
val_data_loader = OFRecordDataLoader(
ofrecord_root=args.ofrecord_path,
mode="val",
dataset_size=3925,
batch_size=args.val_batch_size,
)
# oneflow init
start_t = time.time()
vgg_module = model_dict[args.model]()
if args.load_checkpoint != "":
vgg_module.load_state_dict(flow.load(args.load_checkpoint))
end_t = time.time()
print("init time : {}".format(end_t - start_t))
of_cross_entropy = flow.nn.CrossEntropyLoss()
vgg_module.to("cuda")
of_cross_entropy.to("cuda")
of_sgd = flow.optim.SGD(
vgg_module.parameters(), lr=args.learning_rate, momentum=args.mom
)
of_losses = []
all_samples = len(val_data_loader) * args.val_batch_size
print_interval = 50
for epoch in range(args.epochs):
vgg_module.train()
for b in range(len(train_data_loader)):
image, label = train_data_loader.get_batch()
# oneflow train
start_t = time.time()
image = image.to("cuda")
label = label.to("cuda")
logits = vgg_module(image)
loss = of_cross_entropy(logits, label)
loss.backward()
of_sgd.step()
of_sgd.zero_grad()
end_t = time.time()
if b % print_interval == 0:
l = loss.numpy()
of_losses.append(l)
print(
"epoch {} train iter {} oneflow loss {}, train time : {}".format(
epoch, b, l, end_t - start_t
)
)
print("epoch %d train done, start validation" % epoch)
vgg_module.eval()
correct_of = 0.0
for b in range(len(val_data_loader)):
image, label = val_data_loader.get_batch()
start_t = time.time()
image = image.to("cuda")
with flow.no_grad():
logits = vgg_module(image)
predictions = logits.softmax()
of_predictions = predictions.numpy()
clsidxs = np.argmax(of_predictions, axis=1)
label_nd = label.numpy()
for i in range(args.val_batch_size):
if clsidxs[i] == label_nd[i]:
correct_of += 1
end_t = time.time()
print("epoch %d, oneflow top1 val acc: %f" % (epoch, correct_of / all_samples))
flow.save(
vgg_module.state_dict(),
os.path.join(
args.save_checkpoint_path,
"epoch_%d_val_acc_%f" % (epoch, correct_of / all_samples),
),
)
writer = open("of_losses.txt", "w")
for o in of_losses:
writer.write("%f\n" % o)
writer.close()
if __name__ == "__main__":
args = _parse_args()
main(args)