-
Notifications
You must be signed in to change notification settings - Fork 24
/
shufflenetv2_centernet_V2_SEB.py
158 lines (131 loc) · 7.56 KB
/
shufflenetv2_centernet_V2_SEB.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
from layer_utils import *
import cfg
import loss
class Shufflenetv2_Centernet_SEB():
first_conv_channel = 24
def __init__(self, model_scale=1.0, shuffle_group=2):
self.shuffle_group = shuffle_group
self.channel_sizes = self._select_channel_size(model_scale)
self.inputs = tf.placeholder(shape=[None, cfg.input_image_size, cfg.input_image_size, 3], dtype=tf.float32,
name='inputs')
self.is_training = tf.placeholder(dtype=tf.bool, name='is_training')
self.cls_gt = tf.placeholder(shape=[None, cfg.featuremap_h, cfg.featuremap_w, cfg.num_classes],
dtype=tf.float32)
self.size_gt = tf.placeholder(shape=[None, cfg.featuremap_h, cfg.featuremap_w, 2], dtype=tf.float32)
with tf.variable_scope('shufflenet_centernet'):
with slim.arg_scope([slim.batch_norm], is_training=self.is_training):
self.pred_cls, self.pred_size = self._build_model()
self.build_train()
self.merged_summay = tf.summary.merge_all()
def _select_channel_size(self, model_scale):
# [(out_channel, repeat_times), (out_channel, repeat_times), ...]
if model_scale == 0.5:
return [(48, 4), (96, 8), (192, 4), (1024, 1)]
elif model_scale == 1.0:
return [(116, 4), (232, 8), (464, 4), (1024, 1)]
elif model_scale == 1.5:
return [(176, 4), (352, 8), (704, 4), (1024, 1)]
elif model_scale == 2.0:
return [(244, 4), (488, 8), (976, 4), (2048, 1)]
else:
raise ValueError('Unsupported model size.')
def semantic_embed_block(self, high_level, low_level):
with tf.variable_scope(None, "semantic_up_block"):
n, h, w, c = low_level.get_shape().as_list()
out = conv(high_level, c, 3)
out = upsample_layer(out, (h, w))
out = out * low_level
return out
def _build_model(self):
with tf.variable_scope('stage_4'):
out_2 = conv_bn_relu(self.inputs, self.first_conv_channel, 3, 2) # /2
# out_4 = slim.max_pool2d(out_2, 3, 2 , padding='SAME')#/4
out_4 = conv_bn_relu(out_2, self.first_conv_channel, 3, 2) # /4
with tf.variable_scope('stage_8'):
out_channel, repeat = self.channel_sizes[0]
# First block is downsampling
out_8 = shufflenet_v2_block(out_4, out_channel, 3, 2, shuffle_group=self.shuffle_group) # /8
for i in range(repeat - 1):
out_8 = shufflenet_v2_block(out_8, out_channel, 3, shuffle_group=self.shuffle_group)
with tf.variable_scope('stage_16'):
out_channel, repeat = self.channel_sizes[1]
# First block is downsampling
out_16 = shufflenet_v2_block(out_8, out_channel, 3, 2, shuffle_group=self.shuffle_group) # /16
for i in range(repeat - 1):
out_16 = shufflenet_v2_block(out_16, out_channel, 3, shuffle_group=self.shuffle_group)
with tf.variable_scope('stage_32'):
out_channel, repeat = self.channel_sizes[2]
# First block is downsampling
out_32 = shufflenet_v2_block(out_16, out_channel, 3, 2, shuffle_group=self.shuffle_group) # /32
for i in range(repeat - 1):
out_32 = shufflenet_v2_block(out_32, out_channel, 3, shuffle_group=self.shuffle_group)
with tf.variable_scope('semantic_embed_block'):
out_4 = self.semantic_embed_block(out_8, out_4)
out_4 = conv_bn_relu(out_4, cfg.feature_channels, 3)
out_8 = self.semantic_embed_block(out_16, out_8)
out_8 = conv_bn_relu(out_8, cfg.feature_channels, 3)
out_16 = self.semantic_embed_block(out_32, out_16)
out_16 = conv_bn_relu(out_16, cfg.feature_channels, 3)
with tf.variable_scope('feature_map_fuse'):
deconv1 = deconv_bn_relu(out_32, cfg.feature_channels)
fuse1 = deconv1 + out_16
deconv2 = deconv_bn_relu(fuse1, cfg.feature_channels)
fuse2 = out_8 + deconv2
deconv3 = deconv_bn_relu(fuse2, cfg.feature_channels)
fuse3 = out_4 + deconv3
with tf.variable_scope('detector'):
cls = conv_relu(fuse3, cfg.feature_channels, 3, 1)
cls = conv(cls, cfg.num_classes, 1, 1)
cls = tf.nn.sigmoid(cls, name='cls')
size = conv_relu(fuse3, cfg.feature_channels, 3, 1)
size = conv(size, 2, 1, 1)
size = tf.nn.relu(size, name='size')
return cls, size
def compute_loss(self):
self.cls_loss = loss.focal_loss(self.pred_cls, self.cls_gt)
self.size_loss = loss.reg_l1_loss(self.pred_size, self.size_gt)
self.total_loss = self.cls_loss + 0.1*self.size_loss
def build_train(self):
with tf.variable_scope("loss", "loss"):
self.compute_loss()
self.global_step = tf.Variable(0, trainable=False)
self.lr = cfg.lr
if cfg.lr_type == "exponential":
self.lr = tf.train.exponential_decay(cfg.lr_value,
self.global_step,
cfg.lr_decay_steps,
cfg.lr_decay_rate,
staircase=True) # staircase=True,globstep/decaystep=整数，代表lr突变的，阶梯状
elif cfg.lr_type == "fixed":
self.lr = tf.constant(cfg.lr, dtype=tf.float32)
elif cfg.lr_type == "piecewise":
self.lr = tf.train.piecewise_constant(self.global_step, cfg.lr_boundaries, cfg.lr_values)
if cfg.optimizer == 'adam':
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
elif cfg.optimizer == 'rmsprop':
self.optimizer = tf.train.RMSPropOptimizer(learning_rate=self.lr,
momentum=cfg.momentum)
elif cfg.optimizer == 'adadelta':
self.optimizer = tf.train.AdadeltaOptimizer(learning_rate=self.lr)
elif cfg.optimizer == 'momentum':
self.optimizer = tf.train.MomentumOptimizer(learning_rate=self.lr,
momentum=cfg.momentum)
elif cfg.optimizer == "sgd":
self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.lr)
elif cfg.optimizer == "ftr":
self.optimizer = tf.train.FtrlOptimizer(learning_rate=self.lr)
elif cfg.optimizer == "adagradDA":
self.optimizer = tf.train.AdagradDAOptimizer(learning_rate=self.lr, global_step=self.global_step)
elif cfg.optimizer == "adagrad":
self.optimizer = tf.train.AdagradOptimizer(learning_rate=self.lr)
elif cfg.optimizer == "ProximalAdagrad":
self.optimizer = tf.train.ProximalAdagradOptimizer(learning_rate=self.lr)
elif cfg.optimizer == "ProximalGrad":
self.optimizer = tf.train.ProximalGradientDescentOptimizer(learning_rate=self.lr)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
self.train_op = self.optimizer.minimize(self.total_loss, global_step=self.global_step)
tf.summary.scalar('total_loss', self.total_loss)
tf.summary.scalar('cls_loss', self.cls_loss)
tf.summary.scalar('size_loss', self.size_loss)
tf.summary.scalar("learning_rate", self.lr)