def build_net(self, ):
batch_norm_params = {
'is_training': self.config.is_train,
'center': True,
'scale': True,
'decay': 0.9997,
'epsilon': 1e-3,
}
self.im_input = tf.placeholder(tf.string, name='input')
im, b_scale, = self.handle_im(self.im_input, size=self.config.img_size)
im.set_shape(
tf.TensorShape(
[None, self.config.img_size, self.config.img_size, 3]))
im = im / 255 * 2 - 1
batch_m = tf.shape(im)[0]
with tf.variable_scope('FeatureExtractor'):
with tf.contrib.slim.arg_scope(
mobilenet_v2.training_scope(
is_training=self.config.is_train,
weight_decay=0.00004,
stddev=0.03,
dropout_keep_prob=0.8,
bn_decay=0.9997)):
logits, endpoints = mobilenet_v2.mobilenet(im,
num_classes=None)
net15 = endpoints['layer_15/expansion_output']
net19 = endpoints['layer_19']
print(net15, net19)
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
net20 = new_conv2d(net19, 512, 2, 2)
net21 = new_conv2d(net20, 256, 2, 3)
net22 = new_conv2d(net21, 256, 2, 4)
net23 = new_conv2d(net22, 128, 2, 5)
var_pre = tf.global_variables()[1:]
weights_init = tf.truncated_normal_initializer(stddev=0.03)
net = [net15, net19, net20, net21, net22, net23]
with slim.arg_scope([slim.conv2d],
activation_fn=None,
weights_initializer=weights_init,
weights_regularizer=slim.l2_regularizer(
self.config.weight_decay),
normalizer_fn=None):
for i in range(len(self.config.num_anchors)):
with tf.variable_scope('BoxPredictor_%d' % i):
with tf.variable_scope('ClassPredictor'):
net_cls = slim.conv2d(
net[i], self.config.num_anchors[i] *
(self.config.num_cls + 1), [1, 1])
with tf.variable_scope('BoxEncodingPredictor'):
net_box = slim.conv2d(
net[i],
self.config.num_anchors[i] * 4,
[1, 1],
)
net_cls = tf.reshape(
net_cls,
(batch_m, -1, self.config.num_cls + 1),
)
net_box = tf.reshape(
net_box,
(batch_m, -1, 4),
)
net[i] = tf.concat([net_cls, net_box], axis=-1)
net = tf.concat(net, axis=1, name='net')
pre_bboxes = loc2bbox(
net[..., -4:] *
tf.constant([0.1, 0.1, 0.2, 0.2], dtype=tf.float32), self.anchors)
pre = predict(pre_bboxes[0],
net[0][..., :-4],
size=self.config.img_size,
c_thresh=1e-2,
num_cls=self.config.num_cls)
self.result = tf.concat([pre[..., :4] * b_scale, pre[..., 4:]],
axis=-1,
name='out')
print(self.result)
def build_net(self, Iter):
batch_norm_params = {
'is_training': self.config.is_train,
'center': True,
'scale': True,
'decay': 0.9997,
'epsilon': 1e-3,
}
im, bboxes, nums = Iter.get_next()
im.set_shape(
tf.TensorShape(
[None, self.config.img_size, self.config.img_size, 3]))
im = im / 255 * 2 - 1
batch_m = tf.shape(im)[0]
with tf.variable_scope('FeatureExtractor'):
with tf.contrib.slim.arg_scope(
mobilenet_v2.training_scope(
is_training=self.config.is_train,
weight_decay=0.00004,
stddev=0.03,
dropout_keep_prob=0.8,
bn_decay=0.9997)):
logits, endpoints = mobilenet_v2.mobilenet(im,
num_classes=None)
net15 = endpoints['layer_15/expansion_output']
net19 = endpoints['layer_19']
print(net15, net19)
with slim.arg_scope([slim.batch_norm], **batch_norm_params):
net20 = new_conv2d(net19, 512, 2, 2)
net21 = new_conv2d(net20, 256, 2, 3)
net22 = new_conv2d(net21, 256, 2, 4)
net23 = new_conv2d(net22, 128, 2, 5)
var_pre = tf.global_variables()[1:]
weights_init = tf.truncated_normal_initializer(stddev=0.03)
net = [net15, net19, net20, net21, net22, net23]
with slim.arg_scope([slim.conv2d],
activation_fn=None,
weights_initializer=weights_init,
weights_regularizer=slim.l2_regularizer(
self.config.weight_decay),
normalizer_fn=None):
for i in range(len(self.config.num_anchors)):
with tf.variable_scope('BoxPredictor_%d' % i):
with tf.variable_scope('ClassPredictor'):
net_cls = slim.conv2d(
net[i], self.config.num_anchors[i] *
(self.config.num_cls + 1), [1, 1])
with tf.variable_scope('BoxEncodingPredictor'):
net_box = slim.conv2d(
net[i],
self.config.num_anchors[i] * 4,
[1, 1],
)
net_cls = tf.reshape(
net_cls,
(batch_m, -1, self.config.num_cls + 1),
)
net_box = tf.reshape(
net_box,
(batch_m, -1, 4),
)
net[i] = tf.concat([net_cls, net_box], axis=-1)
net = tf.concat(net, axis=1, name='net')
loss, Num = self.get_loss(net, bboxes, nums)
return tf.reduce_sum(loss) / tf.reduce_sum(Num + 1e-10), var_pre