def forward(self, inputs, batch_norm_decay=0.9, weight_decay=0.0005, isTrain=True, reuse=False):
# set batch norm params
batch_norm_params = {
'decay': batch_norm_decay,
'epsilon': 1e-05,
'scale': True,
'is_training': isTrain,
'fused': None, # Use fused batch norm if possible.
}
with slim.arg_scope([slim.conv2d, slim.batch_norm], reuse=reuse):
# darknet53 特征
# [N, 19, 19, 512], [N, 38, 38, 256], [N, 76, 76, 128]
route_1, route_2, route_3 = module.extraction_feature(inputs, batch_norm_params, weight_decay)
with slim.arg_scope([slim.conv2d],
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
biases_initializer=None,
activation_fn=lambda x: tf.nn.leaky_relu(x, alpha=0.1),
weights_regularizer=slim.l2_regularizer(weight_decay)):
with tf.variable_scope('yolo'):
# 计算 y1 特征
# [N, 76, 76, 128] => [N, 76, 76, 256]
net = module.conv(route_1, 256)
# [N, 76, 76, 256] => [N, 76, 76, 255]
net = slim.conv2d(net, 3*(4+1+self.class_num), 1,
stride=1, normalizer_fn=None,
activation_fn=None, biases_initializer=tf.zeros_initializer())
feature_y3 = net
# 计算 y2 特征
# [N, 76, 76, 128] => [N, 38, 38, 256]
net = module.conv(route_1, 256, stride=2)
# [N, 38, 38, 512]
net = tf.concat([net, route_2], -1)
net = module.yolo_conv_block(net, 512, 2, 1)
route_147 = net
# [N, 38, 38, 256] => [N, 38, 38, 512]
net = module.conv(net, 512)
# [N, 38, 38, 512] => [N, 38, 38, 255]
net = slim.conv2d(net, 3*(4+1+self.class_num), 1,
stride=1, normalizer_fn=None,
activation_fn=None, biases_initializer=tf.zeros_initializer())
feature_y2 = net
# 计算 y3 特征
# [N, 38, 38, 256] => [N, 19, 19, 512]
net = module.conv(route_147, 512, stride=2)
net = tf.concat([net, route_3], -1)
net = module.yolo_conv_block(net, 1024, 3, 0)
# [N, 19, 19, 1024] => [N, 19, 19, 255]
net = slim.conv2d(net, 3*(4+1+self.class_num), 1,
stride=1, normalizer_fn=None,
activation_fn=None, biases_initializer=tf.zeros_initializer())
feature_y1 = net
return feature_y1, feature_y2, feature_y3
def forward(self, inputs, isTrain=True):
# darknet53 特征
# [N, 19, 19, 512], [N, 38, 38, 256], [N, 76, 76, 128]
route_1, route_2, route_3 = module.darknet53(inputs, isTrain=isTrain)
# 计算 y1 特征
# [N, 76, 76, 128] => [N, 76, 76, 256]
net = module.conv(route_3, 256, isTrain=isTrain)
# [N, 76, 76, 256] => [N, 76, 76, 255]
feature_y3 = module.conv(net,
3 * (self.class_num + 4 + 1),
kernel_size=1,
bn=False,
activation=None,
isTrain=isTrain)
# 计算 y2 特征
# [N, 76, 76, 128] => [N, 38, 38, 256]
net = module.conv(route_3, 256, stride=2, isTrain=isTrain)
# [N, 38, 38, 512]
net = tf.concat([net, route_2], -1)
net = module.yolo_conv_block(net, 512, 2, 1, isTrain=isTrain)
route_147 = net
# [N, 38, 38, 256] => [N, 38, 38, 512]
net = module.conv(net, 512, isTrain=isTrain)
feature_y2 = module.conv(net,
3 * (self.class_num + 4 + 1),
kernel_size=1,
bn=False,
activation=None,
isTrain=isTrain)
# 计算 y3 特征
# [N, 38, 38, 256] => [N, 19, 19, 512]
net = module.conv(route_147, 512, stride=2, isTrain=isTrain)
net = tf.concat([net, route_1], -1)
net = module.yolo_conv_block(net, 1024, 3, 0, isTrain=isTrain)
# [N, 19, 19, 1024] => [N, 19, 19, 255]
feature_y1 = module.conv(net,
3 * (self.class_num + 4 + 1),
kernel_size=1,
bn=False,
activation=None,
isTrain=isTrain)
return feature_y1, feature_y2, feature_y3
