一、网络总结构
(1)Decoder 代码
class Decoder(object):
def __init__(self, model_pre_path):
self.weight_vars = []
self.model_pre_path = model_pre_path
with tf.variable_scope('decoder'):
self.weight_vars.append(self._create_variables(64, 64, 3, scope='conv2_1'))
self.weight_vars.append(self._create_variables(64, 32, 3, scope='conv2_2'))
self.weight_vars.append(self._create_variables(32, 16, 3, scope='conv2_3'))
self.weight_vars.append(self._create_variables(16, 1 , 3, scope='conv2_4'))
二、训练时的框架
三、各个框架的卷积结构
四、提出的融合策略
(1)加法策略
(2)l1-norm and soft-max 策略