本人尝试改变一张文章源代码的网络结构，原文是使用四个卷积层进行降噪，我想要把最后一个卷积层变成全连接层，但是我改后似乎输入流只经过了最后一层全连接层，并且输出结果损失函数也不是很好，实在是找不到问题所在，不知道怎么更改，希望各位同僚看一看，感谢万分
这是我的代码：
def build_network(self,
built_for_training=False): # build a network similar with SRCNN, a fully-convolutional netowrk for channel noise estimation
# built_for_train: denote whether the network is build for training or test. If for test, xavier initialization is not needed since the model will be loaded later.
x_in = tf.placeholder(tf.float32, [None, self.net_config.feature_length]) # input data # 576
x_in_reshape = tf.reshape(x_in, (-1, self.net_config.feature_length, 1, 1)) # HY:576
layer_output = {}
for layer in range(self.net_config.total_layers): # construct layers 4
# 给每层命名
self.conv_filter_name[layer] = format("conv_layer%d" % (layer))
self.bias_name[layer] = format("b%d" % (layer))
self.bia_name[layer] = format("b_01%d" % (layer))
self.weight_name[layer] = format("w_01%d" % (layer))
# 给每层设置输入输出维度
if layer == 0:
x_input = x_in_reshape
in_channels = 1
else:
x_input = layer_output[layer - 1]
# in_channels表示每一层的卷积核数量
in_channels = self.net_config.feature_map_nums[layer - 1] # np.array([64,32,16,1])
out_channels = self.net_config.feature_map_nums[layer] # np.array([64,32,16,1])
shape = [self.net_config.filter_sizes[layer], 1, in_channels, out_channels]
if built_for_training:
# xavier initialization for training # np.array([9,3,3,15])，另一维度为1
# get_variable 获取已经存在的名为name的变量
with tf.variable_scope("layer_%d" % layer):
self.conv_filter[layer] = tf.get_variable("w", shape,
tf.float32, tf.contrib.layers.xavier_initializer())
self.bias[layer] = tf.get_variable("b", shape[-1],
tf.float32, tf.contrib.layers.xavier_initializer())
self.bia[layer] = tf.get_variable("b_01", 576,
tf.float32, tf.contrib.layers.xavier_initializer())
self.weight[layer] = tf.get_variable("w_01", [576, 576],
tf.float32, tf.contrib.layers.xavier_initializer())
# Variable 创建变量，定义初始值
self.best_conv_filter[layer] = tf.Variable(
tf.ones([self.net_config.filter_sizes[layer], 1, in_channels, out_channels], tf.float32),
dtype=tf.float32)
self.best_bias[layer] = tf.Variable(tf.ones([out_channels], tf.float32), dtype=tf.float32)
self.best_bia[layer] = tf.Variable(tf.ones(576, tf.float32), dtype=tf.float32)
self.best_weight[layer] = tf.Variable(tf.ones([576, 576], tf.float32), dtype=tf.float32)
self.assign_best_conv_filter[layer] = self.best_conv_filter[layer].assign(self.conv_filter[layer])
self.assign_best_bias[layer] = self.best_bias[layer].assign(self.bias[layer])
self.assign_best_bia[layer] = self.best_bia[layer].assign(self.bia[layer])
self.assign_best_weight[layer] = self.best_weight[layer].assign(self.weight[layer])
else:
# just build tensors for testing and their values will be loaded later.
self.conv_filter[layer] = tf.Variable(
tf.random_normal([self.net_config.filter_sizes[layer], 1, in_channels, out_channels], 0, 1,
tf.float32), dtype=tf.float32,
name=self.conv_filter_name[layer])
self.weight[layer] = tf.Variable(tf.random_normal([576,576], 0, 1,
tf.float32), dtype=tf.float32,
name=self.weight_name[layer])
self.bias[layer] = tf.Variable(tf.random_normal([out_channels], 0, 1, tf.float32), dtype=tf.float32,
name=self.bias_name[layer])
self.bia[layer] = tf.Variable(tf.random_normal([576], 0, 1, tf.float32), dtype=tf.float32,
name=self.bia_name[layer])
# 层定义
if layer == self.net_config.total_layers - 1:
# 原代码
# layer_output[layer] = tf.nn.conv2d(x_input, self.conv_filter[layer], [1, 1, 1, 1], 'SAME') + self.bias[
# layer]
x_in = tf.reshape(x_in, [-1, 576])
layer_output[layer] = tf.nn.relu(tf.matmul(x_in, self.weight[layer]) + self.bia[layer])
else:
# Activation Function
layer_output[layer] = tf.nn.relu(
tf.nn.conv2d(x_input, self.conv_filter[layer], [1, 1, 1, 1], 'SAME') + self.bias[layer])
y_out = layer_output[self.net_config.total_layers - 1]
y_out = tf.reshape(y_out, (-1, self.net_config.label_length)) # +x_in# resnet ?,576
return x_in, y_out
# 存储训练
def save_network_temporarily(self, sess_in):
for layer in range(self.net_config.save_layers):
sess_in.run(self.assign_best_conv_filter[layer])
sess_in.run(self.assign_best_bias[layer])
sess_in.run(self.assign_best_weight[layer])
sess_in.run(self.assign_best_bia[layer])