def build_embedding_graph(self):
self.x = tf.placeholder(tf.float32, shape=[None, None, None, self.channels], name="X")
self.y = tf.placeholder(tf.float32, shape=[None, None, None, self.channels], name="Y")
# H-1 conv
self.Wm1_conv = util.weight([self.cnn_size, self.cnn_size, self.channels, self.feature_num],
stddev=self.weight_dev, name="W-1_conv", initializer=self.initializer)
self.Bm1_conv = util.bias([self.feature_num], name="B-1")
Hm1_conv = util.conv2d_with_bias(self.x, self.Wm1_conv, self.cnn_stride, self.Bm1_conv, add_relu=True, name="H-1")
# H0 conv
self.W0_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="W0_conv", initializer=self.initializer)
self.B0_conv = util.bias([self.feature_num], name="B0")
self.H_conv[0] = util.conv2d_with_bias(Hm1_conv, self.W0_conv, self.cnn_stride, self.B0_conv, add_relu=True,
name="H0")
if self.summary:
# convert to tf.summary.image format [batch_num, height, width, channels]
Wm1_transposed = tf.transpose(self.Wm1_conv, [3, 0, 1, 2])
tf.summary.image("W-1/" + self.model_name, Wm1_transposed, max_outputs=self.log_weight_image_num)
util.add_summaries("B-1", self.model_name, self.Bm1_conv, mean=True, max=True, min=True)
util.add_summaries("W-1", self.model_name, self.Wm1_conv, mean=True, max=True, min=True)
util.add_summaries("B0", self.model_name, self.B0_conv, mean=True, max=True, min=True)
util.add_summaries("W0", self.model_name, self.W0_conv, mean=True, max=True, min=True)
def build_reconstruction_graph(self):
# HD+1 conv
self.WD1_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="WD1_conv", initializer=self.initializer)
self.BD1_conv = util.bias([self.feature_num], name="BD1")
# HD+2 conv
self.WD2_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num+1, self.channels],
stddev=self.weight_dev, name="WD2_conv", initializer=self.initializer)
self.BD2_conv = util.bias([1], name="BD2")
self.Y1_conv = (self.inference_depth) * [None]
self.Y2_conv = (self.inference_depth) * [None]
self.W = tf.Variable(np.full(fill_value=1.0 / self.inference_depth, shape=[self.inference_depth],
dtype=np.float32),name="LayerWeights") # 设置递归层随机变量权重
W_sum = tf.reduce_sum(self.W) # 压缩求和 降维 计算所有递归层权重元素的和
self.y_outputs = self.inference_depth * [None]
for i in range(0, self.inference_depth):
with tf.variable_scope("Y%d" % (i+1)):
self.Y1_conv[i] = util.conv2d_with_bias(self.H_conv[i+1], self.WD1_conv, self.cnn_stride, self.BD1_conv,
add_relu=not self.residual, name="conv_1") # 每个Hd卷积结果输入HD+1中做卷积得到d个结果
y_conv = tf.concat([self.Y1_conv[i], self.x], 3) # 将上面结果与输入X相加得到d个
self.Y2_conv[i] = util.conv2d_with_bias(y_conv, self.WD2_conv, self.cnn_stride, self.BD2_conv,
add_relu=not self.residual, name="conv_2") # 相加结果再输入HD+2中卷积得到d个结果
self.y_outputs[i] = self.Y2_conv[i] * self.W[i] / W_sum # 平均 每个递归层的输出
if self.summary:
util.add_summaries("BD1", self.model_name, self.BD1_conv)
util.add_summaries("WD1", self.model_name, self.WD1_conv, mean=True, max=True, min=True)
util.add_summaries("WD2", self.model_name, self.WD2_conv, mean=True, max=True, min=True)
def build_embedding_graph(self):
self.x = tf.placeholder(tf.float32, shape=[None, None, None, self.channels], name="X") #输入LR图像
self.y = tf.placeholder(tf.float32, shape=[None, None, None, self.channels], name="Y") # HR图像
# H-1 conv
with tf.variable_scope("W-1_conv"):
self.Wm1_conv = util.weight([self.cnn_size, self.cnn_size, self.channels, self.feature_num],
stddev=self.weight_dev, name="conv_W", initializer=self.initializer)
self.Bm1_conv = util.bias([self.feature_num], name="conv_B")
Hm1_conv = util.conv2d_with_bias(self.x, self.Wm1_conv, self.cnn_stride, self.Bm1_conv, add_relu=True, name="H")
# H0 conv
with tf.variable_scope("W0_conv"):
self.W0_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="conv_W", initializer=self.initializer)
self.B0_conv = util.bias([self.feature_num], name="conv_B")
self.H_conv[0] = util.conv2d_with_bias(Hm1_conv, self.W0_conv, self.cnn_stride, self.B0_conv, add_relu=True,name="H")
if self.summary:
Wm1_transposed = tf.transpose(self.Wm1_conv, [3, 0, 1, 2])
#把Wm2_conv格式 [3,3,1,96]转换为tf.summary.image格式 [96,3,3,1][batch_num, height, width, channels]
tf.summary.image("W-1/" + self.model_name, Wm1_transposed, max_outputs=self.log_weight_image_num)
# 为image增加一个summary，这样就能在tensorboard上看到图片了。img=tf.summary.image('input',x,batch_size)
util.add_summaries("B-1", self.model_name, self.Bm1_conv, mean=True, max=True, min=True)
util.add_summaries("W-1", self.model_name, self.Wm1_conv, mean=True, max=True, min=True)
util.add_summaries("B0", self.model_name, self.B0_conv, mean=True, max=True, min=True)
util.add_summaries("W0", self.model_name, self.W0_conv, mean=True, max=True, min=True)
def build_reconstruction_graph(self):
# HD+1 conv
self.WD1_conv = util.weight(
[self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev,
name="WD1_conv",
initializer=self.initializer)
self.BD1_conv = util.bias([self.feature_num], name="BD1")
# HD+2 conv
self.WD2_conv = util.weight(
[self.cnn_size, self.cnn_size, self.feature_num, self.channels],
stddev=self.weight_dev,
name="WD2_conv",
initializer=self.initializer)
self.BD2_conv = util.bias([1], name="BD2")
self.Y1_conv = (self.inference_depth + 1) * [None]
self.Y2_conv = (self.inference_depth + 1) * [None]
self.W = tf.Variable(np.full(fill_value=1.0 /
(self.inference_depth + 1),
shape=[self.inference_depth + 1],
dtype=np.float32),
name="layer_weight")
W_sum = tf.reduce_sum(self.W)
for i in range(0, self.inference_depth + 1):
self.Y1_conv[i] = util.conv2d_with_bias_and_relu(self.H_conv[i],
self.WD1_conv,
self.cnn_stride,
self.BD1_conv,
name="Y%d_1" % i)
self.Y2_conv[i] = util.conv2d_with_bias_and_relu(self.Y1_conv[i],
self.WD2_conv,
self.cnn_stride,
self.BD2_conv,
name="Y%d_2" % i)
y_ = tf.mul(self.W[i], self.Y2_conv[i], name="Y%d_mul" % i)
y_ = tf.div(y_, W_sum, name="Y%d_div" % i)
if i == 0:
self.y_ = y_
else: #skip-connection for layer greater than 0
self.y_ = self.y_ + y_
if self.summary:
util.add_summaries("BD1:" + self.model_name, self.BD1_conv)
util.add_summaries("WD1:" + self.model_name,
self.WD1_conv,
mean=True,
max=True,
min=True)
util.add_summaries("WD2:" + self.model_name,
self.WD2_conv,
mean=True,
max=True,
min=True)
def build_merge_graph(self):
# M-2
self.Wmer2_conv = util.weight([3, 1, 1, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="Wm-2_conv", initializer=self.initializer)
self.Bmer2_conv = util.bias([self.feature_num], name="Bm-2")
self.Mer2_conv = util.conv3d_with_bias(self.net_feature.outputs, self.Wmer2_conv, self.cnn_stride, self.Bmer2_conv, pad="VALID", add_relu=True, name="M-2")
# M-1
self.Wmer1_conv = util.weight([3, 1, 1, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="Wm-1_conv", initializer=self.initializer)
self.Bmer1_conv = util.bias([self.feature_num], name="Bm-1")
self.H_conv[0] = util.conv3d_with_bias(self.Mer2_conv, self.Wmer2_conv, self.cnn_stride, self.Bmer2_conv, pad="VALID", add_relu=True, name="M-1")
self.H_conv[0] = self.H_conv[0][:,0,:,:,:]
def build_inference_graph(self):
if self.inference_depth <= 0:
return
self.W_conv = util.weight(
[self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev,
name="W_conv",
initializer="diagonal")
self.B_conv = util.bias([self.feature_num], name="B")
for i in range(0, self.inference_depth):
self.H_conv[i + 1] = util.conv2d_with_bias(self.H_conv[i],
self.W_conv,
1,
self.B_conv,
name="H%d" % (i + 1))
if self.summary:
util.add_summaries("W",
self.model_name,
self.W_conv,
mean=True,
max=True,
min=True)
util.add_summaries("B",
self.model_name,
self.B_conv,
mean=True,
max=True,
min=True)
def build_inference_graph(self):
if self.inference_depth <= 0:
return
self.WL_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num, self.feature_num],
stddev=self.weight_dev, name="WL_conv", initializer="diagonal")
self.BL_conv = util.bias([self.feature_num], name="BL")
self.WS_conv = util.weight([self.cnn_size, self.cnn_size, 1, self.feature_num],
stddev=self.weight_dev, name="WS_conv", initializer="diagonal")
self.BS_conv = util.bias([self.feature_num], name="BS")
for i in range(0, self.inference_depth):
self.H_conv[i + 1] = util.conv2d_with_bias(self.H_conv[i], self.WL_conv, 1, self.BL_conv, add_relu=True,
name="H%d" % (i + 1))
self.HS_conv[i] = util.conv2d_with_bias(self.net_residual.outputs[:,i,:,:,:], self.WS_conv, 1, self.BS_conv, add_relu=True,
name="HS%d" % (i + 1))
self.H_conv[i + 1] = tf.add(self.H_conv[i + 1], self.HS_conv[i])
# tf.summary.image("Feature_map%d/" % (i+1) + self.model_name, self.R_conv, max_outputs=4)
self.W_conv = util.weight([self.cnn_size, self.cnn_size, self.feature_num, self.channels],
stddev=self.weight_dev, name="W_conv", initializer=self.initializer)
self.B_conv = util.bias([self.channels], name="B")
self.H = util.conv2d_with_bias(self.H_conv[self.inference_depth], self.W_conv, 1, self.B_conv, add_relu=True,
name="H")
self.y_ = self.H
tf.summary.image("prediction/" + self.model_name, self.y_, max_outputs=1)
if self.residual:
self.y_ = tf.add(self.y_, self.net_image.outputs, name="output")
if self.summary:
util.add_summaries("W", self.model_name, self.W_conv, mean=True, max=True, min=True)
util.add_summaries("B", self.model_name, self.B_conv, mean=True, max=True, min=True)
util.add_summaries("BD1", self.model_name, self.BD1_conv)
util.add_summaries("WD1", self.model_name, self.WD1_conv, mean=True, max=True, min=True)
util.add_summaries("WD2", self.model_name, self.WD2_conv, mean=True, max=True, min=True)
def build_new_layer(self):
self.x = tf.placeholder(tf.float32,
shape=[None, None, None, self.channels],
name="X1")
self.y = tf.placeholder(tf.float32,
shape=[None, None, None, self.channels],
name="Y1")
# H-1 conv
with tf.variable_scope("W-3_conv"):
self.Wm1_conv = util.weight([
self.cnn_size, self.cnn_size, self.channels, self.feature_num
],
stddev=self.weight_dev,
name="conv_W1",
initializer=self.initializer)
self.Bm1_conv = util.bias([self.feature_num], name="conv_B1")
Hm1_conv = util.conv2d_with_bias(self.x,
self.Wm1_conv,
self.cnn_stride,
self.Bm1_conv,
add_relu=True,
name="H1")
# H0 conv
with tf.variable_scope("W-2_conv"):
self.W0_conv = util.weight([
self.cnn_size, self.cnn_size, self.feature_num,
self.feature_num
],
stddev=self.weight_dev,
name="conv_W1",
initializer=self.initializer)
self.B0_conv = util.bias([self.feature_num], name="conv_B1")
self.H_conv[0] = util.conv2d_with_bias(Hm1_conv,
self.W0_conv,
self.cnn_stride,
self.B0_conv,
add_relu=True,
name="H1")
if self.summary:
# convert to tf.summary.image format [batch_num, height, width, channels]
Wm1_transposed = tf.transpose(self.Wm1_conv, [3, 0, 1, 2])
tf.summary.image("W-3/" + self.model_name,
Wm1_transposed,
max_outputs=self.log_weight_image_num)
util.add_summaries("B-3",
self.model_name,
self.Bm1_conv,
mean=True,
max=True,
min=True)
util.add_summaries("W-3",
self.model_name,
self.Wm1_conv,
mean=True,
max=True,
min=True)
util.add_summaries("B2",
self.model_name,
self.B0_conv,
mean=True,
max=True,
min=True)
util.add_summaries("W2",
self.model_name,
self.W0_conv,
mean=True,
max=True,
min=True)
