def dense_decoder_fusion4(self, decoder5_feature, encoder4, name=None):#1/16, 1/16
with tf.variable_scope(name):
output_channel = 512
decoder5_feature = Conv_BN_Relu(decoder5_feature, [3, 3, 512, output_channel], "dec5")
encoder4 = Conv_BN_Relu(encoder4, [3, 3, 512, output_channel], "enc4")
output = tf.add(decoder5_feature, encoder4)
return output
def dense_decoder_fusion2(self,
decoder5_feature,
decoder4_feature,
decoder3_feature,
encoder2,
name=None): #1/16, 1/8, 1/4, 1/4
with tf.variable_scope(name):
output_channel = 128
decoder5_feature = Conv_BN_Relu(decoder5_feature,
[3, 3, 512, output_channel],
"dec5")
decoder5_feature = Upsampling(decoder5_feature,
tf.shape(encoder2),
ksize=8,
stride=4,
output_channel=output_channel,
name="up_dec5")
decoder4_feature = Conv_BN_Relu(decoder4_feature,
[3, 3, 256, output_channel],
"dec4")
decoder4_feature = Upsampling(decoder4_feature,
tf.shape(encoder2),
ksize=4,
stride=2,
output_channel=output_channel,
name="up_dec4")
decoder3_feature = Conv_BN_Relu(decoder3_feature,
[3, 3, 128, output_channel],
"dec3")
encoder2 = Conv_BN_Relu(encoder2, [3, 3, 128, output_channel],
"enc2")
output = tf.add(decoder5_feature, decoder4_feature)
output = tf.add(output, decoder3_feature)
output = tf.add(output, encoder2)
return output
def dense_decoder_fusion1(self,
decoder2_feature,
encoder1,
name=None): #1/2, 1/2
with tf.variable_scope(name):
output_channel = 64
decoder2_feature = Conv_BN_Relu(decoder2_feature,
[3, 3, 64, output_channel], "dec2")
encoder1 = Conv_BN_Relu(encoder1, [3, 3, 64, output_channel],
"enc1")
output = tf.add(decoder2_feature, encoder1)
return output
def dense_decoder_fusion2(self,
decoder3_feature,
encoder2,
name=None): #1/4, 1/4
with tf.variable_scope(name):
output_channel = 128
decoder3_feature = Conv_BN_Relu(decoder3_feature,
[3, 3, 128, output_channel],
"dec3")
encoder2 = Conv_BN_Relu(encoder2, [3, 3, 128, output_channel],
"enc2")
output = tf.add(decoder3_feature, encoder2)
return output
def dense_decoder_fusion3(self,
decoder4_feature,
encoder3,
name=None): #1/8, 1/8
with tf.variable_scope(name):
output_channel = 256
decoder4_feature = Conv_BN_Relu(decoder4_feature,
[3, 3, 256, output_channel],
"dec4")
encoder3 = Conv_BN_Relu(encoder3, [3, 3, 256, output_channel],
"enc3")
output = tf.add(decoder4_feature, encoder3)
return output
def build(self, input, input_channel, class_num=2):
self.pool_size = [1,2,2,1]
self.pool_stride = [1,2,2,1]
# encoder
# encoder 1
self.conv1_1 = Conv_BN_Relu(input, [3, 3, input_channel, 64], "conv1_1")
self.conv1_2 = Conv_BN_Relu(self.conv1_1, [3, 3, 64, 64], "conv1_2")
self.pooling1 = tf.nn.max_pool(self.conv1_2, self.pool_size, self.pool_stride, padding="SAME", name="pooling1")
# encoder 2
self.conv2_1 = Conv_BN_Relu(self.pooling1, [3, 3, 64, 128], "conv2_1")
self.conv2_2 = Conv_BN_Relu(self.conv2_1, [3, 3, 128, 128], "conv2_2")
self.pooling2 = tf.nn.max_pool(self.conv2_2, self.pool_size, self.pool_stride, padding="SAME", name="pooling2")
# encoder 3
self.conv3_1 = Conv_BN_Relu(self.pooling2, [3, 3, 128, 256], "conv3_1")
self.conv3_2 = Conv_BN_Relu(self.conv3_1, [3, 3, 256, 256], "conv3_2")
self.conv3_3 = Conv_BN_Relu(self.conv3_2, [3, 3, 256, 256], "conv3_3")
self.pooling3 = tf.nn.max_pool(self.conv3_3, self.pool_size, self.pool_stride, padding="SAME", name="pooling3")
# encoder 4
self.conv4_1 = Conv_BN_Relu(self.pooling3, [3, 3, 256, 512], "conv4_1")
self.conv4_2 = Conv_BN_Relu(self.conv4_1, [3, 3, 512, 512], "conv4_2")
self.conv4_3 = Conv_BN_Relu(self.conv4_2, [3, 3, 512, 512], "conv4_3")
self.pooling4 = tf.nn.max_pool(self.conv4_3, self.pool_size, self.pool_stride, padding="SAME", name="pooling4")
# encoder 5
self.conv5_1 = Conv_BN_Relu(self.pooling4, [3, 3, 512, 512], "conv5_1")
self.conv5_2 = Conv_BN_Relu(self.conv5_1, [3, 3, 512, 512], "conv5_2")
self.conv5_3 = Conv_BN_Relu(self.conv5_2, [3, 3, 512, 512], "conv5_3")
self.pooling5 = tf.nn.max_pool(self.conv5_3, self.pool_size, self.pool_stride, padding="SAME", name="pooling5")
# decoder
# decoder 5
self.upsample5 = Upsampling(self.pooling5, tf.shape(self.pooling4), ksize=4, stride=2, output_channel=512, name="upsample5")
self.deconv5_3 = Conv_BN_Relu(self.upsample5, [3, 3, 512, 512], "deconv5_3")
self.deconv5_2 = Conv_BN_Relu(self.deconv5_3, [3, 3, 512, 512], "deconv5_2")
self.deconv5_1 = Conv_BN_Relu(self.deconv5_2, [3, 3, 512, 512], "deconv5_1")
# decoder 4
self.upsample4 = Upsampling(self.deconv5_1, tf.shape(self.pooling3), ksize=4, stride=2, output_channel=512, name="upsample4")
self.deconv4_3 = Conv_BN_Relu(self.upsample4, [3, 3, 512, 512], "deconv4_3")
self.deconv4_2 = Conv_BN_Relu(self.deconv4_3, [3, 3, 512, 512], "deconv4_2")
self.deconv4_1 = Conv_BN_Relu(self.deconv4_2, [3, 3, 512, 256], "deconv4_1")
# decoder 3
self.upsample3 = Upsampling(self.deconv4_1, tf.shape(self.pooling2), ksize=4, stride=2, output_channel=256, name="upsample3")
self.deconv3_3 = Conv_BN_Relu(self.upsample3, [3, 3, 256, 256], "deconv3_3")
self.deconv3_2 = Conv_BN_Relu(self.deconv3_3, [3, 3, 256, 256], "deconv3_2")
self.deconv3_1 = Conv_BN_Relu(self.deconv3_2, [3, 3, 256, 128], "deconv3_1")
# decoder 2
self.upsample2 = Upsampling(self.deconv3_1, tf.shape(self.pooling1), ksize=4, stride=2, output_channel=128, name="upsample2")
self.deconv2_2 = Conv_BN_Relu(self.upsample2, [3, 3, 128, 128], "deconv2_2")
self.deconv2_1 = Conv_BN_Relu(self.deconv2_2, [3, 3, 128, 64], "deconv2_1")
# decoder 1
self.upsample1 = Upsampling(self.deconv2_1, tf.shape(input), ksize=4, stride=2, output_channel=64, name="upsample1")
self.deconv1_2 = Conv_BN_Relu(self.upsample1, [3, 3, 64, 64], "deconv1_2")
self.deconv1_1 = Conv_BN_Relu(self.deconv1_2, [3, 3, 64, class_num], "deconv1_1")
# deconv1_1 will be processed by softmax and compute loss
return self.deconv1_1
def build(self, input, input_channel, class_num=2):
self.pool_size = [1, 2, 2, 1]
self.pool_stride = [1, 2, 2, 1]
# encoder
# encoder 1
self.conv1_1 = Conv_BN_Relu(input, [3, 3, input_channel, 64],
"conv1_1")
self.conv1_2 = Conv_BN_Relu(self.conv1_1, [3, 3, 64, 64], "conv1_2")
self.pooling1 = tf.nn.max_pool(self.conv1_2,
self.pool_size,
self.pool_stride,
padding="SAME",
name="pooling1")
# encoder 2
self.conv2_1 = Conv_BN_Relu(self.pooling1, [3, 3, 64, 128], "conv2_1")
self.conv2_2 = Conv_BN_Relu(self.conv2_1, [3, 3, 128, 128], "conv2_2")
self.pooling2 = tf.nn.max_pool(self.conv2_2,
self.pool_size,
self.pool_stride,
padding="SAME",
name="pooling2")
# encoder 3
self.inception_block3_1 = _inception_layer(
'inception3_1',
self.pooling2,
in_c=128,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output channel 256
self.inception_block3_2 = _inception_layer(
'inception3_2',
self.inception_block3_1,
in_c=256,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output channel 256
self.inception_block3_3 = _inception_layer(
'inception3_3',
self.inception_block3_2,
in_c=256,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output channel 256
self.pooling3 = tf.nn.max_pool(self.inception_block3_3,
self.pool_size,
self.pool_stride,
padding="SAME",
name="pooling3")
# encoder 4
self.inception_block4_1 = _inception_layer('inception4_1',
self.pooling3,
in_c=256,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.inception_block4_2 = _inception_layer('inception4_2',
self.inception_block4_1,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.inception_block4_3 = _inception_layer('inception4_3',
self.inception_block4_2,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.pooling4 = tf.nn.max_pool(self.inception_block4_3,
self.pool_size,
self.pool_stride,
padding="SAME",
name="pooling4")
# encoder 5
self.inception_block5_1 = _inception_layer('inception5_1',
self.pooling4,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.inception_block5_2 = _inception_layer('inception5_2',
self.inception_block5_1,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.inception_block5_3 = _inception_layer('inception5_3',
self.inception_block5_2,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.pooling5 = tf.nn.max_pool(self.inception_block5_3,
self.pool_size,
self.pool_stride,
padding="SAME",
name="pooling5")
# decoder
# decoder 5
self.upsample5 = Upsampling(self.pooling5,
tf.shape(self.pooling4),
ksize=4,
stride=2,
output_channel=512,
name="upsample5")
self.d_inception_block5_1 = _inception_layer('d_inception5_1',
self.upsample5,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.d_inception_block5_2 = _inception_layer('d_inception5_2',
self.d_inception_block5_1,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.d_inception_block5_3 = _inception_layer('d_inception5_3',
self.d_inception_block5_2,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
# decoder 4
self.upsample4 = Upsampling(self.d_inception_block5_3,
tf.shape(self.pooling3),
ksize=4,
stride=2,
output_channel=512,
name="upsample4")
self.d_inception_block4_1 = _inception_layer('d_inception4_1',
self.upsample4,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.d_inception_block4_2 = _inception_layer('d_inception4_2',
self.d_inception_block4_1,
in_c=512,
out_1=128,
out_21=128,
out_22=128,
out_31=128,
out_32=256,
out_33=128,
out_41=128,
out_42=256,
out_43=128) # output 512
self.d_inception_block4_3 = _inception_layer('d_inception4_3',
self.d_inception_block4_2,
in_c=512,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output 256
# decoder 3
self.upsample3 = Upsampling(self.d_inception_block4_3,
tf.shape(self.pooling2),
ksize=4,
stride=2,
output_channel=256,
name="upsample3")
self.d_inception_block3_1 = _inception_layer(
'd_inception3_1',
self.upsample3,
in_c=256,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output channel 256
self.d_inception_block3_2 = _inception_layer(
'd_inception3_2',
self.d_inception_block3_1,
in_c=256,
out_1=64,
out_21=64,
out_22=64,
out_31=64,
out_32=128,
out_33=64,
out_41=64,
out_42=128,
out_43=64) # output channel 256
self.d_inception_block3_3 = _inception_layer(
'd_inception3_3',
self.d_inception_block3_2,
in_c=256,
out_1=32,
out_21=64,
out_22=32,
out_31=64,
out_32=128,
out_33=32,
out_41=64,
out_42=128,
out_43=32) # output channel 128
# decoder 2
self.upsample2 = Upsampling(self.d_inception_block3_3,
tf.shape(self.pooling1),
ksize=4,
stride=2,
output_channel=128,
name="upsample2")
self.deconv2_2 = Conv_BN_Relu(self.upsample2, [3, 3, 128, 128],
"deconv2_2")
self.deconv2_1 = Conv_BN_Relu(self.deconv2_2, [3, 3, 128, 64],
"deconv2_1")
# decoder 1
self.upsample1 = Upsampling(self.deconv2_1,
tf.shape(input),
ksize=4,
stride=2,
output_channel=64,
name="upsample1")
self.deconv1_2 = Conv_BN_Relu(self.upsample1, [3, 3, 64, 64],
"deconv1_2")
self.deconv1_1 = Conv_BN_Relu(self.deconv1_2, [3, 3, 64, class_num],
"deconv1_1")
# deconv1_1 will be processed by softmax and compute loss
return self.deconv1_1
