def generator_2(self, z):
"""
:param z: random array input dimension (batch_size, z_dim)
:return: image (Gz)
"""
z_, self.h0_w, self.h0_b = linear(z,
64 * 4 * 4 * 8,
'g_h0_lin',
with_w=True)
z_resize = tf.reshape(z_, [-1, 4, 4, 64 * 8]) # add a-relu
z_resize = tf.nn.relu(z_resize)
# up_1 = upsampling(z_resize, [self.batch_size, 8, 8], 512, 1024, 2, name='g_up2')
up_2 = upsampling(z_resize, [self.batch_size, 8, 8],
256,
512,
2,
name='g_up3')
up_3 = upsampling(up_2, [self.batch_size, 16, 16],
128,
256,
2,
name='g_up4')
up_4 = upsampling(up_3, [self.batch_size, 32, 32],
32,
128,
2,
name='g_up5')
up_5 = upsampling(up_4, [self.batch_size, 64, 64],
1,
32,
2,
name='g_up6')
# up_6 = upsampling(up_5, [self.batch_size, 128, 128], 16, 32, 2, name='g_up7')
# up_7 = upsampling(up_6, [self.batch_size, 256, 256], 1,16 , 2, name='g_up8')
return tf.nn.tanh(up_5)
def create_conditional_u_net(input, y, num_classes, keep_prob):
# Pass the image to obtain
input = tf.nn.lrn(input) # normalisation step for intensity scaling
image = tf.concat([input, y], axis=3)
print("Generator input image")
print_shape(image)
conv1_1 = convolution_block(image, [3, 3, 2, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
'g_conv1_1',
batch_normalisation=True,
tanh=False)
conv1_2 = convolution_block(conv1_1, [3, 3, 64, 64], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv1_2',
batch_normalisation=True,
tanh=False)
conv2_1 = convolution_block(conv1_2, [3, 3, 64, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv2_1',
batch_normalisation=True,
tanh=False)
conv2_2 = convolution_block(conv2_1, [3, 3, 128, 128], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv2_2',
batch_normalisation=True,
tanh=False)
conv3_1 = convolution_block(conv2_2, [3, 3, 128, 256], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv3_1',
batch_normalisation=True,
tanh=False)
conv3_2 = convolution_block(conv3_1, [3, 3, 256, 256], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv3_2',
batch_normalisation=True,
tanh=False)
conv4_1 = convolution_block(conv3_2, [3, 3, 256, 512], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv4_1',
batch_normalisation=True,
tanh=False)
conv4_2 = convolution_block(conv4_1, [3, 3, 512, 512], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv4_2',
batch_normalisation=True,
tanh=False)
conv5_1 = convolution_block(conv4_2, [3, 3, 512, 1024], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv5_1',
batch_normalisation=True,
tanh=False)
conv5_2 = convolution_block(conv5_1, [3, 3, 1024, 1024], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv5_2',
batch_normalisation=True,
tanh=False)
up_6 = upsampling(conv5_2, tf.shape(conv4_2), 512, 1024, 2, name='g_up6')
concat_6 = tf.concat([up_6, conv4_2], axis=3)
conv6_1 = convolution_block(concat_6, [3, 3, 1024, 512], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv6_1',
batch_normalisation=True,
tanh=False)
conv6_2 = convolution_block(conv6_1, [3, 3, 512, 512], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv6_2',
batch_normalisation=True,
tanh=False)
up_7 = upsampling(conv6_2, tf.shape(conv3_2), 256, 512, 2, name='g_up7')
concat_7 = tf.concat([up_7, conv3_2], axis=3)
conv7_1 = convolution_block(concat_7, [3, 3, 512, 256], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv7_1',
batch_normalisation=True,
tanh=False)
conv7_2 = convolution_block(conv7_1, [3, 3, 256, 256], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv7_2',
batch_normalisation=True,
tanh=False)
up_8 = upsampling(conv7_2, tf.shape(conv2_2), 128, 256, 2, name='g_up8')
concat_8 = tf.concat([up_8, conv2_2], axis=3)
conv8_1 = convolution_block(concat_8, [3, 3, 256, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv8_1',
batch_normalisation=True,
tanh=False)
conv8_2 = convolution_block(conv8_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv8_2',
batch_normalisation=True,
tanh=False)
up_9 = upsampling(conv8_2, tf.shape(conv1_2), 64, 128, 2, name='g_up9')
concat_9 = tf.concat([up_9, conv1_2], axis=3)
conv9_1 = convolution_block(concat_9, [3, 3, 128, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv9_1',
batch_normalisation=True,
tanh=False)
conv9_2 = convolution_block(conv9_1, [3, 3, 64, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv9_2',
batch_normalisation=True,
tanh=False)
conv_10 = convolution_block(conv9_2, [1, 1, 64, num_classes], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv10',
batch_normalisation=False,
tanh=True)
print("Completed creating U-NET ")
print("Number of classes is:" + str(num_classes))
##with tf.variable_scope('g_logits'):
# logits=tf.nn.softmax(conv_10, name='generator_output') # get logits here
return conv_10
def create_encorder(self):
#create the encorder side
input = tf.nn.lrn(self.X_train) #normalisation step
conv1_1 = convolution(input, [3, 3, 1, 64], [1, 1, 1, 1], 'SAME',
self.keep_prob, 'conv1_1')
conv1_2 = convolution(conv1_1, [3, 3, 64, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv1_2')
max1 = max_pooling(conv1_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max1')
conv2_1 = convolution(max1, [3, 3, 64, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv2_1')
conv2_2 = convolution(conv2_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv2_2')
max2 = max_pooling(conv2_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max2')
conv3_1 = convolution(max2, [3, 3, 128, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv3_1')
conv3_2 = convolution(conv3_1, [3, 3, 256, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv3_2')
max3 = max_pooling(conv3_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max3')
conv4_1 = convolution(max3, [3, 3, 256, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv4_1')
conv4_2 = convolution(conv4_1, [3, 3, 512, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv4_2')
max4 = max_pooling(conv4_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max4')
conv5_1 = convolution(max4, [3, 3, 512, 1024], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv5_1')
conv5_2 = convolution(conv5_1, [3, 3, 1024, 1024], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv5_2')
up_6 = upsampling(conv5_2, tf.shape(conv4_2), 512, 1024, 2, name='up6')
concat_6 = tf.concat([up_6, conv4_2], axis=3)
conv6_1 = convolution(concat_6, [3, 3, 1024, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv6_1')
conv6_2 = convolution(conv6_1, [3, 3, 512, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv6_2')
up_7 = upsampling(conv6_2, tf.shape(conv3_2), 256, 512, 2, name='up7')
concat_7 = tf.concat([up_7, conv3_2], axis=3)
conv7_1 = convolution(concat_7, [3, 3, 512, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv7_1')
conv7_2 = convolution(conv7_1, [3, 3, 256, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv7_2')
up_8 = upsampling(conv7_2, tf.shape(conv2_2), 128, 256, 2, name='up8')
concat_8 = tf.concat([up_8, conv2_2], axis=3)
conv8_1 = convolution(concat_8, [3, 3, 256, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv8_1')
conv8_2 = convolution(conv8_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv8_2')
up_9 = upsampling(conv8_2, tf.shape(conv1_2), 64, 128, 2, name='up9')
concat_9 = tf.concat([up_9, conv1_2], axis=3)
conv9_1 = convolution(concat_9, [3, 3, 128, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv9_1')
conv9_2 = convolution(conv9_1, [3, 3, 64, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv9_2')
conv_10 = convolution(conv9_2, [1, 1, 64, 1], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv10')
print("Down-sampling")
self.print_shape(conv1_2)
self.print_shape(max1)
self.print_shape(conv2_1)
self.print_shape(max2)
self.print_shape(conv3_2)
self.print_shape(max3)
self.print_shape(conv4_1)
self.print_shape(max4)
self.print_shape(conv5_2)
print("Upsampling")
self.print_shape(concat_6)
self.print_shape(conv6_2)
self.print_shape(concat_7)
self.print_shape(conv7_2)
self.print_shape(concat_8)
self.print_shape(conv8_2)
self.print_shape(concat_9)
self.print_shape(conv9_2)
self.print_shape(conv_10)
logits = tf.nn.softmax(conv_10) # get logits here
return logits
def create_u_net_small(input, num_classes, keep_prob):
# u net with 5 block layers
# Pass the image to obtain
# input=tf.nn.lrn(input) #normalisation step for intensity scaling
conv1_1 = convolution_block(input, [3, 3, 1, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
'g_conv1_1',
batch_normalisation=True,
tanh=False)
conv1_2 = convolution_block(conv1_1, [3, 3, 64, 64], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv1_2',
batch_normalisation=True,
tanh=False)
conv2_1 = convolution_block(conv1_2, [3, 3, 64, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv2_1',
batch_normalisation=True,
tanh=False)
conv2_2 = convolution_block(conv2_1, [3, 3, 128, 128], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv2_2',
batch_normalisation=True,
tanh=False)
conv3_1 = convolution_block(conv2_2, [3, 3, 128, 256], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv3_1',
batch_normalisation=True,
tanh=False)
conv3_2 = convolution_block(conv3_1, [3, 3, 256, 256], [1, 2, 2, 1],
'SAME',
keep_prob,
name='g_conv3_2',
batch_normalisation=True,
tanh=False)
"""
conv4_1=convolution_block(conv3_2, [3,3,256,512], [1,1,1,1],'SAME', keep_prob, name='g_conv4_1')
conv4_2=convolution_block(conv4_1,[3,3,512,512], [1,2,2,1],'SAME', keep_prob, name='g_conv4_2')
conv5_1=convolution_block(conv4_2, [3,3,512,1024], [1,1,1,1],'SAME', keep_prob, name='g_conv5_1')
conv5_2=convolution_block(conv5_1,[3,3,1024,1024], [1,2,2,1],'SAME', keep_prob, name='g_conv5_2')
up_6 = upsampling(conv5_2, tf.shape(conv4_2), 512, 1024,2, name='g_up6')
concat_6 = tf.concat([up_6, conv4_2], axis=3)
conv6_1 = convolution_block(concat_6, [3,3,1024,512], [1,1,1,1], 'SAME', keep_prob, name='g_conv6_1')
conv6_2 = convolution_block(conv6_1, [3,3,512,512], [1,1,1,1], 'SAME', keep_prob, name='g_conv6_2')
"""
up_7 = upsampling(conv3_2, tf.shape(conv2_2), 128, 256, 2, name='g_up7')
concat_7 = tf.concat([up_7, conv2_2], axis=3)
conv7_1 = convolution_block(concat_7, [3, 3, 256, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv7_1',
batch_normalisation=True,
tanh=False)
conv7_2 = convolution_block(conv7_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv7_2',
batch_normalisation=True,
tanh=False)
up_8 = upsampling(conv7_2, tf.shape(conv1_2), 64, 128, 2, name='g_up8')
concat_8 = tf.concat([up_8, conv1_2], axis=3)
conv8_1 = convolution_block(concat_8, [3, 3, 128, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv8_1',
batch_normalisation=True,
tanh=False)
conv8_2 = convolution_block(conv8_1, [3, 3, 64, 64], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv8_2',
batch_normalisation=True,
tanh=False)
# up_9 = upsampling(conv8_2, tf.shape(conv1_2), 32, 64, 2, name='g_up9')
# concat_9 = tf.concat([up_9, conv1_2], axis=3)
# conv9_1 = convolution_block(concat_9, [3,3,64,32], [1,1,1,1], 'SAME', keep_prob, name='g_conv9_1', batch_normalisation=True)
# conv9_2 = convolution_block(conv9_1, [3,3,32, 32], [1,1,1,1], 'SAME', keep_prob, name='g_conv9_2', batch_normalisation=True)
conv_10 = convolution_block(conv8_2, [1, 1, 64, num_classes], [1, 1, 1, 1],
'SAME',
keep_prob,
name='g_conv10',
batch_normalisation=False,
tanh=True)
print("Completed creating U-NET (small) ")
##with tf.variable_scope('g_logits'):
## logits=tf.nn.softmax(conv_10, name='g_output') # get logits here
return conv_10
def create_decoder(self, input_tensor):
# create the decorder side
conv11_1 = convolution(input_tensor, [3, 3, 1, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv11_1')
max11 = max_pooling(conv11_1, [1, 2, 2, 1], [1, 2, 2, 1], name='max11')
conv12_1 = convolution(max11, [3, 3, 64, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv12_1')
conv12_2 = convolution(conv12_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv12_2')
max12 = max_pooling(conv12_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max12')
conv13_1 = convolution(max12, [3, 3, 128, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv13_1')
conv13_2 = convolution(conv13_1, [3, 3, 256, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv13_2')
max13 = max_pooling(conv13_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max13')
conv14_1 = convolution(max13, [3, 3, 256, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv14_1')
conv14_2 = convolution(conv14_1, [3, 3, 512, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv14_2')
max14 = max_pooling(conv14_2, [1, 2, 2, 1], [1, 2, 2, 1], name='max14')
conv15_1 = convolution(max14, [3, 3, 512, 1024], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv15_1')
conv15_2 = convolution(conv15_1, [3, 3, 1024, 1024], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv15_2')
up_16 = upsampling(conv15_2,
tf.shape(conv14_2),
512,
1024,
2,
name='up16')
concat_16 = tf.concat([up_16, conv14_2], axis=3)
conv16_1 = convolution(concat_16, [3, 3, 1024, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv16_1')
conv16_2 = convolution(conv16_1, [3, 3, 512, 512], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv16_2')
up_17 = upsampling(conv16_2,
tf.shape(conv13_2),
256,
512,
2,
name='up17')
concat_17 = tf.concat([up_17, conv13_2], axis=3)
conv17_1 = convolution(concat_17, [3, 3, 512, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv17_1')
conv17_2 = convolution(conv17_1, [3, 3, 256, 256], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv17_2')
up_18 = upsampling(conv17_2,
tf.shape(conv12_2),
128,
256,
2,
name='up18')
concat_18 = tf.concat([up_18, conv12_2], axis=3)
conv18_1 = convolution(concat_18, [3, 3, 256, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv18_1')
conv18_2 = convolution(conv18_1, [3, 3, 128, 128], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv18_2')
up_19 = upsampling(conv18_2,
tf.shape(conv11_1),
64,
128,
2,
name='up19')
concat_19 = tf.concat([up_19, conv11_1], axis=3)
conv19_1 = convolution(concat_19, [3, 3, 128, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv19_1')
conv19_2 = convolution(conv19_1, [3, 3, 64, 64], [1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv19_2')
conv20 = convolution(conv19_2, [1, 1, 64, self.num_classes],
[1, 1, 1, 1],
'SAME',
self.keep_prob,
name='conv20')
logits = tf.nn.softmax(conv20)
return logits