def u_res_side_fuse_v3(input_shape=None):
inputs = Input(shape=input_shape) # 512, 512, 1
# Normalization
x = Lambda(lambda x: x / 255, name='pre-process')(inputs)
# Block 1
c1 = conv_block(x, 3, (8, 8, 32), stage=1, block='a', strides=(2, 2))
c1 = identity_block(c1, 3, (8, 8, 32), stage=1, block='b') # 256
# Block 2
c2 = conv_block(c1, 3, (16, 16, 64), stage=2, block='a', strides=(2, 2))
c2 = identity_block(c2, 3, (16, 16, 64), stage=2, block='b') # 128
# Block 3
c3 = conv_block(c2, 3, (32, 32, 128), stage=3, block='a', strides=(2, 2))
c3 = identity_block(c3, 3, (32, 32, 128), stage=3, block='b') # 64
# Block 4
c4 = conv_block(c3, 3, (64, 64, 256), stage=4, block='a', strides=(2, 2))
c4 = identity_block(c4, 3, (64, 64, 256), stage=4, block='b') # 32
# Block 5
c5 = conv_block(c4, 3, (128, 128, 512), stage=5, block='a', strides=(2, 2))
c5 = identity_block(c5, 3, (128, 128, 512), stage=5, block='b') # 16
# Block 6
u6 = Conv2DTranspose(256, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_6',
kernel_initializer=init)(c5)
u6 = concatenate([u6, c4], name='concat_6')
c6 = conv_BN(u6, 256, 3)
c6 = conv_BN(c6, 256, 3) # 32
s1 = side_out(c6, 16)
# Block 7
u7 = Conv2DTranspose(128, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_7',
kernel_initializer=init)(c6)
u7 = concatenate([u7, c3], name='concat_7')
c7 = conv_BN(u7, 128, 3)
c7 = conv_BN(c7, 128, 3) # 64
s2 = side_out(c7, 8)
# Block 8
u8 = Conv2DTranspose(64, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_8',
kernel_initializer=init)(c7)
u8 = concatenate([u8, c2], name='concat_8')
c8 = conv_BN(u8, 64, 3)
c8 = conv_BN(c8, 64, 3) # 128
s3 = side_out(c8, 4)
# Block 9
u9 = Conv2DTranspose(32, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_9',
kernel_initializer=init)(c8)
u9 = concatenate([u9, c1], name='concat_9')
c9 = conv_BN(u9, 32, 3)
c9 = conv_BN(c9, 32, 3) # 256
s4 = side_out(c9, 2)
# fuse
fuse = concatenate(inputs=[s1, s2, s3, s4], axis=-1)
fuse = Conv2D(1, (1, 1), padding='same',
activation=None)(fuse) # 512, 512, 1
# outputs
o1 = Activation('sigmoid', name='o1')(s1)
o2 = Activation('sigmoid', name='o2')(s2)
o3 = Activation('sigmoid', name='o3')(s3)
o4 = Activation('sigmoid', name='o4')(s4)
ofuse = Activation('sigmoid', name='ofuse')(fuse)
model = Model(inputs=[inputs], outputs=[o1, o2, o3, o4, ofuse])
return model
def u_res_side_fuse(input_shape=None):
inputs = Input(shape=input_shape) # 320, 480, 3
# Normalization
x = Lambda(lambda x: x / 255, name='pre-process')(inputs)
x = Conv2D(16, (5, 5), strides=(1, 1), padding='same', name='conv1')(x)
x = BatchNormalization(axis=-1, name='bn_conv1')(x)
x = Activation('relu', name='act1')(x) # 320, 480, 3
# Block 1
c1 = conv_block(x, 3, (8, 8, 32), stage=1, block='a', strides=(1, 1))
c1 = identity_block(c1, 3, (8, 8, 32), stage=1, block='b') # 320, 480, 3
# Block 2
c2 = conv_block(c1, 3, (16, 16, 64), stage=2, block='a', strides=(2, 2))
c2 = identity_block(c2, 3, (16, 16, 64), stage=2, block='b') # 160, 240, 3
# Block 3
c3 = conv_block(c2, 3, (32, 32, 128), stage=3, block='a', strides=(2, 2))
c3 = identity_block(c3, 3, (32, 32, 128), stage=3, block='b') # 80, 120, 3
# Block 4
c4 = conv_block(c3, 3, (64, 64, 256), stage=4, block='a', strides=(2, 2))
c4 = identity_block(c4, 3, (64, 64, 256), stage=4, block='b') # 40, 60, 3
# Block 5
c5 = conv_block(c4, 3, (128, 128, 512), stage=5, block='a', strides=(2, 2))
c5 = identity_block(c5, 3, (128, 128, 512), stage=5,
block='b') # 20, 30, 3
s1 = side_out(c5, 16)
# Block 6
u6 = Conv2DTranspose(128, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_6',
kernel_initializer=init)(c5)
u6 = concatenate([u6, c4], name='concat_6') # 40, 60, 3
c6 = conv_block(u6, 3, (64, 64, 256), stage=6, block='a', strides=(1, 1))
c6 = identity_block(c6, 3, (64, 64, 256), stage=6, block='b')
s2 = side_out(c6, 8)
# Block 7
u7 = Conv2DTranspose(64, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_7',
kernel_initializer=init)(c6)
u7 = concatenate([u7, c3], name='concat_7') # 80, 120, 3
c7 = conv_block(u7, 3, (32, 32, 128), stage=7, block='a', strides=(1, 1))
c7 = identity_block(c7, 3, (32, 32, 128), stage=7, block='b')
s3 = side_out(c7, 4)
# Block 8
u8 = Conv2DTranspose(32, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_8',
kernel_initializer=init)(c7)
u8 = concatenate([u8, c2], name='concat_8') # 160, 240, 3
c8 = conv_block(u8, 3, (16, 16, 64), stage=8, block='a', strides=(1, 1))
c8 = identity_block(c8, 3, (16, 16, 64), stage=8, block='b')
s4 = side_out(c8, 2)
# Block 9
u9 = Conv2DTranspose(16, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_9',
kernel_initializer=init)(c8)
u9 = concatenate([u9, c1], name='concat_9') # 320, 480, 3
c9 = conv_block(u9, 3, (8, 8, 32), stage=9, block='a', strides=(1, 1))
c9 = identity_block(c9, 3, (8, 8, 32), stage=9, block='b')
s5 = side_out(c9, 1)
# fuse
fuse = concatenate(inputs=[s1, s2, s3, s4, s5], axis=-1)
fuse = Conv2D(1, (1, 1), padding='same',
activation=None)(fuse) # 320 480 1
# outputs
o1 = Activation('sigmoid', name='o1')(s1)
o2 = Activation('sigmoid', name='o2')(s2)
o3 = Activation('sigmoid', name='o3')(s3)
o4 = Activation('sigmoid', name='o4')(s4)
o5 = Activation('sigmoid', name='o5')(s5)
ofuse = Activation('sigmoid', name='ofuse')(fuse)
model = Model(inputs=[inputs], outputs=[o1, o2, o3, o4, o5, ofuse])
return model
def get_unet(self, gpu_count):
inputs = Input((self.img_rows, self.img_cols, 1))
x = Conv2D(16, (5, 5), strides=(1, 1), padding='same',
name='conv1')(inputs)
x = BatchNormalization(axis=-1, name='bn_conv1')(x)
x = Activation('relu', name='act1')(x)
# Block 1
c1 = conv_block(x, 3, (8, 8, 32), stage=1, block='a', strides=(1, 1))
c1 = identity_block(c1, 3, (8, 8, 32), stage=1,
block='b') # 320, 480, 3
# Block 2
c2 = conv_block(c1,
3, (16, 16, 64),
stage=2,
block='a',
strides=(2, 2))
c2 = identity_block(c2, 3, (16, 16, 64), stage=2,
block='b') # 160, 240, 3
# Block 3
c3 = conv_block(c2,
3, (32, 32, 128),
stage=3,
block='a',
strides=(2, 2))
c3 = identity_block(c3, 3, (32, 32, 128), stage=3,
block='b') # 80, 120, 3
# Block 4
c4 = conv_block(c3,
3, (64, 64, 256),
stage=4,
block='a',
strides=(2, 2))
c4 = identity_block(c4, 3, (64, 64, 256), stage=4,
block='b') # 40, 60, 3
s1 = side_out(c4, 8)
# Block 5
u5 = Conv2DTranspose(64, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_6',
kernel_initializer=init)(c4)
u5 = concatenate([u5, c3], name='concat_6') # 40, 60, 3
c5 = conv_block(u5,
3, (32, 32, 128),
stage=6,
block='a',
strides=(1, 1))
c5 = identity_block(c5, 3, (32, 32, 128), stage=6, block='b')
s2 = side_out(c5, 4)
# Block 6
u6 = Conv2DTranspose(32, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_7',
kernel_initializer=init)(c5)
u6 = concatenate([u6, c2], name='concat_7') # 80, 120, 3
c6 = conv_block(u6,
3, (16, 16, 64),
stage=7,
block='a',
strides=(1, 1))
c6 = identity_block(c6, 3, (16, 16, 64), stage=7, block='b')
s3 = side_out(c6, 2)
# Block 7
u7 = Conv2DTranspose(16, (2, 2),
strides=(2, 2),
padding='same',
name='upconv_8',
kernel_initializer=init)(c6)
u7 = concatenate([u7, c1], name='concat_8') # 160, 240, 3
c7 = conv_block(u7, 3, (8, 8, 32), stage=8, block='a', strides=(1, 1))
c7 = identity_block(c7, 3, (8, 8, 32), stage=8, block='b')
s4 = side_out(c7, 1)
# fuse
fuse = concatenate(inputs=[s1, s2, s3, s4], axis=-1)
fuse = Conv2D(1, (1, 1), padding='same',
activation=None)(fuse) # 320 480 1
# fuse = to_categorical(fuse,3)
# outputs
# o1 = Activation('sigmoid', name='o1')(s1)
# o2 = Activation('sigmoid', name='o2')(s2)
# o3 = Activation('sigmoid', name='o3')(s3)
# o4 = Activation('sigmoid', name='o4')(s4)
ofuse = Activation('sigmoid', name='ofuse')(fuse)
model = Model(inputs=[inputs], outputs=[ofuse])
# sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
adam = optimizers.Adam(decay=1e-6) #每个mini batch衰减一次
if (gpu_count == 0):
model.compile(optimizer=adam,
loss='binary_crossentropy',
metrics=[f1score])
print('model compile')
return model
else:
from keras.utils import multi_gpu_model
parallel_model = multi_gpu_model(model, gpus=gpu_count)
parallel_model.compile(optimizer=adam,
loss='binary_crossentropy',
metrics=[f1score])
print('parallel model compile')
return parallel_model