texture_unet.py

x = Input(shape=(128,128,3))

n = 16

y0 = Conv2D(n, (3,3), padding='same', activation='relu')(x)
y0 = Conv2D(n, (3,3), padding='same', activation='relu')(y0)
y0 = Conv2D(n, (3,3), padding='same', activation='relu')(y0)

y1 = MaxPooling2D()(y0) # 64
y1 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y1)
y1 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y1)
y1 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y1)

y2 = MaxPooling2D()(y1) # 32
y2 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y2)
y2 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y2)
y2 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y2)

y3 = MaxPooling2D()(y2) # 16
y3 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y3)
y3 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y3)
y3 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y3)

y4 = MaxPooling2D()(y3) # 8
y4 = Conv2D(16*n, (3,3), padding='same', activation='relu')(y4)
y4 = Conv2D(16*n, (3,3), padding='same', activation='relu')(y4)
y4 = Conv2D(16*n, (3,3), padding='same', activation='relu')(y4)

y5 = UpSampling2D()(y4) # 16
y5 = Concatenate()([y3, y5])
y5 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y5)
y5 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y5)
y5 = Conv2D(8*n, (3,3), padding='same', activation='relu')(y5)

y6 = UpSampling2D()(y5) # 32
y6 = Concatenate()([y2, y6])
y6 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y6)
y6 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y6)
y6 = Conv2D(4*n, (3,3), padding='same', activation='relu')(y6)

y7 = UpSampling2D()(y6) # 64
y7 = Concatenate()([y1, y7])
y7 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y7)
y7 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y7)
y7 = Conv2D(2*n, (3,3), padding='same', activation='relu')(y7)

y8 = UpSampling2D()(y7) # 128
y8 = Concatenate()([y0, y8])
y8 = Conv2D(n, (3,3), padding='same', activation='relu')(y8)
y8 = Conv2D(n, (3,3), padding='same', activation='relu')(y8)
y8 = Conv2D(n, (3,3), padding='same', activation='relu')(y8)
y8 = Conv2D(3, (3,3), padding='same', activation='sigmoid')(y8)

mod = Model(inputs=[x], outputs=[y8])