network.py

from keras.models import *
from keras.layers import *
from keras.optimizers import *


IMAGE_SIZE=500

def network(input_size=(IMAGE_SIZE,IMAGE_SIZE,3)):
    #input layer
    inputs = Input(input_size)
    
    #block 1 convnet
    zero = ZeroPadding2D(padding=(100, 100), data_format=None)(inputs)
    conv1 = Conv2D(64, 3,  kernel_initializer='he_normal', activation='relu')(zero)
    conv1_1 = Conv2D(64, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv1)
    pool1 = MaxPooling2D(pool_size=(2, 2))(conv1_1)
    
    
    #block 2 convnet
    conv2_1 = Conv2D(128, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool1)
    conv2_2 = Conv2D(128, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv2_1)
    zero2 = ZeroPadding2D(padding=(1, 1), data_format=None)(conv2_2)
    pool2 = MaxPooling2D(pool_size=(2, 2))(zero2)
    
    
    #block 3 convnet 
    
    conv3_1 = Conv2D(256, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool2)
    conv3_2 = Conv2D(256, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv3_1)
    conv3_3 = Conv2D(256, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv3_2)
    zero3 = ZeroPadding2D(padding=(1, 1), data_format=None)(conv3_3)
    pool3 = MaxPooling2D(pool_size=(2, 2))(zero3)
    
    #block 4 convnet 
    conv4_1 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool3)
    conv4_2 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv4_1)
    conv4_3 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv4_2)
    pool4 = MaxPooling2D(pool_size=(2, 2))(conv4_3)
    
    
    #block 5 convnet 
    conv5_1 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool4)
    conv5_2 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv5_1)
    conv5_3 = Conv2D(512, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(conv5_2)
    pool5 = MaxPooling2D(pool_size=(2, 2))(conv5_3)
    
    #block 6 convnet
    conv6 = Conv2D(4096, 7,  kernel_initializer='he_normal', activation='relu')(pool5)

    
    #dropout    
    dropout1 = Dropout(0.5)(conv6)
    
    #block 7 convnet
    conv7 = Conv2D(4096, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(dropout1)
    
    dropout2 = Dropout(0.5)(conv7)

    #block 8 convnet
    conv8 = Conv2D(3, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool2)

    #block 9 convnet
    conv9 = Conv2D(3, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool3)

    #block 10 convnet
    conv10 = Conv2D(3, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(pool4)
    #block 11 convnet
    conv11 = Conv2D(3, 3,  kernel_initializer='he_normal', activation='relu',padding='same')(dropout2)

    #Deconv1 
    upSample1 = UpSampling2D(size=(2, 2))(dropout2)
    deconv1 = Conv2DTranspose(3,3)(upSample1)

    #fuse1
    fuse1 = concatenate([deconv1,conv10], axis=2)
    model = Model(input=inputs, output=fuse1)


    model.summary()