def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = Lambda(BaseNet.normalize)(input_img)
x0 = Conv2D(depth, (7, 7), padding='same')(xs)
x1 = Conv2D(depth, (5, 5), padding='same')(xs)
x2 = Conv2D(depth, (3, 3), padding='same')(xs)
x3 = Conv2D(depth, (1, 1), padding='same')(xs)
m0 = BaseNet.res_block(x0, 3, depth)
m1 = BaseNet.res_block(x1, 3, depth)
m2 = BaseNet.res_block(x2, 3, depth)
m3 = BaseNet.res_block(x3, 3, depth)
x1 = Add()([x0, x1, x2, x3])
x2 = Add()([m0, m1, m2, m3])
x = Add()([x1, x2])
x = Conv2D(depth * 9, (3, 3), padding='same')(x)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth, (3, 3), padding='same')(x)
x = Conv2D(depth, (1, 1), padding='same')(x)
x = Conv2D(1, (1, 1), padding='same')(x)
x = Lambda(BaseNet.denormalize)(x)
return Model(input_img, x)
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = tf.keras.layers.Lambda(BaseNet.normalize)(input_img)
x0 = Conv2D(depth, (7, 7), padding='same')(xs)
x1 = Conv2D(depth, (5, 5), padding='same')(xs)
x2 = Conv2D(depth, (3, 3), padding='same')(xs)
x3 = Conv2D(depth, (1, 1), padding='same')(xs)
ss = Add()([x0, x1, x2, x3])
x = Conv2D(depth * 9, (5, 5), padding='same')(ss)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
s = Conv2D(depth * 9 * 4, (5, 5), padding='same')(ss)
s = Lambda(BaseNet.pixel_shuffle(3))(s)
s = Conv2D(depth * 4, (5, 5), padding='same')(s)
x = Add()([s, x])
x1 = x
x = Conv2D(depth * 2, (5, 5), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
s = Conv2D(depth * 9 * 2, (5, 5), padding='same')(ss)
s = Lambda(BaseNet.pixel_shuffle(3))(s)
s = Conv2D(depth * 2, (5, 5), padding='same')(s)
x = Add()([s, x])
x2 = x
x = Conv2D(depth * 1, (3, 3), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
s = Conv2D(depth * 9, (5, 5), padding='same')(ss)
s = Lambda(BaseNet.pixel_shuffle(3))(s)
s = Conv2D(depth, (5, 5), padding='same')(s)
x1 = Conv2D(depth, (5, 5), padding='same')(x1)
x2 = Conv2D(depth, (5, 5), padding='same')(x2)
x = Add()([s, x, x1, x2])
x = Conv2D(int(depth * 0.5), (3, 3), padding='same')(x)
x = Conv2D(1, (3, 3), padding='same')(x)
x = tf.keras.layers.Lambda(BaseNet.denormalize)(x)
return Model(input_img, x)
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
x = Conv2D(depth * 9, (5, 5), padding='same')(input_img)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x1 = x
x = Conv2D(depth * 9, (5, 5), padding='same')(x)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x = Conv2D(depth * 9, (5, 5), padding='same')(x)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x = Conv2D(depth * 4, (5, 5), padding='same')(x)
x = MaxPooling2D((3, 3), padding='same')(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x = Conv2D(depth * 4, (5, 5), padding='same')(x)
x = MaxPooling2D((3, 3), padding='same')(x)
x = Conv2D(depth * 4, (7, 7), padding='same')(x)
x = Dropout(0.2)(x)
x = LeakyReLU(0.2)(x)
x = Add()([x, x1])
x = Conv2D(int(depth * 0.5), (3, 3), padding='same')(x)
x = Conv2D(1, (3, 3), padding='same')(x)
return Model(input_img, x)
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = tf.keras.layers.Lambda(BaseNet.normalize)(input_img)
x_a = []
for i in range(0, self.input_im, 2):
x0 = tf.keras.layers.Conv2D(depth, (1, 1),
padding='same')(xs[:, :, :, i:i + 1])
x0 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.LeakyReLU(alpha=0.2)(x0)
x0 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.LeakyReLU(alpha=0.4)(x0)
x1 = tf.keras.layers.Conv2D(depth, (1, 1),
padding='same')(xs[:, :, :,
i + 1:i + 2])
x1 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x1)
x1 = tf.keras.layers.LeakyReLU(alpha=0.2)(x1)
x1 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x1)
x1 = tf.keras.layers.LeakyReLU(alpha=0.4)(x1)
x = tf.keras.layers.Add()([x0, x1])
x = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(x)
x = tf.keras.layers.LeakyReLU(alpha=0.2)(x)
x_a.append(x)
x_l = []
for i in range(len(x_a) - 1):
print(np.shape(x_a[i:i + 1][:]))
x0 = tf.keras.layers.Add()(x_a[i:i + 2])
x0 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.Conv2D(depth * 9, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.Lambda(BaseNet.pixel_shuffle(3))(x0)
x0 = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.LeakyReLU(alpha=0.2)(x0)
x_l.append(x0)
x = tf.keras.layers.Add()(x_l)
x = tf.keras.layers.Conv2D(depth, (2, 2), padding='same')(x)
x = tf.keras.layers.LeakyReLU(alpha=0.2)(x)
x = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x)
x = tf.keras.layers.Conv2D(depth, (1, 1), padding='same')(x)
x = tf.keras.layers.Conv2D(1, (1, 1), padding='same')(x)
x = tf.keras.layers.Lambda(BaseNet.denormalize)(x)
autoencoder = Model(input_img, x)
return autoencoder
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = tf.keras.layers.Lambda(BaseNet.normalize)(input_img)
m = tf.keras.layers.Conv2D(depth * 4, 3, padding='same')(xs)
for j in range(self.num_rez_block):
m = BaseNet.res_block(m, 3, depth * 4)
m = tf.keras.layers.Conv2D(3 * 3**2, 3, padding='same')(m)
m = tf.keras.layers.Lambda(BaseNet.pixel_shuffle(3))(m)
# skip branch
s = tf.keras.layers.Conv2D(3 * 3**2, 5, padding='same')(xs)
s = tf.keras.layers.Lambda(BaseNet.pixel_shuffle(3))(s)
x = tf.keras.layers.Add()([m, s])
x = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(x)
x = tf.keras.layers.Conv2D(1, (1, 1), padding='same')(x)
x = tf.keras.layers.Lambda(BaseNet.denormalize)(x)
return Model(input_img, x)
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = tf.keras.layers.Lambda(BaseNet.normalize)(input_img)
x0 = Conv2D(depth, (7, 7), padding='same')(xs)
x1 = Conv2D(depth, (5, 5), padding='same')(xs)
x2 = Conv2D(depth, (3, 3), padding='same')(xs)
x3 = Conv2D(depth, (1, 1), padding='same')(xs)
x = Add()([x0, x1, x2, x3])
x = Conv2D(depth * 9, (3, 3), padding='same')(x)
x = Lambda(BaseNet.pixel_shuffle(3))(x)
x = Conv2D(depth, (3, 3), padding='same')(x)
x = Conv2D(depth, (1, 1), padding='same')(x)
x = Conv2D(1, (1, 1), padding='same')(x)
x = tf.keras.layers.Lambda(BaseNet.denormalize)(x)
return Model(input_img, x)
def build_net(self):
depth = self.depth
input_img = tf.keras.layers.Input(shape=(128, 128, self.input_im))
xs = tf.keras.layers.Lambda(BaseNet.normalize)(input_img)
x_a = []
for i in range(0, self.input_im):
m = tf.keras.layers.Conv2D(depth, 3, padding='same')(xs[:, :, :,
i:i + 1])
for i in range(self.num_rez_block):
m = BaseNet.res_block(m, 3, depth)
x0 = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(m)
x0 = tf.keras.layers.Conv2D(depth * 4, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.Conv2D(depth * 9, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.Lambda(BaseNet.pixel_shuffle(3))(x0)
x_0 = x0
x1 = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(m)
x1 = tf.keras.layers.Conv2D(depth * 4, (1, 1), padding='same')(x1)
x1 = tf.keras.layers.Conv2D(depth * 9, (1, 1), padding='same')(x1)
x1 = tf.keras.layers.Lambda(BaseNet.pixel_shuffle(3))(x1)
x0 = tf.keras.layers.add([x1, x0])
x0 = tf.keras.layers.Conv2D(depth * 2, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.LeakyReLU(alpha=0.2)(x0)
x1 = tf.keras.layers.add([x_0, x1])
x1 = tf.keras.layers.Conv2D(depth * 2, (1, 1), padding='same')(x1)
x1 = tf.keras.layers.LeakyReLU(alpha=0.2)(x1)
x3 = tf.keras.layers.add([x0, x1])
x = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(x3)
x = tf.keras.layers.LeakyReLU(alpha=0.2)(x)
x = tf.keras.layers.Conv2D(depth * 2, (3, 3), padding='same')(x)
x = tf.keras.layers.add([x, x3])
x_a.append(x)
x_l = []
for i in range(len(x_a) - 1):
x0 = tf.keras.layers.concatenate(x_a[i:i + 2])
x0 = tf.keras.layers.Conv2D(depth * 4, (3, 3), padding='same')(x0)
x0 = tf.keras.layers.Conv2D(depth * 2, (1, 1), padding='same')(x0)
x0 = tf.keras.layers.LeakyReLU(alpha=0.2)(x0)
x_l.append(x0)
x = tf.keras.layers.add(x_l)
encoded = tf.keras.layers.Conv2D(depth * 4, (2, 2), padding='same')(x)
x = tf.keras.layers.LeakyReLU(alpha=0.2)(encoded)
x = tf.keras.layers.Conv2D(depth * 2, (1, 1), padding='same')(x)
x = tf.keras.layers.Conv2D(depth, (5, 5), padding='same')(x)
x = tf.keras.layers.Conv2D(depth, (3, 3), padding='same')(x)
x = tf.keras.layers.Conv2D(1, (1, 1), padding='same')(x)
x = tf.keras.layers.Conv2D(1, (1, 1),
padding='same',
activation='sigmoid')(x)
x = tf.keras.layers.Lambda(BaseNet.denormalize)(x)
autoencoder = Model(input_img, x)
return autoencoder