rafael

gistfile1.txt

from keras.applications import MobileNet

# Construindo a SNN
input_shape=(224,224, 3)   # formato das imagens de entrada
left_input = Input(input_shape)
right_input = Input(input_shape)

# inicializando os kernels
W_init = keras.initializers.RandomNormal(mean = 0.0, stddev = 1e-2)
# inicializando os bias
b_init = keras.initializers.RandomNormal(mean = 0.5, stddev = 1e-2)

# declarando as camadas da SNN


model = keras.models.Sequential([
    MobileNet(include_top=False,input_shape=input_shape),
    keras.layers.Flatten(),
    keras.layers.Dense(256, activation='sigmoid', kernel_initializer=W_init, bias_initializer=b_init)
])

model.layers[0].trainable = False

# pegando a saída (mapa de features) da rede do lado esquerdo
encoded_l = model(left_input)

# pegando a saída (mapa de features) da rede do lado direito
encoded_r = model(right_input)

# Add a customized layer to compute the absolute difference between the encodings
L1_layer = Lambda(lambda tensors:K.abs(tensors[0] - tensors[1]))
L1_distance = L1_layer([encoded_l, encoded_r])

# camada que calcula a diferença entre os mapas de características
#subtracted = keras.layers.Subtract()([encoded_l, encoded_r])

# camada que faz a predição de ser ou não da mesma categoria
prediction = Dense(1, activation='sigmoid', bias_initializer=b_init)(L1_distance)
siamese_net = Model([left_input, right_input], prediction)

# compilando o modelo
optimizer= Adam(learning_rate=0.0006)
siamese_net.compile(loss='binary_crossentropy', optimizer=optimizer)

# plotando o modelo
siamese_net.summary()