joint_optimisation.py 文件源码

def create_joint_model(input_dim, init_w, init_b, gamma, weight_hinge, learning_rate, decay, regulariser=None):
    image_input = Input(shape=(input_dim,), dtype='float32', name='image_input')
    db_input = Input(shape=(input_dim,), dtype='float32', name="db_input")
    shared_layer = Dense(1, input_dim=input_dim, kernel_regularizer=regulariser, kernel_initializer='uniform',
                         activation="linear", use_bias=True, name='shared_layer')
    _ = shared_layer(image_input)
    _ = shared_layer(db_input)
    model = Model(inputs=[image_input, db_input], outputs=[shared_layer.get_output_at(0), shared_layer.get_output_at(1)])
    adam = Adam(lr=learning_rate)  # SGD should also work because convex loss function, but Adam converges faster.
    model.compile(optimizer=adam, loss=['hinge', 'mse'], loss_weights=[weight_hinge, gamma],
                  metrics=[my_accuracy, 'mse'])
    return model