Question

我想为两个损失函数动态更改损失权重。但是，当我在代码中使用Callbacks时，该错误表明损失权重变量未初始化。我已经用纯Keras测试过类似的代码，它可以正常运行并且没有初始化问题。 tensorflow样式的代码如下：

#  dynamic alpha weights for two losses   
class MyCallback(Callback):
    def __init__(self, alpha):
        self.alpha = alpha
    # customize your behavior
    def on_epoch_end(self, epoch, logs={}):
        results = [logs['sumrate_loss_loss'], logs['Wnorm_loss_loss']]
        K.set_value(self.alpha,  results[0] / (results[0] + results[1]) )

# CNN model
input_H = Input(shape=(Nr, Nt))
input_H_expand = Lambda(lambda x: K.expand_dims(x))(input_H)
input_sigma2 = Input(shape=(Nr,))
temp = Conv2D(Nr * Nt, (1, 2), activation='relu', padding='same')(input_H_expand)
temp = Conv2D(Nr * Nt, (1, 2), activation='relu', padding='same')(temp)
temp = Flatten()(temp)
output_W = Dense(Nt * Nr, activation='linear')(concatenate([temp, input_sigma2], axis=-1))
output_W = Lambda(lambda x: K.reshape(x, [-1, Nt, Nr]))(output_W)
input_I_DC = Input(shape=(1, ))

sumrate_loss = Lambda(sumrate, name='sumrate_loss')([input_H, output_W, input_sigma2])
W_norm_reg = Lambda(norms_reg, name='Wnorm_loss')([output_W, input_I_DC])

model = Model(inputs=[input_H, input_sigma2, input_I_DC], outputs=[sumrate_loss, W_norm_reg])

alpha = tf.Variable(0.5)
model.compile(optimizer='adam',
              loss={
                  'sumrate_loss': lambda y_true, y_pred: y_pred,
                  'Wnorm_loss': lambda y_true, y_pred: y_pred},
              loss_weights={
                  'sumrate_loss': alpha,
                  'Wnorm_loss':  1 - alpha},
             )

# fit model
reduce_lr = ReduceLROnPlateau(monitor='val_loss', 
                              factor=0.2,
                              patience=20, 
                              min_lr=0.0005)
checkpoint = ModelCheckpoint('./temp_trained_cnn_maxsumrate_p1_basic.h5', monitor='val_loss',
                              verbose=0, save_best_only=True, mode='min', save_weights_only=True)
input_fit = [H_sample / np.min(np.sqrt(sigma2n_sample[::, ::, 10])), sigma2n_sample[::, ::, 10] / np.min(sigma2n_sample[::, ::, 10]), I_DC_sample[::, 10]]
random_y = [np.random.random(H_sample.shape[0]), np.random.random(H_sample.shape[0])]
history = model.fit(input_fit, random_y,
                    batch_size=16,
                    epochs=10,
                    verbose=1, 
                    validation_split=0.2,
                    callbacks=[MyCallback(alpha)])

错误输出是MyCallback中的Alpha尚未初始化。详细的错误是：

---> 14个callbacks = [MyCallback（alpha）]）

FailedPreconditionError：尝试使用未初始化的值Variable_7 [[{{node Variable_7 / read}} = IdentityT = DT_FLOAT，_device =“ / job：localhost /副本：0 / task：0 / device：CPU：0”]]

我希望有人能告诉我如何在loss_weights中初始化alpha。

如何为张量流中的两个损失函数定义动态损失权重

0 个答案: