Keras自定义损失：想要跟踪每个时代结束时的每个损失值

Question

我想在每个时代结束时检查self.losses['RMSE']，self.loss['CrossEntropy']和self.loss['OtherLoss']的值。目前，我只能检查总损失self.loss['total']。

def train_test(self):
    def custom_loss(y_true, y_pred):
        ## (...) Calculate several losses inside this function
        self.losses['total'] = self.losses['RMSE'] + self.losses['CrossEntropy'] + self.losses['OtherLoss']
        return self.losses['total']


    ## (...) Generate Deep learning model & Read Inputs
    logits = keras.layers.Dense(365, activation=keras.activations.softmax)(concat)
    self.model = keras.Model(inputs=[...], outputs=logits)

    self.model.compile(optimizer=keras.optimizers.Adam(0.001),
                       loss=custom_loss)

    self.history = self.model.fit_generator(
        generator=self.train_data,
        steps_per_epoch=train_data_size//FLAGS.batch_size,
        epochs=5,
        callbacks=[CallbackA(self.losses)])

class TrackTestDataPerformanceCallback(keras.callbacks.Callback):
    def __init__(self, losses):
        self.losses = losses

    def on_epoch_end(self, epoch, logs={}):
        for key in self.losses.keys()
            print('Type of loss: {}, Value: {}'.format(key, K.eval(self.losses[key])))

我将self.loss传递给回调函数CallbackA，以便在每个纪元的末尾打印子损失值。但是，它给出了如下错误消息：

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input_3' with dtype float and shape [?,5]
 [[Node: input_3 = Placeholder[dtype=DT_FLOAT, shape=[?,5], _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]
 [[Node: loss/dense_3_loss/survive_rates/while/LoopCond/_881 = _HostRecv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_360_loss/dense_3_loss/survive_rates/while/LoopCond", tensor_type=DT_BOOL, _device="/job:localhost/replica:0/task:0/device:CPU:0"](^_clooploss/dense_3_loss/survive_rates/while/strided_slice_4/stack_2/_837)]]

我可以再次将列车数据传递给回调函数，并预测自己跟踪每个损失值。但我认为可能还有一个我还不知道的更好的解决方案。

摘要：如何在每个纪元后追踪自定义损失函数中的几个损失值？

约束：为了减少一些计算成本，我想在custom_loss函数中管理几个损失。但是，如果我必须将每个损失包装到每个函数中，那就没问题。

Answer 1

编译时，您可以在列表中使用多个损失。例如，如果你想混合交叉熵和mse，你可以使用：

model.compile(loss=['mse', 'binary_crossentropy'], loss_weights=[0.9, 0.1], optimizer=Adam())

历史记录将包含编译模型时使用的不同损失。

Answer 2

我必须为我们的模型维持一个组合的custom_loss，所以我找到了一种通过输入metrics参数来跟踪几个子损失的方法。每个损失函数都作为函数单独定义。

def custom_loss():
    return subloss1() + subloss2() + subloss3()

def subloss1():
    ...
    return value1

def subloss2():
    ...
    return value2

def subloss3():
    ...
    return value3


self.model.compile(optimizer=keras.optimizers.Adam(0.001),
                       loss=custom_loss,
                       metrics=[subloss1, subloss2, subloss3]