How to log out Keras custom loss component at the end of an epoch?

Question

I have a custom loss function as this format:

def CustomLoss(y_true,y_pred):
    ......
    loss = loss1 + loss2 + loss3
    return loss

How can I return each loss component (loss1, loss2, loss3) at the end of an epoch? By default, I can only observe loss.

Normally, if we have multiple outputs, keras can show it easily. But how to show the value component like what I am mentioned?

Answer 1

You need to create a function for each loss separately and pass them them into metrics when compiling the model as shown below

def CustomLoss():
    ......
    return loss1() + loss2() + loss3()

def Loss1():
    ......
    return value1

def Loss2():
    ......
    return value2

def Loss3():
    ......
    return value3


model.compile(optimizer=Adam(),
                       loss=CustomLoss,
                       metrics=[Loss1, Loss2, Loss3])