Keras custom loss function. How to know which output samples are currently called in the loss funciton

Question

I an building a custom loss function in keras where I want to manipulate the y_true and y_pred before calling mean_absolute_error() . This manipulation involves knowing which sample the current y_true is, therefore, I can easily add another column to the output which will have a sequential index. So I have

• outputs[:,0] which is the acual output variable for all the samples.

• outputs[:,1] a sequential index which I do not want to predict but just use in my custom loss function

   def loss(y_true, y_pred): for i in range(len(y_true)): index = y_true[i, 1] # some manipulations involving the index return mean_absolute_error(new_true, new_pred)

How can I achieve the above in a way that the model ignores the index column output and does not try to predict it as well.

There might be a better solution to my problem. My ultimate goal is to know which sample output(s) are currently in the loss function (by an index).

Answer 1

You can try this:

def customLoss(y_true, y_pred):
    
    # you can do your preprocessing here

    def loss(y_true, y_pred):
        for i in range(len(y_true)):
             index = y_true[i, 1]
             # some manipulations involving the index
        return mean_absolute_error(new_true, new_pred)


    return loss(y_true, y_pred)