Keras Custom Loss Feature

Question

I am working on a class-incremental classifier approach using CNN as a function extractor and a fully connected block for classification.

First, I fine-tuned VGG for each trained network to complete a new task. When the network learns a new task, I keep a few examples for each class so as not to forget when new classes are available.

When some classes are available, I must calculate each output of the instances included in the examples for the new classes. Now adding zeros to the outputs for the old classes and adding a label corresponding to each new class at the output of the new classes, I have my new shortcuts, i.e. if 3 new classes are introduced.

Old type type: [0.1, 0.05, 0.79, ..., 0 0 0]

A new type of type: [0.1, 0.09, 0.3, 0.4, ..., 1 0 0] ** the last outputs correspond to the class.

My question is, how can I change the loss function for a custom one that will train for new classes? The loss function that I want to implement is defined as:

where the distillation losses correspond to the outputs of the old classes to avoid forgetting, and the loss of classification corresponds to the new classes.

If you can provide me a sample code to change the loss function in keras, it would be nice.

Thanks!!!!!

Answer 1

All you have to do is define a function for this, using keras functions for calculations. The function should take the true values ​​and the predicted values ​​of the model.

Now, since I'm not sure what g, q, x, and y are in your function, I’ll just create a basic example here without worrying about what this means or whether it is really a useful function:

 import keras.backend as K def customLoss(yTrue,yPred): return K.sum(K.log(yTrue) - K.log(yPred)) 

All backend functions can be viewed here: https://keras.io/backend/#backend-functions

After that, compile your model using this function instead of the usual one:

 model.compile(loss=customLoss, optimizer = .....)