Tensorflow keras custom loss

Question

I have 1000 XRay images and I'm trying to make some binary predictions (died/lived). I have 900 'lived' pictures and 100 'died' pictures and after some reading (mostly here), I think that I may have to make something about this imbalance of the samples.

I tried the

model.compile(...,loss=mycustom_loss,...)

approach but I think that I can't understand the syntax of the custom function. Reading some other documents, I tried the approach:

my_custom_loss_weights = [ 100/1000, 900/1000 ]  # index 0 is 'died', index 1 is 'lived'
model.compile = (...,loss='binary_crossentropy',loss_weights=my_custom_loss_weights,...)

which gave some pretty good results (loss and val_loss were very close, acc and val_acc were very close and high enough).

So the questions are about how to make a custom loss function returning binary_crossentropy but weighted based on my weights (100/1000 and 900/1000) and second, if the loss_weights parameter is the solution I want and I'm really that lucky!

I tried the oversample approach of the link I mentioned, but the results were terrible.

Thank you in advance!

Answer 1

It's completely right how you are doing it, the loss_weights parameter is the way to go and also implemented correctly. It's also described in the link you posted under the paragraph Weighted loss.