How to compute sample weight in multi-task model?

Question

What's the correct way to calculate sample weight in a multi-task model?

Concretely, I have a model that outputs a 400 class [multi-class] classification, as well as a 5 class multi-label classification. The classes are all disproportional and out of balance. Without any sample weighting, the multi-class classifier gets ~90% top 3 accuracy, and good AUC on some of classes in the multi-label classifications. The goal is to improve performance in one of the classes in the multi-label classification task. This class is fairly unbalanced.

What I'm attempting is to set the sample weights to a fixed value (say, 0.5) when the class of interest is not set and some other arbitrary fixed value (say, 2.5) when it is.

Answer 1

In keras, there is a class_weight argument. Create a dictionary of weights per class and then pass that into the .fit method of the model.

class_weight = {0:.5, 1:2.5, 2:2.5, 3:2.5}

model.fit(
    train_features,
    train_targets,
    batch_size=2048,
    epochs=30,
    verbose=2,
    callbacks=callbacks,
    validation_data=(val_features, val_targets),
    class_weight=class_weight,
)

The weight values can be hand-picked or tuned like any other hyperparameter.