class_weight='auto' for model.fit_generator keras

Question

I have a highly unbalanced dataset and I want to use the class_weight = 'auto' in the model.fit_generator . However, when I do that I see that my model does not learn: training_acc = 0.65 and val_acc = 0.64 starting from epoch 1 up to 50. If I set the class_weight = 'None' then the model starts learning: training_acc = 0.92 and val_acc = 0.88 at epoch 50 .

Did anyone else face this problem? Do I have to define a dictionary to my class weights manually?

Answer 1

For model.fit_generator in keras you can use train_generator.classes for the proper class names for your weighting

Then you simple create a dictionary mapping your classes., eg

class_weights = {'wolf':30 , 'fox':18}

That gives classes 'wolf' weight 30 and 'fox' weight '18'