How do you create a boolean mask for a tensor in Keras?

Question

I am building a custom metric to measure the accuracy of one class in my multi-class dataset during training. I am having trouble selecting the class.

The targets a

Answer 1

Note that when talking about the accuracy of one class one may refer to either of the following (not equivalent) two amounts:

• The recall, which, for class C, is the ratio of examples labelled with class C that are predicted to have class C.
• The precision, which, for class C, is the ratio of examples predicted to be of class C that are in fact labelled with class C.

Instead of doing complex indexing, you can just rely on masking for you computation. Assuming we are talking about precision here (changing to recall would be trivial).

from keras import backend as K

INTERESTING_CLASS_ID = 0  # Choose the class of interest

def single_class_accuracy(y_true, y_pred):
    class_id_true = K.argmax(y_true, axis=-1)
    class_id_preds = K.argmax(y_pred, axis=-1)
    # Replace class_id_preds with class_id_true for recall here
    accuracy_mask = K.cast(K.equal(class_id_preds, INTERESTING_CLASS_ID), 'int32')
    class_acc_tensor = K.cast(K.equal(class_id_true, class_id_preds), 'int32') * accuracy_mask
    class_acc = K.sum(class_acc_tensor) / K.maximum(K.sum(accuracy_mask), 1)
    return class_acc

If you want to be more flexible, you can also have the class of interest parametrised:

from keras import backend as K

def single_class_accuracy(interesting_class_id):
    def fn(y_true, y_pred):
        class_id_true = K.argmax(y_true, axis=-1)
        class_id_preds = K.argmax(y_pred, axis=-1)
        # Replace class_id_preds with class_id_true for recall here
        accuracy_mask = K.cast(K.equal(class_id_preds, interesting_class_id), 'int32')
        class_acc_tensor = K.cast(K.equal(class_id_true, class_id_preds), 'int32') * accuracy_mask
        class_acc = K.sum(class_acc_tensor) / K.maximum(K.sum(accuracy_mask), 1)
        return class_acc
    return fn

And the use it as:

model.compile(..., metrics=[single_class_accuracy(INTERESTING_CLASS_ID)])