How does Keras calculate the categorical cross-entropy?

From https://towardsdatascience.com/cross-entropy-for-classification-d98e7f974451 I understand that categorical cross-entropy is -log(q(x))

When using tf.keras.losses.CategoricalCrossentropy I assume that loss is calculated per batch and per epoch as the weighted sum of the losses of the training instances or is there something else?

(So with k-fold the loss would be the sum of weighted losses of the training folds and the validation loss would be the sum of weighted losses of the validation loss?)

How does Keras deal with a loss of inf?