In addition to a loss function, Keras lets us also use metrics to help judge the performance of a model. While minimizing loss is good, it's not especially obvious how we expect the model to perform given some loss function. Metrics aren't used in training the model, they're just there to help us understand the current state.
While loss might not mean much to us, accuracy does. We humans understand accuracy fairly well.
Keras defines binary accuracy as follows:
def binary_accuracy(y_true, y_pred):
return K.mean(K.equal(y_true, K.round(y_pred)), axis=-1)
This is really just a clever way to simply divide the number of correct answers by the total answers, as we've likely been doing since our very early days in school to figure out our grade on a test.
You might be wondering whether our dataset is balanced because accuracy works so poorly for unbalanced datasets. It's in fact not balanced. Only one-fifth of the dataset is class 1. We will calculate the ROC AUC score as a custom callback to address this. ROC isn't implemented in Keras as a metric because metrics are computed for every mini batch and the ROC AUC score isn't really defined by mini batch.