Beginning Anomaly Detection Using

 Categorical  Accuracy

keras.metrics.categorical_accuracy(y_true, y_pred)

Since most problems tend to involve categorical cross entropy (implying more than 

two classes in the data set), this tends to be the default accuracy metric when ‘accuracy’ 

is passed into the model.compile() function.

Instead of finding all of the instances where the true labels and rounded predictions 

match, categorical accuracy finds all of the instances where the true labels and 

predictions have a maximum value in the same spot.

Recall that for categorical cross entropy, the labels are one-hot encoded. Therefore, 

the truth labels only have one maximum per entry, along with the predictions (though 

again, one value will be really close to one while the others are really close to zero). What 

categorical accuracy does is check if the maximum value in the entry is in the same 

position for both y_true and for y_pred.

Once it’s found all those instances, it finds the mean of the result, leading to an 

accuracy value.

Essentially, it’s a similar equation to the one for binary accuracy, but with a different 

condition regarding y_true and y_pred.

Appendix A   intro to KerAs

---

345

The function is defined by Keras as shown in Figure 

A-24

.