Custom metrics in Keras and how simple they are to use in tensorflow2.2

So lets get down to it. We first make a custom metric class. While there are more steps to this and they are show in the referenced jupyter notebook , the important thing is to implement the API that integrates with the rest of Keras training and testing workflow. That is as simple as implementing and update_state that takes in the true labels and predictions, a reset_states that re-initializes the metric.

class ConfusionMatrixMetric(tf.keras.metrics.Metric):


    def update_state(self, y_true, y_pred,sample_weight=None):
        self.total_cm.assign_add(self.confusion_matrix(y_true,y_pred))
        return self.total_cm

    def result(self):
        return self.process_confusion_matrix()

    def confusion_matrix(self,y_true, y_pred):
        """
        Make a confusion matrix
        """
        y_pred=tf.argmax(y_pred,1)
        cm=tf.math.confusion_matrix(y_true,y_pred,dtype=tf.float32,num_classes=self.num_classes)
        return cm

    def process_confusion_matrix(self):
        "returns precision, recall and f1 along with overall accuracy"
        cm=self.total_cm
        diag_part=tf.linalg.diag_part(cm)
        precision=diag_part/(tf.reduce_sum(cm,0)+tf.constant(1e-15))
        recall=diag_part/(tf.reduce_sum(cm,1)+tf.constant(1e-15))
        f1=2*precision*recall/(precision+recall+tf.constant(1e-15))
        return precision,recall,f1

In the normal Keras workflow, the method result will be called and it will return a number and nothing else needs to be done. However, in our case we have three tensors for precision, recall and f1 being returned and Keras does not know how to handle this out of the box. This is where the new features of tensorflow 2.2 come in.

Related Articles

Search on MC.AI