Keras multi-class average recall (aka. unweighted accuracy) metric.

ua.py

class MulticlassAverageRecall(Layer):
    def __init__(self, name='multiclass_recall', classes=4,
                 output_idx=0, **kwargs):
        super(MulticlassAverageRecall, self).__init__(name=name, **kwargs)
        self.stateful = True
        self.classes = classes

        self.output_idx = output_idx

        self.t = K.variable(np.zeros(classes), dtype='int32')
        self.p = K.variable(np.zeros(classes), dtype='int32')

    def reset_states(self):
        K.set_value(self.t, np.zeros(self.classes))
        K.set_value(self.p, np.zeros(self.classes))

    def __call__(self, y_true, y_pred):
        # Init a bias matrix
        b = K.variable([1 / (v + 1) for v in range(self.classes)],
                       dtype=K.floatx())

        # Simulate to_categorical opeation
        x, y = K.argmax(y_pred, axis=-1), K.argmax(y_true, axis=-1)
        x, y = K.cast(x, K.floatx()), K.cast(y, K.floatx())
        x, y = K.expand_dims(x, axis=-1), K.expand_dims(y, axis=-1)
        x, y = (x + 1) * b - 1, (y + 1) * b - 1

        # Make correct position filled with 1
        x, y = K.cast(x, 'bool'), K.cast(y, 'bool')
        x, y = 1 - K.cast(x, 'int32'), 1 - K.cast(y, 'int32')

        x, y = K.transpose(x), K.transpose(y)

        t = K.sum(y, axis=-1)
        p = K.sum(x * y, axis=-1)

        current_t = self.t * 1
        current_p = self.p * 1

        self.add_update(K.update_add(self.t, t))
        self.add_update(K.update_add(self.p, p))

        return K.mean((current_p + p) / (current_t + t))