I write my own loss function (it is the standard mean absolute error):
def my_mae(y_true, y_pred): return K.max(K.abs(y_true - y_pred), axis=-1, keepdims=True