Or, if files are supplied, always read them?
That way we have actual and predicted from the validation and can compute confusion matrices etc.
Any thoughts, @SelimRABOUDI and @jongbinjung? See what I committed which is something I started to noodle with yesterday. Maybe more to come...