Has anyone been able to train the U-Net model using mixed precision? The “precision” argument for the lightning Trainer seems to have no effect.