Checkpoints are saved while training, but the implementation doesn't offer an option to restart training from them.