A company trains a CNN image model daily in File mode on a single On-Demand EC2 instance with one epoch. They add about 10,000 new images each day and want to speed training and reduce costs without modifying code. Which change will best meet that goal?

Using Spot Instances significantly reduces costs compared to On-Demand instances, as they leverage unused EC2 capacity. Since Spot Instances can be interrupted, enabling model checkpointing is crucial. Checkpointing regularly saves the model's state, allowing training to resume from the last saved point if an interruption occurs, thus preventing loss of progress and ensuring training completes without starting over. This combination directly addresses both cost reduction and speed (by allowing training to complete despite interruptions) without requiring code changes. Switching to Pipe mode or FastFile mode might improve training speed by streaming data, but they don't directly address cost reduction. Replacing On-Demand with Spot Instances alone would reduce costs but risks frequent interruptions, potentially increasing overall training time if checkpoints are not used.