MLS-C01 · Question #328
A machine learning (ML) specialist is using the Amazon SageMaker DeepAR forecasting algorithm to train a model on CPU-based Amazon EC2 On-Demand instances. The model currently takes multiple hours…
The correct answer is C. Replace CPU-based EC2 instances with GPU-based EC2 instances. D. Use multiple training instances. To decrease the training time of a DeepAR forecasting model, the ML specialist should leverage the computational power of GPUs and distribute the training workload across multiple instances.
Question
A machine learning (ML) specialist is using the Amazon SageMaker DeepAR forecasting algorithm to train a model on CPU-based Amazon EC2 On-Demand instances. The model currently takes multiple hours to train. The ML specialist wants to decrease the training time of the model. Which approaches will meet this requirement? (Choose two.)
Options
- AReplace On-Demand Instances with Spot Instances.
- BConfigure model auto scaling dynamically to adjust the number of instances automatically.
- CReplace CPU-based EC2 instances with GPU-based EC2 instances.
- DUse multiple training instances.
- EUse a pre-trained version of the model. Run incremental training.
How the community answered
(44 responses)- A5% (2)
- B9% (4)
- C82% (36)
- E5% (2)
Why each option
To decrease the training time of a DeepAR forecasting model, the ML specialist should leverage the computational power of GPUs and distribute the training workload across multiple instances.
Replacing On-Demand Instances with Spot Instances reduces cost but does not inherently decrease the training duration of the model.
Configuring model auto scaling is for dynamically adjusting the number of instances for inference endpoints based on traffic, not for speeding up a training job.
DeepAR is a neural network-based algorithm that can significantly benefit from GPU acceleration, as GPUs are highly efficient for parallelizing the intensive matrix operations involved in deep learning model training, thereby reducing training time.
Using multiple training instances enables distributed training, where the workload is split across several machines, allowing the DeepAR model to train faster by processing data in parallel.
While incremental training with a pre-trained model can reduce the *effective* time to achieve a good model, it's not a direct method to decrease the computational duration of a fresh training run for the given model.
Concept tested: Optimizing SageMaker training performance
Source: https://docs.aws.amazon.com/sagemaker/latest/dg/deepar.html
Topics
Community Discussion
No community discussion yet for this question.