MLS-C01 Question 378
Select 3You have deployed a machine learning model on Amazon SageMaker, but you notice that the predictions are inconsistent with the expected results. Which of the following steps should you take to debug and troubleshoot the issue? (Select TWO.)
- A
Enable Amazon SageMaker Debugger to monitor model training and analyze detailed metrics.
- B
Examine the input features for the deployed model using Amazon SageMaker Model Monitor to ensure they match the training data distribution.
- C
Immediately increase the instance type used for the endpoint to improve performance.
- D
Retrain the model with a smaller dataset to improve generalization.
- E
Check the endpoint logs in Amazon CloudWatch for errors or anomalies during inference.
Show answer and explanation
Correct answers: A, B, E
Explanation
When debugging a machine learning model deployed on Amazon SageMaker, it is essential to identify the root cause of inconsistencies. Tools like Amazon SageMaker Debugger and Model Monitor help analyze training metrics and detect data drift, while CloudWatch logs provide insight into runtime errors or anomalies during inference. Simply increasing the instance type or retraining the model without understanding the issue is not a recommended approach.
- A. Correct.
Correct. Amazon SageMaker Debugger can help analyze training metrics and detect potential training issues like vanishing gradients or overfitting.
- B. Correct.
Correct. Amazon SageMaker Model Monitor can detect data drift by comparing inference data to the training data distribution, which is critical for debugging issues caused by feature mismatch.
- C. Incorrect.
Incorrect. Increasing the instance type may improve performance but will not address the root cause of inconsistent predictions.
- D. Incorrect.
Incorrect. Retraining with a smaller dataset is unlikely to resolve issues related to debugging or troubleshooting and may lead to underfitting.
- E. Correct.
Correct. CloudWatch logs are essential for identifying runtime errors or anomalies during the inference process, which can help debug the deployed model.