Google Professional Machine Learning Engineer Question 87
Select 3Google Cloud PlatformYou have deployed a trained image classification model on Google Cloud's AI Platform Prediction service. However, during testing, you notice that the model consistently produces incorrect predictions for certain types of images. Upon further investigation, you find that the issue is related to the input image size and preprocessing. How should you resolve this issue?
- A
Update the model to handle varying input sizes by retraining it on augmented data.
- B
Configure a custom prediction routine to resize and preprocess the input images before passing them to the model.
- C
Use AI Platform Prediction's built-in logging feature to analyze input-output pairs and identify preprocessing issues.
- D
Modify the deployed model directly on AI Platform to accept any input size without retraining.
- E
Implement a preprocessing step in the client application to ensure input images match the model's expected format.
Show answer and explanation
Correct answers: A, B, E
Explanation
The problem arises due to mismatched input image sizes, which causes the model to produce incorrect predictions. The best approaches to resolve this issue include retraining the model to handle varying input sizes (1), configuring a custom prediction routine to preprocess inputs on the server side (2), and ensuring inputs are preprocessed correctly on the client side (5). Each of these approaches addresses the root cause of the problem while adhering to best practices for deploying and maintaining machine learning models on Google Cloud.
- A. Correct.
Updating the model to handle varying input sizes by retraining it on augmented data is a valid approach to ensure the model can generalize better for different types of inputs. This is especially helpful if the input variance is common in production scenarios.
- B. Correct.
Configuring a custom prediction routine allows you to preprocess the input images (e.g., resizing, normalizing) directly on the server before the model receives them. This ensures consistency in input format and avoids the need to modify the client application.
- C. Incorrect.
Using AI Platform Prediction's logging feature can help with debugging and identifying preprocessing issues, but it does not directly resolve the issue of incorrect predictions caused by input size mismatches.
- D. Incorrect.
Modifying the deployed model directly on AI Platform is not possible because model architecture or training cannot be altered after deployment. The model needs to be updated and redeployed if changes are required.
- E. Correct.
Adding a preprocessing step in the client application ensures that input images are resized or transformed to match the model’s expected input format. This is an effective and lightweight solution in some cases.