Google Professional Machine Learning Engineer Question 88
Single answerGoogle Cloud PlatformYou have deployed a TensorFlow model to Google Cloud AI Platform (Vertex AI) for online predictions, but the model is returning incorrect predictions. Upon inspecting the logs, you observe a mismatch between the input feature names used during training and the feature names in the incoming prediction requests. What is the best next step to resolve this issue?
- A
Update the client application to ensure that the prediction requests use the correct feature names as expected by the model.
- B
Retrain the model using the feature names provided in the prediction requests.
- C
Enable AutoML to automatically adjust the model to accept any feature names.
- D
Update the deployed model's configuration to map the incoming feature names to the expected feature names.
Show answer and explanation
Correct answer: A
Explanation
The root cause of the issue is a mismatch between the feature names used during training and those in the incoming prediction requests. The best way to resolve this is to ensure that the client application sends prediction requests with the correct feature names, as expected by the model. This avoids unnecessary retraining or reconfiguration.
- A. Correct.
This is correct because ensuring that the client application sends prediction requests with the correct feature names, as used during training, resolves the mismatch without requiring further retraining or reconfiguration.
- B. Incorrect.
This is incorrect because retraining the model is unnecessary for resolving a feature name mismatch. The issue can be fixed by aligning the prediction request feature names with the model's expected names.
- C. Incorrect.
This is incorrect because AutoML does not offer functionality to automatically adjust a deployed TensorFlow model to accept any feature names.
- D. Incorrect.
This is incorrect because Google Cloud AI Platform (Vertex AI) does not provide a built-in feature to map incoming feature names dynamically to the expected feature names for a deployed model.