Google Professional Machine Learning Engineer Question 373
Single answerGoogle Cloud PlatformYou are building a machine learning model on Google Cloud Platform. During training, you apply several pre-processing steps, including normalization and feature encoding. To ensure that the model performs consistently when deployed for serving predictions, what is the best approach to maintain consistency between training and serving pre-processing steps?
- A
Use TensorFlow Transform (tf.Transform) to define a shared pre-processing pipeline for both training and serving.
- B
Manually write separate pre-processing scripts for training and serving, ensuring they both implement the same logic.
- C
Outsource the pre-processing step during serving to a downstream system and use the same approach during training.
- D
Store pre-processed training data in Google Cloud Storage (GCS) and skip pre-processing entirely during serving.
Show answer and explanation
Correct answer: A
Explanation
Ensuring consistent data pre-processing between training and serving is critical for the reliability of machine learning models. TensorFlow Transform (tf.Transform) is a recommended tool that enables a shared definition of pre-processing steps, which can be executed in both training and serving environments. This eliminates the risk of discrepancies and ensures that the model operates on data in the same format during both stages.
- A. Correct.
Correct. TensorFlow Transform (tf.Transform) allows you to define pre-processing logic that can be executed consistently both during training and at serving time, ensuring no discrepancies between the two stages.
- B. Incorrect.
Incorrect. Manually writing scripts for training and serving increases the likelihood of inconsistencies or errors in implementation, which can negatively impact model performance.
- C. Incorrect.
Incorrect. Outsourcing pre-processing to a downstream system introduces dependencies and potential inconsistencies, as the downstream system may not replicate the exact pre-processing logic used during training.
- D. Incorrect.
Incorrect. Pre-processing during training only and skipping it during serving is risky, as the serving pipeline may encounter data that differs from the pre-processed training data, leading to unexpected results.