Google Professional Machine Learning Engineer Question 272
Select 3Google Cloud PlatformYou are training a deep learning model on Google Cloud AI Platform (Vertex AI), but the training job keeps failing after running for a few minutes. Upon reviewing the logs, you notice an 'Out of Memory' (OOM) error. What steps should you take to troubleshoot and resolve this issue?
- A
Reduce the batch size used during training.
- B
Use a larger machine type with more memory for the training job.
- C
Enable preemptible instances to reduce costs during training.
- D
Optimize the model architecture by reducing the number of parameters or layers.
- E
Switch to an on-premise training setup for better memory management.
Show answer and explanation
Correct answers: A, B, D
Explanation
OOM errors during training typically occur when the model or batch size exceeds the available memory of the machine used. To resolve this, you can reduce the batch size, use a machine type with more memory, or optimize the model to reduce its memory requirements. Enabling preemptible instances or switching to an on-premise setup does not directly address the core issue of memory constraints.
- A. Correct.
Reducing the batch size lowers the memory requirements during training, which can help resolve OOM errors.
- B. Correct.
Using a larger machine type with more memory can address the OOM issue by providing sufficient resources for the training job.
- C. Incorrect.
Enabling preemptible instances is cost-effective but does not directly address the OOM error.
- D. Correct.
Optimizing the model architecture by reducing parameters or layers can decrease memory usage, making it a viable solution for OOM errors.
- E. Incorrect.
Switching to an on-premise training setup does not specifically resolve the OOM issue and may introduce additional complexity.