Google Professional Machine Learning Engineer Question 269
Select 3Google Cloud PlatformYou are training a machine learning model on Google Cloud AI Platform and encounter a training failure with the error message: 'Out of memory error on GPU: Device 0'. What steps should you take to resolve this issue?
- A
Reduce the batch size in your training configuration.
- B
Switch to a higher memory GPU, such as from NVIDIA Tesla T4 to NVIDIA Tesla V100.
- C
Increase the learning rate to accelerate convergence and reduce memory usage.
- D
Enable mixed precision training to optimize memory usage on the GPU.
- E
Switch to training on CPUs instead of GPUs to avoid memory issues.
Show answer and explanation
Correct answers: A, B, D
Explanation
Out-of-memory errors on GPUs during model training are often caused by using large batch sizes, insufficient GPU memory, or inefficient memory utilization. Reducing the batch size directly lowers memory requirements. Upgrading to a higher memory GPU provides the necessary resources for larger models or datasets. Mixed precision training is an optimization technique that reduces memory usage while maintaining model accuracy. Increasing the learning rate or switching to CPUs, however, are not effective solutions to address GPU memory errors.
- A. Correct.
Reducing the batch size decreases the amount of data being processed in one forward and backward pass, which reduces memory usage on the GPU.
- B. Correct.
Switching to a higher memory GPU provides more resources, which can address out-of-memory errors in training.
- C. Incorrect.
Increasing the learning rate does not reduce memory usage and can lead to unstable training or convergence issues.
- D. Correct.
Mixed precision training reduces memory usage by using lower precision (e.g., float16) for calculations, which is particularly effective on GPUs.
- E. Incorrect.
Switching to CPUs will not necessarily solve the issue and may significantly slow down the training process.