Google Professional Machine Learning Engineer Question 107
Select 2Google Cloud PlatformYou are training a machine learning model on Google Cloud AI Platform that uses a large dataset stored in Cloud Storage. The training process is taking significantly longer than expected. You want to optimize the training for efficiency while minimizing cost. Which steps should you take?
- A
Use preemptible VMs for training to reduce cost while maintaining performance.
- B
Enable GPU or TPU accelerators to speed up the training process.
- C
Reduce the dataset size by randomly sampling a subset of the data.
- D
Use sharded datasets and TFRecord format to optimize data input pipelines.
- E
Train the model sequentially on a single VM to simplify the setup.
Show answer and explanation
Correct answers: B, D
Explanation
To train a model efficiently on Google Cloud, it's essential to optimize both the computation and data input pipelines. GPUs or TPUs can accelerate computation by leveraging hardware specialized for machine learning workloads. Additionally, using sharded datasets and TFRecord format ensures efficient data loading and preprocessing. These techniques help maximize resource utilization, reduce training time, and improve overall efficiency.
- A. Incorrect.
Preemptible VMs can reduce cost but may interrupt training jobs unexpectedly, leading to inefficiencies for long-running training tasks.
- B. Correct.
Using GPUs or TPUs can significantly accelerate the training process by leveraging specialized hardware for computation.
- C. Incorrect.
Reducing the dataset size may introduce bias or lead to underfitting if the sampled data does not represent the full dataset adequately.
- D. Correct.
Sharded datasets and the TFRecord format improve the performance of data loading pipelines, which is often a bottleneck in training large models with large datasets.
- E. Incorrect.
Training the model sequentially on a single VM is not efficient for large datasets and complex models, as it does not take advantage of distributed training or specialized hardware like GPUs or TPUs.