Model Prototyping Using Notebooks (e.g., Gemini Enterprise Agent Platform Workbench and Colab Enterprise) PROFESSIONAL-MACHINE-LEARNING-ENGINEER Practice Questions (Page 3)

Questions 11–15

1. 11application · medium

   A data scientist is prototyping a PyTorch image classification model in a Workbench notebook. The prototype works well on a small sample, but training on the full dataset takes too long on the current instance. The data scientist wants to scale the training to a larger compute resource without leaving the notebook environment. What should the data scientist do?

   • AManually copy the training script to a new, larger Workbench instance and run it there, monitoring the training progress in the notebook.
   • BUse the Workbench instance's 'Change machine type' option to resize the instance to one with more GPUs, then rerun the notebook.
   • CConvert the PyTorch model to TensorFlow and use a Vertex AI TensorFlow training job, as Vertex AI does not support PyTorch.
   • DSubmit a custom training job to Vertex AI Training from the notebook, using a machine type with more GPUs and the pre-built PyTorch container.
   Reveal Solution Discussion Select an answer first
2. 12expert · hard

   A company wants to prototype a question-answering application. They have a small labeled dataset for fine-tuning. The team is considering using a large foundational model from Model Garden. The primary constraint is to minimize the cost of the prototyping phase while still achieving good performance. What is the most cost-effective approach?

   • AUse a large foundational model from Model Garden via the Vertex AI Prediction API and provide the labeled dataset as few-shot examples in the prompt.
   • BUse the largest foundational model from Model Garden and fine-tune it using Vertex AI Training with multiple GPUs.
   • CUse a large foundational model from Model Garden and fine-tune it using a Colab Enterprise notebook with a CPU-only runtime.
   • DUse a smaller, open-source model from Model Garden and fine-tune it on the labeled dataset in a Workbench notebook with a single GPU.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A data scientist is prototyping a sentiment analysis model in Colab Enterprise. They found a pre-trained BERT model in Model Garden and want to fine-tune it on a custom dataset. The dataset is small, and the data scientist wants to use a GPU for faster training. What is the correct approach?

   • AUse the Vertex AI Prediction API to fine-tune the model by sending the custom dataset as input to the API.
   • BUse the Model Garden 'Deploy to endpoint' option to create a Vertex AI endpoint, then use the endpoint to train the model.
   • CDownload the model weights from Model Garden and manually write all the code to load the model, create the dataset, and train it.
   • DLoad the pre-trained model from Model Garden using the provided sample code, then use a PyTorch or TensorFlow training loop with the custom dataset on a GPU runtime.
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A machine learning engineer wants to prototype a text generation application in a Workbench notebook. They need to quickly test a few different open-source large language models without writing custom model loading code. What is the most efficient way to explore and load these models?

   • ASearch Hugging Face for open-source models and manually download the model weights and tokenizer files into the notebook environment.
   • BUse the Vertex AI Prediction API to call a foundational model directly from the notebook, without loading the model locally.
   • CUse Model Garden to browse and discover suitable models, then use the 'Get sample code' or 'Open in Workbench' option to load the model in the notebook.
   • DCreate a new Workbench instance with a deep learning VM image that includes all open-source models pre-installed.
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A data scientist has a PyTorch training script in a Workbench notebook that runs successfully on a small dataset. They need to train on a much larger dataset that is stored in a Cloud Storage bucket. The training will take several hours. The data scientist wants to avoid keeping the notebook instance running for the entire duration and wants to use a more cost-effective approach. What should they do?

   • AConvert the PyTorch script to TensorFlow and use a Vertex AI TensorFlow training job, as Vertex AI does not support long-running PyTorch jobs.
   • BKeep the Workbench instance running and use a screen session to run the training script in the background, checking on it periodically.
   • CSubmit the training script as a Vertex AI custom training job, which provisions a managed compute resource only for the duration of the job.
   • DUse the Workbench instance's 'Stop' and 'Start' feature to pause the instance during training, resuming it when the training is complete.
   Reveal Solution Discussion Select an answer first