⚡ Free Classes and Scholarships Available for Underprivileged Students -

Model prototyping using Jupyter notebooks

Collaborating within and across teams to manage data and models

10 Questions
No time limit
Practice Mode
0%
Score
0
Correct
0
Incorrect
10
Total Questions
Back to Topics
Question 1 of 10
Your data science team is prototyping a computer vision model in Vertex AI Workbench and wants to leverage a pre-trained foundation model from Model Garden as a starting point. They need to fine-tune the model on a custom dataset of 50,000 labeled images. What is the most efficient approach?
Explanation
Option C (downloading the foundation model from Model Garden into the notebook and fine-tuning it using transfer learning) is correct because Model Garden provides pre-trained models that can be directly accessed from notebooks, and transfer learning is the standard efficient approach for adapting foundation models to custom datasets. This leverages the pre-trained weights while customizing for the specific use case. Option B is incorrect because deploying directly to an endpoint doesn't address the requirement to fine-tune on custom data. Option D is incorrect because training from scratch is inefficient when a suitable foundation model exists, wasting computational resources and time. Option A is incorrect because manually rebuilding the architecture is unnecessarily complex and error-prone when the model can be used directly.
Question 2 of 10
You are developing a natural language processing model in a Vertex AI Workbench notebook using JAX and need GPU acceleration. The notebook instance was created without GPU support. What should you do?
Explanation
Option A (creating a new notebook instance with GPU-enabled machine type and migrating your code) is correct because Vertex AI Workbench instances cannot have their machine type changed after creation to add GPUs. You must create a new instance with the desired GPU configuration. Option C is incorrect because you cannot modify an existing notebook instance's machine type to add GPUs after it has been created; this limitation exists in Vertex AI Workbench. Option B is incorrect because while you can install CUDA drivers, the underlying instance lacks the physical GPU hardware, making this ineffective. Option D is incorrect because there is no setting to enable GPU acceleration on an instance that was provisioned without GPU hardware.
Question 3 of 10
Your team uses Colab Enterprise for prototyping ML models and needs to ensure that all notebook executions use consistent Python package versions across team members. The project uses TensorFlow, Pandas, and several custom internal libraries. What is the recommended approach?
Explanation
Option D (building a custom container image with all dependencies and configuring it as the notebook runtime) is correct because Colab Enterprise supports custom container images as runtimes, ensuring complete environment consistency including system-level dependencies, Python packages, and custom libraries. This provides reproducibility across all team members. Option A is incorrect because while requirements.txt helps, it doesn't guarantee consistency of system-level dependencies or custom internal libraries, and manual installation at startup adds overhead. Option C is incorrect because manual documentation is error-prone and doesn't enforce consistency. Option B is incorrect because it lacks any version control or consistency mechanism and is an inefficient collaboration approach.
Question 4 of 10
You are prototyping a scikit-learn model in Vertex AI Workbench and want to experiment with different feature engineering approaches. You need to track which preprocessing steps produce the best model performance. Which approach best supports this experimentation workflow?
Explanation
Option B (using Vertex AI Experiments SDK to log preprocessing parameters and resulting model metrics) is correct because the Experiments SDK is specifically designed for tracking ML experimentation, allowing you to systematically log parameters, metrics, and artifacts from notebooks. It provides structured comparison and visualization of different approaches. Option D is incorrect because manually named files lack structure, are difficult to query, and don't capture the relationship between preprocessing parameters and outcomes systematically. Option A is incorrect because multiple notebook copies create maintenance overhead, version control issues, and make systematic comparison difficult. Option C is incorrect because while BigQuery can store results, it doesn't provide the ML-specific tracking, lineage, and comparison capabilities that Vertex AI Experiments offers.
Question 5 of 10
Your team is developing ML models in Vertex AI Workbench and needs to implement a workflow where notebook code is regularly committed to a Git repository. Some team members are unfamiliar with Git commands. What is the most user-friendly approach that follows best practices?
Explanation
Option C (using the built-in Git extension in JupyterLab interface to commit and push changes) is correct because Vertex AI Workbench provides JupyterLab with an integrated Git extension that offers a graphical interface for Git operations, making it accessible to users unfamiliar with command-line Git while following proper version control practices. Option A is incorrect because automatic hourly commits don't allow for meaningful commit messages, proper code review, or intentional checkpointing, violating Git best practices. Option D is incorrect because manual file copying is inefficient, error-prone, and defeats the purpose of integrated development workflows. Option B is incorrect because converting to Python scripts adds unnecessary complexity and loses the interactive notebook format that is valuable for ML development.
Question 6 of 10
You need to prototype a PyTorch model using a dataset that is 800GB and stored in Cloud Storage. The data must be preprocessed before training. Your notebook environment should minimize data transfer time and cost. Which Vertex AI Workbench configuration is most appropriate?
Explanation
Option D (creating a notebook instance in the same region as the Cloud Storage bucket) is correct because co-locating compute and storage in the same region minimizes network latency, eliminates cross-region data transfer costs, and provides the fastest access to data. For an 800GB dataset, this is crucial for both performance and cost. Option B is incorrect because while network bandwidth matters, cross-region transfers still incur costs and latency regardless of bandwidth settings. Option A is incorrect because using a different region actually increases latency and incurs egress charges for data transfer. Option C is incorrect because Cloud Storage FUSE doesn't fundamentally change the network transfer requirements and can actually add overhead; region co-location is the primary factor.
Question 7 of 10
Your organization requires that Vertex AI Workbench notebooks use service accounts with minimal permissions rather than user credentials. You need to configure a notebook to read data from BigQuery and write model artifacts to Cloud Storage. What is the recommended approach?
Explanation
Option A (configuring the notebook instance with a custom service account that has BigQuery Data Viewer and Storage Object Admin roles) is correct because it follows the principle of least privilege by granting only the specific permissions needed for the stated requirements, and service accounts are the recommended approach for workload identity. Option B is incorrect because the default service account with owner permissions violates the principle of least privilege and creates unnecessary security risk. Option D is incorrect because the requirement specifically states using service accounts rather than user credentials. Option C is incorrect because API keys are less secure than service accounts, don't support fine-grained IAM permissions, and storing credentials in environment variables is a security anti-pattern.
Question 8 of 10
You are prototyping a recommendation model in Vertex AI Workbench using TensorFlow and need to access data from a private BigQuery dataset that contains customer information. The notebook instance must comply with data governance policies that prevent data exfiltration. Which configuration should you implement?
Explanation
Option B (enabling Private Google Access and disabling external IP addresses on the notebook instance) is correct because this configuration ensures the notebook can access Google Cloud services like BigQuery through Google's internal network while preventing any internet access that could lead to data exfiltration. This directly addresses the data governance requirement. Option A is incorrect because enabling internet access, even with VPC Service Controls, creates potential exfiltration paths and contradicts the security requirement. Option C is incorrect because Cloud NAT provides outbound internet connectivity, which is unnecessary and increases exfiltration risk when working with sensitive data. Option D is incorrect because binary authorization is for container image verification, not for preventing data exfiltration from notebooks.
Question 9 of 10
Your team wants to prototype multiple models using different frameworks (TensorFlow, PyTorch, and Scikit-learn) in Vertex AI Workbench. Each framework requires specific package versions that may conflict. What is the most efficient way to manage these environments?
Explanation
Option D (using Python virtual environments within a single notebook and activating them per framework) is correct because virtual environments isolate package dependencies within the same notebook instance, allowing you to switch between frameworks efficiently without the overhead of multiple instances or constant reinstallation. This is a standard Python best practice. Option B is incorrect because creating separate instances for each framework is unnecessarily costly and adds management overhead when virtual environments can handle isolation. Option C is incorrect because attempting to install conflicting package versions in a single environment will cause errors and unpredictable behavior. Option A is incorrect because reinstalling packages for each notebook is inefficient, time-consuming, and doesn't persist environments properly.
Question 10 of 10
You are experimenting with a large language model from Model Garden in Colab Enterprise. The model requires significant memory and you need to prototype prompt engineering strategies. The model takes several minutes to load. How should you optimize your workflow?
Explanation
Option B (loading the model once at the start and reusing the loaded instance across experiments) is correct because prompt engineering doesn't require reloading the model between variations, and keeping the model in memory eliminates the multi-minute loading overhead for each experiment. This is the standard efficient practice for prompt experimentation. Option C is incorrect because large language models maintain no state between different prompts, so reloading is wasteful and adds unnecessary minutes to each iteration without any benefit. Option A is incorrect because using a smaller model changes the experimental conditions and defeats the purpose of prototyping with the actual target model; results won't transfer. Option D is incorrect because selectively loading layers is not a standard practice for LLMs and would break the model's functionality for inference.