PMLE practice questions 31–40

1. Question 31 · 364 in bank

   You are training an object detection model using a Cloud TPU v2. Training time is taking longer than expected. Based on this simplified trace obtained with a Cloud TPU profile, what action should you take to decrease training time in a cost-efficient way?

   

   • AMove from Cloud TPU v2 to Cloud TPU v3 and increase batch size.
   • BMove from Cloud TPU v2 to 8 NVIDIA V100 GPUs and increase batch size.
   • CRewrite your input function to resize and reshape the input images.
   • DRewrite your input function using parallel reads, parallel processing, and prefetch.

   Show answer

   Answer: D

   Explanation

   Optimizing the data pipeline with parallel reads, processing, and prefetching can significantly improve training speed on TPUs by reducing I/O wait times. This approach utilizes the TPU's capabilities more effectively and avoids extra costs associated with hardware upgrades.

   Source: Adapted from community discussion — fitri001

   References

   • Google Cloud documentation

   Next question

2. Question 32 · 364 in bank

   While performing exploratory data analysis on a dataset, you find that an important categorical feature has 5% null values. You want to minimize the bias that could result from the missing values. How should you handle the missing values?

   • ARemove the rows with missing values, and upsample your dataset by 5%.
   • BReplace the missing values with the feature’s mean.
   • CReplace the missing values with a placeholder category indicating a missing value.
   • DMove the rows with missing values to your validation dataset.

   Show answer

   Answer: C

   Explanation

   Minimizes Bias: Removing rows (A) with missing data can introduce bias if the missingness is not random.expand_more Upsampling the remaining data (A) might not address the underlying cause of missing values. Unsuitable for Categorical Features: Replacing with the mean (B) only works for numerical features. Transparency and Model Interpretation: A placeholder category (C) explicitly acknowledges the missing data and avoids introducing assumptions during model training. It also improves model interpretability. Validation Set Contamination (D): Moving rows with missing values to the validation set (D) contaminates the validation data and hinders its ability to assess model performance on unseen data. Using a placeholder category creates a separate category for missing values, allowing the model to handle them explicitly. This approach is particularly suitable for categorical features with a relatively small percentage of missing values (like 5% in this case).

   Source: Adapted from community discussion — fitri001

   References

   • Google Cloud documentation

   Next question

3. Question 33 · 364 in bank

   You are an ML engineer on an agricultural research team working on a crop disease detection tool to detect leaf rust spots in images of crops to determine the presence of a disease. These spots, which can vary in shape and size, are correlated to the severity of the disease. You want to develop a solution that predicts the presence and severity of the disease with high accuracy. What should you do?

   • ACreate an object detection model that can localize the rust spots.
   • BDevelop an image segmentation ML model to locate the boundaries of the rust spots.
   • CDevelop a template matching algorithm using traditional computer vision libraries.
   • DDevelop an image classification ML model to predict the presence of the disease.

   Show answer

   Answer: B

   Explanation

   B. Develop an image segmentation ML model to locate the boundaries of the rust spots. An image segmentation model is well-suited for this task because it can identify the exact location and shape of the rust spots in the image, which is critical for determining the severity of the disease. Once the rust spots have been identified, other algorithms can be used to analyze the data and predict the severity of the disease. Object detection models are another option, but they may not be as accurate as image segmentation models when it comes to identifying the exact boundaries of the rust spots. Template matching algorithms using traditional computer vision libraries are generally not as accurate as ML models when it comes to image analysis.

   Source: Adapted from community discussion — TNT87

   Next question

4. Question 34 · 364 in bank

   You have been asked to productionize a proof-of-concept ML model built using Keras. The model was trained in a Jupyter notebook on a data scientist’s local machine. The notebook contains a cell that performs data validation and a cell that performs model analysis. You need to orchestrate the steps contained in the notebook and automate the execution of these steps for weekly retraining. You expect much more training data in the future. You want your solution to take advantage of managed services while minimizing cost. What should you do?

   • AMove the Jupyter notebook to a Notebooks instance on the largest N2 machine type, and schedule the execution of the steps in the Notebooks instance using Cloud Scheduler.
   • BWrite the code as a TensorFlow Extended (TFX) pipeline orchestrated with Vertex AI Pipelines. Use standard TFX components for data validation and model analysis, and use Vertex AI Pipelines for model retraining.
   • CRewrite the steps in the Jupyter notebook as an Apache Spark job, and schedule the execution of the job on ephemeral Dataproc clusters using Cloud Scheduler.
   • DExtract the steps contained in the Jupyter notebook as Python scripts, wrap each script in an Apache Airflow BashOperator, and run the resulting directed acyclic graph (DAG) in Cloud Composer.

   Show answer

   Answer: B

   Explanation

   I believe it B. Write the code as a TensorFlow Extended (TFX) pipeline orchestrated with Vertex AI Pipelines. Use standard TFX components for data validation and model analysis, and use Vertex AI Pipelines for model retraining. Because : - Solution A is not scalable and will be expensive to run. It also does not take advantage of managed services. Solution C is more scalable than option A, but it is still not as scalable as using TFX and Vertex AI Pipelines. It also does not take advantage of managed services. - Solution D is the most flexible, but it is also the most complex. It requires more knowledge of Apache Airflow and is more difficult to manage. Overall, the best solution to productionize the proof-of-concept ML model is to use TFX and Vertex AI Pipelines. This solution is scalable, reliable, and easy to manage. It also takes advantage of managed services, which can help to reduce costs.

   Source: Adapted from community discussion — Antmal

   Next question

5. Question 35 · 364 in bank

   You have trained a deep neural network model on Google Cloud. The model has low loss on the training data, but is performing worse on the validation data. You want the model to be resilient to overfitting. Which strategy should you use when retraining the model?

   • AApply a dropout parameter of 0.2, and decrease the learning rate by a factor of 10.
   • BApply a L2 regularization parameter of 0.4, and decrease the learning rate by a factor of 10.
   • CRun a hyperparameter tuning job on AI Platform to optimize for the L2 regularization and dropout parameters.
   • DRun a hyperparameter tuning job on AI Platform to optimize for the learning rate, and increase the number of neurons by a factor of 2.

   Show answer

   Answer: C

   Explanation

   Req: make model resilient A. Apply a dropout parameter of 0.2, and decrease the learning rate by a factor of 10. : Might / might not work . But may not find optimal parameter set since it uses random values B. Apply a L2 regularization parameter of 0.4, and decrease the learning rate by a factor of 10. : Might / might not work . But may not find optimal parameter set since it uses random values C. Run a hyperparameter tuning job on AI Platform to optimize for the L2 regularization and dropout parameters. : l2 and dropout are regularisation method which would work. Let AI find the optimal solution on how extend these parameters should regularise. Yes this would work. D. Run a hyperparameter tuning job on AI Platform to optimize for the learning rate, and increase the number of neurons by a factor of 2 : AIplatform would do but adding neurons would make network nore complex. So we can eliminate this option.

   Source: Adapted from community discussion — harithacML

   Next question

6. Question 36 · 364 in bank

   You are working on a system log anomaly detection model for a cybersecurity organization. You have developed the model using TensorFlow, and you plan to use it for real-time prediction. You need to create a Dataflow pipeline to ingest data via Pub/Sub and write the results to BigQuery. You want to minimize the serving latency as much as possible. What should you do?

   • AContainerize the model prediction logic in Cloud Run, which is invoked by Dataflow.
   • BLoad the model directly into the Dataflow job as a dependency, and use it for prediction.
   • CDeploy the model to a Vertex AI endpoint, and invoke this endpoint in the Dataflow job.
   • DDeploy the model in a TFServing container on Google Kubernetes Engine, and invoke it in the Dataflow job.

   Show answer

   Answer: C

   Explanation

   C. According Google: "Instead of deploying the model to an endpoint, you can use the RunInference API to serve machine learning models in your Apache Beam pipeline. This approach has several advantages, including flexibility and portability. However, deploying the model in Vertex AI offers many additional benefits, such as the platform's built-in tools for model monitoring, TensorBoard, and model registry governance. Vertex AI also provides the ability to use Optimized TensorFlow runtime in your endpoints. To do this, simply specify the TensorFlow runtime container when you deploy your model." https://cloud.google.com/blog/products/ai-machine-learning/streaming-prediction-with-dataflow-and-vertex

   Source: Adapted from community discussion — guilhermebutzke

   References

   • Google Cloud documentation
   • Google Cloud documentation
   • Google Cloud documentation
   • Google Cloud documentation
   • Google Cloud documentation

   Next question

7. Question 37 · 364 in bank

   You are an ML engineer at a mobile gaming company. A data scientist on your team recently trained a TensorFlow model, and you are responsible for deploying this model into a mobile application. You discover that the inference latency of the current model doesn’t meet production requirements. You need to reduce the inference time by 50%, and you are willing to accept a small decrease in model accuracy in order to reach the latency requirement. Without training a new model, which model optimization technique for reducing latency should you try first?

   • AWeight pruning
   • BDynamic range quantization
   • CModel distillation
   • DDimensionality reduction

   Show answer

   Answer: B

   Explanation

   B. Dynamic range quantization The reason for this choice is that dynamic range quantization is a model optimization technique that can significantly reduce model size and inference time while maintaining reasonable model accuracy. Dynamic range quantization uses fewer bits to represent the weights of the model, reducing the memory required to store the model and the time required for inference.

   Source: Adapted from community discussion — TNT87

   Next question

8. Question 38 · 364 in bank

   You work on a data science team at a bank and are creating an ML model to predict loan default risk. You have collected and cleaned hundreds of millions of records worth of training data in a BigQuery table, and you now want to develop and compare multiple models on this data using TensorFlow and Vertex AI. You want to minimize any bottlenecks during the data ingestion state while considering scalability. What should you do?

   • AUse the BigQuery client library to load data into a dataframe, and use tf.data.Dataset.from_tensor_slices() to read it.
   • BExport data to CSV files in Cloud Storage, and use tf.data.TextLineDataset() to read them.
   • CConvert the data into TFRecords, and use tf.data.TFRecordDataset() to read them.
   • DUse TensorFlow I/O’s BigQuery Reader to directly read the data.

   Show answer

   Answer: D

   Explanation

   Vote on D. This will allow to directly access the data from BigQuery without having to first load it into a dataframe or export it to files in Cloud Storage.

   Source: Adapted from community discussion — mil_spyro

   References

   • Google Cloud documentation

   Next question

9. Question 39 · 364 in bank

   You have recently created a proof-of-concept (POC) deep learning model. You are satisfied with the overall architecture, but you need to determine the value for a couple of hyperparameters. You want to perform hyperparameter tuning on Vertex AI to determine both the appropriate embedding dimension for a categorical feature used by your model and the optimal learning rate. You configure the following settings: • For the embedding dimension, you set the type to INTEGER with a minValue of 16 and maxValue of 64. • For the learning rate, you set the type to DOUBLE with a minValue of 10e-05 and maxValue of 10e-02. You are using the default Bayesian optimization tuning algorithm, and you want to maximize model accuracy. Training time is not a concern. How should you set the hyperparameter scaling for each hyperparameter and the maxParallelTrials?

   • AUse UNIT_LINEAR_SCALE for the embedding dimension, UNIT_LOG_SCALE for the learning rate, and a large number of parallel trials.
   • BUse UNIT_LINEAR_SCALE for the embedding dimension, UNIT_LOG_SCALE for the learning rate, and a small number of parallel trials.
   • CUse UNIT_LOG_SCALE for the embedding dimension, UNIT_LINEAR_SCALE for the learning rate, and a large number of parallel trials.
   • DUse UNIT_LOG_SCALE for the embedding dimension, UNIT_LINEAR_SCALE for the learning rate, and a small number of parallel trials.

   Show answer

   Answer: B

   Explanation

   B: Here's why: **Embedding Dimension:** UNIT_LINEAR_SCALE is appropriate for integer hyperparameters with a continuous range like the embedding dimension. It linearly scales the search space from minValue to maxValue. **Learning Rate:** UNIT_LOG_SCALE is generally recommended for hyperparameters with values spanning multiple orders of magnitude like the learning rate (10e-05 - 10e-02). This ensures equal sampling probability across different log-scaled ranges. **Parallel Trials:** as the documentation specifies, parallelization speeds up. However, this speedup comes at the cost of potentially sacrificing the quality of the results. Since training time is not a factor in this case, the benefit of speeding things up with many parallel trials is less valuable. https://cloud.google.com/vertex-ai/docs/training/using-hyperparameter-tuning#parallel-trials

   Source: Adapted from community discussion — guilhermebutzke

   References

   • Google Cloud documentation
   • Google Cloud documentation

   Next question

10. Question 40 · 364 in bank

    You are the Director of Data Science at a large company, and your Data Science team has recently begun using the Kubeflow Pipelines SDK to orchestrate their training pipelines. Your team is struggling to integrate their custom Python code into the Kubeflow Pipelines SDK. How should you instruct them to proceed in order to quickly integrate their code with the Kubeflow Pipelines SDK?

    • AUse the func_to_container_op function to create custom components from the Python code.
    • BUse the predefined components available in the Kubeflow Pipelines SDK to access Dataproc, and run the custom code there.
    • CPackage the custom Python code into Docker containers, and use the load_component_from_file function to import the containers into the pipeline.
    • DDeploy the custom Python code to Cloud Functions, and use Kubeflow Pipelines to trigger the Cloud Function.

    Show answer

    Answer: A

    Explanation

    A. Use the func_to_container_op function to create custom components from the Python code. The func_to_container_op function in the Kubeflow Pipelines SDK is specifically designed to convert Python functions into containerized components that can be executed in a Kubernetes cluster. By using this function, the Data Science team can easily integrate their custom Python code into the Kubeflow Pipelines SDK without having to learn the details of containerization or Kubernetes.

    Source: Adapted from community discussion — TNT87

    Next question

Keyboard:J Kmove between questionsA–Dselect an answerRshow or hide the answer← →change page

PMLE practice, question 31. 364 questions available.