PMLE practice questions 51–60

1. Question 51 · 364 in bank

   You are developing an image recognition model using PyTorch based on ResNet50 architecture. Your code is working fine on your local laptop on a small subsample. Your full dataset has 200k labeled images. You want to quickly scale your training workload while minimizing cost. You plan to use 4 V100 GPUs. What should you do?

   • ACreate a Google Kubernetes Engine cluster with a node pool that has 4 V100 GPUs. Prepare and submit a TFJob operator to this node pool.
   • BCreate a Vertex AI Workbench user-managed notebooks instance with 4 V100 GPUs, and use it to train your model.
   • CPackage your code with Setuptools, and use a pre-built container. Train your model with Vertex AI using a custom tier that contains the required GPUs.
   • DConfigure a Compute Engine VM with all the dependencies that launches the training. Train your model with Vertex AI using a custom tier that contains the required GPUs.

   Show answer

   Answer: C

   Explanation

   Custom trainer , don't overthink 1000%, this is google recommendation. you don't need Vertex AI Workbench user-managed notebooks,Google Kubernetes Engine, Compute Engine at at all , it is a waste of your effort https://cloud.google.com/vertex-ai/docs/training/configure-compute#specifying_gpus You can choose as your want

   Source: Adapted from community discussion — John_Pongthorn

   References

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

   Next question

2. Question 52 · 364 in bank

   You have trained a DNN regressor with TensorFlow to predict housing prices using a set of predictive features. Your default precision is tf.float64, and you use a standard TensorFlow estimator:

   

   Your model performs well, but just before deploying it to production, you discover that your current serving latency is 10ms @ 90 percentile and you currently serve on CPUs. Your production requirements expect a model latency of 8ms @ 90 percentile. You're willing to accept a small decrease in performance in order to reach the latency requirement. Therefore your plan is to improve latency while evaluating how much the model's prediction decreases. What should you first try to quickly lower the serving latency?

   • ASwitch from CPU to GPU serving.
   • BApply quantization to your SavedModel by reducing the floating point precision to tf.float16.
   • CIncrease the dropout rate to 0.8 and retrain your model.
   • DIncrease the dropout rate to 0.8 in _PREDICT mode by adjusting the TensorFlow Serving parameters.

   Show answer

   Answer: B

   Explanation

   Switching from CPU to GPU serving could also improve latency, but it may not be considered a "quick" solution compared to model quantization because it involves additional hardware requirements and potentially more complex deployment changes. Additionally, not all models see a latency improvement on GPUs, especially if the model is not large enough to utilize the GPU effectively or if the infrastructure does not support GPU optimizations. Therefore, the first thing to try would be quantization, which can be done relatively quickly and directly within the TensorFlow framework. After applying quantization, you should evaluate the model to ensure that the decrease in precision does not lead to an unacceptable drop in prediction accuracy.

   Source: Adapted from community discussion — Tayoso

   Next question

3. Question 53 · 364 in bank

   You work on the data science team at a manufacturing company. You are reviewing the company’s historical sales data, which has hundreds of millions of records. For your exploratory data analysis, you need to calculate descriptive statistics such as mean, median, and mode; conduct complex statistical tests for hypothesis testing; and plot variations of the features over time. You want to use as much of the sales data as possible in your analyses while minimizing computational resources. What should you do?

   • AVisualize the time plots in Google Data Studio. Import the dataset into Vertex Al Workbench user-managed notebooks. Use this data to calculate the descriptive statistics and run the statistical analyses.
   • BSpin up a Vertex Al Workbench user-managed notebooks instance and import the dataset. Use this data to create statistical and visual analyses.
   • CUse BigQuery to calculate the descriptive statistics. Use Vertex Al Workbench user-managed notebooks to visualize the time plots and run the statistical analyses.
   • DUse BigQuery to calculate the descriptive statistics, and use Google Data Studio to visualize the time plots. Use Vertex Al Workbench user-managed notebooks to run the statistical analyses.

   Show answer

   Answer: C

   Explanation

   C. Use BigQuery to calculate the descriptive statistics. Use Vertex AI Workbench user-managed notebooks to visualize the time plots and run the statistical analyses. BigQuery is a powerful data analysis tool that can handle massive datasets, making it an ideal solution for calculating descriptive statistics for hundreds of millions of records. It can also perform complex statistical tests for hypothesis testing. For time series analysis, using Vertex AI Workbench user-managed notebooks would be the best solution as it provides a flexible environment for data exploration, visualization, and statistical analysis. By using the two tools together, the data science team can efficiently analyze the sales data while minimizing computational resources. Its C not B

   Source: Adapted from community discussion — TNT87

   References

   • Google Cloud documentation
   • Google Cloud documentation

   Next question

4. Question 54 · 364 in bank

   Your data science team needs to rapidly experiment with various features, model architectures, and hyperparameters. They need to track the accuracy metrics for various experiments and use an API to query the metrics over time. What should they use to track and report their experiments while minimizing manual effort?

   • AUse Vertex Al Pipelines to execute the experiments. Query the results stored in MetadataStore using the Vertex Al API.
   • BUse Vertex Al Training to execute the experiments. Write the accuracy metrics to BigQuery, and query the results using the BigQuery API.
   • CUse Vertex Al Training to execute the experiments. Write the accuracy metrics to Cloud Monitoring, and query the results using the Monitoring API.
   • DUse Vertex Al Workbench user-managed notebooks to execute the experiments. Collect the results in a shared Google Sheets file, and query the results using the Google Sheets API.

   Show answer

   Answer: A

   Explanation

   Vertex AI Pipelines: Pipelines are designed for automating experiment execution. You can define different steps like data preprocessing, training with various configurations, and evaluation. This allows rapid experimentation with minimal manual intervention. Vertex ML Metadata: Vertex AI Pipelines integrate seamlessly with Vertex ML Metadata, which automatically tracks experiment runs, metrics, and artifacts. This eliminates the need for manual data collection in spreadsheets. Vertex AI API: The Vertex AI API allows you to programmatically query the Vertex ML Metadata store. You can retrieve experiment details, including accuracy metrics, for further analysis or visualization.

   Source: Adapted from community discussion — fitri001

   References

   • Google Cloud documentation

   Next question

5. Question 55 · 364 in bank

   You are training an ML model using data stored in BigQuery that contains several values that are considered Personally Identifiable Information (PII). You need to reduce the sensitivity of the dataset before training your model. Every column is critical to your model. How should you proceed?

   • AUsing Dataflow, ingest the columns with sensitive data from BigQuery, and then randomize the values in each sensitive column.
   • BUse the Cloud Data Loss Prevention (DLP) API to scan for sensitive data, and use Dataflow with the DLP API to encrypt sensitive values with Format Preserving Encryption.
   • CUse the Cloud Data Loss Prevention (DLP) API to scan for sensitive data, and use Dataflow to replace all sensitive data by using the encryption algorithm AES-256 with a salt.
   • DBefore training, use BigQuery to select only the columns that do not contain sensitive data. Create an authorized view of the data so that sensitive values cannot be accessed by unauthorized individuals.

   Show answer

   Answer: B

   Explanation

   Not A: Randomizing values alters the data in a way that it can significantly degrade the utility of the data for machine learning purposes (does not preserve original distributions, ...) Not C: Encryption with AES-256 secures the data, but does not preserve the format and would make the data unusable for ML models Not D: This ignores columns with sensitive data, which is not viable here as every column is critical to the model. + Creating an authorized view does not alter the data itself but restricts access, which does not address the need to reduce data sensitivity for model training

   Source: Adapted from community discussion — AnnaR

   References

   • Google Cloud documentation
   • Google Cloud documentation

   Next question

6. Question 56 · 364 in bank

   Select 2

   You recently deployed an ML model. Three months after deployment, you notice that your model is underperforming on certain subgroups, thus potentially leading to biased results. You suspect that the inequitable performance is due to class imbalances in the training data, but you cannot collect more data. What should you do? (Choose two.)

   • ARemove training examples of high-performing subgroups, and retrain the model.
   • BAdd an additional objective to penalize the model more for errors made on the minority class, and retrain the model
   • CRemove the features that have the highest correlations with the majority class.
   • DUpsample or reweight your existing training data, and retrain the model
   • ERedeploy the model, and provide a label explaining the model's behavior to users.

   Show answer

   Answer: BD

   Explanation

   Penalizing Errors on Minority Class (B): This technique, also known as cost-sensitive learning, modifies the loss function during training. Assigning a higher penalty to misclassifications of the minority class steers the model to prioritize learning from those examples. Upsampling/Reweighting Training Data (D): Upsampling increases the representation of the minority class in the training data by duplicating existing data points. Reweighting assigns higher weights to data points from the minority class during training, making their influence more significant.

   Source: Adapted from community discussion — fitri001

   Next question

7. Question 57 · 364 in bank

   You have been asked to develop an input pipeline for an ML training model that processes images from disparate sources at a low latency. You discover that your input data does not fit in memory. How should you create a dataset following Google-recommended best practices?

   • ACreate a tf.data.Dataset.prefetch transformation.
   • BConvert the images to tf.Tensor objects, and then run Dataset.from_tensor_slices().
   • CConvert the images to tf.Tensor objects, and then run tf.data.Dataset.from_tensors().
   • DConvert the images into TFRecords, store the images in Cloud Storage, and then use the tf.data API to read the images for training.

   Show answer

   Answer: D

   Explanation

   My option is D. Cite from Google Pag: to construct a Dataset from data in memory, use tf.data.Dataset.from_tensors() or tf.data.Dataset.from_tensor_slices(). When input data is stored in a file (not in memory), the recommended TFRecord format, you can use tf.data.TFRecordDataset(). tf.data.Dataset is for data in memory. tf.data.TFRecordDataset is for data in non-memory storage.

   Source: Adapted from community discussion — alphard

   References

   • Google Cloud documentation

   Next question

8. Question 58 · 364 in bank

   You are working on a binary classification ML algorithm that detects whether an image of a classified scanned document contains a company’s logo. In the dataset, 96% of examples don’t have the logo, so the dataset is very skewed. Which metric would give you the most confidence in your model?

   • APrecision
   • BRecall
   • CRMSE
   • DF1 score

   Show answer

   Answer: D

   Explanation

   B. Recall In a highly imbalanced dataset like the one described (96% of examples are in the negative class), the metric that would give the most confidence in the model's performance is recall. Recall (also known as sensitivity or true positive rate) is the proportion of actual positive cases that were correctly identified by the model. In this context, it means the percentage of images containing the company's logo that the model correctly classified as positive out of all the actual positive cases. Since the dataset is heavily skewed, a high recall value would indicate that the model is effectively capturing the positive cases (images with the logo) despite the class imbalance. F1 score (D) is a balance between precision and recall and is a useful metric for imbalanced datasets. However, in this specific case, recall is more important because we want to be confident in detecting the logo images, even if it comes at the cost of having some false positives (lower precision).

   Source: Adapted from community discussion — PST21

   Next question

9. Question 59 · 364 in bank

   While running a model training pipeline on Vertex Al, you discover that the evaluation step is failing because of an out-of-memory error. You are currently using TensorFlow Model Analysis (TFMA) with a standard Evaluator TensorFlow Extended (TFX) pipeline component for the evaluation step. You want to stabilize the pipeline without downgrading the evaluation quality while minimizing infrastructure overhead. What should you do?

   • AInclude the flag -runner=DataflowRunner in beam_pipeline_args to run the evaluation step on Dataflow.
   • BMove the evaluation step out of your pipeline and run it on custom Compute Engine VMs with sufficient memory.
   • CMigrate your pipeline to Kubeflow hosted on Google Kubernetes Engine, and specify the appropriate node parameters for the evaluation step.
   • DAdd tfma.MetricsSpec () to limit the number of metrics in the evaluation step.

   Show answer

   Answer: A

   Explanation

   Links already provided below: “That works fine for one hundred records, but what if the goal was to process all 187,002,0025 rows in the dataset? For this, the pipeline is switched from the DirectRunner to the production Dataflow runner.” [Option A] https://blog.tensorflow.org/2020/03/tensorflow-extended-tfx-using-apache-beam-large-scale-data-processing.html. "Metrics to configure (only required if additional metrics are being added outside of those saved with the model).” https://www.tensorflow.org/tfx/guide/evaluator#using_the_evaluator_component will thus add, not “limit the number of metrics in the evaluation step”. [Option D]

   Source: Adapted from community discussion — M25

   Next question

10. Question 60 · 364 in bank

    You are developing an ML model using a dataset with categorical input variables. You have randomly split half of the data into training and test sets. After applying one-hot encoding on the categorical variables in the training set, you discover that one categorical variable is missing from the test set. What should you do?

    • AUse sparse representation in the test set.
    • BRandomly redistribute the data, with 70% for the training set and 30% for the test set
    • CApply one-hot encoding on the categorical variables in the test data
    • DCollect more data representing all categories

    Show answer

    Answer: C

    Explanation

    Answer options analysis: C. Since one categorical variable is missing from the test set, (As I understand: “a categorical variable is in the test but not in train”) apply one hot encoding (trained with the train set?) to the test set, for the variables not present in train we just would obtain an array of all 0’s, so that would be OK. D. That data collection could be not feasible depending on the real-world-problem. B. Randomness would not always fix the problem. A. Not recommended to use different representations for train/test. Sparse representation doesn't magically recover missing categories; it's a way to efficiently store data with a large number of zeros. I would go with C.

    Source: Adapted from community discussion — Nxtgen

    References

    • Google Cloud documentation
    • Google developer documentation

    Next question

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

PMLE practice, question 51. 364 questions available.