Google MLE 1EX25

You've recently created a custom neural network that relies on essential dependencies unique to your organization's framework. Now, you want to train this model using a managed training service in Google Cloud. However, there's a challenge: the ML framework and its related dependencies aren't compatible with AI Platform Training. Additionally, both your model and data exceed the capacity of a single machine's memory. Your preferred ML framework is designed around a distribution structure involving schedulers, workers, and servers.What steps should you take in this situation?

• A. Use a built-in model available on AI Platform Training.
• B. Build your custom container to run jobs on AI Platform Training.
• C. Build your custom containers to run distributed training jobs on AI Platform Training.
• D. Reconfigure your code to a ML framework with dependencies that are supported by AI Platform Training.

Answer explanation

Incorrect Answers:A. Use a built-in model available on AI Platform Training.This option involves abandoning your custom neural network and using a pre-built model available on AI Platform Training. However, this is not ideal as it disregards the custom development work and the unique dependencies that are essential to your organization’s framework.B. Build your custom container to run jobs on AI Platform Training.Building a custom container allows you to package your ML framework, dependencies, and code into a container image that can be run on AI Platform Training. This approach provides flexibility and can accommodate the unique requirements of your custom model.However, this option as stated focuses on a single container, which might not be sufficient for handling the distributed training needs of your large model and dataset.D. Reconfigure your code to a ML framework with dependencies that are supported by AI Platform Training.Reconfiguring your code to a different ML framework that is compatible with AI Platform Training involves a significant amount of work. It means adapting or rewriting your custom neural network to fit within a different framework.While this could make the training process more straightforward in terms of compatibility with AI Platform Training, it might not be feasible or desirable, especially if your custom framework provides specific benefits or is deeply integrated into your organization's processes.Correct answer:C. Build your custom containers to run distributed training jobs on AI Platform Training.This option extends the idea of building a custom container by specifically focusing on distributed training. By creating custom containers that align with your ML framework's distribution structure (involving schedulers, workers, and servers), you can effectively manage the distributed training process.This approach allows you to leverage the scalability and infrastructure management capabilities of AI Platform Training while maintaining the integrity of your custom framework and its unique dependencies.Links:https://cloud.google.com/vertex-ai/docs/training/containers-overview

You're in charge of a data science team within a large international corporation. Your team primarily develops large-scale models using high-level TensorFlow APIs on AI Platform with GPUs. The typical iteration time for a new model version ranges from a few weeks to several months. Recently, there has been a request to assess and reduce your team's Google Cloud compute costs while ensuring that the model's performance remains unaffected.How can you achieve this cost reduction without compromising the model's quality?

• A. Use AI Platform to run distributed training jobs with checkpoints.
• B. Use AI Platform to run distributed training jobs without checkpoints.
• C. Migrate to training with Kuberflow on Google Kubernetes Engine, and use preemptible VMs with checkpoints.
• D. Migrate to training with Kuberflow on Google Kubernetes Engine, and use preemptible VMs without checkpoints.

Answer explanation

Incorrect Answers:A. Use AI Platform to run distributed training jobs with checkpoints.Involving distributed training with checkpoints on AI Platform, doesn't harness the cost-saving potential of preemptible VMs. While it's efficient and ensures safe progress, it lacks the aggressive cost reduction that preemptible VMs offer, as seen in Option C. Preemptible VMs are more cost-effective, and when paired with checkpoints, they provide a balance between cost savings and maintaining training integrity.B. Use AI Platform to run distributed training jobs without checkpoints.While this may reduce storage costs associated with checkpoints, it risks losing progress in training, potentially increasing overall compute time and cost if interruptions occur.D. Migrate to training with Kuberflow on Google Kubernetes Engine, and use preemptible VMs without checkpoints.This option might reduce costs due to the use of preemptible VMs, but without checkpoints, there's a high risk of losing training progress, potentially leading to increased costs in the long run.Correct answer:C. Migrate to training with Kuberflow on Google Kubernetes Engine, and use preemptible VMs with checkpoints.Kubeflow on Google Kubernetes Engine (GKE) with preemptible Virtual Machines (VMs) can significantly cut costs. Preemptible VMs are short-lived, cheaper compute instances. Using checkpoints ensures that you can save and resume model training, mitigating the risk of using these ephemeral resources. So this approach strikes a balance between cost efficiency and ensuring the integrity and continuity of your model training processes. It leverages the cost-effectiveness of preemptible VMs while maintaining progress safety through checkpoints.Links:Reduce the costs of ML workflows with preemptible VMs and GPUsIntroduction to AI Explanations for AI PlatformUsing Checkpoints for Large Models

You have deployed a model on Vertex AI for real-time inference. While processing an online prediction request, you encounter an"Out of Memory"error.What should be your course of action?

• A. Use batch prediction mode instead of online mode.
• B. Send the request again with a smaller batch of instances.
• C. Use base64 to encode your data before using it for prediction.
• D. Apply for a quota increase for the number of prediction requests.

Answer explanation

Incorrect Answers:A. Use batch prediction mode instead of online mode.C. Use base64 to encode your data before using it for prediction.D. Apply for a quota increase for the number of prediction requests.A, C, and D are not directly related to resolving memory issues caused by large data batch sizes in real-time inference. Batch prediction (A) is an alternative approach but doesn't address the memory issue directly. Base64 encoding (C) and quota increase (D) are not relevant to memory limitations.Correct answer:B. Send the request again with a smaller batch of instances.This error often occurs when the data batch size is too large for the allocated resources. Reducing the batch size can help manage memory usage more effectively.Links:HTTP status codes

You are profiling the performance of your TensorFlow model training time and have identified a performance issue caused by inefficiencies in the input data pipeline. This issue is particularly evident when working with a single 5 terabyte CSV file dataset stored on Cloud Storage.What should be your initial action to improve the efficiency of your pipeline?

• A. Preprocess the input CSV file into a TFRecord file.
• B. Randomly select a 10 gigabyte subset of the data to train your model.
• C. Split into multiple CSV files and use a parallel interleave transformation.
• D. Set the reshuffle_each_iteration parameter to true in the tf.data.Dataset.shuffle method

Answer explanation

Incorrect Answers:C. Split into multiple CSV files and use a parallel interleave transformation.This breaks the large file into manageable parts and allows for parallel data loading, which can enhance efficiency.B. Randomly select a 10 gigabyte subset of the data to train your model.D. Set the reshuffle_each_iteration parameter to true in the tf.data.Dataset.shuffle methodB (using a smaller subset) and D (reshuffling each iteration) don't directly address the core issue of handling a large CSV file efficiently in the pipeline.Correct answer:A. Preprocess the input CSV file into a TFRecord file.This format is optimized for TensorFlow and can significantly speed up data loading.Links:MLOps: Continuous delivery and automation pipelines in machine learningLoad CSV data

You are logged into the Vertex AI Pipeline UI and noticed that an automated production TensorFlow training pipeline finished three hours earlier than a typical run. You do not have access to production data for security reasons, but you have verified that no alert was logged in any of the ML system’s monitoring systems and that the pipeline code has not been updated recently. You want to debug the pipeline as quickly as possible so you can determine whether to deploy the trained model.What should you do?

• A. Navigate to Vertex AI Pipelines, and open Vertex AI TensorBoard. Check whether the training regime and metrics converge.
• B. Access the Pipeline run analysis pane from Vertex AI Pipelines, and check whether the input configuration and pipeline steps have the expected values.
• C. Determine the trained model’s location from the pipeline’s metadata in Vertex ML Metadata, and compare the trained model’s size to the previous model.
• D. Request access to production systems. Get the training data’s location from the pipeline’s metadata in Vertex ML Metadata, and compare data volumes of the current run to the previous run.

Answer explanation

Incorrect Answers:B. Access the Pipeline run analysis pane from Vertex AI Pipelines, and check whether the input configuration and pipeline steps have the expected values.Checking the input configuration and pipeline steps is a good practice to ensure that the pipeline ran as expected. However, this method does not directly address the model's performance metrics, such as accuracy and loss, which are crucial for determining whether the model can be deployed. While you can verify if the correct data and configurations were used, it requires more steps to assess the overall training performance. Therefore, it is not as efficient as using TensorBoard for a quick and comprehensive overview of the training metrics.C. Determine the trained model’s location from the pipeline’s metadata in Vertex ML Metadata, and compare the trained model’s size to the previous model.Comparing the model's size can provide insights into whether the model was trained fully or if there were issues during training. However, this metric alone is not sufficient to ensure that the model performs well. Model size is an indirect indicator of health and does not provide detailed information about the model's accuracy, loss, or other performance metrics. Therefore, while it is a useful check, it does not offer the complete picture needed to make a deployment decision.D. Request access to production systems. Get the training data’s location from the pipeline’s metadata in Vertex ML Metadata, and compare data volumes of the current run to the previous run.While data issues are a common cause of anomalies in training times, requesting access to production systems can be time-consuming and might not be immediately feasible. Additionally, comparing data volumes addresses only one aspect of the problem and does not provide direct insights into the model's training performance. This approach might delay the debugging process and is not the most secure or efficient option for quickly assessing whether the model can be deployed.Correct answer:A. Navigate to Vertex AI Pipelines, and open Vertex AI TensorBoard. Check whether the training regime and metrics converge.Navigating to Vertex AI Pipelines and opening Vertex AI TensorBoard allows you to review the training metrics, such as loss and accuracy, over time. This provides a quick and effective way to determine if the training process behaved as expected and if the model has converged with the expected accuracy. If the training metrics indicate normal convergence, it suggests that the model training process might have been valid despite the shorter duration, making it a reliable indicator for whether the model can be deployed.Links:Introduction to Vertex AI TensorBoardIntroduction to Vertex ML MetadataVisualize and analyze pipeline results

You downloaded a TensorFlow language model pre-trained on a proprietary dataset by another company, and you tuned the model with Vertex AI Training by replacing the last layer with a custom dense layer. The model achieves the expected offline accuracy; however, it exceeds the required online prediction latency by20ms. You want to optimize the model to reduce latency while minimizing the offline performance drop before deploying the model to production.What should you do?

• A. Apply post-training quantization on the tuned model, and serve the quantized model.
• B. Use quantization-aware training to tune the pre-trained model on your dataset, and serve the quantized model.
• C. Use pruning to tune the pre-trained model on your dataset, and serve the pruned model after stripping it of training variables.
• D. Use clustering to tune the pre-trained model on your dataset, and serve the clustered model after stripping it of training variables.

Answer explanation

Incorrect Answers:B. Use quantization-aware training to tune the pre-trained model on your dataset, and serve the quantized model.Entails re-tuning the entire model on your dataset with quantization integrated into the training process. While it can improve performance, it may lead to a decrease in offline accuracy due to the complete re-tuning of the model.C. Use pruning to tune the pre-trained model on your dataset, and serve the pruned model after stripping it of training variables.This method removes insignificant weights from the model during re-training. It's effective for reducing the model size, which can indirectly impact latency. However, the latency reduction might not be as significant as with quantization, and the re-tuning process could affect the model's accuracy.D. Use clustering to tune the pre-trained model on your dataset, and serve the clustered model after stripping it of training variables.Groups weights into clusters to compress the model size. Like pruning, it's more focused on model size reduction than latency improvement. Re-tuning the entire model with clustering might also cause a drop in offline performance.Correct answer:A. Apply post-training quantization on the tuned model, and serve the quantized model.Involves reducing the precision of the model's weights and activations post-training. This can significantly lower latency with minimal impact on accuracy. It's effective for models that have already been trained and need quick performance optimization.Links:MLOps: Continuous delivery and automation pipelines in machine learningModel optimizationTransfer learning and fine-tuning

You recently used Vertex AI Prediction to deploy a custom-trained model in production. The automated re-training pipeline made available a new model version that passed all unit and infrastructure tests. You want to define a rollout strategy for the new model version that guarantees an optimal user experience with zero downtime.What should you do?

• A. Release the new model version in the same Vertex AI endpoint. Use traffic splitting in Vertex AI Prediction to route a small random subset of requests to the new version and, if the new version is successful, gradually route the remaining traffic to it.
• B. Release the new model version in a new Vertex AI endpoint. Update the application to send all requests to both Vertex AI endpoints, and log the predictions from the new endpoint. If the new version is successful, route all traffic to the new application.
• C. Deploy the current model version with an Istio resource in Google Kubernetes Engine, and route production traffic to it. Deploy the new model version, and use Istio to route a small random subset of traffic to it. If the new version is successful, gradually route the remaining traffic to it.
• D. Install Seldon Core and deploy an Istio resource in Google Kubernetes Engine. Deploy the current model version and the new model version using the multi-armed bandit algorithm in Seldon to dynamically route requests between the two versions before eventually routing all traffic over to the best-performing version.

Answer explanation

Incorrect Answers:B. Release the new model version in a new Vertex AI endpoint. Update the application to send all requests to both Vertex AI endpoints, and log the predictions from the new endpoint. If the new version is successful, route all traffic to the new application.involves deploying the new model version to a separate Vertex AI endpoint and modifying the application logic to send requests to both endpoints. This strategy increases complexity by requiring changes to the application and managing multiple endpoints. While it allows for performance comparison and rollback if necessary, the operational overhead and potential for errors make it less preferable than directly using Vertex AI's traffic splitting feature.C. Deploy the current model version with an Istio resource in Google Kubernetes Engine, and route production traffic to it. Deploy the new model version, and use Istio to route a small random subset of traffic to it. If the new version is successful, gradually route the remaining traffic to it.introduces the use of Istio within a Google Kubernetes Engine environment for traffic management between different model versions. While Istio offers sophisticated traffic routing rules, including the ability to route a subset of traffic to different services, this approach necessitates managing infrastructure on Kubernetes, which may not be as straightforward as using managed services like Vertex AI. Additionally, it requires setting up and maintaining an additional layer of technology stack that might not be necessary if the workload is already deployed on Vertex AI.D. Install Seldon Core and deploy an Istio resource in Google Kubernetes Engine. Deploy the current model version and the new model version using the multi-armed bandit algorithm in Seldon to dynamically route requests between the two versions before eventually routing all traffic over to the best-performing version.suggests using Seldon Core with an Istio resource on Google Kubernetes Engine, employing a multi-armed bandit algorithm for dynamic traffic routing. While this approach provides a sophisticated mechanism for evaluating multiple versions and optimizing for the best-performing one, it introduces significant complexity and overhead. Managing a Kubernetes environment, along with Seldon Core and Istio, requires deep technical expertise and deviates from the simplicity and managed nature of Vertex AI Prediction, making it a less ideal solution for a scenario focused on deploying and managing machine learning models with minimal operational burden.Correct answer:A. Release the new model version in the same Vertex AI endpoint. Use traffic splitting in Vertex AI Prediction to route a small random subset of requests to the new version and, if the new version is successful, gradually route the remaining traffic to it.is the most streamlined and effective method for implementing a gradual rollout of a new model version with zero downtime, utilizing Vertex AI Prediction's built-in traffic splitting feature. This approach allows you to route a small, random subset of prediction requests to the new model version while monitoring its performance. If the new version proves successful, you can gradually increase the percentage of traffic it handles until it serves all requests. This method ensures a smooth transition with minimal impact on the user experience, leveraging Vertex AI's capabilities for easy traffic management.Links:Data and model validationApplication deployment and testing strategiesChoosing the right strategyRouters in Seldon Core

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?

• A. Use sparse representation in the test set.
• B. Randomly redistribute the data, with 70% for the training set and 30% for the test set
• C. Apply one-hot encoding on the categorical variables in the test data
• D. Collect more data representing all categories

Answer explanation

Incorrect Answers:A. Use sparse representation in the test set.Using sparse representation can efficiently handle the absence of a category, but it may not align with the encoding used in the training set.B. Randomly redistribute the data, with 70% for the training set and 30% for the test setRedistributing the data might result in a better representation of all categories, but there's no guarantee that it will solve the issue of missing categories.D. Collect more data representing all categoriesCollecting additional data ensures representation of all categories. However, it may not be feasible or timely, depending on data availability.Correct answer:C. Apply one-hot encoding on the categorical variables in the test dataIt's crucial to ensure that the one-hot encoding process is consistent across both training and test datasets. This means applying the same encoding transformation used on the training set to the test set, even if some categories are missing. This maintains the structure and comparability of the datasets, which is essential for valid model evaluation.Links:https://www.blog.trainindata.com/one-hot-encoding-categorical-variables/

You have successfully trained a DNN regressor using TensorFlow to predict housing prices, utilizing a set of predictive features. The default precision for your model is tf.float64, and you've employed a standard TensorFlow estimator with the following configuration:estimator=tf.estimator.DNNRegressor(feature_columns=[YOUR_LIST_OF_FEATURES],hidden_units=[1024,512,256],dropout=None)Your model's performance is satisfactory; however, as you prepare to deploy it into production, you notice that your current serving latency on CPUs is 10ms at the 90th percentile. Your production requirements dictate a model latency of 8ms at the 90th percentile, and you are open to a slight decrease in prediction performance to meet this latency requirement.To achieve this, what should be your initial approach to quickly reduce the serving latency?

• A. Switch from CPU to GPU serving.
• B. Apply quantization to your SavedModel by reducing the floating point precision to tf.float16.
• C. Increase the dropout rate to 0.8 and retrain your model.
• D. Increase the dropout rate to 0.8 in _PREDICT mode by adjusting the TensorFlow Serving parameters.

Answer explanation

Incorrect Answers:A. Switch from CPU to GPU serving.Using GPUs can improve performance but might not be feasible if the deployment environment is CPU-based.C. Increase the dropout rate to 0.8 and retrain your model.While dropout can prevent overfitting, it's unlikely to significantly reduce latency.D. Increase the dropout rate to 0.8 in _PREDICT mode by adjusting the TensorFlow Serving parameters.Altering dropout during prediction could affect model consistency but is unlikely to meet the specific latency requirement.Correct answer:B. Apply quantization to your SavedModel by reducing the floating point precision to tf.float16.Reducing the floating point precision of your model to tf.float16, known as quantization, can significantly decrease the serving latency. This approach often involves a trade-off with a slight decrease in prediction accuracyLinks:Post-training float16 quantizationConvert tensorflow saved_model from float32 to float16https://www.tensorflow.org/lite/performance/post_training_quantization#float16_quantization

You work for a retailer that sells clothes to customers around the world. You have been tasked with ensuring that ML models are built in a secure manner. Specifically, you need to protect sensitive customer data that might be used in the models. You have identified four fields containing sensitive data that are being used by your data science team:AGE, IS_EXISTING_CUSTOMER, LATITUDE_LONGITUDE, andSHIRT_SIZE.What should you do with the data before it is made available to the data science team for training purposes?

• A. Tokenize all of the fields using hashed dummy values to replace the real values.
• B. Use principal component analysis (PCA) to reduce the four sensitive fields to one PCA vector.
• C. Coarsen the data by putting AGE into quantiles and rounding LATITUDE_LONGTTUDE into single precision. The other two fields are already as coarse as possible.
• D. Remove all sensitive data fields, and ask the data science team to build their models using non-sensitive data.

Answer explanation

Incorrect Answers:B. Use principal component analysis (PCA) to reduce the four sensitive fields to one PCA vector.Not specifically designed for data privacy, more for dimensionality reduction.C. Coarsen the data by putting AGE into quantiles and rounding LATITUDE_LONGTTUDE into single precision. The other two fields are already as coarse as possible.Balances privacy and data utility but may still retain identifiable information to some extent.D. Remove all sensitive data fields, and ask the data science team to build their models using non-sensitive data.Safest in terms of privacy but could limit the effectiveness of ML models due to the removal of potentially valuable data.Correct answer:A. Tokenize all of the fields using hashed dummy values to replace the real values.Enhances data privacy by replacing sensitive values with tokens. Effective for maintaining usability for ML models, as discussed in theGoogle Cloud blog.Links:Take charge of your data: How tokenization makes data usable without sacrificing privacy