MLA-C01 ML Model Development Practice Test 4 — 15 Questions

Question 1

A company wants to use SageMaker built-in algorithms for a time series forecasting task. Which TWO algorithms are appropriate for this task? (Choose TWO.)

Accepted Answer

DeepAR. DeepAR (A) is a SageMaker built-in algorithm specifically designed for time series forecasting using recurrent neural networks (autoregressive RNNs) that learn from multiple related time series and produce probabilistic forecasts with prediction intervals, making it the canonical choice for this task. Linear Learner (C) is also appropriate because it can be trained on time series data with lagged features and, when configured with the 'forecast' or regression objective, performs linear regression/classification for forecasting numeric sequences. PCA (B) is a dimensionality-reduction algorithm, not a forecasting method, so it cannot predict future values. K-Means (D) is an unsupervised clustering algorithm used for grouping data points, not for time series prediction. XGBoost (E) is a general gradient-boosted trees algorithm for classification/regression on tabular data and, while it can be adapted with feature engineering, it is not a built-in time series forecasting algorithm in the same sense as DeepAR.

Answer

PCA

Answer

K-Means

Answer

XGBoost

Question 2

A machine learning engineer is preparing a dataset for a SageMaker training job. The raw data is stored as a single 200 GB CSV file in Amazon S3. The training script reads the file line by line, but the job fails with an out-of-memory error during the initial data loading phase. The engineer wants to reduce memory usage and improve I/O throughput without changing the training algorithm. Which approach should the engineer take?

Accepted Answer

Convert the CSV file to the RecordIO-protobuf format and store it in Amazon S3.. Converting the CSV to RecordIO-protobuf reduces the memory footprint because the format is binary and blocked, allowing the training algorithm to read data in chunks rather than loading the entire dataset. This directly addresses the out-of-memory error during data loading. Other options either do not change the memory requirement or require additional script modifications that are not specified.

Answer

Store the CSV file in Amazon EFS and mount it to the training instances.

Answer

Enable SageMaker Pipe mode to stream data directly from Amazon S3.

Answer

Increase the instance count of the training cluster to distribute the data loading across more nodes.

Question 3

A data scientist is training a deep learning model using SageMaker and wants to use distributed training to speed up the process. The model has 10 billion parameters and the training data is 1 TB. The data scientist wants to minimize communication overhead between instances while ensuring the model converges correctly. Which distributed training strategy should be used?

Accepted Answer

Model parallelism with SageMaker model parallelism library. Model parallelism with the SageMaker model parallelism library is the correct choice because it partitions the model across instances, reducing memory load and communication overhead tailored for large models. Data parallelism approaches replicate the model, which is infeasible for 10 billion parameters. Pipeline parallelism is a different concept and not applicable here.

Answer

Data parallelism with SageMaker distributed data parallel library

Answer

Data parallelism with Horovod

Answer

Pipeline parallelism with SageMaker Pipelines

Question 4

A machine learning engineer is preparing a dataset for a SageMaker training job that uses Pipe mode. The dataset is stored as many small CSV files in Amazon S3. They want to maximize training throughput and minimize startup latency. Which action should the engineer take?

Accepted Answer

Convert the small CSV files into a single large RecordIO-WebDataset file using the SageMaker open-source tool.. Pipe mode streams data directly from S3, avoiding the download phase of File mode. However, its efficiency depends on the data format: many small files cause excessive S3 requests and slow startup. Converting to a single RecordIO-WebDataset file reduces overhead and enables high-throughput sequential reads, making it the optimal preparation step.

Answer

Keep the files as individual CSVs and rely on SageMaker's automatic file consolidation during training.

Answer

Use File mode instead of Pipe mode and increase the volume size of the training instances.

Answer

Enable SageMaker Enhanced File Mode and set the input mode to FastFile.

Question 5

A data scientist is training a deep learning model using SageMaker and wants to monitor the training job's resource utilization and debug potential issues. The training job is using a custom training script with the TensorFlow framework. The data scientist wants to capture metrics such as GPU utilization, memory usage, and loss values in real time. They also want to visualize these metrics in Amazon CloudWatch. Which of the following should the data scientist do to achieve this?

Accepted Answer

Rely on SageMaker's automatic publishing of system metrics to CloudWatch and use metric_definitions to publish custom metrics.. SageMaker training jobs automatically publish system resource metrics to CloudWatch, so no extra configuration is needed for GPU utilization and memory usage. To capture custom metrics like loss values, the data scientist should define metric_definitions in the estimator, which parses the training logs and publishes the metrics to CloudWatch. This approach is efficient and leverages built-in capabilities.

Answer

Use the SageMaker training job's built-in metrics publishing to CloudWatch by specifying metric_definitions in the estimator.

Answer

Enable SageMaker Debugger with the framework's debugger hook and also enable SageMaker Metrics to capture system metrics.

Answer

Enable SageMaker Debugger by specifying debugger_hook_config in the estimator and use the SageMaker Debugger insights dashboard.

Question 6

A data scientist is preparing a training dataset for a binary classification model in SageMaker. The dataset is stored as CSV in Amazon S3 and contains a mix of numeric and categorical features. They want to use the SageMaker XGBoost built-in algorithm, which requires all features to be numeric. Which preprocessing step is necessary before training?

Accepted Answer

Use SageMaker Data Wrangler to encode categorical features and output the data in CSV format.. SageMaker XGBoost requires all input features to be numeric. Categorical features must be encoded using techniques such as one-hot or ordinal encoding. Data Wrangler provides a no-code way to apply these transformations and output CSV, which XGBoost accepts. Other options either do not address the categorical encoding or introduce unnecessary format conversions.

Answer

Convert the CSV to RecordIO-Protobuf format using the SageMaker SDK.

Answer

Replace missing values with the mean and then standardize all features using Scikit-learn.

Answer

Apply one-hot encoding to the categorical features and convert the CSV to the libsvm format.

Question 7

A machine learning engineer is using SageMaker Automatic Model Tuning to optimize a built-in XGBoost model. They want to reduce the total tuning time and cost while still finding a strong hyperparameter configuration. Which two actions should they take? (Choose two.)

Accepted Answer

Use the Bayesian optimization strategy with early stopping enabled via the SageMaker early stopping mechanism.. To cut tuning time and cost while preserving quality, use strategies that avoid wasting compute on poor trials. Hyperband stops underperforming trials early, and Bayesian optimization with early stopping improves sample efficiency and terminates weak runs. Artificially limiting max_jobs or narrowing ranges risks missing good configurations, and raising parallelism increases cost rather than reducing it.

Answer

Define a narrow hyperparameter range for each tuned parameter.

Answer

Increase the value of max_parallel_jobs to the maximum allowed.

Answer

Set max_jobs to a very small number such as 2.

Question 8

A data scientist is training a deep learning model using the SageMaker PyTorch estimator. The training job is taking too long because the dataset is large and the model is complex. The data scientist wants to reduce training time by using a distributed training strategy that splits the model across multiple GPUs. Which SageMaker distributed training strategy should be used?

Accepted Answer

SageMaker model parallelism. When a model is too large to fit on a single GPU, model parallelism is required. SageMaker model parallelism automatically partitions the model across multiple GPUs, enabling training of large models and reducing per-GPU memory pressure. Data parallelism replicates the model and splits data, which does not solve the model size issue. Pipeline and hybrid parallelism are not valid SageMaker distributed training strategies.

Answer

SageMaker hybrid parallelism

Answer

SageMaker pipeline parallelism

Answer

SageMaker data parallelism

Question 9

A machine learning engineer is using SageMaker Experiments to track multiple training runs. The engineer wants to compare the validation accuracy of runs that used different learning rates. Which SageMaker Experiments component should the engineer use to group and analyze these runs?

Accepted Answer

Experiment. SageMaker Experiments uses experiments to group related trials. By associating all runs with a single experiment, the engineer can compare metrics such as validation accuracy across trials and filter by hyperparameters like learning rate. This provides a structured way to analyze the impact of different hyperparameter settings.

Answer

Trial component

Answer

Lineage tracking

Answer

Trial

Question 10

A company is training a deep learning model using SageMaker distributed data parallelism. They observe that the training job fails with a communication error after a few epochs. The model uses PyTorch and the training script calls the SageMaker distributed data parallel library. Which change should the engineer make to resolve the error?

Accepted Answer

Set the environment variable SM_DDP_USE_EFA to true and ensure the instances are in a cluster placement group.. SageMaker distributed data parallel library supports EFA for improved inter-node communication. Enabling EFA via the environment variable and using a cluster placement group reduces latency and increases bandwidth, which resolves common communication failures in multi-node training. This is a specific configuration step recommended by AWS for distributed training performance and reliability.

Answer

Increase the number of training instances to reduce per-instance batch size.

Answer

Reduce the learning rate to stabilize gradients and prevent the communication error.

Answer

Switch to SageMaker model parallelism to avoid inter-node communication.

Question 11

A machine learning engineer is preparing to train a model using SageMaker and needs to choose an appropriate instance type for the training job. The model is a deep learning model that requires significant GPU memory and compute power. The training dataset is large, and the engineer wants to minimize training time while optimizing cost. The engineer is considering using Amazon EC2 instances with GPUs. Which of the following factors should the engineer consider when selecting the instance type? (Choose two.)

Accepted Answer

The GPU compute capability (e.g., number of CUDA cores) and the availability of Tensor Cores for mixed-precision training.. The two most critical factors are GPU memory and GPU compute capability. GPU memory determines the maximum batch size and model size that can be trained without running out of memory. GPU compute capability, including CUDA cores and Tensor Cores, directly affects training speed, especially with mixed-precision training. These factors are essential for minimizing training time of a compute-intensive deep learning model.

Answer

The network bandwidth available on the instance, as it affects the speed of data transfer from Amazon S3.

Answer

The cost per hour of the instance, as it directly affects the overall training cost.

Answer

The number of vCPUs and the amount of system memory, as these determine the data preprocessing speed.

Question 12

A data scientist is using SageMaker Automatic Model Tuning to optimize a model. The tuning job has completed, and the data scientist wants to identify the best performing training job based on the objective metric. Which action should the data scientist take?

Accepted Answer

Sort the tuning job's training jobs by the objective metric in the SageMaker console.. The SageMaker console provides a sorted list of training jobs by the objective metric, making it easy to identify the best one. While the API can also be used, the console action is the most direct for a data scientist. Manual log comparison or custom scripts are unnecessary and inefficient.

Answer

Use the DescribeHyperParameterTuningJob API to retrieve the BestTrainingJob field.

Answer

Run a separate script that queries Amazon CloudWatch Logs for the objective metric.

Answer

Download all training job logs and manually compare the objective metric values.

Question 13

A machine learning engineer is using SageMaker Automatic Model Tuning (AMT) to optimize a model. They want to limit the total number of training jobs that AMT launches to 50, and they want to ensure that the tuning job stops early if the model is not improving. Which combination of parameters should they configure?

Accepted Answer

Set MaxNumberOfTrainingJobs to 50 and enable early stopping with the Hyperband strategy.. MaxNumberOfTrainingJobs sets the upper limit on the total number of training jobs that the tuning job will launch. Early stopping with Hyperband allows AMT to terminate poorly performing trials before they complete, saving resources. Together, these parameters satisfy the requirement to cap total jobs at 50 and stop early when performance does not improve.

Answer

Set MaxParallelTrainingJobs to 50 and enable early stopping with the Bayesian strategy.

Answer

Set MaxNumberOfTrainingJobs to 50 and use the Bayesian strategy without early stopping.

Answer

Set MaxRuntimeInSeconds to 50 and enable early stopping with the Hyperband strategy.

Question 14

A machine learning engineer is training a model using SageMaker and wants to ensure that the model artifacts are encrypted at rest. The training data is stored in an Amazon S3 bucket that is encrypted with AWS KMS. The engineer needs to specify encryption settings for the training job so that the output model artifacts are also encrypted. Which of the following should the engineer do to encrypt the model artifacts?

Accepted Answer

Specify the KMS key ID in the output_data_config parameter of the SageMaker estimator.. To encrypt model artifacts, the engineer must specify a KMS key in the output_data_config parameter of the SageMaker estimator. This ensures that SageMaker uses that key to encrypt the artifacts before uploading them to S3. Other options do not directly encrypt the model artifacts.

Answer

Use the SageMaker training job's volume KMS key to encrypt the model artifacts.

Answer

Enable encryption on the S3 bucket where the model artifacts are stored, and SageMaker will automatically use that encryption.

Answer

Set the enable_network_isolation parameter to True in the estimator, which encrypts all data in transit and at rest.

Question 15

A machine learning engineer is preparing a dataset for training a model using SageMaker. The dataset contains sensitive personally identifiable information (PII) and is stored in Amazon S3. The engineer must ensure that the data is encrypted at rest and in transit during training. Which two actions should the engineer take? (Choose two.)

Accepted Answer

Enable default encryption on the S3 bucket using AWS KMS (SSE-KMS).. To ensure encryption at rest, enable default encryption on the S3 bucket using SSE-KMS. For encryption in transit, configure the training job to run in a VPC and rely on TLS for data transfer to S3. These two actions together satisfy the requirement for protecting sensitive PII during training.

Answer

Store the data in an Amazon EFS file system mounted to the training instances and enable encryption at rest.

Answer

Use SageMaker Pipe mode to stream data directly from S3 without storing it on the instance.

Answer

Enable SageMaker Debugger to encrypt the training data during transmission.