Exam AWS Certified Machine Learning - Specialty topic 1 question 205 discussion

https://docs.aws.amazon.com/sagemaker/latest/dg/experiments.html

Autopilot is the faster. "Amazon SageMaker Autopilot experiments are now up to 2x faster in Hyperparameter Optimization training mode" . Refer to https://aws.amazon.com/about-aws/whats-new/2022/11/amazon-sagemaker-autopilot-experiments-2x-faster-hyperparameter-optimization-training-mode/?nc1=h_ls

This is the MOST efficient approach. SageMaker's automatic model tuning (HPO) is designed to efficiently search the hyperparameter space and find the best model. By using it for both the custom container image and the built-in algorithm, the ML specialist can directly compare the performance of the two approaches in the least amount of time.

ince AMT works seamlessly with both the built-in image classification algorithm and custom container images, answer D ("tune both models at the same time") might seem feasible at first glance. However, each HPO job is independent—you cannot run a single AMT job for multiple algorithms or containers simultaneously. This limitation makes C the more appropriate choice for managing and comparing experiments systematically

C Question asks "determine whether HPO with the SageMaker built-in image classification algorithm will produce a better model than the model produced by HPO with the custom container image" meaning experiment both option and then determine which is better. "All ML experiments and HPO jobs must be invoked from scripts inside SageMaker Studio notebooks" Sagemake experiments provides more capability https://docs.aws.amazon.com/sagemaker/latest/dg/experiments.html

SageMaker's automatic model tuning (also known as hyperparameter optimization, or HPO) is designed to find the best hyperparameters for your model by running multiple training jobs with different hyperparameter configurations. It supports both built-in algorithms and custom container images, making it a versatile tool for this task.

Option D Not C bcos using SageMaker Experiments to manage multiple training jobs adds an extra layer of management complexity. While it helps in tracking experiments, it does not inherently speed up the HPO process compared to running them concurrently. In summary, Option D provides the most efficient and straightforward approach to determine the best model by leveraging SageMaker’s automatic model tuning capabilities to run HPO on both models simultaneously

Option D stands out as the most effective approach because it leverages SageMaker's automatic model tuning capabilities for both the custom container image and the built-in image classification algorithm. This ensures:

D The question is talking about how to do HPO using AWS Sagemaker for a model in custom image. Experiment is not to do HPO because you need to input parameter manually. So D

Amazon sagemaker experiment is ideal for this.

D: By using SageMaker's automatic model tuning capability to tune both the custom container image model and the built-in image classification algorithm model simultaneously, it leverages the parallel processing capabilities of SageMaker. This approach allows for efficient utilization of compute resources and can potentially complete the tuning process for both models in a shorter amount of time compared to running separate tuning jobs sequentially. Additionally, option D aligns with the requirement of invoking all ML experiments and HPO jobs from scripts inside SageMaker Studio notebooks, as SageMaker's automatic model tuning can be initiated and managed through notebook scripts. While options B and C could potentially work, option D provides the most direct and efficient path to meeting the requirements in the least amount of time by leveraging SageMaker's parallel processing capabilities and avoiding potential development overhead or limitations associated with other approaches.

The best option to meet the requirements in the least amount of time is D. Use the automatic model tuning capability of SageMaker to tune the models of the custom container image and the built-in image classification algorithm at the same time. This approach directly utilizes SageMaker's built-in capabilities for HPO, applies to both custom containers and built-in algorithms, and avoids the inefficiencies associated with local mode or manual management of experiments. It's important to note that while the tuning jobs would not literally run "at the same time" in a single operation, this option represents the most efficient use of SageMaker's capabilities for both scenarios.

Should be C. We are looking at comparing 2 models here, where Sagemaker Experiments fits the bill. D is out because "Amazon SageMaker automatic model tuning (AMT), also known as hyperparameter tuning, finds the best version of a model by running many training jobs on your dataset." https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning.html

D can be done easily https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning.html "You can use SageMaker AMT with built-in algorithms, custom algorithms, or SageMaker pre-built containers for machine learning frameworks."

I will go with D https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning-how-it-works.html

Will go with B and Autopilot supports Image classification as per this link - https://docs.aws.amazon.com/sagemaker/latest/dg/autopilot-automate-model-development.html Autopilot currently supports the following problem types: Regression, binary, and multiclass classification with tabular data formatted as CSV or Parquet files in which each column contains a feature with a specific data type and each row contains an observation. The column data types accepted include numerical, categorical, text, and time series that consists of strings of comma-separated numbers. Text classification with data formatted as CSV or Parquet files in which a column provides the sentences to be classified, while another column should provide the corresponding class label. Image classification with images formats such as PNG, JPEG or a combination of both. Time-series forecasting with time-series data formatted as CSV or as Parquet files.

A. NO - try AMT (=Automatic Model Tuning) before using custom HPO scripts; further, no reason to use the local mode B. NO - Autopilot is not for HPO only, it will also select a model etc. C. NO - requires manual parameter setting for each experiments D. YES - AMT (=Automatic Model Tuning) work with custom containers