Exam AWS Certified Machine Learning - Specialty topic 1 question 201 discussion

A. NO - Must use SageMaker Debugger for visibility into model insights B. NO - Hyperparameters will most likely influence model accuracy but not response time C. YES - SageMaker Debugger is the right tool for model insights; filter (or "kernels") slides in CNN to identify specific features D. NO - SageMaker Model Monitor is for model performance

Pruning is a technique that reduces the complexity of convolutional neural networks (CNNs) by removing unimportant filters or neurons. This can lead to faster inference times and lower memory consumption, which are desirable for self-driving applications. Pruning can be done by ranking the filters based on some criteria, such as the norm of the weights, the activation outputs, or the Taylor expansion of the loss function123.

ChatGPT is an awesome tool, but please ML colleagues: study!

The company should use solution C. Use SageMaker Debugger for visibility into the training weights, gradients, biases, and activation outputs. Compute the filter ranks based on this information. Apply pruning to remove the low-ranking filters. Set the new weights. Run a new training job with the pruned model.

Same example here: https://aws.amazon.com/blogs/machine-learning/pruning-machine-learning-models-with-amazon-sagemaker-debugger-and-amazon-sagemaker-experiments/

Selected Answer: B To reduce the time required for performing inferences in autonomous cars, the automotive company should use SageMaker Debugger for visibility into the training weights, gradients, biases, and activation outputs. They can adjust the model hyperparameters and look for lower inference times. They can also use SageMaker Model Monitor for visibility into the ModelLatency metric and OverheadLatency metric of the model after the model is deployed. However, option C, which suggests computing the filter ranks based on the training outputs and applying pruning to remove the low-ranking filters, is not applicable for transfer learning models since the layers in the pre-trained model are already trained and cannot be changed. Therefore, the correct solution is B.

Even if a better machine could help, the problem is about the model, not about the general or the machine in specific.

Using SageMaker Debugger, the company can monitor the training process and evaluate the performance of the model by computing filter ranks based on information like weights, gradients, biases, and activation outputs. After identifying the low-ranking filters, the company can apply pruning to remove them and set new weights. By doing so, the company can reduce the model size and improve the inference time. Finally, a new training job with the pruned model can be run to verify the performance improvements Not D because Model Monitor is a tool for monitoring the performance of deployed models, and it does not provide any direct feedback or insights into the model training process or ways to improve model inference time. Therefore, while Model Monitor can be useful for monitoring the performance of deployed models, it is not the best choice for evaluating and improving the performance of the models during the training phase, which is what the question is asking for.

It's between C and D. But I think it's C. C. https://aws.amazon.com/blogs/machine-learning/pruning-machine-learning-models-with-amazon-sagemaker-debugger-and-amazon-sagemaker-experiments/ Everything is there. D: https://aws.amazon.com/premiumsupport/knowledge-center/sagemaker-endpoint-latency/ Here it says use Cloudwatch to view ModelLatency and OverheadLatency, not Model Monitor. I think Model Monitor is just for model performance i.e. drift, bias, accuracy etc.

The answer I guess D per below , they should have said Sagemaker model monitor using cloud watch https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html

C, https://aws.amazon.com/blogs/machine-learning/pruning-machine-learning-models-with-amazon-sagemaker-debugger-and-amazon-sagemaker-experiments/

I would say 'D as a more generic approach than C. The problem can be caused not just filters.

The answer is "c" as the question is asking for evaluate and improve the performance of the models? "https://docs.aws.amazon.com/sagemaker/latest/dg/debugger-visualization.html"

I think the answer is D.