20250126_AWS SageMaker model training and deployment resources.md

🟢 Different Levels of AWS Resources for Machine Learning Model Training and Deployment

1. 👉 EC2 Instances: Full User Control (Least Pre-built Content)
   With EC2, you have complete control over the entire setup. You need to:

   • Start an EC2 instance (e.g., GPU-enabled for training deep learning models).
   • Install dependencies manually (e.g., Python, ML libraries like PyTorch or TensorFlow).
   • Copy or configure the training script, and handle the training data management (downloading data from S3 or other sources).
   • Run the training process manually using your own code.
   • Manage all aspects of the environment, scaling, and resource management.

   Use Case: This is suited for advanced users who need fine-grained control over their infrastructure and model training environment.

2. 👉 SageMaker Studio or Notebook Instances: Medium Level (Some Pre-built Content)
   SageMaker Studio or Notebook Instances provide a more managed environment for model training:

   • Pre-installed machine learning frameworks and data science packages (e.g., TensorFlow, PyTorch, Scikit-learn) are available, so you don’t need to install them manually.
   • You still need to choose the instance type (CPU/GPU) and the number of instances for your training.
   • Write your training script, and arrange for input data from S3 or other sources.
   • Run the training job with SageMaker managing infrastructure provisioning, while you configure the runtime.
   • You’ll work within a Jupyter IDE, which provides an interactive experience and allows you to save step-by-step outputs in a notebook, aiding in experimentation and tracking.

   Use Case: This is ideal for users who want a balance between control and convenience, using pre-installed tools in a familiar notebook environment.

3. 👉 SageMaker Pre-built Algorithm Containers (Most Pre-built Content)
   For SageMaker’s pre-built algorithm containers (available in AWS ECR):

   • The algorithm (e.g., Linear Learner, XGBoost, etc.) comes pre-packaged within a Docker container.
   • You only need to arrange the input data (typically in S3) and the location to save the model.
   • In some cases, there are pre-trained model images available to fine-tune or use directly.
   • You still choose the instance type and the number of instances, but SageMaker manages the rest of the infrastructure and runs the pre-built algorithm for you.

   Use Case: This is perfect for users who want to train models quickly without dealing with the complexity of infrastructure management or custom coding. It’s ideal for common algorithms and fast experimentation.

4. 👉 AWS SageMaker JumpStart: No-Code/Low-Code Option
   AWS JumpStart offers a no-code/low-code solution for getting started with machine learning:

   • JumpStart provides pre-built solutions and pre-trained models that can be deployed with minimal or no code.
   • It includes popular use cases (e.g., object detection, sentiment analysis) and can be used to fine-tune models on your own data.
   • JumpStart simplifies the entire process from model selection to deployment with a user-friendly interface.

   Use Case: This is perfect for users or businesses looking to quickly implement machine learning solutions without deep expertise in coding or managing infrastructure.

---

These levels provide a range of options from complete control to minimal involvement, letting you choose the approach that best fits your needs.

✅ 20241219 AWS Machine Learning - Training with EC2 (demo video)
https://www.youtube.com/watch?v=yBq0Fn7Z6RU

✅ AWS SageMaker SDK PyTorch ResNet50 (example project)
https://github.com/nov05/udacity-CD0387-deep-learning-topics-within-computer-vision-nlp-project-starter
  • Training job with debugging and profiling enabled notebook
  • Training script
  • Inference script

✅ AWS ECR Linear learner container (example notebook)
https://github.com/nov05/udacity-aws-mle-nano-course5/blob/main/exercise_3.5/multiinstancestarterfile.ipynb

---

👉 Docker containers for training and deploying models
https://docs.aws.amazon.com/sagemaker/latest/dg/docker-containers.html

👉 How to find AWS SageMaker pre-built container registry paths
AWS > Documentation > Amazon SageMaker > ECR Paths
Docker Registry Paths and Example Code
Click on US East (N. Virginia)
Docker Registry Paths and Example Code for US East (N. Virginia) (us-east-1)
Click on Linear Learner (algorithm)

👉 Get container registry path via SageMaker SDK example Python code

import boto3
from sagemaker import image_uris
image_uris.retrieve(framework='linear-learner', region=boto3.Session().region_name)
## output: 382416733822.dkr.ecr.us-east-1.amazonaws.com/linear-learner:1