The Ultimate AWS MLOps Bootcamp: With Continuous Retraining

Name: The Ultimate AWS MLOps Bootcamp: With Continuous Retraining
Rating: 4.6 (12 reviews)

Build a production MLOps pipeline with SageMaker, Bedrock, CloudWatch, API Gateway, auto-labeling, gated deployment+more

Created byPatrik Szepesi

Last updated 6/2026

English

What you'll learn

Fine-tune a Vision Transformer (ViT) image classifier using Hugging Face on AWS SageMaker
Build Profesional Production Ready Sagemaker Pipelines
Understand the full MLOps lifecycle: train → deploy → monitor → evaluate → gate → promote → repeat
Stage datasets to S3 and feed them into training via SageMaker input channels
Build Monitoring Dashboards
Evaluate model accuracy and F1 score during training
Configure and run managed SageMaker training jobs with GPU instances
Save production-ready model artifacts with embedded inference code
Deploy a real-time SageMaker endpoint from a trained model artifact
AWS Bedrock Guardrails, Quotas, Inference Profiles, Inference Configuration, Model Providers, LLMs, Pricing
Invoke a live endpoint with Sagemaker, Lambda and API Gateway
Enable endpoint data capture to log every request and response to S3
Build a CloudWatch dashboard tracking model confidence, low-confidence rate, and retraining activity
Orchestrate Multi Step Machine Learning Pipelines
Build a safe auto-labeling pipeline with validation, error handling, and audit records
Build an API Lambda + API Gateway that serves predictions and captures low-confidence examples
Batch low-confidence examples and trigger an automated retraining run
Implement a full continuous-improvement loop that runs hands-off in production
Design a serverless inference API with API Gateway and Lambda
Wire SageMaker, Bedrock, Lambda, S3, CloudWatch, and API Gateway into one cohesive system
Apply IAM roles and permissions across the full MLOps stack
Understand the ML Lifecycle
Train Image Classifiers with AWS' Nvidia GPUs
Use HuggingFace Models in Sagemaker
AWS Pricing

Course content

11 sections • 75 lectures • 11h 32m total length

Course Syllabus12:36
How The Retrain Pipeline Gets Kicked Off Demo5:16
Github Source Codes0:02

Signing In To AWS3:20
Understanding IAM Roles and Permissions6:39
Create IAM users and assign roles with specific permissions for SageMaker, S3, and Bedrock, enabling console or programmatic access. Use Identity Center for console access and improve password security.
Setting Up Sagemaker Domains1:38
Understanding SageMaker Notebooks10:58
Open Studio and launch JupyterLab to run experiments, pipelines, and retraining jobs with SageMaker; choose cost-effective MLT3 medium or larger instances based on memory needs.

Downloading The Data From HuggingFace7:08
Download the HuggingFace dataset into the notebook, perform exploratory data analysis, and upload to S3 for training, while optionally loading directly from HuggingFace for large data.
Addressing Rate Limit Errors3:15
Exploratory Data Analysis4:27
Uploading Dataset To S36:06
Upload the Hugging Face recycling dataset to S3 using a defined bucket and data set prefix before training, since SageMaker reads from S3.
Verifying the Dataset in S33:30

Configuring our Vision Transformer Train Setup10:04
Creating Helper Functions and Evaluations for the Train Script14:46
Creating Main Train Function in the Train Script9:13
Setting Up Model Configurations for Training9:23
Getting Metrics During Training6:45
Saving Model Artifacts3:01
Save model weights and config to the SageMaker model directory and save preprocessing config with imageProcessor.savePretrained. Write evalmetrics.json and labels.json for S3, preparing for endpoint inference with inference.py.
Setting up the Sagemaker Estimator8:31

Starting the Training Job on a GPU Server4:55
Learn how to start a SageMaker training job on a GPU server using estimator.fit with an S3 data URI, and monitor training logs in CloudWatch.
Debugging Training Errors3:08
Debug training errors by fixing the missing slash in the S3 dataset path, setting seed to 42, correcting a typo and function defaults, and rerunning estimator.fit after updating train.py.
Evaluating the Training Job3:06

Creating the Inference Script Part 116:51
Deploy a model endpoint and build a minimal inference script for SageMaker. Handle JSON and image input with a pre-trained image processor and GPU or CPU deployment.
Creating Inference Script Part 27:44
Creating the Inference Script Part 39:43
Deploying a Live Endpoint4:00
Testing our Endpoint and our Data Capture Config4:00
Test the deployed SageMaker endpoint by submitting an image URL to verify predictions and data capture for traceability, highlighting soft plastics and top classes for production-ready retraining.
Setting up CloudWatch Dashboards for Monitoring our Pipeline10:02
Shutting Down the Notebook Server1:04

Understanding the Re-Training Pipeline9:31
Creating The Endpoint Update Lambda Function for the SageMaker Pipeline16:47
Endpoint Configuration for Sagemaker Pipeline and A/B test Setup19:04
Finishing The Lambda Deploy Model Function7:23
Comparing Model Metrics in the Re-Train Pipeline9:45
Getting the Candidate Model's Metrics16:26
Brief Clarification regarding the Previous Video3:12
Compare Metrics Function10:20
Understand the compare metrics function that extracts a numeric value from in-memory metrics, supports plain and SageMaker style formats, and compares candidate metrics to current SageMaker metrics loaded from S3.
Saving Comparison Metrics to the Sagemaker Directory9:54
Importing Modules for the Sagemaker Pipeline11:45
Creating Pipeline Definitions14:12
Setting Up the Processing Steps in the Sagemaker Pipeline11:41
Adding The Comparison Step to the Sagemaker Pipeline7:05
Adding the Conditional Lambda Deployment to the Sagemaker Pipeline14:33
Add a conditional lambda deployment step to the SageMaker pipeline that deploys only when the candidate model improves accuracy, with outputs for deploy status, endpoint name, and model name.
Clarification on Condition Step2:12

Understand the Lambda Function that Connects API Gateway with our Pipeline9:06
Connecting Bedrock, API Gateway, Sagemaker Endpoint, and the Pipeline Part 120:07
Connecting Bedrock, API Gateway, Sagemaker Endpoint, and the Pipeline Part 214:26
Clarification Regarding the List Pending Examples Function6:43
Connecting Bedrock, API Gateway, Sagemaker Endpoint, and the Pipeline Part 314:47
Connecting Bedrock, API Gateway, Sagemaker Endpoint, and the Pipeline Part 414:50
Move one pending labeled example into a retraining batch, preserve labels with a safe lock, validate bedrock labels, then start the SageMaker retraining pipeline with the new batch.
Connecting Bedrock, API Gateway, Sagemaker Endpoint, and the Pipeline Part 54:22
Updating our Train Function to Have the Retrain Method15:11
Setting Up API Gateway and the Lambda Integration6:31
Importing Pipeline Definition File to our Sagemaker Notebook5:15
Learn to import and upsert a SageMaker pipeline from a definition, then configure Lambda and API Gateway details for deploying and retraining the model in inference.
Creating our Sagemaker Pipeline and Visualising the Graph3:24
Setting up Deployment Notebook For Professional Deployment16:14
Testing API Gateway and our CloudWatch Dashboard5:33
Checking the Retraining Pipeline8:23
Updating the Sagemaker Pipeline with the Appropriate GPU Server6:03
Adjusts the SageMaker pipeline to use the MLG 5 extra large GPU for training, updates the pipeline definition and deployment notebook, and deploys version 2 to enable a retrain pipeline.
Updating Lambda Function with Appropriate Permissions1:29

Requirements

Basic Python skills
AWS Account

Description

Complete MLOps on AWS: SageMaker, Bedrock & Automated Retraining

Most machine learning courses teach you to train a model and call it done. This course starts where those leave off.

You will build a complete, production-grade MLOps system on AWS from scratch — a self-improving recycling image classifier that trains, deploys, monitors, and retrains itself automatically. Every component is real, every line of code runs in your AWS account, and by the end you have a working closed-loop pipeline that would hold up in a professional environment.

The course opens with a thorough introduction to Amazon Bedrock — model providers, the Bedrock playground, inference profiles, guardrails, and LLM inference configuration — so you understand the foundation before using Claude as an auto-labeling engine inside the retraining loop.

From there you fine-tune a Vision Transformer on SageMaker, deploy it to a live real-time endpoint with data capture enabled, and build a CloudWatch dashboard that tracks model confidence and retraining activity in real time. You then wire up an API Gateway and Lambda function that serves predictions to the outside world, detects low-confidence outputs, and automatically saves those weak cases to S3 for review.

The centerpiece of the course is the SageMaker Pipeline: a multi-step workflow that retrains the model on new Bedrock-labeled data, compares the candidate against the model currently in production, and uses a condition step to gate deployment — the new version ships only if it genuinely beats what is already live. Every model version is tracked, every deployment decision is logged, and the whole system runs without manual intervention.

This is the complete MLOps lifecycle — train, deploy, monitor, evaluate, gate, promote, and repeat — built piece by piece so you understand exactly how every part fits together.

Who this course is for:

Machine Learning Engineers
Devops Engineers
Full Stack Developers
Data Analysts
Data Scientists
Data Engineers
Backend Engineers
Mlops Engineers
Cloud Architects

The Ultimate AWS MLOps Bootcamp: With Continuous Retraining

What you'll learn

Explore related topics

Course content

Course Introduction3 lectures • 18min

Understanding AWS, IAM roles, Domains and Notebooks4 lectures • 23min

Installing Required Packages and Dependencies to our Sagemaker Notebook3 lectures • 24min

Accessing our Dataset and doing Exploratory Data Analysis5 lectures • 24min

Setting up our AWS Sagemaker Training Job with Vision Transformers7 lectures • 1hr 2min

Training our Vision Transformer on Nvidia GPUs in AWS3 lectures • 11min

Deploying our Model to Sagemaker and Setting up Monitoring Dashboards7 lectures • 53min

Creating our Sagemaker Retrain Pipeline15 lectures • 2hr 44min

Connecting the Sagemaker Pipeline with other AWS Services16 lectures • 2hr 32min

Testing our Final Sagemaker Pipeline3 lectures • 22min

Requirements

Description

Who this course is for: