Developer Who Wanted to Learn
Deep Learning on AWS
This curriculum included topics such as model versioning, interactions between application versioning and model versioning, diagnostics of production AI in the presence of data changes, A/B testing, and others. In the code lab, a capstone project was built and published on github.
Description
In this series - we will train you to become a Full Stack Deep Learning Engineer, capable of not just training models but also deploying and managing them in production for business value. You will learn (and showcase) your Full Stack Deep Learning skills by building production Deep Learning microservices in AWS and a custom project that you can showcase to employers and peers.
This curriculum is designed to have 8 hands on sessions where you will build, step by step, production grade Deep Learning services for business applications. You will learn how to build production AI in AWS, how to integrate it with an application, and how to manage an AI through its lifecycle. You will also build from scratch a custom project (where you will gather image data or a time-series data, train and tune a production grade neural network, build a prediction service, and connect that prediction service to create your own image processing AI/time-series forecasting application).
Each session will be 1.5 hours and will cover
-
How AI works for images/time-series data, and an overview of Deep Learning
-
How to build a production Deep Learning AI (Bringing images and time-series data into AWS, Model Training, Model Validation, Endpoints, Gateway/Lambda Integration, Application Integration).
-
One or more AWS tools. Across the 8 sessions we will cover AWS Sagemaker, AWS Sagemaker BuiltIn Algorithms, AWS Sagemaker with with Transfer Learning for Neural Networks, How to select and use GPU instances in AWS, AWS Sagemaker Endpoint, AWS Lambda, AWS API Gateway, AWS Roles and Authentication, AWS Cloudwatch, AWS S3, Python based application integration and testing, Production AI Best Practices. Microservice design pattern for AI deployment. We will also cover Navigator, a life cycle overlay tool that makes AWS tools easier to configure and use.
-
You will build and test Image Classifiers and your custom project.
You will learn how to use these Production AI Cloud Tools:
-
How to configure and use S3 for your data
-
How to bring your data into AWS Sagemaker for Images and Deep Learning
-
How to configure and use AWS Sagemaker. Deep dive on built in AWS Sagemaker built in support for CNN/ResNet and Transfer Learning. How to hyper-parameter tune Sagemaker algorithms for Deep Learning.
-
How to configure and use GPU based AWS compute instances for Deep Learning.
-
Configuring and using a Sagemaker Endpoint.
-
Connecting a Sagemaker Endpoint to a public URL via AWS Gateway and Lambda.
-
Integrating REST microservices with applications. Using Python for API testing with images.
-
Cloud AWS best practices. Cloudwatch for logs, managing endpoints.
-
Navigator for ease of use. How to use Navigator and AWS together.
AI Algorithms, Algorithm Internals and DL Technical Concepts
-
How to build and use production grade DL. How to select an algorithm for a use case, train, deploy and use it in production, and measure how well it is doing.
-
Powerful general purpose algorithms - CNN (ResNet) for images and Transfer Learning for acceleration, DeepAR for time-series forecasting, and how they work internally and how to hyper-parameter tune them for best performance.
-
Metrics and practices for algorithmic evaluation.
-
An intro to advanced aspects of Production AI - live monitoring and diagnosis, model versioning, retraining and others.
Prerequisites
-
Bring an AWS account (you can sign up one for free at AWS). You will use your own AWS account to run the hands on labs and have all artifacts (models, datasets).
-
We assume that you have some coding experience in some language. We will provide examples primarily in Python.
Lesson Plan
Session 01 (1 hour)
Session 02 (1 hour)
Session 03 (1 hour)
Session 04 (1 hour)
Session 05 (1 hour)
Session 06 (1 hour)
Session 07 (1 hour)
Session 08 (1 hour)
How to build and run a production DL in the cloud - Image Classification
In this first session, we will show the steps needed to build and run a production DL in the cloud. We will demo these steps with Image Classification for the user to get an overview of the process. Basics of a neural network and classification type of problem will be covered. In the code lab, the attendees will use sample data and learn how to configure S3 in their AWS account.
-
Overview of the production AI lifecycle and all of its steps. How to go from data to running production AI. Structure you will follow for your custom project.
-
Description of cloud services that can be used for each stage
-
Overview of an image classification problem and how to build an AI for it with AWS with Image Classification Algorithms
-
Theory covering basics of neural network and classification type of problems
-
Code Lab: Get your AWS account started and use S3. We will demo an end to end lifecycle for images. If time permits, we will start on Sagemaker.
AWS Sagemaker for Images and Deep Learning
This session builds upon the first. We cover the basics of a convolution neural network and continue our code lab of building an Image Classifier in AWS. In the code lab, attendees will configure AWS in-built image classification algorithm (based on resNet) and select GPU based AWS Compute Instances for Training.
-
Use Case: Image Classification. We will also share project ideas.
-
Algorithms/Concepts: Deep Learning. Images processing for AWS Sagemaker.
-
Production AI: Data storage in S3, Image processing for Training.
Code Lab: We will configure AWS Sagemaker for DL training. All models and artifacts will be in the attendee’s AWS account for their further use.
Hyper Parameter Tuning Deep Learning in AWS Sagemaker.
In this session, we show how to hyper parameter tune DL algorithms trained in the cloud. We will cover Hyper Parameters of ResNet and the concepts of transfer learning. In code lab, we will continue the AWS DL lifecycle and go through a cycle of hyper parameter configuration. We will also demonstrate how to create Sagemaker Endpoints. You will also decide on your capstone project.
-
Use Case: Image classification. Choose the capstone project.
-
Algorithms/Concepts: Hyper Parameter Tuning for Deep Learning in AWS.
-
Production AI: How to retrain a production DL AI with new information and how to iteratively deploy increasingly accurate models.
Code Lab: Hyper Parameter tuning for DL in Sagemaker. Creation of Sagemaker Endpoints.
Sagemaker Endpoints and AWS Lambda, Going from Model to Prediction Service
In this session, we will continue our development of the Production DL Lifecycle. You will learn how to take your trained DL model from previous sessions and create a working production grade microservice for Predictions. You will also make progress on your projects, and create your first custom dataset.
-
Use Case: Images Classification, Creating a Prediction Service
-
Algorithms/Concepts: How to build a Prediction Service for Deep Learning Predictions. How to assess Your prediction service.
-
Production AI: How to configure and EndPoint and create a publicly accessible URL.
-
Code Lab: Sagemaker Endpoints, AWS Lambda and AWS API Gateway. Y will also interact with this service using python code.
Time-series forecasting and feature engineering
In this session, we will show you how to think about time-series forecasting and the basic concepts of regression. We also show you how to convert a dataset into a format acceptable by SageMaker. We will also cover the basic concepts of Recurrent Neural Networks.
-
Use Case: Time-series forecasting: You will be provided with 2 use cases of electricity consumption data.
-
Algorithms/Concepts: Recurrent neural networks and basics of regression.
Code Lab: Attendees will convert the two publicly available datasets into a format that can be consumed by an AWS forecasting algorithm.
DeepAR Algorithm: Training and hyper-parameters.
In this session, we cover the basics concepts of an LSTM and the parameters of a DeepAR algorithm in this context. Code lab will build on the previous session, where the you will configure a time-series forecasting training job and evaluate its training performance.
What the attendees will learn
-
Use Case: Time-series data. Configuring a DeepAR algorithm, iterative tuning of the hyper-parameters
-
Algorithms/Concepts: DeepAR hyper-parameters
-
Production AI: Your custom project
Code Lab: Training job with DeepAR
Time-series Forecasting: Deployment and evaluation
In this session, we will build on the previous session by deploying this service into production and evaluate its performance using regression metrics. You will learn the details of how a DeepAR algorithm works and what context to use it in. You will also interact with the deployed service using a python snippet.
-
Use Case: Time series dataset. AWS lambda code for it.
-
Algorithms/Concepts: Metrics to evaluate the forecasting model.
-
Production AI: AWS Lambda, API Gateway and Python scripts
Code Lab: Build a time-series forecasting prediction service and use python code to evaluate its performance externally. Continue on Custom Project. Build your project’s prediction service and integrate it with a Python Application.
Advanced topics in Production Cloud AI, Complete and showcase your project.
About AIClub’s Professional Development Program (AIClubPro)
AIClub is an education technology company focused on Artificial Intelligence Literacy. We educate individuals from students to professionals, covering all aspects of Artificial Intelligence and related technologies at appropriate depth from introductory to advanced, depending on the individual's prior knowledge and future interests. Our AIClubPro programs have educated professionals from many industries, with roles ranging from Engineering and Operations to Product Management, Marketing and Executive Leadership.
The AIClubPro program is led by three founders with exceptional depth and expertise in both industry and academia. Together, our leadership team has over 40 years of professional experience in technology, having served in executive roles in public companies and startups, founded four companies, and with over 200 patents and over 100 research publications.
