alexeygrigorev / deep-learning-keras-aws-lambda

Deep Learning with Keras and AWS Lambda

Deep learning achieves the best performance for computer vision, natural language processing, and recommendation tasks.

However, it’s difficult to use deep learning in production: it requires a lot of effort to develop proper infrastructure for serving deep learning models. Platforms for serverless computing, such as AWS Lambda, provide a good alternative: they take care of scaling up and down and offer attractive pricing based only on actual usage.

In this workshop, you'll learn how to do a deep learning project end-to-end. First, we will use Keras to train an image classification model that classifies pictures of clothes. Then we’ll deploy this model with AWS Lambda and TensorFlow-Lite.

After finishing, you’ll be able to train a similar model and host it in the cloud.

Links

• DataTalks.Club - the community of people who love data
• Machine Learning Bookcamp - the book on which the workshop is based
• ML Bookcamp Github Repo - all the code from the book
• Machine Learning Zoomcamp - free course based on the book

Part 1: Image Classification with TensorFlow and Keras

Registration link

Segment 1: Introduction

• About the instructor
• What we’ll cover today
• How to set up the environment — AWS, Anaconda + Jupyter, TensorFlow + Keras, Docker, VS Code
• The first look at the dataset

Segment 2: Training the first model

• Loading and preparing data
• Transfer learning with Keras
• Training a small model on a GPU
• Saving the model and using it

Segment 3: Improving the model

• Adding more layers
• Checkpointing: saving the best version of our model
• Making model generalize better: adding dropout
• Generating more data with data augmentation
• Training a larger model
• Testing the model: applying it to test data

Part 2: Deploying Deep Learning Models with AWS Lambda

TODO: add the registration link

Segment 1: AWS Lambda

• What is AWS Lambda
• Implementing a simple Lambda function
• Lambda and Docker

Segment 2: TenserFlow-Lite

• TensorFlow vs TensorFlow-Lite
• Converting Keras models to TF-Lite
• Making predictions with TF-Lite
• Using keras-image-helper for pre-processing of images

Segment 3: AWS Lambda + TF-Lite

• Putting everything together
• Preparing a Docker file
• Using a pre-compiled version of TF-Lite for AWS Lambda
• Testing Lambda code locally
• Uploading the docker image to ECR
• Creating and testing a Lambda function

Segment 4: API Gateway

• Putting the Lambda function in a web service
• Using the webservice for image classification

Extra resources

• CS231n - a course about neural networks for computer vision
• Creating an AWS account
• Tutorial about Lambda and API Gateway
• Requesting a quota increase for AWS Sagemaker