Skip to content

Aravinda89/fsdl-text-recognizer-project

 
 

Repository files navigation

Full Stack Deep Learning Labs

Welcome!

Project developed during lab sessions of the Full Stack Deep Learning Bootcamp.

  • We will build a handwriting recognition system from scratch, and deploy it as a web service.
  • Uses Keras, but designed to be modular, hackable, and scalable
  • Provides code for training models in parallel and store evaluation in Weights & Biases
  • We will set up continuous integration system for our codebase, which will check functionality of code and evaluate the model about to be deployed.
  • We will package up the prediction system as a REST API, deployable as a Docker container.
  • We will deploy the prediction system as a serverless function to Amazon Lambda.
  • Lastly, we will set up monitoring that alerts us when the incoming data distribution changes.

Schedule for the November 2019 Bootcamp

  • First session (90 min)
    • Setup (10 min): Get set up with jupyterhub.
    • Introduction to problem and project structure (20 min).
    • Gather handwriting data (10 min).
    • Lab 1 (20 min): Introduce EMNIST. Training code details. Train & evaluate character prediction baselines.
    • Lab 2 (30 min): Introduce EMNIST Lines. Overview of CTC loss and model architecture. Train our model on EMNIST Lines.
  • Second session (60 min)
    • Lab 3 (40 min): Weights & Biases + parallel experiments
    • Lab 4 (20 min): IAM Lines and experimentation time (hyperparameter sweeps, leave running overnight).
  • Third session (90 min)
    • Review results from the class on W&B
    • Lab 5 (45 min) Train & evaluate line detection model.
    • Lab 6 (45 min) Label handwriting data generated by the class, download and version results.
  • Fourth session (75 min)
    • Lab 7 (15 min) Add continuous integration that runs linting and tests on our codebase.
    • Lab 8 (60 min) Deploy the trained model to the web using AWS Lambda.

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 97.6%
  • Python 2.4%