MNIST Web Classifier

Changelog

[1.0]

• generated and trained model off of mnist dataset
• put model behind API on webapp with Flask
• javascript function for user to drwa their own digit and have it classified
• application containerized and deployed to AWS

[1.1] (Revitalized)

• app moved from flask to fastAPI
• Jinja and CSS used to make UI 40% less dogshit

[1.2]

• DB and API to store past performance of predictions
• added form to label your own drawing
• Db that stores performance of model
• CRUDs, APIs and schemas for history DB
• JS function that draws MNIST formatted digits
• Downloaded original CSV dataset for use in "explore data" page, view some sample data
  • cut data down to first 1000 rows from MNSIT for faster loading

[1.3]

• imported model generation program from kaggle
• model generation produces modelstats.txt metadata json, which is shown on site
• the data pipeline is [up]
  • each time a user uses the model, a new data object is created: user drawing and user label. this is added to a DB of user created data
  • model trainer launches, opens both CSV MNIST dataset and the user dataset, trains
  • re-generates metadata page, re-deploys model to classify page

[1.4]

• generate requirements.txt with pipreqs
• build docker with docker build -t imagename
• uplodaed to ECR

[1.5]

• added new relic one for basic observability
• user database now dumped to local CSV to simplify database accesses and match format of MNIST dataset
• Update model: switch from an MLP to a convolutional network for better performance on translation-blind features
  • was stuck here for a while, translating data shapes & formats between ML libraries
• model re-training is now completely decoupled from webapp. Good design, but currently no way to re-train network. so needs better implementation

[1.6]

• added makefile to launch easily with optional NR observability
• patched issue with user data not updating into CSV used for training

[todo]

• want to return a percentage distribution of confidences eventually
• dataviz on training convergence would be nice - ROC curve equivalent for non-binary classifier? research
• Deployment:
  • re-containerize & deploy to AWS & add DNS
  • Sagemaker?
  • work queue for training before returning prediction
    • worker pool does training outside the webapp prediction return