mlops/ROADMAP.md

the basic path

• make a jupyterlab docker container!

  • volume mount for notebooks
  • volume for training data?

• get a model working in jupyter notebook

• make a dockerfile to set up the base image

• script to train and validate the model, packaging up the trained model alongside a report of its accuracy etc.

  • include the package requirements in the artifact

• taking that trained model and putting it in a flask or fastAPI container

• how about using my laptop-docker-compose pytest script thing? that's pretty great for what i need

next steps

• get the data into redis for training -- or another data store?
• how about a clean path from jupyter to docker?
• CI/CD: get it working with git and webhooks
• git workflow with webhooks
  • what is a good project model? here's a guess:
    • base dockerfiles for conda and build scripts (?)
    • model definition
    • dataset definitions
    • rest API
    • test container
• whylogs to track model drift?

questions

• how do we get from jupyter to docker? what's a clean path?

• what's a good practice for making training data available? Just have a redis container?

• does it make sense to have each component be a separate CI/CD managed project?

  • data ingest (for each dataset?)
  • feature extraction
  • labeling
  • jupyter notebook for experimentation
  • converter for jupyter-> app model (or do we just rely on people copypasting)
  • model trainer (+verification?)
  • maybe we just define a production application and worry about the rest later?
    • model + API (or messaging system)

• what's a good way to get project dependencies into a jupyter container?

• does a setup like above help with parameter optimization?

  • should an optimizer be its own containerized service?

• how can we reuse feature extraction for regular model input?

  • should a feature extractor be a service?

What needs doing?

• step one is writing out all these steps!

• docker image build pipeline:

  • miniconda image
  • how do we pull in the requirements for a particular model?
    • honestly just ship a Dockerfile alongside the model
      • but it needs to pull from a manifest that an individual developer using something like jupyter would use
      • verify this part of the pipeline according to the training materials

• feature extraction

• training phase

• packaging up a pre-trained model artifact

• CI/CD [ ] webhook plugin for gitea? https://github.com/adnanh/webhook

• missing from the course????

  • training data store!
  • labeling phase (is that the right term?) -- unattended?
  • feature extraction phase

refer:

• https://github.com/docker-science/cookiecutter-docker-science?tab=readme-ov-file
• https://docker-science.github.io/