gt-coar / gt-coar-lab Goto Github PK

Add some use cases to README (or a docs/ site).

• title
• narrative
• screenshot/sketch

• wait for build
• in CI
  • get artifacts
  • create release, upload artifacts
  • verify artifacts
• postmortem
  • tick to new build number #26
  • add post-run job report so we know what our ceiling is for adding new things #26
    • high-water marks:
      • /dev/root 84G 52G 32G 62% / linux-gpu-build
      • /dev/root 84G 37G 47G 45% / release
      • /dev/root 84G 54G 30G 65% / linux-gpu-test
  • release process
    • improve release performance
      • further investigate use of cache vs upload
        • cache invalidation seems quick enough to not be reliable between stages
      • upload files from their downloaded/uncached locations without copy

acme looks like an interesting addition:

• Acme: A new framework for distributed reinforcement learning

A conda-forge staged recipes PR is open, but there on some snags for building some of its dependencies.

Pro:

• it's probably good stuff

Con:

• currently marked "alpha"
• uses lots of nightly/git dependencies

ipydrawio 1.0.1 is up, and ready for integration.

• add to the development environment
• add to the shipped lab

• get jupyterlab-starters on conda-forge
• either...
  • build/cache packages during this build
  • build as separate repos, publish independently as conda-forge packages (probably preferable)
• include in constructor
• test with robot

In CI, after building the installers, on fresh machines (without conda) attempt to run the installers, and use them to do some basic smoke tests:

• list conda packages
• list lab extensions
• start lab
• run some non-trivial notebooks (windows can't do atari_py yet)

Using conda-lock and mamba may give us better cacheability and reproducibility for these increasingly-large builds.

• add mamba and conda-lock to the base env
• use dodo to re-solve envs withconda-lock, based on the state of environment.yml
• check in the lock files
• use the lock files as the input for constructor

https://github.com/rlworkgroup/garage

Let's adopt a long-running poster branch:

• the contents of the poster

Have git HEAD include

• ipydrawio for creating the poster, and making the Binder "waiting room" pretty

Link to the actual poster session:

• use the environment from master
• use the nbgitpuller from poster
• have the poster session jitsi URL

Content Outline

• motivation
• user stories
  • rl-at-home: download offline media, install, run, drill with cards
  • distributed classroom/poster session: use jupyterhub/binder, anonymous class pre-post-assessment
  • offline, small-group classroom e.g. retreat: download offline media, run z2jh, etc.
• screenshots
  • #16 pull from
  • also get some better ones from robotframework
• visualization of the packages used
  • next to the example notebooks provided
• rep the content
  • dennybritz, etc.
• bibliography

dennybritz-2016 --+--- tensorflow
                  |---- scikit.learn

• How did this get built? CI chart

Presentation concerns

• #34 qr codes
  • we can even programatically make pretty qrcodes
  • can these be animated e.g. lunar lander?

Deployment concerns

• binder
• static site on gh-pages with jupyterlite

Using the embedded robotframework, selenium and firefox, test the functionality of gt-coar-lab:

• validate the version
• check each labextension
• run/screenshot some notebooks

Elevator Pitch

Include an in-tool flashcard/spaced-repetition authoring and reviewing tool, with pre-built decks for the included notebook sets (#18).

Motivation

After acquiring new skills and knowledge, the anki approach is effective at helping a learner retain the knowledge.

Design ideas

• Build/adapt a JSON schema for anki cards
• import/export anki decks
  • maybe https://github.com/Stvad/CrowdAnki
• build/adapt a Markdown/python syntax for creating inline cards
  • https://github.com/lukesmurray/markdown-anki-decks
• build/adapt an in-lab deck management, acquisition, and review tool
• add a post_save hook which can extract cards
• build/package decks of cards
  • jupyterlab
  • latex
  • python
  • shipped notebook collections

Build installers that come pre-packed with the GPU versions of key libraries, and necessary support packages.

• tensorflow
• dask
• pytorch

• bump version
• #33 include jupyterlab-starters
  • make starter for britz-2016
• #34 add ipydrawio
• changelog
• merge
• release

Needed for:

• dask.visualize, etc.

• merge #26
• create release
• update release notes when available
• postmortem
  • update release procedure with lessons learned went pretty smoothly
  • merge back to dev, bump version #29

The sagemath stack is available from conda-forge for osx and linux. It's probably worth exploring.