Normal Software vs ML Code

Normal software engineering code All checks have passed 58 successful checks CI lint testing / Success CI build Docker / build-PL (3.6, 3.1) (pull_request) / Successful in 5m CI complete testing / pytest (ubuntu-18.04, 3.6, minimal) (pull_request) / Successful in 15m Check Code Format / Check valid import formatting with isort (pull_request) / Successful in 10s Do a check / sphinx-check (pull_request) / Successful in 1m — sphinx-check Install pkg / pkg-install (ubuntu-20.04, 3.6) (pull_request) / Successful in 1m ML Engineering code tests failing TPU tests failing coverage 73.00% build failing docs failing ML Research code You guys are getting tests?

The image contrasts clean, heavily tested normal software engineering with chaotic machine-learning engineering, where tests, builds, documentation, and coverage are failing while the code burns around the dog. The final reaction asks whether ML research gets tests at all, mocking the stereotype that research code prioritizes experimental results over software quality.

• @profunctor_io / 6218 2021-01-29
• @dev_meme / 2711 2021-02-02