PyTesting the Limits of Machine Learning
summary
Shows concrete techniques: use scikit‑learn Estimator/Transformer, pipelines and FeatureUnion to reduce test surface; fix random_state to make runs repeatable; parametrize tests with pytest to sweep classifiers (MLP, RandomForest, SGD) without duplicating code; use pytest.approx for floating‑point tolerances; format with Black and enforce via pre‑commit; run CI with Jenkins to catch dependency regressions (they report finding version issues); and emit diagnostic images with the yellowbrick visualizer alongside saved predictions.
AI-written summary · about
chapters (14)
- 0:00Start
- 0:46Why We Test Machine Learning
- 1:28Do We Really Need To Test Um Our Machine Learning Code
- 2:18Is It Even Possible To Test Machine Learning Code
- 4:04Daniel Solis
- 5:54Pipelines and Future Unions
- 6:29Example of a Pipeline
- 8:23Importance of Ci Cd
- 9:06Dealing with Non-Determinism
- 9:49Non-Determinism
- 11:35Fix the Random Seed
- 17:19The Prox Function
- 19:42Project Yellow Brick
- 23:59Final Tips
related talks
53mUnleashing the Black Box: Testing Machine Learning Systems
33mMachine Learning Under Test
30mUnit-tests for machine learning pipelines
50mDesigning tests for ML libraries
01h 30mAdvanced Software Testing for Data Scientists
51mHow to Test Artificial Intelligence and Machine Learning Solutions?
30mHow to Properly Test ML Models & Data
01h 33mUnderstanding and diagnosing your machine-learning models