Liz Sander - Software Library APIs: Lessons Learned from scikit-learn

When you think of an API, you’re probably thinking about a web service. But it’s important to think about your developer interface when designing a software library as well! I’ll talk about the scikit-learn package, and how its API makes it easy to construct complex models from simple building blocks, using three basic pieces: transformers, estimators, and meta-estimators. Then I’ll show how this interface enabled us to construct our own meta-estimator for model stacking. This will demonstrate how to implement new modeling techniques in a scikit-learn style, and more generally, the value of writing libraries with the developer interface in mind.

https://us.pycon.org/2018/schedule/presentation/76/

Transcript

2. Software Library APIs: Lessons Learned from scikit-learn Liz Sander, Data Scientist, Civis Analytics GitHub: elsander @sander_liz
3. ◉ Introduction to APIs ◉ Scikit-learn API ◉ Extending scikit-learn AGENDA
4. What is an API?
5. ◉ A website API is an interface between website and developer What is an API?
8. :(
10. :)
11. ◉ Think of an API as the “developer interface” (as opposed to the user interface) APIs are for software too! Code Developer input Output for developer
12. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
13. Software libraries have APIs. It’s worth some upfront time to make them useful.
14. Software libraries have APIs. It’s worth some upfront time to make them useful. Let’s look at a library that does it well!
15. Scikit-Learn
17. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
18. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
19. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
20. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
21. How do I write a class for logistic regression?
22. How do I write a class for logistic regression? … what about a random forest?
23. How do I write a function/class for logistic regression? … what about a random forest? … and a neural network?
24. How do I write a class for logistic regression? … what about a random forest? … and a neural network? How do I create a general framework for modeling?
25. What is a model?
26. What is a model? Model $$$ Data
27. What is a model? Model ETL $$$ Data
28. What is a model? Train ETL Validate Predict $$$ Data
29. What is ETL? ETL Better Data Data
30. What is ETL? Null Imputation Categorical Expansion Feature Scaling Better Data Data
31. What is ETL? Null Imputation Categorical Expansion Feature Scaling Better Data Data “Transformers”
32. Train ETL Validate Predict $$$ Data
33. Model (“Estimator”) Trans- former $$$ Data estimator.fit(X, y) estimator.predict(X)
34. Meta-estimator Model (“Estimator”) Trans- former $$$ Data
35. The Scikit-learn API Train ETL Validate Predict $$$ Data Meta-estimator Estimator Transformer(s) $$$ Data
36. The Scikit-learn API Train ETL Validate Predict $$$ Data Pipeline $$$ Data
42. The right abstraction makes a library easier to use and reason about.
43. ◉ Stable ◉ Integrates with existing tools ◉ Intuitive ◉ Flexible/extendable What makes a good API?
44. ◉ xgboost, keras, lightning ◉ Civis-maintained ◦ python-glmnet (R wrapper) ◦ civisml-extensions ◦ muffnn ◉ Scikit-learn maintains a list of many others Scikit-learn extensions
45. Stacking
46. Stacking Logistic Meta-estimator (Logistic) GBT Random Forest Train data x i Test data x j
47. Train base estimators using {x i , y i } Logistic Meta-estimator (Logistic) GBT Random Forest Train data x i Test data x j
48. Predict base estimators on {x j } Logistic Meta-estimator (Logistic) GBT Random Forest Train data x i Test data x j
49. Use predictions as features to train meta-estimator Logistic Meta-estimator (Logistic) GBT Random Forest Train data x i Test data x j
52. Stacking Pipeline Stacking Transformer(s) $$$ Data Base 1 2 3 M
53. Stacking Model $$$ Data
54. A robust API can give your library life beyond your own ideas.
55. Conclusion ◉ Make your API clear and consistent ◉ Find an abstraction that mirrors your mental model ◉ Think about developers as users
56. Resources https://github.com/civisanalytics/civisml-extensions http://scikit-learn.org/stable/documentation.html
57. Questions?
58. THANK YOU! Liz Sander lsander@civisanalytics.com