Tackling Domain Shift with SKADA: A Hands-On Guide to Domain Adaptation
summary
SKADA is a Python library that brings domain adaptation into scikit-learn and PyTorch integration, providing ~20 shallow methods plus deep approaches (e.g., DeepJDOT and optimal transport pipelines). It implements practical features: scikit-learn–style pipelines, circular (DS) validation implemented for hyperparameter tuning, reweighting for target shift, mapping adapters that move source data then retrain a classifier, and a benchmark showing circular validation works well for shallow methods. The package is built and demonstrably runs end-to-end.
AI-written summary · about
chapters (12)
- 0:00Domain shift in ML
- 3:36Impact on model performance
- 7:16Distribution shifts and adaptation
- 10:28SKADA for domain adaptation
- 14:29Deep learning in SKADA
- 17:42Implementing deep methods
- 20:35Hyperparameter tuning without labels
- 23:05Circular validation
- 24:12SKADA features and future
- 25:13Q&A: Calibration
- 26:17Q&A: Practical applications
- 28:27Q&A: Multi-domain challenges
related talks
50mRobust Deep Learning Under Distribution Shift
37mDomain Adaptation-- 20 years of theory chasing practice
35mScikit-Learn can do THAT?!
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
52mLanguage-guided Adaptation
44mADTK: An open-source Python toolkit for anomaly detection in...
43mDomain-driven Data Science
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning