Discover how to instrument Weights & Biases (W&B) in machine learning pipelines through a practical demonstration using a PII data detection use case. Learn to integrate W&B for experiment tracking, data analysis, and error assessment, applying these skills to a live Kaggle competition. Master the use of W&B Tables for data visualization and error analysis, and explore dataset version control and model checkpoint management with W&B Artifacts. Gain insights into proper validation approaches, cross-validation techniques, and best practices for experiment setup. Delve into token classification, model prediction processes, and effective error identification strategies. Explore dashboard configuration, metrics interpretation, and techniques for ensuring reproducibility in machine learning projects.
Instrumenting Weights & Biases for PII Data Detection - ML Pipeline Tutorial
Overview
Syllabus
Introduction to the Session: Overview of topics covered,
General Approach to Machine Learning Problems
Explanation of the Kaggle Competition
Importance of Evaluation Metric
Overview of Weights and Biases Platform
Proper Validation Approach
Approach to Cross-Validation
Data Visualization and Analysis
Introduction to Best Experiment Setup
Discussion on Scroll Price Competition
Review of Training Script Progress
Monitoring Training Metrics
Overview of Logged Evaluation Metrics
Initial Setup and Dashboard Configuration
Sharing Code and Future Availability
- Explaining Dashboard Views and Metrics Interpretation
Analyzing Model Performance and Error Identification
- Understanding Token Classification and Model Prediction Process
Identifying Prediction Processing Issues and Error Analysis
Explanation of Code for Token Classification and Testing Techniques
Overview of Experiment Tracking, Data Set Versioning, and Reproducibility
Q&A
Outro & Resources to follow
Taught by
Weights & Biases
