Principle:Kubeflow Pipelines Iterative Training Termination
| Sources | KFP Control Flow |
|---|---|
| Domains | Machine_Learning, Control_Flow |
| Last Updated | 2026-02-13 |
Overview
A convergence pattern that terminates an iterative training loop when model quality metrics reach a predefined threshold.
Description
In iterative training, the loop must eventually terminate. The termination pattern uses dsl.Condition to check if the current MSE exceeds a threshold (e.g., 0.01). If it does, the loop continues (recursive call). If not, the condition is not met, the recursive branch is skipped, and the pipeline completes. This combines the recursive sub-pipeline pattern with metric-based convergence checking.
Usage
Use as the termination logic within a recursive training graph component to prevent infinite training and ensure convergence.
Theoretical Basis
Convergence-based termination in iterative algorithms. The error metric (MSE) monotonically decreases with each iteration (in the ideal case). When error < threshold, the model is "good enough" and training stops.