Abstract:As machine learning (ML) increasingly underpins critical applications, credible, comparable, and repeatable experimental results become more important. Everyday workflows should make rigorous experiment specification and controlled execution the default while allowing advanced experimentation when required. In practice, researchers still have to combine tools for configuration, execution, versioning, and evaluation, and repeat common implementation work within their application domain. In this paper, we present a configuration-first design for application-specific ML experimentation frameworks and implement it as LOCALIZE for radio-localization research. Experiments are declared in human-readable configuration files, a workflow orchestrator executes isolated stages with explicit inputs and outputs, and code, data, configurations, environment specifications, and generated artifacts are versioned. LOCALIZE provides preconfigured datasets, processing stages, model-development procedures, and experiment templates while retaining access to the underlying pipeline. Through a qualitative comparison and controlled quantitative studies comparing LOCALIZE with corresponding Jupyter notebook and Kedro implementations, we show that, within the studied localization workflows, LOCALIZE requires fewer codebase edits for changes supported by its localization-specific functionality, while total wall-clock time and peak memory usage remain comparable. In a controlled scaling experiment covering 1x, 5x, and 10x the base dataset volume, we observed sublinear growth in total CPU and wall time over the tested sizes.
From: Tim Strnad [view email]
[v1]
Wed, 29 Oct 2025 16:57:33 UTC (3,441 KB)
[v2]
Sat, 20 Dec 2025 15:53:50 UTC (3,512 KB)
[v3]
Wed, 21 Jan 2026 15:14:36 UTC (3,513 KB)
[v4]
Fri, 14 Aug 2026 09:24:41 UTC (1,768 KB)
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。