serval-uni-lu/drift-robustness

Drift retraining schedules replication package

Introduction

This package contains all the necessary data and scripts to replicate the experiments of our research paper "Predictive Maintenance of Industrial Machine Learning Systems in Production" on public data.

Create the environment

We recommend using miniconda and Python 3.8.10 on a Linux system.

conda create -n drift-study python=3.8.10
conda activate drift-study
pip install -r requirements-lock.txt

For windows, we recommend using the requirements-lock-win.txt requirements file instead.

File structure

• config: configurations files used to prepare and run the experiments
• data: the source data set and the results generated by our experiments
  • optmizer_results: the result of the evaluation of the retraining schedules
• drift_study: python modules to run the experiments
• reports: the final reports of the experiments as tables(.csv) and figures (.html)
• scripts: the script to run an experiment

Unzip the data

./script/unzip_data.sh

Usage

Run the following command to run an experiment $rq among [rq1, rq2, rq3, rq4].

conda activate drift-study
chmod +x ./script/lcld_${rq}.sh
./script/lcld_${rq}.sh

Note that by default, to limit computational costs, only the final reports generation is executed from pre-computed results. Feel free to uncomment python commands in ./script/lcld_${rq}.sh to run the complete experiments.

The reports are found in ./reports/lcld_201317_ds_time/$rq.[csv|html]