Alsatian: Optimizing Model Search for Deep Transfer Learning

This repository contains the code to our SIGMOD 2025 paper. For an overview of our (additional) experiments look into the experiments folder of this repo.

General Info

• We developed our system under the name MOSIX and later renamed it to Alsatian.
• Large files like datasets, model snapshots, or large log files of experiments are not contained in the repo. If you need access please feel free to reach out to us!

Repository Structure

• setup contains information on the hardware and software setup we use.
• model_search contains the implementation of Alsatian(Mosix), the baseline, and the successive halving approach.
• experiments contains the scripts to execute and analyze/plot the experiments in our paper. We structure them into:
  • main experiments: Experiments that are described in detail in our paper.
  • side experiments: Experiments we ran to get additional insights but that do not directly correspond to an experiment in main our paper (you might find them in the appendix).
  • micro benchmarks: Evaluation of sub-operations that helped us making design decisions.
• data contains information on how to download, access, and prepare the datasets used in our experiments.
• custom contains models, data loaders, and short scripts. These files are mostly slight adjustments or minor extensions of PyTorch files and helper scripts that made the development and analysis easier.
• global utils is a collection of constants, helper methods, and scripts we use as part of Alsatian or our evaluation.

Cite our work

If you use Alsatian or reference our findings, please cite us.

@inproceedings{strassenburg_alsatian_2025,
  author = {Strassenburg, Nils and Glavic, Boris and Rabl, Tilmann},
  title = {Alsatian: Optimizing Model Search for Deep Transfer Learning},
  booktitle = {SIGMOD},
  year = {2025},
  istoappear = true,
  venueshort = {{SIGMOD}},
  keywords = {Sys4ML, Machine Learning}
}