Efficiently Managing Deep Learning Models in a Distributed Environment

This repository contains the code to our EDBT '22 paper.

MMlib

MMlib is a library that implements different approaches to save and recover models.

Approach names:

• baseline approach
  • implemented by the BaselineSaveService
• parameter update approach
  • implemented by WeightUpdateSaveService (set improved_version=False)
• improved parameter update approach
  • implemented by WeightUpdateSaveService (set improved_version=True)
• provenance approach
  • implemented by ProvenanceSaveService

Next to the approaches to save and recover models we also implemented a probing tool

• the corresponding code is in probe.py

Examples

• For examples on how to use MMlib and the probing tool checkout the examples directory.

Installation

Option 1: Docker

• Requirements: Docker installed
• Build Library
  • clone this repo
  • run the script generate-archives-docker.sh
    • it runs a docker container and builds the mmlib in it
    • the created dist directory is copied back to repository root
    • it contains the .whl file that can be used to install the library with pip (see below)
• Install
  • to install mmlib run: pip install <PATH>/dist/mmlib-0.0.1-py3-none-any.whl

Option 2: Local Build

• Requirements: Python 3.8 and Python venv
• Build Library
  • run the script generate-archives.sh
    • it creates a virtual environment, activates it, and installs all requirements
    • afterward it builds the library, and a dist directory containing the .whl file is created
• Install
  • to install mmlib run: pip install <PATH>/dist/mmlib-0.0.1-py3-none-any.whl

Cite Our Work

If you use MMlib or insights from the paper, please cite us.

@inproceedings{strassenburg_2022_mmlib,
  author    = {Nils Strassenburg and Ilin Tolovski and Tilmann Rabl},
  title     = {Efficiently Managing Deep Learning Models in a Distributed Environment},
  booktitle = {Proceedings of the 25th International Conference on Extending Database Technology (EDBT 2022) Edinburgh, UK, March 29 - April 1},
  pages     = {234--246},
  publisher = {OpenProceedings.org},
  year      = {2022},
  doi       = {10.48786/edbt.2022.12}
}