Skip to content

Latest commit

 

History

History
43 lines (32 loc) · 1.32 KB

README.md

File metadata and controls

43 lines (32 loc) · 1.32 KB
Raw

DataGossip

DataGossip [paper] is an extension for asynchronous distributed data parallel machine learning that improves the training on imbalanced partitions.

Installation

requires conda:

$ conda env create -f environment.yml
$ conda activate datagossip
$ python setup.py install

Experiment Reproducibility

Download and transform the datasets on your main machine:

$ python prepare_datasets.py

Then, run the following script on each cluster node to start the training. Be aware to set the right ranks and sizes!

$ python experiments/train.py --rank=<rank> --size=<size> --main_address=<main_address>

Afterwards, you can find the results of the experiment in the files (on your machine with rank=0) experiments.pkl and evaluations.pkl which hold pandas DataFrames.

Reference

Please consider citing:

@inproceedings{wenig2022datagossip,
  title={DataGossip: A Data Exchange Extension for Distributed Machine Learning Algorithms},
  author={Wenig, Phillip and Papenbrock, Thorsten},
  booktitle={Proceedings of the International Conference on Extending Database Technology (EDBT)},
  year={2022},
  pages={373--377},
  doi={10.48786/edbt.2022.24},
  url={http://dx.doi.org/10.48786/edbt.2022.24},
}