Practical Low-Rank Communication Compression in Decentralized Deep Learning

Authors: Thijs Vogels, Sai Praneeth Karimireddy, Martin Jaggi

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Out of the box, these compressors perform on par with state-of-the-art tuned compression algorithms in a series of deep learning benchmarks.
Researcher Affiliation Academia Thijs Vogels EPFL Sai Praneeth Karimireddy EPFL Martin Jaggi EPFL
Pseudocode Yes Algorithm 1 Decentralized SGD with edge-wise compression; Algorithm 2 Rank-1 s-step Power Gossip compression for Algorithm 1
Open Source Code Yes This paper s code is available at https://github.com/epfml/powergossip.
Open Datasets Yes We study the algorithm on the Cifar-10 image classification benchmark... We also follow the language modeling experiment on Wiki Text-2... 64 64 images from the Faces Database (AT&T Laboratories Cambridge). (Referenced with URL https://scikit-learn.org/0.19/datasets/olivetti_faces.html)
Dataset Splits No The paper mentions using standard datasets like Cifar-10 and Wiki Text-2 and states 'labeled images that are reshuffled between 8 workers every epoch,' but it does not explicitly provide percentages or sample counts for training, validation, or test splits.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments (e.g., GPU models, CPU types, or memory specifications).
Software Dependencies No The paper does not specify version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes In all experiments, we tune the hyperparameters of our baselines according to Appendix G and use the same learning rate as uncompressed centralized SGD for all instances of Power Gossip. Our compression level is varied through the number of power iterations per gradient update.