Detached Error Feedback for Distributed SGD with Random Sparsification

Authors: An Xu, Heng Huang

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive deep learning experiments show significant empirical improvement of the proposed methods under various settings.
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA.
Pseudocode Yes Algorithm 1 Detached Error Feedback (DEF(-A)).
Open Source Code No The paper does not provide a direct link or explicit statement about releasing the source code for the described methodology.
Open Datasets Yes Extensive deep image classification experiments on CIFAR-10/100 and Image Net show significant improvements of DEF(-A) over existing works with RBGS.
Dataset Splits Yes The model is trained for 200 epochs with a learning rate decay of 0.1 at epoch 100 and 150. Random cropping, random flipping, and standardization are applied as data augmentation techniques.
Hardware Specification Yes Each machine is equipped with 4 NVIDIA P40 GPUs and there are 16 workers (GPUs) in total.
Software Dependencies No The paper mentions 'Py Torch' and 'NCCL as the backend of the Py Torch distributed package' but does not specify version numbers for any software component.
Experiment Setup Yes The base learning rate is tuned from { , 0.1, 0.05, 0.01, } and the batch size is 128. The momentum constant is 0.9 and the weight decay is 5 10 4. The model is trained for 200 epochs with a learning rate decay of 0.1 at epoch 100 and 150. Random cropping, random flipping, and standardization are applied as data augmentation techniques.