Stochastic Gradient Push for Distributed Deep Learning

Authors: Mahmoud Assran, Nicolas Loizou, Nicolas Ballas, Mike Rabbat

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate the performance of SGP on image classification (Res Net-50, Image Net) and machine translation (Transformer, WMT 16 En De) workloads.
Researcher Affiliation Collaboration 1Facebook AI Research, Montr eal, QC, Canada 2Department of Electrical and Computer Engineering, Mc Gill University, Montr eal, QC, Canada 3School of Mathematics, University of Edinburgh, Edinburgh, Scotland.
Pseudocode Yes Pseudocode is shown in Alg. 1.
Open Source Code Yes Our code is available at [https://github.com/facebookresearch/stochastic gradient push].
Open Datasets Yes We train a Res Net-50 (He et al., 2016) on the Image Net classification task (Russakovsky et al., 2015). We train a transformer network (Vaswani et al., 2017) on WMT16-En-De
Dataset Splits Yes We train a Res Net-50 (He et al., 2016) on the Image Net classification task (Russakovsky et al., 2015). We train a transformer network (Vaswani et al., 2017) on WMT16-En-De. The paper mentions 'validation accuracy' and 'validation curves', implying the use of standard benchmark dataset splits.
Hardware Specification Yes Our experiments use 32 NVIDIA DGX-1 servers. Each server has 8 V100 GPUs.
Software Dependencies No The paper mentions that 'All algorithms are implemented in Py Torch (Paszke et al., 2017),' but does not provide a specific version number for PyTorch or any other software component used.
Experiment Setup Yes Every node uses a mini-batch size of 256, so using more nodes corresponds to larger effective mini-batch size. Unless indicated otherwise, all experiments are run for 90 epochs, the learning rate warms up to n 0.1 during the first five epochs following Goyal et al. (2017) and is decayed by a factor of 10 at epochs 30, 60, and 80. All methods use Nesterov momentum.