README.md

[WWW'22] Not All Layers Are Equal: A Layer-Wise Adaptive Approach toward Large-Scale DNN Training

This repository provides an implementation of LENA as described in the paper: Not All Layers Are Equal: A Layer-Wise Adaptive Approach toward Large-Scale DNN Training by Yunyong Ko, Dongwon Lee, and Sang-Wook Kim, In Proceedings of the ACM Web Conference (WWW) 2022

How to run

Run with the 'torch.distributed.launch command':

python3 -m torch.distributed.launch --nproc_per_node=2 --nnodes=1 --node_rank=0 --master_addr="192.168.0.1" run.py \
--num_workers=2 \
--num_iterations=16 \
--method=LENA \
--batch_size=128 \
--num_epochs=100 \
--learning_rate=0.1 \
--warmup=2 \
--warmup_ratio=5 \
--dataset=CIFAR10 \
--model=RESNET18 \
--data_path=/DATAPATH

num_workers: the number of workers (# of GPUs)
num_iterations: the number of local iterations within a node
method: the name of the learning rate scaling scheduler
batch_size: batch size per worker (GPU)
num_epochs: the total number of epochs
learning_rate: base learning rate
warmup: the type of a warmup method (0: no warmup, 1: fixed warmup, 2: layer-wise train-aware warmup)
warmup_ratio: the percentage of the training for warmup period

Citation

Please cite our paper if you have used the code in your work. You can use the following BibTex citation:

@inproceedings{ko2022not,
  title={Not All Layers Are Equal: A Layer-Wise Adaptive Approach Toward Large-Scale DNN Training},
  author={Ko, Yunyong and Lee, Dongwon and Kim, Sang-Wook},
  booktitle={Proceedings of the ACM Web Conference (WWW) 2022},
  pages={1851--1859},
  year={2022}
}