Dragonn

Overview

Dragonn is a randomized hashing algorithm for gradient sparsification in data-parallel distributed training to minimize the compression overhead. DRAGONN can significantly reduce the compression time by up to 70% compared to state-of-the-art GS approaches, and achieve up to 3.52x speedup in total training throughput.

Citation

If you find our project useful in your research, please consider citing:

@InProceedings{wang22aj,
  title = 	 {{DRAGONN}: Distributed Randomized Approximate Gradients of Neural Networks},
  author =       {Wang, Zhuang and Xu, Zhaozhuo and Wu, Xinyu and Shrivastava, Anshumali and Ng, T. S. Eugene},
  booktitle = 	 {Proceedings of the 39th International Conference on Machine Learning (ICML)},
  year = {2022}
}

Dependencies

install dependencies

pip install -r requirements

# install Dragonn
cd extensions/cuda
python setup.py install

download pretrain weights

# ViT
wget https://storage.googleapis.com/vit_models/imagenet21k/ViT-B_16.npz

# MLP-Mixer
wget https://storage.googleapis.com/mixer_models/imagenet1k/Mixer-B_16.npz

# XML
gdown https://drive.google.com/uc?id=14g1ZG1S_Emq2Hu0MBZYKccMYJL1QzI6l

Run

Examples to run ViT

# run ViT on 4 GPUs without compresion. The dataset is cifar10
horovodrun -np 4 /usr/bin/python3.7 main_vit.py --pcie --lr 1e-5 --batch_size 32 --epochs 1 --print_every 10 --pretrained --dataset cifar10

# run ViT on 8 GPUs with DGC as the compressor and the compression ratio is 0.01
horovodrun -np 8 /usr/bin/python3.7 main_vit.py --pcie --compress --compressor dgc --memory none --comm allgather --compress-ratio 0.01 --lr 1e-5 --batch_size 32 --epochs 1 --print_every 10 --pretrained --dataset cifar10

Examples to run MLP-Mixer

# run MLP-Mixer on 8 GPUs with DRAGONN as the compressor and the compression ratio is 0.001.
# --batching is the argument to enable sparse decoding
# --threshold is to specify when to apply compression based on efficiency-aware tensor selection
horovodrun -np 8 python3 main_mixer.py --pcie --compress --compressor atopk --memory none --comm allgather --compress-ratio 1e-3 --lr 1e-5 --batch_size 32 --batching --threshold 204800 --epochs 1 --print_every 1 --pretrained --dataset cifar10

Note: make sure the directory of dataset and pretrained model is correct.

Training with multiple machines

Refer to Horovod to run Dragonn on multiple machines.

How to run Dragonn in your code

It is easy to apply Dragonn to your customized training code. Let's take main_mixer.py as an example to showcase how to add a few lines of code to enable Dragonn.

# Line 24: import required Draggon functions 
from mergeComp_dl.torch.helper import add_parser_arguments, wrap_compress_optimizer

# Line 120: add Dragonn specified arguments, such as compression algorithms and compression ration
parser = add_parser_arguments(parser)

# Line 202: wrap the optimizer with Dragonn's optimizer to support compression
optimizer = wrap_compress_optimizer(model, optimizer, args)