LipGrow

An adaptive training algorithm for residual network based on model Lipschitz

• Our algorithm reduces about 50% time when training ResNet-74 on CIFAR-10

<p align="center"><img width="60%" src="assets/demo.png"/></p>

Install

• Install PyTorch
• Clone recursively

  git clone --recursive https://github.com/shwinshaker/LipGrow.git
  

Setup

• By default, build a ./data directory which includes the datasets
• By default, build a ./checkpoints directory to save the training output

Training

• CIFAR-10/100

  ./launch.sh
  

• Tiny-ImageNet

  ./imagenet-launch.sh
  

• Recipes

  • For vanilla training, set grow=false
  • For training with fixed grow epochs, set grow='fixed', and provide grow epochs dupEpoch
  • For adaptive training, set grow='adapt', and use adaptive cosine learning rate scheduler scheduler='adacosine'

Issues

• ResNet architecture for ImageNet is slightly from the published one. The uneven number of blocks in every subnetwork requires a different grow scheduler for each subnetwork, which demands some extra work

Citation

If you find our algorithm helpful, consider citing our paper

Towards Adaptive Residual Network Training: A Neural-ODE Perspective

@inproceedings{Dong2020TowardsAR,
  title={Towards Adaptive Residual Network Training: A Neural-ODE Perspective},
  author={Chengyu Dong and Liyuan Liu and Zichao Li and Jingbo Shang},
  year={2020}
}