DSXplore: Optimizing Convolutional Neural Networks via Sliding-Channel Convolutions

Accepted at IPDPS-2021 [arxiv]

Author: Yuke Wang, Boyuan Feng, and Yufei Ding
Email: [email protected]

Install

• Dependency: Python 3.7, nvcc 11.1.
• Install Conda to set up an virtual environment Toturial.
• Install Pytorch with GPU support Toturial.
• Go to SCC_conv/, then python setup.py install.

Run

• Avaiable Models [model name]: VGG11, VGG13, VGG16, VGG19, ResNet18, ResNet34, ResNet50, MobileNet.
• Avaiable groups [num_group]: 1,2,4,8.
• Avaiable overlap [overlap_ratio]: 0.25,0.33,0.50,0.75.
• Then execute python main.py --model [Model Name] --groups [num_group] --overlap [overlap_ratio] .
• A detailed example of changing backend convolution implementation is illustrated in models/vgg.py.

Cite

@inproceedings{DSXplore,
  title={DSXplore: Optimizing Convolutional Neural Networks via Sliding-Channel Convolutions},
  author={Yuke Wang and Boyuan Feng and Yufei Ding},
  booktitle={2021 IEEE International Parallel and Distributed Processing Symposium (IPDPS)},
  year={2021}
}

Reference

Train CIFAR10 with PyTorch