C3-SL: Circular Convolution-Based Batch-Wise Compression for Communication-Efficient Split Learning

This repository provides the official Pytorch implementation for C3-SL.

📎 Paper Link ✏️ Citations

• Batch-Wise Compression (A new Compression Paradigm for Split Learning)
• Exploit Circular Convolution and Orthogonality of features to avoid information loss
• Reduce 1152x memory and 2.25x computation overhead compared to the SOTA dimension-wise compression method

---

Presentation Video : Click here

---

Table of Contents

• Prepare Dataset
• Usage
• Experiment Overview
• Citations

---

📚 Prepare Dataset

The source code can be found in CIFAR-10/data preprocess src and CIFAR-100/data preprocess src.

Tasks	Datasets:point_down:
Image Classification	CIFAR-10, CIFAR-100

• Use the following commands:

$ cd CIFAR10  # or cd CIFAR100
$ cd "data preprocess src"    
$ python download_data.py

• The data structure should be formatted like this:

CIFAR10(Current dir)
├──CIFAR
│   ├── train
│   ├── val
├──data preprocess src
│   ├── download_data.py

🏃 Usage - Training

Requirements

• Python 3.6
• Pytorch 1.4.0
• torchvision
• CUDA 10.2
• tensorflow_datasets
• Other dependencies: numpy, matplotlib, tensorflow

Training

Modify parameters in the shell script.

Parameters	Definition
--batch	batch size
--epoch	number of training epochs
--dump_path	save path of experiment logs and checkpoints
--arch	model architecture (resnet50/vgg16)
--split	the split point of model
--bcr	Batch Compression Ratio R

cd CIFAR10/C3-SL
./train.sh

📈 Experiment Overview

• Comparable Accuracy with SOTA
• Greatly Reduce Resource Overhead

Citations

@article{hsieh2022c3,
  title={C3-SL: Circular Convolution-Based Batch-Wise Compression for Communication-Efficient Split Learning},
  author={Hsieh, Cheng-Yen and Chuang, Yu-Chuan and others},
  journal={arXiv preprint arXiv:2207.12397},
  year={2022}
}