Edge_Inference_three-step_framework

This code is for the paper: "communication-computation tradeoffs in resource-constrained edge inference".

Framework

We propose a three-step framework to reduce the end-to-end latency in the edge inference. The structure is shown as follows:

Implementation

The three-step framework is implemented based on ResNet18. The network is split at the end of each building block.

Dependency

Pytorch 1.6.0
Torchvision 0.7.0
Numpy 1.6.12

Dataset

CIFAR-10

How to Run

0. Pretrain a ResNet model from scratch or download from here.
1. model_pruning.py uses magnitude-based pruning method to compress the on-device model with parameter -sp1, -sp2, -sp3, -sp4, and -sp5 to denote the Sparsity ratios in different building blocks. The pruned model is named sp1_xx_sp2_xx_sp3_xx_sp4_xx_test_acc_xxxx.
2. feature_compression.py compresses the intermediate feature, which loads the pruned model by -load, and selects the split point by -split. Besides, -hid_dim determines the intermediate feature dimension, and -bit n means n-bit quantization.

Citation

@article{shao2020communication,
  title={Communication-computation trade-off in resource-constrained edge inference},
  author={Shao, Jiawei and Zhang, Jun},
  journal={IEEE Communications Magazine},
  volume={58},
  number={12},
  pages={20--26},
  year={2020},
  publisher={IEEE}
}