About the project

This repository includes the codes of two papers:

1. Complexity-aware Adaptive Training and Inference for Edge-Cloud Distributed AI Systems
2. Conditionally Deep Hybrid Neural Networks Across Edge and Cloud Our project is built on the framework of Distiller. Please refer to the website (https://github.com/NervanaSystems/distiller.git) for instructions for installation and environment settings. Clone the project repository from github:

$ git clone https://github.com/yinghanlong/Complexity-aware-AI.git

Getting Started

After setting up the enviroment, you can run experiments following the example commands in run.sh. For example, to train an adaptive neural network based on the pretrained ImageNet, you can use the following.

$ source env/bin/activate
$ cd examples/classifier_compression/
$ python Block-Train-ImageNet-pretrain.py --arch=resnet18_p --epoch=110 -b 256 --lr=0.01 -j 1 --out-dir . -n imagenet . --earlyexit_lossweights 0.3 --earlyexit_thresholds 0.8 --deterministic --gpus=1

To train MobileNetV2,

$python Mobilenet-extend.py --arch=mobilenet_v2 --epoch=180 -b 128 --lr=0.01 -j 1 --out-dir . -n mobilenet . --deterministic --gpus=2 --earlyexit_lossweights 0.3 --earlyexit_thresholds 0.8

Use --evaluate and --resume=model_dir to load a trained model and run evaluation.

For more details, there are files you can refer to:

• Models for CIFAR10/100 (modified ResNets into early exiting models and MEANet models which includes main, adaptive and extension blocks): /models/cifar10
• Models for ImageNet (ResNets and MobileNetV2):/models/imagenet
• Codes for training MEANet models which includes main, adaptive and extension blocks: /examples/classifier_compression/Block-Train-extend.py,/examples/classifier_compression/Mobilenet-extend.py,/examples/classifier_compression/Block-Train-ImageNet-pretrained.py
• Codes for training hybrid quantized models with early exits:/examples/classifier_compression/early-exit-classifier.py
• Setting K-bit quantization or binarization for specific layers:/examples/classifier_compression/util_bin.py
• Examples of hard classes of ImageNet/CIFAR100: examples/classifier_compression/mobilenet_imagenet/hard_classes.pickle. examples/classifier_compression/resnet32_hardclass/hard_classes.pickle

We will add more explantions and comments later. Please email me [email protected] if you have any questions regarding the project or the codes.

Built With

• PyTorch - The tensor and neural network framework used by Distiller.

Citation

If you used for your work, please use the following citation:

@INPROCEEDINGS{9546405,

  author={Long, Yinghan and Chakraborty, Indranil and Srinivasan, Gopalakrishnan and Roy, Kaushik},

  booktitle={2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS)}, 

  title={Complexity-aware Adaptive Training and Inference for Edge-Cloud Distributed AI Systems}, 

  year={2021},

  pages={573-583},

  doi={10.1109/ICDCS51616.2021.00061}}

@misc{https://doi.org/10.48550/arxiv.2005.10851,
  doi = {10.48550/ARXIV.2005.10851},
  
  url = {https://arxiv.org/abs/2005.10851},
  
  author = {Long, Yinghan and Chakraborty, Indranil and Roy, Kaushik},
  
  title = {Conditionally Deep Hybrid Neural Networks Across Edge and Cloud},
  
  publisher = {arXiv},
  
  year = {2020},
  
}