Overview

This is the official implementation of IMWUT/Ubicomp paper Context-Aware Compilation of DNN Training Pipelines across Edge and Cloud.

Requirements

python 3.x, pytorch, torchvision, pandas

DataSets

default dataset path is data. CIFAR10 can be automatically downloaded with torchvision tools. IMDB can be automatically downloaded with keras tools. Tiny-ImageNet please refer to http://cs231n.stanford.edu/tiny-imagenet-200.zip. Chair downloading guidelies is in generative/readme.txt. MovieLens 1M dataset is included in DIN.

Running our system

client.py is deployed on the edge and the server.py is deployed on the cloud

Discriminative model

To run the discriminative model (ResNet, Vgg, MobilNet, BI-LSTM), the inplementation is under discriminative

First start server

python3 server.py --ip your-ip --port your-port --hyperport your-port-2

and then start client

python3 client.py --ip your-ip --port your-port --hyperport your-port-2

and

Then you can run our system. For various models and use cases, the setting commands are different, you can refer to them under discriminative folder

Generative model

To run the generative model (Chair), the inplementation is under generative First start server

python3 server.py --ip your-ip --port your-port --hyperport your-port-2

and then start client

python3 client.py --ip your-ip --port your-port --hyperport your-port-2

Cross-Slio Federated Learning

To run the Cross-Slio Federated Leaningr model such as DIN, the inplementation is under DIN

First start server

python3 server.py --ip your-ip --portA port-for-ClientA --portB port-for-ClientB

Then start clientA

python3 clientA.py --ip your-ip --portA port-for-ClientA

Finally start cientB

python3 clientB.py --ip your-ip --portA port-for-ClientB

Other Useful tools

Baselines

We also provide codes for checking our baselines in baseline folder.

base.py is running the original model without any advanced methods, the user command setting for different models and datasets are diffefrent you can refer to them detailedly under baseline

JointDNN can be checked by just set client.py with

--stale_it 0

Ablation Study

To check the impacts of using feature replay error feedback, we can also use base.py to do the abaltion study.

To check feature replay, set base.py with

--stale_it K --split_index le

To check error feed back, set base.py with

--use_EF 1

Also, we can change the compression methods in both baseline and acutal system implementations. To investigate about weights distribution and track weights, set base.py with

--track_weights 1 --track_point 0,1,2,3,....

Profile and Measurements

Energy Consumption and Memory cost are critical for edge devices. So, we provide code to test energy consumption on TX2 in quantify/energy.py and memory test in quantify/memory.py

We also provide our profiling results under profile

Requirements

psutil, tkinter and matpolib (for visualization)

Citation

@article{yao2021context,
  title={Context-Aware Compilation of DNN Training Pipelines across Edge and Cloud},
  author={Yao, Dixi and Xiang, Liyao and Wang, Zifan and Xu, Jiayu and Li, Chao and Wang, Xinbing},
  journal={Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies},
  volume={5},
  number={4},
  pages={1--27},
  year={2021},
  publisher={ACM New York, NY, USA}
}