Distributed micro-split deep learning enables flexible partitioning and distributed computing of deep neural networks in heterogeneous dynamic IoT for low-latency and energy-efficient cooperative Deep Learning operations. DISNET iteratively combines vertical (layerwise) and horizontal DNN partitioning for flexible, distributed, and parallel execution of neural network models in IoT systems with heterogeneous computing capabilities and network conditions without compromising accuracy.
The repository contains the source code of DiSNet. The code is organized as follows:
The folder main contains the files for the core implementation of the DiSNet framework.
Test deep neural network models are included in the models folder.
The file network.py contains the implementation of the IoT mesh networks for testing, saved in the networks folder.
In the file main/model_convert.py modify the first part of the model_path
model_path = "/home/eric/.cache/torch/hub/checkpoints/vgg16_bn-6c64b313.pth".
Create an instance of the Vgg-16 model by running the model_vgg16.py file in the main folder.
python3 models/model_vgg16.py
To generate the parameters of a pre-trained model, run the model_convert.py file.
python3 main/model_convert.py
Run the file run_DiSNet.py to test the inference time, energy consumption, and inference accuracy.
python3 main/run_DiSNet.py
Paper: https://doi.org/10.1109/JIOT.2023.3313514
Please cite the paper as follows: Samikwa, Eric, Antonio Di Maio, and Torsten Braun. "DISNET: Distributed Micro-Split Deep Learning in Heterogeneous Dynamic IoT." IEEE Internet of Things journal (2023).
@article{samikwa2023disnet,
title={Disnet: Distributed micro-split deep learning in heterogeneous dynamic iot},
author={Samikwa, Eric and Di Maio, Antonio and Braun, Torsten},
journal={IEEE Internet of Things journal},
year={2023},
publisher={IEEE}
}