Copyright (C) 2023 Politecnico di Torino, Italy. SPDX-License-Identifier: Apache-2.0. See LICENSE file for details.
Authors: Matteo Risso, Alessio Burrello, Giuseppe Maria Sarda, Luca Benini, Enrico Macii, Massimo Poncino, Marian Verhelst, Daniele Jahier Pagliari
If you use ODiMO in your experiments, please make sure to cite our paper:
@misc{risso2023precisionaware,
title={Precision-aware Latency and Energy Balancing on Multi-Accelerator Platforms for DNN Inference},
author={Matteo Risso and Alessio Burrello and Giuseppe Maria Sarda and Luca Benini and Enrico Macii and Massimo Poncino and Marian Verhelst and Daniele Jahier Pagliari},
year={2023},
eprint={2306.05060},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
To install the latest release:
$ git clone https://github.com/eml-eda/odimo
$ cd odimo
$ python setup.py install
The current version include examples based upon the following datasets:
Visit the directory with the task that you want to run (e.g., image_classification/tiny-imagenet).
To run both the search and fine-tune steps of ODiMO a shell script is provided (e.g., run_ic.sh).
In the case of the Tiny ImageNet task the shell script takes 7 argument as input. The relevant one are detailed within {__}, the other can be leaved as fixed:
$ source run_ic.sh {regularization strenght} {architecture} 64 {regularization target} now search ft
where {regularization strength} is a float number that controls the balance between task loss and regularization loss. {architecture} is the architecture that we want to optimize with the tool. All the supported architectures are listed in models directory. E.g., for Tiny ImageNet task use as architecture res18_pow2_diana_full. {regularization target} represents the hw-related metric that we want to optimize, the supported values are [power, latency, power-naive].
ODiMO is released under Apache 2.0, see the LICENSE file in the root of this repository for details.