ColabNAS

A hardware-aware neural architecture search (HW NAS) algorithm targeting low-RAM microcontrollers. It provides CNNs that run on low-RAM microcontrollers, e.g. 40 kiB, using online GPU programs like Google Colaboratory or Kaggle Kernel.

If you want a faster and similar HW NAS at the expense of being less repeatable and precise, check NanoNAS.

News

• 2023/11 ColabNAS has been accepted for publication in the Future Generation Computer Systems journal! Experimental Data is available in the form of Jupyter Notebook.
• 2023/09 ColabNAS has been cited in a tinyML talk on YouTube.

An overview of its performances

This section evaluates the hardware-aware feature of ColabNAS by providing results for different hardware targets. To highlight the ColabNAS feature of providing lightweight CNNs, three low-RAM STMicroelectronics (STM) MCUs from the Ultra-low Power series have been selected: the L010RBT6, the L151UCY6DTR and the L412KBU3.

STM32 MCU	RAM	Flash	CoreMark
L010RBT6	20 kiB	128 kiB	75
L151UCY6DTR	32 kiB	256 kiB	93
L412KBU3	40 kiB	128 kiB	273

The values in the previous table set the constraints for running ColabNAS on each target. RAM and Flash were provided as is, while the MACC upper bound was obtained by multiplying by $10^{4}$ the CoreMark score of the target, in order to allow a fair exploration of the search space. 50x50x3 input size has been used to cope with MCUs' constrained resources. The results for the Visual Wake Words dataset, a standard TinyML benchmark, follow.

Target	TFlite Model	Accuracy	RAM	Flash	MACC	Search Cost
L010RBT6	link	69.4%	19 kiB	8.02 kiB	227 k	2h:11m
L151UCY6DTR	link	74.5%	22.5 kiB	18.5 kiB	657 k	3h:04m
L412KBU3	link	77.8%	33 kiB	44.9 kiB	2,086 k	2h:47m

For more experimental results refer to the following paper, which you can find here, and to this Google Colaboratory notebook.

How to use

Import the ColabNAS notebook in Google Colaboratory and try it out!

Citation

If you find the project helpful, please consider citing our paper:

@article{GARAVAGNO2024152,
    title = {ColabNAS: Obtaining lightweight task-specific convolutional neural networks following Occam’s razor},
    author = {Andrea Mattia Garavagno and Daniele Leonardis and Antonio Frisoli},
    journal = {Future Generation Computer Systems},
    volume = {152},
    pages = {152-159},
    year = {2024},
    issn = {0167-739X},
    doi = {https://doi.org/10.1016/j.future.2023.11.003}
}