cps2rl/nlp-on-embedded-devices

Benchmarking BERT-based NLP models on Embedded Devices (Raspberry Pi)

How to use the code?

We used the ROS2 platform for conducting Intent Classification (IC) and Name Entity Recognition (NER) tasks. Our implementation employed based on the pub-sub model of robot operation system (ROS). We have implemented the project on Raspberry Pi. Installation guide of ROS2 in Pi: https://docs.ros.org/en/foxy/How-To-Guides/Installing-on-Raspberry-Pi.html . You can find more information about the ROS pub-sub model in the provided link: https://docs.ros.org/en/foxy/Tutorials/Beginner-Client-Libraries/Writing-A-Simple-Py-Publisher-And-Subscriber.html
We have trained the Bert model offline and put it on the BERT folder inside Intent Classification or NER folder.
Terminal command for running pub-sub model:

IC

 ros2 run intenc talker

NER

 ros2 run ner talker

Saved model link:https://figshare.com/account/home#/projects/169256

To measure the energy we used UM25C energy meter.
To measure system memory consumption, we used @profile method of python.
Deatils can be found by analyzing the code of publisher_member_function.py

You can find comprehensive information on working with the ROS2 platform in the provided resource. Their documentation is highly regarded, making it the optimal source for acquiring in-depth knowledge about the ROS platform. Link: https://docs.ros.org/en/foxy/index.html

Data-set link

HuRic Dataset: https://github.com/crux82/huric
Go Emotions Dataset: https://www.kaggle.com/datasets/shivamb/go-emotions-google-emotions-dataset
WNUT'17 Dataset: https://github.com/leondz/emerging_entities_17
CoNLL Dataset: https://ebanalyse.github.io/NERDA/datasets/

Fig: Utterance processing steps of a voice-controlled embedded device.

Fig: Hardware Setup

The arXiv preprint of our findings is available here: https://arxiv.org/abs/2304.11520