This repository contains the code and dataset used for the paper "Structured Nonuniform Pruning for Tiny Angle-of-Arrival Deep Learning Models" presented at IEEE ICC 2025. Our work introduces a novel, structured, nonuniform, block-wise pruning technique that enables efficient deployment of deep learning models for Angle of Arrival (AoA) estimation in resource-constrained IoT environments.
Deploying deep learning (DL) models on IoT devices is challenging due to their limited computational resources. This work proposes a structured, nonuniform pruning approach that selectively prunes model blocks based on their contribution to overall performance. Applied to MobileNetV3 for AoA estimation, our method achieves:
- 35.9X reduction in parameter count
- 11.5X reduction in model size
- 7.7X reduction in multiply-accumulate operations (MACs)
while maintaining minimal impact on accuracy. We validate our approach using a real-world dataset collected from a software-defined radio (SDR) testbed.
The figure below illustrates the structured nonuniform pruning process:
- Structured Block-Wise Pruning: Prunes entire blocks rather than individual parameters to maintain hardware efficiency.
- Nonuniform Importance-Based Pruning: Determines pruning intensity per block using a proposed importance metric.
- Minimal Accuracy Degradation: Ensures a balance between model compression and performance.
- Optimized for IoT & SDR-Based AoA Estimation: Enables real-time inference on edge devices with limited resources.
Our dataset consists of I/Q samples collected from a real-world SDR testbed, covering multiple modulations and sample rates.
- Transmitter: USRP X300 with 5.88 GHz antenna
- Receiver: Two synchronized USRP X300s (four-channel I/Q data)
- 500GB dataset spanning six modulation schemes at two sampling rates (1M & 10M samples/sec)
- Azimuth & elevation metadata included for AoA estimation
To use this repository, install the required dependencies:
pip install -r requirements.txtpython python model_compress.py --device cuda --max_pruning_ratio 0.99python python produce_results.pyKey findings from our structured nonuniform pruning approach:
| Metric | Baseline Model | Pruned Model | Improvement |
|---|---|---|---|
| Parameter Count | 548K | 15.26K | 35.9X Reduction |
| Model Size | 2.1MB | 180KB | 11.5X Reduction |
| MACs | 24.6M | 3.2M | 7.7X Reduction |
If you find this work useful, please cite our paper:
@inproceedings{hallaq2025tinyAoA,
author = {Mohammad Hallaq and Elsayed Mohammed and Fazal Muhammad Ali Khan and Alec Digby and Pasquale Leone and Ashkan Eshaghbeigi and Hatem Abou-Zeid},
title = {Structured Nonuniform Pruning for Tiny Angle-of-Arrival Deep Learning Models},
booktitle = {IEEE International Conference on Communications (ICC)},
year = {2025},
}This research was supported by Qoherent Inc., MITACS Accelerate, and Alberta Innovates. We also thank the University of Calgary WAVES Lab for providing computational resources.
For questions or collaborations, contact Mohammad Hallaq at mohammad.hallaq@ucalgary.ca or create an issue in this repository.