📄 Read Our Paper 📄

High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning

Table of Contents

• Introduction
• Features
• Installation
• Usage
  • Training
• Dataset
• Results
• Citations

Introduction

This repository implements the methods presented in our paper:

"High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning"

Our work focuses on enhancing interference cancellation in edge devices by leveraging depthwise separable convolutions, quantization, and pruning techniques to achieve high throughput and efficiency.

Features

• Depthwise Separable Convolutions: Reduces computational complexity without compromising performance.
• Quantization: Lowers model size and inference latency.
• Pruning: Eliminates redundant parameters for optimized performance.
• Scalable Training: Supports distributed training across multiple devices.
• Custom DataLoader: Tailored for the ICASSP 2024 SP Grand Challenge dataset with unknown interference types.

Installation

1. Clone the repository:

   git clone https://github.com/M0574F4/Fast_CCI.git
   cd Fast_CCI
   

2. Create a virtual environment (optional but recommended):

   python3 -m venv venv
   source venv/bin/activate
   

3. Install the required packages:

   pip install -r requirements.txt
   

Usage

Training

To start training the model, use the train.py script with the desired parameters:

python train.py \
 dataloader.batch_size=<batch_size> \
 model.model_type=<model_type> \
 model.enc_block_type=<enc_block_type> \
 model.bottleneck_type=lstm \
 model.encoder_filters=<encoder_filters> \
 model.hidden_lstm_size=<hidden_lstm_size> \
 model.conv_dim=<conv_dim> \
 trainer.max_steps=<max_steps> \
 trainer.backward_option=<backward_option> \
 dataloader.sig_len=<sig_len> \
 distributed.n_devices=<n_devices> \
 distributed.strategy=<strategy>

Dataset

We utilize the ICASSP 2024 SP Grand Challenge: Data-Driven Signal Separation in Radio Spectrum dataset. This dataset provides a challenging single-channel signal separation problem designed to push the boundaries of data-driven techniques in radio spectrum processing.

Customization

Unlike the original challenge, our approach assumes unknown interference types, making the task more generalized and applicable to real-world scenarios. A custom DataLoader is included in this repository to handle this specific scenario, providing efficient data preprocessing and augmentation.

For more information about the dataset and its usage, visit the ICASSP Challenge page.

Results

Our model demonstrates significant performance improvements across multiple metrics and showcases scalability for edge-device deployments. Below are key results from our research:

Key Findings

• Figure 4: Scatter plot of MSE score versus MACs for different models. The size of the circles is proportional to the number of parameters of each model.

  

• Figure 7: This graph illustrates the increasing inference throughput as the batch size is expanded, highlighting the scalability and efficiency gains achieved during batch processing.

  

For a comprehensive analysis and additional results, please refer to the paper.

Citations

If you use this work in your research, please cite the following paper:

@article{naseri2024high,
  title={High-Throughput Blind Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning},
  author={Naseri, Mostafa and De Poorter, Eli and Moerman, Ingrid and Poor, H Vincent and Shahid, Adnan},
  journal={arXiv preprint arXiv:2411.12541},
  year={2024}
}