Active Deep Decoding with Importance Sampling

Efficient active deep decoding of linear codes using importance sampling and weighted belief propagation.

This repository provides a reproducible TensorFlow implementation of the framework proposed in:

Efficient Active Deep Decoding of Linear Codes Using Importance Sampling IEEE Communications Letters, 2025.

The code implements active weighted belief propagation (WBP) training combined with importance sampling (IS) to improve decoding performance in the high-SNR regime.

---

Overview

This repository contains:

• Weighted belief propagation (WBP) decoder
• Active training procedure
• Importance sampling–based noise generation
• Monte-Carlo BER/FER simulation framework
• Utilities for ALIST → parity-check matrix conversion

The implementation supports arbitrary linear block codes defined by a parity-check matrix.

---

Repository Structure

codes/              # ALIST / parity-check matrix files
Alist2bin.py        # Convert ALIST to binary PCM
BPSio.py            # LDPC / Weighted BP decoder layer
Gnoise.py           # Gaussian + Importance Sampling noise generation
train.py            # Training script
sim.py              # BER/FER simulation script
LICENSE
README.md

---

Requirements

Python 3.8+ recommended.

Main dependencies:

tensorflow >= 2.8
numpy
scipy
matplotlib

Install with:

pip install tensorflow numpy scipy matplotlib

---

Training

The training script implements active learning with importance sampling.

Example:

python train.py

Key configurable parameters inside train.py:

• min_snr, max_snr
• batch_size
• M (importance sampling grid size)
• num_itr (number of BP iterations)

Models are saved in the models/ directory.

---

Simulation (BER / FER)

To evaluate a trained model:

python sim.py

This runs Monte-Carlo simulation over an SNR range and outputs:

• BER curves
• FER curves
• CSV result files
• Plots saved under figs/

The simulation compares:

• Active weighted BP (trained model)
• Standard BP decoding

---

Using Custom Codes

To use a custom code:

1. Place the ALIST file in codes/
2. Load it using:

from Alist2bin import Alist2bin
pcm = Alist2bin("codes/your_code.alist")

Or load .mat files as shown in sim.py.

---

Reproducibility Notes

• All simulations assume BPSK modulation.
• All-zero codeword transmission is used (standard in coding simulations).
• Random seeds can be added for deterministic behavior.

---

Citation

If you use this code in your research, please cite:

@article{ActiveDeepDecoding2025,
  title={Efficient Active Deep Decoding of Linear Codes Using Importance Sampling},
  journal={IEEE Communications Letters},
  year={2025}
}

---

License

This project is released under the MIT License. See LICENSE for details.

---

Contact

For questions, suggestions, or collaboration inquiries, please open a GitHub issue.

For direct communication, please email: [email protected]