lynshao / attentioncode

Source code and well-trained models of AttentionCode

AttentionCode

This repository is the official implementation of paper [AttentionCode: Ultra-Reliable Feedback Codes for Short-Packet Communications] (https://arxiv.org/abs/2205.14955).

If you find this repository useful, please kindly cite as

@article{AttentionCode,

title={Attentioncode: Ultra-reliable feedback codes for short-packet communications},

author={Shao, Yulin and Ozfatura, Emre and Perotti, Alberto and Popovic, Branislav and Gunduz, Deniz},

journal={IEEE Transactions on Communications},

year={2023}

}

Requirements

Experiments were conducted on Python 3.8.5. To install requirements:

pip install -r requirements.txt

Well-trained models

Some well-trained models given. The achieved results are as follows:

Noiseless feedback:

Feedforwad SNR	Feedback SNR	BLER
-0.5 dB	100 dB	4.33e-6
0 dB	100 dB	1.17e-7
0.5 dB	100 dB	4.17e-9

Noisy feedback:

Feedforwad SNR	Feedback SNR	BLER
0 dB	20 dB	1.16e-4
1 dB	20 dB	1.92e-7

To reproduce the results, please run

python main.py --snr1 [input] --snr2 [input] --train 0 --batchSize 100000

Performance

Noiseless feedback:

Noisy feedback:

attentioncode's Issues

You are using K+1 bits, that means you are using zero bit in the end which is not in journal

you have not added curriculum learning and training test SNR mismatch in the code. Please check

Directly adding both noises

When we pass signal to receiver in AWGN channel, noise power gets added to signal power. we need to rebalance this power by some scale at receiver, that you calculated by alpha in remark 1, but in code you are directly adding both forward and feedback noises. In reality, transmitter will calculate some scaled version of forward noise added with feedback noise.

What do the bv and core variables represent in the code?

Do we have to use power reallocation layer in the code to obtain the performance shown in the journal? Which one