README

This is the implementation of paper “Error Correction Coding for One-Bit Quantization with CNN-Based AutoEncoder”.

0、Citation

@ARTICLE{9791347,
  author={Zeng, Rui and Lu, Zhilin and Wang, Jintao and Song, Jian},
  journal={IEEE Communications Letters}, 
  title={Error Correction Coding for One-Bit Quantization With CNN-Based AutoEncoder}, 
  year={2022},
  volume={26},
  number={8},
  pages={1814-1818},
  doi={10.1109/LCOMM.2022.3181502}}

1、Requirements

This project involves the joint programming of MATLAB and Python, and thus package matlab.engine is need for getdata.py (generate data) and system_main.py (test bit error rate).

Find the path like MATLAB\R2017b\extern\engines\python in your MATLAB installation path, and execute the command python setup.py install.

other requirements:

pytorch >= 1.7.1
torchvision >= 0.8.2
python >= 3.6

2、Tree Arrangement

home
├── data
│   ├── mydataset.py
│   ├── gen_data_4.npz                        (Data Files, generated by getdata.py, for QPSK modulation)
│   ├── gen_data_16.npz                       (Data Files, generated by getdata.py, for 16QAM modulation)
│   ├── *.npz                                 (BER Files, generated by system_main.py)
│   ├── model_cnn/                              (Model Files, generated by system_cnn.py)
├── models 
│   ├── ECCNet.py                                   (Encoder and Decoder) 
│   ├── quantization.py                              (soft and hard quantization)
├── setting 
│   ├── settings.py                                  (random seed, gpu)
├── tools
│   ├── draw_BER.py
│   ├── draw_loss.py
│   ├── logger.py
│   ├── parse.py
│   ├── utils.py
├── traditional
│   ├── channel.py
│   ├── match_filtering.py
│   ├── pulse_shaping.py
│   ├── r_filter.py                              (Root sign rising cosine function)
├── getdata.py
├── system_cnn.py
├── system_main.py
...

3、Pipline

A. Data Preparation

execute

python getdata.py -modem_num 4
python getdata.py -modem_num 16

Generated data files are data/gen_data_4.npz and data/gen_data_16.npz. We have generated these two files and you can use them directly.

Modifiable Parameters:

• train_len: train_len × len is total training length.
• test_len: test_len × len is total testing length.
• val_len: val_len × len is total validating length.
• ber_len: ber_len × len is final testing length.
• len: Turbo code input 6144 bit
• code_len: Turbo code output 18444 bit
• train_cut: generate training dataset with dimension of [train_cut, N]
• test_cut: generate testing dataset with dimension of [test_cut, N]
• val_cut: generate validation dataset with dimension of [val_cut, N]

complete command:

python getdata.py -train_len 300 -test_len 200 -val_len 100 -ber_len 100 -len 6144 -code_len 18444 -train_cut 40000 -test_cut 25000 -val_cut 10000 -modem_num 4

B. Train/Test CNN Network

execute

python system_cnn.py

Generate model files are in data/model_cnn_awgn_4/ (or data/model_cnn_awgn_16).

Modifiable Parameters:

• epoch: total epoch
• batch_size: batch size
• mode: 'train' or 'test'
• G: G-fold symbold extension
• N: Block of symbols
• modem_num: modulation order (4 for QPSK, and 16 for 16QAM)
• unit_T: unit increment of soft quantization
• lr: learning rate of Autoencoder
• lr_step: decay step of learning rate
• snr: Eb/n0 when training
• snr_start: start of Eb/n0 when testing
• snr_step: step of Eb/n0 when testing
• snr_end: end of Eb/n0 when testing

complete command:

QPSK

python system_cnn.py -channel_mode awgn -modem_num 4 -snr 5 -lr 5e-4

16QAM

python system_cnn.py -channel_mode awgn -modem_num 16 -snr 10 -lr 1e-4

C. BER measurement

execute

python system_main.py

Result of BER is in data/unquantized_4.npz (or like this).

Modifiable Parameters:

• ber_len: ber_len × len is final testing length
• curve: measure BER based on 'unquantized'、'quantized'、'cnn'
• G: G-fold symbold extension
• N: Block of symbols
• modem_num: modulation order (4 for QPSK, and 16 for 16QAM)
• snr_start: start of Eb/n0 when testing
• snr_step: step of Eb/n0 when testing
• snr_end: end of Eb/n0 when testing
• model_path: path of trained model

complete command:

QPSK

python system_main.py -channel_mode awgn -curve cnn -modem_num 4 -snr_start -1 -snr_end 3 -snr_step 1 -model_path ./data/model_cnn_awgn_4

16QAM

python system_main.py -channel_mode awgn -curve cnn -modem_num 16 -snr_start -2 -snr_end 13 -snr_step 1 -model_path ./data/model_cnn_awgn_16