README

This is the PyTorch implementation of paper “Convolutional Neural Network Assisted Transformer for Automatic Modulation Recognition under Large CFOs and SROs”.

Citation

@ARTICLE{10458301,
  author={Zeng, Rui and Lu, Zhilin and Zhang, Xudong and Wang, Jintao and Wang, Jian},
  journal={IEEE Signal Processing Letters}, 
  title={Convolutional Neural Network Assisted Transformer for Automatic Modulation Recognition Under Large CFOs and SROs}, 
  year={2024},
  volume={31},
  number={},
  pages={741-745},
  keywords={Transformers;Convolution;Modulation;Feature extraction;Task analysis;Receivers;Frequency modulation;Automatic modulation recognition;carrier frequency offsets;sample rate offsets;transformer;group convolution},
  doi={10.1109/LSP.2024.3372770}}

Requirements

pytorch
yacs
h5py
matplotlib
thop  

Architecture

home
├── amr/
│   ├── dataloaders/
│   ├── models/
│   │   ├── losses/
│   │   ├── networks/
│   ├── utils/
│   │   ├── __init__.py
│   │   ├── config.py
│   │   ├── draw.py
│   │   ├── init.py
│   │   ├── logger.py
│   │   ├── solver.py
│   │   ├── static.py
├── configs/  (hyperparameters)
│   ├── *.yaml
├── main.py
├── datasets/
├── results/

Quick Start

If you want to train a network from scratch, please follow these steps:

1. preparing dataset: download the dataset with large-scale offsets dataset.rar, and form the file path like 'dataset/cfo.hdf5'

2. training and testing: run python main.py --config xxx. e.g.python main.py --config configs/transgroupnet_cfo.yaml

3. checking the results: check the well-trained models and the figures in results/

Result Reproduction

1. preparing dataset: download the dataset with large-scale offsets dataset.rar, and form the file path like 'dataset/cfo.hdf5'

2. preparing models: download the prepared models and results in results.rar, and extract to the current path

3. modifying settings: change the state of train From True to False in configs/transgroupnet_cfo.yaml, and run python main.py --config configs/transgroupnet_cfo.yaml to get the expermential results.

   e.g.

   method: 'ours'
   train: False  # change the state: from True to False
   dataset: 'cfo'
   mod_type: ["BPSK", "QPSK", "8PSK", "PAM4", "QAM16", "QAM32", "QAM64", "QAM128", "QAM256", "GFSK", "WBFM", "AM-DSB", "AM-SSB", "OOK", "4ASK", "8ASK", "16PSK", "32PSK","8APSK","GMSK", "DQPSK","16APSK","32APSK","64APSK","128APSK"]
   workers: 8
   seed: 1
   gpu: 0
   cpu: False
   params:
       "network": "TransGroupNet"
       "loss": "loss_CE"
       "batch_size": 1024
       "epochs": 200
       "lr": 5e-3
       "lr_decay": 0.8
       "weight_decay": 5e-2
       "early_stop": False
       "Xmode": [{"type":"APF","options":{"IQ_norm":False, "zero_mask":False}}]