This is the PyTorch implementation of paper “Enhancing Automatic Modulation Recognition through Robust Global Feature Extraction”.

pytorch
yacs
h5py
pickle
matplotlib
thop
ptflops
fvcore

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

1. preparing dataset: download the datasets RadioML2016.10a and RadioML2018.01a on DeepSig's official website, and form the file path like dataset/RML2016.10a_dict.pkl and dataset/RML2018.01.hdf5.

2. training and testing the model: run python main.py --config xxx. e.g.python main.py --config configs/transformerlstm_18.yaml We have devised a training strategy using DistributedDataParallel (DDP). To train using the DDP approach, set the parameter ddp to True in the config and run CUDA_VISIBLE_DEVICES=xxx python -m torch.distributed.launch --nproc_per_node=xxx main.py --config configs/transformerlstm_18.yaml. e.g. CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 main.py --config configs/transformerlstm_18.yaml

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

1. preparing dataset: download the datasets RadioML2016.10a and RadioML2018.01a on DeepSig's official website, and form the file path like 'dataset/RML2016.10a_dict.pkl' and 'dataset/RML2018.01.hdf5'.

2. preparing models: We save the well-trained weights of TLDNN in results/TransformerLSTM/best/RML2016/checkpoints/best_acc.pth for RadioML2016.10a and results/TransformerLSTM/best/RML2018/checkpoints/best_acc.pth for RadioML2018.01a. Additionally, some of the visualized results have also been saved. The prepared models can be used directly.

3. modifying settings: set the parameter train From True to False in configs/transformerlstm_16.yaml and configs/transformerlstm_18.yaml. Then run python main.py --config configs/transformerlstm_16.yaml and python main.py --config configs/transformerlstm_18.yaml to get the expermential results.

   e.g.

   modes:  
       method: 'TransformerLSTM'
       path: 'best' # save address
       loss: 'loss_CE'
       train: False # set the parameter: from True to False
       ddp: False # choose whether to use ddp trainint
   data_settings:  
       dataset: 'RML2016'
       Xmode:
           type: 'AP' # input type for the model
           options:
               IQ_norm: False # choose whether to normalized the input signal
               zero_mask: False
               random_mix: False #choose whether to use the random mixing strategy
       mod_type: ['8PSK', 'AM-DSB', 'AM-SSB', 'BPSK', 'CPFSK', 'GFSK', 'PAM4', 'QAM16', 'QAM64', 'QPSK', 'WBFM']
   opt_params: 
       batch_size: 128
       epochs: 150
       lr: 1e-3
       workers: 8
       seed: 1
       gpu: 0
       cpu: False
       early_stop: False

   modes:  
       method: 'TransformerLSTM'
       path: 'best'
       loss: 'loss_CE'
       train: False # set the parameter: from True to False
       ddp: False
   data_settings:  
       dataset: 'RML2018'
       Xmode:
           type: 'AP'
           options:
               IQ_norm: False 
               zero_mask: False
               random_mix: False
       mod_type: [ '32PSK', '16APSK', '32QAM', 'FM', 'GMSK', '32APSK', 'OQPSK', '8ASK', 'BPSK', '8PSK', 'AM-SSB-SC',
       '4ASK', '16PSK', '64APSK', '128QAM', '128APSK', 'AM-DSB-SC', 'AM-SSB-WC', '64QAM', 'QPSK', '256QAM',
       'AM-DSB-WC', 'OOK', '16QAM' ]
   opt_params:  
       batch_size: 400
       epochs: 150
       lr: 1e-3
       workers: 8
       seed: 1
       gpu: 0
       cpu: False
       early_stop: False