RadioML

主要贡献为gpr去噪，使用“--denoising_method efficient_gpr_per_sample”启用。“per_sample”表示使用每个样本自身数据而非所有样本均值数据，实际上两者速度几乎一样，故只保留了当前方法。

This repository contains machine learning models for radio modulation classification.

Usage

# preparation
pip install -r requirements
cd src

# evaluate
python main.py --models lightweight_hybrid --denoising_method efficient_gpr_per_sample --augment_data --mode evaluate
python main.py --models pet --denoising_method efficient_gpr_per_sample --augment_data --mode evaluate
python main.py --models mcldnn --denoising_method efficient_gpr_per_sample --augment_data --mode evaluate

# train and evaluate
python main.py --models lightweight_hybrid --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models micro_lightweight_hybrid --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models ulcnn --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models pet --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models mcldnn --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models cgdnn --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models mcnet --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models resnet --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data
python main.py --models complex_nn --epochs 200 --denoising_method efficient_gpr_per_sample --batch_size 256 --augment_data

To run the project, use src/main.py. Key command-line arguments include:

• --mode: Mode of operation. Choices: explore, train, evaluate, all. Default: all.
• --model_type: Model architecture to use. Choices: .
• --dataset_path: Path to the RadioML dataset. Default: ../RML2016.10a_dict.pkl.
• --epochs: Number of training epochs. Default: 400.
• --batch_size: Batch size for training. Default: 128.
• --augment_data: Enable data augmentation.
• --denoising_method: Denoising method to apply to the input signals. Default: gpr.
  • efficient_gpr_per_sample: Gaussian Process Regression. (Default kernel is RBF; Matern and RationalQuadratic also available).
  • none: No denoising is applied.

Project Structure

• src/: Source code for all models
  • Models implementations (CNN 1D, CNN 2D, Complex NN, ResNet, Transformer)
  • Utility functions and callbacks
• model_weight_saved/: Saved model weights (managed with Git LFS)
• output/models/: Output model files (managed with Git LFS)
• RML2016.10a_dict.pkl: The RadioML dataset (managed with Git LFS)
• projects/: Contains submodules of related projects
  • ULCNN: Implementation of the ULCNN architecture

Getting Started

To use this code, you will need to:

1. Download the RadioML dataset (RML2016.10a) from the official website
2. Place the dataset file in data
3. Run the training scripts to train the models or use pre-trained weights

Models

MCNET、CGDNN、ULCNN、MCLDNN、PETCGDNN、ComplexCNN、ResNet、Complex-ResNet-mini、Complex-ResNet。

Results

The following table shows the performance of different neural network architectures with GPR denoising and data augmentation techniques:

Model	Accuracy (%)	+Aug (%)	+GPR (%)	+Aug+GPR (%)
MCNET	56.73	57.28	58.82	61.78
CGDNN	53.77	59.46	58.84	61.08
ULCNN	56.15	57.26	58.23	62.28
ComplexCNN	57.11	58.07	61.78	62.89
ResNet	55.37	59.34	62.15	63.24
MCLDNN	60.27	61.71	64.65	65.17
PETCGDNN	60.72	61.98	63.28	65.26
Complex-ResNet-mini	57.06	58.74	60.58	64.94
Complex-ResNet	56.94	59.60	61.49	65.52

Notes:

• Aug: Rotation data augmentation
• GPR: Gaussian Process Regression denoising
• Bold values indicate the best performance in each column
• The best overall accuracy of 65.52% is achieved by Complex-ResNet with both augmentation and GPR denoising