You can deploy this model and start earning money today!

sohelimi/
cvnet-rf

Name: cvnet-rf
Availability: InStock

Views

No views yet

CVNet-RF: Complex-Valued Neural Networks for Automatic Modulation Classification

Overview

CVNet-RF compares two neural network approaches for Automatic Modulation Classification (AMC) — the task of identifying a radio signal's modulation scheme directly from raw I/Q samples, with no demodulation or hand-crafted features.

Model	Parameters	Best Val Acc	Acc / M params
ComplexCNN	379K	50.3%	132
RealCNN (baseline)	934K	54.4%	58

Key finding: ComplexCNN achieves comparable accuracy with 2.5× fewer parameters — delivering 2.3× better parameter efficiency. For edge deployment on FPGAs, UAVs, and embedded sensors, this difference is significant.

What is Automatic Modulation Classification?

Every wireless signal — WiFi, LTE, radar, satellite — uses a modulation scheme to encode information onto a radio wave. AMC is the task of identifying that scheme from raw I/Q samples alone. Applications include:

SIGINT triage — automated first-stage sorting of intercepted signals
Spectrum monitoring — real-time emitter cataloging across wideband spectrum
Electronic warfare — rapid classification for jamming prioritization
Cognitive radio — adaptive waveform selection and coexistence
RF device fingerprinting — extending AMC to emitter-level identification

Dataset

RadioML 2018.01A by DeepSig / GNU Radio

2.56 million samples
24 modulation classes
SNR range: −20 to +30 dB
Input shape: (2, 1024) — I and Q channels × 1024 timesteps

24 Modulation Classes: OOK, 4ASK, 8ASK, BPSK, QPSK, 8PSK, 16PSK, 32PSK, 16APSK, 32APSK, 64APSK, 128APSK, 16QAM, 32QAM, 64QAM, 128QAM, 256QAM, AM-SSB-WC, AM-SSB-SC, AM-DSB-WC, AM-DSB-SC, FM, GMSK, OQPSK

Model Architectures

ComplexCNN

Treats I/Q as a true complex number I + jQ. Uses Wirtinger calculus for backpropagation through complex operations, preserving phase relationships that real-valued networks fundamentally discard.

Input (2, 1024)
    ↓
Complex Conv1D (1→64, k=7)  + Complex BatchNorm + Complex ReLU + Complex MaxPool
    ↓
Complex Conv1D (64→128, k=5) + Complex BatchNorm + Complex ReLU + Complex MaxPool
    ↓
Complex Conv1D (128→256, k=3) + Complex BatchNorm + Complex ReLU + Complex MaxPool
    ↓
Concat [Real, Imag] → Flatten → FC(512) → Dropout → FC(24)
    ↓
Softmax → 24-class prediction

RealCNN (baseline)

Treats I and Q as two independent real-valued channels. Standard 1D CNN — no phase awareness.

Input (2, 1024)
    ↓
Conv1D (2→64, k=7)  + BatchNorm + ReLU + MaxPool
    ↓
Conv1D (64→128, k=5) + BatchNorm + ReLU + MaxPool
    ↓
Conv1D (128→256, k=3) + BatchNorm + ReLU + MaxPool
    ↓
Flatten → FC(512) → Dropout → FC(24)
    ↓
Softmax → 24-class prediction

Training Details

Setting	Value
Optimizer	Adam
Learning rate	1e-3 → 1e-6 (ReduceLROnPlateau)
Batch size	256
Epochs	100
Weight decay	1e-4
Label smoothing	0.1
Dropout	0.5
Augmentation	Phase rotation, AWGN injection
Hardware	Apple M-series (MPS backend)

Quick Start

Install dependencies

pip install torch numpy huggingface_hub

Run inference

python

from huggingface_hub import hf_hub_download
import torch
import numpy as np

# Download model
ckpt_path = hf_hub_download(repo_id="sohelimi/cvnet-rf", filename="complex_best.pt")

# Load and run inference
# See inference.py for full example

Or use the included inference script

bash

# Clone the repo
git clone https://huggingface.co/sohelimi/cvnet-rf
cd cvnet-rf

# Run on a demo synthetic signal
python inference.py --model complex

# Run on your own I/Q data (numpy array, shape (2,1024) or (1024,2))
python inference.py --model complex --iq your_signal.npy

Example output:

Device: mps
Loading complex model...

Top-5 predictions:
Rank   Modulation    Confidence
--------------------------------
  1    16-QAM          91.0%  ███████████████████████████
  2    64-QAM           5.0%  █
  3    QPSK             2.0%
  4    8-PSK            1.0%
  5    AM-DSB           1.0%

Decision: 16-QAM  (91.0% confidence)

Files

File	Description
`complex_best.pt`	ComplexCNN weights — best validation checkpoint (epoch 85)
`real_best.pt`	RealCNN weights — best validation checkpoint (epoch 100)
`inference.py`	Ready-to-run inference script with model definitions
`config.json`	Full training configuration and hyperparameters

Citation

If you use this model in your research, please cite:

bibtex

@misc{cvnet-rf-2025,
  author    = {Sohel Imi},
  title     = {CVNet-RF: Complex-Valued Neural Networks for Automatic Modulation Classification},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/sohelimi/cvnet-rf}
}

Dataset citation:

bibtex

@dataset{radioml2018,
  author    = {DeepSig Inc.},
  title     = {RadioML 2018.01A},
  year      = {2018},
  url       = {https://www.deepsig.ai/datasets}
}

License

MIT License — free to use for research and commercial purposes.

Trained on Apple Silicon (MPS) · PyTorch 2.x · RadioML 2018.01A