HYPERDOA

Hyperdimensional Computing for Direction-of-Arrival/Angle-of-Arrival Estimation

Official implementation of the paper: "HYPERDOA: Robust and Efficient DoA Estimation using Hyperdimensional Computing" (arXiv:2510.10718) - To appear at ICASSP 2026, Barcelona, Spain!

News

• [2025/10] HYPERDOA preprint is released.
• [2026/01] HYPERDOA is accepted at ICASSP 2026.

Overview

HYPERDOA is a lightweight, efficient implementation of HDC-based direction-of-arrival (DOA) estimation for uniform linear arrays.

Installation

Using uv (Recommended)

# Clone the repository
git clone https://github.com/pwh9882/HYPERDOA.git
cd HYPERDOA

# Create virtual environment and install
uv venv
uv pip install -e .

# Or install with dev dependencies
uv pip install -e ".[dev]"

Using pip

git clone https://github.com/pwh9882/HYPERDOA.git
cd HYPERDOA
pip install -e .

Requirements

• Python >= 3.10
• PyTorch >= 2.0
• torch-hd >= 0.6.0
• numpy >= 1.24
• matplotlib >= 3.7

Quick Start

from hyperdoa import HDCAoAModel, DOAConfig, evaluate_hdc
import torch

# Configure system
config = DOAConfig(N=8, M=3, T=100, snr=-5)

# Load your dataset (list of (X, Y) tuples)
# X: Complex tensor (N, T) - sensor observations
# Y: Tensor (M,) - ground truth DOA in radians
train_data = torch.load("data/train_dataset.pt")
test_data = torch.load("data/test_dataset.pt")

# Train and evaluate
loss, model = evaluate_hdc(
    train_data, test_data, config,
    feature_type="lag",  # or "spatial_smoothing"
    return_model=True,
)

print(f"Test MSPE: {loss:.2f} dB")

Dataset Generation

HYPERDOA uses datasets generated by SubspaceNet. The dataset format is directly compatible - no conversion needed.

Dataset Format

Each dataset is a list of (X, Y) tuples:

| Field | Type | Shape | Description | | ----- | ----------------- | -------- | ------------------------------- | | X | torch.complex64 | (N, T) | Sensor observations | | Y | torch.float64 | (M,) | Ground truth DOA in radians |

Generate Datasets Using Our Script

We provide a script that wraps SubspaceNet's data generation:

# Step 1: Clone SubspaceNet
git clone https://github.com/ShlezingerLab/SubspaceNet.git

# Step 2: Generate datasets with paper settings
uv run python scripts/generate_data.py \
    --subspacenet-path ./SubspaceNet \
    --output data/ \
    --N 8 --M 3 --T 100 \
    --snr -5 \
    --train-samples 45000 \
    --test-samples 2250

Example Configuration

The following parameters can be used as a starting point:

| Parameter | Value | Description | | --------------- | ------------------------ | -------------------------- | | N | 8 | Number of sensors | | M | 3 or 4 | Number of sources | | T | 100 | Number of snapshots | | SNR | -5 to -1 dB or 1 to 5 dB | Signal-to-noise ratio (dB) | | signal_nature | non-coherent / coherent | Signal correlation type | | train_samples | 45,000 | Training set size | | test_samples | 2,250 | Test set size |

Alternative: Use SubspaceNet Directly

You can also use SubspaceNet's own data generation:

# In SubspaceNet directory
import sys
sys.path.insert(0, '.')

from src.system_model import SystemModelParams
from src.data_handler import create_dataset

params = SystemModelParams()
params.N, params.M, params.T = 8, 3, 100
params.snr = -5
params.signal_nature = "non-coherent"

# Generate dataset
model_dataset, generic_dataset, samples_model = create_dataset(
    system_model_params=params,
    samples_size=45000,
    model_type="RawX",  # This gives us raw observations
    save_datasets=False,
)

# generic_dataset is directly compatible with HYPERDOA!
# Save it:
import torch
torch.save(generic_dataset, "train_dataset.pt")

Feature Types

| Feature Type | Description | | ------------------- | -------------------------------- | | lag | Mean spatial-lag autocorrelation | | spatial_smoothing | Spatial smoothing covariance |

Examples

# Generate data first
uv run python scripts/generate_data.py --subspacenet-path ./SubspaceNet

# Basic training
uv run python examples/train_basic.py --data-dir data/

API Reference

DOAConfig

Configuration dataclass for system parameters.

DOAConfig(
    N=8,              # Number of sensors
    M=3,              # Number of sources (3 or 4 in paper)
    T=100,            # Number of snapshots
    snr=-5.0,         # SNR in dB (-5 to -1 or 1 to 5 in paper)
    signal_nature="non-coherent"  # or "coherent"
)

HDCAoAModel

Main HDC model for DOA estimation.

model = HDCAoAModel(
    N=8, M=3, T=100,
    feature_type="lag",
    n_dimensions=10000,      # Hypervector dimensionality
    min_angle=-90.0,         # Minimum angle (degrees)
    max_angle=90.0,          # Maximum angle (degrees)
    precision=0.1,           # Angle resolution (degrees)
    min_separation_deg=6.0,  # Minimum peak separation
)

# Methods
model.train_from_dataloader(loader)  # Train from DataLoader
model.fit(X, y)                       # Train from tensors
model.predict(X)                      # Predict angles (radians)
model.predict_multi(X, k=2)           # Multi-source prediction
model.predict_logits(X)               # Raw classification scores
model.compute_mspe_db(loader)         # Compute MSPE (dB)

evaluate_hdc

Convenience function for training and evaluation.

loss, model = evaluate_hdc(
    train_data, test_data, config,
    feature_type="lag",
    return_model=True,
    verbose=True,
)

Citation

If you find HYPERDOA useful in your research, please cite:

@misc{hyperdoa_icassp,
      title={HYPERDOA: Robust and Efficient DoA Estimation using Hyperdimensional Computing},
      author={Rajat Bhattacharjya and Woohyeok Park and Arnab Sarkar and Hyunwoo Oh and Mohsen Imani and Nikil Dutt},
      year={2025},
      eprint={2510.10718},
      archivePrefix={arXiv},
      primaryClass={eess.SP},
      url={https://arxiv.org/abs/2510.10718},
}

Acknowledgements

• Data generation utilities based on SubspaceNet
• HDC encoding using torchhd

License

HYPERDOA is licensed under MIT License. For questions or issues, feel free to open an issue. Thank you for using HYPERDOA!