RayAR

Implementation of RayAR: Neural Autoregressive for Fast Dominant Radio Propagation Path Estimation.

RayAR reconstructs dominant diffraction paths between a specified transmitter and receiver through geometry-constrained autoregressive vertex selection.

Installation

Requires Python 3.10+ and PyTorch 2.1+.

python -m pip install -e .

Data

Use the propagation data released with D²LoS. Place the prepared assets under:

assets/buildings/    Building geometry
assets/cache/        Propagation cache
assets/los/          Visibility and geometry arrays
models/last.pt       Trained checkpoint

Data paths are configurable in configs/. Data and checkpoints are provided separately.

Usage

# Train
python main.py --config configs/full_100map_reference.json train \
  --output output/full --gpu 0

# Evaluate
python main.py --config configs/maps20.json eval \
  --checkpoint models/last.pt --output output/eval.json --gpu 0

# Infer
python infer.py --input examples/query.json \
  --checkpoint models/last.pt --output output/query.json --gpu 0

examples/query.json specifies the map, one Tx position, and a list of Rx positions. The src/los/ directory contains the optional C++ preprocessing backend.

Code is organized directly under src/model, src/data, src/geom, src/train, src/eval, src/runtime, and src/cli.