GRT: Towards Foundational Models for Single-Chip Radar

Implementation of the Generalizable Radar Transformer (GRT) and the experiments shown in Towards Foundational Models for Single-Chip Radar.

Important

This repository is research code, and may contain bugs, outdated links, dependency incompatibilities, and other issues. Getting this code to run in your environment may require a substantial amount of effort, including manually linking to deprecated versions of other dependencies.

Future research should instead build on the Neural Radar Development Kit and RadarML ecosystem, which are being actively developed, maintained, and supported, and includes an official reference implementation of the GRT paper.

Setup

1. Install dependencies:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
   pip install -r requirements.txt
   pip install "roverd[video,ouster]@git+ssh://git@github.com/RadarML/red-rover.git#subdirectory=format"

2. Get data:

   See red-rover for full instructions.

   For each target dataset:

   export SRC=path/to/dataset      # e.g. `radarhd/data` on a network server
   export DST=path/to/destination  # e.g. `data/data` on a local drive
   
   roverp export -p $SRC -o $DST --metadata
   roverp align -p $DST --mode left

Usage

1. Create model in models. See existing examples in models/.

2. Create config. See examples in config.

3. Train:

   python train.py -c your/config/file.yaml -n model/name

4. Observe with tensorboard:

   tensorboard --logdir=path/to/results --host=0.0.0.0