tinygrad implementation of: https://github.com/roboflow/rf-detr (inference)

Setup:

pip install -r requirements.txt

Inference on single image:

python rfdetr.py {link to an image} {model variant}

e.g python rfdetr.py https://i.ytimg.com/vi/w0V4SK21UIE/hq720.jpg l

Live WebGPU inference

python compile_to_webgpu.py
python -m http.server 8080

open localhost:8080

Testing performance

PYTHONPATH=. python test/test_jit.py

for faster inference use tinygrad's BEAM search:

PYTHONPATH=. BEAM=2 python test/test_jit.py

this will result in a longer initial run time as the searches are performed and cached. For visibility on the process use:

PYTHONPATH=. BEAM=2 DEBUG=2 python test/test_jit.py

Speed (M3 Macbook Air)

with BEAM=2:

Model	Resolution	FPS
nano	384	38.01
small	512	21.95
medium	576	15.54
large	704	10.68

(RX7600 via tinygrad egpu)

Model	Resolution	FPS
nano	384	67.10
small	512	23.64
medium	576	23.07
large	704	12.62

without BEAM=2:

Model	Resolution	FPS
nano	384	10.14
small	512	5.56
medium	576	4.15
large	704	2.37