tinygrad implementation of: https://github.com/WongKinYiu/yolov9

Setup:

pip install -r requirements.txt

Inference on single image:

python yolov9.py {link or path to an image} {model variant}

Live WebGPU inference

python compile_to_webgpu.py
python -m http.server 8080

open localhost:8080

Testing performance

python test_jit.py

for faster inference use tinygrad's BEAM search:

BEAM=2 python 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:

BEAM=2 DEBUG=2 python test_jit.py

Speed (M3 Macbook Air)

with BEAM=2:

Model	Resolution	FPS
t	320	198.56
t	640	78.08
t	960	39.41
t	1280	25.27
t	1536	16.48
s	320	97.11
s	640	33.31
s	960	17.85
s	1280	12.24
s	1536	7.94
m	320	46.09
m	640	15.81
m	960	7.73
m	1280	5.01
m	1536	3.37
c	320	35.72
c	640	13.55
c	960	5.82
c	1280	4.22
c	1536	2.58
e	320	20.36
e	640	7.49
e	960	3.25
e	1280	2.25
e	1536	1.43

without BEAM=2:

Model	Resolution	FPS
t	320	139.94
t	640	71.06
t	960	22.38
t	1280	17.31
t	1536	9.72
s	320	63.21
s	640	26.24
s	960	10.09
s	1280	6.09
s	1536	3.31
m	320	27.28
m	640	9.53
m	960	3.61
m	1280	1.49
m	1536	0.89
c	320	17.37
c	640	6.25
c	960	2.50
c	1280	1.27
c	1536	0.74
e	320	8.36
e	640	3.36
e	960	1.38
e	1280	0.74
e	1536	0.42