confidance

Hi there ​@Mahameru, welcome on board!

Ah, I bet this is something to do with quantisation.

When a model runs on the CPU/GPU it's using 16 or 32-bit floating point. But on Metis, the model is quantised to an 8-bit integer to get the optimal performance and efficiency benefits of the hardware. So that quantisation step will probably shift confidence scores slightly compared to the original FP32 model, if that makes sense?

But it’s usually pretty minimal even then. Here’s a bit more info about it, that should help. Is it vastly different of you, or  just a bit of a variance?

Hi, thanks for your explanation!

Yes, I understand quantization can affect confidence scores. 
However, in my case the difference is quite large (e.g. 0.8–0.9 on local vs around 0.3 on Axelera).

Could this be related to scaling or post-processing differences?
Also, should I adjust the confidence threshold when running on Metis?

Thanks!

Ooo, yeah, a drop from 0.8–0.9 down to ~0.3 is bigger than what you'd expect from quantisation. That kind of gap likely points to a preprocessing or pipeline mismatch?

To help narrow this down, could you share:

• What’s your setup? Host system, OS, SDK version etc.
• Are you using the pre-built model zoo YAMLs from the Voyager SDK (e.g. yolov8n-coco-onnx), or did you export your own YOLO model and write a custom YAML?
• The commands you're running on both sides (your local laptop and on Axelera) . That way we can try and spot any differences in the pipeline.

👍