Hey, I’m wondering about your FPS in benchmarks you give. For example, it is 181 for yolov8n for batch size 1. Please tell me how you have performed so well and how I can achieve it. And is it possible to have the same performance on the model retrained on the custom dataset? Thank you!!!
Yes, of course it’s possible. You should download the compiled network (HEF file) to the target system and use the CLI tool to run it:
hailortcli run <file.hef>
Please note that it’s important to test the network out-of-context. Meaning that any other system bottleneck is removed (e.g. video decoding, post-processing, database access etc.). This would give you the uppper limit where it comes to the NNcore engine. After that, when you plug it into a reallife scenario, you should strive to meet that number, albeit, in reality it is hard to get to it 100%.
@Nadav, thank you for the answer; run it on your standard yolov8n from here and get it. What is wrong?
In general, some of the benchmarks in the general page would be higher than when ran on the pi, since the pi has only one PCIe lane.
Specifically, your numbers seems very low. Have you enabled Gen3 on the pi?
When I run I get almost twice the FPS:
@Nadav It was enabled. Try to make it disabled, and allow it once more time and have the results you have; thank you. Maybe it is required to reconfigure it every time the software for PCIe is updated