KSNN cant convert weights

[Agent_kapo]

Good evening everyone

My new problem is that NN shows bad results at inferencing:

I’m getting million of predictions of my first class on image with score 0.5

My opinion that weights are empty (if I’m incorrect please correct me)

After training yolov8 I got best.pt, I checked - it works fine.
Then I changed strings in ultralytics head.py (from site_packages)

and

then I used export:

from ultralytics import YOLO

train_folder = 'train'

model = YOLO(f"/home/titan/Desktop/Project_X/dataset_marking_with_yolov8/runs/detect/{train_folder}/weights/best.pt")
results = model.export(format="onnx")

In netron it looks like this:

Seems strange, because I thought it sgould bee 1x87x8400 (I have 23 classes)
I used aml_npu_sdk and this is what I got:

It has no errors, only warnings:

After it, I tried this weights on your yolov8n-stream.py, you already know the results

Want to mention that my dataset is 3-chanel bw images, but it’s additional info

Waiting for your respond

Also!
torch - 1.9.1
onnx - 1.15
ultralytics - 8.0.228
Vim3 Pro

[Louis-Cheng-Liu]

Hello @Agent_kapo ,

There is nothing wrong. Model has three output, 1×87×80×80, 1×87×40×40, 1×87×20×20. And 8400=80×80+40×40+20×20.

Have you modify LISTSIZE in our YOLOV8n demo?

[Agent_kapo]

Actually I found the reason, but I still don’t know what went wrong.
My colleague removed Lhasa’s in the different room. So when I was running my boy scripts, it was giving me matrixes with 7.5 values only and always predicting the first class.
After I reboot the khadas, mistake have disappeared