lz1004
lz1004
@ArgoHA I also noticed the aspect ratio issue. Let me know if I can help.
My guess is simplicity.
I tried training with letterbox on my custom data starting with the pretrained coco weights for a d-fine nano, and the results were not good.
Yes, in my experience with the repo, I would recommend that. All projects on custom data did well with this combo and no parameter adjusting (except epochs).
@dkurt From debugging I noticed that the problem seems to be related to the output layer names. From openvino > 2024.0, the output layer names end with "_1", "_2" etc.
> > @dkurt From debugging I noticed that the problem seems to be related to the output layer names. From openvino > 2024.0, the output layer names end with "_1",...
I found out a few things: 1. I debugged into that part in net_openvino.cpp and the output layer name is correct, without "_". I also noticed, that this also happens...
Checked it with my older custom *.pb models and various newer IR models (convnext, yolo series) for opencv 4.10+331412d and openvino 2024.4.0. Works just fine, thank you!
Have you tried starting your custom dataset annotations with category id 0 instead of 1?
As I mentioned in my question, I do not have such a checkpoint, but would like to try it, if the authors provide one.