Hi everyone!
Today, I tried parsing a custom pre-trained Yolov8n model we had laying around (I wanted to skip the whole retraining step and “just” compile that model instead).Using the toolchain by Ultralytics, I converted the model from .pt to .onnx easily.
But when parsing, the parser had an error, as some operations are apparently unsupported, and suggested truncating the model to an earlier layer and using that as output. In an older community post, I found a linked YAML-file from the Model Zoo, where the output nodes used in the Model Zoo .hef model were listed.
My question is: how exactly did you come to the conclusion to use the 4 output layers used in your yolov8n conversion? I assume that the layers after are just used for some first steps of postprocessing; however I have no idea how you identified that. Are there some common indicators to identify the nodes? Or do you have to just be very familiar with the model you want to compile?
On a related note: how did you then go ahead and manually implement the postprocessing? Does that require some digging around in the source code of the library?
For the YOLO - family of models, this isn’t that important, as I can just retrain them using the retraining docker images. But in case I want to compile other models whose architectures aren’t part of the Model Zoo, I would have absolutely no idea where to begin in case the parser finds an unsupported operation.
Thanks for the help and insight!