• DepthAI-v3

  • Conversion of YOLO 11 custom trained model using HUBAI for RVC2

I 'm training a model for the OAK-D Lite camera which is RVC2. I have taken some "person" samples from the COCO 2017 dataset and added two more classes for a test using YOLO v11:

model = YOLO("yolo11n.pt")
results = model.train(data=f"{working_directory}/dataset.yaml", epochs=200, imgsz=416, classes=[0,1,4])

I want to convert this using the hub but I'm having some issues: using the old https://tools.luxonis.com/ website everything is good, but I simply cannot make the conversion using the new tools work.

The error is RuntimeError: BlobReader error: File does not seem to be a supported neural network blob when using the superblob. Do I need to switch depthai to v3 for the superblob to work? I've tried directly converting the resulting .pt file on the hub and downloading the blob and downloading the onnx and converting it with modelconverter hub convert rvc2 --path "${working_directory}/best.onnx.tar.xz"

    BobbyFisher
    Yep. Superblobs only work on the v3. For V2, continue using tools. Afaik, there is a way to convert superblob to regular blob by selecting shave count cc @KlemenSkrlj

    Thanks,
    Jaka

      BobbyFisher

      Hello, you can disable the superblob format by adding rvc2.superblob False to the CLI command like this:
      modelconverter hub convert rvc2 --path "${working_directory}/best.onnx.tar.xz" rvc2.superblob False

      You can also control the number of shaves for the regular blob by adding rvc2.number_of_shaves {n}(the default value is 8).