Hi,

On the OAK-FFC-6P with a single RGB camera connected to socket SA, I am able to run the example scripts (tmp_s3 folder).

However, I am having trouble implementing my own network (using my own blob file).

Is the process to create a blob file for the rvc3 different than that listed on this page or the online converter?

If so, what is the proper procedure? Is there a tutorial or comprehensive documentation?

Any help is much appreciated. Thanks in advance!

  • jakaskerl replied to this.
  • Hi @Unknown,

    apologies for the delay in my response. I managed to find the issue, fixed it and deploy the fix. So you should be able now to convert the model (I did it and it worked for me).

    Best,
    Jan

    I am able to run the yolo and mobilenet examples in tmp_s3 folder. However, this is my code:

     Pipeline = dai.Pipeline()
    
     cam = pipeline.create(dai.node.ColorCamera)
    
     cam.setBoardSocket(dai.CameraBoardSocket.CAM_B)
    
     cam.setPreviewSize(512,512)
    
     cam.setRawNumFramesPool(3)
    
     cam.setFps(1)
    
     nn = pipeline.create(dai.node.NeuralNetwork)
    
     nnPath = “MYPATH.blob”
    
     nn.setBlobPath(nnPath)
    
     cam.preview.link(nn.input)
    
     nnXout = pipeline.create(dai.node.XLinkOut)
    
     nnXout.setStreamName(“nn”)
    
     nn.out.link(nnXout.input)
    
     with dai.Device(pipeline) as device:
    
          qNn = device.getOutputQueue("nn")
    
          nnData = qNn.get()

    I get the following error at the last line:

    " RuntimeError: Communication exception - possible device error/misconfiguration. Original message 'Couldn't read data from stream: ‘nn’ (X_LINK_ERROR)' "

    Additionally, the firmware crashes.

    Other notes:

    • I am using the rvc3_develop branch

    • I have updated the OAK-FFC-6P OS to 1.12 and the firmware to rvc3-1.2.0

      Hi ah1
      Do other models work as expected if you switch them inside your code?

      Thanks,
      Jaka

      Yes, the same code with the yolov6n416x416openvino2022.1_vpux.blob model works ok (only change I made was the path and a 416x416 image size).

        Hi ah1
        The AI part on the RVC3 seems to be a little broken so not all models will work with it. I synced with the team and was told that Luxonis model zoo models should work fine, but we can't say the same for custom models.
        Could you do a graph of your layers (using Netron) so we can possibly swap out the layers that are not working.

        Thanks,
        Jaka

        a year later

        Hey @erik
        I have trained a custom model using YOLOv5n. While converting the model to a blob format using RVC2, I encountered an error. However, when I used RVC3 for the conversion, it successfully generated the blob. Unfortunately, when running this blob on DepthAI, I received the following error:

        "RuntimeError: BlobReader error: File does not seem to be a supported neural network blob."

        It's worth noting that a pre-trained YOLOv5n model was converted using RVC2 without issues and ran successfully as a blob. The issue only arises with the custom-trained model.

        Could you please assist in troubleshooting this problem?

          hey @jakaskerl , thankyou.

          i face error while converting yolovn6 pretrained to blob. the error is "Error while converting to openvino". i am using tools.luxonis.com

            Unknown
            Can you send the model .pt? Via email if privacy is an issue.

            Thanks,
            Jaka

            Hey @jakaskerl here is the link to download the model. it will be great if custom model trained using yolov6n6 can also be converted to blob.

              Hi @jakaskerl its a .pt file only. Please click on the black pop out icon above "file size", a new tab will open then you will be able to download the .pt file.

              Thanks.

                10 days later

                Hi @Unknown,

                apologies for the delay in my response. I managed to find the issue, fixed it and deploy the fix. So you should be able now to convert the model (I did it and it worked for me).

                Best,
                Jan