Custom trained yolov5-model "Message has too much metadata"

Johnny-McGee

Hello.

I have trained a yolov5n with ultralytics and converted the best.pt to an .onnx and then to a .blob-file using the BlobConverter-website.

Now if I try to use the model in my preview I constantly get the following error:

"[XLinkOut(3)] [error] Message has too much metadata (511377B) to serialize. Maximum is 51200B. Dropping message"

I don't believe this has anything to do with the model itself since I got the same error with a yolov5s and yolov5m.

A previous question got the recommendation to convert the .pt-file to a .blob-file with this website: https://tools.luxonis.com/[Automatic Yolo export for OAKs](https://tools.luxonis.com/)

But every time I try to convert a .pt-file I get another error that just says "Error while loading model" even though I put in the correct image size.

As further information, I am using an OAK-1 Lite. The .pt-file has a size of 3.72 MB.

Any help would be appreciated since I have been trying to get a custom-yolo model working with an OAK for an entire day at this point.

jakaskerl

Johnny-McGee
Converting through tools is correct; as to why it doesn't work some further debugging is needed.

cc @JanCuhel

Johnny-McGee

Over the weekend I have trained a yolov8-model instead and this worked with the website. However I am still getting the same metadata-error

JanCuhel

Johnny-McGee

Could you please share with me your YOLOv5 model, so that I can investigate it further?

Best,
Jan

Johnny-McGee

JanCuhel I can't share it because of security-reasons of my company. But it was just a regular yolov5 from ultralytics trained after this website: https://docs.ultralytics.com/yolov5/tutorials/train_custom_data/

JanCuhel

Got it, thanks. I'll look into it and let you know.

Johnny-McGee

I have trained the model with yolov8 now and realized that I didn't convert the model from the normal ultralytics-model to a model compatible with an OAK-device as described in this post: luxonis/depthai-ml-trainingblob/master/colab-notebooks/YoloV8_training.ipynb

However every time I try to run the main.py with my .json-config it tells me that the OAK needs to be connected via USB 3.0. Unfortunately the only USB 3.0 I have is on my laptop and if I connect the OAK there it gives me an error that the device crashed with an io error.

I have also tried forcing USB 2.0 connection but then the program can't find the OAK on the Socket anymore.

jakaskerl

Johnny-McGee
Can you give me more info?
What script are you using? What does the io error say?

Thanks,
Jaka

Johnny-McGee

I am using the linked github repo. The last step utilizes the "main.py" from the following repo: luxonis/depthai-experimentstree/master/gen2-yolo/device-decoding

The error is this:

[18443010010B680F00] [1.3] [1725959346.261] [host] [warning] Device crashed, but no crash dump could be extracted.

RuntimeError: Device already closed or disconnected: io error

jakaskerl

Johnny-McGee
Which SDK version are you using? Can I see the .json file?

Thanks,
Jaka

Johnny-McGee

Okay the problem is most likely that my laptop can't output enough power to supply the OAK-device.

Also the last step of the github-repo for custom-training a yolov8 model is written very confusingly. It does not convert the .blob-file to a usable file but you have to use the code in the main.py or main_api.py to run your model.

It would be very very helpful to have a proper tutorial about how to run a custom-trained-yolov8 on your oak and which inputs and outputs to generate and how to link everything. Because after trying to recreate the steps in the main_api.py I get a stream but no detections. Clearer documentation would be appreciated.