Export error using the website

Support
, ,
1

Hi,

Using the website(https://platform.ultralytics.com) I just created my first project, upload the dataset, classified, trainned, and when completed, I try to download the .tflite file by using the correspondent “export” option.

ERROR :cross_mark: /var/tmp/tmplyl7ysxm/best.pt references types outside the supported Ultralytics checkpoint format. Use an official Ultralytics model, i.e. ‘yolo predict model=yolo26n.pt’

I tried with different options, but always the same result.

The integrated AI does not provide to me any solution.

Thanks in advance.

2

Hello.

Could you tell the task and base model you used for training?

Could you also provide the Platform project and model links?

We currently have LiteRT export on Platform, not TFLite. TFLite was deprecated. Are you using LiteRT export?

3

Hi,

I tried LiteRT export, with all the combinations that are offer in the options menu. I also tried to export ONNX same error

Thanks

4

Hi Ernest,

Thank you for reporting this and for sharing the project links. We traced the issue to restricted checkpoint loading on Python 3.13 for classification models. The fix is included in Ultralytics 8.4.99 and has now been deployed to the Platform export workers.

We verified your exact trained checkpoint on the live EU worker: it loads correctly as a classification model and exports to ONNX successfully. Please retry the LiteRT or ONNX export from the same model now; you do not need to retrain it.

If any export still fails, please reply with the format and options you selected and we will investigate further.

5

Now it works!!! thanks so much

6

Great, glad the export is working now! Thanks for confirming, and thanks to the Ultralytics team and YOLO community for helping identify and resolve the issue.