Azure-docs: Export failure

[nexxuno]

Hello, I have a Object Detection trained model (general (compact)) and I want to export it to coreML and Tensorflow Lite but the export fails with only a "Failed" message. Where can I find more logs or some diagnostic data? Where can I report this?

---

Detalji dokumenta

⚠ Ne uređujte ovu sekciju. Ona je neophodna za povezivanje problema slijedom docs.microsoft.com ➟ GitHub.

• ID: 916de6f5-ef41-c82b-5918-0c6149326aab
• Version Independent ID: c7a4bb67-957d-f0a3-4a9f-e71035923f58
• Content: Export your model to mobile - Custom Vision Service - Azure Cognitive Services
• Content Source: articles/cognitive-services/Custom-Vision-Service/export-your-model.md
• Service: cognitive-services
• Sub-service: custom-vision
• GitHub Login: @anrothMSFT
• Microsoft Alias: anroth

[SamaraSoucy-MSFT]

Thanks for the question! We will review and get back to you shortly.

[g-rad]

I'm experiencing the same issue

[nexxuno]

Hello, now I'm able to export, so I guess something changed. But with an object detection model with two labels I get an output shape with these dimensions [1, 13, 13, 35]. Is there any documentation about the output shape obtained using custom vision and how to decipher it? Or the export is just bugged and I need to wait for a fix? When I look into the object returned by the prediction it doesn't make mush sense.

[nexxuno]

https://stackoverflow.com/questions/58308141/how-to-use-customvision-ai-to-create-object-detection-model-for-tensorflow-lite

[GiftA-MSFT]

@nexxuno thanks for your feedback. I am following up internally to gather information. Will followup once I have more details. Thanks.

[shonohs]

Hello, you can look into a sample python script in the exported zip package, which describes how to parse the [1, 13, 13, 35] output vector. Another option is to use "General (compact) [S1]" domain, which has better output format. The detail of the format is in its exported zip file.

[nexxuno]

Ok, I'm using Xamarin, I've tried the compact S1 but as far as I understand it needs a Flex delegate for the tensorflow.lite interpreter to be parsed. Unfortunately it's not available in the Microsoft tensorflow library for Xamarin as far as I know, I've opened a new request here:

https://github.com/xamarin/XamarinComponents/issues/867

I'll try and see if I can port the python post processing included with the non-S1 model export to C#, then again all this trial and error could be avoided with some better documentation (but maybe it's just me).

Thanks for the inputs!

[shonohs]

Thanks! We were not aware of the issue with Xamarin. I'll see if we can replace the Flex op with a new standard op which was introduced in TFLite 2.1.

[nexxuno]

Thank you very much, please let me know as soon as I can test this with a new export.

[GiftA-MSFT]

Thanks for your input @shonohs. Will now proceed to close this thread.

[nexxuno]

@shonohs since this issue has been closed is there any other issue opened on some other repo where I can follow the progress?