Wrong inference results after converting plant disease recognition .tflite model to .nb

[WangWei1]

Dear ST Community,

I’m currently working with an ALIENTEK STM32MP257 development board and ran into an issue when deploying a model converted via the ST Edge AI online tool.

The original model is a plant disease recognition model from the ST model zoo (in .tflite format). I used the ST Edge AI webpage to convert it to an .nb model for deployment on my STM32MP257 board. However, when I test the converted .nb model, the inference output is unexpected: it only predicts two classes — “Background_without_leaves” and “Corn__Northern_Leaf_Blight” — both with relatively high confidence. This happens regardless of the input image.

In contrast, when I run inference directly using the original .tflite file on the same test images, the results are correct and show proper class diversity.

I double-checked my test code and also consulted AI assistance, which suggested that the converted .nb model might be problematic.

I’ve attached a compressed file containing my inference code and the model files for reference. Could someone kindly help verify the converted .nb model, or possibly provide a correctly converted .nb file for this plant disease recognition model? Any guidance would be greatly appreciated.

Thank you very much for your time!

Best regards,
[WangWei]

[Julian E.]

Hi ​@WangWei1,

 

The website changed and some strange things occured.

We maybe have already answered but here is an answer again.

 

With your model it did not work

We have regenerated the .nb (mobilenet_regenerated.nb) model from the original .tflite file. Could you please test this newly generated model on your STM32MP257 board and let me know if the inference results are now correct?

If the issue persists, please share the test outcome so we can investigate further.

 

Have a good day,

Julian