eiQ Toolkit TFLite Converter failing without error message

[georgehill]

I am trying to convert an object detection model from tensorflow to tflite with int8 quantization but find that the converter fails without any error message. The model tool also can only load the model in h5, not in keras v3 or as a saved_model. 

I've seen other posts where people have had success with the tool recently so I don't know what the issue could be, the model is a yolov8 from ultralytics if that is relevant.

Any insight would be appreciated.

[Bio_TICFSL]

Hello,

If your model uses TensorFlow operators not natively supported by TFLite built-ins, enable tf.lite.OpsSet.SELECT_TF_OPS and ensure the Flex delegate is available in your TFLite runtime on the i.MX.

Performance and Accuracy:

Full integer quantization can significantly reduce model size and improve inference speed on hardware optimized for 8-bit integers (like NPUs in i.MX). However, carefully evaluate the impact on model accuracy and adjust the representative dataset or quantization strategy if needed.

You can check the appnote AN13699 for converting to int8.

https://www.nxp.com/docs/en/application-note/AN13699.pdf

Regards