How to convert RNN/LSTM models using the mtk converter?

NeuroPilot NeuroPilot - Analytical AI
,
1

Hi everyone,
I’d like to ask if the mtk converter in the NeuroPilot SDK supports converting RNN/LSTM models to TFLite.

I’ve successfully converted an RNN/LSTM model (pb file) trained using TensorFlow into TFLite using tf.lite.TFLiteConverter(the official TensorFlow Python API).
However, when I try to convert the same models using the mtk converter, I encounter the following error:
RuntimeError: Cannot import the following TensorFlow operators: RandomStandardNormal, TensorListFromTensor, TensorListGetItem, TensorListReserve, TensorListSetItem, TensorListStack.

How can this be resolved? Also, when using the MTK converter to convert RNN/LSTM models, is there a restriction on the framework used to train the model?
For example, is it limited to PyTorch or TensorFlow?

Thanks!

2

Hi Samuel,

Thank you for the inquiry. Could you provide the official TFLite information and the MTK tool version you used in the experiment?

Different MTK Converter versions were designed based on a specific TFLite version. For example, NP8 Converter are designed with TFLite v2.17’s Op scope.

Looking forward to more information that we could assist you in solving the problem :slight_smile:

MTK,
Joying