X-CUBE-AI, tflite LSTM model: Unsupported layer types

[CFrancesco]

Hi,

I’m currently trying to upload the following LSTM model into a STM32L476RG via X-CUBE-AI

Uploading the Keras .h5 model works well, however when I try to upload the TensorFlow Lite converted version of the same model, I get the following prompt:

Any ideas about how to solve this problem?

 

Some information about my setup:

STM32CubeMx version: 6.8.0

X-CUBE-AI version: 8.1.0

TensorFlow version: 2.10.0

 

Thank you very much.

[saralf]

Hi,

I have the same issue with a similar LSTM model. Did you  solve it?

[fauvarque.daniel]

The conversion to TFLite unfolds the LSTM layer in many different ways depending on the paramters of the LSTM and if it is stateless or stateful.

UNIDIRECTIONAL_SEQUENCE_LSTM is supported by X-Cube-AI, you can follow the instructions here https://www.tensorflow.org/lite/models/convert/rnn to properly convert your Keras LSTM model into a tflite model understood by X-CUBE-AI

Regards

[nuvyas]

Hello,

I tried the above mentioned solution but still got the same error....

I am also implementing a CRNN model which consists of CNN layers and LSTM layers inside it, but LSTM part is not being implemented. 

I am doing audio processing over the board and using LSTM can prove to be very efficient for them. 

I have X-Cube-AI 9.0.0, is there any other version I should use? Is there any other way I can work around with LSTM deployment on the board. 

My model description: 

Here, 128 corresponds to the height of the input spectrogram (frequency bins), 600 corresponds to the width (time steps or frames), and 1 corresponds to the number of channels (mono audio).

 

Any help is appreciated really,

Thanking you

N.U.Vyas