STM32 X-Cube-AI won't let me analyze the full integer quantized LSTM model

[Flare]

I have a LSTM model which was quantized using TensorFlowLite full integer quantization. After that, I wanted to try deploy the model on the SensorTile Box. I tried analyzing the model but it throws the following error:

[AI:network] NOT IMPLEMENTED: UNIDIRECTIONAL_SEQUENCE_LSTM with output type (SIGNED, 8 bit, C Size: 8 bits Scales: [0.007843137718737125] Zeros: [-1] Quantizer: UNIFORM) not supported

 

I would be really grateful for ideas or insights on how to solve this!

[jean-michel.d]

Hi Fare,

With the 8.1, the fully quantized version (int8) of the UNIDIRECTIONAL_SEQUENCE_LSTM operator is not supported. It will be supported in the new X-CUBE-AI release with the  cell_state in 16b as converted by TFLite converter.

br

jm

 

[Flare]

Hello Jean,

Thank you for your reply. I found out that in X-CUBE-AI on CubeMX, it provides two different runtime, TFLite Micro and STM32Cube.AI. When I try to analyze it with the TFLite Micro runtime, I am able to analyze the model. Does that mean that this runtime supports the UNIDIRECTIONAL_SEQUENCE_LSTM operator?

What's the difference between these two runtimes?