How to implement the GRU layer with S16 operators?

[Raise-me-up]

Is there any S16 example like Q15 version? (https://github.com/STMicroelectronics/STM32CubeF7/blob/master/Drivers/CMSIS/NN/Examples/ARM/arm_nn_examples/gru/arm_nnexamples_gru.cpp)

It is so difficult for me to implement the GRU layer with S16 operators, and I have spent lots of time.

Any instruction will be grateful!

[ArmRyan]

Hi @Raise-me-up,

We don't own or maintain that example. So I'm not sure if there are any others around gru. Our primary use case is as backend accelerator library for frameworks such as tflite micro which does not seem to support GRU directly.

In saying that, the branch that you linked uses a very very old version of cmsis-nn
`/* ----------------------------------------------------------------------

• Project: CMSIS NN Library
• Title: arm_nnfunctions.h
• Description: Public header file for CMSIS NN Library
• $Date: 13. July 2018
• $Revision: V.1.0.0
• Target Processor: Cortex-M cores
• -------------------------------------------------------------------- *`

The current version on our main is:

• $Date: 04 November 2024
• $Revision: V.18.0.0

If you update to our main branch we have s16 versions for some of the functions mentioned in that example file. I have not checked every function but we do not have support for "arm_fully_connected_q15" anymore for example but we do have an s16 version of fully connected.

I would first see if its possible to update the library version.

Good luck!

[Raise-me-up]

@ArmRyan Thanks for your quick reply! I am doing some works about refactoring the GRU layer by S16 functions, and I have no idea about how to do that. I have made some attempts, but the quantization error is a little big. Could you please give me some guides about the quantization formula and code refactoring?