INTERNAL ERROR: 'FLOAT16' --> Is it that STM doesnot support tf.float16 formats ?

[rickforescue]

I tried to compress a keras model. It is a versy simple Neural network model with just fully connected (dense) layers. I converted it to tflite and compressed the weight to tf.float16 which was originally tf.float32. When I upload the model, it gave me error

Neural Network Tools for STM32AI v1.6.0 (STM.ai v7.1.0-RC3) 

INTERNAL ERROR: 'FLOAT16'

Is it that STM doesnot support float16 formats ?

More Info: Its a tflite model and I used STM32Cube.AI.runtime

[TDK]

Half-precision floats aren't supported by FPUs on Cortex M4 or M7.

[waclawek.jan]

As usually. there's no point in discussing this without knowing STM32 model (core type) and compiler.

Actually, half-precision floats came up recently as a question in a local forum (in Czech language), and there appears to be *some* support in Cortex-M, although compilers probably won't support it seamlessly.

JW

[TDK]

Weird of them to support that instruction and yet not support any add/subtract/mult/divide instructions on FLOAT16.

https://developer.arm.com/documentation/dui0646/a/The-Cortex-M7-Instruction-Set/Multiply-and-divide-instructions