Do the neural networks need to be pre-processed or can they be run directly from Tensorflow?

[Amelie ACKERMANN]

You say the new STM32MP25 can run Neural Networks. Do the neural networks need to be pre-processed (e.g. mode conversion from floating point to integer), or can they be run directly from Tensorflow?

[Roman LUDIN]

You can use directly the TF output without quantization, in which case it will only run on the MP2's CPU. To run on the NPU, you need to use our offline compiler, either standalone version, or available via the Edge AI Developer Cloud. It will let you quantize in INT6 or INT16. More info on stm32ai.st.com