Hello,
Excuse me, has anyone had experience accelerating the Whisper-tiny model for speech-to-text tasks using the NPU on the STM32MP257-EV1 board? From my understanding, the process involves converting the PyTorch or TensorFlow Whisper model to TensorFlow Lite format and then using the ST Edge AI tool to convert it to .nb format for acceleration.
I followed the approach provided by @nyadla-sys on GitHub to obtain the whisper-tiny.en.tflite model (https://github.com/nyadla-sys/whisper.tflite/blob/main/models/generate_tflite_from_whisper.ipynb). However, when I run the following command:
./stedgeai generate --target stm32mp2 --model whisper-tiny.en.tflite
I encounter the following error:
ST Edge AI Core v2.0.0-20049
INTERNAL ERROR: 'NoneType' object has no attribute 'rank'
I am unsure why this error occurs. Could it be due to an internal input being None? I am also uncertain whether this approach is correct, or if I should separate and accelerate the encoder and decoder modules individually.
If anyone has experience accelerating Transformer-based models on this platform or insights into resolving this issue, I would greatly appreciate your guidance.
Thank you!
