Hi,
as described in microTVM mlperf tiny input data, I am experimenting with microTVM and mlperf tiny. I managed to get correct classification with KWS using the quantized int8 model. And also managed to get correct results for visual wake words in both the int8 version and the float32 version.
However when I use microTVM with KWS and the float32 model, I get moslty “unknown” classifications. The way I have built it is identical to the successful VWW runs. I tested the model and the input files directly with tensorflow from python and there it worked fine.
So I assume that the model and the inputs are correct, but I seem to be doing something wrong, when generating the C++ code using microTVM. Can you think of some common pitfalls, I might be a victim of?
Thank you and best regards, Benedikt