RP3: Cause of getting different results?

[Favi0]

Hello i was wondering why im getting different results while doing inference with a model trained in alexeyAB network compared to my RP3 32Bit raspbian?

is it because the 32 bit/64 bit weight or
how the convolution is performed in each framework? <- NNPACK related

im getting for instance 90% for an class in AB
but 60% in DarknetNnpack