Replace dynamic-range int8 with weight-only int8
#3
by mlboydaisuke - opened
Same change as efficientnet_b1 (merged earlier today): the shipped dynamic-range int8 degrades accuracy on this architecture, while weight-only int8 is near-lossless at the same size reduction.
Measured against the torch float reference on real photos plus fixed random tensors at native resolution: dynamic min logit correlation 0.867 (top-1 2/6) vs weight-only 0.999 (top-1 6/6). The replacement file is quantized directly from this repository's published float32 file (litert quantize, recipe weight_only_wi8_afp32), so the weights are identical.
The model card is updated in the same commit: the metric entries for the removed dynamic file are dropped and a short quantized-variant note with the measured numbers is added.
Yoursmiling changed pull request status to merged