litert-community

/

MobileNet-v3-small

like 0

Follow

LiteRT Community (FKA TFLite) 8.3k

This swaps mobilenet_v3_small_dynamic_wi8_afp32.tflite for a weight-only int8 file built from this repo's own float model, because the dynamic file does not reproduce the float model's predictions (0/6 top-1 on real photos; the weight-only file restores 6/6 at the same size). Same shape as the EfficientNet fix you merged earlier. Nothing here is urgent, and if you would rather keep both files or use a different recipe, that is entirely your call — happy to redo it either way.

Details, if useful:

• Checked against this repo's own mobilenet_v3_small.tflite on six natural images (two photos, plus mirrored and center-cropped variants).
• dynamic: 0/6 top-1, logit correlation 0.21–0.46. The misses are unrelated classes rather than near ties — a cat reads as banded gecko, Christmas stocking, safety pin; a baseball scene as wall clock, jigsaw puzzle.
• The float file itself is fine: correlation 1.0000 against torchvision's mobilenet_v3_small, so this is the quantization step and not the checkpoint.
• Likely cause: MobileNetV3's squeeze-excite gates and hard-swish activations have a wide activation range that dynamic-range quantization does not measure. Same shape as the EfficientNet family (SE + SiLU).
• The replacement is quantized directly from this repo's published mobilenet_v3_small.tflite (sha256 c7a8324a…) with the weight_only_wi8_afp32 recipe: 6/6 top-1, logit correlation 0.9918–0.9938, 2.75 MB against the dynamic file's 2.74 MB (3.7x smaller than float32).
• README updated to match: the file table row, plus a short "Quantized variant" note in the same wording as the EfficientNet cards.

If it is useful I can run the same check across the other repos that still ship a dynamic_wi8_afp32 file and send whatever it turns up — just say the word.