Replace dynamic-range int8 with weight-only int8 (fixes collapsed accuracy)
#3
by mlboydaisuke - opened
As discussed in the collaboration space: the bundled efficientnet_b1_dynamic_wi8_afp32.tflite collapses against the torch fp32 reference — Top-1 0/4 with minimum logit correlation 0.32 on fixed random probes (seeds 0-3), 0/10 Top-1 with real photos in the fuller run — which fits EfficientNet-B1's known quantization sensitivity (SE blocks + SiLU activations). The fp32 conversion itself is exact (correlation 1.000000).
This PR replaces it with a weight-only int8 quantization of the same model:
| variant | Top-1 | Top-5 | min corr | size |
|---|---|---|---|---|
| dynamic_wi8_afp32 (removed) | 0/4 | 0/20 | 0.320 | 8.5 MB |
| weight_only_wi8_afp32 (added) | 4/4 | 18/20 | 0.998 | 8.9 MB |
Same ~3.5x size reduction vs fp32. The fp32 model and the Tensor G5 plugin variant are untouched. Happy to share the reproduction harness (torch fp32 reference, fixed-seed probes + photo set) if useful.
Yoursmiling changed pull request status to merged