Replace dynamic-range int8 with weight-only int8 (fixes collapsed accuracy)

README.md

@@ -28,12 +28,6 @@ model-index:
           - name: Top 5 Accuracy (Full Precision)
             type: accuracy
             value: 0.9419
-          - name: Top 1 Accuracy (Dynamic Quantized wi8 afp32)
-            type: accuracy
-            value: 0.7805
-          - name: Top 5 Accuracy (Dynamic Quantized wi8 afp32)
-            type: accuracy
-            value: 0.9392
 ---
 # EfficientNet B1
 
@@ -46,6 +40,15 @@ acc@1 (on ImageNet-1K): 79.838%
 acc@5 (on ImageNet-1K): 94.934%    
 num_params: 7,794,184
 
+### Quantized variant
+
+`efficientnet_b1_weight_only_wi8_afp32.tflite` is a weight-only int8
+quantization of the same weights (about 3.5x smaller than float32). Weight-only
+quantization is used instead of dynamic-range quantization because
+EfficientNet's SE and SiLU layers are sensitive to activation quantization; in
+a 10-image spot check the weight-only model matches the float model's top-1
+prediction on 9 of 10 images (logit correlation 0.996).
+
 ## Intended uses & limitations
 
 The model files were converted from pretrained weights from PyTorch Vision. The models may have their own licenses or terms and conditions derived from PyTorch Vision and the dataset used for training. It is your responsibility to determine whether you have permission to use the models for your use case.

efficientnet_b1_dynamic_wi8_afp32.tflite → efficientnet_b1_weight_only_wi8_afp32.tflite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e2b1af45691f3b18d119d0cf9ba2d278cd04a8fba6fe86908ae4145024fce06a
-size 8540672
+oid sha256:5a4660a2ba9bae994267a7413fe737b57b738328b1e0b1fbab399d181e32327d
+size 8876144