Replace dynamic-range int8 with weight-only int8

README.md

@@ -28,12 +28,6 @@ model-index:
           - name: Top 5 Accuracy (Full Precision)
             type: accuracy
             value: 0.9530
-          - name: Top 1 Accuracy (Dynamic Quantized wi8 afp32)
-            type: accuracy
-            value: 0.8012
-          - name: Top 5 Accuracy (Dynamic Quantized wi8 afp32)
-            type: accuracy
-            value: 0.9501
 ---
 
 # EfficientNet B2
@@ -49,6 +43,16 @@ acc@1 (on ImageNet-1K): 80.608%
 acc@5 (on ImageNet-1K): 95.31%    
 num_params: 9,109,994
 
+### Quantized variant
+
+`efficientnet_b2_weight_only_wi8_afp32.tflite` is a weight-only int8
+quantization of the same weights (about 3.6x 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 spot check against the float model the weight-only
+file keeps the top-1 predictions on real photos with a minimum logit
+correlation of 0.999.
+
 ## 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_b2_dynamic_wi8_afp32.tflite → efficientnet_b2_weight_only_wi8_afp32.tflite

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f45c7230f75d1c2ce26af33e5bff04c3c58669aee53fb16a5dcd4d4d7cd1ca26
-size 10188488
+oid sha256:0e1c14bb8351af4d2421bdf764ce39dc8f983582ebd33ec70438a44b09f2cc68
+size 10249936