add lcd images

Readme.md

@@ -90,7 +90,12 @@ but in most times with delta=0.1
 
 ## Versions
 
-### 1.6 (2002-12-27)
+### 1.6.1 (2023-02-26)
+* new images (lcd)
+* fixed quantization bug - use all images for dataset
+* dig-class-161_s2
+
+### 1.6 (2022-12-27)
 * new images
 * dig-class-160_s2
 

b2n/data/ziffer.py

@@ -26,11 +26,11 @@ def ziffer_data(input_dir='images', input_shape=(32,20,3)):
         base = os.path.basename(aktfile)
 
         # get label from filename (1.2_ new or 1_ old),
-        if (base[1]=="."):
+        if (base[1]=="." and not base[2] == 'j'):
             target = base[0:3]
         else:
             target = base[0:1]
-         
+        #print(base)
         category = float(target)
 
         test_image = Image.open(aktfile).resize((input_shape[1],input_shape[0]))

b2n/models/prune_quantize.py

@@ -78,7 +78,7 @@ def quantization_sparse(model, x_train):
 
     def representative_dataset():
         for n in range(x_train[0].size):
-            data = np.expand_dims(x_train[5], axis=0)
+            data = np.expand_dims(x_train[n], axis=0)
         yield [data.astype(np.float32)]
 
     converter2 = tf.lite.TFLiteConverter.from_keras_model(model)
@@ -103,13 +103,12 @@ def quantization_default(model, x_train):
 
     def representative_dataset():
         for n in range(x_train[0].size):
-            data = np.expand_dims(x_train[5], axis=0)
+            data = np.expand_dims(x_train[n], axis=0)
         yield [data.astype(np.float32)]
 
     converter2 = tf.lite.TFLiteConverter.from_keras_model(model)
     converter2.representative_dataset = representative_dataset
     converter2.optimizations = [tf.lite.Optimize.DEFAULT]
-    converter2.representative_dataset = representative_dataset
     # Ensure that if any ops can't be quantized, the converter throws an error
     converter2.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
     # Set the input and output tensors to uint8 (APIs added in r2.3)