Update backend.py

backend.py

@@ -3,13 +3,44 @@ import numpy as np
 import joblib
 from PIL import Image
 
+# --- THE FIX: Custom Layer Wrappers to ignore version bugs ---
+class SafeDense(tf.keras.layers.Dense):
+    def __init__(self, *args, **kwargs):
+        kwargs.pop('quantization_config', None)
+        super().__init__(*args, **kwargs)
+
+class SafeDropout(tf.keras.layers.Dropout):
+    def __init__(self, *args, **kwargs):
+        kwargs.pop('quantization_config', None)
+        super().__init__(*args, **kwargs)
+
+class SafeConv2D(tf.keras.layers.Conv2D):
+    def __init__(self, *args, **kwargs):
+        kwargs.pop('quantization_config', None)
+        super().__init__(*args, **kwargs)
+
+class SafeMaxPooling2D(tf.keras.layers.MaxPooling2D):
+    def __init__(self, *args, **kwargs):
+        kwargs.pop('quantization_config', None)
+        super().__init__(*args, **kwargs)
+
+# Group them up to feed into the loader
+safe_objects = {
+    'Dense': SafeDense,
+    'Dropout': SafeDropout,
+    'Conv2D': SafeConv2D,
+    'MaxPooling2D': SafeMaxPooling2D
+}
+# -------------------------------------------------------------
+
 print("Loading AI Models... Please wait.")
-# Load all 5 components of your pipeline
-base_cnn = tf.keras.models.load_model('base_cnn.h5')
-encoder = tf.keras.models.load_model('encoder_model.h5') # Make sure you exported this!
+# We use compile=False because we only need to predict, not train. 
+# This also saves memory and prevents further errors!
+base_cnn = tf.keras.models.load_model('base_cnn.h5', custom_objects=safe_objects, compile=False)
+encoder = tf.keras.models.load_model('encoder_model.h5', custom_objects=safe_objects, compile=False)
 pca = joblib.load('pca_model.pkl')
 selector = joblib.load('selector_model.pkl')
-classifier = tf.keras.models.load_model('parkinsons_model.h5')
+classifier = tf.keras.models.load_model('parkinsons_model.h5', custom_objects=safe_objects, compile=False)
 print("All models loaded successfully!")
 
 def process_and_predict(input_img):