Update app.py

app.py

@@ -9,8 +9,8 @@ from PIL import Image
 app = FastAPI(title="DyslexiaLens Prediction API")
 
 # ─── RE-REGISTER CUSTOM ARCHITECTURE COMPONENTS ───
-# Keras requires these custom classes explicitly defined to unpack the .keras file
-@keras.saving.register_keras_serializable()
+# CHANGED: Swapped .saving to .utils to match the tf.keras wrapper ecosystem
+@keras.utils.register_keras_serializable(package="Custom")
 class AdaptiveContrastNorm(layers.Layer):
     def __init__(self, epsilon: float = 1e-6, **kwargs):
         super().__init__(**kwargs)
@@ -33,7 +33,8 @@ class AdaptiveContrastNorm(layers.Layer):
         config.update({'epsilon': self.epsilon})
         return config
 
-@keras.saving.register_keras_serializable()
+# CHANGED: Swapped .saving to .utils here as well
+@keras.utils.register_keras_serializable(package="Custom")
 class MaskedHuberLoss(keras.losses.Loss):
     def __init__(self, delta: float = 0.5, **kwargs):
         super().__init__(**kwargs)
@@ -61,7 +62,6 @@ model = None
 def load_model():
     global model
     try:
-        # Pass custom mapping into custom_objects for accurate loading
         model = keras.models.load_model(
             MODEL_PATH,
             custom_objects={
@@ -91,31 +91,24 @@ async def predict(
         raise HTTPException(status_code=503, detail="Model is loading or uninitialized.")
 
     try:
-        # 1. Image Preprocessing (Matches tf.data Pipeline)
         contents = await file.read()
-        image = Image.open(io.BytesIO(contents)).convert('L') # Force standard Grayscale
-        image = image.resize((128, 128), Image.BILINEAR)      # Downsample from 224x224 source
-        img_array = np.array(image, dtype=np.float32) / 255.0  # Normalize pixels [0, 1]
-        img_tensor = np.expand_dims(img_array, axis=(0, -1))   # Reshape to matching batch: (1, 128, 128, 1)
+        image = Image.open(io.BytesIO(contents)).convert('L')
+        image = image.resize((128, 128), Image.BILINEAR)
+        img_array = np.array(image, dtype=np.float32) / 255.0
+        img_tensor = np.expand_dims(img_array, axis=(0, -1))
 
-        # 2. Clinical Vector Compilation
         feature_vector = np.array([
             stroke_density, center_of_mass_x, center_of_mass_y, 
             bounding_box_ratio, stroke_transitions, horizontal_symmetry
-        ], dtype=np.float32).reshape(1, 6) # Shape: (1, 6)
+        ], dtype=np.float32).reshape(1, 6)
 
-        # 3. Model Pipeline Dual-Head Inference
-        # Keys must explicitly map to the functional API layer names defined during your build phase
         predictions = model.predict({
             'image_input': img_tensor, 
             'feature_input': feature_vector
         })
 
-        # Parse head outputs
         clf_probability = float(predictions[0][0][0])
         severity_score = float(predictions[1][0][0])
-        
-        # Proven threshold division border evaluated via your sweep analysis
         is_dyslexia = clf_probability >= 0.40
 
         return {