Update app.py

app.py

@@ -28,19 +28,14 @@ MODEL = None
 LOAD_ERROR = None  # Store the specific reason for failure
 
 # -------------------- COMPATIBILITY FIX --------------------
-# This function dynamically creates a fixed version of any Keras layer
-# that ignores the Keras 3 specific arguments (like quantization_config)
-# allowing models saved in new versions to load in older environments.
 def fix_layer_config(cls):
     class FixedLayer(cls):
         def __init__(self, *args, **kwargs):
-            # Remove Keras 3 arguments not supported in Keras 2
             kwargs.pop('quantization_config', None)
             kwargs.pop('glitch_filter', None) 
             super().__init__(*args, **kwargs)
     return FixedLayer
 
-# Apply the fix to all layers likely to appear in EfficientNet
 CUSTOM_OBJECTS = {
     'Dense': fix_layer_config(Dense),
     'Dropout': fix_layer_config(Dropout),
@@ -59,61 +54,38 @@ CUSTOM_OBJECTS = {
 # -------------------- LOAD MODEL --------------------
 def init_model():
     global MODEL, LOAD_ERROR
-    # Reset errors
     LOAD_ERROR = None
     
     if os.path.exists(MODEL_FILE):
         print(f"[INIT] Model found: {MODEL_FILE}")
         try:
-            # We pass the custom_objects dictionary to handle the version mismatch
             MODEL = load_model(MODEL_FILE, custom_objects=CUSTOM_OBJECTS)
             print("[INIT] Model loaded successfully.")
         except Exception as e:
             print(f"[ERROR] Failed to load model: {e}")
-            LOAD_ERROR = str(e) # Capture the actual error message
+            LOAD_ERROR = str(e)
             MODEL = None
     else:
         print(f"[ERROR] Model file '{MODEL_FILE}' NOT FOUND on server.")
         MODEL = None
 
-# -------------------- PREPROCESSING --------------------
+# -------------------- PREPROCESSING & TTA --------------------
 def calculate_entropy(img_array):
-    """
-    Calculates Shannon Entropy to measure image texture complexity.
-    Returns a float value (higher = more complex/diseased features).
-    """
     try:
-        # Ensure image is 0-255 uint8 for histogram calculation
         if img_array.dtype != np.uint8:
-            # If normalized 0-1, scale up. If just float 0-255, cast.
-            if np.max(img_array) <= 1.0:
-                calc_img = (img_array * 255).astype(np.uint8)
-            else:
-                calc_img = img_array.astype(np.uint8)
-        else:
-            calc_img = img_array
+            if np.max(img_array) <= 1.0: calc_img = (img_array * 255).astype(np.uint8)
+            else: calc_img = img_array.astype(np.uint8)
+        else: calc_img = img_array
 
-        # Convert to grayscale if it's color (Batch dim, H, W, C) or (H, W, C)
-        if len(calc_img.shape) == 4: # (1, 224, 224, 3)
-            calc_img = calc_img[0]
-        
+        if len(calc_img.shape) == 4: calc_img = calc_img[0]
         gray = cv2.cvtColor(calc_img, cv2.COLOR_RGB2GRAY)
-        
-        # Compute histogram
         hist = cv2.calcHist([gray], [0], None, [256], [0, 256])
-        
-        # Normalize histogram to get probabilities
         hist_norm = hist.ravel() / hist.sum()
-        
-        # Filter zero values to avoid log(0) error
         hist_norm = hist_norm[hist_norm > 0]
-        
-        # Entropy formula: -Sum(p * log2(p))
         entropy_val = -np.sum(hist_norm * np.log2(hist_norm))
         return float(entropy_val)
     except Exception as e:
-        print(f"Entropy Error: {e}")
-        return 4.5 # Fallback average value
+        return 4.5
 
 def process_single_image(image_path):
     try:
@@ -133,7 +105,29 @@ def process_single_image(image_path):
         
         return np.expand_dims(final.astype(np.float32) / 255.0, axis=0)
     except Exception as e:
-        return str(e) # Return error string instead of None
+        return str(e)
+
+def predict_with_tta(model, input_batch):
+    """
+    Test Time Augmentation:
+    Predicts on the original image + flipped versions and averages the results.
+    """
+    img = input_batch[0] # Extract image from batch (224, 224, 3)
+    
+    # Create batch of 3 variants: Original, Horizontal Flip, Vertical Flip
+    # Retina images have no "correct" up/down, so vertical flipping is valid logic.
+    aug_batch = np.array([
+        img,
+        np.fliplr(img),
+        np.flipud(img)
+    ])
+    
+    # Get predictions for all 3 variations
+    preds = model.predict(aug_batch)
+    
+    # Average the probabilities across the 3 views
+    avg_pred = np.mean(preds, axis=0)
+    return avg_pred
 
 # -------------------- ROUTES --------------------
 @app.route("/")
@@ -153,12 +147,11 @@ def analyze():
             if not os.path.exists(MODEL_FILE):
                 return jsonify({
                     "diagnosis": "System Error",
-                    "description": f"Model file '{MODEL_FILE}' not found. Did you upload it?",
+                    "description": f"Model file '{MODEL_FILE}' not found.",
                     "confidence": "0%",
                     "color": "rose", "icon": "alert-octagon"
                 })
             else:
-                # Use the captured specific error message
                 error_msg = LOAD_ERROR if LOAD_ERROR else "Model initialization skipped by server."
                 return jsonify({
                     "diagnosis": "Load Error",
@@ -169,8 +162,6 @@ def analyze():
 
         # 2. Process Image
         input_data = process_single_image(temp_path)
-        
-        # Check if processing failed (returned an error string)
         if isinstance(input_data, str):
             return jsonify({
                 "diagnosis": "OpenCV Error",
@@ -179,8 +170,9 @@ def analyze():
                 "color": "rose", "icon": "alert-triangle"
             })
 
-        # 3. Predict
-        preds = MODEL.predict(input_data)[0]
+        # 3. Predict with TTA (Smart Averaging)
+        preds = predict_with_tta(MODEL, input_data)
+        
         idx = np.argmax(preds)
         label = CLASSES[idx]
         conf = preds[idx] * 100
@@ -215,7 +207,6 @@ def analyze():
         if os.path.exists(temp_path): os.remove(temp_path)
 
 # -------------------- INITIALIZE ON IMPORT --------------------
-# Crucial Fix: Call init_model() globally so Gunicorn runs it!
 init_model()
 
 if __name__ == "__main__":