Update app.py

app.py

@@ -3,7 +3,6 @@ import tensorflow as tf
 import numpy as np
 import json
 from tensorflow.keras.preprocessing.image import img_to_array
-from tensorflow.keras.applications.vgg16 import preprocess_input
 from PIL import Image
 import re
 
@@ -14,7 +13,6 @@ def clean_class_name(raw_name):
     and 'pear_1' to 'Pear'.
     """
     # Remove the trailing underscore and number (e.g., '_1', '_2')
-    # Using regex to remove '_' followed by digits at the end
     cleaned = re.sub(r'_\d+$', '', raw_name)
     # Replace underscores with spaces
     cleaned = cleaned.replace('_', ' ')
@@ -25,7 +23,7 @@ def clean_class_name(raw_name):
 # ---------- Load the Model (.keras format) ----------
 model_path = 'fruits_classifier.keras'
 model = tf.keras.models.load_model(model_path)
-print("Model loaded successfully!")
+print("✅ Model loaded successfully!")
 
 # ---------- Load Class Names ----------
 with open('class_indices.json', 'r') as f:
@@ -33,46 +31,46 @@ with open('class_indices.json', 'r') as f:
 
 # Convert dictionary to a list for easy index-based access
 class_names_list = [class_names_dict[str(i)] for i in range(len(class_names_dict))]
+print(f"✅ Total classes loaded: {len(class_names_list)}")
 
-# ---------- Create a list of cleaned class names for display ----------
-# Format: ["Apple Red", "Apple Braeburn", "Cucumber", ...]
-cleaned_class_names = [clean_class_name(name) for name in class_names_list]
-print(f"Total classes loaded: {len(cleaned_class_names)}")
-
-# ---------- Prediction Function ----------
+# ---------- Prediction Function (Matches Training Preprocessing) ----------
 def predict_image(image):
     """
-    Takes an input image, preprocesses it for VGG16, runs inference,
-    and returns the predicted class name and confidence.
+    Preprocessing exactly matches training:
+    - Resize to (64, 64)
+    - Rescale by 1.0/255.0 (same as ImageDataGenerator rescale)
+    - NO preprocess_input (VGG16 mean subtraction) because training didn't use it
     """
-    # Resize image to match the model's input shape (64x64)
+    # Step 1: Resize image to (64, 64) - matches target_size in training
     img = image.resize((64, 64))
-    # Convert PIL image to numpy array
+    
+    # Step 2: Convert PIL image to numpy array
     img_array = img_to_array(img)
-    # Add batch dimension (1, 64, 64, 3)
+    
+    # Step 3: Rescale pixel values to [0, 1]
+    # This matches: rescale=1.0/255.0 in ImageDataGenerator
+    img_array = img_array / 255.0
+    
+    # Step 4: Add batch dimension (1, 64, 64, 3)
     img_array = np.expand_dims(img_array, axis=0)
-    # Apply VGG16-specific preprocessing (scaling and mean subtraction)
-    img_array = preprocess_input(img_array)
     
-    # Run inference
-    predictions = model.predict(img_array)
+    # Step 5: Run inference (same as model.predict in notebook)
+    predictions = model.predict(img_array, verbose=0)
     predicted_index = np.argmax(predictions, axis=-1)[0]
     confidence = np.max(predictions, axis=-1)[0]
     
-    # Get the raw class name
+    # Step 6: Get the raw class name and clean it for display
     raw_class_name = class_names_list[predicted_index]
-    # Clean the class name for display
     predicted_class = clean_class_name(raw_class_name)
     confidence_percentage = float(confidence) * 100
     
     return predicted_class, f"{confidence_percentage:.2f}%"
 
-# ---------- Create a formatted list of categories for display ----------
-# Format: "Apple, Apple Braeburn, Apple Crimson Snow, ..."
-categories_list = sorted(cleaned_class_names)  # Sort alphabetically
-categories_text = ", ".join(categories_list)   # Join with commas
+# ---------- Create Categories List for Display ----------
+cleaned_class_names = [clean_class_name(name) for name in class_names_list]
+categories_list = sorted(cleaned_class_names)
+categories_text = ", ".join(categories_list)
 
-# Create the description with categories
 description_text = f"""
 ### 🍎 Upload an image of a fruit.
 
@@ -82,10 +80,11 @@ description_text = f"""
 
 ---
 
-*Model: VGG16-based Transfer Learning trained on Fruits-360 dataset.*
+*Model: VGG16-based Transfer Learning*
+*Input size: 64x64 | Preprocessing: Rescale to [0, 1]*
 """
 
-# ---------- Gradio Interface Setup ----------
+# ---------- Gradio Interface ----------
 interface = gr.Interface(
     fn=predict_image,
     inputs=gr.Image(type="pil", label="Upload Fruit Image"),
@@ -97,6 +96,6 @@ interface = gr.Interface(
     description=description_text,
 )
 
-# ---------- Launch the App ----------
+# ---------- Launch ----------
 if __name__ == "__main__":
     interface.launch()