building wrapper sepearately

app.py

@@ -2,6 +2,7 @@ import gradio as gr
 import numpy as np
 import tensorflow as tf
 from PIL import Image
+import efficientnet.tfkeras as efn # <-- Add this import
 
 # ==========================================
 # 1. MRI Model Setup (Your Existing Model)
@@ -29,11 +30,10 @@ def predict_mri(image):
 
 
 # ==========================================
-# 2. X-Ray Model Setup (Reconstructing from Weights)
+# 2. X-Ray Model Setup (Using original EfficientNet library)
 # ==========================================
 print("Building X-Ray model architecture...")
 
-# The 14 classes from the NIH Chest X-Ray dataset
 xray_class_names = [
     'Cardiomegaly', 'Emphysema', 'Effusion', 'Hernia', 'Infiltration', 
     'Mass', 'Nodule', 'Atelectasis', 'Pneumothorax', 'Pleural_Thickening', 
@@ -41,11 +41,11 @@ xray_class_names = [
 ]
 
 def build_xray_model():
-    # Based on the Kaggle dataset description, the weights were trained 
-    # on an EfficientNetB1 with a 128x128 input size.
-    base_model = tf.keras.applications.EfficientNetB1(
+    # Use the 'efn' library instead of tf.keras.applications
+    # This guarantees the architecture has exactly 437 weights as expected.
+    base_model = efn.EfficientNetB1(
         input_shape=(128, 128, 3), 
-        weights=None, # We are loading custom weights next
+        weights=None, 
         include_top=False
     )
     
@@ -53,10 +53,10 @@ def build_xray_model():
         base_model,
         tf.keras.layers.GlobalAveragePooling2D(),
         tf.keras.layers.Dense(1024, activation='relu'),
-        tf.keras.layers.Dense(len(xray_class_names), activation='sigmoid') # Sigmoid for multi-label
+        tf.keras.layers.Dense(len(xray_class_names), activation='sigmoid')
     ])
     
-    # Load the downloaded weights
+    # Load weights should now perfectly match 437 to 437
     model.load_weights("xray.h5")
     return model
 
@@ -68,19 +68,19 @@ def predict_xray(image):
         return None
     
     # Preprocess the X-Ray input
-    img = Image.fromarray(image).convert('RGB') # EfficientNet expects 3 channels
-    img = img.resize((128, 128)) # The Kaggle dataset used 128x128
+    img = Image.fromarray(image).convert('RGB')
+    img = img.resize((128, 128))
     
     img_array = np.array(img)
-    # Keras EfficientNet applications have built-in rescaling, 
-    # so we skip the / 255.0 step here.
+    img_array = np.expand_dims(img_array, axis=0)
     
-    img_array = np.expand_dims(img_array, axis=0) # Add batch dimension
+    # Use the library's built-in preprocessing to match training conditions
+    img_array = efn.preprocess_input(img_array)
     
     # Predict
     predictions = xray_model.predict(img_array)[0]
     
-    # Map probabilities to class names
+    # Map probabilities
     confidences = {xray_class_names[i]: float(predictions[i]) for i in range(len(xray_class_names))}
     return confidences
 
@@ -111,7 +111,6 @@ with gr.Blocks(title="Medical Scan Classification") as interface:
                     xray_input = gr.Image(label="Upload Chest X-Ray")
                     xray_button = gr.Button("Classify X-Ray", variant="primary")
                 with gr.Column():
-                    # Displaying top 5 conditions since there are 14 possible labels
                     xray_output = gr.Label(num_top_classes=5, label="Top 5 Predicted Conditions")
                     
             xray_button.click(fn=predict_xray, inputs=xray_input, outputs=xray_output)