Add application files

app.py

@@ -19,11 +19,15 @@ from scipy.special import kl_div as scipy_kl_div
 from skimage.metrics import structural_similarity as ssim
 import warnings
 
-TASK = "nodules"
+# --- Configuration ---
+# Set the default task.
+TASK = "facades"
 PATH = os.path.join("datasets", TASK, "real")
 images = []
 perceptual_model = None
 
+# --- Model Loading ---
+# Attempt to load the VGG19 model for the perceptual loss metric.
 try:
     vgg = VGG19(weights='imagenet', include_top=False, input_shape=(224, 224, 3))
     vgg.trainable = False
@@ -31,7 +35,9 @@ try:
 except Exception as e:
     perceptual_model = None
 
+# --- Utility Functions ---
 def safe_normalize_heatmap(heatmap):
+    """Safely normalizes a heatmap to a 0-255 range for visualization."""
     if heatmap is None or heatmap.size == 0:
         return np.zeros((64, 64), dtype=np.uint8)
     heatmap = heatmap.astype(np.float32)
@@ -48,7 +54,10 @@ def safe_normalize_heatmap(heatmap):
         normalized_heatmap = np.clip(normalized_heatmap, 0, 255)
     return np.uint8(normalized_heatmap)
 
+# --- Image Comparison Metrics ---
+
 def KL_divergence(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False):
+    """Calculates Kullback-Leibler Divergence between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape:
         return None
     try:
@@ -92,6 +101,7 @@ def KL_divergence(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=
     return img_dict
 
 def L1_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False):
+    """Calculates L1 (Mean Absolute Error) loss between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape: return None
     try:
         img_real_rgb = cv2.cvtColor(img_real, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
@@ -121,6 +131,7 @@ def L1_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False)
     return img_dict
 
 def MSE_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False):
+    """Calculates MSE (Mean Squared Error) loss between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape: return None
     try:
         img_real_rgb = cv2.cvtColor(img_real, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
@@ -150,6 +161,7 @@ def MSE_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False
     return img_dict
 
 def SSIM_loss(img_real, img_fake, block_size=7, sum_channels=False):
+    """Calculates SSIM (Structural Similarity Index) loss between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape: return None
     try:
         img_real_rgb = cv2.cvtColor(img_real, cv2.COLOR_BGR2RGB)
@@ -177,6 +189,7 @@ def SSIM_loss(img_real, img_fake, block_size=7, sum_channels=False):
     return img_dict
 
 def cosine_similarity_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False):
+    """Calculates Cosine Similarity loss between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape: return None
     try:
         img_real_rgb = cv2.cvtColor(img_real, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
@@ -212,6 +225,7 @@ def cosine_similarity_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_
     return img_dict
 
 def TV_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False):
+    """Calculates Total Variation (TV) loss between two images."""
     if img_real is None or img_fake is None or img_real.shape != img_fake.shape: return None
     try:
         img_real_rgb = cv2.cvtColor(img_real, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
@@ -241,6 +255,7 @@ def TV_loss(img_real, img_fake, epsilon=1e-10, block_size=4, sum_channels=False)
     return img_dict
 
 def perceptual_loss(img_real, img_fake, model, block_size=4):
+    """Calculates Perceptual loss using a pre-trained VGG19 model."""
     if img_real is None or img_fake is None or model is None or img_real.shape != img_fake.shape:
         return None
     original_height, original_width, _ = img_real.shape
@@ -264,7 +279,10 @@ def perceptual_loss(img_real, img_fake, model, block_size=4):
     heatmap_original_size = cv2.resize(heatmap_features, (original_width, original_height), interpolation=cv2.INTER_LINEAR)
     return {"SUM": {"SUM": total_loss, "HEATMAP": heatmap_original_size.astype(np.float32)}}
 
+# --- Gradio Core Functions ---
+
 def gather_images(task):
+    """Loads a random pair of real and fake images from the selected dataset."""
     global TASK, PATH, images
     new_path = os.path.join("datasets", task, "real")
     if TASK != task or not images:
@@ -311,6 +329,7 @@ def gather_images(task):
         return placeholder, placeholder, error_msg
 
 def run_comparison(real, fake, measurement, block_size_val):
+    """Runs the selected comparison metric and generates a heatmap."""
     placeholder_heatmap = np.zeros((64, 64, 3), dtype=np.uint8)
     if real is None or fake is None or not isinstance(real, np.ndarray) or not isinstance(fake, np.ndarray):
         return placeholder_heatmap, "Error: Input image(s) missing or invalid. Please load or upload a pair of images."
@@ -353,8 +372,25 @@ def run_comparison(real, fake, measurement, block_size_val):
     return heatmap_rgb, status_msg
 
 def clear_uploads(msg):
+    """Clears the image displays and updates the status message."""
     return None, None, msg
 
+def load_and_compare_initial(task):
+    """Gathers initial images and runs a comparison on them at startup."""
+    # Step 1: Get the initial images
+    real_img, fake_img, gather_status = gather_images(task)
+
+    # Step 2: Run the default comparison
+    # We use the default values from the UI definition
+    default_measurement = "Cosine Similarity"
+    default_block_size = 8
+    heatmap, compare_status = run_comparison(real_img, fake_img, default_measurement, default_block_size)
+
+    # Step 3: Combine status messages and return all initial values
+    final_status = f"{gather_status}\n{compare_status}"
+    return real_img, fake_img, heatmap, final_status
+
+# --- Gradio UI Definition ---
 theme = gr.themes.Soft(primary_hue="blue", secondary_hue="orange")
 with gr.Blocks(theme=theme, css=".gradio-container { max-width: 1400px !important; margin: auto; }") as demo:
     gr.Markdown("# GAN vs Ground Truth Image Comparison")
@@ -368,7 +404,7 @@ with gr.Blocks(theme=theme, css=".gradio-container { max-width: 1400px !importan
             with gr.Tabs():
                 with gr.TabItem("Load from Dataset"):
                     task_dropdown = gr.Dropdown(
-                        ["nodules", "facades"], value=TASK,
+                        ["facades"], value=TASK,
                         info="Select the dataset task.",
                         label="Dataset Task"
                     )
@@ -387,7 +423,7 @@ with gr.Blocks(theme=theme, css=".gradio-container { max-width: 1400px !importan
             with gr.Row():
                 measurement_dropdown = gr.Dropdown(
                     ["Kullback-Leibler Divergence", "L1-Loss", "MSE", "SSIM", "Cosine Similarity", "TV", "Perceptual"],
-                    value="Kullback-Leibler Divergence",
+                    value="Cosine Similarity",
                     info="Select the comparison metric.",
                     label="Comparison Metric",
                     scale=2
@@ -404,6 +440,15 @@ with gr.Blocks(theme=theme, css=".gradio-container { max-width: 1400px !importan
             gr.Markdown("### 3. See Result")
             heatmap_display = gr.Image(type="numpy", label="Comparison Heatmap (Difference)", height=350, interactive=False)
 
+    # --- Event Listeners ---
+    
+    # Load initial sample and run comparison when the app starts
+    demo.load(
+        fn=load_and_compare_initial,
+        inputs=[task_dropdown],
+        outputs=[real_img_display, fake_img_display, heatmap_display, status_message]
+    )
+
     sample_button.click(
         fn=gather_images,
         inputs=[task_dropdown],
@@ -434,7 +479,7 @@ with gr.Blocks(theme=theme, css=".gradio-container { max-width: 1400px !importan
         outputs=[real_img_display, fake_img_display, status_message]
     )
 
-
+# --- Application Entry Point ---
 if __name__ == "__main__":
     print("-------------------------------------------------------------")
     print("Verifying VGG19 model status...")