Update app.py

app.py

@@ -128,46 +128,45 @@ def predict_digit(input_image):
     if input_image is None:
         return "Please draw a number!"
     
-    # 1. Ambil data gambar dari composite layer Gradio
-    if isinstance(input_image, dict) and "composite" in input_image:
-        img_array = input_image["composite"]
-    else:
-        img_array = input_image
-
-    # 2. Ambil channel alpha (transparansi) atau ubah ke grayscale biasa
-    if img_array.shape[-1] == 4:
-        # Jika kuas Gradio menghasilkan RGBA, ambil channel Alpha-nya (mask coretan)
-        grayscale = img_array[:, :, 3]
-    else:
-        # Jika RGB, ubah ke hitam putih manual
-        grayscale = np.dot(img_array[...,:3], [0.2989, 0.5870, 0.1140])
-    
-    # 3. OTOMATIS INVERSI: Paksa coretan jadi PUTIH di atas background HITAM
-    # Kita cek apakah background-nya dominan terang. Jika ya, balik warnanya.
-    if np.mean(grayscale) > 127:
-        grayscale = 255 - grayscale
-
-    # 4. THRESHOLDING (Binarization): Bikin garis tegas dan bersihkan noise semenit
-    # Ini membantu menyamakan ketebalan agar kontrasnya mirip coretan tinta MNIST
-    grayscale = (grayscale > 30).astype(np.uint8) * 255
-
-    # Konversi ke PIL Image untuk proses resize
-    img = Image.fromarray(grayscale).convert('L')
-    
-    # 5. RESIZE & CENTERING: Kecilkan ke 20x20 dulu, lalu beri padding ke 28x28
-    # Ini trik rahasia standar MNIST agar angka kamu posisinya pas di tengah kotak!
-    img = img.resize((20, 20), Image.Resampling.LANCZOS)
-    centered_img = Image.new("L", (28, 28), 0)
-    centered_img.paste(img, (4, 4)) # Taruh di tengah (bukan mepet pojok)
-    
-    # 6. NORMALISASI: Masukkan ke pipeline normalisasi latihan kamu (mean & std)
-    tensor_img = transform_fn(centered_img).unsqueeze(0).to(device)
-    
-    with torch.no_grad():
-        outputs = model(tensor_img)
-        probabilities = F.softmax(outputs, dim=1)[0]
+    try:
+        # 1. Ambil layer coretan pertama langsung seperti kode referensi
+        if isinstance(input_image, dict) and "layers" in input_image and len(input_image["layers"]) > 0:
+            img_array = input_image["layers"][0]
+        elif isinstance(input_image, dict) and "composite" in input_image:
+            img_array = input_image["composite"]
+        else:
+            img_array = input_image
+
+        # 2. Pastikan bentuknya 2D Grayscale
+        if len(img_array.shape) == 3:
+            if img_array.shape[-1] == 4: # Jika RGBA, ambil alpha channel (coretannya)
+                grayscale = img_array[:, :, 3]
+            else: # Jika RGB, konversi ke grayscale
+                grayscale = np.dot(img_array[...,:3], [0.2989, 0.5870, 0.1140])
+        else:
+            grayscale = img_array
+
+        # 3. Cek jika kanvas kosong
+        if np.max(grayscale) == 0:
+            return {str(i): 0.1 for i in range(10)}
+
+        # 4. Konversi ke PIL dan Resize ke 28x28 (Gaya kode referensimu)
+        img = Image.fromarray(grayscale.astype(np.uint8)).convert('L')
+        img = img.resize((28, 28), resample=Image.Resampling.BILINEAR)
+        
+        # 5. Jalankan normalisasi PyTorch sesuai training LookThem V8 kamu
+        tensor_img = transform_fn(img).unsqueeze(0).to(device)
         
-    return {str(i): float(probabilities[i]) for i in range(10)}
+        # 6. Prediksi dengan Otak Sniper LookThem
+        with torch.no_grad():
+            outputs = model(tensor_img)
+            probabilities = F.softmax(outputs, dim=1)[0]
+            
+        return {str(i): float(probabilities[i]) for i in range(10)}
+
+    except Exception as e:
+        return {"Error": 0.0}
+ 
 
 # --- GRADIO INTERFACE CONSTRUCTION ---
 with gr.Blocks() as demo: