diff --git a/models.py b/models.py
index aee33b65..804d3a4a 100755
--- a/models.py
+++ b/models.py
@@ -4,6 +4,7 @@ from utils.parse_config import *
 from utils.utils import *
 
 ONNX_EXPORT = False
+arc = 'normal'  # (normal, uCE, uBCE, uBCEs) detection architectures
 
 
 def create_modules(module_defs, img_size):
@@ -77,8 +78,12 @@ def create_modules(module_defs, img_size):
 
             # Initialize preceding Conv2d() bias (https://arxiv.org/pdf/1708.02002.pdf section 3.3)
             bias = module_list[-1][0].bias.view(len(mask), -1)  # 255 to 3x85
-            bias[:, 4] -= 5.0  # obj
-            bias[:, 5:] -= 4.0  # cls
+            if arc == 'normal':
+                bias[:, 4] -= 5.0  # obj
+                bias[:, 5:] -= 4.0  # cls
+            elif arc == 'uCE':
+                bias[:, 4] += 3.0  # obj
+                bias[:, 5:] -= 4.0  # cls
             module_list[-1][0].bias = torch.nn.Parameter(bias.view(-1))
 
             # for l in model.yolo_layers:  # print pretrained biases
@@ -168,7 +173,6 @@ class YOLOLayer(nn.Module):
             # io[..., 2:4] = ((torch.sigmoid(io[..., 2:4]) * 2) ** 3) * self.anchor_wh  # wh power method
             io[..., :4] *= self.stride
 
-            arc = 'normal'  # (normal, uCE, uBCE, uBCEs) detection architectures
             if arc == 'normal':
                 torch.sigmoid_(io[..., 4:])
             elif arc == 'uCE':  # unified CE (1 background + 80 classes)