diff --git a/models/onnx_export.py b/models/onnx_export.py
index be6faf7..a7132de 100644
--- a/models/onnx_export.py
+++ b/models/onnx_export.py
@@ -13,25 +13,27 @@ from models.common import *
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--weights', type=str, default='./weights/yolov5s.pt', help='weights path')
-    parser.add_argument('--img-size', type=int, default=640, help='inference size (pixels)')
+    parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path')
+    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')
     parser.add_argument('--batch-size', type=int, default=1, help='batch size')
     opt = parser.parse_args()
     print(opt)
 
     # Parameters
     f = opt.weights.replace('.pt', '.onnx')  # onnx filename
-    img = torch.zeros((opt.batch_size, 3, opt.img_size, opt.img_size))  # image size, (1, 3, 320, 192) iDetection
+    img = torch.zeros((opt.batch_size, 3, *opt.img_size))  # image size, (1, 3, 320, 192) iDetection
 
     # Load pytorch model
     google_utils.attempt_download(opt.weights)
     model = torch.load(opt.weights)['model']
     model.eval()
-    # model.fuse()
+    model.fuse()
 
     # Export to onnx
     model.model[-1].export = True  # set Detect() layer export=True
-    torch.onnx.export(model, img, f, verbose=False, opset_version=11)
+    _ = model(img)  # dry run
+    torch.onnx.export(model, img, f, verbose=False, opset_version=11, input_names=['images'],
+                      output_names=['output'])  # output_names=['classes', 'boxes']
 
     # Check onnx model
     model = onnx.load(f)  # load onnx model
diff --git a/models/yolo.py b/models/yolo.py
index 586b054..d0c4ab1 100644
--- a/models/yolo.py
+++ b/models/yolo.py
@@ -20,7 +20,7 @@ class Detect(nn.Module):
         self.export = False  # onnx export
 
     def forward(self, x):
-        x = x.copy()  # for profiling
+        # x = x.copy()  # for profiling
         z = []  # inference output
         self.training |= self.export
         for i in range(self.nl):
@@ -38,6 +38,34 @@ class Detect(nn.Module):
 
         return x if self.training else (torch.cat(z, 1), x)
 
+    def forward_(self, x):
+        if hasattr(self, 'nx'):
+            z = []  # inference output
+            for (y, gi, agi, si, nyi, nxi) in zip(x, self.grid, self.ag, self.stride, self.ny, self.nx):
+                m = self.na * nxi * nyi
+                y = y.view(1, self.na, self.no, nyi, nxi).permute(0, 1, 3, 4, 2).contiguous().view(m, self.no).sigmoid()
+
+                xy = (y[..., 0:2] * 2. - 0.5 + gi) * si  # xy
+                wh = (y[..., 2:4] * 2) ** 2 * agi  # wh
+                p_cls = y[:, 4:5] if self.nc == 1 else y[:, 5:self.no] * y[:, 4:5]  # conf
+                z.append([p_cls, xy, wh])
+
+            z = [torch.cat(x, 0) for x in zip(*z)]
+            return z[0], torch.cat(z[1:3], 1)  # scores, boxes: 3780x80, 3780x4
+
+        else:  # dry run
+            self.nx = [0] * self.nl
+            self.ny = [0] * self.nl
+            self.ag = [0] * self.nl
+            for i in range(self.nl):
+                bs, _, ny, nx = x[i].shape
+                m = self.na * nx * ny
+                self.grid[i] = self._make_grid(nx, ny).repeat(1, self.na, 1, 1, 1).view(m, 2) / torch.tensor([[nx, ny]])
+                self.ag[i] = self.anchor_grid[i].repeat(1, 1, nx, ny, 1).view(m, 2)
+                self.nx[i] = nx
+                self.ny[i] = ny
+            return None
+
     @staticmethod
     def _make_grid(nx=20, ny=20):
         yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])