Init tensor directly on device (#6068)

Slightly more efficient than .to(device)

models/yolo.py

@@ -72,9 +72,9 @@ class Detect(nn.Module):
     def _make_grid(self, nx=20, ny=20, i=0):
         d = self.anchors[i].device
         if check_version(torch.__version__, '1.10.0'):  # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
-            yv, xv = torch.meshgrid([torch.arange(ny).to(d), torch.arange(nx).to(d)], indexing='ij')
+            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)], indexing='ij')
         else:
-            yv, xv = torch.meshgrid([torch.arange(ny).to(d), torch.arange(nx).to(d)])
+            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)])
         grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2)).float()
         anchor_grid = (self.anchors[i].clone() * self.stride[i]) \
             .view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2)).float()