diff --git a/steganogan/encoders.py b/steganogan/encoders.py
index d1a93c8..8742234 100644
--- a/steganogan/encoders.py
+++ b/steganogan/encoders.py
@@ -2,6 +2,20 @@
 
 import torch
 from torch import nn
+import torch
+from torch import nn
+import torch.onnx
+from torchvision.ops.deform_conv import DeformConv2d
+
+input = torch.rand(4, 3, 10, 10)
+kh, kw = 3, 3
+weight = torch.rand(5, 3, kh, kw)
+# offset and mask should have the same spatial size as the output
+# of the convolution. In this case, for an input of 10, stride of 1
+# and kernel size of 3, without padding, the output size is 8
+offset = torch.rand(4, 2 * kh * kw, 8, 8)
+mask = torch.rand(4, kh * kw, 8, 8)
+out = deform_conv2d(input, offset, weight, mask=mask)
 
 
 class BasicEncoder(nn.Module):
@@ -16,7 +30,7 @@ class BasicEncoder(nn.Module):
     add_image = False
 
     def _conv2d(self, in_channels, out_channels):
-        return nn.Conv2d(
+        return DeformConv2d(
             in_channels=in_channels,
             out_channels=out_channels,
             kernel_size=3,