Caution added

model.py

@@ -54,6 +54,8 @@ class Generator(nn.Module):
 
     def forward(self, x, c):
         # Replicate spatially and concatenate domain information.
+        # Note that this type of label conditioning does not work at all if we use reflection padding in Conv2d.
+        # This is because instance normalization ignores the shifting (or bias) effect.
         c = c.view(c.size(0), c.size(1), 1, 1)
         c = c.repeat(1, 1, x.size(2), x.size(3))
         x = torch.cat([x, c], dim=1)
@@ -83,4 +85,4 @@ class Discriminator(nn.Module):
         h = self.main(x)
         out_src = self.conv1(h)
         out_cls = self.conv2(h)
-        return out_src, out_cls.view(out_cls.size(0), out_cls.size(1))
+        return out_src, out_cls.view(out_cls.size(0), out_cls.size(1))