fix bugs in semantic segmentation example (#1824)

* Update unet.py

* Update semantic_segmentation.py

pl_examples/domain_templates/semantic_segmentation.py

@@ -165,7 +165,7 @@ class SegModel(pl.LightningModule):
         return {'val_loss': loss_val}
 
     def validation_epoch_end(self, outputs):
-        loss_val = sum(output['val_loss'] for output in outputs) / len(outputs)
+        loss_val = torch.stack([x['val_loss'] for x in outputs]).mean()
         log_dict = {'val_loss': loss_val}
         return {'log': log_dict, 'val_loss': log_dict['val_loss'], 'progress_bar': log_dict}
 

pl_examples/models/unet.py

@@ -33,7 +33,7 @@ class UNet(nn.Module):
             feats *= 2
 
         for _ in range(num_layers - 1):
-            layers.append(Up(feats, feats // 2), bilinear)
+            layers.append(Up(feats, feats // 2, bilinear))
             feats //= 2
 
         layers.append(nn.Conv2d(feats, num_classes, kernel_size=1))