Better error message if no loss was returned from model.training_step() (#294)

pytorch_lightning/trainer/trainer.py

@@ -1156,6 +1156,10 @@ class Trainer(TrainerIO):
         except Exception:
             if type(output) is torch.Tensor:
                 loss = output
+            else:
+                raise RuntimeError(
+                    'No `loss` value in the dictionary returned from `model.training_step()`.'
+                )
 
         # when using dp need to reduce the loss
         if self.use_dp: