fix deep copying of TFM trainer parameters (#1459)

* fix deep copying of TFM trainer parameters

* fix failing tests

Author: dennisbader

darts/models/forecasting/torch_forecasting_model.py

@@ -475,11 +475,16 @@ def _init_trainer(
         trainer_params: dict, max_epochs: Optional[int] = None
     ) -> pl.Trainer:
         """Initializes a PyTorch-Lightning trainer for training or prediction from `trainer_params`."""
-        trainer_params_copy = {param: val for param, val in trainer_params.items()}
+        trainer_params_copy = {key: val for key, val in trainer_params.items()}
         if max_epochs is not None:
             trainer_params_copy["max_epochs"] = max_epochs
 
-        return pl.Trainer(**trainer_params_copy)
+        # prevent lightning from adding callbacks to the callbacks list in `self.trainer_params`
+        callbacks = trainer_params_copy.pop("callbacks", None)
+        return pl.Trainer(
+            callbacks=[cb for cb in callbacks] if callbacks is not None else callbacks,
+            **trainer_params_copy,
+        )
 
     @abstractmethod
     def _create_model(self, train_sample: Tuple[Tensor]) -> torch.nn.Module:

darts/tests/models/forecasting/test_ptl_trainer.py

@@ -151,7 +151,9 @@ def on_train_epoch_end(self, *args, **kwargs):
 
             # check if callbacks were added
             self.assertEqual(len(model.trainer_params["callbacks"]), 2)
-            model.fit(self.series, epochs=2)
+            model.fit(self.series, epochs=2, verbose=True)
+            # check that lightning did not mutate callbacks (verbosity adds a progress bar callback)
+            self.assertEqual(len(model.trainer_params["callbacks"]), 2)
 
             self.assertEqual(my_counter_0.counter, model.epochs_trained)
             self.assertEqual(my_counter_2.counter, model.epochs_trained + 2)