Warn when encoders are ignored by fit/predict_from_dataset

darts/models/forecasting/torch_forecasting_model.py

@@ -1279,7 +1279,19 @@ def fit_from_dataset(
         -------
         self
             Fitted model.
+
+        Notes
+        -----
+        Encoders passed via ``add_encoders`` at model creation are ignored by this method. Encoders are only
+        applied when training with :func:`fit()`. If you need encoded covariates, the training datasets must
+        already contain them.
         """
+        if self.add_encoders:
+            logger.warning(
+                "Encoders (`add_encoders`) are ignored when training with `fit_from_dataset()`. "
+                "Encoders are only applied when calling `fit()`. If you need encoded covariates, "
+                "make sure the training datasets already contain them."
+            )
         self._train(
             **self._setup_for_train(
                 train_dataset=train_dataset,
@@ -2057,7 +2069,19 @@ def predict_from_dataset(
         -------
         Sequence[TimeSeries]
             Returns one or more forecasts for time series.
+
+        Notes
+        -----
+        Encoders passed via ``add_encoders`` at model creation are ignored by this method. Encoders are only
+        applied when predicting with :func:`predict()`. If you need encoded covariates, the inference dataset must
+        already contain them.
         """
+        if self.add_encoders:
+            logger.warning(
+                "Encoders (`add_encoders`) are ignored when predicting with `predict_from_dataset()`. "
+                "Encoders are only applied when calling `predict()`. If you need encoded covariates, "
+                "make sure the inference dataset already contains them."
+            )
         return self._predict(
             **self._setup_for_predict(
                 n=n,

darts/tests/models/forecasting/test_torch_forecasting_model.py

@@ -2527,6 +2527,45 @@ def test_fit_with_stride(self, stride):
             assert len(train_set) == len(val_set) == math.ceil(3 / stride)
             assert train_set.stride == val_set.stride == stride
 
+    @patch("darts.models.forecasting.torch_forecasting_model.logger.warning")
+    def test_encoders_ignored_in_from_dataset_warns(self, mock_warning):
+        """`fit_from_dataset`/`predict_from_dataset` warn that `add_encoders` is ignored."""
+        model = RNNModel(
+            12,
+            "RNN",
+            10,
+            10,
+            **tfm_kwargs,
+            add_encoders={"cyclic": {"future": ["month"]}},
+        )
+
+        with patch.object(model, "_setup_for_train", return_value={}):
+            with patch.object(model, "_train", return_value=model):
+                model.fit_from_dataset(train_dataset=object())
+        mock_warning.assert_called_once()
+        assert "fit_from_dataset" in mock_warning.call_args.args[0]
+        assert "add_encoders" in mock_warning.call_args.args[0]
+
+        mock_warning.reset_mock()
+        with patch.object(model, "_setup_for_predict", return_value={}):
+            with patch.object(model, "_predict", return_value=[]):
+                model.predict_from_dataset(n=1, dataset=object())
+        mock_warning.assert_called_once()
+        assert "predict_from_dataset" in mock_warning.call_args.args[0]
+        assert "add_encoders" in mock_warning.call_args.args[0]
+
+    def test_encoders_not_set_from_dataset_no_warn(self):
+        """`fit_from_dataset`/`predict_from_dataset` do not warn when no encoders are set."""
+        model = RNNModel(12, "RNN", 10, 10, **tfm_kwargs)
+        with patch("darts.models.forecasting.torch_forecasting_model.logger.warning") as mock_warning:
+            with patch.object(model, "_setup_for_train", return_value={}):
+                with patch.object(model, "_train", return_value=model):
+                    model.fit_from_dataset(train_dataset=object())
+            with patch.object(model, "_setup_for_predict", return_value={}):
+                with patch.object(model, "_predict", return_value=[]):
+                    model.predict_from_dataset(n=1, dataset=object())
+        mock_warning.assert_not_called()
+
     def test_predict_after_fit_from_dataset(self):
         """Test that the model can predict after being trained with `fit_from_dataset` using all covariates."""
         icl, ocl = kwargs["input_chunk_length"], kwargs["output_chunk_length"]