Fix/ptl fixes

darts/models/forecasting/block_rnn_model.py

@@ -213,8 +213,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/models/forecasting/nbeats.py

@@ -546,8 +546,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/models/forecasting/pl_forecasting_module.py

@@ -252,10 +252,17 @@ def _create_from_cls_and_kwargs(cls, kws):
         if self.lr_scheduler_cls is not None:
             lr_sched_kws = {k: v for k, v in self.lr_scheduler_kwargs.items()}
             lr_sched_kws["optimizer"] = optimizer
+
+            # ReduceLROnPlateau requires a metric to "monitor" which must be set separately, most others do not
+            lr_monitor = lr_sched_kws.pop("monitor", None)
+
             lr_scheduler = _create_from_cls_and_kwargs(
                 self.lr_scheduler_cls, lr_sched_kws
             )
-            return [optimizer], [lr_scheduler]
+            return [optimizer], {
+                "scheduler": lr_scheduler,
+                "monitor": lr_monitor if lr_monitor is not None else "val_loss",
+            }
         else:
             return optimizer
 

darts/models/forecasting/rnn_model.py

@@ -279,8 +279,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/models/forecasting/tcn_model.py

@@ -332,8 +332,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/models/forecasting/tft_model.py

@@ -639,8 +639,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/models/forecasting/torch_forecasting_model.py

@@ -161,8 +161,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).
@@ -336,7 +343,9 @@ def __init__(
         self.pl_module_params: Optional[Dict] = None
 
     @staticmethod
-    def _extract_torch_devices(torch_device_str) -> Tuple[str, Optional[list], bool]:
+    def _extract_torch_devices(
+        torch_device_str,
+    ) -> Tuple[str, Optional[Union[list, int]], bool]:
         """This method handles the deprecated `torch_device_str` and should be removed in a future Darts version.
 
         Returns
@@ -372,13 +381,13 @@ def _extract_torch_devices(torch_device_str) -> Tuple[str, Optional[list], bool]
 
         gpus = None
         auto_select_gpus = False
-        accelerator = device_split[0]
-        if len(device_split) == 2 and accelerator == "cuda":
+        accelerator = "gpu" if device_split[0] == "cuda" else device_split[0]
+
+        if len(device_split) == 2 and accelerator == "gpu":
             gpus = device_split[1]
             gpus = [int(gpus)]
         elif len(device_split) == 1:
-            if accelerator == "cuda":
-                accelerator = "gpu"
+            if accelerator == "gpu":
                 gpus = -1
                 auto_select_gpus = True
         else:
@@ -389,9 +398,29 @@ def _extract_torch_devices(torch_device_str) -> Tuple[str, Optional[list], bool]
             )
         return accelerator, gpus, auto_select_gpus
 
-    @staticmethod
-    def _extract_torch_model_params(**kwargs):
+    @classmethod
+    def _validate_model_kwargs(cls, **kwargs):
+        """validate that kwargs used at model creation are part of :class:`TorchForecastingModel`,
+        :class:`PLForecastingModule` or cls __init__ methods.
+        """
+        valid_kwargs = (
+            set(inspect.signature(TorchForecastingModel.__init__).parameters.keys())
+            | set(inspect.signature(PLForecastingModule.__init__).parameters.keys())
+            | set(inspect.signature(cls.__init__).parameters.keys())
+        )
+
+        invalid_kwargs = [kwarg for kwarg in kwargs if kwarg not in valid_kwargs]
+
+        raise_if(
+            len(invalid_kwargs) > 0,
+            f"Invalid model creation parameters. Model `{cls.__name__}` has no args/kwargs `{invalid_kwargs}`",
+            logger=logger,
+        )
+
+    @classmethod
+    def _extract_torch_model_params(cls, **kwargs):
         """extract params from model creation to set up TorchForecastingModels"""
+        cls._validate_model_kwargs(**kwargs)
         get_params = list(
             inspect.signature(TorchForecastingModel.__init__).parameters.keys()
         )
@@ -619,6 +648,13 @@ def fit(
             override Darts' default trainer.
         verbose
             Optionally, whether to print progress.
+
+            .. deprecated:: v0.17.0
+                ``verbose`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, control verbosity with PyTorch Lightning Trainer parameters ``enable_progress_bar``,
+                ``progress_bar_refresh_rate`` and ``enable_model_summary`` in the ``pl_trainer_kwargs`` dict
+                at model creation. See for example here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#enable-progress-bar
         epochs
             If specified, will train the model for ``epochs`` (additional) epochs, irrespective of what ``n_epochs``
             was provided to the model constructor.
@@ -764,6 +800,13 @@ def fit_from_dataset(
             override Darts' default trainer.
         verbose
             Optionally, whether to print progress.
+
+            .. deprecated:: v0.17.0
+                ``verbose`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, control verbosity with PyTorch Lightning Trainer parameters ``enable_progress_bar``,
+                ``progress_bar_refresh_rate`` and ``enable_model_summary`` in the ``pl_trainer_kwargs`` dict
+                at model creation. See for example here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#enable-progress-bar
         epochs
             If specified, will train the model for ``epochs`` (additional) epochs, irrespective of what ``n_epochs``
             was provided to the model constructor.
@@ -965,6 +1008,13 @@ def predict(
             Size of batches during prediction. Defaults to the models' training ``batch_size`` value.
         verbose
             Optionally, whether to print progress.
+
+            .. deprecated:: v0.17.0
+                ``verbose`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, control verbosity with PyTorch Lightning Trainer parameters ``enable_progress_bar``,
+                ``progress_bar_refresh_rate`` and ``enable_model_summary`` in the ``pl_trainer_kwargs`` dict
+                at model creation. See for example here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#enable-progress-bar
         n_jobs
             The number of jobs to run in parallel. ``-1`` means using all processors. Defaults to ``1``.
         roll_size
@@ -1084,7 +1134,14 @@ def predict_from_dataset(
         batch_size
             Size of batches during prediction. Defaults to the models ``batch_size`` value.
         verbose
-            Shows the progress bar for batch predicition. Off by default.
+            Optionally, whether to print progress.
+
+            .. deprecated:: v0.17.0
+                ``verbose`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, control verbosity with PyTorch Lightning Trainer parameters ``enable_progress_bar``,
+                ``progress_bar_refresh_rate`` and ``enable_model_summary`` in the ``pl_trainer_kwargs`` dict
+                at model creation. See for example here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#enable-progress-bar
         n_jobs
             The number of jobs to run in parallel. ``-1`` means using all processors. Defaults to ``1``.
         roll_size

darts/models/forecasting/transformer_model.py

@@ -289,8 +289,15 @@ def __init__(
             Number of epochs to wait before evaluating the validation loss (if a validation
             ``TimeSeries`` is passed to the :func:`fit()` method).
         torch_device_str
-            Optionally, a string indicating the torch device to use. (default: "cuda:0" if a GPU
+            Optionally, a string indicating the torch device to use. (default: ``None``. Set "cuda:0" if a GPU
             is available, otherwise "cpu")
+
+            .. deprecated:: v0.17.0
+                ``torch_device_str`` has been deprecated in v0.17.0 and will be removed in a future version.
+                Instead, specify this with keys ``"accelerator", "gpus", "auto_select_gpus"`` in your
+                ``pl_trainer_kwargs`` dict. For more info, see here:
+                https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#trainer-flags , and
+                https://pytorch-lightning.readthedocs.io/en/stable/advanced/multi_gpu.html#select-gpu-devices
         force_reset
             If set to ``True``, any previously-existing model with the same name will be reset (all checkpoints will
             be discarded).

darts/tests/models/forecasting/test_torch_forecasting_model.py

@@ -12,6 +12,8 @@
 logger = get_logger(__name__)
 
 try:
+    import torch
+
     from darts.models.forecasting.rnn_model import RNNModel
 
     TORCH_AVAILABLE = True
@@ -285,3 +287,80 @@ def test_train_from_10_n_epochs_20_fit_15_epochs(self):
 
             model1.fit(series, epochs=15)
             self.assertEqual(15, model1.epochs_trained)
+
+        def test_optimizers(self):
+            times = pd.date_range("20130101", "20130410")
+            pd_series = pd.Series(range(100), index=times)
+            series = TimeSeries.from_series(pd_series)
+
+            optimizers = [
+                (torch.optim.Adam, {"lr": 0.001}),
+                (torch.optim.SGD, {"lr": 0.001}),
+            ]
+
+            for optim_cls, optim_kwargs in optimizers:
+                model = RNNModel(
+                    12,
+                    "RNN",
+                    10,
+                    10,
+                    optimizer_cls=optim_cls,
+                    optimizer_kwargs=optim_kwargs,
+                )
+                # should not raise an error
+                model.fit(series, epochs=1)
+
+        def test_lr_schedulers(self):
+            times = pd.date_range("20130101", "20130410")
+            pd_series = pd.Series(range(100), index=times)
+            series = TimeSeries.from_series(pd_series)
+
+            lr_schedulers = [
+                (torch.optim.lr_scheduler.LinearLR, {}),
+                (
+                    torch.optim.lr_scheduler.ReduceLROnPlateau,
+                    {"threshold": 0.001, "monitor": "train_loss"},
+                ),
+                (torch.optim.lr_scheduler.ExponentialLR, {"gamma": 0.09}),
+            ]
+
+            for lr_scheduler_cls, lr_scheduler_kwargs in lr_schedulers:
+                model = RNNModel(
+                    12,
+                    "RNN",
+                    10,
+                    10,
+                    lr_scheduler_cls=lr_scheduler_cls,
+                    lr_scheduler_kwargs=lr_scheduler_kwargs,
+                )
+                # should not raise an error
+                model.fit(series, epochs=1)
+
+        def test_devices(self):
+            torch_devices = [
+                ("cpu", ("cpu", None, False)),
+                ("cuda:0", ("gpu", [0], False)),
+                ("cuda", ("gpu", -1, True)),
+                ("auto", ("auto", None, False)),
+            ]
+
+            for torch_device, settings in torch_devices:
+                accelerator, gpus, auto_select_gpus = settings
+                model = RNNModel(12, "RNN", 10, 10, torch_device_str=torch_device)
+
+                self.assertEqual(model.trainer_params["accelerator"], accelerator)
+                self.assertEqual(model.trainer_params["gpus"], gpus)
+                self.assertEqual(
+                    model.trainer_params["auto_select_gpus"], auto_select_gpus
+                )
+
+        def test_wrong_model_creation_params(self):
+            valid_kwarg = {"pl_trainer_kwargs": {}}
+            invalid_kwarg = {"some_invalid_kwarg": None}
+
+            # valid params should not raise an error
+            _ = RNNModel(12, "RNN", 10, 10, **valid_kwarg)
+
+            # invalid params should raise an error
+            with self.assertRaises(ValueError):
+                _ = RNNModel(12, "RNN", 10, 10, **invalid_kwarg)