[lightning] refactored Trainer args

neuralprophet/forecaster.py

@@ -294,16 +294,6 @@ class NeuralProphet:
             Options
                 * ``True``: test data is normalized with global data params even if trained with local data params (global modeling with local normalization)
                 * (default) ``False``: no global modeling with local normalization
-        logger: str
-            Name of logger from pytorch_lightning.loggers to log metrics to.
-
-            Options
-                * TensorBoardLogger
-                * CSVLogger
-                * (MLFlowLogger)
-                * (NeptuneLogger)
-                * (CometLogger)
-                * (WandbLogger)
         trainer_config: dict
             Dictionary of additional trainer configuration parameters.
     """
@@ -346,7 +336,6 @@ def __init__(
         global_normalization=False,
         global_time_normalization=True,
         unknown_data_normalization=False,
-        logger=None,
         trainer_config={},
     ):
         kwargs = locals()
@@ -411,7 +400,6 @@ def __init__(
 
         # Pytorch Lightning Trainer
         self.metrics_logger = MetricsLogger(save_dir=os.getcwd())
-        self.additional_logger = logger
         self.trainer_config = trainer_config
         self.trainer = None
 
@@ -2387,8 +2375,9 @@ def _train(self, df, df_val=None, minimal=False, continue_training=False):
             config_train=self.config_train,
             config=self.trainer_config,
             metrics_logger=self.metrics_logger,
-            additional_logger=self.additional_logger,
             early_stopping_target="Loss_val" if df_val is not None else "Loss",
+            minimal=minimal,
+            num_batches_per_epoch=len(train_loader),
         )
 
         # Set parameters for the learning rate finder
@@ -2449,7 +2438,6 @@ def restore_trainer(self):
             config_train=self.config_train,
             config=self.trainer_config,
             metrics_logger=self.metrics_logger,
-            additional_logger=self.additional_logger,
         )
         self.metrics = metrics.get_metrics(self.collect_metrics)
 

neuralprophet/utils.py

@@ -747,8 +747,9 @@ def configure_trainer(
     config_train: dict,
     config: dict,
     metrics_logger,
-    additional_logger: str = None,
     early_stopping_target: str = "Loss",
+    minimal=False,
+    num_batches_per_epoch=100,
 ):
     """
     Configures the PyTorch Lightning trainer.
@@ -761,10 +762,12 @@ def configure_trainer(
             dictionary containing the custom PyTorch Lightning trainer configuration.
         metrics_logger : MetricsLogger
             MetricsLogger object to log metrics to.
-        additional_logger : str
-            Name of logger from pytorch_lightning.loggers to log metrics to.
         early_stopping_target : str
             Target metric to use for early stopping.
+        minimal : bool
+            If True, no metrics are logged and no progress bar is displayed.
+        num_batches_per_epoch : int
+            Number of batches per epoch.
 
     Returns
     -------
@@ -782,37 +785,32 @@ def configure_trainer(
         if config_train.epochs is not None:
             config["max_epochs"] = config_train.epochs
 
-    # Auto-configure the metric logging frequency
-    if "log_every_n_steps" not in config.keys() and "max_epochs" in config.keys():
-        config["log_every_n_steps"] = math.ceil(config["max_epochs"] / 10)
-
     # Configure the logthing-logs directory
     if "default_root_dir" not in config.keys():
         config["default_root_dir"] = os.getcwd()
 
-    # Configure the loggers
-    # TODO: technically additional loggers work, but somehow the TensorBoard logger interferes with the custom
-    # metrics logger. Resolve before activating this feature. Docs: https://pytorch-lightning.readthedocs.io/en/stable/extensions/logging.html
-    # if additional_logger in pl.loggers.__all__:  # TODO: pl.loggers.__all__ seems to be incomplete
-    #     Logger = importlib.import_module(f"pytorch_lightning.loggers.__init__").__dict__[additional_logger]
-    #     config["logger"] = [Logger(config["default_root_dir"]), metrics_logger]
-    # elif additional_logger is not None:
-    #     log.error(f"Additional logger {additional_logger} not found in pytorch_lightning.loggers")
-    if additional_logger:
-        log.error("Additional loggers are not yet supported")
+    # Configure callbacks
+    callbacks = []
+
+    # Configure the logger
+    if minimal:
+        config["enable_progress_bar"] = False
+        config["enable_model_summary"] = False
+        config["logger"] = False
     else:
         config["logger"] = metrics_logger
-
-    # Configure callbacks
-    config["callbacks"] = []
+        # Configure the progress bar, refresh every 2nd batch
+        prog_bar_callback = pl.callbacks.TQDMProgressBar(refresh_rate=max(1, int(num_batches_per_epoch / 4)))
+        callbacks.append(prog_bar_callback)
 
     # Early stopping monitor
     if config_train.early_stopping:
         early_stop_callback = pl.callbacks.EarlyStopping(
             monitor=early_stopping_target, mode="min", patience=20, divergence_threshold=5.0
         )
-        config["callbacks"].append(early_stop_callback)
+        callbacks.append(early_stop_callback)
 
+    config["callbacks"] = callbacks
     config["num_sanity_val_steps"] = 0
 
     return pl.Trainer(**config)