#47 added _step logic for tendency based learning

Co-authored-by: Jakob Schloer <[email protected]>

src/anemoi/training/train/forecaster.py

@@ -221,6 +221,9 @@ def _step(
     ) -> tuple[torch.Tensor, Mapping[str, torch.Tensor]]:
         return self.step_functions[self.prediction_mode](batch, batch_idx, validation_mode, in_place_proc)
 
+    # NOTE (jakob-schloer): Observation on nomenclature - is this _step_residual function only residual if the "self.model" has a residual structure???
+    # NOTE (rilwan-adewoying): Naming problems can maybe be solved by moving alot of this tendency logic in _step_tendency to advance_input
+
     def _step_residual(
         self,
         batch: torch.Tensor,

src/anemoi/training/train/train.py

@@ -138,6 +138,7 @@ def model(self) -> GraphForecaster:
             "graph_data": self.graph_data,
             "metadata": self.metadata,
             "statistics": self.datamodule.statistics,
+            "statistics_tendencies": self.datamodule.statistics_tendencies,
         }
         if self.load_weights_only:
             LOGGER.info("Restoring only model weights from %s", self.last_checkpoint)
@@ -328,7 +329,7 @@ def train(self) -> None:
             # run a fixed no of batches per epoch (helpful when debugging)
             limit_train_batches=self.config.dataloader.limit_batches.training,
             limit_val_batches=self.config.dataloader.limit_batches.validation,
-            num_sanity_val_steps=4,
+            num_sanity_val_steps=self.config.training.num_sanity_val_steps,
             accumulate_grad_batches=self.config.training.accum_grad_batches,
             gradient_clip_val=self.config.training.gradient_clip.val,
             gradient_clip_algorithm=self.config.training.gradient_clip.algorithm,