PVNet 5.3.2__tar.gz → 5.3.4__tar.gz

{pvnet-5.3.2 → pvnet-5.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PVNet
-Version: 5.3.2
+Version: 5.3.4
 Summary: PVNet
 Author-email: Peter Dudfield <info@openclimatefix.org>
 Requires-Python: <3.14,>=3.11

{pvnet-5.3.2 → pvnet-5.3.4}/PVNet.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: PVNet
-Version: 5.3.2
+Version: 5.3.4
 Summary: PVNet
 Author-email: Peter Dudfield <info@openclimatefix.org>
 Requires-Python: <3.14,>=3.11

{pvnet-5.3.2 → pvnet-5.3.4}/pvnet/optimizers.py

@@ -65,7 +65,7 @@ class AbstractOptimizer(ABC):
     """
 
     @abstractmethod
-    def __call__(self):
+    def __call__(self, model: Module):
         """Abstract call"""
         pass
 
@@ -129,19 +129,18 @@ class EmbAdamWReduceLROnPlateau(AbstractOptimizer):
             {"params": decay, "weight_decay": self.weight_decay},
             {"params": no_decay, "weight_decay": 0.0},
         ]
+        monitor = "quantile_loss/val" if model.use_quantile_regression else "MAE/val"
         opt = torch.optim.AdamW(optim_groups, lr=self.lr, **self.opt_kwargs)
-
         sch = torch.optim.lr_scheduler.ReduceLROnPlateau(
             opt,
             factor=self.factor,
             patience=self.patience,
             threshold=self.threshold,
         )
-        sch = {
-            "scheduler": sch,
-            "monitor": "quantile_loss/val" if model.use_quantile_regression else "MAE/val",
+        return {
+            "optimizer": opt,
+            "lr_scheduler": {"scheduler": sch, "monitor": monitor},
         }
-        return [opt], [sch]
 
 
 class AdamWReduceLROnPlateau(AbstractOptimizer):
@@ -153,15 +152,13 @@ class AdamWReduceLROnPlateau(AbstractOptimizer):
         patience: int = 3, 
         factor: float = 0.5, 
         threshold: float = 2e-4, 
-        step_freq=None, 
         **opt_kwargs,
     ):
         """AdamW optimizer and reduce on plateau scheduler"""
-        self._lr = lr
+        self.lr = lr
         self.patience = patience
         self.factor = factor
         self.threshold = threshold
-        self.step_freq = step_freq
         self.opt_kwargs = opt_kwargs
 
     def _call_multi(self, model):
@@ -169,7 +166,7 @@ class AdamWReduceLROnPlateau(AbstractOptimizer):
 
         group_args = []
 
-        for key in self._lr.keys():
+        for key in self.lr.keys():
             if key == "default":
                 continue
 
@@ -178,43 +175,38 @@ class AdamWReduceLROnPlateau(AbstractOptimizer):
                 if param_name.startswith(key):
                     submodule_params += [remaining_params.pop(param_name)]
 
-            group_args += [{"params": submodule_params, "lr": self._lr[key]}]
+            group_args += [{"params": submodule_params, "lr": self.lr[key]}]
 
         remaining_params = [p for k, p in remaining_params.items()]
         group_args += [{"params": remaining_params}]
-
-        opt = torch.optim.AdamW(
-            group_args, 
-            lr=self._lr["default"] if model.lr is None else model.lr, 
-            **self.opt_kwargs,
+        monitor = "quantile_loss/val" if model.use_quantile_regression else "MAE/val"
+        opt = torch.optim.AdamW(group_args, lr=self.lr["default"], **self.opt_kwargs)
+        sch = torch.optim.lr_scheduler.ReduceLROnPlateau(
+            opt,
+            factor=self.factor,
+            patience=self.patience,
+            threshold=self.threshold,
         )
-        sch = {
-            "scheduler": torch.optim.lr_scheduler.ReduceLROnPlateau(
-                opt,
-                factor=self.factor,
-                patience=self.patience,
-                threshold=self.threshold,
-            ),
-            "monitor": "quantile_loss/val" if model.use_quantile_regression else "MAE/val",
+        return {
+            "optimizer": opt,
+            "lr_scheduler": {"scheduler": sch, "monitor": monitor},
         }
 
-        return [opt], [sch]
 
     def __call__(self, model):
         """Return optimizer"""
-        if not isinstance(self._lr, float):
+        if isinstance(self.lr, dict):
             return self._call_multi(model)
         else:
-            default_lr = self._lr if model.lr is None else model.lr
-            opt = torch.optim.AdamW(model.parameters(), lr=default_lr, **self.opt_kwargs)
+            monitor = "quantile_loss/val" if model.use_quantile_regression else "MAE/val"
+            opt = torch.optim.AdamW(model.parameters(), lr=self.lr, **self.opt_kwargs)
             sch = torch.optim.lr_scheduler.ReduceLROnPlateau(
                 opt,
                 factor=self.factor,
                 patience=self.patience,
                 threshold=self.threshold,
             )
-            sch = {
-                "scheduler": sch,
-                "monitor": "quantile_loss/val" if model.use_quantile_regression else "MAE/val",
+            return {
+                "optimizer": opt,
+                "lr_scheduler": {"scheduler": sch, "monitor": monitor},
             }
-            return [opt], [sch]

{pvnet-5.3.2 → pvnet-5.3.4}/pvnet/training/lightning_module.py

@@ -109,7 +109,7 @@ class PVNetLightningModule(pl.LightningModule):
         losses = self._calculate_common_losses(y, y_hat)
         losses = {f"{k}/train": v for k, v in losses.items()}
 
-        self.log_dict(losses, on_step=True, on_epoch=True)
+        self.log_dict(losses, on_step=True, on_epoch=True, batch_size=y.size(0))
 
         if self.model.use_quantile_regression:
             opt_target = losses["quantile_loss/train"]
@@ -256,11 +256,28 @@ class PVNetLightningModule(pl.LightningModule):
 
         # Calculate the persistance losses - we only need to do this once per training run
         # not every epoch
-        if self.current_epoch == 0:
+        if self.current_epoch==0:
+            # Need to find last valid value before forecast
+            target_data = batch["generation"]
+            history_data = target_data[:, :-(self.model.forecast_len)]
+            
+            # Find where values aren't dropped
+            valid_mask = history_data >= 0
+            
+            # Last valid value index for each sample
+            flipped_mask = valid_mask.float().flip(dims=[1])
+            last_valid_indices_flipped = torch.argmax(flipped_mask, dim=1)
+            last_valid_indices = history_data.shape[1] - 1 - last_valid_indices_flipped
+            
+            # Grab those last valid values
+            batch_indices = torch.arange(
+                history_data.shape[0], 
+                device=history_data.device
+            )
+            last_valid_values = history_data[batch_indices, last_valid_indices]
+
             y_persist = (
-                batch["generation"][:, -(self.model.forecast_len + 1)]
-                .unsqueeze(1)
-                .expand(-1, self.model.forecast_len)
+                last_valid_values.unsqueeze(1).expand(-1, self.model.forecast_len)
             )
             mae_step_persist, mse_step_persist = self._calculate_step_metrics(y, y_persist)
             self._val_persistence_horizon_maes.append(mae_step_persist)
@@ -272,7 +289,7 @@ class PVNetLightningModule(pl.LightningModule):
             )
 
         # Log the metrics
-        self.log_dict(losses, on_step=False, on_epoch=True)
+        self.log_dict(losses, on_step=False, on_epoch=True, batch_size=y.size(0))
 
     def on_validation_epoch_end(self) -> None:
         """Run on epoch end"""