[BUG] Bugfix when no exogenous variable is passed to TFT

pytorch_forecasting/models/temporal_fusion_transformer/sub_modules.py

@@ -336,7 +336,8 @@ def forward(self, x: Dict[str, torch.Tensor], context: torch.Tensor = None):
 
             outputs = var_outputs * sparse_weights
             outputs = outputs.sum(dim=-1)
-        else:  # for one input, do not perform variable selection but just encoding
+        elif self.num_inputs == 1:
+            # for one input, do not perform variable selection but just encoding
             name = next(iter(self.single_variable_grns.keys()))
             variable_embedding = x[name]
             if name in self.prescalers:
@@ -346,6 +347,12 @@ def forward(self, x: Dict[str, torch.Tensor], context: torch.Tensor = None):
                 sparse_weights = torch.ones(outputs.size(0), outputs.size(1), 1, 1, device=outputs.device)  #
             else:  # ndim == 2 -> batch size, hidden size, n_variables
                 sparse_weights = torch.ones(outputs.size(0), 1, 1, device=outputs.device)
+        else:  # for no input
+            outputs = torch.zeros(context.size(), device=context.device)
+            if outputs.ndim == 3:  # -> batch size, time, hidden size, n_variables
+                sparse_weights = torch.zeros(outputs.size(0), outputs.size(1), 1, 0, device=outputs.device)
+            else:  # ndim == 2 -> batch size, hidden size, n_variables
+                sparse_weights = torch.zeros(outputs.size(0), 1, 0, device=outputs.device)
         return outputs, sparse_weights
 
 

tests/test_models/test_temporal_fusion_transformer.py

@@ -6,6 +6,7 @@
 from lightning.pytorch.callbacks import EarlyStopping
 from lightning.pytorch.loggers import TensorBoardLogger
 import numpy as np
+import pandas as pd
 import pytest
 import torch
 
@@ -421,3 +422,50 @@ def test_hyperparameter_optimization_integration(dataloaders_with_covariates, tm
         )
     finally:
         shutil.rmtree(tmp_path, ignore_errors=True)
+
+
+def test_no_exogenous_variable():
+    data = pd.DataFrame(
+        {
+            "target": np.ones(1600),
+            "group_id": np.repeat(np.arange(16), 100),
+            "time_idx": np.tile(np.arange(100), 16),
+        }
+    )
+    training_dataset = TimeSeriesDataSet(
+        data=data,
+        time_idx="time_idx",
+        target="target",
+        group_ids=["group_id"],
+        max_encoder_length=10,
+        max_prediction_length=5,
+        time_varying_unknown_reals=["target"],
+        time_varying_known_reals=[],
+    )
+    validation_dataset = TimeSeriesDataSet.from_dataset(training_dataset, data, stop_randomization=True, predict=True)
+    training_data_loader = training_dataset.to_dataloader(train=True, batch_size=8, num_workers=0)
+    validation_data_loader = validation_dataset.to_dataloader(train=False, batch_size=8, num_workers=0)
+    forecaster = TemporalFusionTransformer.from_dataset(
+        training_dataset,
+        log_interval=1,
+    )
+    from lightning.pytorch import Trainer
+
+    trainer = Trainer(
+        max_epochs=2,
+        limit_train_batches=8,
+        limit_val_batches=8,
+    )
+    trainer.fit(
+        forecaster,
+        train_dataloaders=training_data_loader,
+        val_dataloaders=validation_data_loader,
+    )
+    best_model_path = trainer.checkpoint_callback.best_model_path
+    best_model = TemporalFusionTransformer.load_from_checkpoint(best_model_path)
+    best_model.predict(
+        validation_data_loader,
+        return_x=True,
+        return_y=True,
+        return_index=True,
+    )