fix: only checking pred path if it's a string or path

Signed-off-by: Kin Long Kelvin Lee <[email protected]>

docs/source/inference.rst

@@ -36,7 +36,9 @@ The ``ParityInferenceTask`` helps perform this task by using the PyTorch Lightni
 The default ``Trainer`` settings will create a ``lightning_logs`` directory, followed by an experiment
 number. Within it, once your inference run completes, there will be a ``inference_data.json`` that you
 can then load in. The data is sorted by the name of the target (e.g. ``energy``, ``bandgap``), under
-these keys, ``predictions`` and ``targets``.
+these keys, ``predictions`` and ``targets``. Note that ``pred_split`` does not necessarily have to be
+a completely different hold out: you can pass your training LMDB path if you wish to double check the
+performance of your model after training, or you can use it with unseen samples.
 
 .. note::
 

matsciml/lightning/data_utils.py

@@ -250,7 +250,7 @@ def setup(self, stage: str | None = None) -> None:
                 dset = self._make_dataset(split_path, self.dataset)
                 splits[key] = dset
         # specialty case for 'inference' or prediction runs
-        if hasattr(self.hparams, "pred_split"):
+        if isinstance(self.hparams.pred_split, (str, Path)):
             pred_split_path = self.hparams.pred_split
             if isinstance(pred_split_path, str):
                 pred_split_path = Path(pred_split_path)