Fix: prevent accumulation of SelectFields in PyTorchPredictor (#2951

)

* Prevent redundant accumulation of fields

* update fix

---------

Co-authored-by: Cameronwood611 <cwood611@uab.edu>
Co-authored-by: Lorenzo Stella <stellalo@amazon.com>

src/gluonts/torch/model/predictor.py

@@ -73,12 +73,12 @@ def network(self) -> nn.Module:
  def predict(
  self, dataset: Dataset, num_samples: Optional[int] = None
  ) -> Iterator[Forecast]:
- self.input_transform += SelectFields(
- self.input_names + self.required_fields, allow_missing=True
- )
  inference_data_loader = InferenceDataLoader(
  dataset,
- transform=self.input_transform,
+ transform=self.input_transform
+ + SelectFields(
+ self.input_names + self.required_fields, allow_missing=True
+ ),
  batch_size=self.batch_size,
  stack_fn=lambda data: batchify(data, self.device),
  )