I have found that Trainer.predict returns a prediction using NumPy arrays as output.
However, I believe they likely make use of Torch tensors internally during training and evaluation.
Do you have any insights into why the return value is of type NumPy instead of Torch tensors?
Any help would be appreciated!