Manually reordering predictions with get_preds

I’m currently in the process of migrating some projects to fastai 2 from fastai 1. When I was using v1 of fastai, when getting batch predictions with get_preds, I had to write a function to manually reorder the predictions from the learner to match the idx of the inputs, this was the case when I was doing image segmentation and text classification (using huggingface +fastai v1). This should be taken care of in fastai 2, with get_preds(reorder=True), right? Or do I still have to reorder based on idx of the inputs?

Looking at the code on github (NB: not truly digging in), it looks like reorder=True will only matter if the self.shuffle of the DataLoader you’re using to run inference on is set to True. I would guess that this is set to False by default for the validation DataLoader which is the default DataLoader unless you use a different ds_idx. See here for when then the shuffling occurs.

In my experience, I have always got predictions in the same order that files/rows were represented in my validation set.