Best Practice for Setting Precision In Dataset/DataModule

[fishbotics]

Hi all,

What's considered best practice for making sure the data from the datamodule has the correct precision set? Right now, I get errors thrown because my data is loaded from hdf5 as double and I use torch.as_tensor(...) to cast it. When I set precision=16 in my Trainer, I get errors because the types don't match. I can manually cast everything in my dataset, but then I have to change it anytime I change the precision in my trainer.

Thanks for your help!