So in the tutorial on dataloading skimage is used to read in images from disk in the getitem method which are then transformed into tensors via torch.from_numpy(). As a basis serve pandas dataframes.
Now I read several times that Pil is the format to choose when it comes to loading and saving images in pytorch and in fact there are even dedicated methods for it in the torchvision package. So I wonder
-
Is it most efficient how its shown in the dataloading tutorial to load in the images from disk via paths provided in e.g. dataframes?
-
would it make sense to use data reading methods of PiL instead of skimage in the example? what advantages would that yield?