nconsistency between torchvision.datasets.cifar.CIFAR10.data return and torchvision.datasets.cifar.CIFAR10.__getitem__() about vision HOT 3 CLOSED

Yes, that's what we should have done originally. With more time and resources, we would properly deprecate those and eventually rename them with a leading underscore. But honestly the benefit / cost ratio seems low at this moment

from vision.

Hi @CircuitsOfAIM

Don't access the samples through the .data or .targets attributes. These should not have been made public. Just use __getitem__.

These fields contain the raw data as read from the dataset, but do not contain any transformed data:

https://github.com/pytorch/vision/blob/main/torchvision/datasets/cifar.py#L104-L125

from vision.

Hi @CircuitsOfAIM

Don't access the samples through the .data or .targets attributes. These should not have been made public. Just use __getitem__.

These fields contain the raw data as read from the dataset, but do not contain any transformed data:

https://github.com/pytorch/vision/blob/main/torchvision/datasets/cifar.py#L104-L125

Thanks, that makes sense. So is it better to have them with single leading underscore , ._data and ._targets ?

from vision.