[ImageNet/Pytorch] Cross-validation and/or Holdout Set

[nd26]

Related to Pytorch/ImageNet for any model

Is your feature request related to a problem? Please describe.
It would be a nice add-on to support the aforementioned types of validation (e.g. k-fold cross-validation and holdout set).

Describe the solution you'd like
This would require some code addition on where the dataloaders are defined. For instance, for a pytorch dataloader, it can be implemented using a StratifiedShuffleSplit followed by a SubsetRandomSampler/SubsetSampler which would be passed as a sampler to the training Dataloader (I'd assume it is straightforward for DALI too).

Describe alternatives you've considered
Perhaps this could be added as a preprocessing step through a bash script (i.e. separating a priori the training set in different folders, ie. img_fold_1, img_fold_2, etc. and appropriately if it's just a holdout set).