Hyperparameter tuning in torchdistill

[nighting0le01]

question 1 : Hi @yoshitomo-matsubara how can I specify to using only a small subset of imagenet train for hyperparameter tuning ? I have the dataset locally downloaded. Can i specify it through yaml?

[yoshitomo-matsubara]

You can use Subset class in torchvision and register the class or function to instantiate so that you can define it (as imagenet_subset is the name of registered dataset in the following example) in a yaml config

from torch.utils.data import ImageFolder, Subset
from torchdistill.datasets.registry import register_dataset

@register_dataset
def imagenet_subset(indices, **image_folder_kwargs):
	org_dataset = ImageFolder(**image_folder_kwargs)
	sub_dataset = Subset(org_dataset, indices)
	return sub_dataset

Another option is randomly splitting your dataset like this (dummy_train:dummy_val = 9:1)