I think a good solution can be found here: Changing transforms after creating a dataset - #7 by Brando_Miranda
train_dataset = MyDataset(train_transform)
val_dataset = MyDataset(val_transform)
train_indices, val_indices = sklearn.model_selection.train_test_split(indices)
train_dataset = torch.utils.data.Subset(train_dataset, train_indices)
val_dataset = torch.utils.data.Subset(val_dataset, val_indices)