How to retain the original labels of test/train targets?

[srvCodes]

Hi,

I have been trying to retain the original labels of test/train set targets, for example, the targets in lines 45 and 50 in the protonet training script. Could you please help?

By original labels, I refer to the integer labels before the targets get mapped to the [0, n_way-1] range during few-shot training.

[vivektrivedy]

Any update on this? @srvCodes

[vivektrivedy]

For the existing datasets we can simply pass in target_transform = None to access the original labels.

from torchmeta.datasets.helpers import cifar_fs
from torchmeta.utils.data import BatchMetaDataLoader

cfs = cifar_fs('data', shots=5, ways=5, test_shots=5, meta_train=True, download=True, target_transform = None)
dataloader = BatchMetaDataLoader(cfs, batch_size=1,shuffle=False)

To check:

sample = next(iter(dataloader))
original_labels = sample['train'][1][0][1]