search

snap-stanford / orca

[ICLR 2022] Open-World Semi-Supervised Learning
87 stars 12 forks source link

The dataset of CIFAR100 #2

Closed RikkiXu closed 2 years ago

RikkiXu commented 2 years ago

Hi, Thanks for releasing this code! I notice that the train_unlabel_set and the test_set of CIFAR100 are from the same file: ‘./datasets/cifar-100-python/train’. This confuses me, and I think the test_set should come from ‘./datasets/cifar-100-python/test’ Looking forward to your response.

kaidic commented 2 years ago

Hi,

Thanks for your question. In the paper, we directly test on the unlabeled data in the training set, and it corresponds to the code above. I personally feel either testing on unlabeled training data or unlabeled test data makes sense. So I guess you could always use the validation split to do your own evaluation.