How can I trained on ImageNet-A as IN and evaluate on ImageNet as OOD?

[alirezaabdollahpour]

Hi ;) How can I train on ImageNet-A as the in-distribution (IN) dataset and evaluate on ImageNet as the out-of-distribution (OOD) dataset?

[gabrielilharco]

Hi @alirezaabdollahpour. Unlike ImageNet, ImageNet-A doesn't have a training set (only a test set), so I don't think this would be possible

[alirezaabdollahpour]

Hi @alirezaabdollahpour. Unlike ImageNet, ImageNet-A doesn't have a training set (only a test set), so I don't think this would be possible

Thank you, Gabriel, for your response. Can we assume that the test data represents the entire dataset and then split it into training and testing sets?

[gabrielilharco]

Sure, although I'd keep in mind that the dataset is quite small already, so splitting it like you suggested might lead to training difficulties and more noise in the evaluations

[alirezaabdollahpour]

Thanks