From: "What classifiers know what they don't?"
In a folder with the train/ and val/ ImageNet sub-folders:
mkdir -p imagenot_in-domain/train
mkdir -p imagenot_in-domain/val
cat imagenot_in-domain.txt | xargs -I{} cp -R train/{} imagenot_in-domain/train/
cat imagenot_in-domain.txt | xargs -I{} cp -R val/{} imagenot_in-domain/val/
mkdir -p imagenot_out-domain/train
mkdir -p imagenot_out-domain/val
cat imagenot_out-domain.txt | xargs -I{} cp -R train/{} imagenot_out-domain/train/
cat imagenot_out-domain.txt | xargs -I{} cp -R val/{} imagenot_out-domain/val/
(val/ can be obtained using Soumith's valprep.sh script)
An easy to adapt example to use these folders in PyTorch is provided in dataset.py.