- imagenet.txt -> contains all 1000 classes from ImageNet
- imagenet_dogs.txt -> contains 118 dog classes in ImageNet (obtained from https://arxiv.org/abs/1606.04080)
- imagenet_notdogs.txt -> contains 882 not dog classes in ImageNet
- imagenet_notdogs_subset.txt -> contains 118 not dog classes in ImageNet (subset of above)
imagenet.txt created using:
ls train > imagenet.txt
(train is the folder inside ILSVRC2017_CLS-LOC.tar.gz)
imagenet_notdogs.txt created using:
comm -3 <(sort imagenet.txt) <(sort imagenet_dogs.txt) > imagenet_notdogs.txt
imagenet_notdogs_subset.txt created using:
shuf -n 118 imagenet_notdogs.txt > imagenet_notdogs_subset.txt
In a folder with the train/ and val/ ImageNet sub-folders:
mkdir -p imagenet_notdogs/train
mkdir -p imagenet_notdogs/val
cat imagenet_notdogs.txt | xargs -I{} cp -R train/{} imagenet_notdogs/train/
cat imagenet_notdogs.txt | xargs -I{} cp -R val/{} imagenet_notdogs/val/
mkdir -p imagenet_dogs/train
mkdir -p imagenet_dogs/val
cat imagenet_dogs.txt | xargs -I{} cp -R train/{} imagenet_dogs/train/
cat imagenet_dogs.txt | xargs -I{} cp -R val/{} imagenet_dogs/val/
(val/ can be obtained using Soumith's valprep.sh script)
An example way to use these folders in PyTorch is provided in dataset.py.