Awesome

ImageNet subset generator

Generate a subsets from the original ImageNet1K dataset. Some commonly used subsets:

• SimclrV2 10% subset
• SemiViT 10% subset
• SimclrV2 1% subset
• SemiViT 1% subset
• Extreme low-shot subsets from MSN

Usage

• git clone https://github.com/BenediktAlkin/ImageNetSubsetGenerator
• cd ImageNetSubsetGenerator

Generate subset

• python main_subset.py --in1k_path <ImageNet1K_path> --out_path <out_path> --version in100_sololearn
• this will copy the corresponding samples from the ImageNet1K_path to out_path
• it can then be readily used with e.g. torchvision ImageFolder subset = ImageFolder(root=<out_path>)

For example: python main_subset.py --in1k_path /data/imagenet1k --out_path /data/imagenet1k_10percent_simclrv2 --version in1k_10percent_simclrv2

You can find all supported versions here or via python main_subset.py --help.

Check classes/samples of dataset

python main_statistics.py <path>

train n_classes: 1000
valid n_classes: 1000
train n_samples: 1282169
valid n_samples: 50000
train classes: ['n01440764', ...]
valid classes: ['n01440764', ...]