Our experiments primarily use the ImageNet-1k dataset. Please organize the dataset in the following structure:
data/imagenet/
├── annotations
│ ├── train.json
| └── val.json
├── train
│ ├── n1440764
│ │ ├── 18.JPEG
│ │ └── ...
| └── ...
├── val
│ ├── n1440764
│ │ ├── 293.JPEG
│ │ └── ...
| └── ...
└── synsets.json
Both train.json and val.json have the following structure:
[{"image":"12925.JPEG","synset_id":449},...]
synsets.json maps synset IDs to their corresponding information:
{"1":{"WNID":"n02119789","words":"kit fox, Vulpes macrotis",...},...}
For detailed instructions on downloading and preparing the dataset, refer to https://toddai.readthedocs.io/en/latest/data/imagenet.html#imagenet-1k