Preparation
Install the dependencies
pip install -r requirements-training.txt
pip install -v -e .
Data Preparation
We need to prepare ImageNet-1k datasets to do zero-shot classification task.
- ImageNet-1k
ImageNet-1k contains 1.28 M images for training and 50 K images for validation.
The train set and validation set should be saved as the *.tar archives:
ImageNet/
├── train.tar
└── val.tar
Our code also supports storing images as individual files as follow:
ImageNet/
├── train
│ ├── n01440764
│ │ ├── n01440764_10026.JPEG
│ │ ├── n01440764_10027.JPEG
...
├── val
│ ├── n01440764
│ │ ├── ILSVRC2012_val_00000293.JPEG