Run the following command to install the dependences:
pip install -r requirements.txt
We need to prepare ImageNet-1k and ImageNet-22k datasets from http://www.image-net.org/
.
- 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
- ImageNet-22k
ImageNet-22k contains 14 M images with 21,841 classes, without overlapping part with ImageNet-1k validation set.
The filelist (in22k_image_names.txt
) can be download at here.
Each class is stored as an archive file.
ImageNet-22k/
├── in22k_image_names.txt
├── n00004475.zip
├── n00005787.zip
├── n00006024.zip
...
├── n15102455.zip
└── n15102894.zip
The config DATA.FNAME_FORMAT
defines the image filename format in the archive file, default: {}.jpeg
.
We need IN-1k to evaluate the model, so the folders should be placed like the following (a soft link is available
):
datasets/
├── ImageNet-22k/ # the folder of IN-22k
└── ImageNet/ # the folder of IN-1k