PREPARATION.md

Preparation

Install the requirements

Run the following command to install the dependences:

pip install -r requirements.txt

Data Preparation

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