prepare.md

Preparation

• The only dataset required in this repo is ImageNet, which is enough for pretraining, finetuning, linear evaluation and few-shot evaluation. If you want to evaluate on COCO, LVIS, ADE20k and robustness datasets, please follow the corresponding repos to prepare the data.

Installation

• Python >=3.7
• We recommend to use Pytorch1.11 for a faster training speed.
• timm == 0.6.12
• numpy == 1.21.5
• tensorboard

To run few-shot evaluation, cyanure package is further required. You can install it with

  pip install cyanure-openblas
  # or pip install cyanure-mkl

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/. The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val folder respectively:

  /path/to/imagenet/
      ├── train/
      │   ├── class1/
      │   │   ├── img1.JPEG
      |   │   ├── img2.JPEG
      |   │   ├── img3.JPEG
      |   │   └── ...
      │   ├── class2/
      |   │   └── ...   
      │   ├── class3/
      |   │   └── ...
      |   └── ...
      └─── val
      │   ├── class1/
      │   │   ├── img4.JPEG
      |   │   ├── img5.JPEG
      |   │   ├── img6.JPEG
      |   │   └── ...
      │   ├── class2/
      |   │   └── ...   
      │   ├── class3/
      |   │   └── ...

Note that raw val images are not put into class folders, use this script to get correct layout.