Skip to content

Dataset_preparation.md

Latest commit

 

History

History
34 lines (24 loc) · 2.57 KB

Dataset_preparation.md

File metadata and controls

34 lines (24 loc) · 2.57 KB

Datasets

For DTD, EuroSAT, FGVCAircraft, Flowers102, Food101, ImageNet, Oxford Pets, and Stanford Cars datasets, we use the standard dataset classes that come with torchvision. Creating the dataset directory structures according to official instructions should work with this code base. The value of the --root argument of the main script is passed to the root argument of the torchvision dataset classes directly.

For ImageNetV2, we use this implementation, which should download the dataset itself in the given directory.

Cub

Download the images and annotations from here. Extract the files in the DATA_ROOT directory and pass DATA_ROOT as the --root argument of the main script.

Places365

Download the Places365 dataset from here. In the DATA_ROOT directory, create another directory named places365. In this directory, place the categories_places365.txt and val.txt files as well as the val folder containing the images. Pass DATA_ROOT as the --root argument of the main script.

Stanford Dogs

Set the DATA_ROOT environment variable and run the following script to setup the stanford dogs dataset.

cd $DATA_ROOT
mkdir stanford_dogs
cd stanford_dogs
curl -OLJ http://vision.stanford.edu/aditya86/ImageNetDogs/images.tar
curl -OLJ http://vision.stanford.edu/aditya86/ImageNetDogs/annotation.tar
curl -OLJ http://vision.stanford.edu/aditya86/ImageNetDogs/lists.tar
curl -OLJ http://vision.stanford.edu/aditya86/ImageNetDogs/README.txt
tar -xvf images.tar
tar -xvf lists.tar

Pass DATA_ROOT as the --root argument of the main script.