GeorgeCazenavette / mtt-distillation

Official code for our CVPR '22 paper "Dataset Distillation by Matching Training Trajectories"
https://georgecazenavette.github.io/mtt-distillation/
Other
395 stars 55 forks source link

details about the full real subsets of ImageNet #35

Closed zeyuanyin closed 1 year ago

zeyuanyin commented 1 year ago

Hi,

I am interested about your work, and you use a special dataset called subsets of ImageNet. From the rare information in the paper and this repo, I know in each subsets of ImageNet,

GeorgeCazenavette commented 1 year ago

The resolution is actually 128x128 for the results in the paper.

The number of images per class is the same as in the original ImageNet. Most classes have 1300 images, but some have a few less. All classes have 50 test images.

Note: the original authors of Imagenette and Imagewoof released a v2 version with rebalanced train and test splits. We did not do that here.

zeyuanyin commented 1 year ago

@GeorgeCazenavette Thanks for your detailed reply.