search

VICO-UoE / DatasetCondensation

Dataset Condensation (ICLR21 and ICML21)
MIT License
457 stars 90 forks source link

Hyperparams for large synthetic sets #2

Closed pdejorge closed 3 years ago

pdejorge commented 3 years ago

Congratulations on this interesting work. It seems that in the code you define the hyperparams only for the dataset sizes that appear in the paper (up to 50 ipc), would it be possible to provide the hyperparams for the results reported in your ICLR comments to reviewers? (with 100, 200, 500, 1000 condensed images per class).

Thank you in advance!

PatrickZH commented 3 years ago

Thank you for being interested in our paper. As you see, the corresponding hyper-parameters have been given for these results (≤ 50 images/class) presented in the paper. For extra experiments with > 50 images/class, we pick the best one of “outer_loop, inner_loop = 50, 10” and “outer_loop, inner_loop = 20, 25” for different settings. You may find the better combinations if you try more. Besides, we find that initializing synthetic images from random real images will benefit the performance when learning many images. So, we do it when learning many images.

PatrickZH commented 2 years ago

For 100 img/cls learning on CIFAR10, if you try "outer_loop, inner_loop = 100, 2", you may get 66.7% (DSA).