DatasetFactorization
DatasetFactorization copied to clipboard
Typo in run_cifar10_ipc1_style5.sh
Hi I believe that in run_cifar10_ipc1_style5.sh parameter ipc should be set to the value of 1 in order to reproduce the result from the paper ( 48.3±0.8) and to be consistent with the filename. Currently, it is set to 3.
Hi,
Thanks for your interest. You are correct. To reproduce the results in Tab. 1, ipc should be 1 and the flag '--single_channel' should be turned off. In practice, we found that when ipc is small, using images with single channel instead of RGB 3 channels yields better results. If single-channel image is used, ipc can be 3x given the same storage budget.
I have uploaded scripts of both single-channel 3x ipc and 3-channel 1x ipc.