Question about DM compression time

[Oshikaka]

Hi maintainers,

Thanks for the good work. In the paper, you wrote "previous SOTA method DM [63] needs 28,000 GPU hours to distill ImageNet-1K with 60% data processing." May I wonder how did you get this 28,000 hours (approx. 1166.7 days) result?

Thanks!

[vimar-gu]

Sorry for the late reply. As DM conducts class-wise sample optimization, we estimated the required time for the whole ImageNet-1K dataset based on the subset calculation time.