Open XianzheMa opened 1 month ago
Actually, I want to completely remove the DownsamplingScheduler
, as we also have never used it in experiments.
I think the design of DownsamplingScheduler
, i.e. hard-code specifying "from trigger 3 to trigger 7, I want to use Downsampler X" by a user is hardly practical as the user cannot predict it in advance. We might want to have a more sophisticated scheduler based on data distribution, but this is out of scope for now.
What do you think @MaxiBoether ?
As we have multiple selector instances, the status of
DownsamplerScheduler
, i.e. which sampler is currently applied and which one should be used next, is out of sync. We should fix this.Relevant PR: https://github.com/eth-easl/modyn/pull/496#pullrequestreview-2112449835