Is your feature request related to a problem? Please describe.
Currently, we always deepcopy the cache content for every epoch in CacheDataset, actually it's not necessary for multi-processing DataLoader or the random transforms don't change the cache data at all.