AFAIK you pretty much always want batch-level logging for training and epoch-level for validation (and end-of-epoch training)?

→ View original post on X — @jeremyphoward