Seeking advice for using max_train_steps or using max_train_epochs + num_repeats(dataset_repeats) for classic fine-tuning

[terrificdm]

Any advice or recommendations for choice between using max_train_steps and using max_train_epochs + num_repeats(dataset_repeats) for classic fine-tuning? Usually I configured Kohya's training scrip (fine_tune.py) with max_train_steps e.g. 15000 steps(default repeat number 1, and automatically computed epochs), I am wondering that should I configure max_train_epochs + num_repeats instead? Any different impact towards training result? Any tips? Thanks.