about runner iter

[yuzhobotianzhiyu]

Very good work.

1. I would like to ask if there is any plan to open source on mmdet.
2. If my project sets a smaller number of epochs, such as only 3, is there limited room for acceleration?
3. In theory, can the runner iter method increase the performance?

[mrazhou]

Thanks for your interest and questions!

1. We have proposed a new method that facilitates easier integration of per-sample loss calculation into highly encapsulated code frameworks. This work is currently under review, and we plan to open-source the code upon acceptance.
2. That's an excellent question, and it points to a current limitation of dynamic data pruning methods – their reliance on multi-epoch training. We are actively researching solutions to this issue and already have some promising approaches. We will open-source this work once it's accepted for publication.
3. Theoretically, yes, modifying the runner's iter method could potentially increase performance.