Hi, great work and thanks for sharing the code. May I ask whether should we halve the training epoch concerning DDGSD coming to a double batch size by different DA.