Excuse me, I don't understand the multi-GPU training model using DP in segmentation_train.py and then using DDP in train_util.py. Is there a conflict?