Possible minor issue in process of updating environment

[jinyery]

Hello, this is really a great job and presents a new challenge. But I think there is a small logical flaw in the implementation of the code, specifically in the function "generate_intra_weight()" of the file "train_center_dual. py", because "intra_weight=torch.zeros (total_image).fill_ (0.0)" and "intra_weight[ind]=score" result in partial weights always being zero in the DistributionSampler.

The adjustment of weights is calculated based on the data sampled from the previous epoch. The sampling of DistributionSampler is replacement, so it is highly likely that not all samples will be sampled in the previous epoch. This results in the weight of minority samples in the new environment always being zero after update_env_by_score().

The following modifications may solve this problem: