Zhedong Zheng
Zhedong Zheng
Hi @YJYJLee I found that xformer `MultiHeadDispatch` has supported `key_padding_mask`. But I am not sure how to set the attention for forward. https://facebookresearch.github.io/xformers/components/attentions.html?highlight=attention#module-xformers.components.attention
Dear @absorbguo , I am very sorry about that. Due to the privacy concern from Duke University, DukeMTMC-reID has to be offline.
Hi @kawaiitakagi Could you provide me more details? or try to use another machine?
Thank you @kawaiitakagi I think I have fixed it by https://github.com/layumi/University1652-Baseline/blob/master/folder.py#L102 The issue still exists? Do you find any new solution?
Hi @Noixas Sorry for the late response. 1. Swin transformer needs a fixed input size to ensure the window number. 2. You may use ResNet + Verif-Identif or Swin +...
Hi @BaranYucecengiz Yes. You can combine the reID with tracking algorithm. reID is to extract feature. For instance, you could use human detector first and use the pretrained model in...
Hi @BaranYucecengiz You may just forward once as in the test.py. Pls remember to use Yolo or SSD to detect person first.
Hi @Pherun123 It seems to be the ``torch.compile()'' , which uses dynamo to optimize the graph and accelerate the training speed. https://github.com/layumi/Person_reID_baseline_pytorch/blob/master/train.py#L512 You may consider to remove it in the...
@Stark897 你好。 不是按ID来的，是按图像的ID顺序排列的。 有些ID会有多个query。
Thank you @SIYUAN9446 You are right. Sorry for the mistake. I have updated the README.