DCQ icon indicating copy to clipboard operation
DCQ copied to clipboard

[CVPR21] Dynamic Class Queue for Large Scale Face Recognition in the Wild

Results 4 DCQ issues
Sort by recently updated
recently updated
newest added

Thanks for your great work and released code. Should "q.view(N, C)" be revised to "t.view(N, C)" ?

在DCQ的前向中,对im_k做了一次shuff(对应函数_batch_shuffle_ddp),然后又做了一次unshuff(对应函数_batch_unshuffle_ddp),我认为得到的结果和直接对im_k做前向的结果是一样的。这样做的目的是什么?

![image](https://user-images.githubusercontent.com/78945582/134609516-151b9011-ae4a-4aed-bd98-43120b7c653b.png) 运行DCQ错误,具体如图所示,但是DCQ_train_test_data/MS1MV2/ms1mv2.json我是有的 ![image](https://user-images.githubusercontent.com/78945582/134609631-416fb970-698f-4217-9717-ea7260d64bcd.png)

我觉得这个方法中的weight类似于特征,类似于说今天维护了一个特征的队列,然后加上了triplet loss去做,也就是没有weight。我认为本质上是用triplet loss来训练人脸识别,然后加上momentum。