bilylee / dcq Goto Github PK

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。

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？

Hi, nice work about face recognition! I'm only confused about the following:

In Section 3.1. Preliminaries Classification-based Representation Learning., as the paper says, "During evaluation, the FC layer is removed and only the feature extractor is used."

But in the code

It uses the weight generator network(encoder_k) to get the feature(or weight?) for evaluation; I am very puzzled here. Is there something that I missed?

运行DCQ错误，具体如图所示，但是DCQ_train_test_data/MS1MV2/ms1mv2.json我是有的

1. 想复现一下
2. 这篇论文的方法和 Partial FC: Training 10 Million Identities on a Single Machine 有什么异同吗？@anxiangsir @bilylee