ontheway361 / dul-pytorch Goto Github PK

您好，我使用了分类的dul在webface上进行训练，将batchsize和lr由512和0.1调整为128和0.025，在Lfw上进行测试，但是acc只能达到83%，请问这是什么原因呢

Hi, could you share me the training and tesing dataset?

Wonderful job！I studied your paper these days, which is very enlightening to me. I try to use the data uncertainty in my research.
I have a question about the caculation of the variance. In your code, you use the
var = F.normalize(torch.exp(self.var_head(x))), to process the variance predicted by network. I guess the torch.exp() is used to guarantee the positive of the variance. The F.normalize() I guess is simliar to the L2-normalization for the feature to make the network converge better.
I do not know this understanding is right? Thank you.

Nice project! Any results reported based on your code? Thanks!

Hi, it's really a nice job.
After reading this code, i have a question about the training stage. In line 118~120 of train.py file, if in training stage, the feat_mu represents the sample s(i). But when calculating the loss on line 120, the third parameter of criterion function needs the mu(i), not the sample s(i) named feat_mu.
I wish i made myself clear. Looking forward to your reply, thanks.

Hi, your paper is really great and I am trying to understand the code a bit more. When two images are passed through the network, The network should be producing two wmbeddings right? What are the steps after that? I am a bit confused in this part.

你好，请问一下ms1m_images.txt的内容是什么样的？

logvar? mean(logvar)? norm(logvar)?

Hi, I am learning your paper, and in your released code of faceloss.py. I found some parameters/function could not be found for hardmining.
line 29: label
line 31: batch_size
line 32: ind_sorted
line 33: pred, label

I don't find the gamma and beta at the step in the paper.Why add the gamma and beta here.

你好，我用你的代码在ms1m数据集上训练两个版本，一个kl_lambda设置为0，一个为0.01，发现两种设置对测试精度的影响不大，请问您给一些建议吗 @jielu361

您好，请问ms1m_images.txt的每一行是什么样的结构呢？能否分享一下根据数据集生成.txt文件的代码呢？

看论文里面是这样，是不是需要在model里设置eval()时仅输出mu，不管预测的方差？

Thanks for your code on classificition-based DUL.
And when will you update the code about regression-based DUL.

Hi!

I've been reading through your code base and I got the following question, where is fc_weights parameter is initiated?

I suspect that it should be center attribute. Could you please comment on that?

I have download the dataset'CASIA-webface', but there is no file called 'casia_landmark.txt'

Hi, a really nice repo.
After reading your code, I only found Classification-based DUL for FR , but not Regression-based DUL for FR. do you plan to share this part? thanks a lot.

Nice paper and Nice code! but am confused about how to convert the σ(x) embedding to a std scalar?

请问回归形式的损失相当于

log似然没有上下界，该怎么保证损失收敛呢。
实际上我在用类似损失函数时，得到的损失曲线是

I have used this code for training. The arcface can converge normally, but the kl loss does not decrease. And the trained var-branch has no effect.