tiandunx / ffc Goto Github PK

您好，感谢您的工作。有些疑问希望可以得到您的解答：

1. 原文中阐述了instance-based的两个问题，但是第二个问题是否已经被CVPR 2021 DCQ那篇文章解决？第一个问题在DCQ中仍然存在，但是改进了这个问题好像结果并没有比DCQ好？
2. 原文中阐述的这两个问题，是如何影响实验结果呢？更具体来说，您定义的更新速度这个变量会影响最终识别的结果吗，更新快意味着结果好吗？【1. If the number of identities is severely imbalanced, the update speed of the identities’ centers that have rare number of images is too slow. 2. If we sample M images that belong to M different identities, the DCP may have no positive samples for this iteration】

您好，想问一下您在训练WF42M数据集的具体设置是什么样的呢？

我在训练WF42M的时候，loss下降到70多就很难再下降，学习率调小100倍，也是下降到30多就很难再下降，能否公开一下您训练WF42M的具体设置呢？非常感谢~

Indeed, I wish FFC can solve my problem of GPU memory limitation without damaging the model performance.
so can you describe more details about the difference/advantage of FFC comparing with DCQ?
thanks in advance !

can i get the pretrained weight ?

您好，我有几个疑问需要请教，期待您的回复：

1. 论文中吞吐量的计算，包含了送入g net中的图像吗？
2. 论文中显存占用的对比，是在相同batchsize下吗？该batchsize是否包括送入g net中的数据？
3. 收敛速度和PFC相比如何？我很好奇在训练初期收敛情况是怎么样的？

May I ask where the data enhancement method mentioned in README.md is reflected? Could you help me to guide Xiao Bai
thank you

if world_size=8 then error: argument world_size: invalid int value

for one server which have 8 gpu cart, the value of "world_size" can be setted as what?

thanks

Dear @tiandunx

Thank you for your work. In the Virtual FC paper at CVPR 2021, we have to change the network HEAD and replace it with the Virtual FC layer which takes two parameters in_feat and out_feat. Is your work like them? because I saw your code and didn't get any clear definition. If possible for you, please mention some small examples, because I want to use your layer in my network.

Regards,
Khawar

in Main.py line 110 : optim_config = load_config('config/optim_config')
there is no load_config defined