A question about setting f=f.detach() in ft_netAB

NVlabs / DG-Net

:couple: Joint Discriminative and Generative Learning for Person Re-identification. CVPR'19 (Oral) :couple:

https://www.zdzheng.xyz/publication/Joint-di2019

Other

1.27k stars 230 forks source link

A question about setting f=f.detach() in ft_netAB #59

Open RainbowShirlley opened 3 years ago

RainbowShirlley commented 3 years ago

Hi, thank you for your code. I have a question about the code 'f=f.detach()' in ft_netAB of reIDmodel.py. Since f set to to be detached, then loss(4) in your paper will have two appearance codes which are both inable to backward，thus the loss with be useless. Moreover, since f is used to reconstruct images, then losses for image reconstruction will also have no effect for Ea's updating. I'm so confused, thank you for your kind explanation!

RainbowShirlley commented 3 years ago

Hello, can anybody help me with the question above? I'm really confused, and hope for rational explanation, thanks!