search

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
1.28k stars 228 forks source link

Doubt about dis_opt and gen_opt #68

Open songbo0925 opened 3 years ago

songbo0925 commented 3 years ago

In trainer.py, why only update the parameters of dis_a and gen_a and ignore the parameters of dis_b and gen_b? https://github.com/NVlabs/DG-Net/blob/a067be117b43c7c553275b6570b3a3bf8da465e0/trainer.py#L242-L248

layumi commented 3 years ago

Hi @songbo0925 Thanks for your attention on our paper. The weight is shared between dis_a and dis_b, so as gen_a and gen_b. Therefore, only one optimizer is needed.

songbo0925 commented 3 years ago

Hi @layumi Thanks for your wonderful work and reply. In line 180-181 of trainer.py, gen_b is set to be the same as gen_a just this once. But in each iterations, if just the parameters gen_a is updated, then in next forward gen_b and gen_a may have different parameters. So I want to know how you achieve weight sharing？So as dis_a and dis_b. Maybe I ignored some code, please advise. https://github.com/NVlabs/DG-Net/blob/a067be117b43c7c553275b6570b3a3bf8da465e0/trainer.py#L180-L181