Apply Deep Mutual Learning in OSNet Training

[Thangbluee]

Many thanks to all the contributors for this project. I try DML project in OSNet training and I get the result showed in table 12d in your paper. I want to get the result showed in table 12f, so how to fix the loss function?

[KaiyangZhou]

You can tune the balancing parameter for the losses. I think it should be one of those:

weight_t=0.5,
weight_x=1.,
weight_ml=1.,

Don't remember exactly. I'm old

[Thangbluee]

Thanks for your reply, I really appreciate that. Those parameters were configured the same way as in the config file (https://github.com/KaiyangZhou/deep-person-reid/blob/master/projects/DML/im_osnet_x1_0_dml_256x128_amsgrad_cosine.yaml) but the result seemed to be the same as the 12d table. In your paper, you said that "We combine the effective training techniques, i.e. DML and auxiliary loss learning (with the triplet loss), and show the results in Table 12f." and λt = 0.5 + DML model, but I don't see where it is in your code. Can you explain why? and Do i need to fix something in ImageDMLEngine to get the result shown in Table 12f