Replicate results

[videturfortuna]

Hi,

I attempted to train TransREID with your weights on the original config of Veri-776. Yet using the referred changes in paper, namely lr and batch size I only achieved:
mAP: 83.4%
CMC1: 96.7%

Could you share the exact config used to train TransREID in order to achieve the results published in the paper.
Best regards

[Vehicle-AHU]

vehicelmae_transreid_120ep.log

[Vehicle-AHU]

You can set parameters according to our log file. The best results are at 102 epoch.

[videturfortuna]

In the code you have default 300 epochs to train and in the paper it is written 100. Would you mind sharing the logs of training VehicleMAE in AutoBut1M to confirm this. I'm also having issues with the structural prior with losses not decreasing, I do not know if this is the normal behavior in the begining but would love to see your training log with the loss curves.

Thanks in advance.

[videturfortuna]

I think I found why structural prior is not training. EMA does not update any weights since names_common is an empty list. Can you review this?