Hao Luo

The network initialization has a great influence on MSML, which means an inappropriate initialization may results in NAN. I always trained the model several epochs with softmax loss to initialize...

@ergysr It may depend on the datasets. Without another/softmax loss, i successfully trained the model on Market1501, but failed on CUHK03. I thought that CUHK03 has two images for each...

@XrsSjtu 是通过开源数据集大杂烩随便训出来

@M-Ghorbel You can try the VPN to connect Chinese network, and I will upload the models to Google Drive at once.

@superCman 75%左右吧

@ericxian1997 和数据集有关，大概一天多一点时间能收敛，有点忘记需要多少个epoch了，我是一直训练存checkpoint手动中断的，然后根据Tensorboard挑checkpoint的

Thank you very much. If you have check the code, you can commit a PR and become a contributor for this project.

这个应该是engine设置的问题，可以尝试不用engine的版本 https://github.com/lulujianjie/person-reid-tiny-baseline

Hi，veri776可以参考这个repo https://github.com/damo-cv/TransReID

This repo only supports single GPU. You can refer to https://github.com/DTennant/reid_baseline_with_syncbn.