xbq1994 / meta Goto Github PK

我从您论文摘要所给出的github链接下载了META工程，用四张显存各为12G的2080Ti,按照您配置文件中默认的参数，BASE_LR: 0.04，IMS_PER_BATCH: 64,训练时回报错“FloatingPointError: Loss became infinite or NaN at iteration=510!”，查阅资料后好像时是学习率有点大。我就根据您论文实验部分的参数设置，设置BASE_LR: 0.0003，IMS_PER_BATCH: 64，这样能训练完，但是rank1为39.16，mAP为14.83，达不到论文中的性能。请问是我训练过程有问题吗，能否分享一下您的超参数设置呢？

When loading the CUHK-SYSU dataset, the error "cuhk_sysu/roped_images is not found" appears. Do you know how to solve this problem?

Thanks for your great work! The default setting in your code is protocol 3. How should you set the protocol1 and protocol 2?

Thanks for your great work! But I'm confused about that whether the backbone is frozen in the training procedure?