seathiefwang / MGN-pytorch

Reproduction of paper: Learning Discriminative Features with Multiple Granularities for Person Re-Identification
285 stars 108 forks source link

我有20W的类别标签,每个类别有2张图片以上,训练20 Epoch发现CrossEntropyLoss不下降 #17

Open xf739645524 opened 5 years ago

xf739645524 commented 5 years ago

我有20W的类别标签,每个类别有2张图片以上,训练20 Epoch发现CrossEntropyLoss不下降

lulujianjie commented 5 years ago

类别数太多了,可以用arcface https://github.com/lulujianjie/person-reid-tiny-baseline 试一下

guanyonglai commented 4 years ago

类别数太多了,可以用arcface https://github.com/lulujianjie/person-reid-tiny-baseline 试一下

我的才1W类别标签,训的结果都不如5000的,打算用下你的代码试试,但是arcface和CrossEntropyLoss不是差不多吗