the mAP on cifar10 is only 0.30 about hashnet

[dweiiiiiiiiii]

你好，请问你也只是在cifar-10上得到的结果不太好而在其他数据集上结果正常吗？

[chenshen03]

@dweiiiiiiiiii 是的，应该和你结果差不多。

[chenshen03]

@caozhangjie Can you explain why HashNet behaves so badly on CIFAR10?

[caozhangjie]

Cifar 10 has 10 classes, I recommend replace the class_num to 10. Also, Cifar 10 needs a different backbone network. Please search for a proper backbone.

[zhouxiaohang]

Cifar 10 has 10 classes, I recommend replace the class_num to 10. Also, Cifar 10 needs a different backbone network. Please search for a proper backbone.

Since CIFAR 10 with 10 classes needs class_num to be 10, why COCO dataset with 80 classes is compatible with class_num=1.0? I've tried (64 bits are used.):

COCO dataset with class_num=80.0, mAP=0.53
COCO dataset with class_num=1.0, mAP=0.80 (which is higher than that in paper)

How to explain this?