Skip to content

Potential BUG in the coe & performance reproducibility issues #7

@Sissuire

Description

@Sissuire

您好!非常nice的工作。
我在实践中发现了两个问题,可以商讨下:
(1)网络模型中,这里似乎用到了包含batch在内的全局归一化,这会导致在测试时,测试batch数量不同,测试结果不同。
(2)鉴于mmcv-full 1.4.0需要比较早期的cuda和pytorch,我没有完全复刻您的环境。即便如此,我直接采用您给出的预训练模型进行测试,发现性能似乎也有较大出入,IRSTD-1K的mIoU只有0.675左右。当然,这可能跟上述问题(1)有一定关系。想问下,您是否在其他机器上测过公布的代码和模型性能。

祝好!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions