Can't reach the results on the paper

[ccnankai]

Hi, @yangxue0827 I used my own data set for training, the test results are not good. I have trained 60,000 steps, and The loss function stops and does not reduced( stop at 0.9 ). Can you give some Suggestions for improvement to get better results？Which parameters need to be changed？

Thank you.

@yangJirui

[shihan19911126]

我也上遇到这样的情况，loss降到0.8在有就不再下降,请问你的问题解决了吗？有什么好的建议？

[ccnankai]

@shihan19911126 不知道怎么解决，可能根本就不行。你还有其他的好算法吗？

[shihan19911126]

@ccnankai 我看作者生成的total_loss图也是0.8左右，我用的resnet50可以降到0.73左右，效果还是resnet101好，虽然resnet101 loss高一些

[yangxue0827]

loss的大小并不是衡量结果好坏的标准。 @ccnankai @shihan19911126

[ccnankai]

@yangxue0827 ok，我主要在学习代码，多谢多谢。