I have train almost 80thousand examples within 2000 labels，valid acc almost 92%，but test result all example prob is blew 0.01.

[w279805299]

I have trained almost 80thousand examples within 2000 labels，valid acc almost 92%，but test result all example prob is blew 0.01. I have tried tranning examples to predict.

[qiu-nian]

I have trained almost 80thousand examples within 2000 labels，valid acc almost 92%，but test result all example prob is blew 0.01. I have tried tranning examples to predict.

hello, have you solved this problem?

[bimhud]

When having a large number of labels and the number of actual labels for each sample is small, the loss function needs to be re-considered, otherwise the accuracy is unreliable

[sunxiaowu]

I have trained almost 80thousand examples within 2000 labels，valid acc almost 92%，but test result all example prob is blew 0.01. I have tried tranning examples to predict.

hello have you solved this problem? I meet this problem too

[sunxiaowu]

When having a large number of labels and the number of actual labels for each sample is small, the loss function needs to be re-considered, otherwise the accuracy is unreliable

so I should change the loss function? should I modify the method of accuracy ?

[xiaoyichao]

我也遇到了同样的问题，实际的分类效果非常差

[wl915634785]

是的，我也遇到了，66个类别，准确率不能反映真实情况