Question

I m training a set of images using MobileNet that would do multiclass classification. After training loss-epoch curve looks like this:

我不敢肯定,如果这种解释足够好/可以接受,如何解释? 但是,最终会发生一些激烈的争执,因此我不认为是好的? 如何改进?

Answer 1

It is a common phenomenon for a peak to appear when the model converges. This issue can be observed in many CNN or transformer-based classification models, such as the figure in the "ResNet" Paper.

https://i.stack.imgur.com/HrH7f.png”rel=“nofollow noreferer”>ResNet 损失曲线

各位还可能注意到,许多损失曲线似乎非常顺利;这是因为它们采用平均移动方式来提供更平稳的损失曲线。

因此,我建议为更多的人提供培训,以观察损失曲线的变化或选择较小的学习率。此外,你可以利用检查站,以弥补最低验证损失。

友情链接