Why dose the "loss" reduce first and then increase during the training?

[Yongkai-Tian]

Hi, thank for sharing for the code!
I successfully run it on the samples images with GPU using your pretrained model. Now i'm trying to retrain the models using your code. But the "loss" reduced first and the increased in training, can you explain it?

[xk19971206]

嗨，谢谢分享代码！
我使用您的预训练模型使用GPU在样本图像上成功运行它。现在我正在尝试使用您的代码重新训练模型。但是“损失”首先减少，培训增加，你能解释一下吗？

老哥，请问你的pkl文件怎么打开的啊，我通过教程还是打不开