请问开始训练的时候学习率会特别大么？

songdejia / Siamese-RPN-pytorch

This is a re-implementation of Siamese-RPN with pytorch, which is CVPR2018 spotlight.

MIT License

381 stars 88 forks source link

Closed JingLi513 closed 5 years ago

JingLi513 commented 5 years ago

我最近也在实现这个算法，用了torchvision的alexnet，新添加的卷积层Xavier初始化。用sgd优化，按论文中的学习率刚开始训练就梯度爆炸了，不知道你有没有遇到这样的情况