NAFNet icon indicating copy to clipboard operation
NAFNet copied to clipboard

Dear author,I want to why not use weight decay to train this model?

Open BrotherHappy opened this issue 3 years ago • 1 comments

想知道为什么不使用weight-decay来训练模型,这样不会导致过拟合嘛?

BrotherHappy avatar Apr 29 '22 10:04 BrotherHappy

你好, BrotherHappy, 这在图像恢复模型的训练中属于比较常规的操作,如MPRNet的训练: https://github.com/swz30/MPRNet/blob/main/Deblurring/train.py#L59

而通过观察 train loss / val PSNR, 虽然train loss 在下降, 但是val PSNR并没有出现明显下降的情况,所以可以说并不明显存在传统意义上的过拟合。

mayorx avatar Apr 29 '22 17:04 mayorx