MaybeShewill-CV / attentive-gan-derainnet

Unofficial tensorflow implemention of "Attentive Generative Adversarial Network for Raindrop Removal from A Single Image (CVPR 2018) " model https://maybeshewill-cv.github.io/attentive-gan-derainnet/
MIT License
258 stars 83 forks source link

归一化层 #70

Closed Carlyx closed 4 years ago

Carlyx commented 4 years ago

您好,我目前在研究反射去除问题,其中同样涉及到了normalization层,我发现在conv和ReLU之间加入BN层会较明显的降低性能(batch_size=1),所以目前都是不加入任何normalization层

我看到您在之前的issue中提到图像修复任务中普遍认为归一化会影响结果。请问应该怎么解释这个现象呢,主要原因是batch_size过小吗? 凭借您的经验,如果加入GN层,是否会比不加任何归一化层效果更好呢? @MaybeShewill-CV

MaybeShewill-CV commented 4 years ago

@Carlyx batch size很小的情况下使用norm很影响模型的结果。使用gn替换bn效果没什么改善:)

Carlyx commented 4 years ago

好的,感谢回复~