the switchable normalization method by keras
- the paper Differentiable Learning-to-Normalize via Switchable Normalization introduce the Switchable Normalization(SN).
- the blog 深度剖析 | 可微分学习的自适配归一化 explain SN method.
- the SN adapt to weight 3 different normalization method, including IN/LN/BN. It need 3 trainable mean weights and 3 trainable variance weight to weighted average means and variances respectively.So the trainable parameters of SN equal to 2*channels+6.
- the detail of experiment log is in directory experiments
- compare 3 normalization method: batch_norm(bn) // group_norm(gn) // switchable_norm(sn)
- data: ISBI 2D EM segmentation images
- network: Unet(based on VGG)
- epoch: 5
- environment:GeForce 1080Ti
- consequent:
- training time: sn > gn > bn
- segmentation result: sn > gn > bn