ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders论文解读
论文地址:https://arxiv.org/abs/2301.00808
论文相对ConvNeXt V1的改进主要两点
- 全卷积掩码自编码器(FCMAE)
- Global Response Normalization,去除Layer scale
什么是layer scale
实验
先用FCMAE预训练,然后微调
论文相对ConvNeXt V1的改进主要两点
先用FCMAE预训练,然后微调