onedroid

如果我没理解错的话，CycleMLP 等价于分组shift+ channel shuffle +mlp mlp对前一层的通道有全局依赖，所以channel shuffle是没有必要的，所以cyclemlp 不需要cycle 直接实现为 分组shift+mlp速度会更快，与原来的cyclemlp的差异可以用初始化的方式对齐。

什么智商觉得自己能在技术网站搞这个。 这是技术网站，不是你搞政治宣传的地方，什么牛鬼蛇神啊。在脸书，推特搞不行么，真特么傻叉。What kind of intelligence do you think you can do this on a technical website? This is a technical website, not a place for you to engage in political propaganda. What a monster. Isn’t it possible to do this on Facebook or Twitter? It’s really stupid. https://cirosantilli.com/china-dictatorship/meant-to-be-used

1

1) https://cirosantilli.com/china-dictatorship/meant-to-be-used 1) https://cirosantilli.com/china-dictatorship/brainwashed-by-usa 1) https://cirosantilli.com/china-dictatorship/zhao-heming 1) https://cirosantilli.com/china-dictatorship/stability 1) https://cirosantilli.com/china-dictatorship/is-ciro-anti-communism 1) https://cirosantilli.com/china-dictatorship/most-chinese-people-like-their-dictatorship 1) https://cirosantilli.com/china-dictatorship/censorship 1) https://cirosantilli.com/china-dictatorship/china-has-more-freedom-of-speech-than-the-usa 1) https://cirosantilli.com/china-dictatorship/censorship-circumvention 1) https://cirosantilli.com/china-dictatorship/internal-censorship 1) https://cirosantilli.com/china-dictatorship/github-report 1) https://cirosantilli.com/china-dictatorship/nine-nine-six-icu 1) https://cirosantilli.com/china-dictatorship/i-like-my-dictatorship 1) https://cirosantilli.com/china-dictatorship/funded-by-cia 1) https://cirosantilli.com/china-dictatorship/gfw-to-protect-from-usa...

SE模块可以用Conv2d或者Linear两种方式实现，前者可以少一些view操作。mobilenetv3采用的是Linear，ghostnet使用的是Conv2d，view操作的开销应该不大，但是这两种实现在延时上是有差别的。建议对比这两种模型可以保持相同的实现方式测试延时，避免混入对比变量。

https://github.com/microsoft/X-Decoder/blob/165f8a6314ac84f5c36aaab7216f90dd97e38a43/modeling/modules/criterion.py#L469 As I can not find this loss function in mask2former or maskformer repo,should I remove this loss function if I do not need language encoder part and make x-decoder...