onedroid

Results 2 issues of 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。

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...

shitpost
meant-to-be-used
you-are-stupid-argument