MoreLoRA

Morelora

Lora original:

$ W = W_0 + UV $ y $ rank (uv) leq r $

Mejor inicialización:

$ W = w_0 - u_0 {v_0} + UV $

Aditivo lora:

$ W = w_0 + ui_ {r (1 times frac {n} {r})} + i_ {r ( frac {m} {r} times 1)} v $ dónde $ U in mathbb {r}^{m times r}, v in { mathbb {r}^{r times n}} $ y $ rank (uv) leq 2r $

Hadamard Mul Lora:

$ W = w_0 + odot_ {i = 1}^{i = k} ( delta_i) $ dónde $ Delta_i = u_iv_i $

$ r '= frac {r} {k}, u_i in mathbb {r}^{m times r'} $ , $ V_i in mathbb {r}^{r ' Times n} $ y $ rank ( odot_ {i = 1}^{i = k} ( delta_i^t)) leq ( frac {r} {k})^k $

Hadamard Agregar lora:

$ W = w_0 + odot_ {i = 1}^{i = k} ( delta_i) $ dónde $ Delta_i = u_ii_ {r '(1 Times frac {n} {r'})}+i_ {r '( frac {m} {r'} times 1)} v_i $

$ r '= frac {r} {k} $ , $ U_i in mathbb {r}^{r ' Times n} $ , $ V_i in mathbb {r}^{m times r '} $ y $ rank ( odot_ {i = 1}^{i = k} ( delta_i)) leq ( frac {2r} {k})^k $

Hadamard Lora: activación

$ Delta = odot_ {i = 1}^{i = k} ( tanh (u_iv_i^t)) $

$ Delta = odot_ {i = 1}^{i = k} ( sigma (u_iv_i^t)) $

Dylora:

Actualizar al azar una serie de rangos

Actualizar parte de las capas

Referencia:

 @online { kexuefm-9590 ,
    title = {梯度视角下的LoRA：简介、分析、猜测及推广} ,
    author = {苏剑林} ,
    year = { 2023 } ,
    month = { Apr } ,
    url = { url{https://spaces.ac.cn/archives/9590} } ,
}

 @misc { hyeonwoo2023fedpara ,
      title = { FedPara: Low-Rank Hadamard Product for Communication-Efficient Federated Learning } , 
      author = { Nam Hyeon-Woo and Moon Ye-Bin and Tae-Hyun Oh } ,
      year = { 2023 } ,
      eprint = { 2108.06098 } ,
      archivePrefix = { arXiv } ,
      primaryClass = { cs.LG }
}