Skip to content

[Machine Learning] LayerNorm 層歸一化筆記

LayerNorm 的工作原理如下：

  1. 計算均值（mean）和方差（variance）
mean = \mu =\frac{\sum_{i=1}^{N}x_i}{N} \newline variance = \sigma^2 = \frac{\sum_{i=1}^{N}(x_{i}-\mu)^2}{N}


  1. 使用均值與方差進行標準化
\widehat{x}=\frac{x_i-\mu}{\sqrt{\sigma^2+\epsilon}}

 
ϵ 是一個很小的數字，用來避免分母為 0。

  1. 線性變換

最後，會使用可學習的參數：比例參數 γ 和偏移參數 β（這兩個參數通過訓練得來，初始化 γ = 1 而 β = 0）來對每個輸入元素進行線性變換。

y_i=\gamma\widehat{x_i}+\beta


這種作法是希望模型在一開始不對標準化後的輸出進行任何調整，逐步讓模型自己學習找到最適合的調整參數。

PyTorch 的簡易實現版本可以看成：

import torch


class MyLayerNorm(torch.nn.Module):
    def __init__(self, normalized_shape, eps: float = 1e-5) -> None:
        super().__init__()
        self.eps = eps
        self.gamma = torch.nn.Parameter(torch.ones(normalized_shape))
        self.beta = torch.nn.Parameter(torch.zeros(normalized_shape))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        mean = x.mean(dim=(-1,), keepdim=True)
        std = x.std(dim=(-1,), keepdim=True, unbiased=False)
        x_normalized = (x - mean) / (std + self.eps)
        return self.gamma * x_normalized + self.beta
    

# Inputs
batch_size = 4
dim = 20
x = torch.rand(batch_size, dim)

# My LayerNorm
my_layer_norm = MyLayerNorm(normalized_shape=20)
x_normalized = my_layer_norm(x)

# Official LayerNorm
official_layer_norm = torch.nn.LayerNorm(dim)
x_normalized_official = official_layer_norm(x)


print("Diff:", x_normalized - x_normalized_official)
print("Max:", torch.max(x_normalized - x_normalized_official))


Output:

Diff: tensor([[-4.6015e-05,  7.2122e-06,  1.6093e-05, -2.7657e-05,  2.4617e-05,
          5.2810e-05, -2.3901e-05, -3.0756e-05,  4.5896e-06, -3.6478e-05,
          4.6492e-05,  5.4479e-05,  9.0301e-06, -7.4059e-06, -2.9624e-05,
         -3.9577e-05, -3.8147e-05,  1.8299e-05,  2.9027e-05,  1.1176e-05],
        [ 1.2219e-05, -1.1958e-06,  1.3947e-05, -1.8477e-05,  3.0756e-05,
         -3.0279e-05, -3.4571e-05, -3.5524e-05,  7.5400e-06,  2.5868e-05,
         -3.3379e-05,  2.2292e-05,  2.7537e-05, -9.3132e-07, -1.7941e-05,
          2.4915e-05, -1.5795e-05, -9.8646e-06,  1.9595e-06,  3.0756e-05],
        [ 2.5034e-05, -5.2810e-05,  2.6703e-05,  4.8757e-05,  2.6405e-05,
         -8.8215e-06,  2.7299e-05, -1.7226e-05,  1.4827e-06,  2.6643e-05,
          9.3132e-09, -5.6386e-05,  2.2948e-05, -1.4678e-06,  9.4771e-06,
         -3.9697e-05, -3.8743e-05,  3.2425e-05, -1.5974e-05, -2.0623e-05],
        [-5.6624e-05, -1.9819e-05,  5.9962e-05,  3.7998e-06,  7.2598e-05,
         -7.7128e-05,  1.0788e-05,  4.2319e-05,  2.0355e-05,  1.6630e-05,
          7.0691e-05, -3.1710e-05, -5.1022e-05, -4.5538e-05, -1.3143e-05,
          2.3484e-05, -1.4067e-05, -1.2577e-05, -3.9399e-05,  3.8266e-05]],
       grad_fn=<SubBackward0>)
Max: tensor(7.2598e-05, grad_fn=<MaxBackward1>)


以上在計算 std 時採用了標準差（Standard Deviation）而非方差（Variance）使其跟原始實現更接近。標準差定義為方差之平方根。

References

Read More

Tags:

Leave a Reply