Attention Is All You Need

YoungTimes

发布于 2023-09-01 08:57:00

2260

发布于 2023-09-01 08:57:00

Attention解决了类RNN的长时序依赖问题和计算的并行化的问题，Multi-Head Attention实现了类似RNN多通道的效果。Transformer的整体架构是如下，Encoder和Decoder都使用了Stacked Self-Attention And Point-wise, Fully Connected Layers(MLP)结构。

Encoder

Encoder包含6个Layer，每个Layer包含2个Sub-Layer，Sub-Layer-1是Multi-Head Self-Attention，Sub_Layer-2是MLP，每个Sub-Layer中还包含了Residual Connection和Layer Normalization。

Decoder

Decoder包含6个Layer，每个Layer包含3个Sub-Layer，Sub-Layer-1是Masked Multi-Head Self-Attention，Mask是为了保证【Predictions for position

can depend only on the known outputs at positions less than

】；Sub-Layer-2是Multi-Head Self Attention；Sub-Layer-3是MLP，每个Sub-Layer中还包含了Residual Connection和Layer Normalization。

Layer Normalization

在机器翻译领域，输入的Seq长度往往是不同的，因此相比于在多个样本之间做Normalization的Batch Normalization，在单个样本内部做Layer Normalization是一种更好的策略。

Scaled Dot-Product Attention

Scaled Dot-Product Attention

Attention本质是一种加权和机制，它的计算公式如下：

Attention(Q, K, V) = softmax(\frac{QK}{\sqrt{d_k}}V)

用Q与K的内积衡量二者相似度原理是什么？

来源:https://charlieleee.github.io/post/cmc/

向量的相似性是用两者的角度余弦来度量，角度越小，余弦值愈大，两者愈相似。

为什么要除以

\sqrt{d_k}

“While for small values of

d_k

the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of

d_k

，We suspect that for large values of

d_k

, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients.

简单的说，在Dot Product Attention当

d_k

比较大的时候，点乘的值可能比较大或者比较小，导致向量间相对的差值变大，经过softmax之后，大的值向1靠近，小的值向0靠近，使得网络的梯度变得很小，难以训练。Transformer中的

d_k

一般比较大，所以除以

\sqrt{d_k}

是不错的选择。

Mask是什么

Mask主要避免网络提前看到当前时刻之后的输入。

Multi-Head Attention

Multi-Head Attention

Multi-Head Attention中，Q、K、V先经过Linear层，类似CNN的多Channel机制，给网络h次投影学习机会，让其有机会学到更好的投影。

MultiHead(Q, K, V ) = Concat(head_1, ..., head_h)W^O

其中:

head_i = Attention(QW_i^Q, KW_i^K, VW_i^V)

“In this work we employ h = 8 parallel attention layers, or heads. For each of these we use

d_k = d_v = d_{model} / h = 64

. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.

Embedding & Positional Encoding

Input Embedding将输入单词映射为维度d_{model}的向量，这里将Embedding Weight除以

\sqrt{d_{model}}

，保证Input Embedding与Positional Encoding的Scale基本一致。

Position Encoding采用相对位置编码方式：

PE(pos,2i) = sin(pos/10000^{2i/d_{model}})

PE(pos,2i+1) = cos(pos/10000^{2i/d_{model}})

“where pos is the position and i is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from 2π to 10000 · 2π. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset k,

PE_{pos+k}

can be represented as a linear function of

PE_{pos}

参考材料

1.https://www.zhihu.com/question/592626839/answer/2965200007?utm_campaign=shareopn&utm_medium=social&utm_oi=703332645056049152&utm_psn=1630928281645072385

2.[跟李沐读论文] https://www.youtube.com/watch?v=nzqlFIcCSWQ

本文参与腾讯云自媒体同步曝光计划，分享自微信公众号。

原始发表：2023-05-14，如有侵权请联系 cloudcommunity@tencent.com 删除

机器翻译