Transformer 研究意義:

1、提出了self-attention，拉開了非序列化模型的序幕

2、為預訓練模型的到來打下了堅實的基礎

序列化模型主導(LSTM) <----- 2017 -----> 提出新的attention方式，實現了非序列化的模型并行化，提高效率(self-attention)

?

這篇文章提出來主要是解決機器翻譯問題，機器翻譯指標(BLEU)提升兩個點

機器翻譯指標(BLEU):

Candidate: the the the the the the the

Reference1: the cat is on the mat

Reference2: there is a cat on the mat

在Candidate里，1-gram指的是一個the，2-gram指的是兩個the,.....，相當于一個滑動窗口(N-gram，N相當于一個窗口大小)

count(the) = 7

count1-R(the) = min(7,2) = 2

count2-R(the) = min(7,1) = 1

count-clip(the) = max(2,1) = 2

p1 = count-clip(the)/count(the) = 2/7

只使用1-gram的問題是對每個詞進行翻譯就能得到很高的得分，完全沒有考慮句子的流利性

解決方法: 使用多-gram融合，bleu使用4gram

依據上面計算方法，分別計算出p2 = 0, p3=0, p4=0

P = 0.25*p1 + 0.25*p2 + 0.25*p3 + 0.25*p4

還存在的問題就是長度問題，目前的評價指標對短句有利

解決方法: 對長度增加懲罰

BP = [1 if c>r else e^(1-r/c)]

Then

BLEU = BPexp(sumWnlogPn)

?

論文主要結構:

一、Abstract

? ? ? ? ? 1、常用的序列模型都是基于卷積神經網絡或者循環神經網絡，表現最好的模型也是基于encode-decoder框架上添加attention機制

? ? ? ? ? 2、提出一種基于attention機制的新模型transformer，拋棄了傳統的模型結構

? ? ? ? ? 3、模型在2014WMT翻譯數據集上，比現存的模型的bleu值高兩個點

二、Introduction

? ? ? ? ? 主要論述序列化模型的弊端-長度依賴問題，并提出本文模型

三、The Background

? ? ? ? ? 主要論述傳統卷積模型存在學習遠距離依賴困難的弊端，并提出注意力機制

四、Architecture

? ? ? ? ? 主要論述transformer網絡結構及其內部細節(Scaled Dot-Product Attention、Multi-Head Attention、Positional Encoding)? ? ? ? ? ?

? ? ? ? ? 整體概覽模型結構:

? ? ? ? ? ?

?

主要將的是模型主要為編碼-解碼結構，編碼是將一個特征表示序列x=(x1.....xn)編碼成一個連續的序列表示z=(z1,.....zn), 解碼部分會對z進行解碼得到y=(y1,.....yn)，模型整體結構如下：

?

模型整體結構部分代碼片段class EncoderDecoder(nn.Module):"""A standard Encoder-Decoder architecture. Base for this and many other models."""# super(EncoderDecoder, self).__init__() 等價于 nn.Module.__init__()def __init__(self, encoder, decoder, src_embed, tgt_embed, generator):super(EncoderDecoder, self).__init__()self.encoder = encoderself.decoder = decoderself.src_embed = src_embedself.tgt_embed = tgt_embedself.generator = generator# forward函數調用自身encode方法實現encoder，然后調用decode方式實現decoderdef forward(self, src, tgt, src_mask, tgt_mask):#src的shape=tgt的shape：[batch_size,max_length]"Take in and process masked src and target sequences."return self.decode(self.encode(src, src_mask), src_mask,tgt, tgt_mask)def encode(self, src, src_mask):return self.encoder(self.src_embed(src), src_mask)def decode(self, memory, src_mask, tgt, tgt_mask):return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask)class Generator(nn.Module):"Define standard linear + softmax generation step."def __init__(self, d_model, vocab):super(Generator, self).__init__()self.proj = nn.Linear(d_model, vocab)def forward(self, x):#nn.Linear是一個線性層，x*Wreturn F.log_softmax(self.proj(x), dim=-1)

? ? ? 編碼和解碼

?

編碼器

編碼器	1、編碼器由N=6個完全相同的層堆疊而成	?
	2、每一層都有兩個子層	1、Multi-head self-attention 機制
		2、簡單的、位置完全連接的前饋網絡
	3、對每一個子層采用一個殘差連接、層標準化	?

?

""" 編碼器部分代碼片段 """def clones(module, N):"Produce N identical layers."# copy.deepcopy硬拷貝函數return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])#傳入參數 #Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N) class Encoder(nn.Module):"Core encoder is a stack of N layers"def __init__(self, layer, N):super(Encoder, self).__init__()self.layers = clones(layer, N)self.norm = LayerNorm(layer.size)def forward(self, x, mask):"Pass the input (and mask) through each layer in turn."for layer in self.layers:x = layer(x, mask)return self.norm(x)class LayerNorm(nn.Module):"Construct a layernorm module (See citation for details)."def __init__(self, features, eps=1e-6):super(LayerNorm, self).__init__()# features=layer.size=512self.a_2 = nn.Parameter(torch.ones(features))self.b_2 = nn.Parameter(torch.zeros(features))self.eps = epsdef forward(self, x):mean = x.mean(-1, keepdim=True)std = x.std(-1, keepdim=True)return self.a_2 * (x - mean) / (std + self.eps) + self.b_2class SublayerConnection(nn.Module):"""A residual connection followed by a layer norm.Note for code simplicity the norm is first as opposed to last."""def __init__(self, size, dropout):super(SublayerConnection, self).__init__()self.norm = LayerNorm(size)self.dropout = nn.Dropout(dropout)def forward(self, x, sublayer):"Apply residual connection to any sublayer with the same size."return x + self.dropout(sublayer(self.norm(x)))#傳入參數 #EncoderLayer(d_model, c(attn), c(ff), dropout), N class EncoderLayer(nn.Module):"Encoder is made up of self-attn and feed forward (defined below)"def __init__(self, size, self_attn, feed_forward, dropout):super(EncoderLayer, self).__init__()self.self_attn = self_attnself.feed_forward = feed_forwardself.sublayer = clones(SublayerConnection(size, dropout), 2)self.size = sizedef forward(self, x, mask):"Follow Figure 1 (left) for connections."x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))return self.sublayer[1](x, self.feed_forward)

?

解碼器

解碼器	1、解碼器由N=6個完全相同的層堆疊而成
	2、每一層都有三個子層	1、multi-head self-attention機制
		2、簡單的、位置完全連接的前饋網絡
		3、對編碼器的輸出執行multi-head attention
	3、對每一個子層采用個殘差連接、層標準化
	4、修改解碼器堆棧中的self-attention子層以防止位置關注到后面的位置-> mask 結合將輸出嵌入偏移一個位置，確保對位置的預測 i 只能依賴小于 i 的已知輸出

輸入	input embedding -> positional encoding
encoder	for i in range(6): ? ? ? ?self-attention -> layer normalization -> feed forward -> layer normalization
decoder	for i in range(6): ? ? ? ?self-attention -> layer normalization -> encoder-decoder attention -> layer normalization -> feed forward -> layer normalization

?

Batch Normalization 與 Layer Normalization區別(計算方式不一樣，一個是按列batch計算，一個是按行計算)

? ? ? ? ? ? ? ? Batch Normalization							Layer Normalization						mean	std
Batch	1	2	0	4	5	1	1	2	0	4	5	1	2	2
	2	1	1	6	2	0	2	1	1	6	2	0	2	2
	6	2	5	1	3	1	6	2	5	1	3	1	3	2
mean	3	2	3	4	3	1	? ? ? ? ? ? ? ? ? ? ?-----
std	3	0	3	3	2	1	? ? ? ? ? ? ? ? ? ? ------

""" 解碼部分-代碼片段 """#傳入參數 #Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N) class Decoder(nn.Module):"Generic N layer decoder with masking."def __init__(self, layer, N):super(Decoder, self).__init__()self.layers = clones(layer, N)self.norm = LayerNorm(layer.size)def forward(self, x, memory, src_mask, tgt_mask):for layer in self.layers:x = layer(x, memory, src_mask, tgt_mask)return self.norm(x)#傳入參數 #DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N class DecoderLayer(nn.Module):"Decoder is made of self-attn, src-attn, and feed forward (defined below)"def __init__(self, size, self_attn, src_attn, feed_forward, dropout):super(DecoderLayer, self).__init__()self.size = sizeself.self_attn = self_attnself.src_attn = src_attnself.feed_forward = feed_forwardself.sublayer = clones(SublayerConnection(size, dropout), 3)# memory就是encoder完成之后的結果def forward(self, x, memory, src_mask, tgt_mask):"Follow Figure 1 (right) for connections."m = memory# self-attetion q=k=v,輸入是decoder的embeddingx = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))# soft-attention q!=k=v x是deocder的embedding，m是encoder的輸出x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))return self.sublayer[2](x, self.feed_forward)def subsequent_mask(size):"Mask out subsequent positions."attn_shape = (1, size, size)subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8') # print(subsequent_mask)return torch.from_numpy(subsequent_mask) == 0

?

Attention部分

?

?

""" attention 代碼段 """def attention(query, key, value, mask=None, dropout=None):# shape:query=key=value---->[batch_size,8,max_length,64]d_k = query.size(-1)# k的緯度交換后為：[batch_size,8,64,max_length]# scores的緯度為:[batch_size,8,max_length,max_length]scores = torch.matmul(query, key.transpose(-2, -1)) \/ math.sqrt(d_k)#padding maskif mask is not None:scores = scores.masked_fill(mask == 0, -1e9)p_attn = F.softmax(scores, dim = -1)if dropout is not None:p_attn = dropout(p_attn)return torch.matmul(p_attn, value), p_attn

?

Multi-Head Attention - 將Q、K、V分成h個頭，方便并行計算

?

?

就是由原來

?

變成多頭

?

""" multi-head Attention """class MultiHeadedAttention(nn.Module):def __init__(self, h, d_model, dropout=0.1):"Take in model size and number of heads."super(MultiHeadedAttention, self).__init__()assert d_model % h == 0# We assume d_v always equals d_kself.d_k = d_model // hself.h = hself.linears = clones(nn.Linear(d_model, d_model), 4)self.attn = Noneself.dropout = nn.Dropout(p=dropout)def forward(self, query, key, value, mask=None):# 緯度# shape:query=key=value--->:[batch_size,max_legnth,embedding_dim=512]if mask is not None:# Same mask applied to all h heads.mask = mask.unsqueeze(1)nbatches = query.size(0)#第一步：將q,k,v分別與Wq，Wk，Wv矩陣進行相乘#shape:Wq=Wk=Wv----->[512,512]#第二步：將獲得的Q、K、V在第三個緯度上進行切分#shape:[batch_size,max_length,8,64]#第三部：填充到第一個緯度#shape:[batch_size,8,max_length,64]query, key, value = \[l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2)for l, x in zip(self.linears, (query, key, value))]#進入到attention之后緯度不變，shape:[batch_size,8,max_length,64]x, self.attn = attention(query, key, value, mask=mask, dropout=self.dropout)# 將緯度進行還原# 交換緯度：[batch_size,max_length,8,64]# 緯度還原：[batch_size,max_length,512]x = x.transpose(1, 2).contiguous() \.view(nbatches, -1, self.h * self.d_k)# 最后與WO大矩陣相乘 shape:[512,512]return self.linears[-1](x)

?

Feed-Forward Network

FFN(x) = max(0,xW1 + b1) *W2 + b2

前饋層: 包含兩層 1、線性結構 2、卷積結構 x: 上一層的輸出(一般是self-attention的輸出) w1、w2、b1、b2都是需要學習的參數

?

""" Feedforward """class PositionwiseFeedForward(nn.Module):"Implements FFN equation."def __init__(self, d_model, d_ff, dropout=0.1):super(PositionwiseFeedForward, self).__init__()#[512,2048]self.w_1 = nn.Linear(d_model, d_ff)#[2048,512]self.w_2 = nn.Linear(d_ff, d_model)self.dropout = nn.Dropout(dropout)def forward(self, x):return self.w_2(self.dropout(F.relu(self.w_1(x))))

?

""" Embedding and softmax """"class Embeddings(nn.Module):def __init__(self, d_model, vocab):super(Embeddings, self).__init__()self.lut = nn.Embedding(vocab, d_model)self.d_model = d_modeldef forward(self, x):return self.lut(x) * math.sqrt(self.d_model)

?

Position Encoding -注入序列位置信息

?

?

""" position encoding """ class PositionalEncoding(nn.Module):"Implement the PE function."def __init__(self, d_model, dropout, max_len=5000):super(PositionalEncoding, self).__init__()self.dropout = nn.Dropout(p=dropout)# Compute the positional encodings once in log space.pe = torch.zeros(max_len, d_model)position = torch.arange(0., max_len).unsqueeze(1)div_term = torch.exp(torch.arange(0., d_model, 2) *-(math.log(10000.0) / d_model))pe[:, 0::2] = torch.sin(position * div_term)pe[:, 1::2] = torch.cos(position * div_term)pe = pe.unsqueeze(0)self.register_buffer('pe', pe)def forward(self, x):x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False)return self.dropout(x)

?

""" Full Model """import copy def make_model(src_vocab, tgt_vocab, N=6, d_model=512, d_ff=2048, h=8, dropout=0.1):"Helper: Construct a model from hyperparameters."c = copy.deepcopyattn = MultiHeadedAttention(h, d_model)ff = PositionwiseFeedForward(d_model, d_ff, dropout)position = PositionalEncoding(d_model, dropout)model = EncoderDecoder(Encoder(EncoderLayer(d_model, c(attn), c(ff), dropout), N),Decoder(DecoderLayer(d_model, c(attn), c(attn), c(ff), dropout), N),nn.Sequential(Embeddings(d_model, src_vocab), c(position)),nn.Sequential(Embeddings(d_model, tgt_vocab), c(position)),Generator(d_model, tgt_vocab))# This was important from their code. # Initialize parameters with Glorot / fan_avg.for p in model.parameters():if p.dim() > 1:nn.init.xavier_uniform(p)return model# Small example model. tmp_model = make_model(10, 10, 2)

五、Why self-attention

? ? ? ? ? 主要論述分析self-attention和cnn、lstm的時間復雜度

六、Trainging

? ? ? ? ? 主要論述訓練語料以及硬件、超參數設置介紹

七、Results

? ? ? ? ? 主要介紹transformer的實驗結果對比

八、Discussion

? ? ? ? ? 主要討論將transformer結構應用于除翻譯之外的其它領域

啟發點:

? ? ?1、在進行attention機制的時候，對于padd為0的位置可以mask去掉

? ? ?2、可以在模型中添加殘差網絡和layer normalization提高模型效果

? ? ?3、模型創新的時候可以添加self-attention結構

?

關鍵點:

? ? ?1、scaled Dot-Product Attention原理

? ? ?2、Multi Head Attention 實現

? ? ?3、Mask機制、Layer Normalization、加法attention和dot attention區別

?

九、相關代碼

?

? ? ?本篇論文已開源代碼，開源代碼地址:?https://github.com/tensorflow/tensor2tensor

? ??

?

?

?

?

總結

以上是生活随笔為你收集整理的Attention Is All You Need (transformer)的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。