Transformer模型拆解分析

资源来自：DataWhale 学习资料

最近看了DataWhale 的Transformer图解，突然对Transformer的结构图有了更加清晰的理解，特此记录。

1、大框架

Transformer是由6个encoder和6个decoder组成，模型的具体实现是model变量里边，参数有Encoder[编码器]、Decoder[解码器]、Embedding(src_vocab)[输入文本进行词向量化]、Embedding(tgt_vocab)[目标文本进行词向量化]，Generator[生成器]。

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.deepcopy#多头注意力attn = MultiHeadedAttention(h, d_model)#前馈神经网络ff = PositionwiseFeedForward(d_model, d_ff, dropout)#位置编码position = PositionalEncoding(d_model, dropout)#模型定义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

查看EncoderDecoder函数，搭建了一个seq2seq框架，即包含encoder和decoder，在EncoderDecoder函数中，变量src是输入文本，tgt是输出文本，src_mask是输入文本的掩码，tgt_mask是输出文本的掩码，memory是encoder的最终输出。

class EncoderDecoder(nn.Module):"""基础的Encoder-Decoder结构。A standard Encoder-Decoder architecture. Base for this and many other models."""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 = generatordef forward(self, src, tgt, src_mask, tgt_mask):"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)

2、Encoder

（1）clone

由于Transformer是有6个encoder组成，则使用clone函数进行EncodeLayer层的复制：

def clones(module, N):"产生N个完全相同的网络层""Produce N identical layers."return nn.ModuleList([copy.deepcopy(module) for _ in range(N)])

（解释：nn.ModuleList 函数是保存子模块列表通过for循环，建立6个Encoder）

（2）Encoder

class Encoder(nn.Module):"完整的Encoder包含N层"def __init__(self, layer, N):super(Encoder, self).__init__()self.layers = clones(layer, N)self.norm = LayerNorm(layer.size)def forward(self, x, mask):"每一层的输入是x和mask"for layer in self.layers:x = layer(x, mask)return self.norm(x)

Encoder需要进行“层归一化”，因此是在encoder建立之后进行了LayerNorm操作。

（3）EncoderLayer

先介绍EncoderLayer层（一个编码器encoder），编码器的构成部分是self_Attention->SubLayerConnection(层归一化和残差连接)->FFNN->SubLayerConnection(层归一化和残差连接).

代码中，对SubLayerConnection复制两份，分别加入在self-Attention和FFNN之后。

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)

（4）LayerNorm

‘层归一化’是该层的输入值进行对归一化处理，公式为 $a_2\ast \frac{x-mean}{std+eps}+b_2$ ,层归一化分别在Encoder中的Attention（自身注意力）和FFNN(前馈神经网络)模块后。

class LayerNorm(nn.Module):"Construct a layernorm module (See citation for details)."def __init__(self, features, eps=1e-6):super(LayerNorm, self).__init__()self.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_2

（5）残差连接

所进行的操作时，对输入数据进行层归一化，然后进行sublayer操作，此时sublayer传入的操作是self.attn和self.feed_forward.

class 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)))

因为self-Attention和FFNN在encoder和decoder有异同，下边进行集中梳理。

3、Decoder

（1）Decoder

在Decoder中，也进行了clone操作，此处相较于encoder，多了一个memory和src、tgt的掩码mask。

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)

（2）DecoderLayer

相较于EncoderLayer层，多了一个attention操作，即self_attn是在decoder的注意力机制，即增加了mask机制，src_attn是encoder的输出结果，q是decoder层，k,v是encoder的输出。

（模块1是self_attn,模块2是src_attn）

由于新增一个attention模块，SubLayerConnection就有三层，解码器的构成部分是self_Attention->SubLayerConnection(层归一化和残差连接)->src_Attention->SubLayerConnection(层归一化和残差连接)->FFNN->SubLayerConnection(层归一化和残差连接).

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)def forward(self, x, memory, src_mask, tgt_mask):"Follow Figure 1 (right) for connections."m = memoryx = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))return self.sublayer[2](x, self.feed_forward)

4、Embedding src_vocab & tgt_vocab

Embedding是对文本进行词向量转换，调用函数为nn.Embedding,且进行了math.sqrt(self.d_model)操作。

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)

5、额外实现

（1）self-Attention

Attention计算

目前，Atention机制的演变过程是加法和点积计算，加法计算是计算q,k的相似度，点积是计算q,k的点积，公式为点积计算。

在进行Attention计算时，特别注意mask参数，当mask不为None时，则加入了Mask机制

def attention(query, key, value, mask=None, dropout=None):"Compute 'Scaled Dot Product Attention'"d_k = query.size(-1)scores = torch.matmul(query, key.transpose(-2, -1)) \/ math.sqrt(d_k)if 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

只计算单个Attention很难捕捉输入句中所有空间的讯息，为了优化模型，论文提出了一个multi head的概念，把key,value,query线性映射到不同空间h次，但是在传入Scaled-Dot-Product Attention中时，需要固定的长度，因此再对head进行concat。

代码如下：

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):"Implements Figure 2"if mask is not None:# Same mask applied to all h heads.mask = mask.unsqueeze(1)nbatches = query.size(0)# 1) Do all the linear projections in batch from d_model => h x d_k 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))]# 2) Apply attention on all the projected vectors in batch. x, self.attn = attention(query, key, value, mask=mask, dropout=self.dropout)# 3) "Concat" using a view and apply a final linear. x = x.transpose(1, 2).contiguous() \.view(nbatches, -1, self.h * self.d_k)return self.linears[-1](x)

定义了4个linear层，前三个分别对q,v,k进行分解，维度是(h,d_k,关系是d_model = h*d_k，h是head的数量)，最后一个linear层是对多头的连接之后的数据进行线性变换。

mask机制

mask机制就是防止在训练的时候使用未来的输出的单词，确保对位置i的预测仅依赖于已知的位置i之前的输出，而不会依赖于位置i之后的输出。比如训练时，第一个单词是不能参考第二个单词的生成结果的。 mask就会把这个信息变成0，用来保证预测位置 i 的信息只能基于比 i 小的输出；

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')return torch.from_numpy(subsequent_mask) == 0

生成一个上三角矩阵，令size=3，测试结果为

（2）FFNN

FFNN有两层线性变换，结构是linear->relu->dropout->linear。

class PositionwiseFeedForward(nn.Module):"Implements FFN equation."def __init__(self, d_model, d_ff, dropout=0.1):super(PositionwiseFeedForward, self).__init__()self.w_1 = nn.Linear(d_model, d_ff)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))))

（3）位置编码

encoder的输入层和decoder的输入层是一样的结构，都是token embedding（词向量）+ positional embedding（位置向量），得到最终的输入向量。之所以引入positional embedding主要是解决单单使用token embedding（类似于词袋子），并没有词序的概念的问题。因为该模型并不包括任何的循环或卷积神经网络，所以模型添加了位置编码，为模型提供了关于单词再句子中相对位置的信息。这个向量能决定当前词的位置，或者说在一个句子中不同的词之间的距离。计算方法如下：

pos表示单词的位置，i是指单词的维度，偶数位置用正弦，奇数位置用余弦。

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)

squeeze和unsqueeze函数：对张量Tensor的维度进行压缩或者扩充！！！

6、实现顺序

（1）模拟数据

def data_gen(V, batch, nbatches):"Generate random data for a src-tgt copy task."for i in range(nbatches):data = torch.from_numpy(np.random.randint(1, V, size=(batch, 10)))data[:, 0] = 1src = Variable(data, requires_grad=False)tgt = Variable(data, requires_grad=False)yield Batch(src, tgt, 0)

（2）批处理和掩码

class Batch:"Object for holding a batch of data with mask during training."def __init__(self, src, trg=None, pad=0):self.src = srcself.src_mask = (src != pad).unsqueeze(-2)if trg is not None:self.trg = trg[:, :-1]self.trg_y = trg[:, 1:]self.trg_mask = \self.make_std_mask(self.trg, pad)self.ntokens = (self.trg_y != pad).data.sum()@staticmethoddef make_std_mask(tgt, pad):"Create a mask to hide padding and future words."tgt_mask = (tgt != pad).unsqueeze(-2)tgt_mask = tgt_mask & Variable(subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data))return tgt_mask

（3）模型优化

class NoamOpt:"Optim wrapper that implements rate."def __init__(self, model_size, factor, warmup, optimizer):self.optimizer = optimizerself._step = 0self.warmup = warmupself.factor = factorself.model_size = model_sizeself._rate = 0def step(self):"Update parameters and rate"self._step += 1rate = self.rate()for p in self.optimizer.param_groups:p['lr'] = rateself._rate = rateself.optimizer.step()def rate(self, step = None):"Implement `lrate` above"if step is None:step = self._stepreturn self.factor * \(self.model_size ** (-0.5) *min(step ** (-0.5), step * self.warmup ** (-1.5)))def get_std_opt(model):return NoamOpt(model.src_embed[0].d_model, 2, 4000,torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))

（4）标签平滑

class LabelSmoothing(nn.Module):"Implement label smoothing."def __init__(self, size, padding_idx, smoothing=0.0):super(LabelSmoothing, self).__init__()self.criterion = nn.KLDivLoss(size_average=False)self.padding_idx = padding_idxself.confidence = 1.0 - smoothingself.smoothing = smoothingself.size = sizeself.true_dist = Nonedef forward(self, x, target):assert x.size(1) == self.sizetrue_dist = x.data.clone()true_dist.fill_(self.smoothing / (self.size - 2))true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)true_dist[:, self.padding_idx] = 0mask = torch.nonzero(target.data == self.padding_idx)if mask.dim() > 0:true_dist.index_fill_(0, mask.squeeze(), 0.0)self.true_dist = true_distreturn self.criterion(x, Variable(true_dist, requires_grad=False))

（5）损失函数计算

class SimpleLossCompute:"A simple loss compute and train function."def __init__(self, generator, criterion, opt=None):self.generator = generatorself.criterion = criterionself.opt = optdef __call__(self, x, y, norm):x = self.generator(x)loss = self.criterion(x.contiguous().view(-1, x.size(-1)), y.contiguous().view(-1)) / normloss.backward()if self.opt is not None:self.opt.step()self.opt.optimizer.zero_grad()return loss.item() * norm

（6）批次运行

def run_epoch(data_iter, model, loss_compute):"Standard Training and Logging Function"start = time.time()total_tokens = 0total_loss = 0tokens = 0for i, batch in enumerate(data_iter):out = model.forward(batch.src, batch.trg, batch.src_mask, batch.trg_mask)loss = loss_compute(out, batch.trg_y, batch.ntokens)total_loss += losstotal_tokens += batch.ntokenstokens += batch.ntokensif i % 50 == 1:elapsed = time.time() - startprint("Epoch Step: %d Loss: %f Tokens per Sec: %f" %(i, loss / batch.ntokens, tokens / elapsed))start = time.time()tokens = 0return total_loss / total_tokens

（7）调用

# Train the simple copy task.
V = 11
#标签平滑
criterion = LabelSmoothing(size=V, padding_idx=0, smoothing=0.0)
#定义模型
model = make_model(V, V, N=2)
#模型优化，采用Adam
model_opt = NoamOpt(model.src_embed[0].d_model, 1, 400,torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9))#训练10次，并进行损失函数计算
for epoch in range(10):model.train()run_epoch(data_gen(V, 30, 20), model, SimpleLossCompute(model.generator, criterion, model_opt))model.eval()print(run_epoch(data_gen(V, 30, 5), model, SimpleLossCompute(model.generator, criterion, None)))

参考教程：

1、learn-nlp-with-transformers/2.2.1-Pytorch编写Transformer.md at main · datawhalechina/learn-nlp-with-transformers · GitHub

（原文链接）：The Annotated Transformer
2、Datawhale-零基础入门NLP-新闻文本分类Task06_樱缘之梦-CSDN博客