Pytorch实现Transformer

实现add&Normalize

首先实现这个部分是因为不论解码器还是编码器都需要这一部分。

首先我们假设归一化（LayerNorm）已经做好了。

python 复制代码

class SublayerConnection(nn.Module):
    def __init__(self,size,dropout=0.1):
        super(SublayerConnection,self).__init__()
        self.layer_norm = LayerNorm(size)
        # 假设LayerNorm已经做好，做一个dropout防止过拟合
        self.dropout = nn.Dropout(p=dropout)
    def forward(self,x,sublayer):
        # x+sublayer(x)就是add，sublayer就是上一步的操作
        return self.dropout(self.layer_norm(x+sublayer(x)))

因为我们不确定是Encoder还是Decoder，所以用sublayer传入来作为上一层。

接下来实现LayerNorm

python 复制代码

class LayerNorm(nn.Module):
    def __init__(self,x_size,eps=1e-6):
        super(LayerNorm,self).__init__()
        self.a_2 = nn.Parameter(torch.ones(x_size))
        self.b_2 = nn.Parameter(torch.zeros(x_size))
        self.eps = eps
    def forward(self,x):
        mean = x.mean(-1,keepdim = True)
        std = x.std(-1,keepdim = True)
        return self.a_2*(x-mean)/(math.sqrt(std)+self.eps)+self.b_2

这一串代码很简单，就是公式的实现而已，注意要把a_2和b_2作为可以学习的变量。

实现多头注意力机制和自注意力机制

自注意力机制

python 复制代码

def self_attention(query,key,value,dropout = None,mask = None):
    d_k = query.size(-1)
    sores = np.matmul(query,key.transpose(-2,-1))/math.sqrt(d_k)
    attn_softmax = F.softmax(sores,dim=-1)
    if dropout is not None:
        attn_softmax = dropout(attn_softmax)
    return np.matmul(attn_softmax,value),attn_softmax

实现QKV的相乘，除以 d k \sqrt{d_k} dk 为了便于softmax得出的概率差距不会太极端。

多头自注意力机制

python 复制代码

class MultiHeadattention(nn.Module):
    def __init__(self,head,d_model,dropout = 0.1):
        super(MultiHeadattention,self).__init__()
        assert(d_model % head == 0)
        self.d_k = d_model//head
        self.head = head
        self.d_model = d_model
        self.linear_quary = nn.Linear(d_model,d_model)
        self.linear_key = nn.Linear(d_model,d_model)
        self.linear_value = nn.Linear(d_model,d_model)
        self.linear_out = nn.Linear(d_model,d_model)
        self.dropout = nn.Dropout(p=dropout)
        self.attn = None
    def forward(self,quary,key,value,mask = None):
        n_batch = quary.size(0)
        quary = self.linear_quary(quary).view(n_batch,-1,self.head,self.d_k).transpose(1,2)
        key = self.linear_key(key).view(n_batch,-1,self.head,self.d_k).transpose(1,2)
        value = self.linear_value(value).view(n_batch,-1,self.head,self.d_k).transpose(1,2)
        x,self.attn = self_attention(quary,key,value,dropout=self.dropout,mask=mask)
        x = x.transpose(1,2).contiguous().view(n_batch,-1,self.head*self.d_k)
        return self.linear_out(x)

还有掩码自注意力需要添加，此时没有添加，等讲解到掩码的时候添加。

QKV实现分割并行计算之后，对x进行拼接重塑维度，做一个线性变换输出x。

位置编码

python 复制代码

class PositionalEncoding(nn.Module):
    def __init__(self,dim,dropout,max_len=5000):
        super(PositionalEncoding,self).__init__()
        if dim%2 != 0:
            raise ValueError("Can't use sin/cos positional encoding with"
                             "odd dim (got dim={:d})".format(dim))
        '''
        维度必须是偶数，这个if就是如果维度是奇数的情况下跳出异常，
        只有维度是偶数的情况下才能都有对称性，确保每一个位置都有一个sin一个cos
        能够提供更完整的位置属性   
        '''
        pe = torch.zeros(max_len,dim)
        position = torch.arange(0,max_len).unsqueeze(1)
        div_term = torch.exp((torch.arange(0,dim,2,dtype=torch.float)*-(math.log(10000.0)/dim)))
        pe[:,1::2] = torch.sin(position.float() * div_term)
        pe[:,0::2] = torch.cos(position.float() * div_term)
        pe = pe.unsqueeze(1)
        self.register_buffer('pe',pe)
        # 创建pe的时候用的zeros使得pe是一个可学习的参数，但是位置编码是一直不变的，因此用这个函数使得pe不会改变，但会存储
        self.drop_out = nn.Dropout(p = dropout)
        self.dim = dim
    def forward(self,emb,step=None):
        if step == None:
            emb = emb + self.pe[:emb.size(0)]
        else:
            emb = emb + self.pe[:step]
        emb = self.drop_out(emb)
        return emb

FeedForward实现

python 复制代码

class PositionWiseFeedForward(nn.Module):
    def __init__(self,d_model,d_ff,dropout=0.1) -> None:
        super(PositionWiseFeedForward,self).__init__()
        self.w_1 = nn.Linear(d_model,d_ff)
        self.w_2 = nn.Linear(d_ff,d_model)
        self.relu = nn.ReLU()
        self.layer_norm = nn.LayerNorm(d_model,eps=1e-6)
        self.dropout1 = nn.Dropout(p=dropout)
        self.dropout2 = nn.Dropout(p=dropout)
    def forward(self,x):
        inter = self.dropout1(self.relu(self.w_1(self.layer_norm(x))))
        output = self.dropout2(self.relu(self.w_2(inter)))
        return output

FFN非常简单，就是两次wx+b，加一个激活函数即可，代码里面初始 w 1 , w 2 , r e l u w_1,w_2,relu w1,w2,relu等，为什么没有b呢，pytorch linear在初始化的时候已经帮助我们初始化好了不需要我们自己初始化。

EncoderDecoder层实现

python 复制代码

def clones(module,n):
    return nn.ModuleList([deepcopy(module) for _ in range(n)])

class EncoderLayer(nn.Module):
    def __init__(self,size,attn,feed_forward,dropout=0.1):
        super(EncoderLayer,self).__init__()
        self.attn = attn
        self.feed_forward = feed_forward
        self.sublayer_connection_list = clones(SublayerConnection(size,dropout),2)
    def forward(self,x,mask):
        first_x = self.sublayer_connection_list[0](x,lambda x_attn:self.attn(x,x,x,mask))
        return self.sublayer_connection_list[1](x,self.feed_forward)
        
class Encoder(nn.Module):
    def __init__(self,n,encoder_layer):
        super(Encoder,self).__init__()
        self.encoder_layer_list = clones(encoder_layer,n)
    def forward(self,x,src_mask):
        for encoder_layer in self.encoder_layer_list:
            x = encoder_layer(x,src_mask)
        return x

class DecoderLayer(nn.model):
    def __init__(self,d_model,attn,feed_forward,sublayer_num,dropout=0.1):
        super(DecoderLayer,self).__init__()
        self.attn = attn
        self.feed_forward = feed_forward
        self.sublayer_connection_list = clones(SublayerConnection(d_model,dropout),sublayer_num)
    def forward(self,x,l2r_memory,src_mask,trg_mask,r2l_memory=None,r2l_trg_mask = None):
        first_x = self.sublayer_connection_list[0](x,lambda x_attn: self.attn(x,l2r_memory,l2r_memory,trg_mask))
        second_x = self.sublayer_connection_list[1](first_x,lambda sencond_x_attn:self.attn(first_x,l2r_memory,l2r_memory,src_mask))
        if r2l_memory is not None:
            third_x = self.sublayer_connection_list[-2](second_x,lambda third_x_attn:self.attn(second_x,r2l_memory,r2l_memory,r2l_trg_mask))
            return self.sublayer_connection_list[-1](third_x,self.feed_forward)
        else:
            return self.sublayer_connection_list[-1](second_x,self.feed_forward)


class Decoder(nn.Module):
    def __init__(self,n_layers,decoder_layer):
        super(Decoder,self).__init__()
        self.decoder_layer_list = clones(decoder_layer,n_layers)
    def forward(self,x,memory,src_mask,trg_mask):
        for decoder_layer in self.decoder_layer_list:
            x = decoder_layer(x,memory,src_mask,trg_mask)
        return x

首先实现一个用来复制的函数，因为不管是decoder还是encoder都有两个add和Normalize。

encoder和decoder都先实现一层，再用一个类来多次创建实现多层叠加。

encoderlayer就是一层encoder，就是实现多头+add和Normalize+ffn+add和Normalize，其中传入的参数attn是创建好的多头自注意力模型，ffn同理

decoderlayer就是一层decoder，同理是实现多头+add和Normalize+ffn+add和Normalize，只不过接受的参数不同，x是输入，memory是encoder层的输出，后面两个是掩码，最后两个参数不用管，是反向解码。

拼凑Transformer

python 复制代码

class WordProbGenerator(nn.Module):
    def __init__(self,d_model,vocab_size):
        super(WordProbGenerator,self).__init__()
        self.linear = nn.Linear(d_model,vocab_size)
    def forward(self,x):
        return F.log_softmax(self.linear(x),dim=-1)

class WordEmbedding(nn.Module):
    def __init__(self,vocab_size,d_model) -> None:
        super(WordEmbedding,self).__init__()
        self.embedding = nn.Embedding(vocab_size,d_model)
        self.embed = self.embedding
        self.d_model = d_model
    def forward(self,x):
        return self.embed(x)*math.sqrt(self.d_model)


class ABDTransformer(nn.Module):
    def __init__(self,vocab,d_model,d_ff,n_head,n_layer,dropout,device='cuda'):
        super(ABDTransformer,self).__init__()
        self.vocab = vocab
        self.device = device
        attn = MultiHeadattention(n_head,d_model,dropout)
        feed_forward = PositionWiseFeedForward(d_model,d_ff,dropout)
        self.src_embed = WordEmbedding(vocab,d_model)
        self.pos_embed = PositionalEncoding(d_model,dropout)
        self.encoder = Encoder(n_layer,EncoderLayer(d_model,deepcopy(attn),deepcopy(feed_forward),dropout=dropout))
        self.decoder = Decoder(n_layer,DecoderLayer(d_model,deepcopy(attn),deepcopy(feed_forward),sublayer_num=3,dropout=dropout))
        self.word_prob_generator = WordProbGenerator(d_model,vocab)
    def encode(self,src,src_mask):
        x = self.src_embed(src[0])
        x = self.pos_embed(x)
        x = self.encoder(x,src_mask)
        return x
    def decode(self,trg,memory,src_mask,trg_mask):
        x = self.src_embed(trg)
        x = self.pos_embed(x)
        return self.decoder(x,memory,src_mask,trg_mask)
        
    def forward(self,src,trg,mask):
        src_mask,trg_mask = mask
        encoder_output = self.encode(src,src_mask)
        decoder_output = self.decode(trg,encoder_output,src_mask,trg_mask)
        pred = self.word_prob_generator(decoder_output)
        return pred

第一个函数是得到各个词的概率

第二个函数是将输入词向量的维度，就是字典的维度改成d_model的维度。

组建transformer

第一步的init中将模型都初始化，包括encoder，decoder，ffn，多头自注意力机制，位置编码等。

第二步写decode和encode的输入输出，简单来说我们已经实现了内部的操作，现在写两个函数，将他需要的输入进去即可。

第三步调用函数，最后得到概率输出。