Transformer的pytorch实现

Transformer架构

Transformer模型由encoder和decoder两部分组成，decoder输出的结果，经过一个线性层，然后计算softmax。

d_model = 512  # Embedding Size
d_ff = 2048 # FeedForward dimension
d_k = d_v = 64  # dimension of K(=Q), V
n_layers = 6  # number of Encoder of Decoder Layer
n_heads = 8  # number of heads in Multi-Head Attention

class Transformer(nn.Module):
    def __init__(self):
        super(Transformer, self).__init__()
        self.encoder = Encoder()
        self.decoder = Decoder()
        self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False)
    def forward(self, enc_inputs, dec_inputs):
        enc_outputs, enc_self_attns = self.encoder(enc_inputs)
        dec_outputs, dec_self_attns, dec_enc_attns = self.decoder(dec_inputs, enc_inputs, enc_outputs)
        dec_logits = self.projection(dec_outputs) # dec_logits : [batch_size x src_vocab_size x tgt_vocab_size]
        return dec_logits.view(-1, dec_logits.size(-1)), enc_self_attns, dec_self_attns, dec_enc_attns

实例化一个Transformer模型，输入为enc_inputs[batch_size,src_len]，dec_inputs[batch_size,tgt_len]

model = Transformer()
outputs, enc_self_attns, dec_self_attns, dec_enc_attns = model(enc_inputs, dec_inputs)

Encoder

Encoder分为embedding和6层Encoderlayer，首层的Encoderlayer输入为词embedding和位置embedding相加之和enc_outputs:[batch_size,seq_len,d_model],即(1,5,512)。
Attention_mask:[batch_size,seq_lenq,seq_len],即(1,5,5)，其中one is masking。
用enc_self_attns列表保存每一层encoder的attention，列表长度为6。

class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()
        self.src_emb = nn.Embedding(src_vocab_size, d_model)
        self.pos_emb = nn.Embedding.from_pretrained(get_sinusoid_encoding_table(src_len+1, d_model),freeze=True)
        self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)])

    def forward(self, enc_inputs): # enc_inputs : [batch_size x source_len]
        enc_outputs = self.src_emb(enc_inputs) + self.pos_emb(torch.LongTensor([[1,2,3,4,0]]))
        enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs)
        enc_self_attns = []
        for layer in self.layers:
            enc_outputs, enc_self_attn = layer(enc_outputs, enc_self_attn_mask)
            enc_self_attns.append(enc_self_attn)
        return enc_outputs, enc_self_attns

6层Encoderlayer

每层Encoderlayer包括Multi-head attention和 position-wise Feed Forward，以及这两个子层之间有残差连接。

class EncoderLayer(nn.Module):
    def __init__(self):
        super(EncoderLayer, self).__init__()
        self.enc_self_attn = MultiHeadAttention()
        self.pos_ffn = PoswiseFeedForwardNet()

    def forward(self, enc_inputs, enc_self_attn_mask):
        enc_outputs, attn = self.enc_self_attn(enc_inputs, enc_inputs, enc_inputs, enc_self_attn_mask) # enc_inputs to same Q,K,V
        enc_outputs = self.pos_ffn(enc_outputs) # enc_outputs: [batch_size x len_q x d_model]
        return enc_outputs, attn

Multi-head attention子层

输入的Q,K,V均为enc_inputs:[batch_size,seq_len,d_model],即(1,5,512)
W_Q, W_K, W_V分别经过nn.liner(d_model,d_k*heads) , 虽然维数均为512，但经过线性变换后三者的值不再相等，q_s,k_s,v_s分别将W_Q, W_K, W_V变换为[batch_size,n_heads,seq_len,d_q/d_k/d_v]
Attn_mask加入head后变为[batch_size,n_heads,seq_len,seq_len],即(1,8,5,5)

class MultiHeadAttention(nn.Module):
    def __init__(self):
        super(MultiHeadAttention, self).__init__()
        self.W_Q = nn.Linear(d_model, d_k * n_heads)
        self.W_K = nn.Linear(d_model, d_k * n_heads)
        self.W_V = nn.Linear(d_model, d_v * n_heads)
    def forward(self, Q, K, V, attn_mask):
        # q: [batch_size x len_q x d_model], k: [batch_size x len_k x d_model], v: [batch_size x len_k x d_model]
        residual, batch_size = Q, Q.size(0)
        # (B, S, D) -proj-> (B, S, D) -split-> (B, S, H, W) -trans-> (B, H, S, W)
        q_s = self.W_Q(Q).view(batch_size, -1, n_heads, d_k).transpose(1,2)  # q_s: [batch_size x n_heads x len_q x d_k]
        k_s = self.W_K(K).view(batch_size, -1, n_heads, d_k).transpose(1,2)  # k_s: [batch_size x n_heads x len_k x d_k]
        v_s = self.W_V(V).view(batch_size, -1, n_heads, d_v).transpose(1,2)  # v_s: [batch_size x n_heads x len_k x d_v]

        attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1) # attn_mask : [batch_size x n_heads x len_q x len_k]

        # context: [batch_size x n_heads x len_q x d_v], attn: [batch_size x n_heads x len_q(=len_k) x len_k(=len_q)]
        context, attn = ScaledDotProductAttention()(q_s, k_s, v_s, attn_mask)
        context = context.transpose(1, 2).contiguous().view(batch_size, -1, n_heads * d_v) # context: [batch_size x len_q x n_heads * d_v]
        output = nn.Linear(n_heads * d_v, d_model)(context)
        return nn.LayerNorm(d_model)(output + residual), attn # output: [batch_size x len_q x d_model]

ScaledDotProductAttention：将q_s,k_s经过matmul；然后除以一个缩放因子；再进行sequence masking除去msaking部分的干扰，masking部分不需要给予attention；经过softmax后计算attention 权重；最后与v_s 进行matmul得到atttion后的输出context：[batch_size,n_heads,seq_len,d_v]和atten:[batch_size,n_heads,seq_len,seq_len]矩阵。

class ScaledDotProductAttention(nn.Module):
    def __init__(self):
        super(ScaledDotProductAttention, self).__init__()

    def forward(self, Q, K, V, attn_mask):
        scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k) # scores : [batch_size x n_heads x len_q(=len_k) x len_k(=len_q)]
        scores.masked_fill_(attn_mask, -1e9) # Fills elements of self tensor with value where mask is one.
        attn = nn.Softmax(dim=-1)(scores)
        context = torch.matmul(attn, V)
        return context, attn

将 context经过nn.linear变换为[batch_size,seq_len,d_model],(1,5,512),加上原先的输入（残差），进行 Layer normalization，输出enc_outputs(1,5,512), attn(1,8,5,5）

ontext = context.transpose(1, 2).contiguous().view(batch_size, -1, n_heads * d_v) # context: [batch_size x len_q x n_heads * d_v]
output = nn.Linear(n_heads * d_v, d_model)(context)
return nn.LayerNorm(d_model)(output + residual), attn # output: [batch_size x len_q x d_model]      

Position-wise Feed-Forward network子层

这是一个全连接网络，包含两个线性变换和一个非线性函数（实际上就是ReLU）。公式如下：
FFN(x) = max(0,xW_{1}+b_{1})W_{2}+b_{2}
论文中提到，这个公式还可以用两个核大小为1的一维卷积来解释，卷积的输入输出都是d_{model}=512，中间层的维度是d_{ff}=2048。
最后经过残差连接和 Layer normalization

class PoswiseFeedForwardNet(nn.Module):
    def __init__(self):
        super(PoswiseFeedForwardNet, self).__init__()
        self.conv1 = nn.Conv1d(in_channels=d_model, out_channels=d_ff, kernel_size=1)
        self.conv2 = nn.Conv1d(in_channels=d_ff, out_channels=d_model, kernel_size=1)
    def forward(self, inputs):
        residual = inputs # inputs : [batch_size, len_q, d_model]
        output = nn.ReLU()(self.conv1(inputs.transpose(1, 2)))
        output = self.conv2(output).transpose(1, 2)
        return nn.LayerNorm(d_model)(output + residual)

至此，第一层的Encoderlayer输出为enc_outputs（1,5,512），enc_self_attns.append(attn),其中attn为第一层的attention（1,8,5,5)
enc_outputs作为下一层的输入，经过6层Encoderlayer后，Encoder端的最后输出为：enc_outputs（1,5,512），enc_self_attns(6,1,8,5,5)

Decoder

Decoder的输入为dec_imputs:[batch_size,seq_len]；enc_inputs:[batch_size,seq_len],非embedding,为词典index； enc_outputs:[batch_size,seq_len,d_model],为encoder端输出。

dec_outputs, dec_self_attns, dec_enc_attns = self.decoder(dec_inputs, enc_inputs,enc_outputs)

Decoder分为embedding、6层Decodelayer。
其中embedding与encoder相同，首层的Decoderlayer输入为词embedding和位置embedding相加之和enc_outputs:[batch_size,seq_len,d_model],即(1,5,512)

lass Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        self.tgt_emb = nn.Embedding(tgt_vocab_size, d_model)
        self.pos_emb = nn.Embedding.from_pretrained(get_sinusoid_encoding_table(tgt_len+1, d_model),freeze=True)
        self.layers = nn.ModuleList([DecoderLayer() for _ in range(n_layers)])
    def forward(self, dec_inputs, enc_inputs, enc_outputs): # dec_inputs : [batch_size x target_len]
        dec_outputs = self.tgt_emb(dec_inputs) + self.pos_emb(torch.LongTensor([[5,1,2,3,4]]))
        dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs)
        dec_self_attn_subsequent_mask = get_attn_subsequent_mask(dec_inputs)
        dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask + dec_self_attn_subsequent_mask), 0)
        dec_enc_attn_mask = get_attn_pad_mask(dec_inputs, enc_inputs)
        dec_self_attns, dec_enc_attns = [], []
        for layer in self.layers:
            dec_outputs, dec_self_attn, dec_enc_attn = layer(dec_outputs, enc_outputs, dec_self_attn_mask, dec_enc_attn_mask)
            dec_self_attns.append(dec_self_attn)
            dec_enc_attns.append(dec_enc_attn)
        return dec_outputs, dec_self_attns, dec_enc_attns

6层Decoderlayer

每层Decoderlayer包括Multi-head attention（self-attention和dec-enc-attention）、 position-wise Feed Forward，以及这两个子层之间有残差连接。

lass DecoderLayer(nn.Module):
    def __init__(self):
        super(DecoderLayer, self).__init__()
        self.dec_self_attn = MultiHeadAttention()
        self.dec_enc_attn = MultiHeadAttention()
        self.pos_ffn = PoswiseFeedForwardNet()
    def forward(self, dec_inputs, enc_outputs, dec_self_attn_mask, dec_enc_attn_mask):
        dec_outputs, dec_self_attn = self.dec_self_attn(dec_inputs, dec_inputs, dec_inputs, dec_self_attn_mask)
        dec_outputs, dec_enc_attn = self.dec_enc_attn(dec_outputs, enc_outputs, enc_outputs, dec_enc_attn_mask)
        dec_outputs = self.pos_ffn(dec_outputs)
        return dec_outputs, dec_self_attn, dec_enc_attn

Multi-head attention子层

同Encoderlayer的Multi-head attention

Dec-enc attention子层

相比于Encoder多了dec-enc-aatention;且mask部分包括dec_self_attn_mask和Dec_enc_attn_mask（上1下0上三角矩阵）

Position-wise Feed-Forward network子层

同Encoderlayer的Position-wise Feed-Forward network

Decoder每一层输出为:
dec_outputs： [batch_size,seq_len,d_model] (1,5,512)
dec_self_attn：[batch_size,n_heads,seq_len,seq_len] (1,8,5,5)
dec_enc_attn：[batch_size,n_heads,seq_len,seq_len] (1,8,5,5)

经过6层Decoderlayer后的输出为：
dec_outputs：[batch_size,seq_len,d_model] (1,5,512)
dec_self_attn：[layers,batch_size,n_heads,seq_len,seq_len] (6,1,8,5,5)
dec_enc_attn：[layers,batch_size,n_heads,seq_len,seq_len] (6,1,8,5,5)

Linear

dec_logits = self.projection(dec_outputs) 
 self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False)

Dec_logits为[batch_size x src_vocab_size x tgt_vocab_size]
至此，transformer model的输出为:
Outputs:[batch_size,src_vocab_size,target_vocab_size] —-> [-1,target_vocab_size]
enc_self_attns (6,1,8,5,5)
dec_self_attn：[layers,batch_size,n_heads,seq_len,seq_len] (6,1,8,5,5)
dec_enc_attn：[layers,batch_size,n_heads,seq_len,seq_len] (6,1,8,5,5)

Softmax

计算loss
target_batch为[batch_size,seq_len] (1,5)
loss = criterion(outputs, target_batch.contiguous().view(-1))
criterion = nn.CrossEntropyLoss()

参考代码:link.