关于transformer的原理，这里就不多说，主要还是结合论文中的图来对代码进行一下讲解。
看这张图，其实可以看到最核心的部分就是下面这一块：
关于讲解，我就直接写在代码里面，用中文来对其进行详细的一个介绍。相对应的代码如下：
class ScaledDotProductAttention(nn.Module):
''' Scaled Dot-Product Attention '''
def __init__(self, temperature, attn_dropout=0.1):
super().__init__()
self.temperature = temperature
self.dropout = nn.Dropout(attn_dropout)
def forward(self, q, k, v, mask=None):
attn = torch.matmul(q / self.temperature, k.transpose(2, 3)) # query和每个key进行相似度计算得到权重
if mask is not None:
attn = attn.masked_fill(mask == 0, -1e9)
attn = self.dropout(F.softmax(attn, dim=-1)) # 使用一个softmax函数对这些权重进行归一化
output = torch.matmul(attn, v) # 权重和相应的键值value进行加权求和得到最后的attention
return output, attn
相对应的公式和图，看下面。
除了点积之外，还可以用cosine的相似性、mlp网络等来计算score
class MultiHeadAttention(nn.Module):
''' Multi-Head Attention module '''
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super().__init__()
self.n_head = n_head # 注意力头的数目，说白了就是你想吧hidden_size 分成几部分来分别计算，一般取8/12
self.d_k = d_k # 每个注意力头的大小
self.d_v = d_v # 每个注意力头的大小
# d_model == n_head*d_k
self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False) # 这里的d_model=q的hidden_size
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
self.fc = nn.Linear(n_head * d_v, d_model, bias=False)
self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6) # d_model=hidden_size
def forward(self, q, k, v, mask=None): # 一般q!=(k==v),如果是self_atten,就是q==k==v
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1) # q:(batch, seq_len, hidden_size)
residual = q # 保留原始的q,后面做完attention之后要把原始的q进行相加，具体看上图最左边的黑色箭头
q = self.layer_norm(q) # q:(batch, seq_len, hidden_size)
# Pass through the pre-attention projection: b x lq x (n*dv)
# Separate different heads: b x lq x n x dv
q = self.w_qs(q).view(sz_b, len_q, n_head, d_k) # (batch, seq_len, n_head, d_k)
k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
# Transpose for attention dot product: b x n x lq x dv
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2) # (batch, n_head, seq_len, d_k)
if mask is not None:
mask = mask.unsqueeze(1) # For head axis broadcasting.
q, attn = self.attention(q, k, v, mask=mask) # 要计算atten系数和更新之后的q
# Transpose to move the head dimension back: b x lq x n x dv
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1) # (batch, seq_len, hidden_size)
q = self.dropout(self.fc(q))
q += residual # 将原始的q进行相加
return q, attn # attn里面是所有的系数，其实已经用过了，就没啥作用了，主要保留的是经过atten之后的q
class PositionwiseFeedForward(nn.Module):
''' A two-feed-forward-layer module '''
def __init__(self, d_in, d_hid, dropout=0.1):
super().__init__()
self.w_1 = nn.Linear(d_in, d_hid) # position-wise
self.w_2 = nn.Linear(d_hid, d_in) # position-wise
self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
residual = x
x = self.layer_norm(x)
x = self.w_2(F.relu(self.w_1(x)))
x = self.dropout(x)
x += residual
return x