終於來到transformer了，之前的幾個東西都搞的差不多了，剩下的就是搭積木搭模型了。首先來看一下transformer模型，OK好像就是那一套東西。

transformer是純基於注意力機制的架構，但是也是之前的encoder-decoder架構。

層歸一化

這裡用到了層歸一化，和之前的批量歸一化有區別。

這裡參考了torch文件：

N就是batchsize維，layernorm就是對一個batch裡序列裡的向量做歸一化。

Encoder

import torch
import torch.nn as nn
import torch.nn.functional as F
from d2l import torch as d2l

class add_norm(nn.Module):
    def __init__(self, norm_shape, dropout=0):
        super(add_norm, self).__init__()
        self.norm = nn.LayerNorm(norm_shape)
        self.dropout = nn.Dropout(dropout)
    
    def forward(self, X, Y):
        return self.norm(X + self.dropout(Y)) #這裡預設X， Y的shape一樣

class EncoderBlock(nn.Module):
    def __init__(self, embed_dim, norm_shape):
        super(EncoderBlock, self).__init__()
        self.add_norm1 = add_norm(norm_shape=norm_shape)
        self.attention = nn.MultiheadAttention(embed_dim, 8, batch_first=True) # 這裡將batch_first 設定為了True。
        self.ffn = nn.Sequential(nn.Linear(embed_dim, embed_dim), nn.ReLU(), nn.Linear(embed_dim, embed_dim))
        self.add_norm2 = add_norm(norm_shape=norm_shape)
        
        
    def forward(self, X): 
        Y,_ = self.attention(X, X, X)
        X = self.add_norm1(X, Y)
        Y = self.ffn(X)
        X = self.add_norm2(X, Y)
        return X
    
class Encoder(nn.Module):
    def __init__(self, embed_dim, norm_shape, num_block) -> None:
        super(Encoder, self).__init__()
        self.pos_encoding = d2l.PositionalEncoding(embed_dim, dropout=0)
        self.EncoderBlocks = [EncoderBlock(embed_dim, norm_shape) for _ in range(num_block)]
    
    def forward(self, X):
        X = self.pos_encoding(X)  
        for i in range(len(self.EncoderBlocks)):
            X = self.EncoderBlocks[i](X)
        return X

model = Encoder(128, [35, 128], 2)
s = torch.zeros((64, 35, 128)) 
s = model(s)
 

用torch實現了一個encoder， decoder不想寫，擺爛了，就這樣，愛咋滴咋滴，以後就呼叫框架了。

直接用框架實現了，愛咋滴咋滴吧。