Transformer model (detailed explanation of pytorch code)

Transformer

Basic introduction to the model

Compared with seq2seq, transformer is a purely attention-based architecture (self-attention has the advantages of parallel computing and the shortest maximum path length), and does not use any CNN and RNN.

As shown in the figure below, the transformer is composed of an encoder and a decoder . The transformer's encoder and decoder are superimposed based on self-attention modules . The embedded representations of the source (input) sequence and target (output) sequence will be encoded with position , and then input to the encoder and decoder respectively.

multi-headed attention

For the same key, value, query, it is hoped that different information can be extracted

• Such as short-distance relations and long-distance relations (similar to multiple output channels in convolution)

Multi-head attention uses h independent attention pooling

• Merge the output of each head to get the final result

Specific ideas:

We can transform queries, keys, and values ​​with independently learned h sets of different linear projections . Then, the h sets of transformed queries, keys and values ​​are fed into the attention pool in parallel. Finally, the outputs of these h attention pools are stitched together and transformed by another learnable linear projection to produce the final output. This design is called multi-head attention:

Model:

As shown above, an additional learnable parameter W is added

This parameter maps the dimension of query from Dq to Pq, the dimension of key from Dq to Kq, and the dimension of value from Dq to Vq. (this mapping usually reduces the number)

Finally, the output learnable parameter Wo is multiplied by the splicing result of hi to finally get the output of multi-head attention (Po)

Multi-head attention with mask

When the decoder outputs an element in the sequence, it should not consider the elements after the element

This can be achieved by masking

• That is, when calculating the Xi output, pretend that the current sequence length is i (cover/masking the content after i)

Position-Based Feedforward Networks

The essence is a fully connected layer

• Change the input shape from (b, n, d) to (bn, d) b is batch_size n is the sequence length d is the feature dimension

• Acting on two fully connected layers

• The output shape changes from (bn, d) back to (b, n, d)

• Equivalent to a one-dimensional convolutional layer with a two-layer kernel window of 1

layer normalization

Batch normalization normalizes the elements in each feature/channel (variance changes to 1 and mean changes to 0)

• Not suitable for sequence length will become nlp application (n sequence length in bn is constantly changing)

Layer normalization normalizes the elements of each sample

As shown in FIG:

Batch Normalization processes each b*len matrix in d (the blue part of the left figure) with a variance of 1 and a mean of 0. The range of operations is in each feature dimension.

Layer Normalization deals with the matrix of len*d in each batch (the blue part in the right figure) with a variance of 1 and a mean of 0. The range of operation is within a single sample. More stable than BN when changing length

Information transmission (corresponding to the line connecting the decoder and the encoder in the structure diagram)

Output y1...yn in the encoder

Use it as the key and value of multi-head attention in the ith Transformer block in decoding (query comes from the target sequence)

means that the number of blocks and output dimensions in the encoder and decoder are the same

predict

When predicting t+1 outputs, the decoder inputs the first t predicted values ​​(in self-attention, the first t predicted values ​​are used as key and value, and the t-th predicted value is used as query)

When making the t+1th prediction, the value of the previous t predictions is already known

It can be parallel when training and sequential when predicting.

summary:

• Transformer is a pure attention encoder-decoder
• Both encoder and decoder have n transformer blocks
• Use multi-head (self) attention in each block, position-based feed-forward network and layer normalization

Multi-head attention implementation

Select scaled dot product attention as each attention head

import math
import torch
from torch import nn
from d2l import torch as d2l

#@save
class MultiHeadAttention(nn.Module):
    """多头注意力"""
    def __init__(self, key_size, query_size, value_size, num_hiddens,
                 num_heads, dropout, bias=False, **kwargs):
        super(MultiHeadAttention, self).__init__(**kwargs)
        self.num_heads = num_heads
        self.attention = d2l.DotProductAttention(dropout)
        self.W_q = nn.Linear(query_size, num_hiddens, bias=bias)
        self.W_k = nn.Linear(key_size, num_hiddens, bias=bias)
        self.W_v = nn.Linear(value_size, num_hiddens, bias=bias)
        self.W_o = nn.Linear(num_hiddens, num_hiddens, bias=bias)

    def forward(self, queries, keys, values, valid_lens):
        # queries，keys，values的形状:
        # (batch_size，查询或者“键－值”对的个数，num_hiddens)
        # valid_lens　的形状:
        # (batch_size，)或(batch_size，查询的个数)
        # 经过变换后，输出的queries，keys，values　的形状:
        # (batch_size*num_heads，查询或者“键－值”对的个数，
        # num_hiddens/num_heads)
        queries = transpose_qkv(self.W_q(queries), self.num_heads)
        keys = transpose_qkv(self.W_k(keys), self.num_heads)
        values = transpose_qkv(self.W_v(values), self.num_heads)

        if valid_lens is not None:
            # 在轴0，将第一项（标量或者矢量）复制num_heads次，
            # 然后如此复制第二项，然后诸如此类。
            valid_lens = torch.repeat_interleave(
                valid_lens, repeats=self.num_heads, dim=0)

        # output的形状:(batch_size*num_heads，查询的个数，
        # num_hiddens/num_heads)
        output = self.attention(queries, keys, values, valid_lens)

        # output_concat的形状:(batch_size，查询的个数，num_hiddens)
        output_concat = transpose_output(output, self.num_heads)
        return self.W_o(output_concat)

#@save
def transpose_qkv(X, num_heads):
    """为了多注意力头的并行计算而变换形状"""
    # 输入X的形状:(batch_size，查询或者“键－值”对的个数，num_hiddens)
    # 输出X的形状:(batch_size，查询或者“键－值”对的个数，num_heads，
    # num_hiddens/num_heads)
    X = X.reshape(X.shape[0], X.shape[1], num_heads, -1)

    # 输出X的形状:(batch_size，num_heads，查询或者“键－值”对的个数,
    # num_hiddens/num_heads)
    X = X.permute(0, 2, 1, 3)

    # 最终输出的形状:(batch_size*num_heads,查询或者“键－值”对的个数,
    # num_hiddens/num_heads)
    return X.reshape(-1, X.shape[2], X.shape[3])


#@save
def transpose_output(X, num_heads):
    """逆转transpose_qkv函数的操作"""
    X = X.reshape(-1, num_heads, X.shape[1], X.shape[2])
    X = X.permute(0, 2, 1, 3)
    return X.reshape(X.shape[0], X.shape[1], -1)

num_hiddens, num_heads = 100, 5
attention = MultiHeadAttention(num_hiddens, num_hiddens, num_hiddens,
                               num_hiddens, num_heads, 0.5)
attention.eval()

batch_size, num_queries = 2, 4
num_kvpairs, valid_lens =  6, torch.tensor([3, 2])
X = torch.ones((batch_size, num_queries, num_hiddens))
Y = torch.ones((batch_size, num_kvpairs, num_hiddens))
attention(X, Y, Y, valid_lens).shape

Transformer implementation

## from https://github.com/graykode/nlp-tutorial/tree/master/5-1.Transformer
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import matplotlib.pyplot as plt
import math


def make_batch(sentences):
    input_batch = [[src_vocab[n] for n in sentences[0].split()]]
    output_batch = [[tgt_vocab[n] for n in sentences[1].split()]]
    target_batch = [[tgt_vocab[n] for n in sentences[2].split()]]
    return torch.LongTensor(input_batch), torch.LongTensor(output_batch), torch.LongTensor(target_batch)



## 10
def get_attn_subsequent_mask(seq):
    """
    seq: [batch_size, tgt_len]
    """
    attn_shape = [seq.size(0), seq.size(1), seq.size(1)]
    # attn_shape: [batch_size, tgt_len, tgt_len]
    subsequence_mask = np.triu(np.ones(attn_shape), k=1)  # 生成一个上三角矩阵
    subsequence_mask = torch.from_numpy(subsequence_mask).byte()
    return subsequence_mask  # [batch_size, tgt_len, tgt_len]


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

    def forward(self, Q, K, V, attn_mask):
        ## 输入进来的维度分别是 [batch_size x n_heads x len_q x d_k]  K： [batch_size x n_heads x len_k x d_k]  V: [batch_size x n_heads x len_k x d_v]
        ##首先经过matmul函数得到的scores形状是 : [batch_size x n_heads x len_q x len_k]
        scores = torch.matmul(Q, K.transpose(-1, -2)) / np.sqrt(d_k)

        ## 然后关键词地方来了，下面这个就是用到了我们之前重点讲的attn_mask，把被mask的地方置为无限小，softmax之后基本就是0，对q的单词不起作用
        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


## 6. MultiHeadAttention
class MultiHeadAttention(nn.Module):
    def __init__(self):
        super(MultiHeadAttention, self).__init__()
        ## 输入进来的QKV是相等的，我们会使用映射linear做一个映射得到参数矩阵Wq, Wk,Wv
        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)
        self.linear = nn.Linear(n_heads * d_v, d_model)
        self.layer_norm = nn.LayerNorm(d_model)

    def forward(self, Q, K, V, attn_mask):

        ## 这个多头分为这几个步骤，首先映射分头，然后计算atten_scores，然后计算atten_value;
        ##输入进来的数据形状： 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)

        ##下面这个就是先映射，后分头；一定要注意的是q和k分头之后维度是一致额，所以一看这里都是dk
        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]

        ## 输入进行的attn_mask形状是 batch_size x len_q x len_k，然后经过下面这个代码得到 新的attn_mask : [batch_size x n_heads x len_q x len_k]，就是把pad信息重复了n个头上
        attn_mask = attn_mask.unsqueeze(1).repeat(1, n_heads, 1, 1)


        ##然后我们计算 ScaledDotProductAttention 这个函数，去7.看一下
        ## 得到的结果有两个：context: [batch_size x n_heads x len_q x d_v], attn: [batch_size x n_heads x len_q x len_k]
        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 = self.linear(context)
        return self.layer_norm(output + residual), attn # output: [batch_size x len_q x d_model]


## 8. PoswiseFeedForwardNet
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)
        self.layer_norm = nn.LayerNorm(d_model)

    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 self.layer_norm(output + residual)



## 4. get_attn_pad_mask

## 比如说，我现在的句子长度是5，在后面注意力机制的部分，我们在计算出来QK转置除以根号之后，softmax之前，我们得到的形状
## len_input * len*input  代表每个单词对其余包含自己的单词的影响力

## 所以这里我需要有一个同等大小形状的矩阵，告诉我哪个位置是PAD部分，之后在计算计算softmax之前会把这里置为无穷大；

## 一定需要注意的是这里得到的矩阵形状是batch_size x len_q x len_k，我们是对k中的pad符号进行标识，并没有对k中的做标识，因为没必要

## seq_q 和 seq_k 不一定一致，在交互注意力，q来自解码端，k来自编码端，所以告诉模型编码这边pad符号信息就可以，解码端的pad信息在交互注意力层是没有用到的；

def get_attn_pad_mask(seq_q, seq_k):
    batch_size, len_q = seq_q.size()
    batch_size, len_k = seq_k.size()
    # eq(zero) is PAD token
    pad_attn_mask = seq_k.data.eq(0).unsqueeze(1)  # batch_size x 1 x len_k, one is masking
    return pad_attn_mask.expand(batch_size, len_q, len_k)  # batch_size x len_q x len_k


## 3. PositionalEncoding 代码实现
class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncoding, self).__init__()

        ## 位置编码的实现其实很简单，直接对照着公式去敲代码就可以，下面这个代码只是其中一种实现方式；
        ## 从理解来讲，需要注意的就是偶数和奇数在公式上有一个共同部分，我们使用log函数把次方拿下来，方便计算；
        ## pos代表的是单词在句子中的索引，这点需要注意；比如max_len是128个，那么索引就是从0，1，2，...,127
        ##假设我的demodel是512，2i那个符号中i从0取到了255，那么2i对应取值就是0,2,4...510
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)## 这里需要注意的是pe[:, 0::2]这个用法，就是从0开始到最后面，补长为2，其实代表的就是偶数位置
        pe[:, 1::2] = torch.cos(position * div_term)##这里需要注意的是pe[:, 1::2]这个用法，就是从1开始到最后面，补长为2，其实代表的就是奇数位置
        ## 上面代码获取之后得到的pe:[max_len*d_model]

        ## 下面这个代码之后，我们得到的pe形状是：[max_len*1*d_model]
        pe = pe.unsqueeze(0).transpose(0, 1)

        self.register_buffer('pe', pe)  ## 定一个缓冲区，其实简单理解为这个参数不更新就可以

    def forward(self, x):
        """
        x: [seq_len, batch_size, d_model]
        """
        x = x + self.pe[:x.size(0), :]
        return self.dropout(x)


## 5. EncoderLayer ：包含两个部分，多头注意力机制和前馈神经网络
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):
        ## 下面这个就是做自注意力层，输入是enc_inputs，形状是[batch_size x seq_len_q x d_model] 需要注意的是最初始的QKV矩阵是等同于这个输入的，去看一下enc_self_attn函数 6.
        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


## 2. Encoder 部分包含三个部分：词向量embedding，位置编码部分，注意力层及后续的前馈神经网络

class Encoder(nn.Module):
    def __init__(self):
        super(Encoder, self).__init__()
        self.src_emb = nn.Embedding(src_vocab_size, d_model)  ## 这个其实就是去定义生成一个矩阵，大小是 src_vocab_size * d_model
        self.pos_emb = PositionalEncoding(d_model) ## 位置编码情况，这里是固定的正余弦函数，也可以使用类似词向量的nn.Embedding获得一个可以更新学习的位置编码
        self.layers = nn.ModuleList([EncoderLayer() for _ in range(n_layers)]) ## 使用ModuleList对多个encoder进行堆叠，因为后续的encoder并没有使用词向量和位置编码，所以抽离出来；

    def forward(self, enc_inputs):
        ## 这里我们的 enc_inputs 形状是： [batch_size x source_len]

        ## 下面这个代码通过src_emb，进行索引定位，enc_outputs输出形状是[batch_size, src_len, d_model]
        enc_outputs = self.src_emb(enc_inputs) #把数字索引转化为对应的向量

        ## 这里就是位置编码，把两者相加放入到了这个函数里面，从这里可以去看一下位置编码函数的实现；3.
        enc_outputs = self.pos_emb(enc_outputs.transpose(0, 1)).transpose(0, 1)

        ##get_attn_pad_mask是为了得到句子中pad的位置信息，给到模型后面，在计算自注意力和交互注意力的时候去掉pad符号的影响，去看一下这个函数 4.
        enc_self_attn_mask = get_attn_pad_mask(enc_inputs, enc_inputs)
        enc_self_attns = []
        for layer in self.layers:
            ## 去看EncoderLayer 层函数 5.
            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

## 10.
class 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

## 9. Decoder

class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        self.tgt_emb = nn.Embedding(tgt_vocab_size, d_model)
        self.pos_emb = PositionalEncoding(d_model)
        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)  # [batch_size, tgt_len, d_model]
        dec_outputs = self.pos_emb(dec_outputs.transpose(0, 1)).transpose(0, 1) # [batch_size, tgt_len, d_model]

        ## get_attn_pad_mask 自注意力层的时候的pad 部分
        dec_self_attn_pad_mask = get_attn_pad_mask(dec_inputs, dec_inputs)

        ## get_attn_subsequent_mask 这个做的是自注意层的mask部分，就是当前单词之后看不到，使用一个上三角为1的矩阵
        dec_self_attn_subsequent_mask = get_attn_subsequent_mask(dec_inputs)

        ## 两个矩阵相加，大于0的为1，不大于0的为0，为1的在之后就会被fill到无限小
        dec_self_attn_mask = torch.gt((dec_self_attn_pad_mask + dec_self_attn_subsequent_mask), 0)


        ## 这个做的是交互注意力机制中的mask矩阵，enc的输入是k，我去看这个k里面哪些是pad符号，给到后面的模型；注意哦，我q肯定也是有pad符号，但是这里我不在意的，之前说了好多次了哈
        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


## 1. 从整体网路结构来看，分为三个部分：编码层，解码层，输出层
class Transformer(nn.Module):
    def __init__(self):
        super(Transformer, self).__init__()
        self.encoder = Encoder()  ## 编码层
        self.decoder = Decoder()  ## 解码层
        # 输出层 d_model 是我们解码层每个token输出的维度大小，之后会做一个 tgt_vocab_size 大小的softmax
        self.projection = nn.Linear(d_model, tgt_vocab_size, bias=False)

    def forward(self, enc_inputs, dec_inputs):
        # 这里有两个数据进行输入，一个是enc_inputs 形状为[batch_size, src_len]，主要是作为编码段的输入，一个dec_inputs，形状为[batch_size, tgt_len]，主要是作为解码端的输入
        # enc_inputs作为输入 形状为[batch_size, src_len]，输出由自己的函数内部指定，想要什么指定输出什么，可以是全部tokens的输出，可以是特定每一层的输出；也可以是中间某些参数的输出；
        # enc_outputs就是主要的输出，enc_self_attns这里没记错的是QK转置相乘之后softmax之后的矩阵值，代表的是每个单词和其他单词相关性；
        enc_outputs, enc_self_attns = self.encoder(enc_inputs)

        # dec_outputs 是decoder主要输出，用于后续的linear映射； dec_self_attns类比于enc_self_attns 是查看每个单词对decoder中输入的其余单词的相关性；dec_enc_attns是decoder中每个单词对encoder中每个单词的相关性；
        dec_outputs, dec_self_attns, dec_enc_attns = self.decoder(dec_inputs, enc_inputs, enc_outputs)

        # dec_outputs做映射到词表大小
        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



if __name__ == '__main__':

    ## 句子的输入部分，
    sentences = ['ich mochte ein bier P', 'S i want a beer', 'i want a beer E']


    # Transformer Parameters
    # Padding Should be Zero
    ## 构建词表
    src_vocab = {
    
    'P': 0, 'ich': 1, 'mochte': 2, 'ein': 3, 'bier': 4}
    src_vocab_size = len(src_vocab)

    tgt_vocab = {
    
    'P': 0, 'i': 1, 'want': 2, 'a': 3, 'beer': 4, 'S': 5, 'E': 6}
    tgt_vocab_size = len(tgt_vocab)

    src_len = 5  # length of source
    tgt_len = 5  # length of target

    # 模型参数
    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

    model = Transformer()

    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adam(model.parameters(), lr=0.001)

    enc_inputs, dec_inputs, target_batch = make_batch(sentences)

    for epoch in range(20):
        optimizer.zero_grad()
        outputs, enc_self_attns, dec_self_attns, dec_enc_attns = model(enc_inputs, dec_inputs)
        loss = criterion(outputs, target_batch.contiguous().view(-1))
        print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.6f}'.format(loss))
        loss.backward()
        optimizer.step()