快速学习一个算法，Transformer-51CTO.COM

大家好，我是小寒

今天给大家介绍一个强大的算法模型，Transformer

Transformer 模型是由 Vaswani 等人在 2017 年提出的一种用于自然语言处理的深度学习模型，特别擅长于处理序列到序列的任务，如机器翻译、文本生成等。

今天，我们主要从编码的角度来进行说明。

Transformer 模型架构

Transformer 模型由编码器（Encoder）和解码器（Decoder）两部分组成，每部分包含多个相同的层 (Layer) 堆叠而成。

图片

编码器

编码器由 N 个相同的层组成，每个层包括以下两个子层：

多头自注意力机制（Multi-Head Self-Attention Mechanism）
前馈神经网络（Feed-Forward Neural Network）

每个子层都采用残差连接 (Residual Connection) 和层归一化 (Layer Normalization)。

解码器

解码器也由 N 个相同的层组成，每个层包括以下三个子层：

多头自注意力机制（Masked Multi-Head Self-Attention Mechanism）
编码器-解码器注意力机制（Encoder-Decoder Attention Mechanism）
前馈神经网络（Feed-Forward Neural Network）

同样，每个子层都采用残差连接和层归一化。

Transformer 模型的主要组件

Token Embedding
将输入的离散单词转换为连续的向量表示。
位置编码
在输入序列中引入位置信息，因为 Transformer 本身没有顺序依赖，需要位置编码来帮助模型识别序列的顺序。
多头自注意力
计算 token 之间的注意力分数，使模型能够关注序列中的不同位置，从而捕捉全局信息。
前馈层
对每个位置的表示进行独立的线性变换和激活函数，以增强模型的表达能力。
编码器-解码器注意力机制
解码器中的每个位置可以关注编码器中的所有位置，从而将编码器的信息传递给解码器。
残差连接
在每个子层（子模块）的输入和输出之间添加一个快捷连接（skip connection），以避免梯度消失和梯度爆炸问题，使得训练更稳定。

代码演练

首先，我们导入必要的库；

import numpy as np
import torch
import math
from torch import nn
import torch.nn.functional as F

嵌入将单词转换为 Embedding ；

class Embeddings(nn.Module):
    def __init__(self, vocab_size, d_model):
        super(Embeddings, self).__init__()
        self.embed = nn.Embedding(vocab_size, d_model)
        self.d_model = d_model

    def forward(self, x):
        return self.embed(x) * math.sqrt(self.d_model)

位置编码添加了有关序列顺序的信息。

class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_sequence_length):
        super().__init__()
        self.max_sequence_length = max_sequence_length
        self.d_model = d_model

    def forward(self, x):
        even_i = torch.arange(0, self.d_model, 2).float()
        denominator = torch.pow(10000, even_i/self.d_model)
        position = (torch.arange(self.max_sequence_length)
                          .reshape(self.max_sequence_length, 1))
        even_PE = torch.sin(position / denominator)
        odd_PE = torch.cos(position / denominator)
        stacked = torch.stack([even_PE, odd_PE], dim=2)
        PE = torch.flatten(stacked, start_dim=1, end_dim=2)
        return PE

多头自注意力层；

def scaled_dot_product(q, k, v, mask=None):
        d_k = q.size()[-1]
        scaled = torch.matmul(q, k.transpose(-1, -2)) / math.sqrt(d_k)
        if mask is not None:
            scaled = scaled.permute(1, 0, 2, 3) + mask
            scaled = scaled.permute(1, 0, 2, 3)
        attention = F.softmax(scaled, dim=-1)
        values = torch.matmul(attention, v)
        return values, attention
        
class MultiHeadAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads
        self.qkv_layer = nn.Linear(d_model , 3 * d_model)
        self.linear_layer = nn.Linear(d_model, d_model)
        
    def forward(self, x, mask):
        batch_size, sequence_length, d_model = x.size()
        qkv = self.qkv_layer(x)
        qkv = qkv.reshape(batch_size, sequence_length, self.num_heads, 3 * self.head_dim)
        qkv = qkv.permute(0, 2, 1, 3)
        q, k, v = qkv.chunk(3, dim=-1)
        values, attention = scaled_dot_product(q, k, v, mask)
        values = values.permute(0, 2, 1, 3).reshape(batch_size, sequence_length, self.num_heads * self.head_dim)
        out = self.linear_layer(values)
        return out

前馈层使用 ReLu 激活函数和线性层。

class PositionwiseFeedForward(nn.Module):
    def __init__(self, d_model, hidden, drop_prob=0.1):
        super(PositionwiseFeedForward, self).__init__()
        self.linear1 = nn.Linear(d_model, hidden)
        self.linear2 = nn.Linear(hidden, d_model)
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(p=drop_prob)

    def forward(self, x):
        x = self.linear1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.linear2(x)
        return x

层规范化；

class LayerNormalization(nn.Module):
    def __init__(self, parameters_shape, eps=1e-5):
        super().__init__()
        self.parameters_shape=parameters_shape
        self.eps=eps
        self.gamma = nn.Parameter(torch.ones(parameters_shape))
        self.beta =  nn.Parameter(torch.zeros(parameters_shape))

    def forward(self, inputs):
        dims = [-(i + 1) for i in range(len(self.parameters_shape))]
        mean = inputs.mean(dim=dims, keepdim=True)
        var = ((inputs - mean) ** 2).mean(dim=dims, keepdim=True)
        std = (var + self.eps).sqrt()
        y = (inputs - mean) / std
        out = self.gamma * y + self.beta
        return out

编码器由多个编码器层组成。

class EncoderLayer(nn.Module):
    def __init__(self, d_model, ffn_hidden, num_heads, drop_prob):
        super(EncoderLayer, self).__init__()
        self.attention = MultiHeadAttention(d_model=d_model, num_heads=num_heads)
        self.norm1 = LayerNormalization(parameters_shape=[d_model])
        self.dropout1 = nn.Dropout(p=drop_prob)
        self.ffn = PositionwiseFeedForward(d_model=d_model, hidden=ffn_hidden, drop_prob=drop_prob)
        self.norm2 = LayerNormalization(parameters_shape=[d_model])
        self.dropout2 = nn.Dropout(p=drop_prob)

    def forward(self, x, self_attention_mask):
        residual_x = x.clone()
        x = self.attention(x, mask=self_attention_mask)
        x = self.dropout1(x)
        x = self.norm1(x + residual_x)
        residual_x = x.clone()
        x = self.ffn(x)
        x = self.dropout2(x)
        x = self.norm2(x + residual_x)
        return x

class SequentialEncoder(nn.Sequential):
    def forward(self, *inputs):
        x, self_attention_mask  = inputs
        for module in self._modules.values():
            x = module(x, self_attention_mask)
        return x

class Encoder(nn.Module):
    def __init__(self,
                 d_model,
                 ffn_hidden,
                 num_heads,
                 drop_prob,
                 num_layers,
                 max_sequence_length,
                 language_to_index,
                 START_TOKEN,
                 END_TOKEN,
                 PADDING_TOKEN):
        super().__init__()
        self.sentence_embedding = SentenceEmbedding(max_sequence_length, d_model, language_to_index, START_TOKEN, END_TOKEN, PADDING_TOKEN)
        self.layers = SequentialEncoder(*[EncoderLayer(d_model, ffn_hidden, num_heads, drop_prob)
                                      for _ in range(num_layers)])

    def forward(self, x, self_attention_mask, start_token, end_token):
        x = self.sentence_embedding(x, start_token, end_token)
        x = self.layers(x, self_attention_mask)
        return x

多头交叉注意层；

class MultiHeadCrossAttention(nn.Module):
    def __init__(self, d_model, num_heads):
        super().__init__()
        self.d_model = d_model
        self.num_heads = num_heads
        self.head_dim = d_model // num_heads
        self.kv_layer = nn.Linear(d_model , 2 * d_model)
        self.q_layer = nn.Linear(d_model , d_model)
        self.linear_layer = nn.Linear(d_model, d_model)

    def forward(self, x, y, mask):
        batch_size, sequence_length, d_model = x.size() 
        kv = self.kv_layer(x)
        q = self.q_layer(y)
        kv = kv.reshape(batch_size, sequence_length, self.num_heads, 2 * self.head_dim)
        q = q.reshape(batch_size, sequence_length, self.num_heads, self.head_dim)
        kv = kv.permute(0, 2, 1, 3)
        q = q.permute(0, 2, 1, 3)
        k, v = kv.chunk(2, dim=-1)
        values, attention = scaled_dot_product(q, k, v, mask)
        values = values.permute(0, 2, 1, 3).reshape(batch_size, sequence_length, d_model)
        out = self.linear_layer(values)
        return out

解码器由多个解码器层组成。

class DecoderLayer(nn.Module):
    def __init__(self, d_model, ffn_hidden, num_heads, drop_prob):
        super(DecoderLayer, self).__init__()
        self.self_attention = MultiHeadAttention(d_model=d_model, num_heads=num_heads)
        self.layer_norm1 = LayerNormalization(parameters_shape=[d_model])
        self.dropout1 = nn.Dropout(p=drop_prob)

        self.encoder_decoder_attention = MultiHeadCrossAttention(d_model=d_model, num_heads=num_heads)
        self.layer_norm2 = LayerNormalization(parameters_shape=[d_model])
        self.dropout2 = nn.Dropout(p=drop_prob)

        self.ffn = PositionwiseFeedForward(d_model=d_model, hidden=ffn_hidden, drop_prob=drop_prob)
        self.layer_norm3 = LayerNormalization(parameters_shape=[d_model])
        self.dropout3 = nn.Dropout(p=drop_prob)

    def forward(self, x, y, self_attention_mask, cross_attention_mask):
        _y = y.clone()
        y = self.self_attention(y, mask=self_attention_mask)
        y = self.dropout1(y)
        y = self.layer_norm1(y + _y)

        _y = y.clone()
        y = self.encoder_decoder_attention(x, y, mask=cross_attention_mask)
        y = self.dropout2(y)
        y = self.layer_norm2(y + _y)

        _y = y.clone()
        y = self.ffn(y)
        y = self.dropout3(y)
        y = self.layer_norm3(y + _y)
        return y


class SequentialDecoder(nn.Sequential):
    def forward(self, *inputs):
        x, y, self_attention_mask, cross_attention_mask = inputs
        for module in self._modules.values():
            y = module(x, y, self_attention_mask, cross_attention_mask)
        return y

class Decoder(nn.Module):
    def __init__(self,
                 d_model,
                 ffn_hidden,
                 num_heads,
                 drop_prob,
                 num_layers,
                 max_sequence_length,
                 language_to_index,
                 START_TOKEN,
                 END_TOKEN,
                 PADDING_TOKEN):
        super().__init__()
        self.sentence_embedding = SentenceEmbedding(max_sequence_length, d_model, language_to_index, START_TOKEN, END_TOKEN, PADDING_TOKEN)
        self.layers = SequentialDecoder(*[DecoderLayer(d_model, ffn_hidden, num_heads, drop_prob) for _ in range(num_layers)])

    def forward(self, x, y, self_attention_mask, cross_attention_mask, start_token, end_token):
        y = self.sentence_embedding(y, start_token, end_token)
        y = self.layers(x, y, self_attention_mask, cross_attention_mask)
        return y

def forward(self, x, y, self_attention_mask, cross_attention_mask, start_token, end_token):
    y = self.sentence_embedding(y, start_token, end_token)
    y = self.layers(x, y, self_attention_mask, cross_attention_mask)
    return y

transformer 模型

class Transformer(nn.Module):
    def __init__(self,
                d_model,
                ffn_hidden,
                num_heads,
                drop_prob,
                num_layers,
                max_sequence_length,
                spn_vocab_size,
                english_to_index,
                spanish_to_index,
                START_TOKEN,
                END_TOKEN,
                PADDING_TOKEN
                ):
        super().__init__()
        self.encoder = Encoder(d_model, ffn_hidden, num_heads, drop_prob, num_layers, max_sequence_length, english_to_ind, START_TOKEN, END_TOKEN, PADDING_TOKEN)
        self.decoder = Decoder(d_model, ffn_hidden, num_heads, drop_prob,num_layers, max_sequence_length, spanish_to_ind, START_TOKEN, END_TOKEN, PADDING_TOKEN)
        self.linear = nn.Linear(d_model, spn_vocab_size)
        self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')

    def forward(self,
                x,
                y,
                encoder_self_attention_mask=None,
                decoder_self_attention_mask=None,
                decoder_cross_attention_mask=None,
                enc_start_token=False,
                enc_end_token=False,
                dec_start_token=False, # We should make this true
                dec_end_token=False): # x, y are batch of sentences
        x = self.encoder(x, encoder_self_attention_mask, start_token=enc_start_token, end_token=enc_end_token)
        out = self.decoder(x, y, decoder_self_attention_mask, decoder_cross_attention_mask, start_token=dec_start_token, end_token=dec_end_token)
        out = self.linear(out)
        return out