前言

机器翻译（Machine Translation, MT）是一类将某种语言（源语言，source language）的句子 $x$ 翻译成另一种语言（目标语言，target language）的句子 $y$ 的任务。机器翻译的相关研究早在上世纪50年代美苏冷战时期就开始了，当时的机器翻译系统是基于规则的，利用两种语言的单词、短语对应关系将俄语翻译成英语。

在早期的机器翻译主要是依靠统计学模型，使用一种叫统计机器翻译（Statistical Machine Translation, SMT）的方法，在1990年至2010年间是较为主流的方法，也取得了不错的效果。统计机器翻译方法的核心思想是从数据中统计出概率模型。假设需要构建一个将法语翻译成英语的模型，我们可以将任务描述为给定法语句子 $x$ ，找到最有可能是 $x$ 的翻译的英语句子 $y$ ，即 $argmax_y P(y|x) = argmax_y P(x|y)P(y)$
上式中的贝叶斯公式可分为两部分，其中 $P (x ∣ y)$ 可以看作翻译模型，保证翻译的正确性; $P (y)$ 则是语言模型，保证翻译结果语言流畅。在当时的技术条件下，这样拆分后两个任务相对独立，分开学习会更容易一些。

在2014年，机器翻译研究迎来了重大的突破——神经机器翻译（Neural Machine Translation, NMT）。最简单的NMT系统使用单个神经网络就可以完成机器翻译任务，这个神经网络结构就是Seq2Seq（sequence-to-sequence，序列到序列），它由两个RNN组成，其中一个作为Encoder（编码器），另一个则是Decoder（解码器），Encoder的输入为源语言的语句，经过RNN处理并在最后一步得到输出，这个输出就可以看作输入句子的向量编码。

本篇文章中用于机器翻译的网络结构仍然是Seq2Seq网络结构，但是与两个RNN或者RNN+attention方法不同，经过简单的修改的transformer会是这个序列模型的主要结构。

transformer这个模型，大家都比较熟悉了，由论文《Attention is All You Need》提出，主要是可以实现并行化的训练，网络结构主要由全连接层、注意力机制和normalization层组成。关于transformer的结构在这里就不详细介绍了，大家感兴趣的可以去看看别的博主的文章，这里主要介绍如何将transformer用于机器翻译的模型训练。

代码

数据处理

数据主要使用的是torchtext中含英语和德语句子对的知名数据集Multi30k dataset，其包含大约30,000个英语和德语的句子对（平均长度约为13）。下面代码是数据处理部分和相关依赖库导入：

import torch
import torch.nn as nn
import torch.optim as optim

import torchtext
from torchtext.datasets import  Multi30k
#from torchtext.legacy.data import Field, BucketIterator   # pytorch1.9+  torchtext==0.10.0
from torchtext.data  import Field, TabularDataset, BucketIterator, Iterator  ## torchtext==0.6.0 torch==1.7

import matplotlib.pyplot as plt
import matplotlib.ticker as ticker

import numpy as np

import random
import math
import time

SEED = 1234

random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True


def tokenize_de(text):
    """
    Tokenizes German text from a string into a list of strings
    """
    return [w for w in text.split() if w]

def tokenize_en(text):
    """
    Tokenizes English text from a string into a list of strings
    """
    return [w for w in text.split() if w]


SRC = Field(tokenize=tokenize_de,
            init_token='<sos>',
            eos_token='<eos>',
            lower=True,
            batch_first=True)

TRG = Field(tokenize=tokenize_en,
            init_token='<sos>',
            eos_token='<eos>',
            lower=True,
            batch_first=True)



#train_data, valid_data, test_data = Multi30k()   # pytorch1.9+  torchtext==0.10.0

train_data, valid_data, test_data = Multi30k.splits(exts = ('.de', '.en'),    ## torchtext==0.6.0 torch==1.7
                                                    fields = (SRC, TRG))

SRC.build_vocab(train_data, min_freq = 2)
TRG.build_vocab(train_data, min_freq = 2)


device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('device:', device)


BATCH_SIZE = 64

train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
    (train_data, valid_data, test_data),
    batch_sizes=(BATCH_SIZE, BATCH_SIZE, BATCH_SIZE),
    device=device,
    sort_key=lambda x: len(x.src), # 根据源语句的长度
    sort_within_batch=False,   # 是否需要对批处理的内容进行排序
    repeat=False
)

特别说明一下，数据导入，是直接从torchtext.datasets import的，但是torchtext版本不一样，api接口会有所变化，详细看代码里面注释，笔者这里用的torchtext 0.6.0， torch 1.7。

如果由于网络的原因，下载数据失败，可以在手动下载该数据集并放到与代码同级的.data/目录下即可。

模型定义

class Encoder(nn.Module):
    def __init__(self, 
                 input_dim, 
                 hid_dim, 
                 n_layers, 
                 n_heads, 
                 pf_dim,
                 dropout, 
                 device,
                 max_length = 100):
        super().__init__()

        self.device = device
        
        self.tok_embedding = nn.Embedding(input_dim, hid_dim)
        self.pos_embedding = nn.Embedding(max_length, hid_dim)
        
        self.layers = nn.ModuleList([EncoderLayer(hid_dim, 
                                                  n_heads, 
                                                  pf_dim,
                                                  dropout, 
                                                  device) 
                                     for _ in range(n_layers)])
        
        self.dropout = nn.Dropout(dropout)
        
        self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
        
    def forward(self, src, src_mask):
        
        #src = [batch size, src len]
        #src_mask = [batch size, src len]
        
        batch_size = src.shape[0]
        src_len = src.shape[1]
        
        pos = torch.arange(0, src_len).unsqueeze(0).repeat(batch_size, 1).to(self.device)
        
        #pos = [batch size, src len]
        
        src = self.dropout((self.tok_embedding(src) * self.scale) + self.pos_embedding(pos))
        
        #src = [batch size, src len, hid dim]
        
        for layer in self.layers:
            src = layer(src, src_mask)
            
        #src = [batch size, src len, hid dim]
            
        return src



class EncoderLayer(nn.Module):
    def __init__(self, 
                 hid_dim, 
                 n_heads, 
                 pf_dim,  
                 dropout, 
                 device):
        super().__init__()
        
        self.layer_norm = nn.LayerNorm(hid_dim)
        self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads, dropout, device)
        self.positionwise_feedforward = PositionwiseFeedforwardLayer(hid_dim, 
                                                                     pf_dim, 
                                                                     dropout)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, src, src_mask):
        
        #src = [batch size, src len, hid dim]
        #src_mask = [batch size, src len]
                
        #self attention
        _src, _ = self.self_attention(src, src, src, src_mask)
        
        #dropout, residual connection and layer norm
        src = self.layer_norm(src + self.dropout(_src))
        
        #src = [batch size, src len, hid dim]
        
        #positionwise feedforward
        _src = self.positionwise_feedforward(src)
        
        #dropout, residual and layer norm
        src = self.layer_norm(src + self.dropout(_src))
        
        #src = [batch size, src len, hid dim]
        
        return src


class MultiHeadAttentionLayer(nn.Module):
    def __init__(self, hid_dim, n_heads, dropout, device):
        super().__init__()
        
        assert hid_dim % n_heads == 0
        
        self.hid_dim = hid_dim
        self.n_heads = n_heads
        self.head_dim = hid_dim // n_heads
        
        self.fc_q = nn.Linear(hid_dim, hid_dim)
        self.fc_k = nn.Linear(hid_dim, hid_dim)
        self.fc_v = nn.Linear(hid_dim, hid_dim)
        
        self.fc_o = nn.Linear(hid_dim, hid_dim)
        
        self.dropout = nn.Dropout(dropout)
        
        self.scale = torch.sqrt(torch.FloatTensor([self.head_dim])).to(device)
        
    def forward(self, query, key, value, mask = None):
        
        batch_size = query.shape[0]
        
        #query = [batch size, query len, hid dim]
        #key = [batch size, key len, hid dim]
        #value = [batch size, value len, hid dim]
                
        Q = self.fc_q(query)
        K = self.fc_k(key)
        V = self.fc_v(value)
        
        #Q = [batch size, query len, hid dim]
        #K = [batch size, key len, hid dim]
        #V = [batch size, value len, hid dim]
                
        Q = Q.view(batch_size, -1, self.n_heads, self.head_dim).permute(0, 2, 1, 3)
        K = K.view(batch_size, -1, self.n_heads, self.head_dim).permute(0, 2, 1, 3)
        V = V.view(batch_size, -1, self.n_heads, self.head_dim).permute(0, 2, 1, 3)
        
        #Q = [batch size, n heads, query len, head dim]
        #K = [batch size, n heads, key len, head dim]
        #V = [batch size, n heads, value len, head dim]
                
        energy = torch.matmul(Q, K.permute(0, 1, 3, 2)) / self.scale
        
        #energy = [batch size, n heads, query len, key len]
        
        if mask is not None:
            energy = energy.masked_fill(mask == 0, -1e10)
        
        attention = torch.softmax(energy, dim = -1)
                
        #attention = [batch size, n heads, query len, key len]
                
        x = torch.matmul(self.dropout(attention), V)
        
        #x = [batch size, n heads, query len, head dim]
        
        x = x.permute(0, 2, 1, 3).contiguous()
        
        #x = [batch size, query len, n heads, head dim]
        
        x = x.view(batch_size, -1, self.hid_dim)
        
        #x = [batch size, query len, hid dim]
        
        x = self.fc_o(x)
        
        #x = [batch size, query len, hid dim]
        
        return x, attention

class PositionwiseFeedforwardLayer(nn.Module):
    def __init__(self, hid_dim, pf_dim, dropout):
        super().__init__()
        
        self.fc_1 = nn.Linear(hid_dim, pf_dim)
        self.fc_2 = nn.Linear(pf_dim, hid_dim)
        
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, x):
        
        #x = [batch size, seq len, hid dim]
        
        x = self.dropout(torch.relu(self.fc_1(x)))
        
        #x = [batch size, seq len, pf dim]
        
        x = self.fc_2(x)
        
        #x = [batch size, seq len, hid dim]
        
        return x


class Decoder(nn.Module):
    def __init__(self, 
                 output_dim, 
                 hid_dim, 
                 n_layers, 
                 n_heads, 
                 pf_dim, 
                 dropout, 
                 device,
                 max_length = 100):
        super().__init__()
        
        self.device = device
        
        self.tok_embedding = nn.Embedding(output_dim, hid_dim)
        self.pos_embedding = nn.Embedding(max_length, hid_dim)
        
        self.layers = nn.ModuleList([DecoderLayer(hid_dim, 
                                                  n_heads, 
                                                  pf_dim, 
                                                  dropout, 
                                                  device)
                                     for _ in range(n_layers)])
        
        self.fc_out = nn.Linear(hid_dim, output_dim)
        
        self.dropout = nn.Dropout(dropout)
        
        self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
        
    def forward(self, trg, enc_src, trg_mask, src_mask):
        
        #trg = [batch size, trg len]
        #enc_src = [batch size, src len, hid dim]
        #trg_mask = [batch size, trg len]
        #src_mask = [batch size, src len]
                
        batch_size = trg.shape[0]
        trg_len = trg.shape[1]
        
        pos = torch.arange(0, trg_len).unsqueeze(0).repeat(batch_size, 1).to(self.device)
                            
        #pos = [batch size, trg len]
            
        trg = self.dropout((self.tok_embedding(trg) * self.scale) + self.pos_embedding(pos))
                
        #trg = [batch size, trg len, hid dim]
        
        for layer in self.layers:
            trg, attention = layer(trg, enc_src, trg_mask, src_mask)
        
        #trg = [batch size, trg len, hid dim]
        #attention = [batch size, n heads, trg len, src len]
        
        output = self.fc_out(trg)
        
        #output = [batch size, trg len, output dim]
            
        return output, attention


class DecoderLayer(nn.Module):
    def __init__(self, 
                 hid_dim, 
                 n_heads, 
                 pf_dim, 
                 dropout, 
                 device):
        super().__init__()
        
        self.layer_norm = nn.LayerNorm(hid_dim)
        self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads, dropout, device)
        self.encoder_attention = MultiHeadAttentionLayer(hid_dim, n_heads, dropout, device)
        self.positionwise_feedforward = PositionwiseFeedforwardLayer(hid_dim, 
                                                                     pf_dim, 
                                                                     dropout)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, trg, enc_src, trg_mask, src_mask):
        
        #trg = [batch size, trg len, hid dim]
        #enc_src = [batch size, src len, hid dim]
        #trg_mask = [batch size, trg len]
        #src_mask = [batch size, src len]
        
        #self attention
        _trg, _ = self.self_attention(trg, trg, trg, trg_mask)
        
        #dropout, residual connection and layer norm
        trg = self.layer_norm(trg + self.dropout(_trg))
            
        #trg = [batch size, trg len, hid dim]
            
        #encoder attention
        _trg, attention = self.encoder_attention(trg, enc_src, enc_src, src_mask)
        
        #dropout, residual connection and layer norm
        trg = self.layer_norm(trg + self.dropout(_trg))
                    
        #trg = [batch size, trg len, hid dim]
        
        #positionwise feedforward
        _trg = self.positionwise_feedforward(trg)
        
        #dropout, residual and layer norm
        trg = self.layer_norm(trg + self.dropout(_trg))
        
        #trg = [batch size, trg len, hid dim]
        #attention = [batch size, n heads, trg len, src len]
        
        return trg, attention


class Seq2Seq(nn.Module):
    def __init__(self, 
                 encoder, 
                 decoder, 
                 src_pad_idx, 
                 trg_pad_idx, 
                 device):
        super().__init__()
        
        self.encoder = encoder
        self.decoder = decoder
        self.src_pad_idx = src_pad_idx
        self.trg_pad_idx = trg_pad_idx
        self.device = device
        
    def make_src_mask(self, src):
        
        #src = [batch size, src len]
        
        src_mask = (src != self.src_pad_idx).unsqueeze(1).unsqueeze(2)

        #src_mask = [batch size, 1, 1, src len]

        return src_mask
    
    def make_trg_mask(self, trg):
        
        #trg = [batch size, trg len]
        
        trg_pad_mask = (trg != self.trg_pad_idx).unsqueeze(1).unsqueeze(2)
        
        #trg_pad_mask = [batch size, 1, 1, trg len]
        
        trg_len = trg.shape[1]
        
        trg_sub_mask = torch.tril(torch.ones((trg_len, trg_len), device = self.device)).bool()
        
        #trg_sub_mask = [trg len, trg len]
            
        trg_mask = trg_pad_mask & trg_sub_mask
        
        #trg_mask = [batch size, 1, trg len, trg len]
        
        return trg_mask

    def forward(self, src, trg):
        
        #src = [batch size, src len]
        #trg = [batch size, trg len]
                
        src_mask = self.make_src_mask(src)
        trg_mask = self.make_trg_mask(trg)
        
        #src_mask = [batch size, 1, 1, src len]
        #trg_mask = [batch size, 1, trg len, trg len]
        
        enc_src = self.encoder(src, src_mask)
        
        #enc_src = [batch size, src len, hid dim]
                
        output, attention = self.decoder(trg, enc_src, trg_mask, src_mask)
        
        #output = [batch size, trg len, output dim]
        #attention = [batch size, n heads, trg len, src len]
        
        return output, attention

模型训练

INPUT_DIM = len(SRC.vocab)
OUTPUT_DIM = len(TRG.vocab)
HID_DIM = 256
ENC_LAYERS = 3
DEC_LAYERS = 3
ENC_HEADS = 8
DEC_HEADS = 8
ENC_PF_DIM = 512
DEC_PF_DIM = 512
ENC_DROPOUT = 0.1
DEC_DROPOUT = 0.1

enc = Encoder(INPUT_DIM, 
              HID_DIM, 
              ENC_LAYERS, 
              ENC_HEADS, 
              ENC_PF_DIM, 
              ENC_DROPOUT, 
              device)

dec = Decoder(OUTPUT_DIM, 
              HID_DIM, 
              DEC_LAYERS, 
              DEC_HEADS, 
              DEC_PF_DIM, 
              DEC_DROPOUT, 
              device)

SRC_PAD_IDX = SRC.vocab.stoi[SRC.pad_token]
TRG_PAD_IDX = TRG.vocab.stoi[TRG.pad_token]

model = Seq2Seq(enc, dec, SRC_PAD_IDX, TRG_PAD_IDX, device).to(device)

def count_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)

print(f'The model has {
      
      count_parameters(model):,} trainable parameters')

def initialize_weights(m):
    if hasattr(m, 'weight') and m.weight.dim() > 1:
        nn.init.xavier_uniform_(m.weight.data)

model.apply(initialize_weights)

LEARNING_RATE = 0.0005

optimizer = torch.optim.Adam(model.parameters(), lr = LEARNING_RATE)

criterion = nn.CrossEntropyLoss(ignore_index = TRG_PAD_IDX)


def train(model, iterator, optimizer, criterion, clip):
    
    model.train()
    
    epoch_loss = 0
    
    for i, batch in enumerate(iterator):
        
        src = batch.src
        trg = batch.trg
        
        optimizer.zero_grad()
        
        output, _ = model(src, trg[:,:-1])
                
        #output = [batch size, trg len - 1, output dim]
        #trg = [batch size, trg len]
            
        output_dim = output.shape[-1]
            
        output = output.contiguous().view(-1, output_dim)
        trg = trg[:,1:].contiguous().view(-1)
                
        #output = [batch size * trg len - 1, output dim]
        #trg = [batch size * trg len - 1]
            
        loss = criterion(output, trg)
        
        loss.backward()
        
        torch.nn.utils.clip_grad_norm_(model.parameters(), clip)
        
        optimizer.step()
        
        epoch_loss += loss.item()
        
    return epoch_loss / len(iterator)

def evaluate(model, iterator, criterion):
    
    model.eval()
    
    epoch_loss = 0
    
    with torch.no_grad():
    
        for i, batch in enumerate(iterator):

            src = batch.src
            trg = batch.trg

            output, _ = model(src, trg[:,:-1])
            
            #output = [batch size, trg len - 1, output dim]
            #trg = [batch size, trg len]
            
            output_dim = output.shape[-1]
            
            output = output.contiguous().view(-1, output_dim)
            trg = trg[:,1:].contiguous().view(-1)
            
            #output = [batch size * trg len - 1, output dim]
            #trg = [batch size * trg len - 1]
            
            loss = criterion(output, trg)

            epoch_loss += loss.item()
        
    return epoch_loss / len(iterator)


def epoch_time(start_time, end_time):
    elapsed_time = end_time - start_time
    elapsed_mins = int(elapsed_time / 60)
    elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
    return elapsed_mins, elapsed_secs

N_EPOCHS = 10
CLIP = 1

best_valid_loss = float('inf')

for epoch in range(N_EPOCHS):
    
    start_time = time.time()
    
    train_loss = train(model, train_iterator, optimizer, criterion, CLIP)
    valid_loss = evaluate(model, valid_iterator, criterion)
    
    end_time = time.time()
    
    epoch_mins, epoch_secs = epoch_time(start_time, end_time)
    
    if valid_loss < best_valid_loss:
        best_valid_loss = valid_loss
        torch.save(model.state_dict(), 'tut6-model.pt')
    
    print(f'Epoch: {
      
      epoch+1:02} | Time: {
      
      epoch_mins}m {
      
      epoch_secs}s')
    print(f'\tTrain Loss: {
      
      train_loss:.3f} | Train PPL: {
      
      math.exp(train_loss):7.3f}')
    print(f'\t Val. Loss: {
      
      valid_loss:.3f} |  Val. PPL: {
      
      math.exp(valid_loss):7.3f}')

最后训练输出：
在这里插入图片描述

基于transformer的Seq2Seq机器翻译模型训练、预测教程

前言

代码

数据处理

模型定义

模型训练

猜你喜欢