【经典模型实现】一个简单的基于transformer的机器翻译

背景

最近对序列模型比较感兴趣，正好使用pytorch实现一个简单的机器翻译模型，在gpu运行环境下，在epoch=50轮左右的时候，训练效果，看似还不错。这里记录下。

效果

• 训练后的效果如下
• 正常情况下，翻译的结果和正确答案差异不大。
• 但是在某些特殊情况下，比如字符在测试集合中有，但是训练集合中没有的情况下，效果会比较差。
  

代码实现

数据预处理

import numpy as np
import unicodedata
from torch.autograd import Variable
import torch
import re
from collections import Counter
UNK = 0  # 未登录词的标识符对应的词典id
PAD = 1  # padding占位符对应的词典id
import os
# os.environ["CUDA_VISIBLE_DEVICES"] = '2'
# DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")


def seq_padding(X, padding=0):
    """
    对一个batch批次(以单词id表示)的数据进行padding填充对齐长度
    """
    # 计算该批次数据各条数据句子长度
    L = [len(x) for x in X]
    # 获取该批次数据最大句子长度
    ML = max(L)
    # 对X中各条数据x进行遍历，如果长度短于该批次数据最大长度ML，则以padding id填充缺失长度ML-len(x)
    return np.array([
        np.concatenate([x, [padding] * (ML - len(x))]) if len(x) < ML else x for x in X
    ])

def subsequent_mask(size):
    "Mask out subsequent positions."
    # 设定subsequent_mask矩阵的shape
    attn_shape = (1, size, size)

    # TODO: 生成一个右上角(不含主对角线)为全1，左下角(含主对角线)为全0的subsequent_mask矩阵
    subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')

    # TODO: 返回一个右上角(不含主对角线)为全False，左下角(含主对角线)为全True的subsequent_mask矩阵
    return torch.from_numpy(subsequent_mask) == 0


class Batch:
    "Object for holding a batch of data with mask during training."

    def __init__(self, src, trg=None, pad=0, DEVICE=None):
        # 将输入与输出的单词id表示的数据规范成整数类型
        src = torch.from_numpy(src).to(DEVICE).long()
        trg = torch.from_numpy(trg).to(DEVICE).long()
        self.src = src
        # 对于当前输入的句子非空部分进行判断成bool序列
        # 并在seq length前面增加一维，形成维度为 1×seq length 的矩阵
        self.src_mask = (src != pad).unsqueeze(-2).to(DEVICE)
        # 如果输出目标不为空，则需要对decoder要使用到的target句子进行mask
        if trg is not None:
            # decoder要用到的target输入部分
            self.trg = trg[:, :-1]
            # decoder训练时应预测输出的target结果
            self.trg_y = trg[:, 1:]
            # 将target输入部分进行attention mask
            self.trg_mask = self.make_std_mask(self.trg, pad).to(DEVICE)
            # 将应输出的target结果中实际的词数进行统计
            self.ntokens = (self.trg_y != pad).data.sum().to(DEVICE)

    # Mask掩码操作
    @staticmethod
    def make_std_mask(tgt, pad):
        "Create a mask to hide padding and future words."
        tgt_mask = (tgt != pad).unsqueeze(-2)
        tgt_mask = tgt_mask & Variable(subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data))
        return tgt_mask


class PrepareData:
    def __init__(self, train_file, dev_file, batch_size, DEVICE):
        self.batch_size = batch_size
        # 读取数据 并分词
        self.train_en, self.train_cn = self.load_data(train_file)
        self.dev_en, self.dev_cn = self.load_data(dev_file)

        # 构建单词表
        self.en_word_dict, self.en_total_words, self.en_index_dict = self.build_dict(self.train_en)
        self.cn_word_dict, self.cn_total_words, self.cn_index_dict = self.build_dict(self.train_cn)
        self.tgt_vocab = len(self.en_word_dict)
        self.src_vocab = len(self.cn_word_dict)

        # id化
        self.train_en, self.train_cn = self.wordToID(self.train_en, self.train_cn, self.en_word_dict, self.cn_word_dict)
        self.dev_en, self.dev_cn = self.wordToID(self.dev_en, self.dev_cn, self.en_word_dict, self.cn_word_dict)

        # 划分batch + padding + mask
        self.train_data = self.splitBatch(self.train_en, self.train_cn, batch_size, DEVICE)
        self.dev_data = self.splitBatch(self.dev_en, self.dev_cn, batch_size, DEVICE)

    def unicodeToAscii(self, s):
        return ''.join(
            c for c in unicodedata.normalize('NFD', s)
            if unicodedata.category(c) != 'Mn'
        )

    def normalizeString(self, s):
        s = self.unicodeToAscii(s.lower().strip())
        s = re.sub(r"([.!?])", r" \1", s)
        # s = re.sub(r"[^a-zA-Z.!?]+", r" ", s)
        return s

    def load_data(self, path):
        """
        读取翻译前(英文)和翻译后(中文)的数据文件
        每条数据都进行分词，然后构建成包含起始符(BOS)和终止符(EOS)的单词(中文为字符)列表
        形式如：en = [['BOS', 'i', 'love', 'you', 'EOS'], ['BOS', 'me', 'too', 'EOS'], ...]
                cn = [['BOS', '我', '爱', '你', 'EOS'], ['BOS', '我', '也', '是', 'EOS'], ...]
        """
        en = []
        cn = []
        # TODO ...
        with open(path, 'r', encoding='utf-8') as fin:
            for line in fin:
                list_content = line.split('\t')
                # print(list_content)
                # engs = ['BOS'] + word_tokenize(list_content[0]) + ['EOS']
                engs = ['BOS'] + self.normalizeString(list_content[0]).split(' ') + ['EOS']
                # print('engs=', engs)
                en.append(engs)
                # print('list_content[1]=', list_content[1])
                # lists = word_tokenize(" ".join(list_content[1]))
                lists = words = [w for w in self.normalizeString(list_content[1])]
                # print('lists=', lists)
                chins = ['BOS'] + lists + ['EOS']
                # print('chins=', chins)
                cn.append(chins)

        return en, cn

    def build_dict(self, sentences, max_words=50000):
        """
        传入load_data构造的分词后的列表数据
        构建词典(key为单词，value为id值)
        """
        # 对数据中所有单词进行计数
        # print('sentences=', sentences)
        word_count = Counter()

        for sentence in sentences:
            for s in sentence:
                word_count[s] += 1
        # 只保留最高频的前max_words数的单词构建词典
        # 并添加上UNK和PAD两个单词，对应id已经初始化设置过
        ls = word_count.most_common(max_words)
        # 统计词典的总词数
        total_words = len(ls) + 2

        word_dict = {w[0]: index + 2 for index, w in enumerate(ls)}
        word_dict['UNK'] = UNK
        word_dict['PAD'] = PAD
        # 再构建一个反向的词典，供id转单词使用
        index_dict = {v: k for k, v in word_dict.items()}

        return word_dict, total_words, index_dict

    def wordToID(self, en, cn, en_dict, cn_dict, sort=True):
        """
        该方法可以将翻译前(英文)数据和翻译后(中文)数据的单词列表表示的数据
        均转为id列表表示的数据
        如果sort参数设置为True，则会以翻译前(英文)的句子(单词数)长度排序
        以便后续分batch做padding时，同批次各句子需要padding的长度相近减少padding量
        """
        # 计算英文数据条数
        length = len(en)

        # TODO: 将翻译前(英文)数据和翻译后(中文)数据都转换为id表示的形式
        out_en_ids = [[en_dict.get(w, 0) for w in sent] for sent in en]
        out_cn_ids = [[cn_dict.get(w, 0) for w in sent] for sent in cn]

        # 构建一个按照句子长度排序的函数
        def len_argsort(seq):
            """
            传入一系列句子数据(分好词的列表形式)，
            按照句子长度排序后，返回排序后原来各句子在数据中的索引下标
            """
            return sorted(range(len(seq)), key=lambda x: len(seq[x]))

        # 把中文和英文按照同样的顺序排序
        if sort:
            # 以英文句子长度排序的(句子下标)顺序为基准
            sorted_index = len_argsort(out_en_ids)

            # TODO: 对翻译前(英文)数据和翻译后(中文)数据都按此基准进行排序
            out_en_ids = [out_en_ids[i] for i in sorted_index]
            out_cn_ids = [out_cn_ids[i] for i in sorted_index]

        return out_en_ids, out_cn_ids

    def splitBatch(self, en, cn, batch_size, shuffle=True, DEVICE=None):
        """
        将以单词id列表表示的翻译前(英文)数据和翻译后(中文)数据
        按照指定的batch_size进行划分
        如果shuffle参数为True，则会对这些batch数据顺序进行随机打乱
        """
        # 在按数据长度生成的各条数据下标列表[0, 1, ..., len(en)-1]中
        # 每隔指定长度(batch_size)取一个下标作为后续生成batch的起始下标
        idx_list = np.arange(0, len(en), batch_size)
        # 如果shuffle参数为True，则将这些各batch起始下标打乱
        if shuffle:
            np.random.shuffle(idx_list)
        # 存放各个batch批次的句子数据索引下标
        batch_indexs = []
        for idx in idx_list:
            # 注意，起始下标最大的那个batch可能会超出数据大小
            # 因此要限定其终止下标不能超过数据大小

            batch_indexs.append(np.arange(idx, min(idx + batch_size, len(en))))

        # 按各batch批次的句子数据索引下标，构建实际的单词id列表表示的各batch句子数据
        batches = []
        for batch_index in batch_indexs:
            # 按当前batch的各句子下标(数组批量索引)提取对应的单词id列表句子表示数据
            batch_en = [en[index] for index in batch_index]
            batch_cn = [cn[index] for index in batch_index]
            # 对当前batch的各个句子都进行padding对齐长度
            # 维度为：batch数量×batch_size×每个batch最大句子长度
            batch_cn = seq_padding(batch_cn)
            batch_en = seq_padding(batch_en)
            # 将当前batch的英文和中文数据添加到存放所有batch数据的列表中
            batches.append(Batch(batch_cn, batch_en, DEVICE=DEVICE))

        return batches

序列化模型实现

import torch.nn as nn
import torch
import copy
import torch.nn.functional as F
from torch.autograd import Variable
import math
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def clones(module, N):
    """
    克隆模型块，克隆的模型块参数不共享
    """
    return nn.ModuleList([copy.deepcopy(module).to(DEVICE) for _ in range(N)])

class MultiHeadedAttention(nn.Module):
    def __init__(self, h, d_model, dropout=0.1):
        super(MultiHeadedAttention, self).__init__()
        # 保证可以整除
        assert d_model % h == 0
        # 得到一个head的attention表示维度
        self.d_k = d_model // h
        # head数量
        self.h = h
        # 定义4个全连接函数，供后续作为WQ，WK，WV矩阵和最后h个多头注意力矩阵concat之后进行变换的矩阵
        self.linears = clones(nn.Linear(d_model, d_model), 4)
        self.attn = None
        self.dropout = nn.Dropout(p=dropout)
        self.to(DEVICE)

    def forward(self, query, key, value, mask=None):
        if mask is not None:
            mask = mask.unsqueeze(1)
        # query的第一个维度值为batch size
        nbatches = query.size(0)
        # print('nbatches=', nbatches)
        # 将embedding层乘以WQ，WK，WV矩阵(均为全连接)
        # 并将结果拆成h块，然后将第二个和第三个维度值互换(具体过程见上述解析)
        query, key, value = [l(x).view(nbatches, -1, self.h, self.d_k).transpose(1, 2) for l, x in zip(self.linears, (query, key, value))]
        # 调用上述定义的attention函数计算得到h个注意力矩阵跟value的乘积，以及注意力矩阵
        x, self.attn = self.attention(query, key, value, mask=mask, dropout=self.dropout)
        # 将h个多头注意力矩阵concat起来（注意要先把h变回到第三维的位置）
        x = x.transpose(1, 2).contiguous().view(nbatches, -1, self.h * self.d_k)
        # 使用self.linears中构造的最后一个全连接函数来存放变换后的矩阵进行返回
        return self.linears[-1](x)

    def attention(self, query, key, value, mask=None, dropout=None):
        # 将query矩阵的最后一个维度值作为d_k
        d_k = query.size(-1)

        # TODO: 将key的最后两个维度互换(转置)，才能与query矩阵相乘，乘完了还要除以d_k开根号
        scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)

        # 如果存在要进行mask的内容，则将那些为0的部分替换成一个很大的负数
        scores = scores.to(DEVICE)
        mask = mask.to(DEVICE)
        if mask is not None:
            scores = scores.masked_fill(mask == 0, -1e9)

        # TODO: 将mask后的attention矩阵按照最后一个维度进行softmax
        p_attn = F.softmax(scores, dim=-1)

        # 如果dropout参数设置为非空，则进行dropout操作
        if dropout is not None:
            p_attn = dropout(p_attn)
        # 最后返回注意力矩阵跟value的乘积，以及注意力矩阵
        return torch.matmul(p_attn, value), p_attn

class PositionwiseFeedForward(nn.Module):
    def __init__(self, d_model, d_ff, dropout=0.1):
        super(PositionwiseFeedForward, self).__init__()
        self.w_1 = nn.Linear(d_model, d_ff)
        self.w_2 = nn.Linear(d_ff, d_model)
        self.dropout = nn.Dropout(dropout)
        self.to(DEVICE)

    def forward(self, x):
        # TODO: 请利用init中的成员变量实现Feed Forward层的功能
        return self.w_2(self.dropout(F.relu(self.w_1(x))))


class LayerNorm(nn.Module):
    def __init__(self, features, eps=1e-6):
        super(LayerNorm, self).__init__()
        # 初始化α为全1, 而β为全0
        self.a_2 = nn.Parameter(torch.ones(features))
        self.b_2 = nn.Parameter(torch.zeros(features))
        # 平滑项
        self.eps = eps
        self.to(DEVICE)

    def forward(self, x):
        # TODO: 请利用init中的成员变量实现LayerNorm层的功能
        # 按最后一个维度计算均值和方差
        mean = x.mean(-1, keepdim=True)
        std = x.std(-1, keepdim=True)

        # TODO: 返回Layer Norm的结果
        return self.a_2 * (x - mean) / torch.sqrt(std ** 2 + self.eps) + self.b_2

class SublayerConnection(nn.Module):
    """
    SublayerConnection的作用就是把Multi-Head Attention和Feed Forward层连在一起
    只不过每一层输出之后都要先做Layer Norm再残差连接
    sublayer是lamda函数
    """
    def __init__(self, size, dropout):
        super(SublayerConnection, self).__init__()
        self.norm = LayerNorm(size)
        self.dropout = nn.Dropout(dropout)
        self.to(DEVICE)

    def forward(self, x, sublayer):
        # TODO: 请利用init中的成员变量实现LayerNorm和残差连接的功能
        # 返回Layer Norm和残差连接后结果
        return x + self.dropout(sublayer(self.norm(x)))

class EncoderLayer(nn.Module):
    def __init__(self, h, d_model, d_ff, dropout):
        super(EncoderLayer, self).__init__()
        self.self_attn = MultiHeadedAttention(h, d_model).to(DEVICE)
        self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout).to(DEVICE)
        # SublayerConnection的作用就是把multi和ffn连在一起
        # 只不过每一层输出之后都要先做Layer Norm再残差连接
        self.sublayer = clones(SublayerConnection(d_model, dropout), 2)
        # d_model
        self.size = d_model
        self.to(DEVICE)

    def forward(self, x, mask):
        # 将embedding层进行Multi head Attention
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, mask))
        # 注意到attn得到的结果x直接作为了下一层的输入
        return self.sublayer[1](x, self.feed_forward)

class Encoder(nn.Module):
    # layer = EncoderLayer
    # N = 6
    def __init__(self, h, d_model, d_ff, dropout, N):
        super(Encoder, self).__init__()
        # 复制N个encoder layer
        layer = EncoderLayer(h, d_model, d_ff, dropout).to(DEVICE)
        self.layers = clones(layer, N)
        # Layer Norm
        self.norm = LayerNorm(d_model).to(DEVICE)
        self.to(DEVICE)

    def forward(self, x, mask):
        """
        使用循环连续eecode N次(这里为6次)
        这里的Eecoderlayer会接收一个对于输入的attention mask处理
        """
        for layer in self.layers:
            x = layer(x, mask)
        return self.norm(x)

class DecoderLayer(nn.Module):
    def __init__(self, h, d_model, d_ff, dropout):
        super(DecoderLayer, self).__init__()
        self.size = d_model
        # Self-Attention
        self.self_attn = MultiHeadedAttention(h, d_model).to(DEVICE)
        # 与Encoder传入的Context进行Attention
        self.src_attn = MultiHeadedAttention(h, d_model).to(DEVICE)
        self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
        self.sublayer = clones(SublayerConnection(d_model, dropout), 3)
        self.to(DEVICE)

    def forward(self, x, memory, src_mask, tgt_mask):
        # 用m来存放encoder的最终hidden表示结果
        m = memory

        # TODO: 参照EncoderLayer完成DecoderLayer的forwark函数
        # Self-Attention：注意self-attention的q，k和v均为decoder hidden
        x = self.sublayer[0](x, lambda x: self.self_attn(x, x, x, tgt_mask))
        # Context-Attention：注意context-attention的q为decoder hidden，而k和v为encoder hidden
        x = self.sublayer[1](x, lambda x: self.src_attn(x, m, m, src_mask))
        return self.sublayer[2](x, self.feed_forward)

class Decoder(nn.Module):
    def __init__(self, h, d_model, d_ff, dropout, N, vocab):
        super(Decoder, self).__init__()
        layer = DecoderLayer(h, d_model, d_ff, dropout)
        # 复制N个encoder layer
        self.layers = clones(layer, N)
        # Layer Norm
        self.norm = LayerNorm(layer.size)
        self.proj = nn.Linear(d_model, vocab)
        self.to(DEVICE)

    def forward(self, x, memory, src_mask, tgt_mask, voc_out=False):
        """
        使用循环连续decode N次(这里为6次)
        这里的Decoderlayer会接收一个对于输入的attention mask处理
        和一个对输出的attention mask + subsequent mask处理
        """
        for layer in self.layers:
            x = layer(x, memory, src_mask, tgt_mask)
        x = self.norm(x)
        if not voc_out:
            return x
        else:
            return F.log_softmax(self.proj(x), dim=-1)

class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout, DEVICE, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        # 初始化一个size为 max_len(设定的最大长度)×embedding维度 的全零矩阵
        # 来存放所有小于这个长度位置对应的porisional embedding
        pe = torch.zeros(max_len, d_model, device=DEVICE)
        # 生成一个位置下标的tensor矩阵(每一行都是一个位置下标)
        """
        形式如：
        tensor([[0.],
                [1.],
                [2.],
                [3.],
                [4.],
                ...])
        """
        position = torch.arange(0., max_len, device=DEVICE).unsqueeze(1)
        # 这里幂运算太多，我们使用exp和log来转换实现公式中pos下面要除以的分母（由于是分母，要注意带负号）
        div_term = torch.exp(torch.arange(0., d_model, 2, device=DEVICE) * -(math.log(10000.0) / d_model))

        # TODO: 根据公式，计算各个位置在各embedding维度上的位置纹理值，存放到pe矩阵中
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)

        # 加1个维度，使得pe维度变为：1×max_len×embedding维度
        # (方便后续与一个batch的句子所有词的embedding批量相加)
        pe = pe.unsqueeze(0)
        # 将pe矩阵以持久的buffer状态存下(不会作为要训练的参数)
        self.register_buffer('pe', pe)
        self.to(DEVICE)

    def forward(self, x):
        # 将一个batch的句子所有词的embedding与已构建好的positional embeding相加
        # (这里按照该批次数据的最大句子长度来取对应需要的那些positional embedding值)
        x = x + Variable(self.pe[:, :x.size(1)], requires_grad=False)
        return self.dropout(x)

class Embeddings(nn.Module):
    def __init__(self, d_model, vocab, DEVICE):
        super(Embeddings, self).__init__()
        # Embedding层
        self.lut = nn.Embedding(vocab, d_model).to(DEVICE)
        # Embedding维数
        self.d_model = d_model
        self.DEVICE = DEVICE
        self.to(DEVICE)

    def forward(self, x):
        # 返回x对应的embedding矩阵（需要乘以math.sqrt(d_model)）
        x = x.to(self.DEVICE)
        return self.lut(x) * math.sqrt(self.d_model)

# class Generator(nn.Module):
#     # vocab: tgt_vocab
#     def __init__(self, d_model, vocab):
#         super(Generator, self).__init__()
#         # decode后的结果，先进入一个全连接层变为词典大小的向量
#         self.proj = nn.Linear(d_model, vocab)
#
#     def forward(self, x):
#         # 然后再进行log_softmax操作(在softmax结果上再做多一次log运算)
#         return F.log_softmax(self.proj(x), dim=-1)


class Transformer(nn.Module):
    def __init__(self, h, d_model, d_ff, dropout, N, src_vocab, tgt_vocab, DEVICE):
        super(Transformer, self).__init__()
        self.encoder = Encoder(h, d_model, d_ff, dropout, N).to(DEVICE)
        self.decoder = Decoder(h, d_model, d_ff, dropout, N, tgt_vocab).to(DEVICE)
        self.src_embed = nn.Sequential(Embeddings(d_model, src_vocab, DEVICE).to(DEVICE), PositionalEncoding(d_model, dropout, DEVICE).to(DEVICE))
        self.tgt_embed = nn.Sequential(Embeddings(d_model, tgt_vocab, DEVICE).to(DEVICE), PositionalEncoding(d_model, dropout, DEVICE).to(DEVICE))
        # self.generator = Generator(d_model, tgt_vocab)

        for p in self.parameters():
            if p.dim() > 1:
                # 这里初始化采用的是nn.init.xavier_uniform
                nn.init.xavier_uniform_(p)
        self.to(DEVICE)

    def encode(self, src, src_mask):
        return self.encoder(self.src_embed(src), src_mask)

    def decode(self, memory, src_mask, tgt, tgt_mask):
        return self.decoder(self.tgt_embed(tgt), memory, src_mask, tgt_mask, voc_out=True)

    def forward(self, src, tgt, src_mask, tgt_mask):
        # encoder的结果作为decoder的memory参数传入，进行decode
        return self.decode(self.encode(src, src_mask), src_mask, tgt, tgt_mask)

模型训练

import time
import torch
import torch.nn as nn
from torch.autograd import Variable
import os

import random
from data_generator import *
import numpy as np
from transformer_models import *
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

class CustomLearnRateOptimizer:
    "Optim wrapper that implements rate."

    def __init__(self, model_size, factor, warmup, model):
        self.optimizer = torch.optim.Adam(model.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9)
        self._step = 0
        self.warmup = warmup
        self.factor = factor
        self.model_size = model_size
        self._rate = 0

    def step(self):
        "Update parameters and rate"
        self._step += 1
        rate = self.rate()
        for p in self.optimizer.param_groups:
            p['lr'] = rate
        self._rate = rate
        self.optimizer.step()

    def rate(self, step=None):
        "Implement `lrate` above"
        if step is None:
            step = self._step
        return self.factor * (self.model_size ** (-0.5) * min(step ** (-0.5), step * self.warmup ** (-1.5)))

class LabelSmoothingLoss(nn.Module):
    """标签平滑处理"""

    def __init__(self, size, padding_idx=0, smoothing=0.0):
        super(LabelSmoothingLoss, self).__init__()
        self.criterion = nn.KLDivLoss(reduction='sum')
        self.padding_idx = padding_idx
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing
        self.size = size
        self.true_dist = None

    def forward(self, x, target):
        assert x.size(1) == self.size
        x = x.to(DEVICE)
        target = target.to(DEVICE)
        true_dist = x.data.clone()

        true_dist.fill_(self.smoothing / (self.size - 2))
        true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
        true_dist[:, self.padding_idx] = 0
        mask = torch.nonzero(target.data == self.padding_idx)
        if mask.dim() > 0:
            true_dist.index_fill_(0, mask.squeeze(), 0.0)
        self.true_dist = true_dist
        return self.criterion(x, Variable(true_dist, requires_grad=False))

class Trainer:
    def __init__(self, epochs, tgt_vocab, D_MODEL, model, SAVE_FILE='model1.pt', MAX_LENGTH=100):
        self.epochs = epochs
        self.loss = LabelSmoothingLoss(tgt_vocab)
        self.optimizer = CustomLearnRateOptimizer(D_MODEL, 1, 2000, model)
        self.SAVE_FILE = SAVE_FILE
        self.MAX_LENGTH = MAX_LENGTH
        self.model = model

    def compute_loss(self, x, y, norm):
        # 计算loss，并且反向传播
        loss = self.loss(x.contiguous().view(-1, x.size(-1)), y.contiguous().view(-1)) / norm
        # print('loss=', loss.shape, loss)
        loss.backward()
        if self.optimizer is not None:
            self.optimizer.step()
            self.optimizer.optimizer.zero_grad()
        return loss.data.item() * norm.float()

    def train(self, data):
        """
        训练并保存模型
        """
        for p in self.model.parameters():
            if p.dim() > 1:
                # 这里初始化采用的是nn.init.xavier_uniform
                nn.init.xavier_uniform_(p)
        # 初始化模型在dev集上的最优Loss为一个较大值
        best_dev_loss = 1e5

        for epoch in range(self.epochs):
            # 模型训练
            self.model.train()
            self.run_epoch(data.train_data, self.model, epoch)
            self.model.eval()

            # 在dev集上进行loss评估
            print('>>>>> Evaluate')
            dev_loss = self.run_epoch(data.dev_data, self.model, epoch, max_step=200)
            print('<<<<< Evaluate loss: %f' % dev_loss)
            # # TODO: 如果当前epoch的模型在dev集上的loss优于之前记录的最优loss则保存当前模型，并更新最优loss值
            if dev_loss < best_dev_loss:
                print('saving model...')
                torch.save(self.model.state_dict(), self.SAVE_FILE)
                best_dev_loss = dev_loss
            print('>>>>>>>>>>>>>>> Evaluate case, epach=%d' % epoch)
            self.evaluate(data)
            print('>>>>>>>>>>>>>>> end evaluate case, epach=%d' % epoch)

    def run_epoch(self, data, model, epoch, max_step=200):
        start = time.time()
        total_tokens = 0.
        total_loss = 0.
        tokens = 0.
        # print('data=', len(data), data)
        for i, batch in enumerate(data):
            if i>=max_step:
                break
            out = model(batch.src, batch.trg, batch.src_mask, batch.trg_mask)
            loss = self.compute_loss(out, batch.trg_y, batch.ntokens)

            total_loss += loss
            total_tokens += batch.ntokens
            tokens += batch.ntokens

            if i % 100 == 1:
                elapsed = time.time() - start
                print("Epoch %d Batch: %d Loss: %f Tokens per Sec: %fs" % (epoch, i - 1, loss / batch.ntokens, (tokens.float() / elapsed / 1000.)))
                start = time.time()
                tokens = 0

        return total_loss / total_tokens




    def evaluate(self, data):
        """
        在data上用训练好的模型进行预测，打印模型翻译结果
        """
        # 梯度清零
        with torch.no_grad():
            # 在data的英文数据长度上遍历下标
            for _ in range(10):
                i = random.randint(0, len(data.dev_en) - 1)
                # print('i=', i)
                # TODO: 打印待翻译的src句子
                cn_sent = " ".join([data.cn_index_dict[w] for w in data.dev_cn[i]])
                print("\n" + cn_sent)

                # TODO: 打印对应的句子答案
                en_sent = " ".join([data.en_index_dict[w] for w in data.dev_en[i]])
                print("".join(en_sent))

                # 将当前以单词id表示的英文句子数据转为tensor，并放如DEVICE中
                src = torch.from_numpy(np.array(data.dev_cn[i])).long().to(DEVICE)
                # 增加一维
                src = src.unsqueeze(0)
                # 设置attention mask
                src_mask = (src != 0).unsqueeze(-2)
                # 用训练好的模型进行decode预测
                out = self.greedy_decode(src, src_mask, max_len=self.MAX_LENGTH, start_symbol=data.en_word_dict["BOS"])
                # 初始化一个用于存放模型翻译结果句子单词的列表
                translation = []
                # 遍历翻译输出字符的下标（注意：开始符"BOS"的索引0不遍历）
                for j in range(1, out.size(1)):
                    # 获取当前下标的输出字符
                    sym = data.en_index_dict[out[0, j].item()]
                    # 如果输出字符不为'EOS'终止符，则添加到当前句子的翻译结果列表
                    if sym != 'EOS':
                        translation.append(sym)
                    # 否则终止遍历
                    else:
                        break
                # 打印模型翻译输出的中文句子结果
                print("translation: %s" % " ".join(translation))

    def greedy_decode(self, src, src_mask, max_len, start_symbol):
        """
        传入一个训练好的模型，对指定数据进行预测
        """
        # 先用encoder进行encode
        memory = self.model.encode(src, src_mask)
        # print('src=', src.shape, src)
        # print('src_mask=', src_mask.shape, src_mask)
        # print('memory=', memory.shape, memory)
        # 初始化预测内容为1×1的tensor，填入开始符('BOS')的id，并将type设置为输入数据类型(LongTensor)
        ys = torch.ones(1, 1).fill_(start_symbol).type_as(src.data)
        # print('ys=', ys.shape, ys)
        # 遍历输出的长度下标
        for i in range(max_len - 1):
            # decode得到隐层表示
            tgt_msk = subsequent_mask(ys.size(1)).type_as(src.data)
            # print('i=',i,'tgt_msk=', tgt_msk)
            out = self.model.decode(memory,
                               src_mask,
                               Variable(ys),
                               Variable(tgt_msk))

            # 获取当前位置最大概率的预测词id

            _, next_word = torch.max(out, dim=2)
            # print('i = ', i, ', out=', out.shape, out)
            # print('i = ', i, ', next_word=', next_word.shape, next_word)
            next_word = next_word.data[0, 0]
            # 将当前位置预测的字符id与之前的预测内容拼接起来
            ys = torch.cat([ys, torch.ones(1, 1).type_as(src.data).fill_(next_word)], dim=1)
            # print('i = ', i, ' ,next_word=', next_word, ', ys=', ys)
        return ys

if __name__=='__main__':
    TRAIN_FILE = '../datas/eng-cmn-train.txt'
    DEV_FILE = '../datas/eng-cmn-eval.txt'
    TEST_FILE = ''
    SAVE_FILE = 'model.pt'
    h = 8
    d_model = 512
    d_ff = 2048
    dropout = 0.1
    N = 6
    epochs = 52
    batch_size = 64


    print('<<<<<<<<<<<<<<>>>>>>>>>>>>>>>>>>>>\n1.start prepare data++++++++++++++++++++++\n')
    data = PrepareData(TRAIN_FILE, DEV_FILE, batch_size, DEVICE)
    print("src_vocab %d" % data.src_vocab)
    print("tgt_vocab %d" % data.tgt_vocab)
    print('self.train_en.len = ', len(data.train_en))
    print('self.dev_en.len = ', len(data.dev_en))

    print('\n2.start to train++++++++++++++++++++++++\n')
    transformer = Transformer(h, d_model, d_ff, dropout, N, data.src_vocab, data.tgt_vocab, DEVICE).to(DEVICE)
    trainer = Trainer(epochs, data.tgt_vocab, d_model, transformer, SAVE_FILE='model.pt', MAX_LENGTH=20)
    trainer.train(data)

    print('\n3.start to eval++++++++++++++++++++++++\n')
    transformer.load_state_dict(torch.load(SAVE_FILE))
    # trainer = Trainer(epochs, data.tgt_vocab, d_model, transformer, SAVE_FILE='model1.pt', MAX_LENGTH=5)
    trainer.evaluate(data)