ELMo代码详解(一):数据准备

ELMo代码解读笔记


1.数据准备

  数据准备包括:1.生成word的词汇表类; 2.生成字符的词汇表类; 3.以word-ids作为输入的训练batch生成类; 4.以char-ids作为输入的训练batch生成类; 5.生成语言模型输入的数据集类

1.1 word词汇表类(Vocabulary)

  根据一个词汇表文件,生成word和索引的相互对应关系,即_id_to_word和_word_to_id,前者是一个数组,后者是一个字典。当然,我们也需要加上一个特殊的词,比如, (分别表示句首,句尾和不知词)。主要的代码如下:

def __init__(self, filename, validate_file=False):
    '''
    filename = the vocabulary file.  It is a flat text file with one
        (normalized) token per line.  In addition, the file should also
        contain the special tokens , ,  (case sensitive).
        vocab文件,是一个纯文本,每一行只有一个词。另外,这个文件应该包含特殊词,
        比如, , 等
    '''
    self._id_to_word = []
    self._word_to_id = {}
    self._unk = -1
    self._bos = -1
    self._eos = -1

    with open(filename) as f:
        idx = 0
        for line in f:                                                        #词汇表中一行就是一个单词
            word_name = line.strip()
            if word_name == '':
                self._bos = idx
            elif word_name == '':
                self._eos = idx
            elif word_name == '':
                self._unk = idx
            if word_name == '!!!MAXTERMID':
                continue

            self._id_to_word.append(word_name)
            self._word_to_id[word_name] = idx
            idx += 1

    # check to ensure file has special tokens
    if validate_file:
        if self._bos == -1 or self._eos == -1 or self._unk == -1:
            raise ValueError("Ensure the vocabulary file has "
                             ", ,  tokens")

当然,类中还有两个很实用的函数,一个是编码函数encode,另一个是解码函数decode。编码器encode的作用是将一条句子sentence转化为一个word-ids列表,注意要加上句首和句尾token。当然包括反转选项,用来做双向的LSTM。而解码器decode就是将word-ids列表转化为相应的单词。

def encode(self, sentence, reverse=False, split=True):
    """Convert a sentence to a list of ids, with special tokens added.
    Sentence is a single string with tokens separated by whitespace.

    If reverse, then the sentence is assumed to be reversed, and
        this method will swap the BOS/EOS tokens appropriately.
       将一个sentenct转化为ids序列
       并提供句子反转的功能
    """

    if split:
        word_ids = [
            self.word_to_id(cur_word) for cur_word in sentence.split()
        ]
    else:
        word_ids = [self.word_to_id(cur_word) for cur_word in sentence]

    if reverse:
        return np.array([self.eos] + word_ids + [self.bos], dtype=np.int32) #在每一条句子首位加上了
    else:
        return np.array([self.bos] + word_ids + [self.eos], dtype=np.int32)

def decode(self, cur_ids):
    """Convert a list of ids to a sentence, with space inserted.
       将一个ids序列转化为word序列
    """
    return ' '.join([self.id_to_word(cur_id) for cur_id in cur_ids])

1.2 字符词汇表(UnicodeCharsVocabulary)

  注意这个类是上面word词汇表Vocabulary的子类,这意味着这个字符类包含了Vocabulary的所有变量和方法!
  每个字符(character)的id是用该字符对应的utf-8编码,这样也就可以形成id和char之间的转换,因为使用utf-8编码,这将限制char词汇表中所有可能的id数量为256。当然,我们也需要加入5个额外的特殊字符,包括:句首,句尾,词头,词尾和padding。通过词汇表文件,形成字符词汇表的_word_char_ids的代码为:

#将词转化为char_ids
def _convert_word_to_char_ids(self, word):
    code = np.zeros([self.max_word_length], dtype=np.int32)
    code[:] = self.pad_char

    #将word中每一个字符转化为utf-8编码,然后用数组存起来,例如:
    #english中,e:101, n:110, g:103, l:108, h:105, s:115, h:104
    word_encoded = word.encode('utf-8', 'ignore')[:(self.max_word_length-2)]
    code[0] = self.bow_char                                      #加上词开始和结尾的编码
    for k, chr_id in enumerate(word_encoded, start=1):
        code[k] = chr_id
    code[k + 1] = self.eow_char

    return code


def __init__(self, filename, max_word_length, **kwargs):
    #调用父类Vocabulary,生成word和id之间的转换等
    super(UnicodeCharsVocabulary, self).__init__(filename, **kwargs)
    self._max_word_length = max_word_length                             #每个词对应最大字符长

    # char ids 0-255 come from utf-8 encoding bytes
    # assign 256-300 to special chars
    self.bos_char = 256  # 
    self.eos_char = 257  # 
    self.bow_char = 258  # 
    self.eow_char = 259  # 
    self.pad_char = 260 # 

    num_words = len(self._id_to_word)                                   #单词的个数,父类中的属性

    #每个词都会对应一个char_ids列表
    self._word_char_ids = np.zeros([num_words, max_word_length],
        dtype=np.int32)

    # the charcter representation of the begin/end of sentence characters
    # 对句首或者句尾的token来一个字符的表示
    def _make_bos_eos(c):
        r = np.zeros([self.max_word_length], dtype=np.int32)
        r[:] = self.pad_char
        r[0] = self.bow_char                                            #词的开始
        r[1] = c
        r[2] = self.eow_char                                            #词的结束
        return r
    self.bos_chars = _make_bos_eos(self.bos_char)                       #句子开始对应的char_ids
    self.eos_chars = _make_bos_eos(self.eos_char)                       #句子的结尾对应的char_ids

    for i, word in enumerate(self._id_to_word):                         #遍历id2word数组,得到每一个词的char_ids
        self._word_char_ids[i] = self._convert_word_to_char_ids(word)

    self._word_char_ids[self.bos] = self.bos_chars                      #将句子开头和结尾当作一个word处理
    self._word_char_ids[self.eos] = self.eos_chars

通过以上两个函数,我们就可以得到每个单词(word)对应的字符id序列(char-ids),包括句首和句尾的字符id序列表示。
  这个类还提供将句子转化为相应的char-ids数组的功能,它首先查词汇表字典_word_char_ids来得到每个词的char_ids表示,然后组成句子,返回的是一个二维数组。实现如下:

#返回word对应的char_ids数组
def word_to_char_ids(self, word):
    if word in self._word_to_id:
        return self._word_char_ids[self._word_to_id[word]]
    else:
        return self._convert_word_to_char_ids(word)

def encode_chars(self, sentence, reverse=False, split=True):
    '''
    Encode the sentence as a white space delimited string of tokens.
    对一整句话进行编码,编码成chars
    '''
    if split:                                                             #如果切割了句子
        chars_ids = [self.word_to_char_ids(cur_word)       
                 for cur_word in sentence.split()]
    else:
        chars_ids = [self.word_to_char_ids(cur_word)
                 for cur_word in sentence]
    if reverse:
        return np.vstack([self.eos_chars] + chars_ids + [self.bos_chars]) #在每一条句子上都加了       
    else:
        return np.vstack([self.bos_chars] + chars_ids + [self.eos_chars])

1.3 生成word-ids输入的batch类(TokenBatcher)

  将一个batch的句子文本转化为相应的word-ids形式。主要代码如下:

def batch_sentences(self, sentences: List[List[str]]):
    '''
    Batch the sentences as character ids
    确定是character_ids?而不是word_ids
    Each sentence is a list of tokens without  or , e.g.
    [['The', 'first', 'sentence', '.'], ['Second', '.']]
    '''
    n_sentences = len(sentences)
    max_length = max(len(sentence) for sentence in sentences) + 2

    X_ids = np.zeros((n_sentences, max_length), dtype=np.int64)          #word_ids是二维的,[batch_size, max_len]

    for k, sent in enumerate(sentences):
        length = len(sent) + 2
        ids_without_mask = self._lm_vocab.encode(sent, split=False)
        # add one so that 0 is the mask value
        X_ids[k, :length] = ids_without_mask + 1                         #0表示mask值

    return X_ids

1.4 生成char-ids输入的类(Batcher)

  和上面类似,只是这里生成的是一个batch的句子文本的char-ids的表示,形成的是一个三维数组。主要代码为:

def batch_sentences(self, sentences: List[List[str]]):
    '''
    Batch the sentences as character ids
    Each sentence is a list of tokens without  or , e.g.
    [['The', 'first', 'sentence', '.'], ['Second', '.']]
    '''
    n_sentences = len(sentences)                                      #句子个数
    max_length = max(len(sentence) for sentence in sentences) + 2     #句子最大长度,加上句首和句尾?

    X_char_ids = np.zeros(                                            #三维数组,每条句子中每个单词对应的char_ids数组
        (n_sentences, max_length, self._max_token_length),
        dtype=np.int64
    )

    #遍历数组
    for k, sent in enumerate(sentences):
        length = len(sent) + 2
        char_ids_without_mask = self._lm_vocab.encode_chars(          #对每个sentence得到char_ids数组
            sent, split=False)
        # add one so that 0 is the mask value, 加上1,所以0是mask值
        X_char_ids[k, :length, :] = char_ids_without_mask + 1         #直接复制粘贴?将对应值加1,其他值填0

    return X_char_ids

  接着定义了一个生成各种数据的batch的方法,该方法每次从输入中读取一个batch的数据,batch中每个数据条目就是一条句子,每个条目包括句子的word-ids表示,char-ids表示和targets(即句子每个词要预测的下一个词)。该方法中有一个生成器(generator),每次会产生一条句子的数据,包括句子的word-ids和char-ids表示,所有只要重复调用该generator的next方法batch_size次就能够构造出一个batch的数据,代码如下:

def _get_batch(generator, batch_size, num_steps, max_word_length):
"""Read batches of input.
   都一个batch的输入
"""
cur_stream = [None] * batch_size                                         #None表示任意大小

no_more_data = False
while True:
    inputs = np.zeros([batch_size, num_steps], np.int32)                 #batch中word_ids          
    if max_word_length is not None:                                      #batch中每条句子每个word对应的char_ids
        char_inputs = np.zeros([batch_size, num_steps, max_word_length],
                            np.int32)
    else:
        char_inputs = None
    targets = np.zeros([batch_size, num_steps], np.int32)                #我们的目标是预测下一个词来优化emlo,所以我们以向右滑动的1个词作为target

    for i in range(batch_size):                                          #每一条句子
        cur_pos = 0                                                      #这个值?

        while cur_pos < num_steps:                                       #循环是不是有点多余, 毫无意义
            if cur_stream[i] is None or len(cur_stream[i][0]) <= 1:
                try:
                    cur_stream[i] = list(next(generator))                #一个生成器一次只生成一条句子信息
                except StopIteration:
                    # No more data, exhaust current streams and quit     
                    no_more_data = True
                    break
            #感觉cur_stream是这样一个东西,[i][0]代表的是word_ids,[i][1]代表的是char_ids?
            #你的猜测是完全正确的,num_steps是一个窗口大小吗?
            #所以下面的一次是,读一个窗口的数据?
            how_many = min(len(cur_stream[i][0]) - 1, num_steps - cur_pos)
            next_pos = cur_pos + how_many

            inputs[i, cur_pos:next_pos] = cur_stream[i][0][:how_many]
            if max_word_length is not None:
                char_inputs[i, cur_pos:next_pos] = cur_stream[i][1][
                                                                :how_many]
            targets[i, cur_pos:next_pos] = cur_stream[i][0][1:how_many+1]     #后一个词是预测对象

            cur_pos = next_pos

            cur_stream[i][0] = cur_stream[i][0][how_many:]                    #cur_stream也跟着往后移动?
            if max_word_length is not None:
                cur_stream[i][1] = cur_stream[i][1][how_many:]

    if no_more_data:
        # There is no more data.  Note: this will not return data
        # for the incomplete batch
        break

    X = {'token_ids': inputs, 'tokens_characters': char_inputs,
             'next_token_id': targets}

    yield X

1.5 语言模型的数据集类(LMDataset)

  数据集类为语言模型训练提供相应的数据输入。它是随机的从数据文件列表中选取一个文件(数据不是仅仅在一个文件里面,而是很多文件),一次读取所有数据到内存中,然后提供一个句子生成器,再调用上面定义的_get_batch()函数来每次产生一个batch的数据集。具体实现代码如下:

def get_sentence(self):
    """
    构造一个生成器吗?
    """
    while True:
        if self._i == self._nids:
            self._ids = self._load_random_shard()                             #重新加载文件读取
        ret = self._ids[self._i]                                              #一次仅仅训练一条句子?
        self._i += 1
        yield ret


def iter_batches(self, batch_size, num_steps):
    """一个生成数据的迭代器"""
    for X in _get_batch(self.get_sentence(), batch_size, num_steps,
                       self.max_word_length):

        # token_ids = (batch_size, num_steps)
        # char_inputs = (batch_size, num_steps, 50) of character ids
        # targets = word ID of next word (batch_size, num_steps)
        yield X

  上面的语言模型只是普通的语言模型的输入,为了构建双向的LSTM模型,我们得将正常的数据反转,得到反向LSTM的输入。于是有了BidirectionalLMDataset类,其核心代码如下:

def __init__(self, filepattern, vocab, test=False, shuffle_on_load=False):
    '''
    bidirectional version of LMDataset
    前向的LSTM传播过程数据正常取
    反向的LSTM传播过程只需要将数据反转就好了
    '''
    self._data_forward = LMDataset(                                            #正向数据集
        filepattern, vocab, reverse=False, test=test,
        shuffle_on_load=shuffle_on_load)
    self._data_reverse = LMDataset(
        filepattern, vocab, reverse=True, test=test,                           #反向数据集
        shuffle_on_load=shuffle_on_load)


def iter_batches(self, batch_size, num_steps):
    """
    将二者合成一个数据集?
    """
    max_word_length = self._data_forward.max_word_length

    for X, Xr in zip(
        _get_batch(self._data_forward.get_sentence(), batch_size,
                  num_steps, max_word_length),
        _get_batch(self._data_reverse.get_sentence(), batch_size,
                  num_steps, max_word_length)
        ):

        for k, v in Xr.items():                                               #都合并到X中去
            #形成token_ids_reverse, token_characters_reverse等
            X[k + '_reverse'] = v                                             

        yield X

你可能感兴趣的:(Python,NLP)