NLP学习3——Word2Vec训练技巧：负采样

负采样是一种训练技巧，通过观察上一节中的Skip-Gram和CBOW的目标函数就会发现，公式的最后都有一个求和项，而求和的范围是词汇表中的所有词，如果词汇表非常巨大，那么在对目标函数进行反向传播时的计算量会非常大。而通过负采样可以减少输入维度，对于Skip-Gram模型来说就是选定几个中心词之外的词和中心词一起训练，这比整个词汇表一起训练的计算量要小很多。
想了解具体公式推导的读者可以参考这篇博客
下面给出代码实现：

import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.optim as optim
import torch.nn.functional as F
import nltk
import random
import numpy as np
from collections import Counter
flatten = lambda l: [item for sublist in l for item in sublist]
random.seed(1024)

USE_CUDA = torch.cuda.is_available()
gpus = [0]
torch.cuda.set_device(gpus[0])

FloatTensor = torch.cuda.FloatTensor if USE_CUDA else torch.FloatTensor
LongTensor = torch.cuda.LongTensor if USE_CUDA else torch.LongTensor
ByteTensor = torch.cuda.ByteTensor if USE_CUDA else torch.ByteTensor

def getBatch(batch_size, train_data):
    random.shuffle(train_data)
    sindex = 0
    eindex = batch_size
    while eindex < len(train_data):
        batch = train_data[sindex: eindex]
        temp = eindex
        eindex = eindex + batch_size
        sindex = temp
        yield batch
    
    if eindex >= len(train_data):
        batch = train_data[sindex:]
        yield batch

def prepare_sequence(seq, word2index):
    idxs = list(map(lambda w: word2index[w] if word2index.get(w) is not None else word2index["<UNK>"], seq))
    return Variable(LongTensor(idxs))

def prepare_word(word, word2index):
    return Variable(LongTensor([word2index[word]]) if word2index.get(word) is not None else LongTensor([word2index["<UNK>"]]))

corpus = list(nltk.corpus.gutenberg.sents('melville-moby_dick.txt'))[:500]
corpus = [[word.lower() for word in sent] for sent in corpus]
# 以上加载数据集和一些函数的内容和上一节一样
# 构造停用词
word_count = Counter(flatten(corpus))
MIN_COUNT = 3
exclude = []
for w, c in word_count.items():
    if c < MIN_COUNT:
        exclude.append(w)
# 构造训练集的方式也和上一节一样
vocab = list(set(flatten(corpus)) - set(exclude))    
word2index = {}
for vo in vocab:
    if word2index.get(vo) is None:
        word2index[vo] = len(word2index)
        
index2word = {v:k for k, v in word2index.items()}
WINDOW_SIZE = 5
windows =  flatten([list(nltk.ngrams(['<DUMMY>'] * WINDOW_SIZE + c + ['<DUMMY>'] * WINDOW_SIZE, WINDOW_SIZE * 2 + 1)) for c in corpus])
train_data = []
for window in windows:
    for i in range(WINDOW_SIZE * 2 + 1):
        if window[i] in exclude or window[WINDOW_SIZE] in exclude: 
            continue # min_count
        if i == WINDOW_SIZE or window[i] == '<DUMMY>': 
            continue
        train_data.append((window[WINDOW_SIZE], window[i]))
X_p = []
y_p = []
for tr in train_data:
    X_p.append(prepare_word(tr[0], word2index).view(1, -1))
    y_p.append(prepare_word(tr[1], word2index).view(1, -1))
# 下面介绍选择负样本词的方法
# 简单来说就是计算一个词在词汇表中的权重，公式：(word_count/num_total_words)**0.75
# 具体情况可以参考文章开头给出的博客  
Z = 0.001    
word_count = Counter(flatten(corpus))
num_total_words = sum([c for w, c in word_count.items() if w not in exclude])
unigram_table = []
for vo in vocab:
    unigram_table.extend([vo] * int(((word_count[vo]/num_total_words)**0.75)/Z))
# 定义负采样函数，该函数返回选择出来的负样本词
def negative_sampling(targets, unigram_table, k):
    batch_size = targets.size(0)
    neg_samples = []
    for i in range(batch_size):
        nsample = []
        target_index = targets[i].data.cpu().tolist()[0] if USE_CUDA else targets[i].data.tolist()[0]
        while len(nsample) < k:   # num of sampling
            neg = random.choice(unigram_table)
            if word2index[neg] == target_index:
                continue
            nsample.append(neg)
        neg_samples.append(prepare_sequence(nsample, word2index).view(1, -1))
    
    return torch.cat(neg_samples)
# 定义模型  
class SkipgramNegSampling(nn.Module):
    
    def __init__(self, vocab_size, projection_dim):
        super(SkipgramNegSampling, self).__init__()
        self.embedding_v = nn.Embedding(vocab_size, projection_dim) # center embedding
        self.embedding_u = nn.Embedding(vocab_size, projection_dim) # out embedding
        self.logsigmoid = nn.LogSigmoid()
                
        initrange = (2.0 / (vocab_size + projection_dim))**0.5 # Xavier init
        self.embedding_v.weight.data.uniform_(-initrange, initrange) # init
        self.embedding_u.weight.data.uniform_(-0.0, 0.0) # init
        
    def forward(self, center_words, target_words, negative_words):
        center_embeds = self.embedding_v(center_words) # B x 1 x D
        target_embeds = self.embedding_u(target_words) # B x 1 x D
        
        neg_embeds = -self.embedding_u(negative_words) # B x K x D
        
        positive_score = target_embeds.bmm(center_embeds.transpose(1, 2)).squeeze(2) # Bx1
        negative_score = torch.sum(neg_embeds.bmm(center_embeds.transpose(1, 2)).squeeze(2), 1).view(negs.size(0), -1) # BxK -> Bx1
        
        loss = self.logsigmoid(positive_score) + self.logsigmoid(negative_score)
        
        return -torch.mean(loss)
    
    def prediction(self, inputs):
        embeds = self.embedding_v(inputs)
        
        return embeds 
EMBEDDING_SIZE = 30 
BATCH_SIZE = 256
EPOCH = 100
NEG = 10 # Num of Negative Sampling
model = SkipgramNegSampling(len(word2index), EMBEDDING_SIZE)
if USE_CUDA:
    model = model.cuda()
optimizer = optim.Adam(model.parameters(), lr=0.001)
for epoch in range(EPOCH):
    for i,batch in enumerate(getBatch(BATCH_SIZE, train_data)):
        
        inputs, targets = zip(*batch)
        
        inputs = torch.cat(inputs) # B x 1
        targets = torch.cat(targets) # B x 1
        negs = negative_sampling(targets, unigram_table, NEG)
        model.zero_grad()

        loss = model(inputs, targets, negs)
        
        loss.backward()
        optimizer.step()
    
        losses.append(loss.data.tolist()[0])
    if epoch % 10 == 0:
        print("Epoch : %d, mean_loss : %.02f" % (epoch, loss))
# 计算余弦相似度
def word_similarity(target, vocab):
    if USE_CUDA:
        target_V = model.prediction(prepare_word(target, word2index))
    else:
        target_V = model.prediction(prepare_word(target, word2index))
    similarities = []
    for i in range(len(vocab)):
        if vocab[i] == target: 
            continue
        
        if USE_CUDA:
            vector = model.prediction(prepare_word(list(vocab)[i], word2index))
        else:
            vector = model.prediction(prepare_word(list(vocab)[i], word2index))
        
        cosine_sim = F.cosine_similarity(target_V, vector).data.tolist()[0]
        similarities.append([vocab[i], cosine_sim])
    return sorted(similarities, key=lambda x: x[1], reverse=True)[:10]