导入包:
1 import torch
2 import torch.nn as nn
3 import torch.nn.functional as F
4 import torch.utils.data as tud
5
6 from torch.nn.parameter import Parameter #参数更新和优化函数
7
8 from collections import Counter #Counter 计数器
9 import numpy as np
10 import random
11 import math
12
13 import pandas as pd
14 import scipy #SciPy是基于NumPy开发的高级模块,它提供了许多数学算法和函数的实现
15 import sklearn
16 from sklearn.metrics.pairwise import cosine_similarity #余弦相似度函数
参数设定:
1 USE_CUDA = torch.cuda.is_available() #有GPU可以用
2
3 # 为了保证实验结果可以复现,我们经常会把各种random seed固定在某一个值
4 random.seed(53113)
5 np.random.seed(53113)
6 torch.manual_seed(53113)
7 if USE_CUDA:
8 torch.cuda.manual_seed(53113)
9
10 # 设定一些超参数
11 K = 100 # number of negative samples 负样本随机采样数量
12 C = 3 # nearby words threshold 指定周围三个单词进行预测
13 NUM_EPOCHS = 2 # The number of epochs of training 迭代轮数
14 MAX_VOCAB_SIZE = 30000 # the vocabulary size 词汇表多大
15 BATCH_SIZE = 128 # the batch size 每轮迭代1个batch的数量
16 LEARNING_RATE = 0.2 # the initial learning rate #学习率
17 EMBEDDING_SIZE = 100 #词向量维度
18
19
20 LOG_FILE = "word-embedding.log"
21
22 # tokenize函数,把一篇文本转化成一个个单词
23 def word_tokenize(text):
24 return text.split()
1.创建vocabulary
- 从文本文件中读取所有的文字,通过这些文本创建一个vocabulary;
- 由于单词数量可能太大,我们只选取最常见的MAX_VOCAB_SIZE个单词;
- 我们添加一个UNK单词表示所有不常见的单词;
- 我们需要记录单词到index的mapping,以及index到单词的mapping,单词的count,单词的(normalized) frequency,以及单词总数。
1 with open("sample_data/text8.train.txt", "r") as fin: #读入文件
2 text = fin.read()
1 text = [w for w in word_tokenize(text.lower())] #分词,在这里类似于text.split()
2
3 vocab = dict(Counter(text).most_common(MAX_VOCAB_SIZE-1)) #字典格式,把(MAX_VOCAB_SIZE-1)个最频繁出现的单词取出来,-1是留给不常见的单词
4 vocab["<unk>"] = len(text) - np.sum(list(vocab.values())) #不常见单词数=总单词数-常见单词数,这里计算的vocab["<unk>"]=29999
5
6 idx_to_word = [word for word in vocab.keys()] #取出字典的所有单词key
7 word_to_idx = {word:i for i, word in enumerate(idx_to_word)} #取出所有单词和对应的索引,索引值与单词出现次数相反,最常见单词索引为0。
8
9 word_counts = np.array([count for count in vocab.values()], dtype=np.float32) #所有单词的频数values
10 word_freqs = word_counts / np.sum(word_counts) #所有单词的频率
11
12 word_freqs = word_freqs ** (3./4.) #论文里频率乘以3/4次方
13 word_freqs = word_freqs / np.sum(word_freqs) #被选作negative sampling的单词概率
14
15 VOCAB_SIZE = len(idx_to_word) #词汇表单词数30000=MAX_VOCAB_SIZE
2.实现Dataloader
一个dataloader需要以下内容:
- 把所有text编码成数字,然后用subsampling预处理这些文字。
- 保存vocabulary,单词count,normalized word frequency
- 每个iteration sample一个中心词
- 根据当前的中心词返回context单词
- 根据中心词sample一些negative单词
- 返回单词的counts
这里有一个好的tutorial介绍如何使用PyTorch dataloader. 为了使用dataloader,我们需要定义以下两个function:
-
__len__
function需要返回整个数据集中有多少个item -
__get__
根据给定的index返回一个item
有了dataloader之后,我们可以轻松随机打乱整个数据集,拿到一个batch的数据等等。
1 class WordEmbeddingDataset(tud.Dataset): #tud.Dataset父类
2 def __init__(self, text, word_to_idx, idx_to_word, word_freqs, word_counts):
3 ''' text: a list of words, all text from the training dataset
4 word_to_idx: the dictionary from word to idx
5 idx_to_word: idx to word mapping
6 word_freq: the frequency of each word
7 word_counts: the word counts
8 '''
9 super(WordEmbeddingDataset, self).__init__() #初始化模型
10 self.text_encoded = [word_to_idx.get(t, VOCAB_SIZE-1) for t in text] #取出text里每个单词word_to_idx字典里对应的索引,不在字典里返回"<unk>"的索引,get括号里第二个参数应该写word_to_idx["<unk>"]
11 self.text_encoded = torch.LongTensor(self.text_encoded) #变成Longtensor类型
12
13 self.word_to_idx = word_to_idx #以下皆为保存数据
14 self.idx_to_word = idx_to_word
15 self.word_freqs = torch.Tensor(word_freqs)
16 self.word_counts = torch.Tensor(word_counts)
17
18 def __len__(self):
19 ''' 返回整个数据集(所有单词)的长度
20 '''
21 return len(self.text_encoded)
22
23 def __getitem__(self, idx): #这里__getitem__函数是个迭代器,idx代表了所有的单词索引
24 ''' 这个function返回以下数据用于训练
25 - 中心词
26 - 这个单词附近的(positive)单词
27 - 随机采样的K个单词作为negative sample
28 '''
29 center_word = self.text_encoded[idx] #中心词索引
30
31 pos_indices = list(range(idx-C, idx)) + list(range(idx+1, idx+C+1)) #除中心词外,周围词的索引,比如idx=0时,pos_indices = [-3, -2, -1, 1, 2, 3]
32 pos_indices = [i%len(self.text_encoded) for i in pos_indices] #idx可能超出词汇总数,需要取余
33 pos_words = self.text_encoded[pos_indices] #周围词索引,是正例单词
34
35 #负例采样单词索引,torch.multinomial作用是对self.word_freqs做K * pos_words.shape[0](正确单词数量)次取值,输出的是self.word_freqs对应的下标
36 #取样方式采用有放回的采样,并且self.word_freqs数值越大,取样概率越大
37 neg_words = torch.multinomial(self.word_freqs, K * pos_words.shape[0], True)
38
39 return center_word, pos_words, neg_words
创建dataset和dataloader
1 dataset = WordEmbeddingDataset(text, word_to_idx, idx_to_word, word_freqs, word_counts)
2 dataloader = tud.DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4)
测试一下dataloader的内容(太大,显示shape吧)
1 for i, (input_labels, pos_labels, neg_labels) in enumerate(dataloader):
2 print(input_labels.shape, pos_labels.shape, neg_labels.shape)
3 break
torch.Size([128]) torch.Size([128, 6]) torch.Size([128, 600])
3.Pytorch模型
1 class EmbeddingModel(nn.Module):
2 def __init__(self, vocab_size, embed_size):
3 ''' 初始化输入和输出
4 '''
5 super(EmbeddingModel, self).__init__()
6 self.vocab_size = vocab_size #30000
7 self.embed_size = embed_size #100
8
9 initrange = 0.5 / self.embed_size
10 self.out_embed = nn.Embedding(self.vocab_size, self.embed_size, sparse=False) #模型输出nn.Embedding(30000, 100)
11 self.out_embed.weight.data.uniform_(-initrange, initrange) #权重初始化的一种方法
12
13 self.in_embed = nn.Embedding(self.vocab_size, self.embed_size, sparse=False) #模型输入nn.Embedding(30000, 100)
14 self.in_embed.weight.data.uniform_(-initrange, initrange)
15
16
17 def forward(self, input_labels, pos_labels, neg_labels): #输入为tud.DataLoader()返回的三个数据
18 '''
19 input_labels: [batch_size],中心词
20 pos_labels: [batch_size, (C * 2)],C=window_size,中心词周围context window出现过的单词
21 neg_labelss: [batch_size, (C * 2 * K)],中心词周围没有出现过的单词,从negative sampling得到
22
23 return: loss, [batch_size]
24 '''
25 input_embedding = self.in_embed(input_labels) #[batch_size, embed_size],(128,30000)*(30000,100)= 128 * 100
26 pos_embedding = self.out_embed(pos_labels) #[batch_size, 2*C, embed_size],增加了维度(2*C),表示一个batch有B组周围词单词,一组周围词有(2*C)个单词,每个单词有embed_size个维度
27 neg_embedding = self.out_embed(neg_labels) #[batch_size, 2*C*K, embed_size],增加了维度(2*C*K)
28
29 #torch.bmm()为batch间的矩阵相乘(b,n.m)*(b,m,p)=(b,n,p)
30 pos_dot = torch.bmm(pos_embedding, input_embedding.unsqueeze(2)).squeeze() #input_embedding.unsqueeze(2)的维度[batch_size, embed_size, 1],调用bmm之后[batch, 2*C, 1],再压缩掉最后一维
31 neg_dot = torch.bmm(neg_embedding, -input_embedding.unsqueeze(2)).squeeze() #[batch_size, 2*C*K]
32
33 #下面loss计算就是论文里的公式
34 log_pos = F.logsigmoid(pos_dot).sum(1) #batch_size
35 log_neg = F.logsigmoid(neg_dot).sum(1)
36 loss = log_pos + log_neg
37
38 return -loss
39
40 def input_embeddings(self): #取出self.in_embed数据参数
41 return self.in_embed.weight.data.cpu().numpy()
详细见的negSamplingLossAndGradient函数部分。
定义一个模型以及把模型移动到GPU
1 model = EmbeddingModel(VOCAB_SIZE, EMBEDDING_SIZE) #得到model,有参数,有loss,可以优化了
2
3 if USE_CUDA:
4 model = model.cuda()
4.评估模型
评估的文件类似如下结构(word1 word2 相似度分值):
1 def evaluate(filename, embedding_weights):
2 if filename.endswith(".csv"):
3 data = pd.read_csv(filename, sep=",")
4 else:
5 data = pd.read_csv(filename, sep="\t")
6 human_similarity = []
7 model_similarity = []
8 for i in data.iloc[:, 0:2].index: #data.iloc[:, 0:2]取所有行索引为0、1的数据
9 word1, word2 = data.iloc[i, 0], data.iloc[i, 1]
10 if word1 not in word_to_idx or word2 not in word_to_idx:
11 continue
12 else:
13 word1_idx, word2_idx = word_to_idx[word1], word_to_idx[word2]
14 word1_embed, word2_embed = embedding_weights[[word1_idx]], embedding_weights[[word2_idx]]
15 model_similarity.append(float(sklearn.metrics.pairwise.cosine_similarity(word1_embed, word2_embed))) #模型计算的相似度
16 human_similarity.append(float(data.iloc[i, 2])) #已知的相似度
17
18 return scipy.stats.spearmanr(human_similarity, model_similarity) #两者相似度的差异
19
20 def find_nearest(word):
21 index = word_to_idx[word]
22 embedding = embedding_weights[index]
23 cos_dis = np.array([scipy.spatial.distance.cosine(e, embedding) for e in embedding_weights])
24 return [idx_to_word[i] for i in cos_dis.argsort()[:10]]
5.训练模型
- 模型一般需要训练若干个epoch
- 每个epoch我们都把所有的数据分成若干个batch
- 把每个batch的输入和输出都包装成cuda tensor
- forward pass,通过输入的句子预测每个单词的下一个单词
- 用模型的预测和正确的下一个单词计算cross entropy loss
- 清空模型当前gradient
- backward pass
- 更新模型参数
- 每隔一定的iteration输出模型在当前iteration的loss,以及在验证数据集上做模型的评估
1 optimizer = torch.optim.SGD(model.parameters(), lr=LEARNING_RATE)
2 #随机梯度下降
3
4 for e in range(NUM_EPOCHS): #开始迭代
5 for i, (input_labels, pos_labels, neg_labels) in enumerate(dataloader):
6
7 input_labels = input_labels.long() #longtensor
8 pos_labels = pos_labels.long()
9 neg_labels = neg_labels.long()
10 if USE_CUDA:
11 input_labels = input_labels.cuda()
12 pos_labels = pos_labels.cuda()
13 neg_labels = neg_labels.cuda()
14
15 optimizer.zero_grad() #梯度归零
16 loss = model(input_labels, pos_labels, neg_labels).mean()
17 loss.backward()
18 optimizer.step()
19
20 #打印结果
21 if i % 100 == 0:
22 with open(LOG_FILE, "a") as fout:
23 fout.write("epoch: {}, iter: {}, loss: {}\n".format(e, i, loss.item()))
24 print("epoch: {}, iter: {}, loss: {}".format(e, i, loss.item()))
25
26
27 if i % 2000 == 0:
28 embedding_weights = model.input_embeddings()
29 sim_simlex = evaluate("sample_data/simlex-999.txt", embedding_weights)
30 sim_men = evaluate("sample_data/men.txt", embedding_weights)
31 sim_353 = evaluate("sample_data/wordsim353.csv", embedding_weights)
32 with open(LOG_FILE, "a") as fout:
33 print("epoch: {}, iteration: {}, simlex-999: {}, men: {}, sim353: {}, nearest to monster: {}\n".format(
34 e, i, sim_simlex, sim_men, sim_353, find_nearest("monster")))
35 fout.write("epoch: {}, iteration: {}, simlex-999: {}, men: {}, sim353: {}, nearest to monster: {}\n".format(
36 e, i, sim_simlex, sim_men, sim_353, find_nearest("monster")))
37
38 embedding_weights = model.input_embeddings()
39 np.save("embedding-{}".format(EMBEDDING_SIZE), embedding_weights)
40 torch.save(model.state_dict(), "embedding-{}.th".format(EMBEDDING_SIZE))
跑不动,训练过程先不放了。
保存状态
1 model.load_state_dict(torch.load("embedding-{}.th".format(EMBEDDING_SIZE)))
6.在 MEN 和 Simplex-999 数据集上做评估
1 embedding_weights = model.input_embeddings()
2 print("simlex-999", evaluate("simlex-999.txt", embedding_weights))
3 print("men", evaluate("men.txt", embedding_weights))
4 print("wordsim353", evaluate("wordsim353.csv", embedding_weights))
simlex-999 SpearmanrResult(correlation=0.17251697429101504, pvalue=7.863946056740345e-08)
men SpearmanrResult(correlation=0.1778096817088841, pvalue=7.565661657312768e-20)
wordsim353 SpearmanrResult(correlation=0.27153702278146635, pvalue=8.842165885381714e-07)
7.寻找nearest neighbors
1 for word in ["good", "fresh", "monster", "green", "like", "america", "chicago", "work", "computer", "language"]:
2 print(word, find_nearest(word))
good ['good', 'bad', 'perfect', 'hard', 'questions', 'alone', 'money', 'false', 'truth', 'experience']
fresh ['fresh', 'grain', 'waste', 'cooling', 'lighter', 'dense', 'mild', 'sized', 'warm', 'steel']
monster ['monster', 'giant', 'robot', 'hammer', 'clown', 'bull', 'demon', 'triangle', 'storyline', 'slogan']
green ['green', 'blue', 'yellow', 'white', 'cross', 'orange', 'black', 'red', 'mountain', 'gold']
like ['like', 'unlike', 'etc', 'whereas', 'animals', 'soft', 'amongst', 'similarly', 'bear', 'drink']
america ['america', 'africa', 'korea', 'india', 'australia', 'turkey', 'pakistan', 'mexico', 'argentina', 'carolina']
chicago ['chicago', 'boston', 'illinois', 'texas', 'london', 'indiana', 'massachusetts', 'florida', 'berkeley', 'michigan']
work ['work', 'writing', 'job', 'marx', 'solo', 'label', 'recording', 'nietzsche', 'appearance', 'stage']
computer ['computer', 'digital', 'electronic', 'audio', 'video', 'graphics', 'hardware', 'software', 'computers', 'program']
language ['language', 'languages', 'alphabet', 'arabic', 'grammar', 'pronunciation', 'dialect', 'programming', 'chinese', 'spelling']
8.单词之间的关系
1 man_idx = word_to_idx["man"]
2 king_idx = word_to_idx["king"]
3 woman_idx = word_to_idx["woman"]
4 embedding = embedding_weights[woman_idx] - embedding_weights[man_idx] + embedding_weights[king_idx]
5 cos_dis = np.array([scipy.spatial.distance.cosine(e, embedding) for e in embedding_weights])
6 for i in cos_dis.argsort()[:20]:
7 print(idx_to_word[i])
king henry charles pope queen iii prince elizabeth alexander constantine edward son iv louis emperor mary james joseph frederick francis