文章目录

简介

Word2Vec 学习（Skip-gram方法）
参考资料：https://wmathor.com/index.php/archives/1443/

代码

import torchimport numpy as np import unicodedataimport stringimport reimport torch.nn as nnimport torch.optim as optimimport torch.utils.data as data

字典库

class Dictionary:    def __init__(lf, name):        lf.name=name        lf.words=[]        def addWord(lf, word):        if word not in lf.words:            lf.words.append(word)        def getSize(lf):        return len(lf.words)

字符串转换函数，可去除标点符号，统一为小写
（参照pytorch中文文档教程）

def unicodeToAscii(s):    return ''.join(        c for c in unicodedata.normalize('NFD', s)        if unicodedata.category(c) != 'Mn'    )def normalizeString(s):    s = unicodeToAscii(s.lower())    s = re.sub(r"([.!?])", r" ", s)    s = re.sub(r"[^a-zA-Z.!?]+", r" ", s)    s=s.strip()    return s

从文件中读取字符串

lines=[]with open('data\\news.txt',encoding='utf-8') as f:    phras=f.read().strip().split('\n')    for phra in phras:        phra=phra.strip().split('.')        for p in phra:            if p!='':                lines+=normalizeString(p定语从句三要素).split(' ')

建立中心/上下文词对

engDictionary=Dictionary('English')for word in lines:    engDictionary.addWord(word)def make_list(lines):    words_list=[]    for i in range(2, len(l乞ines)-2):        centre=engDictionary.words.index(lines[i])        context=[]        for t in range(i-2, i+3):            if t!=i:                context.append(engDictionary.words.index(lines[t]))        for w in context:            words_list.append([centre, w])    return words_list words_list=make_list(lines)

转换为训练数据
ps: 在造one-hot向量时要用numpy的narray，直接用torch.tensor会报错

input_data=[]output_data=[]def make_data(words_list):    for w in words_list:        k=np.zeros(engDictionary.getSize())        k[w[0]]=1        input_data.append(k)        output_data.append(w[1])make_data(words_list)input_data=torch.Tensor(input_data)output_data=torch.LongTensor(output_data)datat=data.TensorDatat(input_data, output_data)dataloader=data.DataLoader(datat, batch_size=8, shuffle=True)

模型
ps: 最后不需要softmax，crosntropy自带softmax

class Skip_gram(nn.Module):    def __init__(lf, input_size, hidden_size):        super(Skip_gram, lf).__init__()        lf.hidden_size=hidden_size        lf.W=nn.Linear(input_size, hidden_size, bias=Fal)        lf.V=nn.Linear(hidden_size, input_size, bias=Fal)            def forward(lf, x):        x=s2017版 我的前半生 elf.W(x)        x=lf.V(x)        return x    model=Skip_gram(engDi醒酒的食物ctionary.getSize(), 10)

训练
ps: 1. optimizer用Adam， 不用SGD
2. crosntropy的输入一个为概率向量，另一个为分类标签的值，不是向量

criterion=nn.CrossEntropyLoss()optimizer=optim.Adam(model.parameters(), lr=0.001)def train(epoches):    for epoch in range(1, epoches+1):        sumloss=0        for x, y in iter(dataloader):作文写作            optimizer.zero_grad()            o=model(x)            loss=criterion(o, y)            sumloss+=loss.item()            loss.backward()            optimizer.step()        if epoch%100==0:            print('epoch {}: loss:{:.2f}'.format(epoch, sumloss))train(1000)

保存网络数据

PATH='./word2vec.pth'torch.save(model.state_dict(), PATH)  

本文地址：https://blog.csdn.net/qq_32401661/article/details/109960321