python实现word2vec（不使用框架）

参考两篇博客
基础知识
源码解析

import time
import numpy as np
import math

wordHash = {}
wordNum = 0
window = 2
words = []
vecSize = 100
u = 0.1
t = 500

#将单词存入map
def read_file():
    global wordNum,wordHash,words
    f =open("test.txt",encoding="utf-8")
    sentences = f.readlines()
    for sentence in sentences:
        words = sentence.split(" ")
        for word in words:
            if word in wordHash:
                wordHash[word] += 1
            else:
                wordHash[word] = 1
            wordNum+=1



#对单词构建哈弗曼编码
def buildHFMTree():
    global wordHash
    vocab = sorted(wordHash.items(), key=lambda item: item[1], reverse=True)
    length = len(vocab)*2
    weight = [None]*length
    parent = [None]*length
    pos = [None]*length
    for i in range(length):
        if i < length/2:
            wordHash[vocab[i][0]] = i
            weight[i] = vocab[i][1]
        else:
            weight[i] = wordNum
    lp = len(vocab)-1
    rp = lp+1
    addp = lp+1
    while True:
        if lp<0:
            if rp+1==addp:
                break
            weight[addp] = weight[rp]+weight[rp+1]
            pos[rp] = 0
            pos[rp+1] = 1
            parent[rp] = addp
            parent[rp+1] = addp
            addp+=1
            rp+=2
            continue
        if weight[lp] < weight[rp]:
            if lp-1>=0 and weight[lp-1] < weight[rp]:
                min = lp
                max = lp-1
                lp = lp-2
            else:
                min = lp
                max = rp
                lp-=1
                rp+=1
        else:
            if weight[rp+1] > weight[lp]:
                min = rp
                max = lp
                lp -= 1
                rp+=1
            else:
                min = rp
                max = rp+1
                rp+=2
        weight[addp] = weight[min]+weight[max]
        pos[min] = 0
        pos[max] = 1
        parent[min] = addp
        parent[max] = addp
        addp+=1
    return pos,parent

def sigmiod(n):
    return np.exp(n)/(1+np.exp(n))

def getHFMCode(word,pos,parent):
    global wordHash
    i = wordHash[word]
    code = []
    while parent[i]!=None:
        code.append(pos[i])
        i = parent[i]
    print("单词'"+word+"'的哈弗曼编码:"+str(code))
    return code

def updataParam(word,pos,parent,ansVec,projVec,paramVec):
    global wordHash
    i = wordHash[word]
    ll = 0
    paramChange = np.zeros((vecSize,wordNum-1))
    projChange = np.zeros((vecSize))
    while parent[i] != None:
        d = pos[i]
        n = ansVec[parent[i]-wordNum]
        try:
            ll += (1-d)*math.log(sigmiod(n))+d*math.log(1-sigmiod(n))
        except Exception:
            ll += 0
        m = 1-d-sigmiod(n)
        gradProj = m*projVec
        gradParam = m * paramVec[:,parent[i] - wordNum]
        paramChange[:,parent[i]-wordNum] += gradProj*u
        projChange += gradParam*u
        i = parent[i]
    return projChange,paramChange,ll


def initVec():
    global wordNum,vecSize
    wordVec = np.random.random((wordNum,vecSize))
    paramVec = np.zeros((vecSize,wordNum-1))
    for i in range(wordNum):
        for j in range(vecSize):
            wordVec[i][j] = (wordVec[i][j]-0.5)/vecSize
    return wordVec,paramVec

def train(wordVec,paramVec,pos,parent):
    global vecSize
    for k in range(t+1):
        paramChange = np.zeros((vecSize, wordNum - 1))
        wordChange = np.zeros((wordNum,vecSize))
        for i in range(len(words)):
            projVec = np.zeros(vecSize)
            n = 0
            for j in range(i-window,i):
                if j<0:
                   continue
                projVec += wordVec[wordHash[words[j]]]
                n+=1
            for j in range(i+1,i+window):
                if j>=len(words):
                    continue
                projVec += wordVec[wordHash[words[j]]]
                n+=1
            projVec = projVec/n
            ansVec = projVec.dot(paramVec)
            projChange1,paramChange1,ll = updataParam(words[i],pos,parent,ansVec,projVec,paramVec)
            for j in range(i - window, i):
                if j < 0:
                    continue
                wordChange[wordHash[words[j]]]+=projChange1
            for j in range(i + 1, i + window):
                if j >= len(words):
                    continue
                wordChange[wordHash[words[j]]]+=projChange1
            paramChange+=paramChange1
            if k%100==0:
                print("第"+str(k)+"轮训练中，单词"+words[i]+"的损失为："+str(ll))
        wordVec+=wordChange
        paramVec+=paramChange

print("开始读取单词")
read_file()
print("读取单词结束")
print("开始构建哈弗曼树")
pos,parent = buildHFMTree()
print("构建完成")
#getHFMCode('interpretations',pos,parent)
print("开始初始化单词向量")
wordVec,paramVec = initVec()
print("单词向量初始完成")
print("准备训练参数")
train(wordVec,paramVec,pos,parent)

语料：

In the near future the translation history will only be viewable when you log in to your account and it will be centrally managed in my activity record. This upgrade will clear the previous history so if you want the system to record certain translations for future review please be sure to save the translation results

运行结果：


不足
相对于源码，未实现负采样，多线程和指数运算近似来加快性能的功能