# _*_ coding: utf-8 _*_ """ python_lda.py by xianhu """ import os import numpy import logging from collections import defaultdict # 全局变量 MAX_ITER_NUM = 10000 # 最大迭代次数 VAR_NUM = 20 # 自动计算迭代次数时,计算方差的区间大小 class BiDictionary(object): """ 定义双向字典,通过key可以得到value,通过value也可以得到key """ def __init__(self): """ :key: 双向字典初始化 """ self.dict = {} # 正向的数据字典,其key为self的key self.dict_reversed = {} # 反向的数据字典,其key为self的value return def __len__(self): """ :key: 获取双向字典的长度 """ return len(self.dict) def __str__(self): """ :key: 将双向字典转化为字符串对象 """ str_list = ["%s\t%s" % (key, self.dict[key]) for key in self.dict] return "\n".join(str_list) def clear(self): """ :key: 清空双向字典对象 """ self.dict.clear() self.dict_reversed.clear() return def add_key_value(self, key, value): """ :key: 更新双向字典,增加一项 """ self.dict[key] = value self.dict_reversed[value] = key return def remove_key_value(self, key, value): """ :key: 更新双向字典,删除一项 """ if key in self.dict: del self.dict[key] del self.dict_reversed[value] return def get_value(self, key, default=None): """ :key: 通过key获取value,不存在返回default """ return self.dict.get(key, default) def get_key(self, value, default=None): """ :key: 通过value获取key,不存在返回default """ return self.dict_reversed.get(value, default) def contains_key(self, key): """ :key: 判断是否存在key值 """ return key in self.dict def contains_value(self, value): """ :key: 判断是否存在value值 """ return value in self.dict_reversed def keys(self): """ :key: 得到双向字典全部的keys """ return self.dict.keys() def values(self): """ :key: 得到双向字典全部的values """ return self.dict_reversed.keys() def items(self): """ :key: 得到双向字典全部的items """ return self.dict.items() class CorpusSet(object): """ 定义语料集类,作为LdaBase的基类 """ def __init__(self): """ :key: 初始化函数 """ # 定义关于word的变量 self.local_bi = BiDictionary() # id和word之间的本地双向字典,key为id,value为word self.words_count = 0 # 数据集中word的数量（排重之前的） self.V = 0 # 数据集中word的数量（排重之后的） # 定义关于article的变量 self.artids_list = [] # 全部article的id的列表,按照数据读取的顺序存储 self.arts_Z = [] # 全部article中所有词的id信息,维数为 M * art.length() self.M = 0 # 数据集中article的数量 # 定义推断中用到的变量（可能为空） self.global_bi = None # id和word之间的全局双向字典,key为id,value为word self.local_2_global = {} # 一个字典,local字典和global字典之间的对应关系 return def init_corpus_with_file(self, file_name): """ :key: 利用数据文件初始化语料集数据。文件每一行的数据格式: id[tab]word1 word2 word3...... """ with open(file_name, "r", encoding="utf-8") as file_iter: self.init_corpus_with_articles(file_iter) return def init_corpus_with_articles(self, article_list): """ :key: 利用article的列表初始化语料集。每一篇article的格式为: id[tab]word1 word2 word3...... """ # 清理数据--word数据 self.local_bi.clear() self.words_count = 0 self.V = 0 # 清理数据--article数据 self.artids_list.clear() self.arts_Z.clear() self.M = 0 # 清理数据--清理local到global的映射关系 self.local_2_global.clear() # 读取article数据 for line in article_list: frags = line.strip().split() if len(frags) < 2: continue # 获取article的id art_id = frags[0].strip() # 获取word的id art_wordid_list = [] for word in [w.strip() for w in frags[1:] if w.strip()]: local_id = self.local_bi.get_key(word) if self.local_bi.contains_value(word) else len(self.local_bi) # 这里的self.global_bi为None和为空是有区别的 if self.global_bi is None: # 更新id信息 self.local_bi.add_key_value(local_id, word) art_wordid_list.append(local_id) else: if self.global_bi.contains_value(word): # 更新id信息 self.local_bi.add_key_value(local_id, word) art_wordid_list.append(local_id) # 更新local_2_global self.local_2_global[local_id] = self.global_bi.get_key(word) # 更新类变量: 必须article中word的数量大于0 if len(art_wordid_list) > 0: self.words_count += len(art_wordid_list) self.artids_list.append(art_id) self.arts_Z.append(art_wordid_list) # 做相关初始计算--word相关 self.V = len(self.local_bi) logging.debug("words number: " + str(self.V) + ", " + str(self.words_count)) # 做相关初始计算--article相关 self.M = len(self.artids_list) logging.debug("articles number: " + str(self.M)) return def save_wordmap(self, file_name): """ :key: 保存word字典,即self.local_bi的数据 """ with open(file_name, "w", encoding="utf-8") as f_save: f_save.write(str(self.local_bi)) return def load_wordmap(self, file_name): """ :key: 加载word字典,即加载self.local_bi的数据 """ self.local_bi.clear() with open(file_name, "r", encoding="utf-8") as f_load: for _id, _word in [line.strip().split() for line in f_load if line.strip()]: self.local_bi.add_key_value(int(_id), _word.strip()) self.V = len(self.local_bi) return class LdaBase(CorpusSet): """ LDA模型的基类,相关说明: 》article的下标范围为[0, self.M), 下标为 m 》wordid的下标范围为[0, self.V), 下标为 w 》topic的下标范围为[0, self.K), 下标为 k 或 topic 》article中word的下标范围为[0, article.size()), 下标为 n """ def __init__(self): """ :key: 初始化函数 """ CorpusSet.__init__(self) # 基础变量--1 self.dir_path = "" # 文件夹路径,用于存放LDA运行的数据、中间结果等 self.model_name = "" # LDA训练或推断的模型名称,也用于读取训练的结果 self.current_iter = 0 # LDA训练或推断的模型已经迭代的次数,用于继续模型训练过程 self.iters_num = 0 # LDA训练或推断过程中Gibbs抽样迭代的总次数,整数值或者"auto" self.topics_num = 0 # LDA训练或推断过程中的topic的数量,即self.K值 self.K = 0 # LDA训练或推断过程中的topic的数量,即self.topics_num值 self.twords_num = 0 # LDA训练或推断结束后输出与每个topic相关的word的个数 # 基础变量--2 self.alpha = numpy.zeros(self.K) # 超参数alpha,K维的float值,默认为50/K self.beta = numpy.zeros(self.V) # 超参数beta,V维的float值,默认为0.01 # 基础变量--3 self.Z = [] # 所有word的topic信息,即Z(m, n),维数为 M * article.size() # 统计计数(可由self.Z计算得到) self.nd = numpy.zeros((self.M, self.K)) # nd[m, k]用于保存第m篇article中第k个topic产生的词的个数,其维数为 M * K self.ndsum = numpy.zeros((self.M, 1)) # ndsum[m, 0]用于保存第m篇article的总词数,维数为 M * 1 self.nw = numpy.zeros((self.K, self.V)) # nw[k, w]用于保存第k个topic产生的词中第w个词的数量,其维数为 K * V self.nwsum = numpy.zeros((self.K, 1)) # nwsum[k, 0]用于保存第k个topic产生的词的总数,维数为 K * 1 # 多项式分布参数变量 self.theta = numpy.zeros((self.M, self.K)) # Doc-Topic多项式分布的参数,维数为 M * K,由alpha值影响 self.phi = numpy.zeros((self.K, self.V)) # Topic-Word多项式分布的参数,维数为 K * V,由beta值影响 # 辅助变量,目的是提高算法执行效率 self.sum_alpha = 0.0 # 超参数alpha的和 self.sum_beta = 0.0 # 超参数beta的和 # 先验知识,格式为{word_id: [k1, k2, ...], ...} self.prior_word = defaultdict(list) # 推断时需要的训练模型 self.train_model = None return # --------------------------------------------------辅助函数--------------------------------------------------------- def init_statistics_document(self): """ :key: 初始化关于article的统计计数。先决条件: self.M, self.K, self.Z """ assert self.M > 0 and self.K > 0 and self.Z # 统计计数初始化 self.nd = numpy.zeros((self.M, self.K), dtype=numpy.int) self.ndsum = numpy.zeros((self.M, 1), dtype=numpy.int) # 根据self.Z进行更新,更新self.nd[m, k]和self.ndsum[m, 0] for m in range(self.M): for k in self.Z[m]: self.nd[m, k] += 1 self.ndsum[m, 0] = len(self.Z[m]) return def init_statistics_word(self): """ :key: 初始化关于word的统计计数。先决条件: self.V, self.K, self.Z, self.arts_Z """ assert self.V > 0 and self.K > 0 and self.Z and self.arts_Z # 统计计数初始化 self.nw = numpy.zeros((self.K, self.V), dtype=numpy.int) self.nwsum = numpy.zeros((self.K, 1), dtype=numpy.int) # 根据self.Z进行更新,更新self.nw[k, w]和self.nwsum[k, 0] for m in range(self.M): for k, w in zip(self.Z[m], self.arts_Z[m]): self.nw[k, w] += 1 self.nwsum[k, 0] += 1 return def init_statistics(self): """ :key: 初始化全部的统计计数。上两个函数的综合函数。 """ self.init_statistics_document() self.init_statistics_word() return def sum_alpha_beta(self): """ :key: 计算alpha、beta的和 """ self.sum_alpha = self.alpha.sum() self.sum_beta = self.beta.sum() return def calculate_theta(self): """ :key: 初始化并计算模型的theta值(M*K),用到alpha值 """ assert self.sum_alpha > 0 self.theta = (self.nd + self.alpha) / (self.ndsum + self.sum_alpha) return def calculate_phi(self): """ :key: 初始化并计算模型的phi值(K*V),用到beta值 """ assert self.sum_beta > 0 self.phi = (self.nw + self.beta) / (self.nwsum + self.sum_beta) return # ---------------------------------------------计算Perplexity值------------------------------------------------------ def calculate_perplexity(self): """ :key: 计算Perplexity值,并返回 """ # 计算theta和phi值 self.calculate_theta() self.calculate_phi() # 开始计算 preplexity = 0.0 for m in range(self.M): for w in self.arts_Z[m]: preplexity += numpy.log(numpy.sum(self.theta[m] * self.phi[:, w])) return numpy.exp(-(preplexity / self.words_count)) # --------------------------------------------------静态函数--------------------------------------------------------- @staticmethod def multinomial_sample(pro_list): """ :key: 静态函数,多项式分布抽样,此时会改变pro_list的值 :param pro_list: [0.2, 0.7, 0.4, 0.1],此时说明返回下标1的可能性大,但也不绝对 """ # 将pro_list进行累加 for k in range(1, len(pro_list)): pro_list[k] += pro_list[k-1] # 确定随机数 u 落在哪个下标值,此时的下标值即为抽取的类别（random.rand()返回: [0, 1.0)） u = numpy.random.rand() * pro_list[-1] return_index = len(pro_list) - 1 for t in range(len(pro_list)): if pro_list[t] > u: return_index = t break return return_index # ----------------------------------------------Gibbs抽样算法-------------------------------------------------------- def gibbs_sampling(self, is_calculate_preplexity): """ :key: LDA模型中的Gibbs抽样过程 :param is_calculate_preplexity: 是否计算preplexity值 """ # 计算preplexity值用到的变量 pp_list = [] pp_var = numpy.inf # 开始迭代 last_iter = self.current_iter + 1 iters_num = self.iters_num if self.iters_num != "auto" else MAX_ITER_NUM for self.current_iter in range(last_iter, last_iter+iters_num): info = "......" # 是否计算preplexity值 if is_calculate_preplexity: pp = self.calculate_perplexity() pp_list.append(pp) # 计算列表最新VAR_NUM项的方差 pp_var = numpy.var(pp_list[-VAR_NUM:]) if len(pp_list) >= VAR_NUM else numpy.inf info = (", preplexity: " + str(pp)) + ((", var: " + str(pp_var)) if len(pp_list) >= VAR_NUM else "") # 输出Debug信息 logging.debug("\titeration " + str(self.current_iter) + info) # 判断是否跳出循环 if self.iters_num == "auto" and pp_var < (VAR_NUM / 2): break # 对每篇article的每个word进行一次抽样,抽取合适的k值 for m in range(self.M): for n in range(len(self.Z[m])): w = self.arts_Z[m][n] k = self.Z[m][n] # 统计计数减一 self.nd[m, k] -= 1 self.ndsum[m, 0] -= 1 self.nw[k, w] -= 1 self.nwsum[k, 0] -= 1 if self.prior_word and (w in self.prior_word): # 带有先验知识,否则进行正常抽样 k = numpy.random.choice(self.prior_word[w]) else: # 计算theta值--下边的过程为抽取第m篇article的第n个词w的topic,即新的k theta_p = (self.nd[m] + self.alpha) / (self.ndsum[m, 0] + self.sum_alpha) # 计算phi值--判断是训练模型,还是推断模型（注意self.beta[w_g]） if self.local_2_global and self.train_model: w_g = self.local_2_global[w] phi_p = (self.train_model.nw[:, w_g] + self.nw[:, w] + self.beta[w_g]) / \ (self.train_model.nwsum[:, 0] + self.nwsum[:, 0] + self.sum_beta) else: phi_p = (self.nw[:, w] + self.beta[w]) / (self.nwsum[:, 0] + self.sum_beta) # multi_p为多项式分布的参数,此时没有进行标准化 multi_p = theta_p * phi_p # 此时的topic即为Gibbs抽样得到的topic,它有较大的概率命中多项式概率大的topic k = LdaBase.multinomial_sample(multi_p) # 统计计数加一 self.nd[m, k] += 1 self.ndsum[m, 0] += 1 self.nw[k, w] += 1 self.nwsum[k, 0] += 1 # 更新Z值 self.Z[m][n] = k # 抽样完毕 return # -----------------------------------------Model数据存储、读取相关函数------------------------------------------------- def save_parameter(self, file_name): """ :key: 保存模型相关参数数据,包括: topics_num, M, V, K, words_count, alpha, beta """ with open(file_name, "w", encoding="utf-8") as f_param: for item in ["topics_num", "M", "V", "K", "words_count"]: f_param.write("%s\t%s\n" % (item, str(self.__dict__[item]))) f_param.write("alpha\t%s\n" % ",".join([str(item) for item in self.alpha])) f_param.write("beta\t%s\n" % ",".join([str(item) for item in self.beta])) return def load_parameter(self, file_name): """ :key: 加载模型相关参数数据,和上一个函数相对应 """ with open(file_name, "r", encoding="utf-8") as f_param: for line in f_param: key, value = line.strip().split() if key in ["topics_num", "M", "V", "K", "words_count"]: self.__dict__[key] = int(value) elif key in ["alpha", "beta"]: self.__dict__[key] = numpy.array([float(item) for item in value.split(",")]) return def save_zvalue(self, file_name): """ :key: 保存模型关于article的变量,包括: arts_Z, Z, artids_list等 """ with open(file_name, "w", encoding="utf-8") as f_zvalue: for m in range(self.M): out_line = [str(w) + ":" + str(k) for w, k in zip(self.arts_Z[m], self.Z[m])] f_zvalue.write(self.artids_list[m] + "\t" + " ".join(out_line) + "\n") return def load_zvalue(self, file_name): """ :key: 读取模型的Z变量。和上一个函数相对应 """ self.arts_Z = [] self.artids_list = [] self.Z = [] with open(file_name, "r", encoding="utf-8") as f_zvalue: for line in f_zvalue: frags = line.strip().split() art_id = frags[0].strip() w_k_list = [value.split(":") for value in frags[1:]] # 添加到类中 self.artids_list.append(art_id) self.arts_Z.append([int(item[0]) for item in w_k_list]) self.Z.append([int(item[1]) for item in w_k_list]) return def save_twords(self, file_name): """ :key: 保存模型的twords数据,要用到phi的数据 """ self.calculate_phi() out_num = self.V if self.twords_num > self.V else self.twords_num with open(file_name, "w", encoding="utf-8") as f_twords: for k in range(self.K): words_list = sorted([(w, self.phi[k, w]) for w in range(self.V)], key=lambda x: x[1], reverse=True) f_twords.write("Topic %dth:\n" % k) f_twords.writelines(["\t%s %f\n" % (self.local_bi.get_value(w), p) for w, p in words_list[:out_num]]) return def load_twords(self, file_name): """ :key: 加载模型的twords数据,即先验数据 """ self.prior_word.clear() topic = -1 with open(file_name, "r", encoding="utf-8") as f_twords: for line in f_twords: if line.startswith("Topic"): topic = int(line.strip()[6:-3]) else: word_id = self.local_bi.get_key(line.strip().split()[0].strip()) self.prior_word[word_id].append(topic) return def save_tag(self, file_name): """ :key: 输出模型最终给数据打标签的结果,用到theta值 """ self.calculate_theta() with open(file_name, "w", encoding="utf-8") as f_tag: for m in range(self.M): f_tag.write("%s\t%s\n" % (self.artids_list[m], " ".join([str(item) for item in self.theta[m]]))) return def save_model(self): """ :key: 保存模型数据 """ name_predix = "%s-%05d" % (self.model_name, self.current_iter) # 保存训练结果 self.save_parameter(os.path.join(self.dir_path, "%s.%s" % (name_predix, "param"))) self.save_wordmap(os.path.join(self.dir_path, "%s.%s" % (name_predix, "wordmap"))) self.save_zvalue(os.path.join(self.dir_path, "%s.%s" % (name_predix, "zvalue"))) #保存额外数据 self.save_twords(os.path.join(self.dir_path, "%s.%s" % (name_predix, "twords"))) self.save_tag(os.path.join(self.dir_path, "%s.%s" % (name_predix, "tag"))) return def load_model(self): """ :key: 加载模型数据 """ name_predix = "%s-%05d" % (self.model_name, self.current_iter) # 加载训练结果 self.load_parameter(os.path.join(self.dir_path, "%s.%s" % (name_predix, "param"))) self.load_wordmap(os.path.join(self.dir_path, "%s.%s" % (name_predix, "wordmap"))) self.load_zvalue(os.path.join(self.dir_path, "%s.%s" % (name_predix, "zvalue"))) return class LdaModel(LdaBase): """ LDA模型定义,主要实现训练、继续训练、推断的过程 """ def init_train_model(self, dir_path, model_name, current_iter, iters_num=None, topics_num=10, twords_num=200, alpha=-1.0, beta=0.01, data_file="", prior_file=""): """ :key: 初始化训练模型,根据参数current_iter（是否等于0）决定是初始化新模型,还是加载已有模型 :key: 当初始化新模型时,除了prior_file先验文件外,其余所有的参数都需要,且current_iter等于0 :key: 当加载已有模型时,只需要dir_path, model_name, current_iter（不等于0）, iters_num, twords_num即可 :param iters_num: 可以为整数值或者“auto” """ if current_iter == 0: logging.debug("init a new train model") # 初始化语料集 self.init_corpus_with_file(data_file) # 初始化部分变量 self.dir_path = dir_path self.model_name = model_name self.current_iter = current_iter self.iters_num = iters_num self.topics_num = topics_num self.K = topics_num self.twords_num = twords_num # 初始化alpha和beta self.alpha = numpy.array([alpha if alpha > 0 else (50.0/self.K) for k in range(self.K)]) self.beta = numpy.array([beta if beta > 0 else 0.01 for w in range(self.V)]) # 初始化Z值,以便统计计数 self.Z = [[numpy.random.randint(self.K) for n in range(len(self.arts_Z[m]))] for m in range(self.M)] else: logging.debug("init an existed model") # 初始化部分变量 self.dir_path = dir_path self.model_name = model_name self.current_iter = current_iter self.iters_num = iters_num self.twords_num = twords_num # 加载已有模型 self.load_model() # 初始化统计计数 self.init_statistics() # 计算alpha和beta的和值 self.sum_alpha_beta() # 初始化先验知识 if prior_file: self.load_twords(prior_file) # 返回该模型 return self def begin_gibbs_sampling_train(self, is_calculate_preplexity=True): """ :key: 训练模型,对语料集中的所有数据进行Gibbs抽样,并保存最后的抽样结果 """ # Gibbs抽样 logging.debug("sample iteration start, iters_num: " + str(self.iters_num)) self.gibbs_sampling(is_calculate_preplexity) logging.debug("sample iteration finish") # 保存模型 logging.debug("save model") self.save_model() return def init_inference_model(self, train_model): """ :key: 初始化推断模型 """ self.train_model = train_model # 初始化变量: 主要用到self.topics_num, self.K self.topics_num = train_model.topics_num self.K = train_model.K # 初始化变量self.alpha, self.beta,直接沿用train_model的值 self.alpha = train_model.alpha # K维的float值,训练和推断模型中的K相同,故可以沿用 self.beta = train_model.beta # V维的float值,推断模型中用于计算phi的V值应该是全局的word的数量,故可以沿用 self.sum_alpha_beta() # 计算alpha和beta的和 # 初始化数据集的self.global_bi self.global_bi = train_model.local_bi return def inference_data(self, article_list, iters_num=100, repeat_num=3): """ :key: 利用现有模型推断数据 :param article_list: 每一行的数据格式为: id[tab]word1 word2 word3...... :param iters_num: 每一次迭代的次数 :param repeat_num: 重复迭代的次数 """ # 初始化语料集 self.init_corpus_with_articles(article_list) # 初始化返回变量 return_theta = numpy.zeros((self.M, self.K)) # 重复抽样 for i in range(repeat_num): logging.debug("inference repeat_num: " + str(i+1)) # 初始化变量 self.current_iter = 0 self.iters_num = iters_num # 初始化Z值,以便统计计数 self.Z = [[numpy.random.randint(self.K) for n in range(len(self.arts_Z[m]))] for m in range(self.M)] # 初始化统计计数 self.init_statistics() # 开始推断 self.gibbs_sampling(is_calculate_preplexity=False) # 计算theta self.calculate_theta() return_theta += self.theta # 计算结果,并返回 return return_theta / repeat_num if __name__ == "__main__": """ 测试代码 """ logging.basicConfig(level=logging.DEBUG, format="%(asctime)s\t%(levelname)s\t%(message)s") # train或者inference test_type = "train" # test_type = "inference" # 测试新模型 if test_type == "train": model = LdaModel() # 由prior_file决定是否带有先验知识 model.init_train_model("data/", "model", current_iter=0, iters_num="auto", topics_num=10, data_file="corpus.txt") # model.init_train_model("data/", "model", current_iter=0, iters_num="auto", topics_num=10, data_file="corpus.txt", prior_file="prior.twords") model.begin_gibbs_sampling_train() elif test_type == "inference": model = LdaModel() model.init_inference_model(LdaModel().init_train_model("data/", "model", current_iter=134)) data = [ "cn 咪咕 漫画 咪咕 漫画 漫画 更名 咪咕 漫画 资源 偷星 国漫 全彩 日漫 实时 在线看 随心所欲 登陆 漫画 资源 黑白 全彩 航海王", "co aircloud aircloud 硬件 设备 wifi 智能 手要 平板电脑 电脑 存储 aircloud 文件 远程 型号 aircloud 硬件 设备 wifi" ] result = model.inference_data(data) # 退出程序 exit()