学习笔记TF034:实现Word2Vec
卷积神经网络发展趋势。Perceptron(感知机),1957年,Frank Resenblatt提出,始祖。Neocognitron(神经认知机),多层级神经网络,日本科学家Kunihiko fukushima,20世纪80年代提出,一定程度视觉认知功能,启发卷积神经网络。LeNet-5,CNN之父,Yann LeCun,1997年提出,首次多层级联卷积结构,手写数字有效识别。2012年,Hinton学生Alex,8层卷积神经网络,ILSVRC 2012比赛冠军。AlexNet 成功应用ReLU激活函数、Dropout、最大覆盖池化、LRN层、GPU加速,启发后续技术创新,卷积神经网络研究进入快车道。
AlexNetx后,卷积神经网络,一类网络结构改进调整,一类网络深度增加。
Perceptron(1957) Neocognitron(198x) NIN(2013) VGG(2014) Incepiton V1(2014) MSRANet(2014) Incepiton V2(2015) ResNet(2015) Incepiton V3(2015) ResNet V2(2015) Inception ResNet V2(2016)
2013年,颜水成教授,Network in Network首次发表,优化卷积神经网络结构,推广1x1卷积结构。2014年,Google Incepiton Net V1,Inception Module,反复堆叠高效卷积网络结构,ILSVRC 2014冠军。2015年初,Incepiton V2,Batch Normalization,加速训练过程,提升网络性能。2015年末,Inception V3,Factorization in Small Convolutions思想,分解大尺寸卷积为多个小卷积或一维卷积。
2014年,ILSVRC亚军,VGGNet,全程3x3卷积,19层网络。季军MSRA-Net(微软)也是深层网络。2015年,微软ResNet,152层网络,ILSVRC 2015冠军,top-5错误率3.46%。ResNet V2,Batch Normalization,去除激活层,用Identity Mapping或Preactivation,提升网络性能。Inception ResNet V2,融合Inception Net网络结构,和ResNet训练极深网络残差学习模块。
GPU计算资源,开源工具。
循环神经网络(RNN),NLP(Nature Language Processing,自然语言处理)最常用神经网络结构。Word2Vec,语言字词转化稠密向量(Dense Vector)。
Word2Vec,Word Embeddings,词向量或词嵌入。语言字词转向量形式表达(Vector Representations)模型。图片,像素点稠密矩阵,音频,声音信号频谱数据。
One-Hot Encoder,字词转离散单独符号。一个词对应一个向量,整篇文章对应一个稀疏矩阵。文本分类模型,Bag of Words,稀疏矩阵合并为一个向量,每个词对应向量计数,统计词出现次数,作为特征。特征编码随机,没有关联信息,没有字词关系。稀疏向量需要更多数据训练,训练效率低,计算麻烦。
向量表达(Vector Representations),向量空间模型(Vector Space Models),字词转连续值向量表达,意思相近词映射向量空量空间相近位置。向量空间模型在NLP依赖假设Distributional Hypothesis,相同语境词语义相近。向量空间模型,分两类,计数模型(Latent Semantic Analysis),预测模型(Neural Probabilistic Language Models)。计数模型统计语料库相邻词频率,计数统计结果转小稠密矩阵,预测模型根据词周围相邻词推测出这个词和空间向量。
Word2Vec,计算非常高效,从原始语料学习字词空间向量预测模型。CBOW(Continuous Bag of Words)模式从原始语句推测目标字词,适合小型数据。Skip-Gram从目标字词推测原始语句,适合大型语料。意思相近词向量空间位置接近。
预测模型(Neural Probabilistic Language Models),用最大似然方法,给定前语句h,最大化目标词汇Wt概率。计算量大,需计算词汇表所有单词出现可能性。Word2Vec CBOw模型,只需训练二元分类模型,区分真实目标词汇、编造词汇(噪声)两类。少量噪声词汇估计,类似蒙特卡洛模拟。
模型预测真实目标词汇高概率,预测其他噪声词汇低概率,训练学习目标最优化。编造噪声词汇训练,Negative Sampling,计算loss fuction效率非常高,只需计算随机选择k个词汇,训练速度快。Noise_contrastive Estimation(NCE) Loss,TensorFlow tf.nn.nce_loss。
Word2Vec Skip-Gram模式。构造语境与目标词汇映射关系。语境包括单词左边和右边词汇。滑窗尺寸 1。Skip-Gram模型,从目标词汇预测语境。制造随机词汇作负样本(噪声)。预测概率分布,正样本尽可能大,随机产生负样本尺可能小。优化算法(SGD)更新模型Word Embedding参数,概率分布损失函数(NCE Loss)尽可能小。单词Embedded Vector随训练过程调整,直到最适合语料空间位置。损失函数最小,最符合语料,预测正确单词概率最高。
载入依赖库。
定义下载广西数据函数,urllib.request.urlretrieve下载数据压缩文件核文件尺寸。已下载跳过。
解压下载压缩文件,tf.compat.as_str 数据转单词列表。数据转为17005207单词列表。
创建vocabulary词汇表,collections.Counter统计单词列表单词频数,most_common方法取top 50000频数单词作vocabulary。创建dict,top 50000词汇vocabulary放入dictionary,快速查询。Python dict查询复杂度O(1),性能好。全部单词转编号(频数排序编号)。top50000以外单词,认定为Unkown(未知),编号0,统计数量。遍历单词列表,每个单词,判断是否出现在dictionary,是转编号,不是编0。返回转换编码(data)、单词频数统计count、词汇表(dictionary)、反转形式(reverse_dictionary)。
删除原始单词列表,节约内存。打印vocabulary最高频词汇、数量(包括Unknow词汇)。“UNK”类418391个。“the”1061396个。“of”593677个。data前10单词['anarchism','originated','as','a','term','of','abuse','first','used','against'],编号[5235,3084,12,6,195,2,3137,46,59,156]。
生成Word2Vec训练样本。Skip-Gram模式(从目标单词反推语境)。定义函数generate_batch生成训练batch数据。参数batch_size为batch大小。skip_window单词最远可联系距离,设1只能跟紧邻两个单词生成样本。num_skips单词生成样本个数,不能大于skip_window两倍,batch_size是它的整数倍,确保batch包含词汇所有样本。
单词序号data_index为global变量,反复调用generate_batch,确保data_index可以在函数genetate_batch修改。assert确保num_skips、batch_size满足条件。np.ndarray初始化batch、labels为数组。定义span 单词创建相关样本单词数量,包括目标单词和前后单词,span=2*skip_window+1。创建最大容量span deque,双向队列,deque append方法添加变量,只保留最后插入span个变量。
从序号data_index开始,span个单词顺序读入buffer作初始值。buffer容量为span deque,已填满,后续数据替换前面数据。
第一层循环(次数batch_size//num_skips),循环内目标单词生成样本。buffer目标单词和所有相关单词,定义target-skip_window,buffer第skip_window个变量为目标单词。定义生成样本需避免单词列表,tagets_to_avoid,列表开始包括第skip_window个单词(目标单词),预测语境单词,不包括目标单词。
第二层循环(次数num_skips),循环语境单词生成样本,先产生随机数,直到随机数不在targets_to_avoid中,代表可用语境单词,生成样本,feature目标词汇buffer[skip_window],label是buffer[target]。语境单词使用,添加到targets_to_avoid过滤。目标单词所有样本生成完(num_skips个),读入下一个单词,抛掉buffer第一个单词,滑窗向后移动一位,目标单词向后移动一个,语境单词整体后移,开始生成下一个目标单词训练样本。
两层循环完成,获得batch_size个训练样本。返回batch、labels。
调用generate_batch函数测试。参数batch_size设8,num_skips设2,skip_window设1,执行generate_batch获得batch、labels,打印。
定义训练batch_size 128,embedding_size 128。embedding_size,单词转稠密向量维度,50〜1000。skip_window单词间最远联系距离设1,num_skips目标单词提取样本数设2.生成验证数据valid_examples。随机抽取频数最高单词,看向量空间最近单词是否相关性高。valid_size设16抽取验证单词数。valid_window设100验证单词频为最高100个单词抽取。np.random.choice函数随机抽取。num_sampled训练负样本噪声单词数量。
定义Skip_Gram Word2Vec模型网络结构。创建f.Graph,设置为默认graph。创建训练数据inputs、labels placeholder,随机产生valid_examples转TensorFlow constant。with tf.device('/cpu:0')限定所有计算在CPU执行。tf.random_uniform随机生成所有单词词向量embeddings,单词表大小50000,向量维度128,tf.nn.embedding_lookup查找输入train_inputs对应赂理embed。tf.truncated_normal初始化训练优化目标NCE Loss的权重参数nce_weights,nce_biases初始化0。tf.nn.nce_loss计算学习词向量embedding训练数据loss,tf.reduce_mean汇总。
定义优化器SGD ,学习速率1.0。计算嵌入向量embeddings L2范数norm,embeddings除L2范数得标准化normalized_embeddings。tf.nn.embedding_lookup查询验证单词嵌入向量,计算验证单词嵌入同与词汇表所有单词相似性。tf.global_variables_initializer初始化所有模型参数。
定义最大迭代次数10万次,创建设置默认session,执行参数初始化。迭代中,generate_batch生成batch inputs、labels数据,创建feed_dict。session.run()执行优化器运算(参数更新)和损失计算,训练loss累积到avegage_loss。
每2000次循环,计算平均loss,显示。
每10000次循环,计算验证单词和全部单词相似度,验证单词最相似8个单词展示。
训练模型对名词、动词、形容词类型单词相似词汇识别非常准确。Skip-Gram Word2Vec 向量空间表达(Vetor Representations)质量非常高,近义词在向量空间位置非常靠近。
定义可视化Word2Vec效果函数。low_dim_embs降给到2维单词空间向量,图表展示单词位置。plt.scatter(matplotlib.pyplot)显示散点图(单词位置),plt.annotate展示单词本身。plt.savefig保存图片到本地文件。
sklearn.manifold.TSNe实现降维,原始128维嵌入同量降到2维,plot_sith_labels函数展示。只展示词频最高100个单词可视化结果。
距离相近单词,语义高相似性。左上角单个字母聚集地。冠词聚集在左边中部。Word2Vec性能评价,可视化观察,Analogical Reasoning直接预测语义、语境关系。回答填空问题。大规模语料库,参数调试选取最适合值。
import collections import math import os import random import zipfile import numpy as np import urllib import tensorflow as tf # Step 1: Download the data. url = 'http://mattmahoney.net/dc/' def maybe_download(filename, expected_bytes): if not os.path.exists(filename): filename, _ = urllib.request.urlretrieve(url + filename, filename) statinfo = os.stat(filename) if statinfo.st_size == expected_bytes: print('Found and verified', filename) else: print(statinfo.st_size) raise Exception( 'Failed to verify ' + filename + '. Can you get to it with a browser?') return filename filename = maybe_download('text8.zip', 31344016) # Read the data into a list of strings. def read_data(filename): with zipfile.ZipFile(filename) as f: data = tf.compat.as_str(f.read(f.namelist()[0])).split() return data words = read_data(filename) print('Data size', len(words)) # Step 2: Build the dictionary and replace rare words with UNK token. vocabulary_size = 50000 def build_dataset(words): count = [['UNK', -1]] count.extend(collections.Counter(words).most_common(vocabulary_size - 1)) dictionary = dict() for word, _ in count: dictionary[word] = len(dictionary) data = list() unk_count = 0 for word in words: if word in dictionary: index = dictionary[word] else: index = 0 # dictionary['UNK'] unk_count += 1 data.append(index) count[0][1] = unk_count reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys())) return data, count, dictionary, reverse_dictionary data, count, dictionary, reverse_dictionary = build_dataset(words) del words # Hint to reduce memory. print('Most common words (+UNK)', count[:5]) print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]]) data_index = 0 # Step 3: Function to generate a training batch for the skip-gram model. def generate_batch(batch_size, num_skips, skip_window): global data_index assert batch_size % num_skips == 0 assert num_skips <= 2 * skip_window batch = np.ndarray(shape=(batch_size), dtype=np.int32) labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32) span = 2 * skip_window + 1 # [ skip_window target skip_window ] buffer = collections.deque(maxlen=span) for _ in range(span): buffer.append(data[data_index]) data_index = (data_index + 1) % len(data) for i in range(batch_size // num_skips): target = skip_window # target label at the center of the buffer targets_to_avoid = [ skip_window ] for j in range(num_skips): while target in targets_to_avoid: target = random.randint(0, span - 1) targets_to_avoid.append(target) batch[i * num_skips + j] = buffer[skip_window] labels[i * num_skips + j, 0] = buffer[target] buffer.append(data[data_index]) data_index = (data_index + 1) % len(data) return batch, labels batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1) for i in range(8): print(batch[i], reverse_dictionary[batch[i]], '->', labels[i, 0], reverse_dictionary[labels[i, 0]]) # Step 4: Build and train a skip-gram model. batch_size = 128 embedding_size = 128 # Dimension of the embedding vector. skip_window = 1 # How many words to consider left and right. num_skips = 2 # How many times to reuse an input to generate a label. valid_size = 16 # Random set of words to evaluate similarity on. valid_window = 100 # Only pick dev samples in the head of the distribution. valid_examples = np.random.choice(valid_window, valid_size, replace=False) num_sampled = 64 # Number of negative examples to sample. graph = tf.Graph() with graph.as_default(): # Input data. train_inputs = tf.placeholder(tf.int32, shape=[batch_size]) train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1]) valid_dataset = tf.constant(valid_examples, dtype=tf.int32) # Ops and variables pinned to the CPU because of missing GPU implementation with tf.device('/cpu:0'): # Look up embeddings for inputs. embeddings = tf.Variable( tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0)) embed = tf.nn.embedding_lookup(embeddings, train_inputs) # Construct the variables for the NCE loss nce_weights = tf.Variable( tf.truncated_normal([vocabulary_size, embedding_size], stddev=1.0 / math.sqrt(embedding_size))) nce_biases = tf.Variable(tf.zeros([vocabulary_size])) loss = tf.reduce_mean( tf.nn.nce_loss(weights=nce_weights, biases=nce_biases, labels=train_labels, inputs=embed, num_sampled=num_sampled, num_classes=vocabulary_size)) # Construct the SGD optimizer using a learning rate of 1.0. optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss) # Compute the cosine similarity between minibatch examples and all embeddings. norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True)) normalized_embeddings = embeddings / norm valid_embeddings = tf.nn.embedding_lookup( normalized_embeddings, valid_dataset) similarity = tf.matmul( valid_embeddings, normalized_embeddings, transpose_b=True) # Add variable initializer. init = tf.global_variables_initializer() # Step 5: Begin training. num_steps = 100001 with tf.Session(graph=graph) as session: init.run() print("Initialized") average_loss = 0 for step in range(num_steps): batch_inputs, batch_labels = generate_batch( batch_size, num_skips, skip_window) feed_dict = {train_inputs : batch_inputs, train_labels : batch_labels} _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict) average_loss += loss_val if step % 2000 == 0: if step > 0: average_loss /= 2000 # The average loss is an estimate of the loss over the last 2000 batches. print("Average loss at step ", step, ": ", average_loss) average_loss = 0 # Note that this is expensive (~20% slowdown if computed every 500 steps) if step % 10000 == 0: sim = similarity.eval() for i in range(valid_size): valid_word = reverse_dictionary[valid_examples[i]] top_k = 8 # number of nearest neighbors nearest = (-sim[i, :]).argsort()[1:top_k+1] log_str = "Nearest to %s:" % valid_word for k in range(top_k): close_word = reverse_dictionary[nearest[k]] log_str = "%s %s," % (log_str, close_word) print(log_str) final_embeddings = normalized_embeddings.eval() # Step 6: Visualize the embeddings. def plot_with_labels(low_dim_embs, labels, filename='tsne.png'): assert low_dim_embs.shape[0] >= len(labels), "More labels than embeddings" plt.figure(figsize=(18, 18)) #in inches for i, label in enumerate(labels): x, y = low_dim_embs[i,:] plt.scatter(x, y) plt.annotate(label, xy=(x, y), xytext=(5, 2), textcoords='offset points', ha='right', va='bottom') plt.savefig(filename) #%% try: from sklearn.manifold import TSNE import matplotlib.pyplot as plt tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000) plot_only = 200 low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only,:]) labels = [reverse_dictionary[i] for i in range(plot_only)] plot_with_labels(low_dim_embs, labels) except ImportError: print("Please install sklearn, matplotlib, and scipy to visualize embeddings.")
参考资料:《TensorFlow实践》
欢迎付费咨询(150元每小时),我的微信:qingxingfengzi