本文对谷歌开源的词向量生成工具 word2vec_basic.py 进行了注释，以备自己和需要了解其运行原理的同学理解其背后的机理，有注解不当之处还请及时提出，本人会不断修正，争取每一行代码都有正确和清晰的注解。
可以分为如下几个步骤。
- 下载数据，将数据读取为列表格式
from __future__ import absolute_import from __future__ import print_function import collections import math import os import random import zipfile import numpy as np from six.moves import urllib from six.moves import xrange # pylint: disable=redefined-builtin import tensorflow as tf # Step 1: 下载数据 url = 'http://mattmahoney.net/dc/' # 检查期望的文件大小和实际大小是否符合 def maybe_download(filename, expected_bytes): """Download a file if not present, and make sure it's the right size.""" 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) # 将文件内容读取为列表格式 def read_data(filename): f = zipfile.ZipFile(filename) for name in f.namelist(): return f.read(name).split() f.close() words = read_data(filename) print('Data size', len(words)) - 构造词典和逆向词典等训练数据，定义miniBatch生成方法
vocabulary_size = 50000 # Step 2: Build the dictionary and replace rare words with UNK token. ''' 构建数据，为后面训练模型做准备 输入：words, 单词列表 输出：data, words中每个词在dictionary中对应的值, 即词的索引 count, 词与词频的元组列表，其中'UNK'表示所有非高频词的数量 dictionary, 词与词索引的映射 reverse_dictionary, dictionary的逆向映射 ''' def build_dataset(words): count = [['UNK', -1]] # 选取words中词频最高的49999个词，以(word, frequency)的格式加入count # 此时count内部的元素排列如：[['UNK',-1], ('is',2), ('name',5), ...] count.extend(collections.Counter(words).most_common(vocabulary_size - 1)) # 创建词到词索引的映射，如: {'UNK' : 1, 'is' : 200, 'name' : 30, ... } dictionary = dict() for word, _ in count: dictionary[word] = len(dictionary) # 对words中的每个词，将其在dictionary中的索引取出，放到data列表中 # 最终的data列表，如[1,3,2,5,20,0,2,0,0,9,...]，0表示某个词不在词频最高的49999个词当中 # 同时，计数'UNK'的数量 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 # 构建从词索引到词的逆向字典 reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys())) return data, count, dictionary, reverse_dictionary data, count, dictionary, reverse_dictionary = build_dataset(words) # Hint to reduce memory. del words # 输出最高频的5个词，打印data中的前10个索引值 print('Most common words (+UNK)', count[:5]) print('Sample data', data[:10]) data_index = 0 ''' 为skip-gram模型生成一个batch的训练样本集合 输入：batch_size, batch大小 num_skips, 每个窗口产生多少样本 skip_window, 窗口大小 输出：batch, 中心词列表 labels, 标签列表 ''' def generate_batch(batch_size, num_skips, skip_window): global data_index # batch_size需是num_skips的整数倍 assert batch_size % num_skips == 0 # num_skips的上限，不能超过2倍窗口大小 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) # 窗口左边界到右边界之间的长度 # 样例：[skip_window target skip_window] # 若skip_window是1，则span是3；若skip_window是2，则span是5 span = 2 * skip_window + 1 # 建立一个容量为span的队列，当容量大于span时，若有插入操作，最老的元素出队 # 队列容量始终保持为span buffer = collections.deque(maxlen=span) # 将词索引列表data中[data_index : data_index+span]范围内的词索引加入buffer # buffer本质就是一个滑动窗口 # 若data为[1,4,3,5,2,5], 则当前窗口为[1,4,3], 中心词索引为4 for _ in range(span): buffer.append(data[data_index]) data_index = (data_index + 1) % len(data) # 对batch_size / num_skips个窗口进行样本采集 # 注：一个样本由一个中心词以及该中心词的标签(即中心词要预测的词)构成，例如 # 在句子I like to eat apple中，(eat, apple)为生成的样本，eat为中心词，apple为该中心词 对应的标签 # 注：中心词总是对应当前窗口的中心位置 for i in range(batch_size // num_skips): target = skip_window # target label at the center of the buffer targets_to_avoid = [ skip_window ] # 在第i个窗口中采集num_skips个样本 for j in range(num_skips): # 在[0,span)区间(该区间的大小对应窗口的大小)进行随机无放回采样， # 直至发现不在targets_to_avoid的值为止， # 搜索到的target下次将不再采样. while target in targets_to_avoid: target = random.randint(0, span - 1) targets_to_avoid.append(target) # 加入中心词以及中心词对应的标签 batch[i * num_skips + j] = buffer[skip_window] labels[i * num_skips + j, 0] = buffer[target] # 窗口右移一个位置，假设data为[1,4,3,5,2,5] # 若当前窗口为[1,4,3], 则更新后的窗口为[4,3,5], 中心词的索引为3 buffer.append(data[data_index]) data_index = (data_index + 1) % len(data) return batch, labels # 生成batch, 并采样打印 batch, labels = generate_batch(batch_size=8, num_skips=2, skip_window=1) for i in range(8): print(batch[i], '->', labels[i, 0]) print(reverse_dictionary[batch[i]], '->', reverse_dictionary[labels[i, 0]]) - 定义词向量训练过程
其中，关于NCE损失的介绍不在本文范围内，代码注释中给出了一个比较不错的说明文档，可供参考。
batch_size = 128 # batch大小 embedding_size = 128 # 嵌入层大小 skip_window = 1 # 窗口大小 # 如果窗口大小为2，则某个词附近一共有4个词可取到，最多可以组成4个训练样本， # 但如果你只需要2个样本的话则可通过num_skip来设置 # 如：I like to eat apple, skip_window = 1, num_skips = 2, 则 # 生成的样本是 (like, I), (like, to), (to, like), (to, eat), (eat, to), (eat, apple) num_skips = 2 # We pick a random validation set to sample nearest neighbors. Here we limit the # validation samples to the words that have a low numeric ID, which by # construction are also the most frequent. # 验证集构造，仅从词频最高的100个词中随机选择16个 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.array(random.sample(range(valid_window), valid_size)) # 负采样数量 num_sampled = 64 # Step 4: 构建并训练skip-gram模型. # 实例化一个用于tensorflow计算和表示用的数据流图 graph = tf.Graph() # 定义属于计算图graph的张量和操作 with graph.as_default(): # 中心词，标签，验证集 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. # 生成维度为[vocabulary_size, embedding_size], 且范围为[-1,1]的均匀分布随机浮点嵌入张量 embeddings = tf.Variable( tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0)) # 根据词索引查询词向量，本质就是一个查询映射定义 embed = tf.nn.embedding_lookup(embeddings, train_inputs) # Construct the variables for the NCE loss # 截断的产生正态分布的随机数，即随机数与均值的差值若大于两倍的标准差，则重新生成 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])) # Compute the average NCE loss for the batch. # tf.nce_loss automatically draws a new sample of the negative labels each # time we evaluate the loss. # 定义nce平均损失值计算 # 关于nce损失可参见：https://blog.csdn.net/qq_36092251/article/details/79684721 loss = tf.reduce_mean( tf.nn.nce_loss(nce_weights, nce_biases, embed, train_labels, num_sampled, vocabulary_size)) # Construct the SGD optimizer using a learning rate o # 定义随机梯度下降优化器最小化loss，初始学习率为1.0 optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss) # Compute the cosine similarity between minibatch examples and all embeddings. # 计算余弦相似度 # 对embeddings中的每个值规范化 norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True)) normalized_embeddings = embeddings / norm # 根据normalized_embeddings查询valid_dataset中每个词的词嵌入 valid_embeddings = tf.nn.embedding_lookup( normalized_embeddings, valid_dataset) # 计算相似度矩阵，首先要对normalized_embeddings进行矩阵转置 # 其中，valid_embeddings维度为16 * 128, normalized_embeddings维度为50000 * 128 # 这一步的输出维度是16 * 50000，每行表示某个词与词表中每个词的相似度 similarity = tf.matmul( valid_embeddings, normalized_embeddings, transpose_b=True) - 词向量训练
# 训练轮数 num_steps = 100001 with tf.Session(graph=graph) as session: # 训练之前，初始化所有需要用到的变量 tf.initialize_all_variables().run() print("Initialized") average_loss = 0 for step in xrange(num_steps): # 生成一个batch数据 batch_inputs, batch_labels = generate_batch( batch_size, num_skips, skip_window) feed_dict = {train_inputs : batch_inputs, train_labels : batch_labels} # We perform one update step by evaluating the optimizer op (including it # in the list of returned values for session.run() _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict) average_loss += loss_val # 每隔2000轮打印一次目前的平均损失 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) # 当训练完10000轮时，输出与验证集valid_examples中每个词最相近的8个词 # 因为这一步特别耗时，所以整个训练过程仅进行了10次 if step % 10000 == 0: sim = similarity.eval() for i in xrange(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 xrange(top_k): close_word = reverse_dictionary[nearest[k]] log_str = "%s %s," % (log_str, close_word) print(log_str) # 得到最终的词嵌入(经过了正规化) final_embeddings = normalized_embeddings.eval() - 可视化词嵌入
# 可视化词嵌入
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 = 500
low_dim_embs = tsne.fit_transform(final_embeddings[:plot_only,:])
labels = [reverse_dictionary[i] for i in xrange(plot_only)]
plot_with_labels(low_dim_embs, labels)
except ImportError:
print("Please install sklearn and matplotlib to visualize embeddings.")
1276
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
