实现word2vec中skip_gram模型
建立图形
先定义一个嵌套参数矩阵。我们用唯一的随机值来初始化这个大矩阵。
embeddings = tf.Variable(
tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
对噪声-比对的损失计算就使用一个逻辑回归模型。对此,我们需要对语料库中的每个单词定义一个权重值和偏差值。(也可称之为输出权重 与之对应的 输入嵌套值)。定义如下。
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]))
Skip-Gram模型有两个输入。一个是一组用整型表示的上下文单词,另一个是目标单词。给这些输入建立占位符节点,之后就可以填入数据了。
# 建立输入占位符
train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
然后我们需要对批数据中的单词建立嵌套向量,TensorFlow提供了方便的工具函数。
embed = tf.nn.embedding_lookup(embeddings, train_inputs)
好了,现在我们有了每个单词的嵌套向量,接下来就是使用噪声-比对的训练方式来预测目标单词。
# 计算 NCE 损失函数, 每次使用负标签的样本.
loss = tf.reduce_mean(
tf.nn.nce_loss(nce_weights, nce_biases, embed, train_labels,num_sampled, vocabulary_size))
我们对损失函数建立了图形节点,然后我们需要计算相应梯度和更新参数的节点,比如说在这里我们会使用随机梯度下降法,TensorFlow也已经封装好了该过程。
# 使用 SGD 控制器.
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0).minimize(loss)
训练模型
训练的过程很简单,只要在循环中使用feed_dict不断给占位符填充数据,同时调用 session.run即可。
for step in range(num_steps):
# 先使用generate_batch生成一个batch的inputs和labels数据
batch_inputs, batch_labels = generate_batch(batch_size, num_skips, skip_window)
# 并用他们创建feed_dict
feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}
# 然后执行一次优化器运算(即一次参数更新)和损失计算
_, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
完整代码如下:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import collections
import math
import os
import random
from tempfile import gettempdir
import zipfile
import numpy as np
import urllib
# from six.moves import urllib
# from six.moves import xrange # pylint: disable=redefined-builtin
import tensorflow as tf
# 解压下载的压缩文件,并使用tf.compat.as_str将数据转成单词的列表。
def read_data(filename):
with zipfile.ZipFile(filename) as f:
data = tf.compat.as_str(f.read(f.namelist()[0])).split()
return data
# vocabulary是将文本分词后的所有单词,按照文本顺序保存
vocabulary = read_data("/Users/iris/Documents/数据集/text8.zip")
print('Data size', len(vocabulary))
vocabulary_size = 50000
# 创建vocabulary词汇表
def build_dataset(words, n_words):
# 使用collections.Counter统计单词列表中单词的频数,然后使用most_common方法提取top 50000频数的单词作为vocabulary
count = [['UNK', -1]]
count.extend(collections.Counter(words).most_common(n_words - 1)) # count[['UNK', -1],[word,word频数],。。。]
# 将top 50000 词汇的vocabulary转为编号,以频数排序
dictionary = dict()
for word, _ in count:
dictionary[word] = len(dictionary) # dictionary [word: 编号]
# 遍历单词列表,判断每个单词是否出现在dictionary中,是则转换为dictionary中按频数排序的编号,不是则转为编号0
data = list()
unk_count = 0
for word in words:
index = dictionary.get(word, 0) # 查找word,若不存在返回默认值0
if index == 0: # dictionary['UNK']
unk_count += 1
data.append(index) # data (w1编号,w2编号,。。。,wn编号)
# 更新count的第一个数据
count[0][1] = unk_count
reversed_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
# 最后返回转换后的编码(data)、每个单词的频数统计(count)、词汇表(dictionary)及其反转的形式(reverse_dictionary)
return data, count, dictionary, reversed_dictionary
data, count, dictionary, reverse_dictionary = build_dataset(vocabulary, vocabulary_size)
# 删除原始单词列表,可以节约内存,再打印出最高频出现的词汇及其数量
del vocabulary # 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 # 定义data_index为全局变量,因为我们会反复调用generate_batch
# 生成训练用的batch数据:
# batch_size为batch的大小
# num_skips为对每个单词生成多少个样本(不能大于skip_window值的两倍,batch_size必须是它的整数倍(确保每个batch包含了一个词汇对应的所有样本))
# skip_window指单词最远可以联系的距离
def generate_batch(batch_size, num_skips, skip_window):
global data_index
# 断言是声明其布尔值必须为真的判定,如果发生异常就说明表达示为假
assert batch_size % num_skips == 0
assert num_skips <= 2 * skip_window
# np.ndarray 初始化为数组,shape 生成数组形状,dtype 数据类型
batch = np.ndarray(shape=(batch_size), dtype=np.int32)
labels = np.ndarray(shape=(batch_size, 1), dtype=np.int32)
# span为对某个单词创建相关样本时会使用到的单词数量,包括目标单词本身和它前后的单词
span = 2 * skip_window + 1
# deque是double-ended queue的缩写,双端队列结构。最大容量为span,在对deque执行append方法添加变量时,只会保留最后插入的span个变量
buffer = collections.deque(maxlen=span)
if data_index + span > len(data):
data_index = 0
buffer.extend(data[data_index:data_index + span])
data_index += span
# 每次循环对一个目标单词生成样本
for i in range(batch_size // num_skips): # //整数除法
context_words = [w for w in range(span) if w != skip_window] # 输出0~span中所有数值,除了skip_window
# print(context_words)
words_to_use = random.sample(context_words, num_skips)
# print(words_to_use)
for j, context_word in enumerate(words_to_use):
batch[i * num_skips + j] = buffer[skip_window]
labels[i * num_skips + j, 0] = buffer[context_word]
if data_index == len(data):
buffer[:] = data[:span]
data_index = span
else:
buffer.append(data[data_index])
data_index += 1
data_index = (data_index + len(data) - span) % len(data)
return batch, labels
# 这里调用generate_batch函数测试一下其功能
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]])
##########################################################################################################
# Step1、2 定义输入和定义算法公式
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.
num_sampled = 64 # Number of negative examples to sample.
valid_size = 16 # Random set of words to evaluate similarity on.用来抽取的验证单词数
valid_window = 100 # Only pick dev samples in the head of the distribution.指验证单词只从频数最高的100个单词中抽取
valid_examples = np.random.choice(valid_window, valid_size, replace=False) # 使用np.random.choice函数进行随机抽取
# 下面开始定义Skip-Gram WordVec模型的网络结构
graph = tf.Graph()
# 返回一个上下文管理器,使得这个Graph对象成为当前默认的graph.
# with关键字让这个代码块内创建的从操作(ops)添加到这个新的图里面.
# 让with下面代码块的操作都执行在这个图中,作为默认graph
with graph.as_default():
# Input data. 创建训练数据中的inputs和labels的placeholder,placeholder给节点输入数据
# placeholder是占位符,相当于java中的定义
train_inputs = tf.placeholder(tf.int32, shape=[batch_size])
train_labels = tf.placeholder(tf.int32, shape=[batch_size, 1])
# 将前面产生的valid_examples转为TensorFlow中的constant
valid_dataset = tf.constant(valid_examples, dtype=tf.int32)
# 限定所有计算再CPU上执行
# 以下代码块操作是初始化所有参数
with tf.device('/cpu:0'):
# 随机生成所有单词的词向量embeddings
embeddings = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
# embedding_lookup(params, ids)其实就是按照ids顺序返回params中的第ids行。
# 在embeddings中查找train_inputs对应的表示,赋值给embed
embed = tf.nn.embedding_lookup(embeddings, train_inputs)
# NCE Loss作为训练的优化目标
# 使用truncated_normal初始化NCE Loss中的权重参数nce_weights
nce_weights = tf.Variable(tf.truncated_normal([vocabulary_size, embedding_size], stddev=1.0 / math.sqrt(embedding_size)))
# 将nce_biases初始化为0
nce_biases = tf.Variable(tf.zeros([vocabulary_size]))
# Step3 定义损失函数
# 使用tf.nn.nce_loss计算学习出的词向量embedding在训练数据上的loss,并使用tf.reduce_mean进行汇总
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))
# Step4 定义优化算法
# 定义优化器为SGD并且学习速率为1.0
optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)
# Compute the cosine similarity between minibatch examples and all embeddings.
# 计算embeddings的L2范数norm
norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
# embeddings除以L2范数到标准化后的normalized_embeddings
normalized_embeddings = embeddings / norm
# 查询验证单词的嵌入向量embeddings
valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings, valid_dataset)
# 计算并验证单词的嵌入向量与词汇表中所有单词的相似性
similarity = tf.matmul(valid_embeddings, normalized_embeddings, transpose_b=True)
# Add variable initializer.
init = tf.global_variables_initializer() # 最后初始化所有模型参数
# Step5 迭代执行训练操作
num_steps = 100001
with tf.Session(graph=graph) as session:
# We must initialize all variables before we use them.
init.run()
print('Initialized')
average_loss = 0
for step in range(num_steps):
# 先使用generate_batch生成一个batch的inputs和labels数据
batch_inputs, batch_labels = generate_batch(batch_size, num_skips, skip_window)
# 并用他们创建feed_dict
feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}
# 然后执行一次优化器运算(即一次参数更新)和损失计算
_, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
average_loss += loss_val # 并将这一步训练的loss累积到average_loss
# 之后每2000次循环,计算一下平均loss并显示出来
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
# 每10000次循环,计算一次验证单词与全部单词的相似度,并将与每个验证单词最相似的8个单词展示出来
# 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.
# pylint: disable=missing-docstring
# Function to draw visualization of distance between embeddings.
# 下面定义一个用来可视化Word2Vec效果的函数,low_dim_embs是降维到2维的单词的空间向量,将在图表中展示每个单词的位置
def plot_with_labels(low_dim_embs, labels, filename):
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, :]
# 显示散点图
plt.scatter(x, y)
# 展示单词本身
plt.annotate(label,
xy=(x, y),
xytext=(5, 2),
textcoords='offset points',
ha='right',
va='bottom')
# 保存图片到本地文件
plt.savefig(filename)
# 实现降维,将原始的128维降到2维,再使用plot_with_labels进行展示,这里只展示词频最高的100个单词的可视化结果
try:
# pylint: disable=g-import-not-at-top
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=5000, method='exact')
plot_only = 500
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, os.path.join(gettempdir(), 'tsne.png'))
except ImportError as ex:
print('Please install sklearn, matplotlib, and scipy to show embeddings.')
print(ex)