tensorflow实现RNN及Word2Vec

参考：《tensorflow实战》

首先介绍一下Word2Vec

Word2Vec：从原始语料中学习字词空间向量的预测模型。主要分为CBOW（Continue Bags of Words）连续词袋模型和Skip-Gram两种模式

CBOW：从原始语句（中国的首都是___）推测目标字词（北京）。Skip-Gram正好相反，从目标词反推原始语句。

预测模型使用最大似然的方法。在给定前面的语句h的情况下，最大化目标词汇的概率。比如（中国的___是北京），首都就是我们的目标词汇。

使用NCE（噪声对比估计 Noise-Contrastive Estimation）作为损失函数。

NCE：把上下文h对应的正确的词汇标记为正样本D=1，再抽取一些错误的词汇作为负样本（D=0），然后最大化目标函数的值

基于Skip-Gram的Word2Vec

import collections
import math
import os
import random
import zipfile
import numpy as np
import urllib
import tensorflow as tf
from six.moves import xrange  # pylint: disable=redefined-builtin
import matplotlib as plt

一、下载数据

下载完数据之后，当前文件夹下有text8.zip文件

# 下载数据
# 如何下载过了，就不需要执行这段代码
url = 'http://mattmahoney.net/dc/'

def maybe_download(filename, expected_bytes):
    if not os.path.exists(filename):
        filename, _ = urllib.request.urlretrieve(url + filename, filename)
    # 获取文件相关属性
    statinfo = os.stat(filename)
    # 比对文件的大小是否正确
    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)

Found and verified text8.zip

二、解压下载的压缩文件

# 解压下载的压缩文件
def read_data(filename):
    with zipfile.ZipFile(filename) as f:
        data = tf.compat.as_str(f.read(f.namelist()[0])).split()
    return data

# 单词表
words = read_data(filename)
print('Data size', len(words))

Data size 17005207

三、创建vocabulary词汇表，取top50000频数的单词

# 只留50000个单词，其他的词都归为UNK
# UNK ： 不认识的词
vocabulary_size = 50000

def build_dataset(words, vocabulary_size):
    count = [['UNK', -1]]
    # extend追加一个列表
    # Counter用来统计没个词出现的次数
    # most_common返回一个Top列表，只留50000个单词包括UNK
    # c = Counter('abracadabra')
    # c.most_commom()
    # [('a', 5), ('r', 2), ('b', 2), ('c', 1), ('d', 1)]
    # c.most_common(3)
    # [('a', 5), ('r', 2), ('b', 2)]
    # 前50000个出现次数最多的词
    count.extend(collections.Counter(words).most_common(vocabulary_size - 1))
    # 生成 dictionary，词对应编号, word:id(0-49999)
    # 词频越高编号越小
    dictionary = dict()
    for word, _ in count:
        dictionary[word] = len(dictionary)
    # data把数据集的词都编号
    data = list()
    unk_count = 0
    for word in words:
        if word in dictionary:
            index = dictionary[word]
        else:
            index = 0
            unk_count += 1  # dictionary['UNK']
        data.append(index)
    # 记录UNK词的数量
    count[0][1] = unk_count
    # 编号对应词的字典
    reverse_dictionary = dict(zip(dictionary.values(), dictionary.keys()))
    return data, count, dictionary, reverse_dictionary

# data 数据集，编号形式
# count 前50000个出现次数最多的词
# dictionary 词对应编号
# reverse_dictionary 编号对应词
data, count, dictionary, reverse_dictionary = build_dataset(words, vocabulary_size)

# 删除原始单词列表，节约内存
# 打印最高频出现的词汇及其数量
del words
print('Most common words (+UNK)', count[:5])
print('Sample data', data[:10], [reverse_dictionary[i] for i in data[:10]])

Most common words (+UNK) [['UNK', 418391], ('the', 1061396), ('of', 593677), ('and', 416629), ('one', 411764)]
Sample data [5234, 3081, 12, 6, 195, 2, 3134, 46, 59, 156] ['anarchism', 'originated', 'as', 'a', 'term', 'of', 'abuse', 'first', 'used', 'against']

四、生成Word2Vec的训练样本

1. 使用Skip-Gram模式（从目标单词反推语境）

data_index = 0

def generate_batch(batch_size, num_skips, skip_window):
    # skip_window ：单词最远可以联系的距离，设为1代表只能跟紧邻的两个单词生成样本
    # num_skips：每个单词生成多少样本
    # batch_size必须是num_skips的整数倍（确保每个batch包含了一个词汇对应的所有样本）
    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
    buffer = collections.deque(maxlen=span)
    
    for _ in range(span):
        buffer.append(data[data_index])
        data_index = (data_index + 1) % len(data)
    # 获取batch和label
    for i in range(batch_size // num_skips):
        target = skip_window
        targets_to_avoid = [skip_window]
        # 循环2次，一个目标单词对应两个上下文单词
        for j in range(num_skips):
            while target in targets_to_avoid:
                # 可能先拿到前面的单词也可能先拿到后面的单词
                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)
    # 回溯3个词。因为执行完一个batch的操作之后，data_index会往右多偏移span个位置
    data_index = (data_index + len(data) - span) % len(data)
    return batch, labels

# 打印sample data
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]])

3081 originated -> 5234 anarchism
3081 originated -> 12 as
12 as -> 6 a
12 as -> 3081 originated
6 a -> 195 term
6 a -> 12 as
195 term -> 2 of
195 term -> 6 a

五、建立和训练一个skip-gram模型

# 建立和训练一个skip-gram模型
batch_size = 128
# 词向量维度
embedding_size = 128
skip_windows = 1
num_skips = 2

valid_size = 16
valid_window = 100
# 从0-100抽取16个整数，无放回抽样
valid_examples = np.random.choice(valid_window, valid_size, replace=False)
# 负采样样本数
num_sampled = 64

graph = tf.Graph()
with graph.as_default():
    # 输入数据
    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)

    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
    # 比如说，ids=[1,7,4],就是返回params中第1,7,4行。返回结果为由params的1,7,4行组成的tensor
    # 提取要训练的词
    embed = tf.nn.embedding_lookup(embeddings, train_inputs)
    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))
    optimizer = tf.train.GradientDescentOptimizer(1).minimize(loss)

    # embeddings 的L2范数
    norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keep_dims=True))
    # embeddings除以L2范数得到标准化后的 normalized_embeddings
    normalized_embeddings = embeddings / norm
    # 抽取一些常用词来测试余弦相似度
    valid_embeddings = tf.nn.embedding_lookup(normalized_embeddings, valid_dataset)
    # valid_size == 16
    # [16,1] * [1*50000] = [16,50000]
    similarity = tf.matmul(valid_embeddings, normalized_embeddings, transpose_b=True)

    init = tf.global_variables_initializer()

优化器为SGD，然后计算嵌入向量embedding的L2范数norm，在将embedding除以其L2范数得到标准化后的normalized_embeddings。使用tf.nn.embedding_lookup查询单词的嵌入向量，并计算验证单词的嵌入向量与词汇表中所有单词的相似性。

# Step 5: Begin training.
num_steps = 100001
final_embeddings = []

with tf.Session(graph=graph) as session:
    init.run()
    print("Initialized")

    average_loss = 0
    for step in xrange(num_steps):
        # 获取一个批次的target，以及对应的labels，都是编号形式的
        batch_inputs, batch_labels = generate_batch(batch_size, num_skips, skip_windows)
        feed_dict = {train_inputs:batch_inputs, train_labels:batch_labels}

        _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
        average_loss += loss_val

        # 计算训练2000次的平均loss
        if step % 2000 == 0:
            if step > 0:
                average_loss /= 2000
            print("Average loss at step ", step, ": ", average_loss)
            average_loss = 0

        if step & 2000 == 0:
            sim = similarity.eval()
            # 计算验证集的余弦相似度最高的词
            for i in xrange(valid_size):
                # 根据id拿到对应单词
                valid_word = reverse_dictionary[valid_examples[i]]
                top_k = 8
                # 从大到小排序，排除自己本身，取前top_k个值
                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)
    # 训练结束得到的词向量
    final_embeddings = normalized_embeddings.eval()

Initialized
Average loss at step  0 :  296.63623046875
Nearest to b: game, ampersand, odour, reorganize, relented, missiles, svetlana, sustains,
......

Nearest to to: would, can, through, ursus, renouf, abet, circ, for,
Nearest to also: which, often, now, still, operatorname, apatosaurus, capitalists, not,

Average loss at step  100000 :  4.69689565706253

六、可视化效果函数

# 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"
    # 设置图片大小
    plt.figure(figsize = (15, 15))
    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, method='exact')# mac：method='exact'
    # 画500个点
    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, matplotlib, and scipy to visualize embeddings."

最后生成的图片如下，效果还可以。距离相近的单词在语义上具有很高的相似度。