Tensorflow— word2vec

代码:

# 导入包
import tensorflow as tf
# encoding=utf8  
from __future__ import absolute_import
from __future__ import division
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

代码:

# Step 1: Download the data.
url = 'http://mattmahoney.net/dc/'

# 下载数据集
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)
    # 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)

# Read the data into a list of strings.
def read_data(filename):
    """Extract the first file enclosed in a zip file as a list of words"""
    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)

# Data size
print('Data size', len(words))

运行结果:

Found and verified text8.zip
Data size 17005207

代码:

# Step 2: Build the dictionary and replace rare words with UNK token.
# 建立字典
# 只留50000个单词,其他的词都归为UNK
vocabulary_size = 50000

def build_dataset(words, vocabulary_size):
    count = [['UNK', -1]]
    # extend追加一个列表
    # Counter用来统计每个词出现的次数
    # most_common返回一个TopN列表,只留50000个单词包括UNK  
    # c = Counter('abracadabra')
    # c.most_common()
    # [('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  # dictionary['UNK']
            unk_count += 1
        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)
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

运行结果:

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']

代码:

# 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 ]  3
    # 双向队列
    buffer = collections.deque(maxlen=span)
    # [ skip_window target skip_window ]
            # [ skip_window target skip_window ]
                    # [ skip_window target skip_window ]
            
#     [0 1 2 3 4 5 6 7 8 9 ...]
#            t     i  
    # 循环3次
    for _ in range(span):
        buffer.append(data[data_index])
        data_index = (data_index + 1) % len(data)
    # 获取batch和labels
    for i in range(batch_size // num_skips):
        target = skip_window  # target label at the center of the buffer
        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)
    # Backtrack a little bit to avoid skipping words in the end of a batch
    # 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]])

运行结果:

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

代码:

# Step 4: Build and train a skip-gram model.
batch_size = 128
# 词向量维度
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.

# 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.
valid_size = 16     # Random set of words to evaluate similarity on.
valid_window = 100  # Only pick dev samples in the head of the distribution.
# 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    # 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'):
        # 词向量
        # Look up embeddings for inputs.
    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)

    # Construct the variables for the noise-contrastive estimation(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]))

    # 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.
    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).minimize(loss)

    # Compute the cosine similarity between minibatch examples and all embeddings.
    norm = tf.sqrt(tf.reduce_sum(tf.square(embeddings), 1, keepdims=True))
    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)

    # Add variable initializer.
    init = tf.global_variables_initializer()

代码:

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

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 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_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

        # 计算训练2000次的平均loss
        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 % 20000 == 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  # number of nearest neighbors
                # 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()

运行结果:

Initialized
Average loss at step  0 :  302.383636475
Nearest to one: mitochondrial, succumbed, heligoland, ump, slowed, forges, exquisite, hardly,
Nearest to of: conformal, holbach, binitarianism, woodland, globe, coeur, artery, legacy,
Nearest to three: rooks, notorious, geelong, macrinus, airstrip, macy, fled, route,
Nearest to war: kadyrov, burbank, groin, geoffrey, machines, announcers, vascular, accidents,
Nearest to in: omari, scourge, esr, ethica, osman, sex, typhoid, panacea,
Nearest to are: ilium, ordain, reproductive, progesterone, calais, alphabetically, direct, porsche,
Nearest to were: moreau, biochemistry, innsbruck, slum, potencies, uncertainty, capacitive, autumn,
Nearest to the: singularities, differentials, hurriedly, kiribati, neverwinter, kabbalists, owe, jonathan,
Nearest to new: tamar, synthesized, midrashim, klux, tian, veronica, cervix, thermopylae,
Nearest to american: quarterly, psychopathic, motala, nudity, indentured, cooking, melancholia, groningen,
Nearest to they: dct, frye, theirs, athenaeus, obscura, andromache, rial, austere,
Nearest to its: shuffled, socratic, masterminded, post, among, painter, rifles, petty,
Nearest to seven: cautions, expounded, dm, doings, containing, afro, isle, cadre,
Nearest to which: risen, inevitable, toe, fis, geniuses, jim, dictatorship, insufficient,
Nearest to up: colliding, crushes, pritchard, mirza, communicate, soundhole, heir, jail,
Nearest to has: brenner, bulges, scandalous, doorman, hermeticism, escap, decrees, salvator,
Average loss at step  2000 :  113.390164988
Average loss at step  4000 :  52.6187440126
Average loss at step  6000 :  33.286626014
Average loss at step  8000 :  23.7020297694
Average loss at step  10000 :  17.7729295442
Average loss at step  12000 :  14.1117788121
Average loss at step  14000 :  11.771555624
Average loss at step  16000 :  9.96325901306
Average loss at step  18000 :  8.48758016729
Average loss at step  20000 :  8.13723055959
Nearest to one: two, operatorname, eight, six, nine, three, four, five,
Nearest to of: and, in, for, dasyprocta, with, between, nine, s,
Nearest to three: eight, four, two, nine, zero, seven, six, operatorname,
Nearest to war: vocals, feast, hundreds, machines, aoc, voluntarily, geoffrey, coimbra,
Nearest to in: and, of, on, for, at, from, with, by,
Nearest to are: were, is, was, ilium, zero, tiny, by, would,
Nearest to were: are, was, is, capable, and, transportation, arkham, modestly,
Nearest to the: a, dasyprocta, one, his, operatorname, this, their, circ,
Nearest to new: tamar, random, synthesized, veronica, tian, of, mathbf, readable,
Nearest to american: quarterly, s, bckgr, feminist, and, d, subsistence, helps,
Nearest to they: there, frye, he, anglicans, theirs, creating, elite, often,
Nearest to its: the, his, dasyprocta, eichmann, circ, a, en, ancestors,
Nearest to seven: nine, eight, four, zero, six, three, two, five,
Nearest to which: that, and, this, agincourt, tissue, the, dictatorship, toe,
Nearest to up: module, his, blacks, partners, alien, agouti, adding, austin,
Nearest to has: had, is, was, have, scandalous, decrees, kubitzki, marlon,
Average loss at step  22000 :  7.06678244722
Average loss at step  24000 :  6.85123083913
Average loss at step  26000 :  6.81250965095
Average loss at step  28000 :  6.34419031024
Average loss at step  30000 :  5.9245999701
Average loss at step  32000 :  5.93673320675
Average loss at step  34000 :  5.70171094501
Average loss at step  36000 :  5.74694340336
Average loss at step  38000 :  5.50405748427
Average loss at step  40000 :  5.25322429836
Nearest to one: two, eight, four, six, three, seven, zero, operatorname,
Nearest to of: zero, and, in, dasyprocta, for, agouti, recitative, eight,
Nearest to three: four, six, eight, five, seven, two, zero, one,
Nearest to war: feast, brass, machines, vocals, automobile, voluntarily, vascular, aoc,
Nearest to in: zero, and, at, on, dasyprocta, from, during, of,
Nearest to are: were, is, zero, was, progesterone, abet, have, calais,
Nearest to were: are, was, is, be, zero, have, had, by,
Nearest to the: its, dasyprocta, his, their, zero, agouti, operatorname, circ,
Nearest to new: tamar, veronica, synthesized, random, midrashim, ancestor, dasyprocta, elephant,
Nearest to american: and, quarterly, abakan, zero, indentured, bckgr, feminist, vma,
Nearest to they: there, he, it, we, not, deport, i, discard,
Nearest to its: the, their, his, dasyprocta, a, circ, some, zero,
Nearest to seven: six, eight, four, five, nine, zero, three, two,
Nearest to which: that, this, also, it, and, tissue, agincourt, one,
Nearest to up: module, mirza, recitative, enabling, partners, abandoning, cyanobacteria, mg,
Nearest to has: had, was, is, have, scandalous, amalthea, decrees, aba,
Average loss at step  42000 :  5.36403241181
Average loss at step  44000 :  5.27934718394
Average loss at step  46000 :  5.25050886309
Average loss at step  48000 :  5.24700605953
Average loss at step  50000 :  4.9966404134
Average loss at step  52000 :  5.03326895094
Average loss at step  54000 :  5.17822365785
Average loss at step  56000 :  5.04268380868
Average loss at step  58000 :  5.06483457124
Average loss at step  60000 :  4.93359541976
Nearest to one: two, four, three, six, five, eight, operatorname, seven,
Nearest to of: and, for, in, nine, dasyprocta, eight, including, six,
Nearest to three: four, five, two, six, eight, seven, operatorname, one,
Nearest to war: machines, boreal, feast, brass, automobile, rebellious, kadyrov, denigrating,
Nearest to in: from, during, at, dasyprocta, on, and, microsite, kapoor,
Nearest to are: were, is, have, was, zero, michelob, other, be,
Nearest to were: are, was, had, have, is, be, by, zero,
Nearest to the: its, dasyprocta, their, circ, recitative, his, a, operatorname,
Nearest to new: tamar, random, veronica, synthesized, member, xb, midrashim, tian,
Nearest to american: and, abakan, british, quarterly, abercrombie, indentured, feminist, bckgr,
Nearest to they: he, there, we, it, you, i, not, who,
Nearest to its: their, his, the, dasyprocta, bckgr, some, circ, dddddd,
Nearest to seven: eight, six, five, nine, four, three, zero, operatorname,
Nearest to which: this, that, also, it, but, ursus, one, wct,
Nearest to up: module, them, mirza, recitative, partners, enabling, cyanobacteria, abandoning,
Nearest to has: had, have, was, is, ursus, wct, decrees, amalthea,
Average loss at step  62000 :  4.99505268264
Average loss at step  64000 :  4.82697634709
Average loss at step  66000 :  4.59925288892
Average loss at step  68000 :  4.98079027224
Average loss at step  70000 :  4.89412822211
Average loss at step  72000 :  4.74675208092
Average loss at step  74000 :  4.80340922415
Average loss at step  76000 :  4.72690085912
Average loss at step  78000 :  4.79733606535
Average loss at step  80000 :  4.80540977299
Nearest to one: seven, six, two, five, four, operatorname, ursus, three,
Nearest to of: mico, in, dasyprocta, including, abet, kapoor, abakan, original,
Nearest to three: six, four, two, five, seven, eight, operatorname, lymphoma,
Nearest to war: machines, boreal, automobile, feast, kadyrov, brass, rebellious, geophysical,
Nearest to in: during, at, dasyprocta, from, ursus, and, of, under,
Nearest to are: were, is, have, be, was, michelob, britney, these,
Nearest to were: are, was, have, had, be, by, been, ursus,
Nearest to the: their, dasyprocta, a, kapoor, its, operatorname, iit, circ,
Nearest to new: tamar, random, member, veronica, synthesized, xb, dasyprocta, electrical,
Nearest to american: british, abakan, abercrombie, mico, indentured, bckgr, nunnery, indian,
Nearest to they: there, he, we, you, it, who, she, not,
Nearest to its: their, his, the, dasyprocta, dddddd, bckgr, her, iit,
Nearest to seven: six, eight, five, four, nine, three, one, two,
Nearest to which: that, this, also, but, it, ursus, what, and,
Nearest to up: filmfour, them, module, mirza, him, out, abandoning, recitative,
Nearest to has: had, have, was, is, ursus, decrees, wct, amalthea,
Average loss at step  82000 :  4.75804372787
Average loss at step  84000 :  4.75755859768
Average loss at step  86000 :  4.77840362
Average loss at step  88000 :  4.74728782678
Average loss at step  90000 :  4.73435067379
Average loss at step  92000 :  4.66841691899
Average loss at step  94000 :  4.72742706275
Average loss at step  96000 :  4.69911255908
Average loss at step  98000 :  4.60135727322
Average loss at step  100000 :  4.70121149051
Nearest to one: two, six, seven, five, four, operatorname, three, eight,
Nearest to of: mico, in, dasyprocta, including, cebus, globemaster, kapoor, same,
Nearest to three: four, five, six, seven, two, eight, operatorname, lymphoma,
Nearest to war: boreal, automobile, brass, feast, machines, kadyrov, hanna, cotswold,
Nearest to in: during, at, on, dasyprocta, microsite, from, under, within,
Nearest to are: were, is, have, these, be, michelob, britney, while,
Nearest to were: are, was, have, had, be, is, by, been,
Nearest to the: dasyprocta, their, a, its, agouti, kapoor, iit, this,
Nearest to new: tamar, random, veronica, member, synthesized, xb, mishnayot, dasyprocta,
Nearest to american: british, abakan, indian, abercrombie, mico, and, quarterly, bckgr,
Nearest to they: he, there, we, you, it, she, not, who,
Nearest to its: their, his, the, dasyprocta, her, elwes, bckgr, some,
Nearest to seven: eight, six, five, four, nine, three, zero, two,
Nearest to which: that, this, but, also, what, it, ursus, and,
Nearest to up: them, filmfour, out, him, module, abandoning, recitative, mirza,
Nearest to has: had, have, was, is, ursus, stationary, wct, globemaster,

代码:

# 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))  # 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, 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.")

运行结果:


代码:

with tf.Session() as session:
    valid_word = "one"
    valid_examples = dictionary[valid_word]
    valid_dataset = tf.constant([valid_examples], dtype=tf.int32)
    valid_embeddings = tf.nn.embedding_lookup(final_embeddings, valid_dataset)
    similarity = tf.matmul(valid_embeddings, final_embeddings, transpose_b=True)
    sim = similarity.eval()

    top_k = 8  # number of nearest neighbors
    nearest = (-sim[0]).argsort()[1:top_k + 1]# 排除自己本身,从小到大排序
    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)

运行结果:

Nearest to one: two, six, seven, five, four, operatorname, three, eight,


评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值