tensorlayer学习日志14_chapter5_5.3

第5.3节讲的是文本简单处理，课程源码在这里： https://github.com/tensorlayer/tensorlayer/blob/master/example/tutorial_word2vec_basic.py

但是Step 7: Evaluate by analogy questions.我没能实现，主要是不知道，questions-words.txt是个什么东东，有机会再来试试吧~~

import sys
import time
import numpy as np
import tensorflow as tf
import tensorlayer as tl
from six.moves import xrange  


words = tl.files.load_matt_mahoney_text8_dataset()
data_size = len(words)
print(data_size)  
print(words[0:10])  

resume = False  
_UNK = "_UNK"

vocabulary_size = 50000 
batch_size = 128
embedding_size = 128  
skip_window = 1  
num_skips = 2  

num_sampled = 64 
learning_rate = 1.0
n_epoch = 12
model_file_name = "model_word2vec_50k_128"

num_steps = int((data_size / batch_size) * n_epoch)  

print('%d Steps in a Epoch, total Epochs %d' % (int(data_size / batch_size), n_epoch))
print('   learning_rate: %f' % learning_rate)
print('   batch_size: %d' % batch_size)

if resume:
    print("Load existing data and dictionaries" + "!" * 10)
    all_var = tl.files.load_npy_to_any(name=model_file_name + '.npy')
    data = all_var['data']
    count = all_var['count']
    dictionary = all_var['dictionary']
    reverse_dictionary = all_var['reverse_dictionary']
else:
    data, count, dictionary, reverse_dictionary = tl.nlp.build_words_dataset(words, vocabulary_size, True, _UNK)

print('~~~~~~~~~~~~~Most 5 common words (+UNK)~~~~~~~~~~~~')
print(count[:5])
print('~~~~~~~~~~~~~~~~~~~~~~Sample data~~~~~~~~~~~~~~~~~')
print(data[:10], [reverse_dictionary[i] for i in data[:10]])

print('~~~~~~~~~~~~~只看周围两个单词，左右各一个~~~~~~~~~~~~')
batch, labels, data_index = tl.nlp.generate_skip_gram_batch(data=data, \
    batch_size=8, num_skips=2, skip_window=1, data_index=0)
for i in range(8):
    print(batch[i], reverse_dictionary[batch[i]], '->', labels[i, 0], reverse_dictionary[labels[i, 0]])

print('~~~~~~~~~~~~~只看周围四个单词，左右各两个~~~~~~~~~~~~')

batch, labels, data_index = tl.nlp.generate_skip_gram_batch(data=data, \
    batch_size=8, num_skips=4, skip_window=2, data_index=0)
for i in range(8):
    print(batch[i], reverse_dictionary[batch[i]], '->', labels[i, 0], reverse_dictionary[labels[i, 0]])

print('~~~~~~~~~~~~~~Build Skip-Gram model~~~~~~~~~~~~~')

valid_size = 16  
valid_window = 100  
valid_examples = np.random.choice(valid_window, valid_size, replace=False)

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)

emb_net = tl.layers.Word2vecEmbeddingInputlayer(
    inputs=train_inputs,
    train_labels=train_labels,
    vocabulary_size=vocabulary_size,
    embedding_size=embedding_size,
    num_sampled=num_sampled,    
    name='word2vec_layer')

cost = emb_net.nce_cost
train_params = emb_net.all_params
# train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost, var_list=train_params)
train_op = tf.train.AdagradOptimizer(learning_rate).minimize(cost, var_list=train_params)

normalized_embeddings = emb_net.normalized_embeddings
valid_embed = tf.nn.embedding_lookup(normalized_embeddings, valid_dataset)
similarity = tf.matmul(valid_embed, normalized_embeddings, transpose_b=True)

print('~~~~~~~~~~~~~~~~~Start training~~~~~~~~~~~~~~~~')
sess = tf.InteractiveSession()
tl.layers.initialize_global_variables(sess)
if resume:
    print("Load existing model" + "!" * 10)
    # Load from ckpt or npz file
    # saver = tf.train.Saver()
    # saver.restore(sess, model_file_name+'.ckpt')
    tl.files.load_and_assign_npz_dict(name=model_file_name + '.npz', sess=sess)

emb_net.print_params()
emb_net.print_layers()

# save vocabulary to txt
tl.nlp.save_vocab(count, name='vocab_text8.txt')

average_loss = 0
step = 0
print_freq=100000

while step < num_steps:
    start_time = time.time()
    batch_inputs, batch_labels, data_index = tl.nlp.generate_skip_gram_batch(data=data, \
        batch_size=batch_size, num_skips=num_skips, skip_window=skip_window, data_index=data_index)
    feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}
    _, loss_val = sess.run([train_op, cost], feed_dict=feed_dict)
    average_loss += loss_val

    if step % print_freq == 0:
        if step > 0:
            average_loss /= print_freq
        print("~~~~~~~~Average loss at step %d/%d. loss: %f took: %fs~~~~~~~~~" % \
            (step, num_steps, average_loss, time.time() - start_time))
        average_loss = 0

    if step % (print_freq * 2) == 0:
        sim = similarity.eval(session=sess)
        for i in xrange(valid_size):
            valid_word = reverse_dictionary[valid_examples[i]]
            top_k = 8  
            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)

    if (step % (print_freq * 5) == 0) and (step != 0):
        print("******Save model, data and dictionaries***" + "!" * 10)
        # Save to ckpt or npz file
        # saver = tf.train.Saver()
        # save_path = saver.save(sess, model_file_name+'.ckpt')
        tl.files.save_npz_dict(emb_net.all_params, name=model_file_name + '.npz', sess=sess)
        tl.files.save_any_to_npy(
            save_dict={
                'data': data,
                'count': count,
                'dictionary': dictionary,
                'reverse_dictionary': reverse_dictionary
            }, name=model_file_name + '.npy'
        )
    step += 1

print('~~~~~~~~~~~~~~~Visualize t-SNE~~~~~~~~~~~~~~~~~~~~~~')

final_embeddings = sess.run(normalized_embeddings)  
tl.visualize.tsne_embedding(final_embeddings, reverse_dictionary, plot_only=500, \
    second=10, saveable=True, name='word2vec_basic',fig_idx=1988)

我是缩水训练了n_epoch = 12轮，嫌输出太多，就改成print_freq=100000，输出如下：

 

[TL] Load or Download matt_mahoney_text8 Dataset> data\mm_test8
17005207
['anarchism', 'originated', 'as', 'a', 'term', 'of', 'abuse', 'first', 'used', 'against']
132853 Steps in a Epoch, total Epochs 12
   learning_rate: 1.000000
   batch_size: 128
[TL] Real vocabulary size    253854
[TL] Limited vocabulary size 50000
~~~~~~~~~~~~~Most 5 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']
~~~~~~~~~~~~~只看周围两个单词，左右各一个~~~~~~~~~~~~
3081 originated -> 12 as
3081 originated -> 5234 anarchism
12 as -> 6 a
12 as -> 3081 originated
6 a -> 195 term
6 a -> 12 as
195 term -> 2 of
195 term -> 6 a
~~~~~~~~~~~~~只看周围四个单词，左右各两个~~~~~~~~~~~~
12 as -> 5234 anarchism
12 as -> 6 a
12 as -> 195 term
12 as -> 3081 originated
6 a -> 3081 originated
6 a -> 12 as
6 a -> 2 of
6 a -> 195 term
~~~~~~~~~~~~~~Build Skip-Gram model~~~~~~~~~~~~~
[TL] Word2vecEmbeddingInputlayer word2vec_layer: (50000, 128)
~~~~~~~~~~~~~~~~~Start training~~~~~~~~~~~~~~~~
[TL]   param   0: word2vec_layer/embeddings:0 (50000, 128)       float32_ref (mean: -0.00022887902741786093, median: -0.0002027750015258789, std: 0.5774627327919006)   
[TL]   param   1: word2vec_layer/nce_weights:0 (50000, 128)       float32_ref (mean: -9.809378752834164e-06, median: -1.1922868907277007e-05, std: 0.026393720880150795)   
[TL]   param   2: word2vec_layer/nce_biases:0 (50000,)           float32_ref (mean: 0.0               , median: 0.0               , std: 0.0               )   
[TL]   num of params: 12850000
[TL]   layer   0: word2vec_layer/embedding_lookup:0 (128, 128)         float32
[TL] 50000 vocab saved to vocab_text8.txt in C:\bbbb\学习\python教材\jfj\一起玩转Tensorlayer
~~~~~~~~Average loss at step 0/1594238. loss: 285.750641 took: 0.015627s~~~~~~~~~
------Nearest to over: antonia, note, jennifer, digested, rates, louise, mutinied, dalglish,
------Nearest to his: took, allman, planckian, motorways, adolph, bader, contraindicated, overworked,
------Nearest to when: sunni, eugen, intruder, jeanette, laconia, recorded, comte, tautology,
------Nearest to its: templar, trouser, gough, gini, menstrual, sizable, bawerk, sag,
------Nearest to eight: improvement, jamiroquai, duodecimal, amalthea, reflections, assay, duarte, tov,
------Nearest to use: dagmar, appeared, frame, sv, starboard, bypass, phenomenalism, amygdalin,
------Nearest to nine: recognise, wilmington, wankel, biathletes, cdot, mem, infantile, panel,
------Nearest to are: yoga, brandy, govern, quadrangle, penitence, add, dieter, ostia,
------Nearest to three: transfiguration, disbandment, contains, mini, campus, tezuka, dilute, rob,
------Nearest to some: false, axe, kalat, continuation, tortuous, fixture, rebuttal, berthold,
------Nearest to united: equivalently, dili, juvenile, prototype, domenico, today, sardinian, bush,
------Nearest to first: mets, chieftain, keiretsu, wonder, schwarzschild, federer, responsibility, cryptanalysis,
------Nearest to up: ditko, weicker, subspace, maia, upright, acrobatics, amperes, alvaro,
------Nearest to time: menthol, antares, dale, countrymen, foyle, dao, omniscient, nobles,
------Nearest to th: pilgrim, higham, msg, keralite, configured, mcdonalds, generalizing, katakana,
------Nearest to who: generality, balfour, rede, farad, recessive, suspicions, copyleft, diluted,
~~~~~~~~Average loss at step 100000/1594238. loss: 9.816838 took: 0.000000s~~~~~~~~~
~~~~~~~~Average loss at step 200000/1594238. loss: 4.424414 took: 0.015627s~~~~~~~~~
------Nearest to over: off, through, about, marienburg, within, around, behind, monies,
------Nearest to his: her, their, its, our, your, my, whose, the,
------Nearest to when: although, if, while, though, where, before, however, after,
------Nearest to its: their, his, the, her, our, whose, your, any,
------Nearest to eight: six, seven, nine, four, five, three, zero, two,
------Nearest to use: support, cause, need, allow, using, representation, practice, share,
------Nearest to nine: eight, seven, six, four, five, three, zero, two,
------Nearest to are: were, have, is, contain, include, tend, require, exist,
------Nearest to three: four, five, six, seven, two, eight, nine, zero,
------Nearest to some: many, several, any, certain, these, both, all, various,
------Nearest to united: european, confederate, senate, southern, university, northern, including, national,
------Nearest to first: last, second, next, third, fourth, final, same, original,
------Nearest to up: off, out, down, them, him, forward, together, back,
------Nearest to time: thing, nosferatu, period, shiki, day, credo, macphail, stars,
------Nearest to th: bc, fourth, six, one, rd, fifth, zero, four,
------Nearest to who: never, actually, she, often, still, frequently, also, which,
~~~~~~~~Average loss at step 300000/1594238. loss: 4.273012 took: 0.000000s~~~~~~~~~
~~~~~~~~Average loss at step 400000/1594238. loss: 4.185988 took: 0.000000s~~~~~~~~~
------Nearest to over: around, within, about, through, across, approximately, off, killing,
------Nearest to his: her, their, my, its, your, our, the, whose,
------Nearest to when: if, before, while, after, although, though, where, unless,
------Nearest to its: their, his, the, her, our, whose, my, some,
------Nearest to eight: nine, six, seven, five, four, three, zero, two,
------Nearest to use: practice, need, share, usage, applications, reach, support, allow,
------Nearest to nine: eight, seven, six, five, four, zero, three, two,
------Nearest to are: were, have, is, contain, include, tend, remain, require,
------Nearest to three: four, five, six, seven, two, eight, nine, one,
------Nearest to some: many, several, any, numerous, certain, these, various, all,
------Nearest to united: confederate, southern, communist, grand, yugoslav, senate, european, baltic,
------Nearest to first: second, last, next, third, oldest, fourth, previous, best,
------Nearest to up: off, out, down, him, back, them, together, forth,
------Nearest to time: day, nosferatu, observed, length, measures, thing, times, writing,
------Nearest to th: rd, fourth, sixth, fifth, nine, nd, seventh, zero,
------Nearest to who: he, she, never, actually, still, they, often, which,
~~~~~~~~Average loss at step 500000/1594238. loss: 4.121214 took: 0.000000s~~~~~~~~~
******Save model, data and dictionaries***!!!!!!!!!!
[TL] [*] Model saved in npz_dict model_word2vec_50k_128.npz
~~~~~~~~Average loss at step 600000/1594238. loss: 4.112394 took: 0.015627s~~~~~~~~~
------Nearest to over: around, across, through, about, within, throughout, approximately, off,
------Nearest to his: her, their, my, your, its, our, the, whose,
------Nearest to when: if, before, while, although, unless, after, though, where,
------Nearest to its: their, his, the, her, our, whose, your, any,
------Nearest to eight: seven, six, nine, five, four, three, zero, two,
------Nearest to use: need, practice, allow, usage, sense, support, applications, share,
------Nearest to nine: eight, seven, six, five, four, zero, three, births,
------Nearest to are: were, have, remain, is, contain, tend, although, represent,
------Nearest to three: five, four, two, six, seven, eight, nine, one,
------Nearest to some: many, several, certain, any, numerous, these, various, this,
------Nearest to united: confederate, communist, yugoslav, baltic, arab, frankish, usa, battle,
------Nearest to first: second, last, next, third, oldest, previous, earliest, fourth,
------Nearest to up: off, out, down, forth, together, them, him, back,
------Nearest to time: xli, renminbi, week, length, dreidel, eras, hours, opportunity,
------Nearest to th: rd, nd, sixth, fifth, fourth, seventh, nine, twentieth,
------Nearest to who: actually, often, he, never, she, still, reportedly, nor,
~~~~~~~~Average loss at step 700000/1594238. loss: 4.070077 took: 0.015626s~~~~~~~~~
~~~~~~~~Average loss at step 800000/1594238. loss: 4.040599 took: 0.015627s~~~~~~~~~
------Nearest to over: across, about, approximately, around, off, beyond, past, throughout,
------Nearest to his: her, their, my, your, its, our, the, whose,
------Nearest to when: if, before, where, while, unless, after, though, although,
------Nearest to its: their, his, the, her, whose, our, some, inferential,
------Nearest to eight: six, seven, nine, four, five, three, zero, two,
------Nearest to use: need, practice, using, usage, allow, sense, cause, offer,
------Nearest to nine: eight, six, seven, five, four, zero, three, two,
------Nearest to are: were, have, is, contain, remain, include, tend, although,
------Nearest to three: four, five, six, seven, two, eight, nine, zero,
------Nearest to some: many, several, certain, numerous, these, any, various, this,
------Nearest to united: confederate, communist, yugoslav, baltic, u, university, neighboring, caribbean,
------Nearest to first: second, last, third, next, fourth, oldest, previous, earliest,
------Nearest to up: off, down, him, them, out, forth, back, forward,
------Nearest to time: period, year, nosferatu, successful, observed, segment, saying, sense,
------Nearest to th: nd, rd, fifth, seventh, fourth, sixth, ninth, st,
------Nearest to who: actually, which, reportedly, never, he, often, she, whom,
~~~~~~~~Average loss at step 900000/1594238. loss: 4.047786 took: 0.015627s~~~~~~~~~
~~~~~~~~Average loss at step 1000000/1594238. loss: 4.000875 took: 0.000000s~~~~~~~~~
------Nearest to over: about, around, approximately, across, beyond, off, throughout, through,
------Nearest to his: her, their, my, your, its, the, our, this,
------Nearest to when: if, although, where, though, while, before, after, unless,
------Nearest to its: their, his, the, whose, her, our, carom, any,
------Nearest to eight: seven, nine, four, six, five, three, zero, two,
------Nearest to use: need, usage, applications, sense, allow, practice, mention, share,
------Nearest to nine: eight, seven, six, five, four, zero, three, two,
------Nearest to are: were, have, contain, is, include, tend, require, these,
------Nearest to three: four, five, two, six, seven, eight, nine, zero,
------Nearest to some: many, several, any, certain, numerous, these, various, most,
------Nearest to united: confederate, communist, yugoslav, u, frankish, usa, republican, holden,
------Nearest to first: second, last, third, next, fourth, oldest, earliest, latest,
------Nearest to up: off, down, out, them, back, together, forth, him,
------Nearest to time: hours, period, lengths, observed, sense, day, length, saying,
------Nearest to th: nd, sixth, rd, ninth, fifth, seventh, fourth, st,
------Nearest to who: which, never, reportedly, actually, whom, he, she, initially,
******Save model, data and dictionaries***!!!!!!!!!!
[TL] [*] Model saved in npz_dict model_word2vec_50k_128.npz
~~~~~~~~Average loss at step 1100000/1594238. loss: 4.019906 took: 0.000000s~~~~~~~~~
~~~~~~~~Average loss at step 1200000/1594238. loss: 4.011443 took: 0.000000s~~~~~~~~~
------Nearest to over: about, across, around, approximately, throughout, through, off, within,
------Nearest to his: her, their, my, your, its, our, the, whose,
------Nearest to when: if, before, unless, where, while, although, after, though,
------Nearest to its: their, his, the, whose, her, our, any, some,
------Nearest to eight: seven, nine, six, five, four, three, zero, two,
------Nearest to use: usage, applications, using, practice, mention, need, sense, employ,
------Nearest to nine: seven, eight, six, five, four, three, zero, two,
------Nearest to are: were, have, is, contain, include, remain, represent, exist,
------Nearest to three: five, four, seven, two, six, eight, nine, one,
------Nearest to some: many, several, numerous, certain, any, all, these, various,
------Nearest to united: confederate, yugoslav, communist, frankish, neighboring, oas, u, holden,
------Nearest to first: last, second, next, third, oldest, earliest, fourth, latest,
------Nearest to up: off, down, out, together, them, forth, him, back,
------Nearest to time: day, segment, year, period, month, length, times, hours,
------Nearest to th: nd, rd, sixth, fifth, fourth, ninth, seventh, twentieth,
------Nearest to who: which, never, reportedly, he, often, actually, they, she,
~~~~~~~~Average loss at step 1300000/1594238. loss: 3.963830 took: 0.015627s~~~~~~~~~
~~~~~~~~Average loss at step 1400000/1594238. loss: 3.993019 took: 0.000000s~~~~~~~~~
------Nearest to over: across, around, about, approximately, through, throughout, within, off,
------Nearest to his: her, their, my, its, our, the, your, whose,
------Nearest to when: if, where, before, while, unless, after, although, though,
------Nearest to its: their, his, the, whose, her, our, any, some,
------Nearest to eight: seven, nine, six, five, four, three, zero, two,
------Nearest to use: practice, usage, sense, need, applications, mention, reach, share,
------Nearest to nine: eight, seven, six, five, four, zero, three, two,
------Nearest to are: were, have, is, contain, include, remain, comprise, tend,
------Nearest to three: four, five, six, two, seven, eight, nine, zero,
------Nearest to some: many, several, numerous, certain, any, various, these, most,
------Nearest to united: confederate, usa, u, communist, arab, frankish, holden, oas,
------Nearest to first: second, last, third, fourth, next, earliest, oldest, latest,
------Nearest to up: off, out, down, forth, together, them, him, back,
------Nearest to time: day, period, times, life, periods, eras, masses, length,
------Nearest to th: nd, rd, ninth, seventh, fourth, sixth, fifth, st,
------Nearest to who: reportedly, which, never, he, nevertheless, often, seldom, actually,
~~~~~~~~Average loss at step 1500000/1594238. loss: 3.958364 took: 0.000000s~~~~~~~~~
******Save model, data and dictionaries***!!!!!!!!!!
[TL] [*] Model saved in npz_dict model_word2vec_50k_128.npz
~~~~~~~~~~~~~~~Visualize t-SNE~~~~~~~~~~~~~~~~~~~~~~
[Finished in 3668.3s]

 从结果来看相当不错，特别是 数字 和 his 两类：

------Nearest to his: her, their, my, its, our, the, your, whose,
------Nearest to when: if, where, before, while, unless, after, although, though,
------Nearest to its: their, his, the, whose, her, our, any, some,
------Nearest to eight: seven, nine, six, five, four, three, zero, two,
------Nearest to use: practice, usage, sense, need, applications, mention, reach, share,
------Nearest to nine: eight, seven, six, five, four, zero, three, two,