-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path14863637393852.html
executable file
·526 lines (339 loc) · 18.9 KB
/
14863637393852.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
<!doctype html>
<html class="no-js" lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>
低素质弹幕分类器的CNN实现 - 雪地
</title>
<link href="atom.xml" rel="alternate" title="雪地" type="application/atom+xml">
<link rel="stylesheet" href="asset/css/foundation.min.css" />
<link rel="stylesheet" href="asset/css/docs.css" />
<link rel="icon" href="asset/img/favicon.ico" />
<script src="asset/js/vendor/modernizr.js"></script>
<script src="asset/js/vendor/jquery.js"></script>
<script src="asset/highlightjs/highlight.pack.js"></script>
<link href="asset/highlightjs/styles/github.css" media="screen, projection" rel="stylesheet" type="text/css">
<script>hljs.initHighlightingOnLoad();</script>
<script type="text/javascript">
function before_search(){
var searchVal = 'site:yinzo.github.io ' + document.getElementById('search_input').value;
document.getElementById('search_q').value = searchVal;
return true;
}
</script>
</head>
<body class="antialiased hide-extras">
<div class="marketing off-canvas-wrap" data-offcanvas>
<div class="inner-wrap">
<nav class="top-bar docs-bar hide-for-small" data-topbar>
<section class="top-bar-section">
<div class="row">
<div style="position: relative;width:100%;"><div style="position: absolute; width:100%;">
<ul id="main-menu" class="left">
<li id="menu_item_index"><a href="index.html">Blog</a></li>
<li id="menu_item_archives"><a href="archives.html">Archives</a></li>
<li id="menu_item_about"><a href="http://yinz.xyz/">Home</a></li>
</ul>
<ul class="right" id="search-wrap">
<li>
<form target="_blank" onsubmit="return before_search();" action="http://google.com/search" method="get">
<input type="hidden" id="search_q" name="q" value="" />
<input tabindex="1" type="search" id="search_input" placeholder="Search"/>
</form>
</li>
</ul>
</div></div>
</div>
</section>
</nav>
<nav class="tab-bar show-for-small">
<a href="javascript:void(0)" class="left-off-canvas-toggle menu-icon">
<span> 雪地</span>
</a>
</nav>
<aside class="left-off-canvas-menu">
<ul class="off-canvas-list">
<li><a href="index.html">Blog</a></li>
<li><a href="archives.html">Archives</a></li>
<li><a href="http://yinz.xyz/">Home</a></li>
<li><label>Categories</label></li>
<li><a href="Security%20Info.html">Security Info</a></li>
<li><a href="Adversary%20Learning.html">Adversary Learning</a></li>
<li><a href="TCPIP.html">TCP/IP</a></li>
<li><a href="Pattern%20Recognition.html">Pattern Recognition</a></li>
<li><a href="Python.html">Python</a></li>
<li><a href="OS.html">OS</a></li>
<li><a href="Deep%20Learning.html">Deep Learning</a></li>
<li><a href="Machine%20Learning.html">Machine Learning</a></li>
</ul>
</aside>
<a class="exit-off-canvas" href="#"></a>
<section id="main-content" role="main" class="scroll-container">
<script type="text/javascript">
$(function(){
$('#menu_item_index').addClass('is_active');
});
</script>
<div class="row">
<div class="large-8 medium-8 columns">
<div class="markdown-body article-wrap">
<div class="article">
<h1>低素质弹幕分类器的CNN实现</h1>
<div class="read-more clearfix">
<span class="date">2017/2/6 14:48 下午</span>
<span>posted in </span>
<span class="posted-in"><a href='Deep%20Learning.html'>Deep Learning</a></span>
<span class="comments">
</span>
</div>
</div><!-- article -->
<div class="article-content">
<h2 id="toc_0">整体架构</h2>
<p>对于一条弹幕，首先进行分词，然后通过 word2vec 转换为词向量，再填充至固定长度，作为卷积神经网络的输入。</p>
<p>卷积神经网络的结构如下：</p>
<pre><code class="language-python">model = Sequential()
model.add(Convolution1D(100, 4, border_mode='valid', input_shape=(100, word_model.vector_size)))
model.add(Activation('relu'))
model.add(Convolution1D(100, 4, border_mode='valid', input_shape=(100, word_model.vector_size)))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(2, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy']
)
</code></pre>
<p>最终输出为2位的 categorical result，直接使用第一项，即骂人弹幕的概率作为输出。</p>
<p>然后通过代理，在弹幕服务器与播放器之间插入一层，实现弹幕的分类与屏蔽。最终实现了有效的骂人弹幕自动屏蔽，但是误伤的情况依然存在。</p>
<h2 id="toc_1">搭建过程</h2>
<span id="more"></span><!-- more -->
<p>使用游戏区的所有弹幕来训练 word2vec model。这里我是用的是 word2vec 的 Python 实现 gensim</p>
<p>训练脚本来自这篇文章 <a href="http://www.52nlp.cn/%E4%B8%AD%E8%8B%B1%E6%96%87%E7%BB%B4%E5%9F%BA%E7%99%BE%E7%A7%91%E8%AF%AD%E6%96%99%E4%B8%8A%E7%9A%84word2vec%E5%AE%9E%E9%AA%8C">中英文维基百科语料上的Word2Vec实验</a></p>
<pre><code class="language-python"># -*- coding: utf-8 -*-
"""
build a word2vec model by text file, each sentence for a line.
usage: [input file] [gensim model filename] [word2vec model filename]
example: output.txt life_damku.model life_damku.vector
"""
import logging
import os.path
import sys
import multiprocessing
from gensim.models import Word2Vec
from gensim.models.word2vec import LineSentence
if __name__ == '__main__':
program = os.path.basename(sys.argv[0])
logger = logging.getLogger(program)
logging.basicConfig(format='%(asctime)s: %(levelname)s: %(message)s')
logging.root.setLevel(level=logging.INFO)
logger.info("running %s" % ' '.join(sys.argv))
# check and process input arguments
if len(sys.argv) < 4:
print(globals()['__doc__'] % locals())
sys.exit(1)
inp, outp1, outp2 = sys.argv[1:4]
model = Word2Vec(LineSentence(inp), size=400, window=5, min_count=5,
workers=multiprocessing.cpu_count())
# trim unneeded model memory = use(much) less RAM
# model.init_sims(replace=True)
model.save(outp1)
model.save_word2vec_format(outp2, binary=False)
</code></pre>
<p>然后，我从所有弹幕中随机抽取了5000条，进行人工标注分类，其中有63条骂人弹幕。由于骂人弹幕太少，我又通过关键词搜索加人工筛选的方式，增加了4000条骂人弹幕。</p>
<p>以这约9000条弹幕作为训练样本，80%的弹幕作为 train set， 其余的20%作为 validation set</p>
<p>对训练样本进行预处理：</p>
<ol>
<li>分词</li>
<li>转换为词向量</li>
<li>填充至100位长，其中填充的位的词向量全部置零。</li>
</ol>
<p>开始构建卷积神经网络，我选用的框架是使用 TensorFlow 后端的 keras。最终经过调试，得到这样一个结构</p>
<p><a href="https://yinzo.github.io/14863637259966.html">具体的训练过程可以看这里</a></p>
<pre><code class="language-python">model = Sequential()
model.add(Convolution1D(100, 4, border_mode='valid', input_shape=(100, word_model.vector_size)))
model.add(Activation('relu'))
model.add(Convolution1D(100, 4, border_mode='valid', input_shape=(100, word_model.vector_size)))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(2, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy']
)
</code></pre>
<p>需要注意的是，<code>model.fit</code> 指定了 class weight</p>
<pre><code class="language-python">model.fit(X_train, Y_train, nb_epoch=1, batch_size=300, class_weight={
0: len(positive_sample)/len(negative_sample),
1: 1
})
</code></pre>
<p>单次训练迭代时间约25秒，我总共迭代了20次。其中每次迭代我都将 model 保存到一个列表中，并使用当前 model 对验证集进行测试，输出测试结果。</p>
<p>然后根据测试结果，选用了第3次迭代的 model ，尝试使用新的，不在样本中的视频弹幕进行人工检验识别效果。</p>
<p>选用视频 <a href="http://www.bilibili.com/video/av8365806/">av8365806</a> 里面存在大量对骂的弹幕，适合进行测试</p>
<p>这里我花了小半天时间，用 tornado 搭了个服务器，模拟 comment.bilibili.com 的所有请求，其中对于弹幕的请求，插入一层封装好的神经网络模型进行分类，再返回给用户请求，其余请求直接转发。搭建好后通过 Surge 的 URL Rewrite 将请求转到本地服务器上。</p>
<p><img src="media/14856271844324/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-06%2013.24.12.png" alt="屏幕快照 2017-02-06 13.24.12"/></p>
<p>然后此时访问弹幕都会先经过本地服务器进行处理了。查看 av8365806 的弹幕分类情况：</p>
<p>首先是比较欣慰的结果，大部分直接用脏字喷人的弹幕都被高确信地识别出来了<br/>
<img src="media/14856271844324/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-06%2013.26.37.png" alt="屏幕快照 2017-02-06 13.26.37"/></p>
<p>其次，存在一些短弹幕被误伤的情况，而且确信度还莫名的挺高<br/>
<img src="media/14856271844324/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-06%2013.26.51.png" alt="屏幕快照 2017-02-06 13.26.51"/></p>
<p>然后，依然存在一部分弹幕脏字不明显，以及一些反讽的语句没有识别出来<br/>
<img src="media/14856271844324/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-06%2013.27.39.png" alt="屏幕快照 2017-02-06 13.27.39"/></p>
<p>另外由于3000弹幕的识别时间在5秒左右，再加上网络延迟以及一些预处理后处理的时间，整体延迟约10秒，所以这个代理服务器可能还没能达到投入大规模使用的效率，比较可惜。</p>
<p>但是其识别效果，对于日常观感绝对是有所提升的。</p>
<p>脑海中对于屏蔽模式可以分为以下三挡：</p>
<ol>
<li>【仁慈模式】被分类器识别为骂人的弹幕，弹幕内容前填充100个空格，颜色变为白色，字号变小，统一为滚动弹幕。其中，填充空格能够使得这条弹幕将会以极快的速度飘过视野。</li>
<li>【常规模式】被分类器识别为骂人的弹幕直接删除，该弹幕的发送者的其他弹幕使用仁慈模式进行修正。</li>
<li>【灭杀模式】被分类器识别为骂人的弹幕，以及其发送者发送的其他弹幕，统一删除。</li>
</ol>
<p>原本对于仁慈模式是想进行高透明度处理的，但是由于B站弹幕不支持分开透明度所以就变成了现在的方案。不过由于可以直接修改弹幕类型，不知道高级弹幕是否支持透明度选项，以及高级弹幕在各个平台上的支持度如何。</p>
<h2 id="toc_2">结论</h2>
<p>这次构建卷积神经网络，是学习深度学习方向以来第一次完全自己设计网络结构，并独立完成所有的过程，包括样本收集、样本处理、搭建网络、训练，并且实现了模型的真实可用。作为深度学习的阶段性成果，也是给了我很大的鼓励。这个假期接下来的时间，准备继续学习 RNN 以及 LSTM 等模型，以及其他优化技巧。</p>
</div>
<div class="row">
<div class="large-6 columns">
<p class="text-left" style="padding:15px 0px;">
<a href="14883590547961.html"
title="Previous Post: 原始模型优化笔记">« 原始模型优化笔记</a>
</p>
</div>
<div class="large-6 columns">
<p class="text-right" style="padding:15px 0px;">
<a href="14863637259966.html"
title="Next Post: 低素质弹幕分类器 CNN 训练笔记">低素质弹幕分类器 CNN 训练笔记 »</a>
</p>
</div>
</div>
<div class="comments-wrap">
<div class="share-comments">
<div id="disqus_thread"></div>
<script>
/**
* RECOMMENDED CONFIGURATION VARIABLES: EDIT AND UNCOMMENT THE SECTION BELOW TO INSERT DYNAMIC VALUES FROM YOUR PLATFORM OR CMS.
* LEARN WHY DEFINING THESE VARIABLES IS IMPORTANT: https://disqus.com/admin/universalcode/#configuration-variables
*/
/*
var disqus_config = function () {
this.page.url = PAGE_URL; // Replace PAGE_URL with your page's canonical URL variable
this.page.identifier = PAGE_IDENTIFIER; // Replace PAGE_IDENTIFIER with your page's unique identifier variable
};
*/
(function() { // DON'T EDIT BELOW THIS LINE
var d = document, s = d.createElement('script');
s.src = '//yinzo.disqus.com/embed.js';
s.setAttribute('data-timestamp', +new Date());
(d.head || d.body).appendChild(s);
})();
</script>
<noscript>Please enable JavaScript to view the <a href="https://disqus.com/?ref_noscript" rel="nofollow">comments powered by Disqus.</a></noscript>
</div>
</div>
</div><!-- article-wrap -->
</div><!-- large 8 -->
<div class="large-4 medium-4 columns">
<div class="hide-for-small">
<div id="sidebar" class="sidebar">
<div id="site-info" class="site-info">
<div class="site-a-logo"><img src="asset/img/3.png" /></div>
<h1>雪地</h1>
<div class="site-des"></div>
<div class="social">
<a class="github" target="_blank" href="https://github.com/Yinzo" title="GitHub">GitHub</a>
<a class="email" href="mailto:[email protected]" title="Email">Email</a>
<a class="rss" href="atom.xml" title="RSS">RSS</a>
</div>
</div>
<div id="site-categories" class="side-item ">
<div class="side-header">
<h2>Categories</h2>
</div>
<div class="side-content">
<p class="cat-list">
<a href="Security%20Info.html"><strong>Security Info</strong></a>
<a href="Adversary%20Learning.html"><strong>Adversary Learning</strong></a>
<a href="TCPIP.html"><strong>TCP/IP</strong></a>
<a href="Pattern%20Recognition.html"><strong>Pattern Recognition</strong></a>
<a href="Python.html"><strong>Python</strong></a>
<a href="OS.html"><strong>OS</strong></a>
<a href="Deep%20Learning.html"><strong>Deep Learning</strong></a>
<a href="Machine%20Learning.html"><strong>Machine Learning</strong></a>
</p>
</div>
</div>
<div id="site-categories" class="side-item">
<div class="side-header">
<h2>Recent Posts</h2>
</div>
<div class="side-content">
<ul class="posts-list">
<li class="post">
<a href="14968173531750.html">CS229 学习笔记 Part3</a>
</li>
<li class="post">
<a href="14965964854250.html">CS229 学习笔记 Part2</a>
</li>
<li class="post">
<a href="14946020792948.html">CS229 学习笔记 Part 1</a>
</li>
<li class="post">
<a href="14883590547961.html">原始模型优化笔记</a>
</li>
<li class="post">
<a href="14863637393852.html">低素质弹幕分类器的CNN实现</a>
</li>
</ul>
</div>
</div>
<div id="site-link" class="side-item">
<div class="side-header">
<h2>Link</h2>
</div>
<div class="side-content">
<p class="link-list">
<a href="http://blog.winkidney.com/">阿毛</a>
</p>
</div>
</div>
</div><!-- sidebar -->
</div><!-- hide for small -->
</div><!-- large 4 -->
</div><!-- row -->
<div class="page-bottom clearfix">
<div class="row">
<p class="copyright">Copyright © 2016
Powered by <a target="_blank" href="http://www.mweb.im">MWeb</a>,
Theme used <a target="_blank" href="http://github.com">GitHub CSS</a>.
Modified by <a target="_blank" href="http://yinz.xyz">Yinzo</a>.</p>
</div>
</div>
</section>
</div>
</div>
<script src="asset/js/foundation.min.js"></script>
<script>
$(document).foundation();
function fixSidebarHeight(){
var w1 = $('.markdown-body').height();
var w2 = $('#sidebar').height();
if (w1 > w2) { $('#sidebar').height(w1); };
}
$(function(){
fixSidebarHeight();
})
$(window).load(function(){
fixSidebarHeight();
});
</script>
<script src="asset/chart/all-min.js"></script><script type="text/javascript">$(function(){ var mwebii=0; var mwebChartEleId = 'mweb-chart-ele-'; $('pre>code').each(function(){ mwebii++; var eleiid = mwebChartEleId+mwebii; if($(this).hasClass('language-sequence')){ var ele = $(this).addClass('nohighlight').parent(); $('<div id="'+eleiid+'"></div>').insertAfter(ele); ele.hide(); var diagram = Diagram.parse($(this).text()); diagram.drawSVG(eleiid,{theme: 'simple'}); }else if($(this).hasClass('language-flow')){ var ele = $(this).addClass('nohighlight').parent(); $('<div id="'+eleiid+'"></div>').insertAfter(ele); ele.hide(); var diagram = flowchart.parse($(this).text()); diagram.drawSVG(eleiid); } });});</script>
<script type="text/javascript" src="https://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script><script type="text/x-mathjax-config">MathJax.Hub.Config({TeX: { equationNumbers: { autoNumber: "AMS" } }});</script>
<script src="asset/js/instantclick.min.js" data-no-instant></script>
<script data-no-instant>InstantClick.on('change', function() {
MathJax.Hub.Queue(["Typeset",MathJax.Hub]);
});</script>
<script data-no-instant>InstantClick.init();</script>
</body>
</html>