-
Notifications
You must be signed in to change notification settings - Fork 0
/
Deep Learning.html
executable file
·487 lines (313 loc) · 15.7 KB
/
Deep Learning.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
<!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>
Deep Learning - 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 home-categories">
<div class="article">
<a class="clearlink" href="14883590547961.html">
<h1>原始模型优化笔记</h1>
<div class="a-content">
<div class="a-content-text">
<p>对于原始弹幕分类CNN模型进行优化。</p>
<h2 id="toc_0">修改 word2vec model 的 vector size</h2>
<ul>
<li>400:
Nice at epoch 38, validation acc 96.56%</li>
<li>200:
Nice at epoch 37, validation acc 95.22%</li>
<li>100:
Nice at epoch 34, validation acc 94.78%
单轮训练时间与50维相近，测试样例测试耗时 0.92secs</li>
<li>50:
Nice at epoch 40, validation acc 94.39%
单轮训练时间在7秒左右，测试样例(av 8365806)测试耗时 0.7secs</li>
</ul>
<h2 id="toc_1">尝试加入dropout</h2>
<p>在两个 conv 层之间和两个 fc 层之间各加入了一个 \(p=0.5\) 的 dropout</p>
<p>40 epoch 时只有 89.1 acc， 和预想的一样，会导致 达到最佳效果的 epoch 数上升。</p>
<p>用了 dropout 后一个很明显的变化是，原本训练过程中通常是train acc 高于 validation acc，现在通常是 validation acc 高于 train acc，训练后期才基本持平或反超</p>
<p>vector在 epoch 90 左右 达到了96.50%上下的 acc，最终在epoch 300 以上能达到 97.10% 左右的 acc</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-24%2022.15.27.png" alt="屏幕快照 2017-02-24 22.15.27"/></p>
</div>
</div>
</a>
<div class="read-more clearfix">
<div class="more-left left">
<a href="14883590547961.html">Read more</a>
<span class="date">2017/3/1 17:4 下午</span>
<span>posted in </span>
<span class="posted-in"><a href='Deep%20Learning.html'>Deep Learning</a></span>
</div>
<div class="more-right right">
<span class="comments">
</span>
</div>
</div>
</div><!-- article -->
<div class="article">
<a class="clearlink" href="14863637393852.html">
<h1>低素质弹幕分类器的CNN实现</h1>
<div class="a-content">
<div class="a-content-text">
<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>
</div>
</div>
</a>
<div class="read-more clearfix">
<div class="more-left left">
<a href="14863637393852.html">Read more</a>
<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>
</div>
<div class="more-right right">
<span class="comments">
</span>
</div>
</div>
</div><!-- article -->
<div class="article">
<a class="clearlink" href="14863637259966.html">
<h1>低素质弹幕分类器 CNN 训练笔记</h1>
<div class="a-content">
<div class="a-content-text">
<p>一开始使用这个结构，迭代10次。</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(5, 4, border_mode='valid'))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dense(32, activation='relu'))
model.add(Dense(2, activation='softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy']
)
</code></pre>
<p>完成训练后，乍一看准确率很高，结果 print 出来看一下，低素质弹幕完全没有被过滤，完全是将分类全部丢给 positive 达到的高准确率 (0.98) 的确是 meaningless classification<br/>
并且这个结果在loss里看得很清楚，loss一直是处于15+的</p>
</div>
</div>
</a>
<div class="read-more clearfix">
<div class="more-left left">
<a href="14863637259966.html">Read more</a>
<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>
</div>
<div class="more-right right">
<span class="comments">
</span>
</div>
</div>
</div><!-- article -->
<div class="row">
<div class="large-6 columns">
<p class="text-left" style="padding-top:25px;">
</p>
</div>
<div class="large-6 columns">
<p class="text-right" style="padding-top:25px;">
</p>
</div>
</div>
</div>
</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:yinz995-1@yahoo.com" 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>