-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path14883590547961.html
executable file
·485 lines (304 loc) · 18 KB
/
14883590547961.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
<!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>
原始模型优化笔记 - 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>原始模型优化笔记</h1>
<div class="read-more clearfix">
<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>
<span class="comments">
</span>
</div>
</div><!-- article -->
<div class="article-content">
<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>
<span id="more"></span><!-- more -->
<p>人工检查实际识别效果,仍有少量漏网。果然几个百分点的区别,人简单扫视还是很难看出区别的,而且还要排除安慰剂效应。</p>
<p>实验证明 dropout 确实有效防止了过拟合,并且提高了一定的分类准确度。</p>
<p>接着将几个 word2vec 长度的模型训练图都画出来对比:</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-03-01%2016.57.28.png" alt="屏幕快照 2017-03-01 16.57.28"/></p>
<p>可以从左下角的 tooltip 看到,从上至下分别是词向量长度为 400、200、100、50、25 的模型,在相同迭代次数下的准确度排序。由于选取的是上述模型都仍有收敛空间的迭代数,所以这个排名一定程度上可以代表模型训练所需的迭代次数排序。</p>
<p>由此可以得出结论,词向量维数越多,模型收敛所需的迭代次数越少,但是最终收敛的效果没有变化,这可能是目前训练样本较少的原因。</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-03-01%2017.01.36.png" alt="屏幕快照 2017-03-01 17.01.36"/></p>
<p>切换到训练时间为横轴的图来看,收敛速度实际上是100维最快,这倒是出乎意外,原本以为会是低维模型最快,结果发现,100维由于单次迭代提升更大,虽然迭代速度稍慢但是总体收敛最快。我还特地检查了100维的data graph,确定的确是在用100维的词向量来训练的。</p>
<h3 id="toc_2">关于卷积层是否应该加入 dropout 的问题</h3>
<p>这个问题我找到了 <a href="https://www.reddit.com/r/MachineLearning/comments/42nnpe/why_do_i_never_see_dropout_applied_in/">Reddit 的这个讨论串</a></p>
<h4 id="toc_3">里面提到了以下这些说法:</h4>
<ul>
<li>卷积层的参数数量没有全连接层那么多,所以不那么需要 regularizaion</li>
<li>卷积层 filter map 的梯度是对于整个样本进行平均化的<sup id="fnref1"><a href="#fn1" rel="footnote">1</a></sup>,这样会使得卷积核原本存在相关性的参数,在样本的不同位置使用了不同的 dropout mask,导致 dropout 无效。当然,你可以想办法使卷积核的 dropout mask 在同一层中固定,但是这又会导致 regularizaion 过强。</li>
<li>Srivastava/Hinton 在 dropout 的论文中也有提到:在卷积层加入 dropout 的效果等于没有 \((3.02\% \rightarrow 2.55\%)\),因为卷积层的参数太少了,不存在过拟合的问题,所以 dropout 几乎没有效果。但是 dropout 在较低的层仍是有用的,它的效果相当于产生一点噪声,使得后面层数较高的全连接层避免过拟合。</li>
<li><p>在卷积层使用 dropout 也不是绝对没有的,以下几篇论文中就有用到:</p>
<ul>
<li><a href="http://arxiv.org/pdf/1511.07289v3.pdf">http://arxiv.org/pdf/1511.07289v3.pdf</a></li>
<li><a href="http://torch.ch/blog/2015/07/30/cifar.html">http://torch.ch/blog/2015/07/30/cifar.html</a></li>
<li><a href="http://danielnouri.org/notes/2014/12/17/using-convolutional-neural-nets-to-detect-facial-keypoints-tutorial/">http://danielnouri.org/notes/2014/12/17/using-convolutional-neural-nets-to-detect-facial-keypoints-tutorial/</a></li>
</ul>
<p>但是他们都有卷积层的 dropout 『keep_prob 较大』的特点。</p></li>
</ul>
<h4 id="toc_4">以下是测试的结果:</h4>
<p>先测试了卷积层不使用 dropout 的效果<br/>
<img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-27%2013.11.43.png" alt="屏幕快照 2017-02-27 13.11.43"/></p>
<p>绿线是没有 卷积层dropout 的 acc, 紫线是卷积层 p=0.5 dropout 的 acc,两者都有一个 p=0.5 的全连接 dropout。由此可得卷积层的 dropout 还是有效果的,首先防止过拟合的程度要更高,其次对于准确度的提升也是有的 \((\uparrow2.3\%)\)</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-03-01%2015.45.56.png" alt="屏幕快照 2017-03-01 15.45.56"/></p>
<p>卷积层 p=0.6 和 p=0.4 dropout 的测试,相比 p=0.5 的模型,收敛速度、收敛精度都没有明显的区别。</p>
<h2 id="toc_5">尝试加入 max_pooling</h2>
<p>先在卷积层和全连接层之间加入一个大小为2的 <code>max_pooling1d</code><br/>
<img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-26%2016.54.23.png" alt="屏幕快照 2017-02-26 16.54.23"/></p>
<p>可以看到,这个 max_pool 对于最终收敛精度没有影响,在前期略微加快了收敛,但是中期减慢了收敛。</p>
<p>尝试移动该 pooling 层到两个卷积层之间。</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-28%2013.42.28.png" alt="屏幕快照 2017-02-28 13.42.28"/></p>
<p>上图为多次测试的结果。收敛速度和收敛准确率没有较为明显的提高或降低 \((96.623\% \rightarrow 96.694\%)\),只有训练速度有略微的提高。</p>
<p>值得一提的是,加入了 max-pool 的模型在高迭代次数的时候,标准差在逐渐增大。我认为这是因为 max-pool 一定程度上降低了训练样本的精度,相当于训练样本变少了,于是乎少量增加了整个模型在高迭代次数过拟合的风险。这里我选择使用84次迭代的模型,也就是图中横坐标约为 1k 的位置。相对来说过拟合的程度应该是非常小的。</p>
<p><img src="media/14879250453025/%E5%B1%8F%E5%B9%95%E5%BF%AB%E7%85%A7%202017-02-26%2018.49.29.png" alt="屏幕快照 2017-02-26 18.49.29"/></p>
<div class="footnotes">
<hr/>
<ol>
<li id="fn1">
<blockquote>
<p>the gradients are averaged over the spatial extent of the feature maps <a href="#fnref1" rev="footnote">↩</a></p>
</blockquote>
</li>
</ol>
</div>
</div>
<div class="row">
<div class="large-6 columns">
<p class="text-left" style="padding:15px 0px;">
<a href="14946020792948.html"
title="Previous Post: CS229 学习笔记 Part 1">« CS229 学习笔记 Part 1</a>
</p>
</div>
<div class="large-6 columns">
<p class="text-right" style="padding:15px 0px;">
<a href="14863637393852.html"
title="Next Post: 低素质弹幕分类器的CNN实现">低素质弹幕分类器的CNN实现 »</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>