-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathchapter43.html
220 lines (180 loc) · 21.6 KB
/
chapter43.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
<!DOCTYPE HTML>
<html lang="en" class="sidebar-visible no-js light">
<head>
<!-- Book generated using mdBook -->
<meta charset="UTF-8">
<title>人工数据合成 - Machine Learning Yearning</title>
<!-- Custom HTML head -->
<meta content="text/html; charset=utf-8" http-equiv="Content-Type">
<meta name="description" content="">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="theme-color" content="#ffffff" />
<link rel="icon" href="favicon.svg">
<link rel="shortcut icon" href="favicon.png">
<link rel="stylesheet" href="css/variables.css">
<link rel="stylesheet" href="css/general.css">
<link rel="stylesheet" href="css/chrome.css">
<link rel="stylesheet" href="css/print.css" media="print">
<!-- Fonts -->
<link rel="stylesheet" href="FontAwesome/css/font-awesome.css">
<link rel="stylesheet" href="fonts/fonts.css">
<!-- Highlight.js Stylesheets -->
<link rel="stylesheet" href="highlight.css">
<link rel="stylesheet" href="tomorrow-night.css">
<link rel="stylesheet" href="ayu-highlight.css">
<!-- Custom theme stylesheets -->
</head>
<body>
<!-- Provide site root to javascript -->
<script type="text/javascript">
var path_to_root = "";
var default_theme = window.matchMedia("(prefers-color-scheme: dark)").matches ? "navy" : "light";
</script>
<!-- Work around some values being stored in localStorage wrapped in quotes -->
<script type="text/javascript">
try {
var theme = localStorage.getItem('mdbook-theme');
var sidebar = localStorage.getItem('mdbook-sidebar');
if (theme.startsWith('"') && theme.endsWith('"')) {
localStorage.setItem('mdbook-theme', theme.slice(1, theme.length - 1));
}
if (sidebar.startsWith('"') && sidebar.endsWith('"')) {
localStorage.setItem('mdbook-sidebar', sidebar.slice(1, sidebar.length - 1));
}
} catch (e) { }
</script>
<!-- Set the theme before any content is loaded, prevents flash -->
<script type="text/javascript">
var theme;
try { theme = localStorage.getItem('mdbook-theme'); } catch(e) { }
if (theme === null || theme === undefined) { theme = default_theme; }
var html = document.querySelector('html');
html.classList.remove('no-js')
html.classList.remove('light')
html.classList.add(theme);
html.classList.add('js');
</script>
<!-- Hide / unhide sidebar before it is displayed -->
<script type="text/javascript">
var html = document.querySelector('html');
var sidebar = 'hidden';
if (document.body.clientWidth >= 1080) {
try { sidebar = localStorage.getItem('mdbook-sidebar'); } catch(e) { }
sidebar = sidebar || 'visible';
}
html.classList.remove('sidebar-visible');
html.classList.add("sidebar-" + sidebar);
</script>
<nav id="sidebar" class="sidebar" aria-label="Table of contents">
<div class="sidebar-scrollbox">
<ol class="chapter"><li class="chapter-item expanded affix "><a href="index.html">引言</a></li><li class="chapter-item expanded "><a href="chapter1.html"><strong aria-hidden="true">1.</strong> 机器学习策略的原因</a></li><li class="chapter-item expanded "><a href="chapter2.html"><strong aria-hidden="true">2.</strong> 如何使用本书来帮助您的团队</a></li><li class="chapter-item expanded "><a href="chapter3.html"><strong aria-hidden="true">3.</strong> 预备知识和注释</a></li><li class="chapter-item expanded "><a href="chapter4.html"><strong aria-hidden="true">4.</strong> 规模推动机器学习进步</a></li><li class="chapter-item expanded "><a href="chapter5.html"><strong aria-hidden="true">5.</strong> 您的开发和测试集</a></li><li class="chapter-item expanded "><a href="chapter6.html"><strong aria-hidden="true">6.</strong> 你的开发集和测试集应该来自相同的分布</a></li><li class="chapter-item expanded "><a href="chapter7.html"><strong aria-hidden="true">7.</strong> 开发集/测试集需要多大</a></li><li class="chapter-item expanded "><a href="chapter8.html"><strong aria-hidden="true">8.</strong> 为您的团队建立单一数字的评估指标以进行优化</a></li><li class="chapter-item expanded "><a href="chapter9.html"><strong aria-hidden="true">9.</strong> 优化指标和满足指标</a></li><li class="chapter-item expanded "><a href="chapter10.html"><strong aria-hidden="true">10.</strong> 通过开发集和评估标准加速迭代</a></li><li class="chapter-item expanded "><a href="chapter11.html"><strong aria-hidden="true">11.</strong> 何时更改开发/测试集和评估指标</a></li><li class="chapter-item expanded "><a href="chapter12.html"><strong aria-hidden="true">12.</strong> 小结：建立开发集和测试集</a></li><li class="chapter-item expanded "><a href="chapter13.html"><strong aria-hidden="true">13.</strong> 快速构建您的第一个系统，然后迭代</a></li><li class="chapter-item expanded "><a href="chapter14.html"><strong aria-hidden="true">14.</strong> 误差分析：查看开发集样本以评估想法</a></li><li class="chapter-item expanded "><a href="chapter15.html"><strong aria-hidden="true">15.</strong> 在误差分析期间并行评估多个想法</a></li><li class="chapter-item expanded "><a href="chapter16.html"><strong aria-hidden="true">16.</strong> 清理错误标注的开发和测试集样本</a></li><li class="chapter-item expanded "><a href="chapter17.html"><strong aria-hidden="true">17.</strong> 如果你有一个大的开发集，将其分成两个子集，只着眼于其中的一个</a></li><li class="chapter-item expanded "><a href="chapter18.html"><strong aria-hidden="true">18.</strong> Eyeball 和 Blackbox 开发集应该多大？</a></li><li class="chapter-item expanded "><a href="chapter19.html"><strong aria-hidden="true">19.</strong> 小贴士：基本误差分析</a></li><li class="chapter-item expanded "><a href="chapter20.html"><strong aria-hidden="true">20.</strong> 偏差和方差：误差的两大来源</a></li><li class="chapter-item expanded "><a href="chapter21.html"><strong aria-hidden="true">21.</strong> 偏差和方差的例子</a></li><li class="chapter-item expanded "><a href="chapter22.html"><strong aria-hidden="true">22.</strong> 比较最优错误率</a></li><li class="chapter-item expanded "><a href="chapter23.html"><strong aria-hidden="true">23.</strong> 处理偏差和方差</a></li><li class="chapter-item expanded "><a href="chapter24.html"><strong aria-hidden="true">24.</strong> 偏差和方差间的权衡</a></li><li class="chapter-item expanded "><a href="chapter25.html"><strong aria-hidden="true">25.</strong> 减少可避免偏差的方法</a></li><li class="chapter-item expanded "><a href="chapter26.html"><strong aria-hidden="true">26.</strong> 训练集上的误差分析</a></li><li class="chapter-item expanded "><a href="chapter27.html"><strong aria-hidden="true">27.</strong> 减少方差的方法</a></li><li class="chapter-item expanded "><a href="chapter28.html"><strong aria-hidden="true">28.</strong> 诊断偏差和方差：学习曲线</a></li><li class="chapter-item expanded "><a href="chapter29.html"><strong aria-hidden="true">29.</strong> 绘制训练误差曲线</a></li><li class="chapter-item expanded "><a href="chapter30.html"><strong aria-hidden="true">30.</strong> 解读学习曲线：高偏差</a></li><li class="chapter-item expanded "><a href="chapter31.html"><strong aria-hidden="true">31.</strong> 解释学习曲线：其他情况</a></li><li class="chapter-item expanded "><a href="chapter32.html"><strong aria-hidden="true">32.</strong> 绘制学习曲线</a></li><li class="chapter-item expanded "><a href="chapter33.html"><strong aria-hidden="true">33.</strong> 为何我们要与人类水平的表现作对比</a></li><li class="chapter-item expanded "><a href="chapter34.html"><strong aria-hidden="true">34.</strong> 如何定义人类水平的表现</a></li><li class="chapter-item expanded "><a href="chapter35.html"><strong aria-hidden="true">35.</strong> 超越人类水平表现</a></li><li class="chapter-item expanded "><a href="chapter36.html"><strong aria-hidden="true">36.</strong> 何时应该在不同的分布下训练和测试</a></li><li class="chapter-item expanded "><a href="chapter37.html"><strong aria-hidden="true">37.</strong> 如何决定是否使用所有数据</a></li><li class="chapter-item expanded "><a href="chapter38.html"><strong aria-hidden="true">38.</strong> 如何决定是否包含不一致的数据</a></li><li class="chapter-item expanded "><a href="chapter39.html"><strong aria-hidden="true">39.</strong> 加权数据</a></li><li class="chapter-item expanded "><a href="chapter40.html"><strong aria-hidden="true">40.</strong> 从训练集到开发集的泛化</a></li><li class="chapter-item expanded "><a href="chapter41.html"><strong aria-hidden="true">41.</strong> 识别偏差、方差和数据不匹配误差</a></li><li class="chapter-item expanded "><a href="chapter42.html"><strong aria-hidden="true">42.</strong> 处理数据不匹配</a></li><li class="chapter-item expanded "><a href="chapter43.html" class="active"><strong aria-hidden="true">43.</strong> 人工数据合成</a></li><li class="chapter-item expanded "><a href="chapter44.html"><strong aria-hidden="true">44.</strong> 优化验证测试</a></li><li class="chapter-item expanded "><a href="chapter45.html"><strong aria-hidden="true">45.</strong> 优化验证集的一般形式</a></li><li class="chapter-item expanded "><a href="chapter46.html"><strong aria-hidden="true">46.</strong> 强化学习样本</a></li><li class="chapter-item expanded "><a href="chapter47.html"><strong aria-hidden="true">47.</strong> 端到端学习的兴起</a></li><li class="chapter-item expanded "><a href="chapter48.html"><strong aria-hidden="true">48.</strong> 更多端到端学习示例</a></li><li class="chapter-item expanded "><a href="chapter49.html"><strong aria-hidden="true">49.</strong> 端到端学习的优点和缺点</a></li><li class="chapter-item expanded "><a href="chapter50.html"><strong aria-hidden="true">50.</strong> 选择流水线组件：数据可用性</a></li><li class="chapter-item expanded "><a href="chapter51.html"><strong aria-hidden="true">51.</strong> 选择流水线组件：任务简单</a></li><li class="chapter-item expanded "><a href="chapter52.html"><strong aria-hidden="true">52.</strong> 直接学习丰富的输出</a></li><li class="chapter-item expanded "><a href="chapter53.html"><strong aria-hidden="true">53.</strong> 组件错误分析</a></li><li class="chapter-item expanded "><a href="chapter54.html"><strong aria-hidden="true">54.</strong> 将错误归因于某个组件</a></li><li class="chapter-item expanded "><a href="chapter55.html"><strong aria-hidden="true">55.</strong> 错误归因的一般情况</a></li><li class="chapter-item expanded "><a href="chapter56.html"><strong aria-hidden="true">56.</strong> 组件错误分析和与人类水平的对比</a></li><li class="chapter-item expanded "><a href="chapter57.html"><strong aria-hidden="true">57.</strong> 发现有瑕疵的ML流水线</a></li><li class="chapter-item expanded "><a href="chapter58.html"><strong aria-hidden="true">58.</strong> 组建一个超级英雄团队——让你的队友阅读本书</a></li></ol>
</div>
<div id="sidebar-resize-handle" class="sidebar-resize-handle"></div>
</nav>
<div id="page-wrapper" class="page-wrapper">
<div class="page">
<div id="menu-bar-hover-placeholder"></div>
<div id="menu-bar" class="menu-bar sticky bordered">
<div class="left-buttons">
<button id="sidebar-toggle" class="icon-button" type="button" title="Toggle Table of Contents" aria-label="Toggle Table of Contents" aria-controls="sidebar">
<i class="fa fa-bars"></i>
</button>
<button id="theme-toggle" class="icon-button" type="button" title="Change theme" aria-label="Change theme" aria-haspopup="true" aria-expanded="false" aria-controls="theme-list">
<i class="fa fa-paint-brush"></i>
</button>
<ul id="theme-list" class="theme-popup" aria-label="Themes" role="menu">
<li role="none"><button role="menuitem" class="theme" id="light">Light (default)</button></li>
<li role="none"><button role="menuitem" class="theme" id="rust">Rust</button></li>
<li role="none"><button role="menuitem" class="theme" id="coal">Coal</button></li>
<li role="none"><button role="menuitem" class="theme" id="navy">Navy</button></li>
<li role="none"><button role="menuitem" class="theme" id="ayu">Ayu</button></li>
</ul>
<button id="search-toggle" class="icon-button" type="button" title="Search. (Shortkey: s)" aria-label="Toggle Searchbar" aria-expanded="false" aria-keyshortcuts="S" aria-controls="searchbar">
<i class="fa fa-search"></i>
</button>
</div>
<h1 class="menu-title">Machine Learning Yearning</h1>
<div class="right-buttons">
<a href="print.html" title="Print this book" aria-label="Print this book">
<i id="print-button" class="fa fa-print"></i>
</a>
</div>
</div>
<div id="search-wrapper" class="hidden">
<form id="searchbar-outer" class="searchbar-outer">
<input type="search" id="searchbar" name="searchbar" placeholder="Search this book ..." aria-controls="searchresults-outer" aria-describedby="searchresults-header">
</form>
<div id="searchresults-outer" class="searchresults-outer hidden">
<div id="searchresults-header" class="searchresults-header"></div>
<ul id="searchresults">
</ul>
</div>
</div>
<!-- Apply ARIA attributes after the sidebar and the sidebar toggle button are added to the DOM -->
<script type="text/javascript">
document.getElementById('sidebar-toggle').setAttribute('aria-expanded', sidebar === 'visible');
document.getElementById('sidebar').setAttribute('aria-hidden', sidebar !== 'visible');
Array.from(document.querySelectorAll('#sidebar a')).forEach(function(link) {
link.setAttribute('tabIndex', sidebar === 'visible' ? 0 : -1);
});
</script>
<div id="content" class="content">
<main>
<h2 id="chapter-43artificial-data-synthesis"><a class="header" href="#chapter-43artificial-data-synthesis">Chapter 43、Artificial data synthesis</a></h2>
<p><strong>人工数据合成</strong></p>
<p>你的语音系统需要更多听起来像是从车内获取的数据。相比在驾车时手机许多数据，可能有一个更简单的方法去获取数据：通过人工合成它。</p>
<p>假设你获得了大量的车/道路噪音音频剪辑。你可以在多个网站上下载这些数据。假设你也有一个大的训练集，数据是人们在安静的房间里说话。如果你将一个人说话的音频剪辑和车/道路噪声音频剪辑“加”在一起，你将获得听起来像是人在吵闹的车内说话的音频剪辑。使用该过程，你可以“合成”大量听起来像是在车内收集的数据。</p>
<p>更一般的说，在几种情况下，人工数据合成可以让你创造出一个合理匹配开发集的庞大数据集。让我们用猫图检测器作为第二个例子。你注意到开发集图片有更多的运动模糊，因为它们倾向于来自于手机用户，用户在拍照时会轻微移动手机。你可以从互联网图片训练集中得到不模糊的图片，并为它们添加伪造运动模糊，从而使它们和开发集更类似。</p>
<p>记住人工数据合成有其挑战性：有时候创建对人来说看起来很逼真的合成数据比创建对电脑来说看起来很逼真的数据要容易。例如，假设你有1000小时的语音训练数据，但只有一小时的车噪声。如果你对原始1000小时训练数据的不同部分重复使用同样的这一小时的汽车噪声，你最终得到的合成数据集有着相同的汽车噪声，不断的重复。虽然听这种音频的人可能不能区分（对我们大多数人来说所有的汽车噪声都一样），但学习算法可能会“过拟合”这一个小时的汽车噪声。因此，对于汽车噪声听起来不同的新音频剪辑，算法可能泛化能力较差。</p>
<p>或者，假设你有1000个独特的汽车噪声小时，但它们全部来自仅10中不同的车。在这种情况下，算法那可能会“过拟合”这些车，如果在不同的汽车音频上测试，表现较差。不幸的是，这些问题很难被发现。</p>
<p><img src="img/myl-c43-0.jpg" alt="43" /></p>
<p>再看一个例子，假设你在构建一个识别车的计算机视觉系统。假设你和一个视频游公司合作，该公司拥有一些车的计算机图形模型。为了训练算法那，你使用这些模型生成汽车的合成图像。即使合成图片看起来很逼真，该方法（已被很多人独立提车）可能不会工作的很好。在整个视频游戏中可能有约20辆车设计。构建一个3D的汽车模型成本非常高；如果你在玩游戏，你可能不会注意到你反复看到同样的车，也许只是油漆色不同，也就是说，这些数据看起来对你非常逼真。但是和路上所有车的集合相比（因此你可能会在开发/测试集中看到），这20辆合成的车仅仅是世界汽车分布的一小部分。因此，如果你10W训练样本都来自这20辆车，系统将“过拟合”这20辆特定车的设计，并不能很好的泛化到包括其他车型设计的开发/测试集。</p>
<p>当合成数据的时，考虑一下你是否真的合成了一组具有代表性的样例。试图避免提供合成数据的属性，使得学习算法可以区分合成和非合成样本，例如如果所有的合成数据都来自20辆车设计中的一个，或所有的合成音频都来自车一个小时的噪声。这建议很难遵循。</p>
<p>当开展数据合成工作时，我的团队有时在生成数据上会花费数周，生成数据的详细信息足够贴近真实分布的，以便会有显著的影响。但如果你能正确获取细节，你会瞬间获取远比以前大的训练集。</p>
</main>
<nav class="nav-wrapper" aria-label="Page navigation">
<!-- Mobile navigation buttons -->
<a rel="prev" href="chapter42.html" class="mobile-nav-chapters previous" title="Previous chapter" aria-label="Previous chapter" aria-keyshortcuts="Left">
<i class="fa fa-angle-left"></i>
</a>
<a rel="next" href="chapter44.html" class="mobile-nav-chapters next" title="Next chapter" aria-label="Next chapter" aria-keyshortcuts="Right">
<i class="fa fa-angle-right"></i>
</a>
<div style="clear: both"></div>
</nav>
</div>
</div>
<nav class="nav-wide-wrapper" aria-label="Page navigation">
<a rel="prev" href="chapter42.html" class="nav-chapters previous" title="Previous chapter" aria-label="Previous chapter" aria-keyshortcuts="Left">
<i class="fa fa-angle-left"></i>
</a>
<a rel="next" href="chapter44.html" class="nav-chapters next" title="Next chapter" aria-label="Next chapter" aria-keyshortcuts="Right">
<i class="fa fa-angle-right"></i>
</a>
</nav>
</div>
<!-- Livereload script (if served using the cli tool) -->
<script type="text/javascript">
var socket = new WebSocket("ws://localhost:3000/__livereload");
socket.onmessage = function (event) {
if (event.data === "reload") {
socket.close();
location.reload();
}
};
window.onbeforeunload = function() {
socket.close();
}
</script>
<script type="text/javascript">
window.playground_copyable = true;
</script>
<script src="elasticlunr.min.js" type="text/javascript" charset="utf-8"></script>
<script src="mark.min.js" type="text/javascript" charset="utf-8"></script>
<script src="searcher.js" type="text/javascript" charset="utf-8"></script>
<script src="clipboard.min.js" type="text/javascript" charset="utf-8"></script>
<script src="highlight.js" type="text/javascript" charset="utf-8"></script>
<script src="book.js" type="text/javascript" charset="utf-8"></script>
<!-- Custom JS scripts -->
</body>
</html>