-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathchapter57.html
226 lines (186 loc) · 20.4 KB
/
chapter57.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
<!DOCTYPE HTML>
<html lang="en" class="sidebar-visible no-js light">
<head>
<!-- Book generated using mdBook -->
<meta charset="UTF-8">
<title>发现有瑕疵的ML流水线 - 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"><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" class="active"><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-57spotting-a-flawed-ml-pipeline"><a class="header" href="#chapter-57spotting-a-flawed-ml-pipeline">Chapter 57、Spotting a flawed ML pipeline</a></h2>
<p><strong>发现有瑕疵的ML流水线</strong></p>
<p>假如你的ML流水线的每个独立组件表现在人类水平或接近人类水平，但是整体流水线却远远低于人类水平，该怎么办？这通常意味着流水线有瑕疵，需要重新设计。错误分析也可以帮助你理解你是否需要重新设计管道。</p>
<p><img src="img/myl-c57-0.jpg" alt="图片1" /></p>
<p>在前一章节，我们提出了三个组件中的每个组件是否达到人类水平的问题。假设所有三个问题的答案是肯定的。也就是说：</p>
<ol>
<li>用于检测从摄像机图像中检测汽车的汽车检测组件（大致）是人类水平的表现。</li>
<li>用于从摄像机图像中检测行人的行人检测组件（大致）是人类水平的表现。</li>
<li>与仅根据前两个组件的输出（而不是访问摄像机图像）规划汽车路径的人相比，路径规划组件的表现和人类似。</li>
</ol>
<p>然而，自动驾驶汽车的整体表现远低于人类水平。即，给人类访问摄像机图片，可以为汽车规划出明显更好的路径。你能得出什么结论？</p>
<p>唯一可能的结论是ML流水线有瑕疵。在这种情况下，规划路径组件在给定输入的情况下已经做的尽可能好了，但输入没有包含足够的信息。你应该问你自己，除了两个早先流水线组件的输出外，还需要哪些其他信息来很好地规划汽车驾驶路径。换句话说，一个熟练的人类司机需要哪些其他信息？</p>
<p>例如，假设你意识到人类司机还需要知道道路标记的位置。这表明你应该重新设计流水线如下【1】：</p>
<p><img src="img/myl-c57-1.jpg" alt="图片2" /></p>
<p>最终，如果你不认为流水线整体将会达到人类水平的表现，即使每个独立组件具备人类水平表现（记住你正在和被给予组件相同输入的人类进行比较），那么流水线是有瑕疵的并应该被重新设计。</p>
<p>——————</p>
<p>【1】在上面的自动驾驶示例中，理论上可以通过将原始相机图像馈送到路径规划组件中来解决该问题。但这违反了51章中描述的“任务简单性”的设计原则，因为路径规划模块现在需要输入原始图片，并要解决非常复杂的任务。这就是为什么增加一个检测道路标记组件是更好的选择——它有助于得到重要的和之前路径规划模块丢失的关于道路标记信息，但你可避免使用任何特定模块过于复杂而无法构建/训练。</p>
</main>
<nav class="nav-wrapper" aria-label="Page navigation">
<!-- Mobile navigation buttons -->
<a rel="prev" href="chapter56.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="chapter58.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="chapter56.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="chapter58.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>