-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path20210821102845-regularization_with_gan_output.html
484 lines (459 loc) · 21.9 KB
/
20210821102845-regularization_with_gan_output.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
<!DOCTYPE html>
<html lang="zh">
<head>
<!-- 2021-12-09 四 17:18 -->
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>正则化对 GAN 输出结果的影响</title>
<meta name="author" content="shirui" />
<meta name="generator" content="Org Mode" />
<style>
#content { max-width: 60em; margin: auto; }
.title { text-align: center;
margin-bottom: .2em; }
.subtitle { text-align: center;
font-size: medium;
font-weight: bold;
margin-top:0; }
.todo { font-family: monospace; color: red; }
.done { font-family: monospace; color: green; }
.priority { font-family: monospace; color: orange; }
.tag { background-color: #eee; font-family: monospace;
padding: 2px; font-size: 80%; font-weight: normal; }
.timestamp { color: #bebebe; }
.timestamp-kwd { color: #5f9ea0; }
.org-right { margin-left: auto; margin-right: 0px; text-align: right; }
.org-left { margin-left: 0px; margin-right: auto; text-align: left; }
.org-center { margin-left: auto; margin-right: auto; text-align: center; }
.underline { text-decoration: underline; }
#postamble p, #preamble p { font-size: 90%; margin: .2em; }
p.verse { margin-left: 3%; }
pre {
border: 1px solid #e6e6e6;
border-radius: 3px;
background-color: #f2f2f2;
padding: 8pt;
font-family: monospace;
overflow: auto;
margin: 1.2em;
}
pre.src {
position: relative;
overflow: auto;
}
pre.src:before {
display: none;
position: absolute;
top: -8px;
right: 12px;
padding: 3px;
color: #555;
background-color: #f2f2f299;
}
pre.src:hover:before { display: inline; margin-top: 14px;}
/* Languages per Org manual */
pre.src-asymptote:before { content: 'Asymptote'; }
pre.src-awk:before { content: 'Awk'; }
pre.src-authinfo::before { content: 'Authinfo'; }
pre.src-C:before { content: 'C'; }
/* pre.src-C++ doesn't work in CSS */
pre.src-clojure:before { content: 'Clojure'; }
pre.src-css:before { content: 'CSS'; }
pre.src-D:before { content: 'D'; }
pre.src-ditaa:before { content: 'ditaa'; }
pre.src-dot:before { content: 'Graphviz'; }
pre.src-calc:before { content: 'Emacs Calc'; }
pre.src-emacs-lisp:before { content: 'Emacs Lisp'; }
pre.src-fortran:before { content: 'Fortran'; }
pre.src-gnuplot:before { content: 'gnuplot'; }
pre.src-haskell:before { content: 'Haskell'; }
pre.src-hledger:before { content: 'hledger'; }
pre.src-java:before { content: 'Java'; }
pre.src-js:before { content: 'Javascript'; }
pre.src-latex:before { content: 'LaTeX'; }
pre.src-ledger:before { content: 'Ledger'; }
pre.src-lisp:before { content: 'Lisp'; }
pre.src-lilypond:before { content: 'Lilypond'; }
pre.src-lua:before { content: 'Lua'; }
pre.src-matlab:before { content: 'MATLAB'; }
pre.src-mscgen:before { content: 'Mscgen'; }
pre.src-ocaml:before { content: 'Objective Caml'; }
pre.src-octave:before { content: 'Octave'; }
pre.src-org:before { content: 'Org mode'; }
pre.src-oz:before { content: 'OZ'; }
pre.src-plantuml:before { content: 'Plantuml'; }
pre.src-processing:before { content: 'Processing.js'; }
pre.src-python:before { content: 'Python'; }
pre.src-R:before { content: 'R'; }
pre.src-ruby:before { content: 'Ruby'; }
pre.src-sass:before { content: 'Sass'; }
pre.src-scheme:before { content: 'Scheme'; }
pre.src-screen:before { content: 'Gnu Screen'; }
pre.src-sed:before { content: 'Sed'; }
pre.src-sh:before { content: 'shell'; }
pre.src-sql:before { content: 'SQL'; }
pre.src-sqlite:before { content: 'SQLite'; }
/* additional languages in org.el's org-babel-load-languages alist */
pre.src-forth:before { content: 'Forth'; }
pre.src-io:before { content: 'IO'; }
pre.src-J:before { content: 'J'; }
pre.src-makefile:before { content: 'Makefile'; }
pre.src-maxima:before { content: 'Maxima'; }
pre.src-perl:before { content: 'Perl'; }
pre.src-picolisp:before { content: 'Pico Lisp'; }
pre.src-scala:before { content: 'Scala'; }
pre.src-shell:before { content: 'Shell Script'; }
pre.src-ebnf2ps:before { content: 'ebfn2ps'; }
/* additional language identifiers per "defun org-babel-execute"
in ob-*.el */
pre.src-cpp:before { content: 'C++'; }
pre.src-abc:before { content: 'ABC'; }
pre.src-coq:before { content: 'Coq'; }
pre.src-groovy:before { content: 'Groovy'; }
/* additional language identifiers from org-babel-shell-names in
ob-shell.el: ob-shell is the only babel language using a lambda to put
the execution function name together. */
pre.src-bash:before { content: 'bash'; }
pre.src-csh:before { content: 'csh'; }
pre.src-ash:before { content: 'ash'; }
pre.src-dash:before { content: 'dash'; }
pre.src-ksh:before { content: 'ksh'; }
pre.src-mksh:before { content: 'mksh'; }
pre.src-posh:before { content: 'posh'; }
/* Additional Emacs modes also supported by the LaTeX listings package */
pre.src-ada:before { content: 'Ada'; }
pre.src-asm:before { content: 'Assembler'; }
pre.src-caml:before { content: 'Caml'; }
pre.src-delphi:before { content: 'Delphi'; }
pre.src-html:before { content: 'HTML'; }
pre.src-idl:before { content: 'IDL'; }
pre.src-mercury:before { content: 'Mercury'; }
pre.src-metapost:before { content: 'MetaPost'; }
pre.src-modula-2:before { content: 'Modula-2'; }
pre.src-pascal:before { content: 'Pascal'; }
pre.src-ps:before { content: 'PostScript'; }
pre.src-prolog:before { content: 'Prolog'; }
pre.src-simula:before { content: 'Simula'; }
pre.src-tcl:before { content: 'tcl'; }
pre.src-tex:before { content: 'TeX'; }
pre.src-plain-tex:before { content: 'Plain TeX'; }
pre.src-verilog:before { content: 'Verilog'; }
pre.src-vhdl:before { content: 'VHDL'; }
pre.src-xml:before { content: 'XML'; }
pre.src-nxml:before { content: 'XML'; }
/* add a generic configuration mode; LaTeX export needs an additional
(add-to-list 'org-latex-listings-langs '(conf " ")) in .emacs */
pre.src-conf:before { content: 'Configuration File'; }
table { border-collapse:collapse; }
caption.t-above { caption-side: top; }
caption.t-bottom { caption-side: bottom; }
td, th { vertical-align:top; }
th.org-right { text-align: center; }
th.org-left { text-align: center; }
th.org-center { text-align: center; }
td.org-right { text-align: right; }
td.org-left { text-align: left; }
td.org-center { text-align: center; }
dt { font-weight: bold; }
.footpara { display: inline; }
.footdef { margin-bottom: 1em; }
.figure { padding: 1em; }
.figure p { text-align: center; }
.equation-container {
display: table;
text-align: center;
width: 100%;
}
.equation {
vertical-align: middle;
}
.equation-label {
display: table-cell;
text-align: right;
vertical-align: middle;
}
.inlinetask {
padding: 10px;
border: 2px solid gray;
margin: 10px;
background: #ffffcc;
}
#org-div-home-and-up
{ text-align: right; font-size: 70%; white-space: nowrap; }
textarea { overflow-x: auto; }
.linenr { font-size: smaller }
.code-highlighted { background-color: #ffff00; }
.org-info-js_info-navigation { border-style: none; }
#org-info-js_console-label
{ font-size: 10px; font-weight: bold; white-space: nowrap; }
.org-info-js_search-highlight
{ background-color: #ffff00; color: #000000; font-weight: bold; }
.org-svg { width: 90%; }
</style>
<link rel="stylesheet" href="/static/style.css" type="text/css"/>
<link rel="stylesheet" href="/static/org.css" type="text/css"/>
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js"></script>
<meta http-equiv="content-type" content="application/xhtml+xml; charset=UTF-8" />
<meta name="viewport" content="initial-scale=1,width=device-width,minimum-scale=1">
</head>
<body>
<div id="preamble" class="status">
<div class="header">
<a href="https://lampze.github.io">lampze's Blog</a>
<span id="post-date" style="float: right">2021-08-21 六 00:00</span>
</div>
</div>
<div id="content" class="content">
<header>
<h1 class="title">正则化对 GAN 输出结果的影响</h1>
</header>
<div id="outline-container-orgcb0898b" class="outline-2">
<h2 id="orgcb0898b">前言</h2>
<div class="outline-text-2" id="text-orgcb0898b">
<p>
在我尝试使用正则化解决 <code>GAN</code> 模式崩溃的问题时，我发现了一个有意思的现象。我给模型添加了 <code>L2</code> 正则项后，生成的图片比以往要模糊许多，所以我决定测试一下，模糊是不是由 <code>L2</code> 导致的。<br>
</p>
</div>
</div>
<div id="outline-container-orgaabc24f" class="outline-2">
<h2 id="orgaabc24f">测试</h2>
<div class="outline-text-2" id="text-orgaabc24f">
<p>
本模型除了 <code>L1</code> 、 <code>L2</code> 、 <code>Dropout</code> 还使用了其他的优化方法， <code>LayerNorm</code> 、 <code>GELU</code> 、 <code>Adam</code> 。每种模型的训练次数相同，没有给图片加标签，让模式崩溃的效果更明显些。<br>
</p>
</div>
<div id="outline-container-orgf4725a3" class="outline-3">
<h3 id="orgf4725a3">基础代码</h3>
<div class="outline-text-3" id="text-orgf4725a3">
<p>
后面的测试都是基于这份代码之上修改，只放了关键的模型定义与训练代码。<br>
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="font-weight: bold;">class</span> <span style="font-weight: bold;">Discriminator</span>(nn.Module):
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">__init__</span>(<span style="font-weight: bold;">self</span>):
<span style="font-weight: bold;">super</span>().__init__()
<span style="font-weight: bold;">self</span>.<span style="color: #383a42;">model</span> = nn.Sequential(
nn.Flatten(),
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">28</span> * <span style="color: #8a3b3c; font-weight: bold;">28</span>, <span style="color: #8a3b3c; font-weight: bold;">200</span>),
nn.GELU(),
nn.LayerNorm(<span style="color: #8a3b3c; font-weight: bold;">200</span>),
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">200</span>, <span style="color: #8a3b3c; font-weight: bold;">1</span>),
nn.Sigmoid(),
)
<span style="font-weight: bold;">self</span>.<span style="color: #383a42;">loss_function</span> = nn.BCELoss()
<span style="font-weight: bold;">self</span>.<span style="color: #383a42;">optimiser</span> = torch.optim.Adam(
<span style="font-weight: bold;">self</span>.parameters(), lr=<span style="color: #8a3b3c; font-weight: bold;">0.0001</span>
)
<span style="font-weight: bold;">self</span>.progress = []
<span style="font-weight: bold;">pass</span>
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">forward</span>(<span style="font-weight: bold;">self</span>, inputs):
<span style="font-weight: bold;">return</span> <span style="font-weight: bold;">self</span>.model(inputs).reshape(<span style="color: #8a3b3c; font-weight: bold;">1</span>)
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">train</span>(<span style="font-weight: bold;">self</span>, inputs, targets):
outputs = <span style="font-weight: bold;">self</span>.forward(inputs)
loss = <span style="font-weight: bold;">self</span>.loss_function(outputs, targets)
<span style="font-weight: bold;">self</span>.optimiser.zero_grad()
loss.backward()
<span style="font-weight: bold;">self</span>.optimiser.step()
<span style="font-weight: bold;">pass</span>
<span style="font-weight: bold;">class</span> <span style="font-weight: bold;">Generator</span>(nn.Module):
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">__init__</span>(<span style="font-weight: bold;">self</span>):
<span style="font-weight: bold;">super</span>().__init__()
<span style="font-weight: bold;">self</span>.model = nn.Sequential(
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">200</span>, <span style="color: #8a3b3c; font-weight: bold;">400</span>),
nn.GELU(),
nn.LayerNorm(<span style="color: #8a3b3c; font-weight: bold;">400</span>),
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">400</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span> * <span style="color: #8a3b3c; font-weight: bold;">28</span>),
nn.Sigmoid(),
)
<span style="font-weight: bold;">self</span>.optimiser = torch.optim.Adam(
<span style="font-weight: bold;">self</span>.parameters(), lr=<span style="color: #8a3b3c; font-weight: bold;">0.0001</span>
)
<span style="font-weight: bold;">self</span>.progress = []
<span style="font-weight: bold;">pass</span>
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">forward</span>(<span style="font-weight: bold;">self</span>, inputs):
<span style="font-weight: bold;">return</span> <span style="font-weight: bold;">self</span>.model(inputs)
<span style="font-weight: bold;">def</span> <span style="font-weight: bold;">train</span>(<span style="font-weight: bold;">self</span>, D, inputs, targets):
g_output = <span style="font-weight: bold;">self</span>.forward(inputs)
d_output = D.forward(g_output.reshape(<span style="color: #8a3b3c; font-weight: bold;">1</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>))
loss = D.loss_function(d_output, targets)
<span style="font-weight: bold;">self</span>.optimiser.zero_grad()
loss.backward()
<span style="font-weight: bold;">self</span>.optimiser.step()
<span style="font-weight: bold;">pass</span>
D = Discriminator()
G = Generator()
<span style="font-weight: bold;">for</span> epoch <span style="font-weight: bold;">in</span> <span style="font-weight: bold;">range</span>(<span style="color: #8a3b3c; font-weight: bold;">4</span>):
<span style="font-weight: bold;">for</span> image_data_tensor, label <span style="font-weight: bold;">in</span> mnist_train:
D.train(image_data_tensor, torch.tensor([<span style="color: #8a3b3c; font-weight: bold;">1.0</span>]))
D.train(
G.forward(generate_random_seed(<span style="color: #8a3b3c; font-weight: bold;">200</span>)).detach().reshape(<span style="color: #8a3b3c; font-weight: bold;">1</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>),
torch.tensor([<span style="color: #8a3b3c; font-weight: bold;">0.0</span>]),
)
G.train(D, generate_random_seed(<span style="color: #8a3b3c; font-weight: bold;">200</span>), torch.tensor([<span style="color: #8a3b3c; font-weight: bold;">1.0</span>]))
<span style="font-weight: bold;">pass</span>
<span style="font-weight: bold;">pass</span>
</pre>
</div>
<p>
后面的测试只放修改位置的代码。<br>
</p>
</div>
<div id="outline-container-org476dc8d" class="outline-4">
<h4 id="org476dc8d">训练结果</h4>
<div class="outline-text-4" id="text-org476dc8d">
<figure id="orgaf957af">
<img src="./static/img/regularization_with_gan_output/normal.png" alt="normal.png"><br>
</figure>
</div>
</div>
<div id="outline-container-orgfd9449d" class="outline-4">
<h4 id="orgfd9449d">评价</h4>
<div class="outline-text-4" id="text-orgfd9449d">
<p>
有些许的模式崩溃现象，只有几种类型的数字，数字的颗粒感比较严重，不够平滑。<br>
</p>
</div>
</div>
</div>
<div id="outline-container-orgbe1d6fd" class="outline-3">
<h3 id="orgbe1d6fd">L1</h3>
<div class="outline-text-3" id="text-orgbe1d6fd">
<p>
更改训练函数即可<br>
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="font-weight: bold;">def</span> <span style="font-weight: bold;">train</span>(<span style="font-weight: bold;">self</span>, D, inputs, targets):
<span style="color: #383a42;">g_output</span> = <span style="font-weight: bold;">self</span>.forward(inputs)
<span style="color: #383a42;">d_output</span> = D.forward(g_output.reshape(<span style="color: #8a3b3c; font-weight: bold;">1</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span>))
<span style="color: #383a42;">loss</span> = D.loss_function(d_output, targets)
<span style="color: #383a42;">l1_loss</span> = <span style="color: #8a3b3c; font-weight: bold;">0</span>
<span style="font-weight: bold;">for</span> param <span style="font-weight: bold;">in</span> <span style="font-weight: bold;">self</span>.parameters():
<span style="color: #383a42;">l1_loss</span> += torch.<span style="font-weight: bold;">sum</span>(torch.<span style="font-weight: bold;">abs</span>(param))
<span style="font-weight: bold;">pass</span>
<span style="color: #383a42;">loss</span> += <span style="color: #8a3b3c; font-weight: bold;">0.005</span> * l1_loss
<span style="font-weight: bold;">self</span>.optimiser.zero_grad()
loss.backward()
<span style="font-weight: bold;">self</span>.optimiser.step()
<span style="font-weight: bold;">pass</span>
</pre>
</div>
</div>
<div id="outline-container-org0ee3427" class="outline-4">
<h4 id="org0ee3427">训练结果</h4>
<div class="outline-text-4" id="text-org0ee3427">
<figure id="orgda7de16">
<img src="./static/img/regularization_with_gan_output/l1.png" alt="l1.png"><br>
</figure>
</div>
</div>
<div id="outline-container-org09291a8" class="outline-4">
<h4 id="org09291a8">评价</h4>
<div class="outline-text-4" id="text-org09291a8">
<p>
几乎没有可以识别的数字，可能是训练次数不够，但数字可以明显看出比较平滑，也就是那种模糊的效果。<br>
</p>
</div>
</div>
</div>
<div id="outline-container-org401c451" class="outline-3">
<h3 id="org401c451">L2</h3>
<div class="outline-text-3" id="text-org401c451">
<p>
<code>pytorch</code> 的优化器通过参数 <code>weight_decay</code> 可以实现 <code>L2</code> 正则。<br>
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="font-weight: bold;">self</span>.<span style="color: #383a42;">optimiser</span> = torch.optim.Adam(
<span style="font-weight: bold;">self</span>.parameters(), lr=<span style="color: #8a3b3c; font-weight: bold;">0.0001</span>, weight_decay=<span style="color: #8a3b3c; font-weight: bold;">0.005</span>
)
</pre>
</div>
</div>
<div id="outline-container-org4d2ccc8" class="outline-4">
<h4 id="org4d2ccc8">训练结果</h4>
<div class="outline-text-4" id="text-org4d2ccc8">
<figure id="org00f7cc7">
<img src="./static/img/regularization_with_gan_output/l2.png" alt="l2.png"><br>
</figure>
</div>
</div>
<div id="outline-container-org6275533" class="outline-4">
<h4 id="org6275533">评价</h4>
<div class="outline-text-4" id="text-org6275533">
<p>
相比 <code>L1</code> 数字周围更加干净，更加清晰一些，但模糊的效果还在，与 <code>L1</code> 的差异不排除随机误差的可能。<br>
</p>
</div>
</div>
</div>
<div id="outline-container-orgc8a7ec6" class="outline-3">
<h3 id="orgc8a7ec6">Dropout</h3>
<div class="outline-text-3" id="text-orgc8a7ec6">
<p>
这个添加一个 <code>Dropout</code> 层即可<br>
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="font-weight: bold;">self</span>.<span style="color: #383a42;">model</span> = nn.Sequential(
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">200</span>, <span style="color: #8a3b3c; font-weight: bold;">400</span>),
nn.Dropout(<span style="color: #8a3b3c; font-weight: bold;">0.5</span>),
nn.GELU(),
nn.LayerNorm(<span style="color: #8a3b3c; font-weight: bold;">400</span>),
nn.Linear(<span style="color: #8a3b3c; font-weight: bold;">400</span>, <span style="color: #8a3b3c; font-weight: bold;">28</span> * <span style="color: #8a3b3c; font-weight: bold;">28</span>),
nn.Sigmoid(),
)
</pre>
</div>
</div>
<div id="outline-container-org3c0eb7b" class="outline-4">
<h4 id="org3c0eb7b">训练结果</h4>
<div class="outline-text-4" id="text-org3c0eb7b">
<figure id="org67078d9">
<img src="./static/img/regularization_with_gan_output/dropout.png" alt="dropout.png"><br>
</figure>
</div>
</div>
<div id="outline-container-org6a94aad" class="outline-4">
<h4 id="org6a94aad">评价</h4>
<div class="outline-text-4" id="text-org6a94aad">
<p>
没有了模糊效果，但数字都不成形，可能是训练次数不够。<br>
</p>
</div>
</div>
</div>
<div id="outline-container-orga0f0edb" class="outline-3">
<h3 id="orga0f0edb">Dropout + L2</h3>
<div class="outline-text-3" id="text-orga0f0edb">
</div>
<div id="outline-container-org0995401" class="outline-4">
<h4 id="org0995401">训练结果</h4>
<div class="outline-text-4" id="text-org0995401">
<figure id="org0057163">
<img src="./static/img/regularization_with_gan_output/dropout_l2.png" alt="dropout_l2.png"><br>
</figure>
</div>
</div>
<div id="outline-container-org538c2aa" class="outline-4">
<h4 id="org538c2aa">评价</h4>
<div class="outline-text-4" id="text-org538c2aa">
<p>
数字很平滑，但都不成型。<br>
</p>
</div>
</div>
</div>
</div>
<div id="outline-container-org69fd22e" class="outline-2">
<h2 id="org69fd22e">结论</h2>
<div class="outline-text-2" id="text-org69fd22e">
<p>
经过我不严谨的实验可以发现， <code>L1</code> 与 <code>L2</code> 正则化生成的图片确实会有模糊的效果，具体的原因可能是因为 <code>L1</code> 会让神经网络的参数矩阵更加稀疏， <code>L2</code> 不会让单个参数过大，也就不会因为一个神经元大范围改变结果，而尖锐的效果就是相邻的像素之间差别过大，故会产生模糊的效果。<br>
</p>
</div>
</div>
</div>
<div id="postamble" class="status">
<script type="text/javascript" src="/static/main.js"></script>
<center id="modi-date">2021-12-09 四 17:18</center>
</div>
</body>
</html>