-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcs231n_lec8.html
executable file
·1060 lines (717 loc) · 48.9 KB
/
cs231n_lec8.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
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html>
<head>
<!-- Document Settings -->
<meta charset="utf-8" />
<meta http-equiv="X-UA-Compatible" content="IE=edge" />
<!-- On Post front-matter YAML, set "use_math: true" to use LaTex -->
<script type="text/x-mathjax-config">
MathJax.Hub.Config({
TeX: {
equationNumbers: {
autoNumber: "AMS"
}
},
tex2jax: {
inlineMath: [ ['$', '$'], ["\\(","\\)"] ],
displayMath: [ ['$$', '$$'], ["\\[","\\]"] ],
processEscapes: true,
}
});
MathJax.Hub.Register.MessageHook("Math Processing Error",function (message) {
alert("Math Processing Error: "+message[1]);
});
MathJax.Hub.Register.MessageHook("TeX Jax - parse error",function (message) {
alert("Math Processing Error: "+message[1]);
});
</script>
<script type="text/javascript" async
src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/MathJax.js?config=TeX-MML-AM_CHTML">
</script>
<!-- Base Meta -->
<!-- dynamically fixing the title for tag/author pages -->
<title>cs231n - Lecture 8. Training Neural Networks II</title>
<meta name="HandheldFriendly" content="True" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<!-- Styles'n'Scripts -->
<link rel="stylesheet" type="text/css" href="/assets/built/screen.css" />
<link rel="stylesheet" type="text/css" href="/assets/built/screen.edited.css" />
<link rel="stylesheet" type="text/css" href="/assets/built/syntax.css" />
<!-- syntax.css -->
<link rel="stylesheet" type="text/css" href="/assets/built/syntax.css" />
<!-- highlight.js -->
<link rel="stylesheet" href="//cdnjs.cloudflare.com/ajax/libs/highlight.js/9.12.0/styles/default.min.css">
<style>.hljs { background: none; }</style>
<!--[if IE]>
<style>
p, ol, ul{
width: 100%;
}
blockquote{
width: 100%;
}
</style>
<![endif]-->
<!-- This tag outputs SEO meta+structured data and other important settings -->
<meta name="description" content="" />
<link rel="shortcut icon" href="http://0.0.0.0:4000/assets/built/images/favicon.jpg" type="image/png" />
<link rel="canonical" href="http://0.0.0.0:4000/cs231n_lec8" />
<meta name="referrer" content="no-referrer-when-downgrade" />
<!--title below is coming from _includes/dynamic_title-->
<meta property="og:site_name" content="Darron's Devlog" />
<meta property="og:type" content="website" />
<meta property="og:title" content="cs231n - Lecture 8. Training Neural Networks II" />
<meta property="og:description" content="Optimization Problems with SGD What if loss changes quickly in one direction and slowly in another? What does gradient descent do? Very slow progress along shallow dimension, jitter along steep direction What if the loss function has a local minima or saddle point? Zero gradient, gradient descent gets stuck(more common" />
<meta property="og:url" content="http://0.0.0.0:4000/cs231n_lec8" />
<meta property="og:image" content="http://0.0.0.0:4000/assets/built/images/blog-cover1.png" />
<meta property="article:publisher" content="https://www.facebook.com/" />
<meta property="article:author" content="https://www.facebook.com/" />
<meta property="article:published_time" content="2021-12-27T15:00:00+00:00" />
<meta property="article:modified_time" content="2021-12-27T15:00:00+00:00" />
<meta property="article:tag" content="cs231n" />
<meta name="twitter:card" content="summary_large_image" />
<meta name="twitter:title" content="cs231n - Lecture 8. Training Neural Networks II" />
<meta name="twitter:description" content="Optimization Problems with SGD What if loss changes quickly in one direction and slowly in another? What does gradient descent do? Very slow progress along shallow dimension, jitter along steep direction What if the loss function has a local minima or saddle point? Zero gradient, gradient descent gets stuck(more common" />
<meta name="twitter:url" content="http://0.0.0.0:4000/" />
<meta name="twitter:image" content="http://0.0.0.0:4000/assets/built/images/blog-cover1.png" />
<meta name="twitter:label1" content="Written by" />
<meta name="twitter:data1" content="Darron's Devlog" />
<meta name="twitter:label2" content="Filed under" />
<meta name="twitter:data2" content="cs231n" />
<meta name="twitter:site" content="@" />
<meta name="twitter:creator" content="@" />
<meta property="og:image:width" content="1400" />
<meta property="og:image:height" content="933" />
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Website",
"publisher": {
"@type": "Organization",
"name": "Darron's Devlog",
"logo": "http://0.0.0.0:4000/"
},
"url": "http://0.0.0.0:4000/cs231n_lec8",
"image": {
"@type": "ImageObject",
"url": "http://0.0.0.0:4000/assets/built/images/blog-cover1.png",
"width": 2000,
"height": 666
},
"mainEntityOfPage": {
"@type": "WebPage",
"@id": "http://0.0.0.0:4000/cs231n_lec8"
},
"description": "Optimization Problems with SGD What if loss changes quickly in one direction and slowly in another? What does gradient descent do? Very slow progress along shallow dimension, jitter along steep direction What if the loss function has a local minima or saddle point? Zero gradient, gradient descent gets stuck(more common"
}
</script>
<!-- <script type="text/javascript" src="https://demo.ghost.io/public/ghost-sdk.min.js?v=724281a32e"></script>
<script type="text/javascript">
ghost.init({
clientId: "ghost-frontend",
clientSecret: "f84a07a72b17"
});
</script> -->
<meta name="generator" content="Jekyll 3.6.2" />
<link rel="alternate" type="application/rss+xml" title="cs231n - Lecture 8. Training Neural Networks II" href="/feed.xml" />
</head>
<body class="post-template">
<div class="site-wrapper">
<!-- All the main content gets inserted here, index.hbs, post.hbs, etc -->
<!-- default -->
<!-- The tag above means: insert everything in this file
into the {body} of the default.hbs template -->
<header class="site-header outer">
<div class="inner">
<nav class="site-nav">
<div class="site-nav-left">
<a class="site-nav-logo" href="/">Darron's Devlog</a>
<ul class="nav" role="menu">
<li class="nav-home" role="menuitem"><a href="/">Home</a></li>
<li class="nav-about" role="menuitem"><a href="/about/">About</a></li>
<li class="nav-projects" role="menuitem"><a href="/tag/projects/">Projects</a></li>
<li class="nav-studies" role="menuitem"><a href="/tag/studies/">Studies</a></li>
<li class="nav-blog" role="menuitem"><a href="/tag/blog/">Blog</a></li>
<li class="nav-archive" role="menuitem">
<a href="/archive.html">All Posts</a>
</li>
</ul>
</div>
<div class="site-nav-right">
<div class="social-links">
</div>
<a class="subscribe-button" href="#subscribe">Search</a>
</div>
</nav>
</div>
</header>
<!-- Everything inside the #post tags pulls data from the post -->
<!-- #post -->
<main id="site-main" class="site-main outer" role="main">
<div class="inner">
<article class="post-full tag-cs231n no-image">
<header class="post-full-header">
<section class="post-full-meta">
<time class="post-full-meta-date" datetime="27 December 2021">27 December 2021</time>
<span class="date-divider">/</span>
<a href='/tag/cs231n/'>CS231N</a>
</section>
<h1 class="post-full-title">cs231n - Lecture 8. Training Neural Networks II</h1>
</header>
<!--
-->
<section class="post-full-content">
<div class="kg-card-markdown">
<h2 id="optimization">Optimization</h2>
<h3 id="problems-with-sgd">Problems with SGD</h3>
<ol>
<li>
<p>What if loss changes quickly in one direction and slowly in another? What does gradient descent do?
Very slow progress along shallow dimension, jitter along steep direction</p>
</li>
<li>
<p>What if the loss function has a local minima or saddle point?<br />
Zero gradient, gradient descent gets stuck(more common in high dimension)</p>
</li>
<li>
<p>Gradients come from minibatches can be noisy</p>
</li>
</ol>
<h3 id="sgd--momentum">SGD + Momentum</h3>
<ul>
<li>To avoid local minima, combine gradient at current point with <em>velocity</em> to get step used to update weights; continue moving in the general direction as the previous iterations<br />
\(v_{t+1}=\rho v_t + \nabla f(x_t)\)<br />
\(x_{t+1}=x_t - \alpha v_{t+1}\)<br />
with <em>rho</em> giving “friction”; typically <em>0.9</em> or <em>0.99</em></li>
</ul>
<div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="n">vx</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">while</span> <span class="bp">True</span><span class="p">:</span>
<span class="n">dx</span> <span class="o">=</span> <span class="n">compute_gradient</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">vx</span> <span class="o">=</span> <span class="n">rho</span> <span class="o">*</span> <span class="n">vx</span> <span class="o">+</span> <span class="n">dx</span>
<span class="n">x</span> <span class="o">-=</span> <span class="n">learning_rate</span> <span class="o">*</span> <span class="n">vx</span>
</code></pre></div></div>
<h3 id="nesterov-momentum">Nesterov Momentum</h3>
<ul>
<li>“Look ahead” to the point where updating using velocity would take us; compute gradient there and mix it with velocity to get actual update direction<br />
\(v_{t+1}=\rho v_t - \alpha\nabla f(x_t + \rho v_t)\)<br />
\(x_{t+1}=x_t + v_{t+1}\)<br />
rearrange with \(\tilde{x}_t = x_t + \rho v_t\),<br />
\(v_{t+1}=\rho v_t - \alpha\nabla f(\tilde{x}_t)\)<br />
\(\begin{align*}
\tilde{x}_{t+1} &= \tilde{x}_t - \rho v_t + (1+\rho)v_{t+1}
&= \tilde{x}_t + v_{t+1} + \rho(v_{t+1}-v_t)
\end{align*}\)</li>
</ul>
<h3 id="adagrad">AdaGrad</h3>
<div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="n">grad_squared</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">while</span> <span class="bp">True</span><span class="p">:</span>
<span class="n">dx</span> <span class="o">=</span> <span class="n">compute_gradient</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">grad_squared</span> <span class="o">+=</span> <span class="n">dx</span> <span class="o">*</span> <span class="n">dx</span>
<span class="n">x</span> <span class="o">-=</span> <span class="n">learning_rate</span> <span class="o">*</span> <span class="n">dx</span> <span class="o">/</span> <span class="p">(</span><span class="n">np</span><span class="p">.</span><span class="n">sqrt</span><span class="p">(</span><span class="n">grad_squared</span><span class="p">)</span> <span class="o">+</span> <span class="mf">1e-7</span><span class="p">)</span>
</code></pre></div></div>
<ul>
<li>Added element-wise scaling of the gradient based on the historical sum of squares in each dimension<br />
“Per-parameter learning rates” or “adaptive learning rates”<br />
Progress along “steep” directions is damped and “flat” directions is accelerated<br />
Step size decays to zero over time</li>
</ul>
<h3 id="rmsprop-leaky-adagrad">RMSProp: “Leaky AdaGrad”</h3>
<div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="n">grad_squared</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">while</span> <span class="bp">True</span><span class="p">:</span>
<span class="n">dx</span> <span class="o">=</span> <span class="n">compute_gradient</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">grad_squared</span> <span class="o">=</span> <span class="n">decay_rate</span> <span class="o">*</span> <span class="n">grad_squared</span> <span class="o">+</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">decay_rate</span><span class="p">)</span> <span class="o">*</span> <span class="n">dx</span> <span class="o">*</span> <span class="n">dx</span>
<span class="n">x</span> <span class="o">-=</span> <span class="n">learning_rate</span> <span class="o">*</span> <span class="n">dx</span> <span class="o">/</span> <span class="p">(</span><span class="n">np</span><span class="p">.</span><span class="n">sqrt</span><span class="p">(</span><span class="n">grad_squared</span><span class="p">)</span> <span class="o">+</span> <span class="mf">1e-7</span><span class="p">)</span>
</code></pre></div></div>
<h3 id="adam">Adam</h3>
<div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="n">first_moment</span> <span class="o">=</span> <span class="mi">0</span>
<span class="n">second_moment</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">t</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">num_iterations</span><span class="p">):</span>
<span class="n">dx</span> <span class="o">=</span> <span class="n">compute_gradient</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">first_moment</span> <span class="o">=</span> <span class="n">beta1</span> <span class="o">*</span> <span class="n">first_moment</span> <span class="o">+</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">beta1</span><span class="p">)</span> <span class="o">*</span> <span class="n">dx</span> <span class="c1"># Momentum
</span> <span class="n">second_moment</span> <span class="o">=</span> <span class="n">beta2</span> <span class="o">*</span> <span class="n">second_moment</span> <span class="o">+</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">beta2</span><span class="p">)</span> <span class="o">*</span> <span class="n">dx</span> <span class="o">*</span> <span class="n">dx</span>
<span class="n">first_unbias</span> <span class="o">=</span> <span class="n">first_moment</span> <span class="o">/</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">beta1</span> <span class="o">**</span> <span class="n">t</span><span class="p">)</span> <span class="c1"># Bias correction
</span> <span class="n">second_unbias</span> <span class="o">=</span> <span class="n">second_moment</span> <span class="o">/</span> <span class="p">(</span><span class="mi">1</span> <span class="o">-</span> <span class="n">beta2</span> <span class="o">**</span> <span class="n">t</span><span class="p">)</span>
<span class="n">x</span> <span class="o">-=</span> <span class="n">learning_rate</span> <span class="o">*</span> <span class="n">first_unbias</span> <span class="o">/</span> <span class="p">(</span><span class="n">np</span><span class="p">.</span><span class="n">sqrt</span><span class="p">(</span><span class="n">second_unbias</span><span class="p">)</span> <span class="o">+</span> <span class="mf">1e-7</span><span class="p">)</span> <span class="c1"># AdaGrad/ RMSProp
</span></code></pre></div></div>
<ul>
<li>Sort of like RMSProp with momentum
Bias correction for the fact that first and second moment estimates start at zero<br />
Adam with <em>beta1 = 0.9, beta2 = 0.999, and learning_rate = 1e-3 or 5e-4</em> is a great starting point</li>
</ul>
<h2 id="learning-rate-schedules">Learning rate schedules</h2>
<h3 id="learning-rate-decays-over-time">Learning rate decays over time</h3>
<ul>
<li>Reduce learning rate by a certain value at a few fixed points(after some epochs)</li>
</ul>
<h3 id="learning-rate-decay">Learning Rate Decay</h3>
<ul>
<li>
<p>Reduce learning rate gradually, e.g.<br />
Cosine: $\alpha_t = \frac{1}{2}\alpha_0(1+\mbox{cos}(t\pi / T))$<br />
Linear: $\alpha_t = \alpha_0(1-t/T)$<br />
Inverse sqrt: $\alpha_t = \alpha_0 / \sqrt{t}$<br />
while $\alpha_0$ is the initial learning rate, $\alpha_t$ is one at epoch <em>t</em>, and <em>T</em> is the total number of epochs</p>
</li>
<li>
<p>Linear Warmup<br />
High initial learning rates can make loss explode; linearly increasing learning rate from <em>0</em> over the first <em>~5000</em> iterations can prevent this<br />
Empirical rule of thumb: If you increase the batch size by <em>N</em>, also scale the initial learning rate by <em>N</em></p>
</li>
</ul>
<h3 id="first-order-optimization">First-Order Optimization</h3>
<ol>
<li>Use gradient from linear approximation</li>
<li>Step to minimize the approximation</li>
</ol>
<h3 id="second-order-optimization">Second-Order Optimization</h3>
<ol>
<li>Use gradient and <strong>Hessian</strong> to form <strong>quadratic</strong> approximation</li>
<li>Step to the <strong>minima</strong> of the (quadratic) approximation</li>
</ol>
<ul>
<li>But Hessian has <em>O(N^2)</em> elements and inverting takes <em>O(N^3)</em>, <em>N</em> is extremely large
<ul>
<li>Quasi-Newton methods (BGFS most popular):<br />
instead of inverting the Hessian, approximate inverse Hessian with rank 1 updates over time</li>
<li>L-BFGS (Limited memory BFGS):<br />
Does not form/store the full inverse Hessian. Usually works very well in full batch, deterministic mode, but does not transfer very well to mini-batch setting. Large-scale, stochastic setting is an active area of research.</li>
</ul>
</li>
</ul>
<h3 id="summary">Summary</h3>
<ul>
<li>Adam is a good default choice in many cases; even with constant learning rate</li>
<li>SGD+Momentum can outperform Adam but may equire more tuning of LR and schedule. Cosine schedule preferred, since it has very few hyperparameters.</li>
<li>L-BFGS is good if you can afford to do full batch updates.</li>
</ul>
<h2 id="improve-test-error">Improve test error</h2>
<ul>
<li>Better optimization algorithms help reduce <strong>training</strong> loss, but what we really care is about error on new data - how to reduce the gap?</li>
</ul>
<h3 id="early-stopping-always-do-this">Early Stopping: Always do this</h3>
<ul>
<li>Stop training the model when accuracy on the validation set decreases. Or train for a long time, but always keep track of the model snapshot that worked best on val.</li>
</ul>
<h3 id="model-ensembles">Model Ensembles</h3>
<ol>
<li>Train multiple independent models</li>
<li>At test time average their results<br />
(Take average of predicted probability distributions, then choose argmax)</li>
</ol>
<h3 id="regularization">Regularization</h3>
<ul>
<li>
<p>To improve single-model performance, add terms to loss<br />
e.g. L1, L2, Elastic net.</p>
</li>
<li>
<p>or, use Dropout:<br />
In each forward pass, randomly set some neurons to zero. Probability of dropping is a hyperparameter; 0.5 is common.<br />
It forces the network to have a redundant representation; Prevents co-adaptation of features. Dropout can be interpreted as training a large ensemble of models (that share parameters).</p>
</li>
</ul>
<div class="language-python highlighter-rouge"><div class="highlight"><pre class="highlight"><code><span class="n">p</span> <span class="o">=</span> <span class="mf">0.5</span> <span class="c1"># dropout rate
</span>
<span class="k">def</span> <span class="nf">train_step</span><span class="p">(</span><span class="n">X</span><span class="p">):</span> <span class="c1"># drop in train time
</span> <span class="n">H1</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">maximum</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W1</span><span class="p">,</span> <span class="n">X</span><span class="p">)</span> <span class="o">+</span> <span class="n">b1</span><span class="p">)</span>
<span class="n">U1</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">random</span><span class="p">.</span><span class="n">rand</span><span class="p">(</span><span class="o">*</span><span class="n">H1</span><span class="p">.</span><span class="n">shape</span><span class="p">)</span> <span class="o"><</span> <span class="n">p</span>
<span class="n">H1</span> <span class="o">*=</span> <span class="n">U1</span>
<span class="n">H2</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">maximum</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W2</span><span class="p">,</span> <span class="n">H1</span><span class="p">)</span> <span class="o">+</span> <span class="n">b2</span><span class="p">)</span>
<span class="n">U2</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">random</span><span class="p">.</span><span class="n">rand</span><span class="p">(</span><span class="o">*</span><span class="n">H2</span><span class="p">.</span><span class="n">shape</span><span class="p">)</span> <span class="o"><</span> <span class="n">p</span>
<span class="n">H2</span> <span class="o">*=</span> <span class="n">U2</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W3</span><span class="p">,</span> <span class="n">H2</span><span class="p">)</span> <span class="o">+</span> <span class="n">b3</span>
<span class="c1"># backward pass: compute gradients...
</span> <span class="c1"># perform parameter update...
</span>
<span class="k">def</span> <span class="nf">predict</span><span class="p">(</span><span class="n">X</span><span class="p">):</span> <span class="c1"># scale at test time
</span> <span class="n">H1</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">maximum</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W1</span><span class="p">,</span> <span class="n">X</span><span class="p">)</span> <span class="o">+</span> <span class="n">b1</span><span class="p">)</span> <span class="o">*</span> <span class="n">p</span>
<span class="n">H2</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">maximum</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W2</span><span class="p">,</span> <span class="n">H1</span><span class="p">)</span> <span class="o">+</span> <span class="n">b2</span><span class="p">)</span> <span class="o">*</span> <span class="n">p</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">np</span><span class="p">.</span><span class="n">dot</span><span class="p">(</span><span class="n">W3</span><span class="p">,</span> <span class="n">H2</span><span class="p">)</span> <span class="o">+</span> <span class="n">b3</span>
</code></pre></div></div>
<ul>
<li>
<p>more common: “Inverted dropout”<br />
<code class="language-plaintext highlighter-rouge">U1 = (np.random.rand(*H1.shape) < p) / p</code> in train time and no scaling in test time</p>
</li>
<li>
<p>A common pattern of regularization<br />
Training: Add some kind of randomness<br />
$y = fw(x,z)$<br />
Testing: Average out randomness (sometimes approximate)<br />
\(y = f(x) = E_z[f(x,z)] = \int p(z)f(x,z)\, dz\)</p>
</li>
<li>
<p>Data Augmentation:<br />
Addes <em>transformed</em> data to train model<br />
e.g. translation, rotation, stretching, shearing, lens distortions.</p>
</li>
<li>
<p>DropConnect:<br />
Training: Drop connections between neurons (set weights to 0)<br />
Testing: Use all the connections</p>
</li>
<li>
<p>Fractional Pooling:<br />
Training: Use randomized pooling regions<br />
Testing: Average predictions from several regions</p>
</li>
<li>
<p>Stochastic Depth:<br />
Training: Skip some layers in the network<br />
Testing: Use all the layer</p>
</li>
<li>
<p>Cutout:<br />
Training: Set random image regions to zero<br />
Testing: Use full image</p>
</li>
<li>
<p>Mixup:<br />
Training: Train on random blends of images<br />
Testing: Use original images<br />
e.g. Randomly blend the pixels of pairs of training images, say 40% cat and 60% dog, and set the target label as cat:0.4 and dog:0.6.</p>
</li>
<li>
<p>Summary
Consider dropout for large fully-connected layers<br />
Batch normalization and data augmentation almost always a good idea<br />
Try cutout and mixup especially for small classification datasets</p>
</li>
</ul>
<h2 id="choosing-hyperparameters">Choosing Hyperparameters</h2>
<ul>
<li>
<p>Step 1: Check initial loss<br />
Turn off weight decay, sanity check loss at initialization<br />
e.g. <em>log(C)</em> for softmax with <em>C</em> classes</p>
</li>
<li>
<p>Step 2: Overfit a small sample<br />
Try to train to 100% training accuracy on a small sample of training data (~5-10 minibatches); fiddle with architecture, learning rate, weight initialization<br />
If loss is not going down, LR too low or bad initialization. If loss explodes, then LR is too high or bad initialization.</p>
</li>
<li>
<p>Step 3: Find LR that makes loss go down<br />
Use the architecture from the previous step, use all training data, turn on small weight decay, find a learning rate that makes the loss drop significantly within ~100 iterations.</p>
</li>
<li>
<p>Step 4: Coarse grid, train for ~1-5 epochs<br />
Choose a few values of learning rate and weight decay around what worked from Step 3, train a few models for ~1-5 epochs.</p>
</li>
<li>
<p>Step 5: Refine grid, train longer<br />
Pick best models from Step 4, train them for longer (~10-20 epochs) without learning rate decay</p>
</li>
<li>
<p>Step 6: Look at loss and accuracy curves<br />
If accuracy still going up, you need to train longer. If it goes down, huge train / val gap means overfitting. You need to increase regularization or get more data. If there’s no gap between train / val, it means underfitting. Train longer or use a bigger model.</p>
</li>
<li>
<p>Look at learning curves<br />
Losses may be noisy, use a scatter plot and also plot moving average to see trends better. Cross-validation is useful too.</p>
</li>
<li>
<p>Step 7: <strong>GO TO Step 5</strong></p>
</li>
<li>
<p>Hyperparameters to play with:<br />
network architecture,<br />
learning rate, its decay schedule, update type,<br />
regularization (L2/Dropout strength)</p>
</li>
<li>
<p>for Hyper-Parameter Optimization, consider both Random Search and Grid Search</p>
</li>
</ul>
<h3 id="summary-1">Summary</h3>
<ul>
<li>Improve your training error:
<ul>
<li>Optimizers</li>
<li>Learning rate schedules</li>
</ul>
</li>
<li>Improve your test error:
<ul>
<li>Regularization</li>
<li>Choosing Hyperparameters</li>
</ul>
</li>
</ul>
</div>
</section>
<!-- Email subscribe form at the bottom of the page -->
<!--
<section class="subscribe-form">
<h3 class="subscribe-form-title">Subscribe to Darron's Devlog</h3>
<p>Get the latest posts delivered right to your inbox</p>
<span id="searchform" method="post" action="/subscribe/" class="">
<input class="confirm" type="hidden" name="confirm" />
<input class="location" type="hidden" name="location" />
<input class="referrer" type="hidden" name="referrer" />
<div class="form-group">
<input class="subscribe-email" onkeyup="myFunc()"
id="searchtext" type="text" name="searchtext"
placeholder="Search..." />
</div>
<script type="text/javascript">
function myFunc() {
if(event.keyCode == 13) {
var url = encodeURIComponent($("#searchtext").val());
location.href = "/search.html?query=" + url;
}
}
</script>
</span>
</section>
-->
<footer class="post-full-footer">
<!-- Everything inside the #author tags pulls data from the author -->
<!-- #author-->
<!-- /author -->
</footer>
<!-- If you use Disqus comments, just uncomment this block.
The only thing you need to change is "test-apkdzgmqhj" - which
should be replaced with your own Disqus site-id. -->
<section class="post-full-comments">
<div id="disqus_thread"></div>
<script>
var disqus_config = function () {
var this_page_url = 'http://0.0.0.0:4000/cs231n_lec8';
var this_page_identifier = '/cs231n_lec8';
var this_page_title = 'cs231n - Lecture 8. Training Neural Networks II';
};
(function() {
var d = document, s = d.createElement('script');
s.src = 'https://.disqus.com/embed.js';
s.setAttribute('data-timestamp', +new Date());
(d.head || d.body).appendChild(s);
})();
</script>
</section>
</article>
</div>
</main>
<!-- Links to Previous/Next posts -->
<aside class="read-next outer">
<div class="inner">
<div class="read-next-feed">
<article class="read-next-card"
style="background-image: url(/assets/built/images/blog-cover1.png)"
>
<header class="read-next-card-header">
<small class="read-next-card-header-sitetitle">— Darron's Devlog —</small>
<h3 class="read-next-card-header-title"><a href="/tag/cs231n/">Cs231n</a></h3>
</header>
<div class="read-next-divider"><svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path d="M13 14.5s2 3 5 3 5.5-2.463 5.5-5.5S21 6.5 18 6.5c-5 0-7 11-12 11C2.962 17.5.5 15.037.5 12S3 6.5 6 6.5s4.5 3.5 4.5 3.5"/></svg>
</div>
<div class="read-next-card-content">
<ul>
<li><a href="/cs231n_lec15">cs231n - Lecture 15. Detection and Segmentation</a></li>
<li><a href="/cs231n_lec14">cs231n - Lecture 14. Visualizing and Understanding</a></li>
<li><a href="/cs231n_lec13">cs231n - Lecture 13. Self-Supervised Learning</a></li>
</ul>
</div>
<footer class="read-next-card-footer">
<a href="/tag/cs231n/">
See all 13 posts →
</a>
</footer>
</article>
<!-- If there's a next post, display it using the same markup included from - partials/post-card.hbs -->
<article class="post-card post-template no-image">
<div class="post-card-content">
<a class="post-card-content-link" href="/cs231n_lec9">
<header class="post-card-header">
<span class="post-card-tags">Cs231n</span>
<h2 class="post-card-title">cs231n - Lecture 9. CNN Architectures</h2>
</header>
<section class="post-card-excerpt">
<p>Review LeCun et al., 1998 $5\times 5$ Conv filters applied at stride 1 $2\times 2$ Subsampling (Pooling) layers applied at stride 2 i.e. architecture is [CONV-POOL-CONV-POOL-FC-FC] Stride: Downsample output activations Padding: Preserve input</p>
</section>
</a>
<footer class="post-card-meta">
</footer>
</div>
</article>
<!-- If there's a previous post, display it using the same markup included from - partials/post-card.hbs -->
<article class="post-card post-template no-image">
<div class="post-card-content">
<a class="post-card-content-link" href="/cs231n_lec7">
<header class="post-card-header">
<span class="post-card-tags">Cs231n</span>
<h2 class="post-card-title">cs231n - Lecture 7. Training Neural Networks I</h2>
</header>
<section class="post-card-excerpt">
<p>Activation Functions Sigmoid $\sigma(x)=1/(1+e^{-x})$ Squashes numbers to range [0,1] Historically popular since they have nice interpretation as a saturating “firing rate” of a neuron. Problem: Gradient Vanishing: Saturated neurons “kill” the gradients; If</p>
</section>
</a>
<footer class="post-card-meta">
</footer>
</div>
</article>
</div>
</div>
</aside>
<!-- Floating header which appears on-scroll, included from includes/floating-header.hbs -->
<div class="floating-header">
<div class="floating-header-logo">
<a href="/">
<span>Darron's Devlog</span>
</a>
</div>
<span class="floating-header-divider">—</span>
<div class="floating-header-title">cs231n - Lecture 8. Training Neural Networks II</div>
<div class="floating-header-share">
<div class="floating-header-share-label">Share this <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24">
<path d="M7.5 15.5V4a1.5 1.5 0 1 1 3 0v4.5h2a1 1 0 0 1 1 1h2a1 1 0 0 1 1 1H18a1.5 1.5 0 0 1 1.5 1.5v3.099c0 .929-.13 1.854-.385 2.748L17.5 23.5h-9c-1.5-2-5.417-8.673-5.417-8.673a1.2 1.2 0 0 1 1.76-1.605L7.5 15.5zm6-6v2m-3-3.5v3.5m6-1v2"/>
</svg>
</div>
<a class="floating-header-share-tw" href="https://twitter.com/share?text=cs231n+-+Lecture+8.+Training+Neural+Networks+II&url=https://12kdh43.github.io/cs231n_lec8"
onclick="window.open(this.href, 'share-twitter', 'width=550,height=235');return false;">
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32"><path d="M30.063 7.313c-.813 1.125-1.75 2.125-2.875 2.938v.75c0 1.563-.188 3.125-.688 4.625a15.088 15.088 0 0 1-2.063 4.438c-.875 1.438-2 2.688-3.25 3.813a15.015 15.015 0 0 1-4.625 2.563c-1.813.688-3.75 1-5.75 1-3.25 0-6.188-.875-8.875-2.625.438.063.875.125 1.375.125 2.688 0 5.063-.875 7.188-2.5-1.25 0-2.375-.375-3.375-1.125s-1.688-1.688-2.063-2.875c.438.063.813.125 1.125.125.5 0 1-.063 1.5-.25-1.313-.25-2.438-.938-3.313-1.938a5.673 5.673 0 0 1-1.313-3.688v-.063c.813.438 1.688.688 2.625.688a5.228 5.228 0 0 1-1.875-2c-.5-.875-.688-1.813-.688-2.75 0-1.063.25-2.063.75-2.938 1.438 1.75 3.188 3.188 5.25 4.25s4.313 1.688 6.688 1.813a5.579 5.579 0 0 1 1.5-5.438c1.125-1.125 2.5-1.688 4.125-1.688s3.063.625 4.188 1.813a11.48 11.48 0 0 0 3.688-1.375c-.438 1.375-1.313 2.438-2.563 3.188 1.125-.125 2.188-.438 3.313-.875z"/></svg>
</a>
<a class="floating-header-share-fb" href="https://www.facebook.com/sharer/sharer.php?u=https://12kdh43.github.io/cs231n_lec8"
onclick="window.open(this.href, 'share-facebook','width=580,height=296');return false;">
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 32 32"><path d="M19 6h5V0h-5c-3.86 0-7 3.14-7 7v3H8v6h4v16h6V16h5l1-6h-6V7c0-.542.458-1 1-1z"/></svg>
</a>
</div>
<progress class="progress" value="0">
<div class="progress-container">
<span class="progress-bar"></span>
</div>
</progress>
</div>
<!-- /post -->
<!-- The #contentFor helper here will send everything inside it up to the matching #block helper found in default.hbs -->
<!-- Previous/next page links - displayed on every page -->
<!-- The footer at the very bottom of the screen -->
<footer class="site-footer outer">
<div class="site-footer-content inner">
<section class="copyright"><a href="/">Darron's Devlog</a> © 2022</section>
<!--
<section class="poweredby">Proudly published with <a href="https://jekyllrb.com/">Jekyll</a> &
<a href="https://pages.github.com/" target="_blank" rel="noopener">GitHub Pages</a> using
<a href="https://github.com/jekyllt/jasper2" target="_blank" rel="noopener">Jasper2</a></section>
-->
<nav class="site-footer-nav">
<a href="/">Latest Posts</a>
</nav>
</div>
</footer>
</div>
<!-- The big email subscribe modal content -->
<div id="subscribe" class="subscribe-overlay">
<a class="subscribe-overlay-close" href="#"></a>
<div class="subscribe-overlay-content">
<h1 class="subscribe-overlay-title">Search Darron's Devlog</h1>
<p class="subscribe-overlay-description">
</p>
<span id="searchform" method="post" action="/subscribe/" class="">
<input class="confirm" type="hidden" name="confirm" />
<input class="location" type="hidden" name="location" />
<input class="referrer" type="hidden" name="referrer" />
<div class="form-group">
<input class="subscribe-email" onkeyup="myFunc()"
id="searchtext" type="text" name="searchtext"
placeholder="Search..." />
</div>
<script type="text/javascript">
function myFunc() {
if(event.keyCode == 13) {
var url = encodeURIComponent($("#searchtext").val());
location.href = "/search.html?query=" + url;
}
}
</script>
</span>
</div>
</div>
<!-- highlight.js -->
<script src="https://cdnjs.cloudflare.com/ajax/libs/prism/1.10.0/components/prism-abap.min.js"></script>
<script>$(document).ready(function() {
$('pre code').each(function(i, block) {
hljs.highlightBlock(block);
});
});</script>
<!-- jQuery + Fitvids, which makes all video embeds responsive -->
<script
src="https://code.jquery.com/jquery-3.2.1.min.js"
integrity="sha256-hwg4gsxgFZhOsEEamdOYGBf13FyQuiTwlAQgxVSNgt4="
crossorigin="anonymous">
</script>
<script type="text/javascript" src="/assets/js/jquery.fitvids.js"></script>
<script type="text/javascript" src="https://demo.ghost.io/assets/js/jquery.fitvids.js?v=724281a32e"></script>
<!-- Paginator increased to "infinit" in _config.yml -->
<!-- if paginator.posts -->
<!-- <script>
var maxPages = parseInt('');
</script>
<script src="/assets/js/infinitescroll.js"></script> -->
<!-- /endif -->
<!-- Add Google Analytics -->
<!-- Google Analytics Tracking code -->
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','//www.google-analytics.com/analytics.js','ga');
ga('create', '', 'auto');
ga('send', 'pageview');
</script>
<!-- The #block helper will pull in data from the #contentFor other template files. In this case, there's some JavaScript which we only want to use in post.hbs, but it needs to be included down here, after jQuery has already loaded. -->
<script>
// NOTE: Scroll performance is poor in Safari
// - this appears to be due to the events firing much more slowly in Safari.
// Dropping the scroll event and using only a raf loop results in smoother
// scrolling but continuous processing even when not scrolling
$(document).ready(function () {
// Start fitVids
var $postContent = $(".post-full-content");
$postContent.fitVids();
// End fitVids