Repository navigation
Expand file tree
/
Copy pathbert-to-lstm.html
More file actions
1239 lines (949 loc) · 48.6 KB
/
Copy pathbert-to-lstm.html
File metadata and controls
1239 lines (949 loc) · 48.6 KB
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 lang="zh-CN">
<head>
<meta charset="utf-8">
<meta name="keywords" content="Distilling Task-Specific Knowledge from BERT into Simple Neural Networks, AI,NLP,LLM,深度学习,预训练模型,知识蒸馏,模型微调,工程实践">
<meta name="baidu-site-verification" content="code-0cpWDYunkM" />
<meta name="google-site-verification" content="ERyD9x2BzhwQPlITH_IY7aZZUGfb1ZWGrZSDraIYlWw" />
<meta name="360-site-verification" content="b7c11a830ef90fd1464ad6206bb7b6e7" />
<meta name="description" content="论文Distilling Task-Specific Knowledge from BERT into Simple Neural Networks
1 介绍2 相关工作模型压缩 一项突出的工作致力于压缩大型神经网络以加速推理.早期的开创性">
<meta http-equiv="X-UA-Compatible" content="IE=edge">
<meta name="viewport" content="width=device-width, initial-scale=1.0, user-scalable=no">
<meta name="renderer" content="webkit|ie-stand|ie-comp">
<meta name="mobile-web-app-capable" content="yes">
<meta name="format-detection" content="telephone=no">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="black-translucent">
<title>Distilling Task-Specific Knowledge from BERT into Simple Neural Networks | iMountTai's Blog</title>
<link rel="icon" type="image/png" href="/favicon.png">
<link rel="stylesheet" type="text/css" href="/libs/awesome/css/font-awesome.min.css">
<link rel="stylesheet" type="text/css" href="/libs/materialize/materialize.min.css">
<link rel="stylesheet" type="text/css" href="/libs/aos/aos.css">
<link rel="stylesheet" type="text/css" href="/libs/animate/animate.min.css">
<link rel="stylesheet" type="text/css" href="/libs/lightGallery/css/lightgallery.min.css">
<link rel="stylesheet" type="text/css" href="/css/matery.css">
<link rel="stylesheet" type="text/css" href="/css/my.css">
<style type="text/css">
</style>
<script src="/libs/jquery/jquery-2.2.0.min.js"></script>
<script src="https://sdk.jinrishici.com/v2/browser/jinrishici.js" charset="utf-8"></script>
<script>
var _hmt = _hmt || [];
(function() {
var hm = document.createElement("script");
hm.src = "https://hm.baidu.com/hm.js?77b89fafcbf8d98b4a20ea9a700a3436";
var s = document.getElementsByTagName("script")[0];
s.parentNode.insertBefore(hm, s);
})();
</script>
<script>
(function(){
var bp = document.createElement('script');
var curProtocol = window.location.protocol.split(':')[0];
if (curProtocol === 'https') {
bp.src = 'https://zz.bdstatic.com/linksubmit/push.js';
}
else {
bp.src = 'http://push.zhanzhang.baidu.com/push.js';
}
var s = document.getElementsByTagName("script")[0];
s.parentNode.insertBefore(bp, s);
})();
</script>
<script>
(function(){
var src = "https://jspassport.ssl.qhimg.com/11.0.1.js?d182b3f28525f2db83acfaaf6e696dba";
document.write('<script src="' + src + '" id="sozz"><\/script>');
})();
</script>
<meta name="baidu-site-verification" content="" />
<style type="text/css" lang="css">
#loading-container{
position: fixed;
top: 0;
left: 0;
min-height: 100vh;
width: 100vw;
z-index: 9999;
display: flex;
flex-direction: column;
justify-content: center;
align-items: center;
background: #FFF;
text-align: center;
/* loaderҳ����ʧ���ý����ķ�ʽ*/
-webkit-transition: opacity 1s ease;
-moz-transition: opacity 1s ease;
-o-transition: opacity 1s ease;
transition: opacity 1s ease;
}
.loading-image{
width: 120px;
height: 50px;
transform: translate(-50%);
}
.loading-image div:nth-child(2) {
-webkit-animation: pacman-balls 1s linear 0s infinite;
animation: pacman-balls 1s linear 0s infinite
}
.loading-image div:nth-child(3) {
-webkit-animation: pacman-balls 1s linear .33s infinite;
animation: pacman-balls 1s linear .33s infinite
}
.loading-image div:nth-child(4) {
-webkit-animation: pacman-balls 1s linear .66s infinite;
animation: pacman-balls 1s linear .66s infinite
}
.loading-image div:nth-child(5) {
-webkit-animation: pacman-balls 1s linear .99s infinite;
animation: pacman-balls 1s linear .99s infinite
}
.loading-image div:first-of-type {
width: 0;
height: 0;
border: 25px solid #49b1f5;
border-right-color: transparent;
border-radius: 25px;
-webkit-animation: rotate_pacman_half_up .5s 0s infinite;
animation: rotate_pacman_half_up .5s 0s infinite;
}
.loading-image div:nth-child(2) {
width: 0;
height: 0;
border: 25px solid #49b1f5;
border-right-color: transparent;
border-radius: 25px;
-webkit-animation: rotate_pacman_half_down .5s 0s infinite;
animation: rotate_pacman_half_down .5s 0s infinite;
margin-top: -50px;
}
@-webkit-keyframes rotate_pacman_half_up {0% {transform: rotate(270deg)}50% {transform: rotate(1turn)}to {transform: rotate(270deg)}}
@keyframes rotate_pacman_half_up {0% {transform: rotate(270deg)}50% {transform: rotate(1turn)}to {transform: rotate(270deg)}}
@-webkit-keyframes rotate_pacman_half_down {0% {transform: rotate(90deg)}50% {transform: rotate(0deg)}to {transform: rotate(90deg)}}
@keyframes rotate_pacman_half_down {0% {transform: rotate(90deg)}50% {transform: rotate(0deg)}to {transform: rotate(90deg)}}
@-webkit-keyframes pacman-balls {75% {opacity: .7}to {transform: translate(-100px, -6.25px)}}
@keyframes pacman-balls {75% {opacity: .7}to {transform: translate(-100px, -6.25px)}}
.loading-image div:nth-child(3),
.loading-image div:nth-child(4),
.loading-image div:nth-child(5),
.loading-image div:nth-child(6){
background-color: #49b1f5;
width: 15px;
height: 15px;
border-radius: 100%;
margin: 2px;
width: 10px;
height: 10px;
position: absolute;
transform: translateY(-6.25px);
top: 25px;
left: 100px;
}
.loading-text{
margin-bottom: 20vh;
text-align: center;
color: #2c3e50;
font-size: 2rem;
box-sizing: border-box;
padding: 0 10px;
text-shadow: 0 2px 10px rgba(0,0,0,0.2);
}
@media only screen and (max-width: 500px) {
.loading-text{
font-size: 1.5rem;
}
}
.fadeout {
opacity: 0;
filter: alpha(opacity=0);
}
/* logo���ֶ��� */
@-webkit-keyframes fadeInDown{0%{opacity:0;-webkit-transform:translate3d(0,-100%,0);transform:translate3d(0,-100%,0)}100%{opacity:1;-webkit-transform:none;transform:none}}
@keyframes fadeInDown{0%{opacity:0;-webkit-transform:translate3d(0,-100%,0);}}
</style>
<script>
(function () {
const loaded = function(){
setTimeout(function(){
const loader = document.getElementById("loading-container");
loader.className="fadeout" ;//ʹ�ý����ķ�������loading page
// document.getElementById("body-wrap").style.display="flex";
setTimeout(function(){
loader.style.display="none";
},1000);
},1000);//ǿ����ʾloading page 1s
};
loaded();
})()
</script><meta name="generator" content="Hexo 6.0.0"><link rel="alternate" href="/atom.xml" title="iMountTai's Blog" type="application/atom+xml">
<link rel="stylesheet" href="/css/prism-tomorrow.css" type="text/css">
<link rel="stylesheet" href="/css/prism-line-numbers.css" type="text/css"></head>
<div id="loading-container">
<p class="loading-text"></p>
<div class="loading-image">
<div></div>
<div></div>
<div></div>
<div></div>
<div></div>
</div>
</div><body>
<header class="navbar-fixed">
<nav id="headNav" class="bg-color nav-transparent">
<div id="navContainer" class="nav-wrapper container">
<div class="brand-logo">
<a href="/" class="waves-effect waves-light">
<span class="logo-span">iMountTai's Blog</span>
</a>
</div>
<a href="#" data-target="mobile-nav" class="sidenav-trigger button-collapse"><i class="fa fa-navicon"></i></a>
<ul class="right">
<li class="hide-on-med-and-down">
<a href="/" class="waves-effect waves-light">
<span>首页</span>
</a>
</li>
<li class="hide-on-med-and-down">
<a href="/tags" class="waves-effect waves-light">
<span>标签</span>
</a>
</li>
<li class="hide-on-med-and-down">
<a href="/categories" class="waves-effect waves-light">
<span>分类</span>
</a>
</li>
<li class="hide-on-med-and-down">
<a href="/archives" class="waves-effect waves-light">
<span>归档</span>
</a>
</li>
<li class="hide-on-med-and-down">
<a href="/about" class="waves-effect waves-light">
<span>关于</span>
</a>
</li>
<li class="hide-on-med-and-down">
<a href="/contact" class="waves-effect waves-light">
<span>留言板</span>
</a>
</li>
<li>
<a href="#searchModal" class="modal-trigger waves-effect waves-light">
<i id="searchIcon" class="fa fa-search" title="搜索"></i>
</a>
</li>
</ul>
<div id="mobile-nav" class="side-nav sidenav">
<div class="mobile-head bg-color">
<div class="logo-name">iMountTai's Blog</div>
<div class="logo-desc">
记录自然语言处理、预训练模型、大语言模型和工程实践的个人技术博客。
</div>
</div>
<ul class="menu-list mobile-menu-list">
<li>
<a href="/" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
首页
</a>
</li>
<li>
<a href="/tags" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
标签
</a>
</li>
<li>
<a href="/categories" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
分类
</a>
</li>
<li>
<a href="/archives" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
归档
</a>
</li>
<li>
<a href="/about" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
关于
</a>
</li>
<li>
<a href="/contact" class="waves-effect waves-light">
<i class="fa fa-fw fa-link"></i>
留言板
</a>
</li>
<li><div class="divider"></div></li>
<li>
<a href="https://github.com/iMountTai" class="waves-effect waves-light" target="_blank">
<i class="fa fa-github-square fa-fw"></i>Fork Me
</a>
</li>
</ul>
</div>
</div>
<style>
.nav-transparent .github-corner {
display: none !important;
}
.github-corner {
position: absolute;
z-index: 10;
top: 0;
right: 0;
border: 0;
transform: scale(1.1);
}
.github-corner svg {
color: #0f9d58;
fill: #fff;
height: 64px;
width: 64px;
}
.github-corner:hover .octo-arm {
animation: a 0.56s ease-in-out;
}
.github-corner .octo-arm {
animation: none;
}
@keyframes a {
0%,
to {
transform: rotate(0);
}
20%,
60% {
transform: rotate(-25deg);
}
40%,
80% {
transform: rotate(10deg);
}
}
</style>
<a href="https://github.com/iMountTai" class="github-corner tooltipped hide-on-med-and-down" target="_blank"
data-tooltip="Fork Me" data-position="left" data-delay="50">
<svg viewBox="0 0 250 250" aria-hidden="true">
<path d="M0,0 L115,115 L130,115 L142,142 L250,250 L250,0 Z"></path>
<path d="M128.3,109.0 C113.8,99.7 119.0,89.6 119.0,89.6 C122.0,82.7 120.5,78.6 120.5,78.6 C119.2,72.0 123.4,76.3 123.4,76.3 C127.3,80.9 125.5,87.3 125.5,87.3 C122.9,97.6 130.6,101.9 134.4,103.2"
fill="currentColor" style="transform-origin: 130px 106px;" class="octo-arm"></path>
<path d="M115.0,115.0 C114.9,115.1 118.7,116.5 119.8,115.4 L133.7,101.6 C136.9,99.2 139.9,98.4 142.2,98.6 C133.8,88.0 127.5,74.4 143.8,58.0 C148.5,53.4 154.0,51.2 159.7,51.0 C160.3,49.4 163.2,43.6 171.4,40.1 C171.4,40.1 176.1,42.5 178.8,56.2 C183.1,58.6 187.2,61.8 190.9,65.4 C194.5,69.0 197.7,73.2 200.1,77.6 C213.8,80.2 216.3,84.9 216.3,84.9 C212.7,93.1 206.9,96.0 205.4,96.6 C205.1,102.4 203.0,107.8 198.3,112.5 C181.9,128.9 168.3,122.5 157.7,114.1 C157.9,116.9 156.7,120.9 152.7,124.9 L141.0,136.5 C139.8,137.7 141.6,141.9 141.8,141.8 Z"
fill="currentColor" class="octo-body"></path>
</svg>
</a>
</nav>
</header>
<script src="/libs/cryptojs/crypto-js.min.js"></script>
<script>
(function() {
let pwd = '';
if (pwd && pwd.length > 0) {
if (pwd !== CryptoJS.SHA256(prompt('请输入访问本文章的密码')).toString(CryptoJS.enc.Hex)) {
alert('密码错误，将返回主页！');
location.href = '/';
}
}
})();
</script>
<div class="bg-cover pd-header post-cover" style="background-image: url('/medias/featureimages/8.jpg')">
<div class="container">
<div class="row">
<div class="col s12 m12 l12">
<div class="brand">
<div class="description center-align post-title">
Distilling Task-Specific Knowledge from BERT into Simple Neural Networks
</div>
</div>
</div>
</div>
</div>
</div>
<main class="post-container content">
<link rel="stylesheet" href="/libs/tocbot/tocbot.css">
<style>
#articleContent h1::before,
#articleContent h2::before,
#articleContent h3::before,
#articleContent h4::before,
#articleContent h5::before,
#articleContent h6::before {
display: block;
content: " ";
height: 100px;
margin-top: -100px;
visibility: hidden;
}
#articleContent :focus {
outline: none;
}
.toc-fixed {
position: fixed;
top: 64px;
}
.toc-widget {
padding-left: 20px;
}
.toc-widget .toc-title {
margin: 35px 0 15px 0;
padding-left: 17px;
font-size: 1.5rem;
font-weight: bold;
line-height: 1.5rem;
}
.toc-widget ol {
padding: 0;
list-style: none;
}
#toc-content ol {
padding-left: 10px;
}
#toc-content ol li {
padding-left: 10px;
}
#toc-content .toc-link:hover {
color: #42b983;
font-weight: 700;
text-decoration: underline;
}
#toc-content .toc-link::before {
background-color: transparent;
max-height: 25px;
}
#toc-content .is-active-link {
color: #42b983;
}
#toc-content .is-active-link::before {
background-color: #42b983;
}
#floating-toc-btn {
position: fixed;
right: 20px;
bottom: 76px;
padding-top: 15px;
margin-bottom: 0;
z-index: 998;
}
#floating-toc-btn .btn-floating {
width: 48px;
height: 48px;
}
#floating-toc-btn .btn-floating i {
line-height: 48px;
font-size: 1.4rem;
}
</style>
<div class="row">
<div id="main-content" class="col s12 m12 l9">
<!-- 文章内容详情 -->
<div id="artDetail">
<div class="card">
<div class="card-content article-info">
<div class="row tag-cate">
<div class="col s7">
<div class="article-tag">
<a href="/tags/%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F/" target="_blank">
<span class="chip bg-color">知识蒸馏</span>
</a>
<a href="/tags/BERT/" target="_blank">
<span class="chip bg-color">BERT</span>
</a>
<a href="/tags/KD/" target="_blank">
<span class="chip bg-color">KD</span>
</a>
</div>
</div>
<div class="col s5 right-align">
<div class="post-cate">
<i class="fa fa-bookmark fa-fw icon-category"></i>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/" class="post-category" target="_blank">
人工智能
</a>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/NLP/" class="post-category" target="_blank">
NLP
</a>
</div>
</div>
</div>
<div class="post-info">
<div class="post-date info-break-policy">
<i class="fa fa-calendar-minus-o fa-fw"></i>发布日期:
2022-07-26
</div>
<div class="post-author info-break-policy">
<i class="fa fa-user-o fa-fw"></i>作者:
iMountTai
</div>
<div class="info-break-policy">
<i class="fa fa-file-word-o fa-fw"></i>文章字数:
1.5k
</div>
<div class="info-break-policy">
<i class="fa fa-clock-o fa-fw"></i>阅读时长:
5 分
</div>
</div>
</div>
<hr class="clearfix">
<div class="card-content article-card-content">
<div id="articleContent">
<p>论文<br><a target="_blank" rel="noopener" href="https://arxiv.org/pdf/1903.12136.pdf">Distilling Task-Specific Knowledge from BERT into Simple Neural Networks</a></p>
<h1 id="1-介绍"><a href="#1-介绍" class="headerlink" title="1 介绍"></a>1 介绍</h1><h1 id="2-相关工作"><a href="#2-相关工作" class="headerlink" title="2 相关工作"></a>2 相关工作</h1><p><strong>模型压缩</strong> 一项突出的工作致力于压缩大型神经网络以加速推理.早期的开创性工作包括LeCun等人(1990)，他们提出了一种基于局部误差的方法来修剪不重要的权重。最近，Han等人(2015)提出了一种简单的压缩pipeline，在不损害精度的情况下，实现了模型尺寸的40倍缩小。不幸的是，这些技术导致不规则的权重稀疏，这排除了高度优化的计算例程。</p>
<h1 id="3-文章的工作"><a href="#3-文章的工作" class="headerlink" title="3 文章的工作"></a>3 文章的工作</h1><h2 id="3-1-模型结构"><a href="#3-1-模型结构" class="headerlink" title="3.1 模型结构"></a>3.1 模型结构</h2><p>教师模型采用预训练微调的BERT，对于输入的句子对，通过BERT计算一个特征向量$h$,在$h$的基础上构建任务分类器。对于单一句子分类，直接构建softmax层，预测概率为$y^{B}=softmax(Wh)$,$W \in \Bbb{R}^{k \times d}$是softmax的权重矩阵,$k$为标签数量。对于句子对任务，拼接两个句子的BERT特征向量送进softmax层。在训练过程中，我们利用交叉熵损失最大化正确标签的概率，联合微调BERT和分类器的参数</p>
<p>相比之下，我们的学生模型是单层的具有非线性分类器的BiLSTM。将输入词嵌入输入BiLSTM后，将每个方向上最后一步的隐藏状态拼接，输入带有ReLUs的全连接层，该层的输出再传递到softmax层进行分类.</p>
<img src="/bert-to-lstm/figure1.jpg" class>
<p>对于句子对任务，在两个句子编码（$h_{s1}$和$h_{s2}$）的连体结构中共享BiLSTM的编码器，如下图所示</p>
<img src="/bert-to-lstm/f2.jpg" class>
<p>然后对两个句子向量应用标准的concatenate–compare操作$f(h_{s1},h_{s2})=[h_{s1},h_{s2},h_{s1} \odot h_{s2},|h_{s1}-h_{s2}|]$，$\odot$表示按元素乘法。然后把该输出送到ReLU激活的分类器中。</p>
<p>应该强调的是，为了重新审视BiLSTM本身的表示能力，约束了工程实现。我们避免了任何额外的技巧，比如注意和层归一化。</p>
<h2 id="3-2-蒸馏目的"><a href="#3-2-蒸馏目的" class="headerlink" title="3.2 蒸馏目的"></a>3.2 蒸馏目的</h2><p>给定数据，学生模型模仿教师模型的输出。除了最后的标签，教师模型的预测概率同样重要。在二分类情感分析中，一些句子具有较强的情感属性，一些句子则是中性的。如果只用最后的预测标签训练学生模型，会丢失不确定性预测的有价值信息。softmax的参数认为是logits。学生模型在logits上训练使得训练更容易，因为教师模型在所有目标上学到的关系是同等重要的。</p>
<p>使用MSE损失实现$\mathcal{L}_{distill} = ||z^{(B)}-z^{(S)}||_2^2$，$z^{(B)}$和$z^{(S)}$分别是教师模型和学生的logits.交叉熵损失其实也行，但是本文实验证明MSE更好一点。</p>
<p>训练损失：<br>$\mathcal{L}=\alpha \cdot \mathcal{L}_{CE}+(1-\alpha) \cdot \mathcal{L}_{distill} = -\alpha \sum_{i} t_i \log {y_i^{(S)}} - (1 - \alpha)||z^{(B)}-z^{(S)}||_2^2$<br>其中$t$是ground_truth_label。对于没有标签的数据，预测label使用教师模型预测的label，即$t_i=1$如果$i=\operatorname{argmax} {y^{(B)}}$否则为0</p>
<h2 id="3-3-数据增强"><a href="#3-3-数据增强" class="headerlink" title="3.3 数据增强"></a>3.3 数据增强</h2><p>利用教师提供的伪标签（pseudo-labels），用一个大的、未标记的数据集来增强训练集，以帮助有效地提取知识。具体来说，随机执行以下操作：<br><strong>masking</strong>：给定一个概率$p_{mask}$，随机用toekn <em>[MASK]</em> 代替,对应于我们的模型中的未知标记和BERT中的掩码字标记.文章认为该工作能阐明每个词对标签的贡献。<br><strong>POS-guided word replacement</strong>，给定一个概率$p_{post}$,将该单词替换为相同词性（POS）的标签。为了保证原始训练分布，新单词从由词性 (POS) 标签重新归一化的一元词分布中采样。例如将 “What do pigs eat” 替换成 “How do pigs eat”。<br><strong>n-gram sampling</strong>：给定一个概率$p_{ng}$,在 1-5 中随机采样得到 n，然后采用 n-gram，去掉其他单词。这个规则在概念上等同于去掉示例中的所有其他单词，这是一种更激进的屏蔽形式。<br>流程如下：给定一个训练例子{$w_1,…w_n$},迭代单词$w_i$制uniform distribution,如果$x_i<p_{mask}$，对$w_i$进行masking操作，如果$p_{mask} \le x_i<p_{mask}+p_{pos}$，则进行POS-guided word replacement，这两种操作是相互排斥的，即一旦应用了一条规则，另一条就会被忽略。遍历单词之后，在整个合成的例子中采用n-gram采样。这个最后合成的例子加入到增强的无监督数据集中。<br>对单一样本进行n次迭代处理生成n个采样结果。对于句子对任务，我们通过只增强第一个句子(保持第二个固定值)、只增强第二个句子(保持第一个固定值)和两个句子进行循环。</p>
<h1 id="4-实验设置"><a href="#4-实验设置" class="headerlink" title="4 实验设置"></a>4 实验设置</h1><p>使用$BERT_{LARGE}$作为教师模型，加载预训练权重并遵循原始、特定任务的微调。文章微调是，选择用Adam优化器并设置学习率为{2，3，4，5}*$10^{-5}$,然后选择验证集上效果最好的模型。<br>学生模型，将原始数据及合成数据送给教师模型预测获得预测标签（soft logit targets），用以蒸馏$BiLSTM_{SOFT}$，即设置$\alpha = 0$，这种效果是最好的。</p>
<h2 id="4-1-数据集"><a href="#4-1-数据集" class="headerlink" title="4.1 数据集"></a>4.1 数据集</h2><p>SST-2,MNLI,QQP</p>
<h2 id="4-2-超参"><a href="#4-2-超参" class="headerlink" title="4.2 超参"></a>4.2 超参</h2><p>看论文吧</p>
<h1 id="5-结论"><a href="#5-结论" class="headerlink" title="5 结论"></a>5 结论</h1><p>效果比不上大模型，但是比原始的LSTM好，相对大模型就是参数少，响应快。文章认为较浅的bilstm对自然语言任务的表达能力比之前认为的更强</p>
</div>
<hr />
<link rel="stylesheet" type="text/css" href="/libs/share/css/share.min.css">
<div id="article-share">
<div class="social-share" data-disabled="qzone" data-wechat-qrcode-helper="<p>微信里点“发现”->“扫一扫”二维码便可查看分享。</p>"></div>
</div>
<script src="/libs/share/js/social-share.min.js"></script>
<div class="reprint" id="reprint-statement">
<p class="reprint-tip">
<i class="fa fa-exclamation-triangle"></i>
<span>转载规则</span>
</p>
<div class="center-align">
<a rel="license noopener" target="_blank" href="https://creativecommons.org/licenses/by/4.0/deed.zh">
<img alt=""
style="border-width:0"
src="https://i.creativecommons.org/l/by/4.0/88x31.png"/>
</a>
</div>
<br/>
<span xmlns:dct="http://purl.org/dc/terms/" href="http://purl.org/dc/dcmitype/Text"
property="dct:title" rel="dct:type">
《Distilling Task-Specific Knowledge from BERT into Simple Neural Networks》
</span> 由
<a xmlns:cc="http://creativecommons.org/ns#" href="/bert-to-lstm.html" property="cc:attributionName"
rel="cc:attributionURL">
iMountTai
</a> 采用
<a rel="license noopener" target="_blank" href="https://creativecommons.org/licenses/by/4.0/deed.zh">
知识共享署名 4.0 国际许可协议
</a>进行许可。
</div>
<script async defer>
document.addEventListener("copy", function (e) {
let toastHTML = '<span>复制成功，请遵循本文的转载规则</span><button class="btn-flat toast-action" onclick="navToReprintStatement()" style="font-size: smaller">查看</a>';
M.toast({html: toastHTML})
});
function navToReprintStatement() {
$("html, body").animate({scrollTop: $("#reprint-statement").offset().top - 80}, 800);
}
</script>
</div>
</div>
<article id="prenext-posts" class="prev-next articles">
<div class="row article-row">
<div class="article col s12 m6" data-aos="fade-up">
<div class="article-badge left-badge text-color">
<i class="fa fa-chevron-left"></i> 上一篇</div>
<div class="card">
<a href="/mobilebert.html">
<div class="card-image">
<img src="/medias/featureimages/13.jpg" class="responsive-img" alt="MobileBERT a Compact Task-Agnostic BERT for Resource-Limited Devices">
<span class="card-title">MobileBERT a Compact Task-Agnostic BERT for Resource-Limited Devices</span>
</div>
</a>
<div class="card-content article-content">
<div class="summary block-with-text">
论文MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices
introductionMobileBERT被设计成和$BERT_{large}$一样深，而每一
</div>
<div class="publish-info">
<span class="publish-date">
<i class="fa fa-clock-o fa-fw icon-date"></i>2022-08-07
</span>
<span class="publish-author">
<i class="fa fa-bookmark fa-fw icon-category"></i>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/" class="post-category" target="_blank">
人工智能
</a>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/NLP/" class="post-category" target="_blank">
NLP
</a>
</span>
</div>
</div>
<div class="card-action article-tags">
<a href="/tags/%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F/" target="_blank">
<span class="chip bg-color">知识蒸馏</span>
</a>
<a href="/tags/BERT/" target="_blank">
<span class="chip bg-color">BERT</span>
</a>
<a href="/tags/KD/" target="_blank">
<span class="chip bg-color">KD</span>
</a>
</div>
</div>
</div>
<div class="article col s12 m6" data-aos="fade-up">
<div class="article-badge right-badge text-color">
本篇 <i class="fa fa-dot-circle-o"></i>
</div>
<div class="card">
<a href="/bert-to-lstm.html">
<div class="card-image">
<img src="/medias/featureimages/8.jpg" class="responsive-img" alt="Distilling Task-Specific Knowledge from BERT into Simple Neural Networks">
<span class="card-title">Distilling Task-Specific Knowledge from BERT into Simple Neural Networks</span>
</div>
</a>
<div class="card-content article-content">
<div class="summary block-with-text">
论文Distilling Task-Specific Knowledge from BERT into Simple Neural Networks
1 介绍2 相关工作模型压缩 一项突出的工作致力于压缩大型神经网络以加速推理.早期的开创性
</div>
<div class="publish-info">
<span class="publish-date">
<i class="fa fa-clock-o fa-fw icon-date"></i>2022-07-26
</span>
<span class="publish-author">
<i class="fa fa-bookmark fa-fw icon-category"></i>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/" class="post-category" target="_blank">
人工智能
</a>
<a href="/categories/%E4%BA%BA%E5%B7%A5%E6%99%BA%E8%83%BD/NLP/" class="post-category" target="_blank">
NLP
</a>
</span>
</div>
</div>
<div class="card-action article-tags">
<a href="/tags/%E7%9F%A5%E8%AF%86%E8%92%B8%E9%A6%8F/" target="_blank">
<span class="chip bg-color">知识蒸馏</span>
</a>
<a href="/tags/BERT/" target="_blank">
<span class="chip bg-color">BERT</span>
</a>
<a href="/tags/KD/" target="_blank">
<span class="chip bg-color">KD</span>
</a>
</div>
</div>
</div>
</div>
</article>
</div>
<script>
$('#articleContent').on('copy', function (e) {
// IE8 or earlier browser is 'undefined'
if (typeof window.getSelection === 'undefined') return;
var selection = window.getSelection();
// if the selection is short let's not annoy our users.
if (('' + selection).length < Number.parseInt('120')) {
return;
}
// create a div outside of the visible area and fill it with the selected text.
var bodyElement = document.getElementsByTagName('body')[0];
var newdiv = document.createElement('div');
newdiv.style.position = 'absolute';
newdiv.style.left = '-99999px';
bodyElement.appendChild(newdiv);
newdiv.appendChild(selection.getRangeAt(0).cloneContents());
// we need a <pre> tag workaround.
// otherwise the text inside "pre" loses all the line breaks!
if (selection.getRangeAt(0).commonAncestorContainer.nodeName === 'PRE') {
newdiv.innerHTML = "<pre>" + newdiv.innerHTML + "</pre>";
}
var url = document.location.href;
newdiv.innerHTML += '<br />'
+ '来源: iMountTai's Blog<br />'
+ '作者: iMountTai<br />'
+ '链接: <a href="' + url + '">' + url + '</a><br />'
+ '本文章著作权归作者所有，任何形式的转载都请注明出处。';
selection.selectAllChildren(newdiv);
window.setTimeout(function () { bodyElement.removeChild(newdiv); }, 200);
});
</script>
</div>
<div id="toc-aside" class="expanded col l3 hide-on-med-and-down">
<div class="toc-widget">
<div class="toc-title"><i class="fa fa-list-alt"></i> 目录</div>
<div id="toc-content"></div>
</div>
</div>
</div>
<!-- TOC 悬浮按钮. -->
<div id="floating-toc-btn" class="hide-on-med-and-down">
<a class="btn-floating btn-large bg-color">
<i class="fa fa-list"></i>
</a>
</div>
<script src="/libs/tocbot/tocbot.min.js"></script>
<script>
$(function () {
tocbot.init({
tocSelector: '#toc-content',
contentSelector: '#articleContent',
headingsOffset: -($(window).height() * 0.4 - 45),
// headingsOffset: -205,
headingSelector: 'h1, h2, h3, h4'
});
// modify the toc link href to support Chinese.
let i = 0;
let tocHeading = 'toc-heading-';
$('#toc-content a').each(function () {
$(this).attr('href', '#' + tocHeading + (++i));
});
// modify the heading title id to support Chinese.
i = 0;
$('#articleContent').children('h1, h2, h3, h4').each(function () {
$(this).attr('id', tocHeading + (++i));
});
// Set scroll toc fixed.
let tocHeight = parseInt($(window).height() * 0.4 - 64);
let $tocWidget = $('.toc-widget');
$(window).scroll(function () {
let scroll = $(window).scrollTop();
/* add post toc fixed. */
if (scroll > tocHeight) {
$tocWidget.addClass('toc-fixed');
} else {
$tocWidget.removeClass('toc-fixed');
}
});
/* 修复文章卡片 div 的宽度. */
let fixPostCardWidth = function (srcId, targetId) {
let srcDiv = $('#' + srcId);
if (srcDiv.length === 0) {
return;
}
let w = srcDiv.width();
if (w >= 450) {
w = w + 21;
} else if (w >= 350 && w < 450) {
w = w + 18;
} else if (w >= 300 && w < 350) {
w = w + 16;
} else {
w = w + 14;
}
$('#' + targetId).width(w);
};
// 切换TOC目录展开收缩的相关操作.
const expandedClass = 'expanded';
let $tocAside = $('#toc-aside');
let $mainContent = $('#main-content');
$('#floating-toc-btn .btn-floating').click(function () {
if ($tocAside.hasClass(expandedClass)) {
$tocAside.removeClass(expandedClass).slideUp(500);
$mainContent.removeClass('l9');
} else {
$tocAside.addClass(expandedClass).slideDown(500);
$mainContent.addClass('l9');
}
fixPostCardWidth('artDetail', 'prenext-posts');
});
});
</script>
</main>
<script src="//cdn.jsdelivr.net/npm/mathjax@2/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
<script>
MathJax.Hub.Config({
tex2jax: {inlineMath: [['$', '$'], ['\(', '\)']]}
});
</script>