-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathquantized_transfer_learning_tutorial.html
1128 lines (911 loc) · 91.3 KB
/
quantized_transfer_learning_tutorial.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>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning) — PyTorch Tutorials 1.10.2+cu102 documentation</title>
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/gallery.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.13.11/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="(beta) Static Quantization with Eager Mode in PyTorch" href="../advanced/static_quantization_tutorial.html" />
<link rel="prev" title="(베타) BERT 모델 동적 양자화하기" href="dynamic_quantization_bert_tutorial.html" />
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.kr/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.kr/get-started">시작하기</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr">튜토리얼</a>
</li>
<li>
<a href="https://pytorch.kr/hub">허브</a>
</li>
<li>
<a href="https://discuss.pytorch.kr">커뮤니티</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
1.10.2+cu102
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Tutorials" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 레시피</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../recipes/recipes_index.html">모든 레시피 보기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../prototype/prototype_index.html">모든 프로토타입 레시피 보기</a></li>
</ul>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 시작하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/intro.html">파이토치(PyTorch) 기본 익히기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/quickstart_tutorial.html">빠른 시작(Quickstart)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/tensorqs_tutorial.html">텐서(Tensor)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/data_tutorial.html">Dataset과 DataLoader</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/transforms_tutorial.html">변형(Transform)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/buildmodel_tutorial.html">신경망 모델 구성하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/autogradqs_tutorial.html"><code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code>를 사용한 자동 미분</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/optimization_tutorial.html">모델 매개변수 최적화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/saveloadrun_tutorial.html">모델 저장하고 불러오기</a></li>
</ul>
<p class="caption"><span class="caption-text">Introduction to PyTorch on YouTube</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt.html">Introduction to PyTorch - YouTube Series</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">Introduction to PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Introduction to PyTorch Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/autogradyt_tutorial.html">The Fundamentals of Autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/modelsyt_tutorial.html">Building Models with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensorboardyt_tutorial.html">PyTorch TensorBoard Support</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/trainingyt.html">Training with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/captumyt.html">Model Understanding with Captum</a></li>
</ul>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 배우기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deep_learning_60min_blitz.html">PyTorch로 딥러닝하기: 60분만에 끝장내기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/pytorch_with_examples.html">예제로 배우는 파이토치(PyTorch)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/nn_tutorial.html"><cite>torch.nn</cite> 이 <em>실제로</em> 무엇인가요?</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_tutorial.html">TensorBoard로 모델, 데이터, 학습 시각화하기</a></li>
</ul>
<p class="caption"><span class="caption-text">이미지/비디오</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision_tutorial.html">TorchVision 객체 검출 미세조정(Finetuning) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transfer_learning_tutorial.html">컴퓨터 비전(Vision)을 위한 전이학습(Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/fgsm_tutorial.html">적대적 예제 생성(Adversarial Example Generation)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dcgan_faces_tutorial.html">DCGAN 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">배포를 위한 비전 트랜스포머(Vision Transformer) 모델 최적화하기</a></li>
</ul>
<p class="caption"><span class="caption-text">오디오</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_io_tutorial.html">Audio I/O</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_resampling_tutorial.html">Audio Resampling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_data_augmentation_tutorial.html">Audio Data Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_extractions_tutorial.html">Audio Feature Extractions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_augmentation_tutorial.html">Audio Feature Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_datasets_tutorial.html">Audio Datasets</a></li>
<li class="toctree-l1"><a class="reference internal" href="speech_recognition_pipeline_tutorial.html">Speech Recognition with Wav2Vec2</a></li>
<li class="toctree-l1"><a class="reference internal" href="speech_command_classification_with_torchaudio_tutorial.html">Speech Command Classification with torchaudio</a></li>
<li class="toctree-l1"><a class="reference internal" href="text_to_speech_with_torchaudio.html">Text-to-speech with torchaudio</a></li>
<li class="toctree-l1"><a class="reference internal" href="forced_alignment_with_torchaudio_tutorial.html">Forced Alignment with Wav2Vec2</a></li>
</ul>
<p class="caption"><span class="caption-text">텍스트</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transformer_tutorial.html">nn.Transformer 와 TorchText 로 시퀀스-투-시퀀스(Sequence-to-Sequence) 모델링하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_classification_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 분류하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_generation_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 생성하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="seq2seq_translation_tutorial.html">기초부터 시작하는 NLP: Sequence to Sequence 네트워크와 Attention을 이용한 번역</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/text_sentiment_ngrams_tutorial.html">torchtext 라이브러리로 텍스트 분류하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/translation_transformer.html">nn.Transformer와 torchtext로 언어 번역하기</a></li>
</ul>
<p class="caption"><span class="caption-text">강화학습</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_q_learning.html">강화 학습 (DQN) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="mario_rl_tutorial.html">Train a Mario-playing RL Agent</a></li>
</ul>
<p class="caption"><span class="caption-text">PyTorch 모델을 프로덕션 환경에 배포하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="flask_rest_api_tutorial.html">Flask를 사용하여 Python에서 PyTorch를 REST API로 배포하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/Intro_to_TorchScript_tutorial.html">TorchScript 소개</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_export.html">C++에서 TorchScript 모델 로딩하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/super_resolution_with_onnxruntime.html">(선택) PyTorch 모델을 ONNX으로 변환하고 ONNX 런타임에서 실행하기</a></li>
</ul>
<p class="caption"><span class="caption-text">Code Transforms with FX</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="fx_conv_bn_fuser.html">(베타) FX에서 합성곱/배치 정규화(Convolution/Batch Norm) 결합기(Fuser) 만들기</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx_profiling_tutorial.html">(beta) Building a Simple CPU Performance Profiler with FX</a></li>
</ul>
<p class="caption"><span class="caption-text">프론트엔드 API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="memory_format_tutorial.html">(베타) PyTorch를 사용한 Channels Last 메모리 형식</a></li>
<li class="toctree-l1"><a class="reference internal" href="forward_ad_usage.html">Forward-mode Automatic Differentiation (Beta)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_frontend.html">PyTorch C++ 프론트엔드 사용하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch-script-parallelism.html">TorchScript의 동적 병렬 처리(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_autograd.html">C++ 프론트엔드의 자동 미분 (autograd)</a></li>
</ul>
<p class="caption"><span class="caption-text">PyTorch 확장하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="custom_function_double_backward_tutorial.html">Double Backward with Custom Functions</a></li>
<li class="toctree-l1"><a class="reference internal" href="custom_function_conv_bn_tutorial.html">Fusing Convolution and Batch Norm using Custom Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_extension.html">Custom C++ and CUDA Extensions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_ops.html">Extending TorchScript with Custom C++ Operators</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch_script_custom_classes.html">커스텀 C++ 클래스로 TorchScript 확장하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dispatcher.html">Registering a Dispatched Operator in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/extend_dispatcher.html">Extending dispatcher for a new backend in C++</a></li>
</ul>
<p class="caption"><span class="caption-text">모델 최적화</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch 모듈 프로파일링 하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_profiler_tutorial.html">PyTorch Profiler With TensorBoard</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Hyperparameter tuning with Ray Tune</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">배포를 위한 비전 트랜스포머(Vision Transformer) 모델 최적화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="parametrizations.html">Parametrizations Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="pruning_tutorial.html">가지치기 기법(Pruning) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dynamic_quantization_tutorial.html">(베타) LSTM 기반 단어 단위 언어 모델의 동적 양자화</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_bert_tutorial.html">(베타) BERT 모델 동적 양자화하기</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/static_quantization_tutorial.html">(beta) Static Quantization with Eager Mode in PyTorch</a></li>
</ul>
<p class="caption"><span class="caption-text">병렬 및 분산 학습</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_parallel_tutorial.html">단일 머신을 사용한 모델 병렬화 모범 사례</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_tutorial.html">분산 데이터 병렬 처리 시작하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_tuto.html">PyTorch로 분산 어플리케이션 개발하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_tutorial.html">Getting Started with Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_param_server_tutorial.html">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism Using RPC</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_async_execution.html">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/rpc_ddp_tutorial.html">분산 데이터 병렬(DDP)과 분산 RPC 프레임워크 결합</a></li>
<li class="toctree-l1"><a class="reference internal" href="pipeline_tutorial.html">파이프라인 병렬화로 트랜스포머 모델 학습시키기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/ddp_pipeline.html">분산 데이터 병렬 처리와 병렬 처리 파이프라인을 사용한 트랜스포머 모델 학습</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/generic_join.html">Distributed Training with Uneven Inputs Using the Join Context Manager</a></li>
</ul>
<p class="caption"><span class="caption-text">Mobile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deeplabv3_on_ios.html">iOS에서의 이미지 분할 DeepLapV3</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deeplabv3_on_android.html">안드로이드에서의 이미지 분할 DeepLapV3</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="../index.html">
Tutorials
</a> >
</li>
<li>(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning)</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/quantized_transfer_learning_tutorial.rst.txt" rel="nofollow"><img src="../_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="pytorch-call-to-action-links">
<div id="tutorial-type">intermediate/quantized_transfer_learning_tutorial</div>
<div id="google-colab-link">
<img class="call-to-action-img" src="../_static/images/pytorch-colab.svg"/>
<div class="call-to-action-desktop-view">Run in Google Colab</div>
<div class="call-to-action-mobile-view">Colab</div>
</div>
<div id="download-notebook-link">
<img class="call-to-action-notebook-img" src="../_static/images/pytorch-download.svg"/>
<div class="call-to-action-desktop-view">Download Notebook</div>
<div class="call-to-action-mobile-view">Notebook</div>
</div>
<div id="github-view-link">
<img class="call-to-action-img" src="../_static/images/pytorch-github.svg"/>
<div class="call-to-action-desktop-view">View on GitHub</div>
<div class="call-to-action-mobile-view">GitHub</div>
</div>
</div>
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="quantized-transfer-learning">
<h1>(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning)<a class="headerlink" href="#quantized-transfer-learning" title="Permalink to this headline">¶</a></h1>
<div class="admonition tip">
<p class="first admonition-title">Tip</p>
<p class="last">이 튜토리얼을 최대한 활용하시려면, 다음의 링크를 이용하시길 추천합니다.
<a class="reference external" href="https://colab.research.google.com/github/pytorch/tutorials/blob/gh-pages/_downloads/quantized_transfer_learning_tutorial.ipynb">Colab 버전</a>.
이를 통해 아래에 제시된 정보로 실험을 해 볼 수 있습니다.</p>
</div>
<p><strong>Author</strong>: <a class="reference external" href="https://github.com/z-a-f">Zafar Takhirov</a>
<strong>Reviewed by</strong>: <a class="reference external" href="https://github.com/raghuramank100">Raghuraman Krishnamoorthi</a>
<strong>Edited by</strong>: <a class="reference external" href="https://github.com/jlin27">Jessica Lin</a>
<strong>번역</strong>: <a class="reference external" href="https://github.com/jjeamin">정재민</a></p>
<p>이 튜토리얼은 <a class="reference external" href="https://chsasank.github.io/">Sasank Chilamkurthy</a> 가 작성한
<a class="reference internal" href="../beginner/transfer_learning_tutorial.html"><span class="doc">컴퓨터 비전(Vision)을 위한 전이학습(Transfer Learning)</span></a> 을 기반으로 합니다.</p>
<p>전이학습(Transfer learning)은 다른 데이터셋에 적용하기 위해서 미리 학습된 모델을 사용하는 기술을 말합니다.
전이학습을 사용하는 2가지 주요 방법이 있습니다.</p>
<ol class="arabic">
<li><p class="first"><strong>고정 된 특징 추출기로써 ConvNet</strong>: 여기서는 마지막 몇개의 계층(일명 “헤드(the head)”, 일반적으로 완전히 연결된 계층)
을 제외하고 네트워크의 모든 매개 변수 가중치를 <a class="reference external" href="https://arxiv.org/abs/1706.04983">“고정(freeze)”</a> 합니다.
마지막 계층은 임의의 가중치로 초기화된 새로운 계층으로 대체되며 오직 이 계층만 학습됩니다.</p>
</li>
<li><p class="first"><strong>ConvNet 미세조정(Finetuning)</strong>: 랜덤 초기화 대신, 미리 학습된 네트워크를 이용하여 모델을 초기화합니다.
이후 평소처럼 학습을 진행하지만 다른 데이터셋을 사용합니다.
평소처럼 학습이 진행되지만 다른 데이터셋을 사용합니다.</p>
<p>출력의 수가 다를 수 있기 때문에, 일반적으로 신경망에서 헤드(또는 그 일부)는 교체됩니다.
이 방법에서는 학습률을 더 작은 수로 설정하는 것이 일반적입니다.
이는 네트워크가 이미 학습되었기 때문이며 새로운 데이터셋으로 “미세조정(finetuning)”하려면 약간의 변경만이 필요합니다.</p>
</li>
</ol>
<p>또한 위의 두 방법을 결합할 수도 있습니다.
먼저 특징 추출기를 고정(freeze)하고 헤드(the head)를 학습시킵니다.
그런 다음 특징 추출기(또는 그 일부)를 고정해제(unfreeze)하고 학습률을
더 작은 수로 설정한 다음 학습을 계속할 수 있습니다.</p>
<p>이번 파트에서는 첫번째 방법을 사용해 양자화된 모델로 특징을 추출해봅시다.</p>
<div class="section" id="id2">
<h2>파트 0. 요구사항<a class="headerlink" href="#id2" title="Permalink to this headline">¶</a></h2>
<p>전이 학습(transfer learning)을 시작하기 전에,
데이터 불러오기 / 시각화와 같은 “요구사항(prerequisites)”을 검토하겠습니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Imports</span>
<span class="kn">import</span> <span class="nn">copy</span>
<span class="kn">import</span> <span class="nn">matplotlib.pyplot</span> <span class="k">as</span> <span class="nn">plt</span>
<span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
<span class="kn">import</span> <span class="nn">os</span>
<span class="kn">import</span> <span class="nn">time</span>
<span class="n">plt</span><span class="o">.</span><span class="n">ion</span><span class="p">()</span>
</pre></div>
</div>
<div class="section" id="nightly-build">
<h3>Nightly Build 설치하기<a class="headerlink" href="#nightly-build" title="Permalink to this headline">¶</a></h3>
<p>PyTorch의 베타(beta)를 사용할 것이므로 최신 버전의 <code class="docutils literal notranslate"><span class="pre">torch</span></code> 와 <code class="docutils literal notranslate"><span class="pre">torchvision</span></code> 을 설치하는 것을 권장합니다.
로컬(local) 설치에 대한 최신 지침은 <a class="reference external" href="https://pytorch.org/get-started/locally/">여기</a> 에서 찾을 수 있습니다.
예를 들어 GPU 지원 없이 설치하려면 :</p>
<div class="code shell highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">pip</span> <span class="n">install</span> <span class="n">numpy</span>
<span class="n">pip</span> <span class="n">install</span> <span class="o">--</span><span class="n">pre</span> <span class="n">torch</span> <span class="n">torchvision</span> <span class="o">-</span><span class="n">f</span> <span class="n">https</span><span class="p">:</span><span class="o">//</span><span class="n">download</span><span class="o">.</span><span class="n">pytorch</span><span class="o">.</span><span class="n">org</span><span class="o">/</span><span class="n">whl</span><span class="o">/</span><span class="n">nightly</span><span class="o">/</span><span class="n">cpu</span><span class="o">/</span><span class="n">torch_nightly</span><span class="o">.</span><span class="n">html</span>
<span class="c1"># CUDA 지원은 https://download.pytorch.org/whl/nightly/cu101/torch_nightly.html를 사용하세요.</span>
</pre></div>
</div>
</div>
<div class="section" id="id4">
<h3>데이터 불러오기<a class="headerlink" href="#id4" title="Permalink to this headline">¶</a></h3>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">이번 섹션은 원본 전이학습(Transfer Learning) 튜토리얼과 동일합니다.</p>
</div>
<p><code class="docutils literal notranslate"><span class="pre">torchvision</span></code> 과 <code class="docutils literal notranslate"><span class="pre">torch.utils.data</span></code> 패키지를 사용하여 데이터를 불러옵니다.</p>
<p>여기서 풀고자 하는 문제는 이미지로부터 <strong>개미</strong> 와 <strong>벌</strong> 을 분류하는 것입니다.
이 데이터셋은 개미와 벌에 대해 각각 120장의 학습용 이미지, 75개의 검증용 이미지를 포함합니다.
이는 일반화하기에는 아주 작은 데이터셋입니다.
하지만 우리는 전이학습(Transfer Learning)을 사용하기 때문에, 일반화를 꽤 잘 할 수 있을 것입니다.</p>
<p>이 데이터셋은 imagenet의 아주 작은 일부입니다.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last"><a class="reference external" href="https://download.pytorch.org/tutorial/hymenoptera_data.zip">여기</a> 에서 데이터를 다운로드 받아 <code class="docutils literal notranslate"><span class="pre">data</span></code> 디렉토리에 압축을 푸세요.</p>
</div>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">torchvision</span> <span class="kn">import</span> <span class="n">transforms</span><span class="p">,</span> <span class="n">datasets</span>
<span class="c1"># 학습을 위한 데이터 보강(Data augmentation)과 정규화</span>
<span class="c1"># 검증을 위한 정규화</span>
<span class="n">data_transforms</span> <span class="o">=</span> <span class="p">{</span>
<span class="s1">'train'</span><span class="p">:</span> <span class="n">transforms</span><span class="o">.</span><span class="n">Compose</span><span class="p">([</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Resize</span><span class="p">(</span><span class="mi">224</span><span class="p">),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">RandomCrop</span><span class="p">(</span><span class="mi">224</span><span class="p">),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">RandomHorizontalFlip</span><span class="p">(),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">(),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Normalize</span><span class="p">([</span><span class="mf">0.485</span><span class="p">,</span> <span class="mf">0.456</span><span class="p">,</span> <span class="mf">0.406</span><span class="p">],</span> <span class="p">[</span><span class="mf">0.229</span><span class="p">,</span> <span class="mf">0.224</span><span class="p">,</span> <span class="mf">0.225</span><span class="p">])</span>
<span class="p">]),</span>
<span class="s1">'val'</span><span class="p">:</span> <span class="n">transforms</span><span class="o">.</span><span class="n">Compose</span><span class="p">([</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Resize</span><span class="p">(</span><span class="mi">224</span><span class="p">),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">CenterCrop</span><span class="p">(</span><span class="mi">224</span><span class="p">),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">(),</span>
<span class="n">transforms</span><span class="o">.</span><span class="n">Normalize</span><span class="p">([</span><span class="mf">0.485</span><span class="p">,</span> <span class="mf">0.456</span><span class="p">,</span> <span class="mf">0.406</span><span class="p">],</span> <span class="p">[</span><span class="mf">0.229</span><span class="p">,</span> <span class="mf">0.224</span><span class="p">,</span> <span class="mf">0.225</span><span class="p">])</span>
<span class="p">]),</span>
<span class="p">}</span>
<span class="n">data_dir</span> <span class="o">=</span> <span class="s1">'data/hymenoptera_data'</span>
<span class="n">image_datasets</span> <span class="o">=</span> <span class="p">{</span><span class="n">x</span><span class="p">:</span> <span class="n">datasets</span><span class="o">.</span><span class="n">ImageFolder</span><span class="p">(</span><span class="n">os</span><span class="o">.</span><span class="n">path</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">data_dir</span><span class="p">,</span> <span class="n">x</span><span class="p">),</span>
<span class="n">data_transforms</span><span class="p">[</span><span class="n">x</span><span class="p">])</span>
<span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="p">[</span><span class="s1">'train'</span><span class="p">,</span> <span class="s1">'val'</span><span class="p">]}</span>
<span class="n">dataloaders</span> <span class="o">=</span> <span class="p">{</span><span class="n">x</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">image_datasets</span><span class="p">[</span><span class="n">x</span><span class="p">],</span> <span class="n">batch_size</span><span class="o">=</span><span class="mi">16</span><span class="p">,</span>
<span class="n">shuffle</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">8</span><span class="p">)</span>
<span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="p">[</span><span class="s1">'train'</span><span class="p">,</span> <span class="s1">'val'</span><span class="p">]}</span>
<span class="n">dataset_sizes</span> <span class="o">=</span> <span class="p">{</span><span class="n">x</span><span class="p">:</span> <span class="nb">len</span><span class="p">(</span><span class="n">image_datasets</span><span class="p">[</span><span class="n">x</span><span class="p">])</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="p">[</span><span class="s1">'train'</span><span class="p">,</span> <span class="s1">'val'</span><span class="p">]}</span>
<span class="n">class_names</span> <span class="o">=</span> <span class="n">image_datasets</span><span class="p">[</span><span class="s1">'train'</span><span class="p">]</span><span class="o">.</span><span class="n">classes</span>
<span class="n">device</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span><span class="s2">"cuda:0"</span> <span class="k">if</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">is_available</span><span class="p">()</span> <span class="k">else</span> <span class="s2">"cpu"</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="id6">
<h3>일부 이미지 시각화하기<a class="headerlink" href="#id6" title="Permalink to this headline">¶</a></h3>
<p>데이터 보강을 이해하기 위해 일부 학습용 이미지를 시각화 해보겠습니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torchvision</span>
<span class="k">def</span> <span class="nf">imshow</span><span class="p">(</span><span class="n">inp</span><span class="p">,</span> <span class="n">title</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">ax</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">figsize</span><span class="o">=</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">)):</span>
<span class="sd">"""Imshow for Tensor."""</span>
<span class="n">inp</span> <span class="o">=</span> <span class="n">inp</span><span class="o">.</span><span class="n">numpy</span><span class="p">()</span><span class="o">.</span><span class="n">transpose</span><span class="p">((</span><span class="mi">1</span><span class="p">,</span> <span class="mi">2</span><span class="p">,</span> <span class="mi">0</span><span class="p">))</span>
<span class="n">mean</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array</span><span class="p">([</span><span class="mf">0.485</span><span class="p">,</span> <span class="mf">0.456</span><span class="p">,</span> <span class="mf">0.406</span><span class="p">])</span>
<span class="n">std</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">array</span><span class="p">([</span><span class="mf">0.229</span><span class="p">,</span> <span class="mf">0.224</span><span class="p">,</span> <span class="mf">0.225</span><span class="p">])</span>
<span class="n">inp</span> <span class="o">=</span> <span class="n">std</span> <span class="o">*</span> <span class="n">inp</span> <span class="o">+</span> <span class="n">mean</span>
<span class="n">inp</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">clip</span><span class="p">(</span><span class="n">inp</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">if</span> <span class="n">ax</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">fig</span><span class="p">,</span> <span class="n">ax</span> <span class="o">=</span> <span class="n">plt</span><span class="o">.</span><span class="n">subplots</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">figsize</span><span class="o">=</span><span class="n">figsize</span><span class="p">)</span>
<span class="n">ax</span><span class="o">.</span><span class="n">imshow</span><span class="p">(</span><span class="n">inp</span><span class="p">)</span>
<span class="n">ax</span><span class="o">.</span><span class="n">set_xticks</span><span class="p">([])</span>
<span class="n">ax</span><span class="o">.</span><span class="n">set_yticks</span><span class="p">([])</span>
<span class="k">if</span> <span class="n">title</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">ax</span><span class="o">.</span><span class="n">set_title</span><span class="p">(</span><span class="n">title</span><span class="p">)</span>
<span class="c1"># 학습 데이터의 배치를 얻습니다.</span>
<span class="n">inputs</span><span class="p">,</span> <span class="n">classes</span> <span class="o">=</span> <span class="nb">next</span><span class="p">(</span><span class="nb">iter</span><span class="p">(</span><span class="n">dataloaders</span><span class="p">[</span><span class="s1">'train'</span><span class="p">]))</span>
<span class="c1"># 배치로부터 격자 형태의 이미지를 만듭니다.</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">torchvision</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">make_grid</span><span class="p">(</span><span class="n">inputs</span><span class="p">,</span> <span class="n">nrow</span><span class="o">=</span><span class="mi">4</span><span class="p">)</span>
<span class="n">fig</span><span class="p">,</span> <span class="n">ax</span> <span class="o">=</span> <span class="n">plt</span><span class="o">.</span><span class="n">subplots</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">figsize</span><span class="o">=</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">10</span><span class="p">))</span>
<span class="n">imshow</span><span class="p">(</span><span class="n">out</span><span class="p">,</span> <span class="n">title</span><span class="o">=</span><span class="p">[</span><span class="n">class_names</span><span class="p">[</span><span class="n">x</span><span class="p">]</span> <span class="k">for</span> <span class="n">x</span> <span class="ow">in</span> <span class="n">classes</span><span class="p">],</span> <span class="n">ax</span><span class="o">=</span><span class="n">ax</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="id7">
<h3>모델 학습을 위한 지원 함수<a class="headerlink" href="#id7" title="Permalink to this headline">¶</a></h3>
<p>다음은 모델을 학습하기 위한 일반 함수 입니다.</p>
<ul class="simple">
<li>학습률(learning rate)을 관리합니다(schedules).</li>
<li>최적의 모델을 저장합니다.</li>
</ul>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">train_model</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">criterion</span><span class="p">,</span> <span class="n">optimizer</span><span class="p">,</span> <span class="n">scheduler</span><span class="p">,</span> <span class="n">num_epochs</span><span class="o">=</span><span class="mi">25</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s1">'cpu'</span><span class="p">):</span>
<span class="sd">"""</span>
<span class="sd"> Support function for model training.</span>
<span class="sd"> 모델 학습을 위한 지원 함수</span>
<span class="sd"> 매개변수:</span>
<span class="sd"> model: 학습할 모델</span>
<span class="sd"> criterion: 최적화 기준(손실)</span>
<span class="sd"> optimizer: 학습에 사용할 옵티마이저</span>
<span class="sd"> scheduler: ``torch.optim.lr_scheduler`` 의 인스턴스</span>
<span class="sd"> num_epochs: 에폭의 수</span>
<span class="sd"> device: 학습을 동작시킬 장치. 'cpu' 또는 'cuda'여야 합니다.</span>
<span class="sd"> """</span>
<span class="n">since</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
<span class="n">best_model_wts</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">state_dict</span><span class="p">())</span>
<span class="n">best_acc</span> <span class="o">=</span> <span class="mf">0.0</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">num_epochs</span><span class="p">):</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'Epoch </span><span class="si">{}</span><span class="s1">/</span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">epoch</span><span class="p">,</span> <span class="n">num_epochs</span> <span class="o">-</span> <span class="mi">1</span><span class="p">))</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'-'</span> <span class="o">*</span> <span class="mi">10</span><span class="p">)</span>
<span class="c1"># 각 에폭에는 학습 및 검증 단계가 있습니다.</span>
<span class="k">for</span> <span class="n">phase</span> <span class="ow">in</span> <span class="p">[</span><span class="s1">'train'</span><span class="p">,</span> <span class="s1">'val'</span><span class="p">]:</span>
<span class="k">if</span> <span class="n">phase</span> <span class="o">==</span> <span class="s1">'train'</span><span class="p">:</span>
<span class="n">model</span><span class="o">.</span><span class="n">train</span><span class="p">()</span> <span class="c1"># 모델을 학습 모드로 설정하기</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">model</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span> <span class="c1"># 모델을 평가 모드로 설정하기</span>
<span class="n">running_loss</span> <span class="o">=</span> <span class="mf">0.0</span>
<span class="n">running_corrects</span> <span class="o">=</span> <span class="mi">0</span>
<span class="c1"># 데이터 반복하기</span>
<span class="k">for</span> <span class="n">inputs</span><span class="p">,</span> <span class="n">labels</span> <span class="ow">in</span> <span class="n">dataloaders</span><span class="p">[</span><span class="n">phase</span><span class="p">]:</span>
<span class="n">inputs</span> <span class="o">=</span> <span class="n">inputs</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">labels</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span>
<span class="c1"># 매개 변수 기울기를 0으로 설정하기</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="c1"># 순전파</span>
<span class="c1"># 학습 동안만 연산 기록을 추적하기</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">set_grad_enabled</span><span class="p">(</span><span class="n">phase</span> <span class="o">==</span> <span class="s1">'train'</span><span class="p">):</span>
<span class="n">outputs</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="n">inputs</span><span class="p">)</span>
<span class="n">_</span><span class="p">,</span> <span class="n">preds</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">criterion</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="n">labels</span><span class="p">)</span>
<span class="c1"># 역전파 + 학습 단계에서만 최적화</span>
<span class="k">if</span> <span class="n">phase</span> <span class="o">==</span> <span class="s1">'train'</span><span class="p">:</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="c1"># 통계 보기</span>
<span class="n">running_loss</span> <span class="o">+=</span> <span class="n">loss</span><span class="o">.</span><span class="n">item</span><span class="p">()</span> <span class="o">*</span> <span class="n">inputs</span><span class="o">.</span><span class="n">size</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="n">running_corrects</span> <span class="o">+=</span> <span class="n">torch</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">preds</span> <span class="o">==</span> <span class="n">labels</span><span class="o">.</span><span class="n">data</span><span class="p">)</span>
<span class="k">if</span> <span class="n">phase</span> <span class="o">==</span> <span class="s1">'train'</span><span class="p">:</span>
<span class="n">scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="n">epoch_loss</span> <span class="o">=</span> <span class="n">running_loss</span> <span class="o">/</span> <span class="n">dataset_sizes</span><span class="p">[</span><span class="n">phase</span><span class="p">]</span>
<span class="n">epoch_acc</span> <span class="o">=</span> <span class="n">running_corrects</span><span class="o">.</span><span class="n">double</span><span class="p">()</span> <span class="o">/</span> <span class="n">dataset_sizes</span><span class="p">[</span><span class="n">phase</span><span class="p">]</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'</span><span class="si">{}</span><span class="s1"> Loss: </span><span class="si">{:.4f}</span><span class="s1"> Acc: </span><span class="si">{:.4f}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span>
<span class="n">phase</span><span class="p">,</span> <span class="n">epoch_loss</span><span class="p">,</span> <span class="n">epoch_acc</span><span class="p">))</span>
<span class="c1"># 모델 복사하기</span>
<span class="k">if</span> <span class="n">phase</span> <span class="o">==</span> <span class="s1">'val'</span> <span class="ow">and</span> <span class="n">epoch_acc</span> <span class="o">></span> <span class="n">best_acc</span><span class="p">:</span>
<span class="n">best_acc</span> <span class="o">=</span> <span class="n">epoch_acc</span>
<span class="n">best_model_wts</span> <span class="o">=</span> <span class="n">copy</span><span class="o">.</span><span class="n">deepcopy</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">state_dict</span><span class="p">())</span>
<span class="nb">print</span><span class="p">()</span>
<span class="n">time_elapsed</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span> <span class="o">-</span> <span class="n">since</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'Training complete in </span><span class="si">{:.0f}</span><span class="s1">m </span><span class="si">{:.0f}</span><span class="s1">s'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span>
<span class="n">time_elapsed</span> <span class="o">//</span> <span class="mi">60</span><span class="p">,</span> <span class="n">time_elapsed</span> <span class="o">%</span> <span class="mi">60</span><span class="p">))</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'Best val Acc: </span><span class="si">{:4f}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">best_acc</span><span class="p">))</span>
<span class="c1"># 최적의 모델 가중치 불러오기</span>
<span class="n">model</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">best_model_wts</span><span class="p">)</span>
<span class="k">return</span> <span class="n">model</span>
</pre></div>
</div>
</div>
<div class="section" id="id8">
<h3>모델 예측을 시각화하기 위한 지원 함수<a class="headerlink" href="#id8" title="Permalink to this headline">¶</a></h3>
<p>일부 이미지에 대한 예측을 출력하는 일반 함수</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">visualize_model</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">rows</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">cols</span><span class="o">=</span><span class="mi">3</span><span class="p">):</span>
<span class="n">was_training</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">training</span>
<span class="n">model</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">current_row</span> <span class="o">=</span> <span class="n">current_col</span> <span class="o">=</span> <span class="mi">0</span>
<span class="n">fig</span><span class="p">,</span> <span class="n">ax</span> <span class="o">=</span> <span class="n">plt</span><span class="o">.</span><span class="n">subplots</span><span class="p">(</span><span class="n">rows</span><span class="p">,</span> <span class="n">cols</span><span class="p">,</span> <span class="n">figsize</span><span class="o">=</span><span class="p">(</span><span class="n">cols</span><span class="o">*</span><span class="mi">2</span><span class="p">,</span> <span class="n">rows</span><span class="o">*</span><span class="mi">2</span><span class="p">))</span>
<span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="k">for</span> <span class="n">idx</span><span class="p">,</span> <span class="p">(</span><span class="n">imgs</span><span class="p">,</span> <span class="n">lbls</span><span class="p">)</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">dataloaders</span><span class="p">[</span><span class="s1">'val'</span><span class="p">]):</span>
<span class="n">imgs</span> <span class="o">=</span> <span class="n">imgs</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
<span class="n">lbls</span> <span class="o">=</span> <span class="n">lbls</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
<span class="n">outputs</span> <span class="o">=</span> <span class="n">model</span><span class="p">(</span><span class="n">imgs</span><span class="p">)</span>
<span class="n">_</span><span class="p">,</span> <span class="n">preds</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">outputs</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">for</span> <span class="n">jdx</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">imgs</span><span class="o">.</span><span class="n">size</span><span class="p">()[</span><span class="mi">0</span><span class="p">]):</span>
<span class="n">imshow</span><span class="p">(</span><span class="n">imgs</span><span class="o">.</span><span class="n">data</span><span class="p">[</span><span class="n">jdx</span><span class="p">],</span> <span class="n">ax</span><span class="o">=</span><span class="n">ax</span><span class="p">[</span><span class="n">current_row</span><span class="p">,</span> <span class="n">current_col</span><span class="p">])</span>
<span class="n">ax</span><span class="p">[</span><span class="n">current_row</span><span class="p">,</span> <span class="n">current_col</span><span class="p">]</span><span class="o">.</span><span class="n">axis</span><span class="p">(</span><span class="s1">'off'</span><span class="p">)</span>
<span class="n">ax</span><span class="p">[</span><span class="n">current_row</span><span class="p">,</span> <span class="n">current_col</span><span class="p">]</span><span class="o">.</span><span class="n">set_title</span><span class="p">(</span><span class="s1">'predicted: </span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">class_names</span><span class="p">[</span><span class="n">preds</span><span class="p">[</span><span class="n">jdx</span><span class="p">]]))</span>
<span class="n">current_col</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="k">if</span> <span class="n">current_col</span> <span class="o">>=</span> <span class="n">cols</span><span class="p">:</span>
<span class="n">current_row</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="n">current_col</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">if</span> <span class="n">current_row</span> <span class="o">>=</span> <span class="n">rows</span><span class="p">:</span>
<span class="n">model</span><span class="o">.</span><span class="n">train</span><span class="p">(</span><span class="n">mode</span><span class="o">=</span><span class="n">was_training</span><span class="p">)</span>
<span class="k">return</span>
<span class="n">model</span><span class="o">.</span><span class="n">train</span><span class="p">(</span><span class="n">mode</span><span class="o">=</span><span class="n">was_training</span><span class="p">)</span>
</pre></div>
</div>
</div>
</div>
<div class="section" id="quantized-feature-extractor">
<h2>파트 1. 양자화된 특징 추출기(Quantized Feature Extractor)를 기반으로 사용자 지정 분류기 훈련하기<a class="headerlink" href="#quantized-feature-extractor" title="Permalink to this headline">¶</a></h2>
<p>이번 섹션에서는 “고정된(frozen)” 양자화 특징 추출기를 사용하고 그 위에 사용자 지정 분류기 헤드를
학습합니다. 부동 소수점 모델과 다르게 양자화된 모델에는 학습 가능한 매개 변수가 없으므로
requires_grad = False를 설정할 필요가 없습니다. 자세한 내용은 <a class="reference external" href="https://pytorch.org/docs/stable/quantization.html">설명서</a> 를 참조하세요.</p>
<p>미리 학습된 모델을 불러옵니다: 이번 예제에서는 <a class="reference external" href="https://pytorch.org/hub/pytorch_vision_resnet/">ResNet-18</a> 을 사용할 것입니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torchvision.models.quantization</span> <span class="k">as</span> <span class="nn">models</span>
<span class="c1"># 나중에 사용할 수 있게 `fc` 에 필터의 수가 필요합니다.</span>
<span class="c1"># 여기서 각 출력 샘플의 크기는 2로 설정합니다.</span>
<span class="c1"># 또한, nn.Linear(num_ftrs, len(class_names))로 일반화 할 수 있습니다.</span>
<span class="n">model_fe</span> <span class="o">=</span> <span class="n">models</span><span class="o">.</span><span class="n">resnet18</span><span class="p">(</span><span class="n">pretrained</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">progress</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">quantize</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">num_ftrs</span> <span class="o">=</span> <span class="n">model_fe</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">in_features</span>
</pre></div>
</div>
<p>이 시점에서 미리 학습된 모델을 수정해야 합니다. 모델의 시작과 끝에는 양자화/역양자화 블록이 있습니다.
그러나 특징 추출기만 사용하기 때문에 역양자화(dequantization) 계층은 선형 계층(헤드) 바로 전으로 이동시켜야 합니다.
가장 쉬운 방법은 모델을 <code class="docutils literal notranslate"><span class="pre">nn.Sequential</span></code> 모듈로 감싸는 것입니다.</p>
<p>첫번째 단계는 ResNet 모델에서 특징 추출기를 분리하는 것입니다.
이 예제에서는 <code class="docutils literal notranslate"><span class="pre">fc</span></code> 를 제외한 모든 계층을 특징 추출기로 사용해야 하지만, 실제로는 필요한 만큼 많은 부분을 사용할 수 있습니다.
이것은 합성곱 계층 중 일부를 교체하려는 경우에도 유용합니다.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">양자화 모델에서 특징 추출기를 분리할 때 양자화를 유지하려는 부분의 시작과 끝에 수동으로 양자화/역양자화를 배치해야 합니다.</p>
</div>
<p>아래 함수는 사용자 지정 헤드로 모델을 생성하는 함수입니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
<span class="k">def</span> <span class="nf">create_combined_model</span><span class="p">(</span><span class="n">model_fe</span><span class="p">):</span>
<span class="c1"># 1 단계. 특징 추출기를 분리합니다.</span>
<span class="n">model_fe_features</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">quant</span><span class="p">,</span> <span class="c1"># Quantize the input</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">conv1</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">bn1</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">relu</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">maxpool</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">layer1</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">layer2</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">layer3</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">layer4</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">avgpool</span><span class="p">,</span>
<span class="n">model_fe</span><span class="o">.</span><span class="n">dequant</span><span class="p">,</span> <span class="c1"># 출력을 역양자화하기</span>
<span class="p">)</span>
<span class="c1"># 2 단계. 새로운 "헤드(head)"를 만듭니다.</span>
<span class="n">new_head</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Dropout</span><span class="p">(</span><span class="n">p</span><span class="o">=</span><span class="mf">0.5</span><span class="p">),</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="n">num_ftrs</span><span class="p">,</span> <span class="mi">2</span><span class="p">),</span>
<span class="p">)</span>
<span class="c1"># 3 단계. 결합하고 양자 스텁(stubs)을 잊으면 안됩니다.</span>
<span class="n">new_model</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span>
<span class="n">model_fe_features</span><span class="p">,</span>
<span class="n">nn</span><span class="o">.</span><span class="n">Flatten</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span>
<span class="n">new_head</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">new_model</span>
</pre></div>
</div>
<div class="admonition warning">
<p class="first admonition-title">Warning</p>
<p class="last">현재 양자화된 모델은 CPU에서만 실행할 수 있습니다.
그러나 모델의 양자화 되지 않은 부분은 GPU로 보낼 수 있습니다.</p>
</div>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch.optim</span> <span class="k">as</span> <span class="nn">optim</span>
<span class="n">new_model</span> <span class="o">=</span> <span class="n">create_combined_model</span><span class="p">(</span><span class="n">model_fe</span><span class="p">)</span>
<span class="n">new_model</span> <span class="o">=</span> <span class="n">new_model</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="s1">'cpu'</span><span class="p">)</span>
<span class="n">criterion</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">CrossEntropyLoss</span><span class="p">()</span>
<span class="c1"># 헤드(the head)만 훈련 한다는 점을 유의하세요</span>
<span class="n">optimizer_ft</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">new_model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
<span class="c1"># 7 에폭마다 0.1배씩 학습률이 감소</span>
<span class="n">exp_lr_scheduler</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">lr_scheduler</span><span class="o">.</span><span class="n">StepLR</span><span class="p">(</span><span class="n">optimizer_ft</span><span class="p">,</span> <span class="n">step_size</span><span class="o">=</span><span class="mi">7</span><span class="p">,</span> <span class="n">gamma</span><span class="o">=</span><span class="mf">0.1</span><span class="p">)</span>
</pre></div>
</div>
<div class="section" id="id10">
<h3>학습과 평가<a class="headerlink" href="#id10" title="Permalink to this headline">¶</a></h3>
<p>이 단계는 CPU에서 약 15 ~ 25분 걸립니다. 양자화된 모델은 CPU에서만 실행되기 때문에
GPU에서는 훈련을 실행할 수 없습니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">new_model</span> <span class="o">=</span> <span class="n">train_model</span><span class="p">(</span><span class="n">new_model</span><span class="p">,</span> <span class="n">criterion</span><span class="p">,</span> <span class="n">optimizer_ft</span><span class="p">,</span> <span class="n">exp_lr_scheduler</span><span class="p">,</span>
<span class="n">num_epochs</span><span class="o">=</span><span class="mi">25</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s1">'cpu'</span><span class="p">)</span>
<span class="n">visualize_model</span><span class="p">(</span><span class="n">new_model</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">tight_layout</span><span class="p">()</span>
</pre></div>
</div>
</div>
</div>
<div class="section" id="finetuning">
<h2>파트 2. 양자화 가능한 모델 미세조정(Finetuning)<a class="headerlink" href="#finetuning" title="Permalink to this headline">¶</a></h2>
<p>이번 파트에서는 전이학습(Transfer Learning)을 사용하여 특징 추출기(Feature Extractor)를
미세조정(Finetuning) 합니다. 파트 1과 2 모두에서 특징 추출기는 양자화됩니다. 차이점은 파트 1에서
미리 학습 된 양자화 모델을 사용합니다. 이번 파트에서, 우리는 관심있는 데이터셋으로 미세조정(Finetuning)한 후
양자화된 특징 추출기를 생성하므로, 양자화의 장점을 가지면서 전이 학습(Transfer Learning)으로 더 나은 정확도를
얻을 수 있습니다. 특정한 예제에서는 학습용 셋은 매우 작기 때문에(120개의 이미지) 전체 모델을
미세조정(Finetuning)하는 장점이 불분명 합니다. 그러나 여기에 표시된 절차는 더욱 더 큰 데이터셋을 사용한 전이 학습(Transfer Learning)의
정확도를 향상시킵니다.</p>
<p>미리 학습된 특징 추출기는 양자화가 가능해야 합니다.
양자화가 가능한지 확인하기 위해서 다음 단계를 수행하세요:</p>
<blockquote>
<div><ol class="arabic simple">
<li><code class="docutils literal notranslate"><span class="pre">torch.quantization.fuse_modules</span></code> 를 사용하여 <code class="docutils literal notranslate"><span class="pre">(Conv,</span> <span class="pre">BN,</span> <span class="pre">ReLU)</span></code> ,
<code class="docutils literal notranslate"><span class="pre">(Conv,</span> <span class="pre">BN)</span></code>, 그리고 <code class="docutils literal notranslate"><span class="pre">(Conv,</span> <span class="pre">ReLU)</span></code> 를 융합합니다.</li>
<li>특징 추출기를 사용자 지정 헤드와 연결합니다. 이를 위해서 특징 추출기의 출력을 역으로 양자화 해야합니다.</li>
<li>특징 추출기의 적합한 위치에 가짜 양자화 모듈을 삽입하여 학습하는 동안에 양자화를 모방합니다.</li>
</ol>
</div></blockquote>
<p>(1) 단계의 경우 멤버 메서드(member method) <code class="docutils literal notranslate"><span class="pre">fuse_model</span></code> 이 있는
<code class="docutils literal notranslate"><span class="pre">torchvision/models/quantization</span></code> 의 모델을 사용합니다.
이 함수는 모든 <code class="docutils literal notranslate"><span class="pre">conv</span></code> , <code class="docutils literal notranslate"><span class="pre">bn</span></code> , 그리고 <code class="docutils literal notranslate"><span class="pre">relu</span></code> 모듈을 통합합니다.
사용자 지정 모델의 경우, 수동으로 통합할 모듈의 목록과 함께 <code class="docutils literal notranslate"><span class="pre">torch.quantization.fuse_modules</span></code> API를 호출해야합니다.</p>
<ol class="arabic simple" start="2">
<li>단계는 이전 섹션에서 사용한 <code class="docutils literal notranslate"><span class="pre">create_combined_model</span></code> 함수에 의해서 수행됩니다.</li>
<li>단계는 가짜 양자화 모듈을 삽입하는 <code class="docutils literal notranslate"><span class="pre">torch.quantization.prepare_qat</span></code> 를 사용하여 수행됩니다.</li>
<li>단계로 모델을 “미세조정(Finetuning)”한 후, 완전하게 양자화된 버전으로 변환(5단계) 할 수 있습니다.</li>
</ol>
<p>미세조정(Finetuning) 모델을 양자화된 모델로 변환하려면 <code class="docutils literal notranslate"><span class="pre">torch.quantization.convert</span></code> 함수를
호출 할 수 있습니다. (이 경우 특징 추출기만 양자화 됩니다.)</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">랜덤 초기화 때문에 여러분의 결과가 튜토리얼에 표시된 결과와 다를 수 있습니다.</p>
</div>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># `quantize=False` 를 주목하세요</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">models</span><span class="o">.</span><span class="n">resnet18</span><span class="p">(</span><span class="n">pretrained</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">progress</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">quantize</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
<span class="n">num_ftrs</span> <span class="o">=</span> <span class="n">model</span><span class="o">.</span><span class="n">fc</span><span class="o">.</span><span class="n">in_features</span>
<span class="c1"># 1 단계</span>
<span class="n">model</span><span class="o">.</span><span class="n">train</span><span class="p">()</span>
<span class="n">model</span><span class="o">.</span><span class="n">fuse_model</span><span class="p">()</span>
<span class="c1"># 2 단계</span>
<span class="n">model_ft</span> <span class="o">=</span> <span class="n">create_combined_model</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="n">model_ft</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">qconfig</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">default_qat_qconfig</span> <span class="c1"># Use default QAT configuration</span>
<span class="c1"># 3 단계</span>
<span class="n">model_ft</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">prepare_qat</span><span class="p">(</span><span class="n">model_ft</span><span class="p">,</span> <span class="n">inplace</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
</pre></div>
</div>
<div class="section" id="id11">
<h3>모델 미세조정<a class="headerlink" href="#id11" title="Permalink to this headline">¶</a></h3>
<p>현재 튜토리얼에서는 전체 모델이 미세조정 되었습니다.
일반적으로 이것은 더 높은 정확도로 이어질 것입니다.
그러나 여기서는 크기가 작은 학습용 데이터셋을 사용했기 때문에 결국 과적합하게 됩니다.</p>
<p>4 단계. 모델 미세조정하기</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">model_ft</span><span class="o">.</span><span class="n">parameters</span><span class="p">():</span>
<span class="n">param</span><span class="o">.</span><span class="n">requires_grad</span> <span class="o">=</span> <span class="kc">True</span>
<span class="n">model_ft</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="p">)</span> <span class="c1"># GPU에서 미세조정(Finetuning) 할 수 있습니다.</span>
<span class="n">criterion</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">CrossEntropyLoss</span><span class="p">()</span>
<span class="c1"># 이미 모든 것이 학습된 상태이므로 학습률이 낮습니다.</span>
<span class="c1"># 더 작은 Learning rate에 주목하세요</span>
<span class="n">optimizer_ft</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model_ft</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">1e-3</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">,</span> <span class="n">weight_decay</span><span class="o">=</span><span class="mf">0.1</span><span class="p">)</span>
<span class="c1"># 학습률을 몇 에폭마다 0.3배 감소시키기</span>
<span class="n">exp_lr_scheduler</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">lr_scheduler</span><span class="o">.</span><span class="n">StepLR</span><span class="p">(</span><span class="n">optimizer_ft</span><span class="p">,</span> <span class="n">step_size</span><span class="o">=</span><span class="mi">5</span><span class="p">,</span> <span class="n">gamma</span><span class="o">=</span><span class="mf">0.3</span><span class="p">)</span>
<span class="n">model_ft_tuned</span> <span class="o">=</span> <span class="n">train_model</span><span class="p">(</span><span class="n">model_ft</span><span class="p">,</span> <span class="n">criterion</span><span class="p">,</span> <span class="n">optimizer_ft</span><span class="p">,</span> <span class="n">exp_lr_scheduler</span><span class="p">,</span>
<span class="n">num_epochs</span><span class="o">=</span><span class="mi">25</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">)</span>
</pre></div>
</div>
<p>5 단계. 양자화된 모델로 변환하기</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch.quantization</span> <span class="kn">import</span> <span class="n">convert</span>
<span class="n">model_ft_tuned</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
<span class="n">model_quantized_and_trained</span> <span class="o">=</span> <span class="n">convert</span><span class="p">(</span><span class="n">model_ft_tuned</span><span class="p">,</span> <span class="n">inplace</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
</pre></div>
</div>
<p>양자화된 모델이 일부 이미지에서 어떻게 동작하는지 살펴보겠습니다.</p>
<div class="code python highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">visualize_model</span><span class="p">(</span><span class="n">model_quantized_and_trained</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">ioff</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">tight_layout</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
</div>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="../advanced/static_quantization_tutorial.html" class="btn btn-neutral float-right" title="(beta) Static Quantization with Eager Mode in PyTorch" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="dynamic_quantization_bert_tutorial.html" class="btn btn-neutral" title="(베타) BERT 모델 동적 양자화하기" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr class="rating-hr hr-top">
<div class="rating-container">
<div class="rating-prompt">Rate this Tutorial</div>
<div class="stars-outer">
<i class="far fa-star" title="1 Star" data-behavior="tutorial-rating" data-count="1"></i>
<i class="far fa-star" title="2 Stars" data-behavior="tutorial-rating" data-count="2"></i>
<i class="far fa-star" title="3 Stars" data-behavior="tutorial-rating" data-count="3"></i>
<i class="far fa-star" title="4 Stars" data-behavior="tutorial-rating" data-count="4"></i>
<i class="far fa-star" title="5 Stars" data-behavior="tutorial-rating" data-count="5"></i>
</div>
</div>
<hr class="rating-hr hr-bottom"/>
<div role="contentinfo">
<p>
© Copyright 2021, PyTorch & PyTorch Korea Community.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning)</a><ul>
<li><a class="reference internal" href="#id2">파트 0. 요구사항</a><ul>
<li><a class="reference internal" href="#nightly-build">Nightly Build 설치하기</a></li>
<li><a class="reference internal" href="#id4">데이터 불러오기</a></li>
<li><a class="reference internal" href="#id6">일부 이미지 시각화하기</a></li>
<li><a class="reference internal" href="#id7">모델 학습을 위한 지원 함수</a></li>
<li><a class="reference internal" href="#id8">모델 예측을 시각화하기 위한 지원 함수</a></li>
</ul>
</li>
<li><a class="reference internal" href="#quantized-feature-extractor">파트 1. 양자화된 특징 추출기(Quantized Feature Extractor)를 기반으로 사용자 지정 분류기 훈련하기</a><ul>
<li><a class="reference internal" href="#id10">학습과 평가</a></li>
</ul>
</li>
<li><a class="reference internal" href="#finetuning">파트 2. 양자화 가능한 모델 미세조정(Finetuning)</a><ul>
<li><a class="reference internal" href="#id11">모델 미세조정</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script type="text/javascript" src="../_static/jquery.js"></script>
<script type="text/javascript" src="../_static/underscore.js"></script>
<script type="text/javascript" src="../_static/doctools.js"></script>
<script type="text/javascript" src="../_static/clipboard.min.js"></script>
<script type="text/javascript" src="../_static/copybutton.js"></script>
<script type="text/javascript" src="https://cdn.jsdelivr.net/npm/katex@0.13.11/dist/katex.min.js"></script>
<script type="text/javascript" src="https://cdn.jsdelivr.net/npm/katex@0.13.11/dist/contrib/auto-render.min.js"></script>
<script type="text/javascript" src="../_static/katex_autorenderer.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script>
//add microsoft link
if(window.location.href.indexOf("/beginner/basics/")!= -1)
{
var url="https://docs.microsoft.com/learn/paths/pytorch-fundamentals/?wt.mc_id=aiml-7486-cxa";
switch(window.location.pathname.split("/").pop().replace('.html',''))
{
case"quickstart_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/9-quickstart?WT.mc_id=aiml-7486-cxa";
break;
case"tensorqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/2-tensors?WT.mc_id=aiml-7486-cxa";
break;
case"data_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/3-data?WT.mc_id=aiml-7486-cxa";
break;
case"transforms_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/4-transforms?WT.mc_id=aiml-7486-cxa";
break;
case"buildmodel_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/5-model?WT.mc_id=aiml-7486-cxa";
break;
case"autogradqs_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/6-autograd?WT.mc_id=aiml-7486-cxa";
break;
case"optimization_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/7-optimization?WT.mc_id=aiml-7486-cxa";
break;
case"saveloadrun_tutorial":
url="https://docs.microsoft.com/learn/modules/intro-machine-learning-pytorch/8-inference?WT.mc_id=aiml-7486-cxa";
}
$(".pytorch-call-to-action-links").children().first().before("<a href="+url+' data-behavior="call-to-action-event" data-response="Run in Microsoft Learn" target="_blank"><div id="microsoft-learn-link" style="padding-bottom: 0.625rem;border-bottom: 1px solid #f3f4f7;padding-right: 2.5rem;display: -webkit-box; display: -ms-flexbox; isplay: flex; -webkit-box-align: center;-ms-flex-align: center;align-items: center;"><img class="call-to-action-img" src="../../_static/images/microsoft-logo.svg"/><div class="call-to-action-desktop-view">Run in Microsoft Learn</div><div class="call-to-action-mobile-view">Learn</div></div></a>')
}
</script>
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-71919972-3"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-71919972-3');
</script>
<script>
$("[data-behavior='call-to-action-event']").on('click', function(){
ga('send', {
hitType: 'event',
eventCategory: $(this).attr("data-response"),
eventAction: 'click',
eventLabel: window.location.href
});
gtag('event', 'click', {
'event_category': $(this).attr("data-response"),
'event_label': $("h1").first().text(),
'tutorial_link': window.location.href
});
});
$("[data-behavior='tutorial-rating']").on('click', function(){
gtag('event', 'click', {
'event_category': 'Tutorial Rating',
'event_label': $("h1").first().text(),
'value': $(this).attr("data-count")
});
});
if (location.pathname == "/") {
$(".rating-container").hide();
$(".hr-bottom").hide();
}
</script>