-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathreinforcement_q_learning.html
1141 lines (921 loc) · 92.1 KB
/
reinforcement_q_learning.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>강화 학습 (DQN) 튜토리얼 — 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="Train a Mario-playing RL Agent" href="mario_rl_tutorial.html" />
<link rel="prev" title="nn.Transformer와 torchtext로 언어 번역하기" href="../beginner/translation_transformer.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 class="current">
<li class="toctree-l1 current"><a class="current reference internal" href="#">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>
<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"><a class="reference internal" href="quantized_transfer_learning_tutorial.html">(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>강화 학습 (DQN) 튜토리얼</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/reinforcement_q_learning.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/reinforcement_q_learning</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="sphx-glr-download-link-note admonition note">
<p class="first admonition-title">Note</p>
<p class="last">Click <a class="reference internal" href="#sphx-glr-download-intermediate-reinforcement-q-learning-py"><span class="std std-ref">here</span></a> to download the full example code</p>
</div>
<div class="sphx-glr-example-title section" id="dqn">
<span id="sphx-glr-intermediate-reinforcement-q-learning-py"></span><h1>강화 학습 (DQN) 튜토리얼<a class="headerlink" href="#dqn" title="Permalink to this headline">¶</a></h1>
<dl class="docutils">
<dt><strong>Author</strong>: <a class="reference external" href="https://github.com/apaszke">Adam Paszke</a></dt>
<dd><strong>번역</strong>: <a class="reference external" href="https://github.com/adonisues">황성수</a></dd>
</dl>
<p>이 튜토리얼에서는 <a class="reference external" href="https://gym.openai.com/">OpenAI Gym</a> 의
CartPole-v0 태스크에서 DQN (Deep Q Learning) 에이전트를 학습하는데
PyTorch를 사용하는 방법을 보여드립니다.</p>
<p><strong>태스크</strong></p>
<p>에이전트는 연결된 막대가 똑바로 서 있도록 카트를 왼쪽이나 오른쪽으로
움직이는 두 가지 동작 중 하나를 선택해야 합니다.
다양한 알고리즘과 시각화 기능을 갖춘 공식 순위표를
<a class="reference external" href="https://gym.openai.com/envs/CartPole-v0">Gym 웹사이트</a> 에서 찾을 수 있습니다.</p>
<div class="figure" id="id7">
<img alt="cartpole" src="../_images/cartpole.gif" />
<p class="caption"><span class="caption-text">cartpole</span></p>
</div>
<p>에이전트가 현재 환경 상태를 관찰하고 행동을 선택하면,
환경이 새로운 상태로 <em>전환</em> 되고 작업의 결과를 나타내는 보상도 반환됩니다.
이 태스크에서 매 타임스텝 증가마다 보상이 +1이 되고, 만약 막대가 너무 멀리
떨어지거나 카트가 중심에서 2.4 유닛 이상 멀어지면 환경이 중단됩니다.
이것은 더 좋은 시나리오가 더 오랫동안 더 많은 보상을 축적하는 것을 의미합니다.</p>
<p>카트폴 태스크는 에이전트에 대한 입력이 환경 상태(위치, 속도 등)를 나타내는
4개의 실제 값이 되도록 설계되었습니다. 그러나 신경망은 순수하게 그 장면을 보고
태스크를 해결할 수 있습니다 따라서 카트 중심의 화면 패치를 입력으로 사용합니다.
이 때문에 우리의 결과는 공식 순위표의 결과와 직접적으로 비교할 수 없습니다.
우리의 태스크는 훨씬 더 어렵습니다.
불행히도 모든 프레임을 렌더링해야되므로 이것은 학습 속도를 늦추게됩니다.</p>
<p>엄밀히 말하면, 현재 스크린 패치와 이전 스크린 패치 사이의 차이로 상태를 표시할 것입니다.
이렇게하면 에이전트가 막대의 속도를 한 이미지에서 고려할 수 있습니다.</p>
<p><strong>패키지</strong></p>
<p>먼저 필요한 패키지를 가져옵니다. 첫째, 환경을 위해
<a class="reference external" href="https://gym.openai.com/docs">gym</a> 이 필요합니다.
(<cite>pip install gym</cite> 을 사용하여 설치하십시오).
또한 PyTorch에서 다음을 사용합니다:</p>
<ul class="simple">
<li>신경망 (<code class="docutils literal notranslate"><span class="pre">torch.nn</span></code>)</li>
<li>최적화 (<code class="docutils literal notranslate"><span class="pre">torch.optim</span></code>)</li>
<li>자동 미분 (<code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code>)</li>
<li>시각 태스크를 위한 유틸리티들 (<code class="docutils literal notranslate"><span class="pre">torchvision</span></code> - <a class="reference external" href="https://github.com/pytorch/vision">a separate
package</a>).</li>
</ul>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">gym</span>
<span class="kn">import</span> <span class="nn">math</span>
<span class="kn">import</span> <span class="nn">random</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">matplotlib</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">from</span> <span class="nn">collections</span> <span class="kn">import</span> <span class="n">namedtuple</span><span class="p">,</span> <span class="n">deque</span>
<span class="kn">from</span> <span class="nn">itertools</span> <span class="kn">import</span> <span class="n">count</span>
<span class="kn">from</span> <span class="nn">PIL</span> <span class="kn">import</span> <span class="n">Image</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="kn">import</span> <span class="nn">torch.optim</span> <span class="k">as</span> <span class="nn">optim</span>
<span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="kn">import</span> <span class="nn">torchvision.transforms</span> <span class="k">as</span> <span class="nn">T</span>
<span class="n">env</span> <span class="o">=</span> <span class="n">gym</span><span class="o">.</span><span class="n">make</span><span class="p">(</span><span class="s1">'CartPole-v0'</span><span class="p">)</span><span class="o">.</span><span class="n">unwrapped</span>
<span class="c1"># matplotlib 설정</span>
<span class="n">is_ipython</span> <span class="o">=</span> <span class="s1">'inline'</span> <span class="ow">in</span> <span class="n">matplotlib</span><span class="o">.</span><span class="n">get_backend</span><span class="p">()</span>
<span class="k">if</span> <span class="n">is_ipython</span><span class="p">:</span>
<span class="kn">from</span> <span class="nn">IPython</span> <span class="kn">import</span> <span class="n">display</span>
<span class="n">plt</span><span class="o">.</span><span class="n">ion</span><span class="p">()</span>
<span class="c1"># GPU를 사용할 경우</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"</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 class="section" id="replay-memory">
<h2>재현 메모리(Replay Memory)<a class="headerlink" href="#replay-memory" title="Permalink to this headline">¶</a></h2>
<p>우리는 DQN 학습을 위해 경험 재현 메모리를 사용할 것입니다.
에이전트가 관찰한 전환(transition)을 저장하고 나중에 이 데이터를
재사용할 수 있습니다. 무작위로 샘플링하면 배치를 구성하는 전환들이
비상관(decorrelated)하게 됩니다. 이것이 DQN 학습 절차를 크게 안정시키고
향상시키는 것으로 나타났습니다.</p>
<p>이를 위해서 두개의 클래스가 필요합니다:</p>
<ul class="simple">
<li><code class="docutils literal notranslate"><span class="pre">Transition</span></code> - 우리 환경에서 단일 전환을 나타내도록 명명된 튜플.
그것은 화면의 차이인 state로 (state, action) 쌍을 (next_state, reward) 결과로 매핑합니다.</li>
<li><code class="docutils literal notranslate"><span class="pre">ReplayMemory</span></code> - 최근 관찰된 전이를 보관 유지하는 제한된 크기의 순환 버퍼.
또한 학습을 위한 전환의 무작위 배치를 선택하기위한
<code class="docutils literal notranslate"><span class="pre">.sample</span> <span class="pre">()</span></code> 메소드를 구현합니다.</li>
</ul>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Transition</span> <span class="o">=</span> <span class="n">namedtuple</span><span class="p">(</span><span class="s1">'Transition'</span><span class="p">,</span>
<span class="p">(</span><span class="s1">'state'</span><span class="p">,</span> <span class="s1">'action'</span><span class="p">,</span> <span class="s1">'next_state'</span><span class="p">,</span> <span class="s1">'reward'</span><span class="p">))</span>
<span class="k">class</span> <span class="nc">ReplayMemory</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">capacity</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">memory</span> <span class="o">=</span> <span class="n">deque</span><span class="p">([],</span><span class="n">maxlen</span><span class="o">=</span><span class="n">capacity</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">push</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="o">*</span><span class="n">args</span><span class="p">):</span>
<span class="sd">"""transition 저장"""</span>
<span class="bp">self</span><span class="o">.</span><span class="n">memory</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">Transition</span><span class="p">(</span><span class="o">*</span><span class="n">args</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">sample</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">batch_size</span><span class="p">):</span>
<span class="k">return</span> <span class="n">random</span><span class="o">.</span><span class="n">sample</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">memory</span><span class="p">,</span> <span class="n">batch_size</span><span class="p">)</span>
<span class="k">def</span> <span class="fm">__len__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">memory</span><span class="p">)</span>
</pre></div>
</div>
<p>이제 모델을 정의합시다. 그러나 먼저 DQN이 무엇인지 간단히 요약해 보겠습니다.</p>
</div>
<div class="section" id="id2">
<h2>DQN 알고리즘<a class="headerlink" href="#id2" title="Permalink to this headline">¶</a></h2>
<p>우리의 환경은 결정론적이므로 여기에 제시된 모든 방정식은 단순화를 위해
결정론적으로 공식화됩니다. 강화 학습 자료은 환경에서 확률론적 전환에
대한 기대값(expectation)도 포함할 것입니다.</p>
<p>우리의 목표는 할인된 누적 보상 (discounted cumulative reward)을
극대화하려는 정책(policy)을 학습하는 것입니다.
<span class="math">\(R_{t_0} = \sum_{t=t_0}^{\infty} \gamma^{t - t_0} r_t\)</span>, 여기서
<span class="math">\(R_{t_0}\)</span> 는 <em>반환(return)</em> 입니다. 할인 상수,
<span class="math">\(\gamma\)</span>, 는 <span class="math">\(0\)</span> 과 <span class="math">\(1\)</span> 의 상수이고 합계가
수렴되도록 보장합니다. 에이전트에게 불확실한 먼 미래의 보상이
가까운 미래의 것에 비해 덜 중요하게 만들고, 이것은 상당히 합리적입니다.</p>
<p>Q-learning의 주요 아이디어는 만일 함수 <span class="math">\(Q^*: State \times Action \rightarrow \mathbb{R}\)</span> 를
가지고 있다면 반환이 어떻게 될지 알려줄 수 있고,
만약 주어진 상태(state)에서 행동(action)을 한다면, 보상을 최대화하는
정책을 쉽게 구축할 수 있습니다:</p>
<div class="math">
\[\pi^*(s) = \arg\!\max_a \ Q^*(s, a)
\]</div>
<p>그러나 세계(world)에 관한 모든 것을 알지 못하기 때문에,
<span class="math">\(Q^*\)</span> 에 도달할 수 없습니다. 그러나 신경망은
범용 함수 근사자(universal function approximator)이기 때문에
간단하게 생성하고 <span class="math">\(Q^*\)</span> 를 닮도록 학습할 수 있습니다.</p>
<p>학습 업데이트 규칙으로, 일부 정책을 위한 모든 <span class="math">\(Q\)</span> 함수가
Bellman 방정식을 준수한다는 사실을 사용할 것입니다:</p>
<div class="math">
\[Q^{\pi}(s, a) = r + \gamma Q^{\pi}(s', \pi(s'))
\]</div>
<p>평등(equality)의 두 측면 사이의 차이는
시간차 오류(temporal difference error), <span class="math">\(\delta\)</span> 입니다.:</p>
<div class="math">
\[\delta = Q(s, a) - (r + \gamma \max_a Q(s', a))
\]</div>
<p>오류를 최소화하기 위해서 <a class="reference external" href="https://en.wikipedia.org/wiki/Huber_loss">Huber
loss</a> 를 사용합니다.
Huber loss 는 오류가 작으면 평균 제곱 오차( mean squared error)와 같이
동작하고 오류가 클 때는 평균 절대 오류와 유사합니다.
- 이것은 <span class="math">\(Q\)</span> 의 추정이 매우 혼란스러울 때 이상 값에 더 강건하게 합니다.
재현 메모리에서 샘플링한 전환 배치 <span class="math">\(B\)</span> 에서 이것을 계산합니다:</p>
<div class="math">
\[\mathcal{L} = \frac{1}{|B|}\sum_{(s, a, s', r) \ \in \ B} \mathcal{L}(\delta)\]</div>
<div class="math">
\[\text{where} \quad \mathcal{L}(\delta) = \begin{cases}
\frac{1}{2}{\delta^2} & \text{for } |\delta| \le 1, \\
|\delta| - \frac{1}{2} & \text{otherwise.}
\end{cases}\]</div>
<div class="section" id="q">
<h3>Q-네트워크<a class="headerlink" href="#q" title="Permalink to this headline">¶</a></h3>
<p>우리 모델은 현재와 이전 스크린 패치의 차이를 취하는
CNN(convolutional neural network) 입니다. 두가지 출력 <span class="math">\(Q(s, \mathrm{left})\)</span> 와
<span class="math">\(Q(s, \mathrm{right})\)</span> 가 있습니다. (여기서 <span class="math">\(s\)</span> 는 네트워크의 입력입니다)
결과적으로 네트워크는 주어진 현재 입력에서 각 행동의 <em>기대값</em> 을 예측하려고 합니다.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">DQN</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="n">w</span><span class="p">,</span> <span class="n">outputs</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">DQN</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">16</span><span class="p">,</span> <span class="n">kernel_size</span><span class="o">=</span><span class="mi">5</span><span class="p">,</span> <span class="n">stride</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">bn1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="mi">16</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">16</span><span class="p">,</span> <span class="mi">32</span><span class="p">,</span> <span class="n">kernel_size</span><span class="o">=</span><span class="mi">5</span><span class="p">,</span> <span class="n">stride</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">bn2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="mi">32</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">conv3</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Conv2d</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">32</span><span class="p">,</span> <span class="n">kernel_size</span><span class="o">=</span><span class="mi">5</span><span class="p">,</span> <span class="n">stride</span><span class="o">=</span><span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">bn3</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="mi">32</span><span class="p">)</span>
<span class="c1"># Linear 입력의 연결 숫자는 conv2d 계층의 출력과 입력 이미지의 크기에</span>
<span class="c1"># 따라 결정되기 때문에 따로 계산을 해야합니다.</span>
<span class="k">def</span> <span class="nf">conv2d_size_out</span><span class="p">(</span><span class="n">size</span><span class="p">,</span> <span class="n">kernel_size</span> <span class="o">=</span> <span class="mi">5</span><span class="p">,</span> <span class="n">stride</span> <span class="o">=</span> <span class="mi">2</span><span class="p">):</span>
<span class="k">return</span> <span class="p">(</span><span class="n">size</span> <span class="o">-</span> <span class="p">(</span><span class="n">kernel_size</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span> <span class="o">-</span> <span class="mi">1</span><span class="p">)</span> <span class="o">//</span> <span class="n">stride</span> <span class="o">+</span> <span class="mi">1</span>
<span class="n">convw</span> <span class="o">=</span> <span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">w</span><span class="p">)))</span>
<span class="n">convh</span> <span class="o">=</span> <span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">conv2d_size_out</span><span class="p">(</span><span class="n">h</span><span class="p">)))</span>
<span class="n">linear_input_size</span> <span class="o">=</span> <span class="n">convw</span> <span class="o">*</span> <span class="n">convh</span> <span class="o">*</span> <span class="mi">32</span>
<span class="bp">self</span><span class="o">.</span><span class="n">head</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="n">linear_input_size</span><span class="p">,</span> <span class="n">outputs</span><span class="p">)</span>
<span class="c1"># 최적화 중에 다음 행동을 결정하기 위해서 하나의 요소 또는 배치를 이용해 호촐됩니다.</span>
<span class="c1"># ([[left0exp,right0exp]...]) 를 반환합니다.</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">x</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">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">bn1</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">conv1</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">bn2</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">conv2</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">bn3</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">conv3</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">head</span><span class="p">(</span><span class="n">x</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">x</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="o">-</span><span class="mi">1</span><span class="p">))</span>
</pre></div>
</div>
</div>
<div class="section" id="id3">
<h3>입력 추출<a class="headerlink" href="#id3" title="Permalink to this headline">¶</a></h3>
<p>아래 코드는 환경에서 렌더링 된 이미지를 추출하고 처리하는 유틸리티입니다.
이미지 변환을 쉽게 구성할 수 있는 <code class="docutils literal notranslate"><span class="pre">torchvision</span></code> 패키지를 사용합니다.
셀(cell)을 실행하면 추출한 예제 패치가 표시됩니다.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">resize</span> <span class="o">=</span> <span class="n">T</span><span class="o">.</span><span class="n">Compose</span><span class="p">([</span><span class="n">T</span><span class="o">.</span><span class="n">ToPILImage</span><span class="p">(),</span>
<span class="n">T</span><span class="o">.</span><span class="n">Resize</span><span class="p">(</span><span class="mi">40</span><span class="p">,</span> <span class="n">interpolation</span><span class="o">=</span><span class="n">Image</span><span class="o">.</span><span class="n">CUBIC</span><span class="p">),</span>
<span class="n">T</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">()])</span>
<span class="k">def</span> <span class="nf">get_cart_location</span><span class="p">(</span><span class="n">screen_width</span><span class="p">):</span>
<span class="n">world_width</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">x_threshold</span> <span class="o">*</span> <span class="mi">2</span>
<span class="n">scale</span> <span class="o">=</span> <span class="n">screen_width</span> <span class="o">/</span> <span class="n">world_width</span>
<span class="k">return</span> <span class="nb">int</span><span class="p">(</span><span class="n">env</span><span class="o">.</span><span class="n">state</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">*</span> <span class="n">scale</span> <span class="o">+</span> <span class="n">screen_width</span> <span class="o">/</span> <span class="mf">2.0</span><span class="p">)</span> <span class="c1"># MIDDLE OF CART</span>
<span class="k">def</span> <span class="nf">get_screen</span><span class="p">():</span>
<span class="c1"># gym이 요청한 화면은 400x600x3 이지만, 가끔 800x1200x3 처럼 큰 경우가 있습니다.</span>
<span class="c1"># 이것을 Torch order (CHW)로 변환한다.</span>
<span class="n">screen</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">render</span><span class="p">(</span><span class="n">mode</span><span class="o">=</span><span class="s1">'rgb_array'</span><span class="p">)</span><span class="o">.</span><span class="n">transpose</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="mi">1</span><span class="p">))</span>
<span class="c1"># 카트는 아래쪽에 있으므로 화면의 상단과 하단을 제거하십시오.</span>
<span class="n">_</span><span class="p">,</span> <span class="n">screen_height</span><span class="p">,</span> <span class="n">screen_width</span> <span class="o">=</span> <span class="n">screen</span><span class="o">.</span><span class="n">shape</span>
<span class="n">screen</span> <span class="o">=</span> <span class="n">screen</span><span class="p">[:,</span> <span class="nb">int</span><span class="p">(</span><span class="n">screen_height</span><span class="o">*</span><span class="mf">0.4</span><span class="p">):</span><span class="nb">int</span><span class="p">(</span><span class="n">screen_height</span> <span class="o">*</span> <span class="mf">0.8</span><span class="p">)]</span>
<span class="n">view_width</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">screen_width</span> <span class="o">*</span> <span class="mf">0.6</span><span class="p">)</span>
<span class="n">cart_location</span> <span class="o">=</span> <span class="n">get_cart_location</span><span class="p">(</span><span class="n">screen_width</span><span class="p">)</span>
<span class="k">if</span> <span class="n">cart_location</span> <span class="o"><</span> <span class="n">view_width</span> <span class="o">//</span> <span class="mi">2</span><span class="p">:</span>
<span class="n">slice_range</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">view_width</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">cart_location</span> <span class="o">></span> <span class="p">(</span><span class="n">screen_width</span> <span class="o">-</span> <span class="n">view_width</span> <span class="o">//</span> <span class="mi">2</span><span class="p">):</span>
<span class="n">slice_range</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="o">-</span><span class="n">view_width</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">slice_range</span> <span class="o">=</span> <span class="nb">slice</span><span class="p">(</span><span class="n">cart_location</span> <span class="o">-</span> <span class="n">view_width</span> <span class="o">//</span> <span class="mi">2</span><span class="p">,</span>
<span class="n">cart_location</span> <span class="o">+</span> <span class="n">view_width</span> <span class="o">//</span> <span class="mi">2</span><span class="p">)</span>
<span class="c1"># 카트를 중심으로 정사각형 이미지가 되도록 가장자리를 제거하십시오.</span>
<span class="n">screen</span> <span class="o">=</span> <span class="n">screen</span><span class="p">[:,</span> <span class="p">:,</span> <span class="n">slice_range</span><span class="p">]</span>
<span class="c1"># float 으로 변환하고, rescale 하고, torch tensor 로 변환하십시오.</span>
<span class="c1"># (이것은 복사를 필요로하지 않습니다)</span>
<span class="n">screen</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">ascontiguousarray</span><span class="p">(</span><span class="n">screen</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">np</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span> <span class="o">/</span> <span class="mi">255</span>
<span class="n">screen</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">screen</span><span class="p">)</span>
<span class="c1"># 크기를 수정하고 배치 차원(BCHW)을 추가하십시오.</span>
<span class="k">return</span> <span class="n">resize</span><span class="p">(</span><span class="n">screen</span><span class="p">)</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">figure</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">imshow</span><span class="p">(</span><span class="n">get_screen</span><span class="p">()</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">squeeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">permute</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="o">.</span><span class="n">numpy</span><span class="p">(),</span>
<span class="n">interpolation</span><span class="o">=</span><span class="s1">'none'</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">title</span><span class="p">(</span><span class="s1">'Example extracted screen'</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 class="section" id="id4">
<h2>학습<a class="headerlink" href="#id4" title="Permalink to this headline">¶</a></h2>
<div class="section" id="id5">
<h3>하이퍼 파라미터와 유틸리티<a class="headerlink" href="#id5" title="Permalink to this headline">¶</a></h3>
<p>이 셀은 모델과 최적화기를 인스턴스화하고 일부 유틸리티를 정의합니다:</p>
<ul class="simple">
<li><code class="docutils literal notranslate"><span class="pre">select_action</span></code> - Epsilon Greedy 정책에 따라 행동을 선택합니다.
간단히 말해서, 가끔 모델을 사용하여 행동을 선택하고 때로는 단지 하나를
균일하게 샘플링할 것입니다. 임의의 액션을 선택할 확률은
<code class="docutils literal notranslate"><span class="pre">EPS_START</span></code> 에서 시작해서 <code class="docutils literal notranslate"><span class="pre">EPS_END</span></code> 를 향해 지수적으로 감소할 것입니다.
<code class="docutils literal notranslate"><span class="pre">EPS_DECAY</span></code> 는 감쇠 속도를 제어합니다.</li>
<li><code class="docutils literal notranslate"><span class="pre">plot_durations</span></code> - 지난 100개 에피소드의 평균(공식 평가에서 사용 된 수치)에 따른
에피소드의 지속을 도표로 그리기 위한 헬퍼. 도표는 기본 훈련 루프가
포함 된 셀 밑에 있으며, 매 에피소드마다 업데이트됩니다.</li>
</ul>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">BATCH_SIZE</span> <span class="o">=</span> <span class="mi">128</span>
<span class="n">GAMMA</span> <span class="o">=</span> <span class="mf">0.999</span>
<span class="n">EPS_START</span> <span class="o">=</span> <span class="mf">0.9</span>
<span class="n">EPS_END</span> <span class="o">=</span> <span class="mf">0.05</span>
<span class="n">EPS_DECAY</span> <span class="o">=</span> <span class="mi">200</span>
<span class="n">TARGET_UPDATE</span> <span class="o">=</span> <span class="mi">10</span>
<span class="c1"># AI gym에서 반환된 형태를 기반으로 계층을 초기화 하도록 화면의 크기를</span>
<span class="c1"># 가져옵니다. 이 시점에 일반적으로 3x40x90 에 가깝습니다.</span>
<span class="c1"># 이 크기는 get_screen()에서 고정, 축소된 렌더 버퍼의 결과입니다.</span>
<span class="n">init_screen</span> <span class="o">=</span> <span class="n">get_screen</span><span class="p">()</span>
<span class="n">_</span><span class="p">,</span> <span class="n">_</span><span class="p">,</span> <span class="n">screen_height</span><span class="p">,</span> <span class="n">screen_width</span> <span class="o">=</span> <span class="n">init_screen</span><span class="o">.</span><span class="n">shape</span>
<span class="c1"># gym 행동 공간에서 행동의 숫자를 얻습니다.</span>
<span class="n">n_actions</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">action_space</span><span class="o">.</span><span class="n">n</span>
<span class="n">policy_net</span> <span class="o">=</span> <span class="n">DQN</span><span class="p">(</span><span class="n">screen_height</span><span class="p">,</span> <span class="n">screen_width</span><span class="p">,</span> <span class="n">n_actions</span><span class="p">)</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">target_net</span> <span class="o">=</span> <span class="n">DQN</span><span class="p">(</span><span class="n">screen_height</span><span class="p">,</span> <span class="n">screen_width</span><span class="p">,</span> <span class="n">n_actions</span><span class="p">)</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">target_net</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">policy_net</span><span class="o">.</span><span class="n">state_dict</span><span class="p">())</span>
<span class="n">target_net</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">RMSprop</span><span class="p">(</span><span class="n">policy_net</span><span class="o">.</span><span class="n">parameters</span><span class="p">())</span>
<span class="n">memory</span> <span class="o">=</span> <span class="n">ReplayMemory</span><span class="p">(</span><span class="mi">10000</span><span class="p">)</span>
<span class="n">steps_done</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">def</span> <span class="nf">select_action</span><span class="p">(</span><span class="n">state</span><span class="p">):</span>
<span class="k">global</span> <span class="n">steps_done</span>
<span class="n">sample</span> <span class="o">=</span> <span class="n">random</span><span class="o">.</span><span class="n">random</span><span class="p">()</span>
<span class="n">eps_threshold</span> <span class="o">=</span> <span class="n">EPS_END</span> <span class="o">+</span> <span class="p">(</span><span class="n">EPS_START</span> <span class="o">-</span> <span class="n">EPS_END</span><span class="p">)</span> <span class="o">*</span> \
<span class="n">math</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="o">-</span><span class="mf">1.</span> <span class="o">*</span> <span class="n">steps_done</span> <span class="o">/</span> <span class="n">EPS_DECAY</span><span class="p">)</span>
<span class="n">steps_done</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="k">if</span> <span class="n">sample</span> <span class="o">></span> <span class="n">eps_threshold</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="c1"># t.max (1)은 각 행의 가장 큰 열 값을 반환합니다.</span>
<span class="c1"># 최대 결과의 두번째 열은 최대 요소의 주소값이므로,</span>
<span class="c1"># 기대 보상이 더 큰 행동을 선택할 수 있습니다.</span>
<span class="k">return</span> <span class="n">policy_net</span><span class="p">(</span><span class="n">state</span><span class="p">)</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="mi">1</span><span class="p">)[</span><span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([[</span><span class="n">random</span><span class="o">.</span><span class="n">randrange</span><span class="p">(</span><span class="n">n_actions</span><span class="p">)]],</span> <span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">long</span><span class="p">)</span>
<span class="n">episode_durations</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">def</span> <span class="nf">plot_durations</span><span class="p">():</span>
<span class="n">plt</span><span class="o">.</span><span class="n">figure</span><span class="p">(</span><span class="mi">2</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">clf</span><span class="p">()</span>
<span class="n">durations_t</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="n">episode_durations</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">title</span><span class="p">(</span><span class="s1">'Training...'</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">xlabel</span><span class="p">(</span><span class="s1">'Episode'</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">ylabel</span><span class="p">(</span><span class="s1">'Duration'</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">plot</span><span class="p">(</span><span class="n">durations_t</span><span class="o">.</span><span class="n">numpy</span><span class="p">())</span>
<span class="c1"># 100개의 에피소드 평균을 가져 와서 도표 그리기</span>
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">durations_t</span><span class="p">)</span> <span class="o">>=</span> <span class="mi">100</span><span class="p">:</span>
<span class="n">means</span> <span class="o">=</span> <span class="n">durations_t</span><span class="o">.</span><span class="n">unfold</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="mi">100</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">mean</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
<span class="n">means</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">((</span><span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mi">99</span><span class="p">),</span> <span class="n">means</span><span class="p">))</span>
<span class="n">plt</span><span class="o">.</span><span class="n">plot</span><span class="p">(</span><span class="n">means</span><span class="o">.</span><span class="n">numpy</span><span class="p">())</span>
<span class="n">plt</span><span class="o">.</span><span class="n">pause</span><span class="p">(</span><span class="mf">0.001</span><span class="p">)</span> <span class="c1"># 도표가 업데이트되도록 잠시 멈춤</span>
<span class="k">if</span> <span class="n">is_ipython</span><span class="p">:</span>
<span class="n">display</span><span class="o">.</span><span class="n">clear_output</span><span class="p">(</span><span class="n">wait</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">display</span><span class="o">.</span><span class="n">display</span><span class="p">(</span><span class="n">plt</span><span class="o">.</span><span class="n">gcf</span><span class="p">())</span>
</pre></div>
</div>
</div>
<div class="section" id="id6">
<h3>학습 루프<a class="headerlink" href="#id6" title="Permalink to this headline">¶</a></h3>
<p>최종적으로 모델 학습을 위한 코드.</p>
<p>여기서, 최적화의 한 단계를 수행하는 <code class="docutils literal notranslate"><span class="pre">optimize_model</span></code> 함수를 찾을 수 있습니다.
먼저 배치 하나를 샘플링하고 모든 Tensor를 하나로 연결하고
<span class="math">\(Q(s_t, a_t)\)</span> 와 <span class="math">\(V(s_{t+1}) = \max_a Q(s_{t+1}, a)\)</span> 를 계산하고
그것들을 손실로 합칩니다. 우리가 설정한 정의에 따르면 만약 <span class="math">\(s\)</span> 가
마지막 상태라면 <span class="math">\(V(s) = 0\)</span> 입니다.
또한 안정성 추가 위한 <span class="math">\(V(s_{t+1})\)</span> 계산을 위해 목표 네트워크를 사용합니다.
목표 네트워크는 대부분의 시간 동결 상태로 유지되지만, 가끔 정책
네트워크의 가중치로 업데이트됩니다.
이것은 대개 설정한 스텝 숫자이지만 단순화를 위해 에피소드를 사용합니다.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">optimize_model</span><span class="p">():</span>
<span class="k">if</span> <span class="nb">len</span><span class="p">(</span><span class="n">memory</span><span class="p">)</span> <span class="o"><</span> <span class="n">BATCH_SIZE</span><span class="p">:</span>
<span class="k">return</span>
<span class="n">transitions</span> <span class="o">=</span> <span class="n">memory</span><span class="o">.</span><span class="n">sample</span><span class="p">(</span><span class="n">BATCH_SIZE</span><span class="p">)</span>
<span class="c1"># Transpose the batch (see https://stackoverflow.com/a/19343/3343043 for</span>
<span class="c1"># detailed explanation). 이것은 batch-array의 Transitions을 Transition의 batch-arrays로</span>
<span class="c1"># 전환합니다.</span>
<span class="n">batch</span> <span class="o">=</span> <span class="n">Transition</span><span class="p">(</span><span class="o">*</span><span class="nb">zip</span><span class="p">(</span><span class="o">*</span><span class="n">transitions</span><span class="p">))</span>
<span class="c1"># 최종이 아닌 상태의 마스크를 계산하고 배치 요소를 연결합니다</span>
<span class="c1"># (최종 상태는 시뮬레이션이 종료 된 이후의 상태)</span>
<span class="n">non_final_mask</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="nb">tuple</span><span class="p">(</span><span class="nb">map</span><span class="p">(</span><span class="k">lambda</span> <span class="n">s</span><span class="p">:</span> <span class="n">s</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">,</span>
<span class="n">batch</span><span class="o">.</span><span class="n">next_state</span><span class="p">)),</span> <span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">bool</span><span class="p">)</span>
<span class="n">non_final_next_states</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">([</span><span class="n">s</span> <span class="k">for</span> <span class="n">s</span> <span class="ow">in</span> <span class="n">batch</span><span class="o">.</span><span class="n">next_state</span>
<span class="k">if</span> <span class="n">s</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">])</span>
<span class="n">state_batch</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">state</span><span class="p">)</span>
<span class="n">action_batch</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">action</span><span class="p">)</span>
<span class="n">reward_batch</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">batch</span><span class="o">.</span><span class="n">reward</span><span class="p">)</span>
<span class="c1"># Q(s_t, a) 계산 - 모델이 Q(s_t)를 계산하고, 취한 행동의 열을 선택합니다.</span>
<span class="c1"># 이들은 policy_net에 따라 각 배치 상태에 대해 선택된 행동입니다.</span>
<span class="n">state_action_values</span> <span class="o">=</span> <span class="n">policy_net</span><span class="p">(</span><span class="n">state_batch</span><span class="p">)</span><span class="o">.</span><span class="n">gather</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">action_batch</span><span class="p">)</span>
<span class="c1"># 모든 다음 상태를 위한 V(s_{t+1}) 계산</span>
<span class="c1"># non_final_next_states의 행동들에 대한 기대값은 "이전" target_net을 기반으로 계산됩니다.</span>
<span class="c1"># max(1)[0]으로 최고의 보상을 선택하십시오.</span>
<span class="c1"># 이것은 마스크를 기반으로 병합되어 기대 상태 값을 갖거나 상태가 최종인 경우 0을 갖습니다.</span>
<span class="n">next_state_values</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">BATCH_SIZE</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">)</span>
<span class="n">next_state_values</span><span class="p">[</span><span class="n">non_final_mask</span><span class="p">]</span> <span class="o">=</span> <span class="n">target_net</span><span class="p">(</span><span class="n">non_final_next_states</span><span class="p">)</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="mi">1</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">detach</span><span class="p">()</span>
<span class="c1"># 기대 Q 값 계산</span>
<span class="n">expected_state_action_values</span> <span class="o">=</span> <span class="p">(</span><span class="n">next_state_values</span> <span class="o">*</span> <span class="n">GAMMA</span><span class="p">)</span> <span class="o">+</span> <span class="n">reward_batch</span>
<span class="c1"># Huber 손실 계산</span>
<span class="n">criterion</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">SmoothL1Loss</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">state_action_values</span><span class="p">,</span> <span class="n">expected_state_action_values</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">))</span>
<span class="c1"># 모델 최적화</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</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="k">for</span> <span class="n">param</span> <span class="ow">in</span> <span class="n">policy_net</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">grad</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">clamp_</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>아래에서 주요 학습 루프를 찾을 수 있습니다. 처음으로 환경을
재설정하고 <code class="docutils literal notranslate"><span class="pre">상태</span></code> Tensor를 초기화합니다. 그런 다음 행동을
샘플링하고, 그것을 실행하고, 다음 화면과 보상(항상 1)을 관찰하고,
모델을 한 번 최적화합니다. 에피소드가 끝나면 (모델이 실패)
루프를 다시 시작합니다.</p>
<p>아래에서 <cite>num_episodes</cite> 는 작게 설정됩니다. 노트북을 다운받고
의미있는 개선을 위해서 300 이상의 더 많은 에피소드를 실행해 보십시오.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">num_episodes</span> <span class="o">=</span> <span class="mi">50</span>
<span class="k">for</span> <span class="n">i_episode</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">num_episodes</span><span class="p">):</span>
<span class="c1"># 환경과 상태 초기화</span>
<span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">()</span>
<span class="n">last_screen</span> <span class="o">=</span> <span class="n">get_screen</span><span class="p">()</span>
<span class="n">current_screen</span> <span class="o">=</span> <span class="n">get_screen</span><span class="p">()</span>
<span class="n">state</span> <span class="o">=</span> <span class="n">current_screen</span> <span class="o">-</span> <span class="n">last_screen</span>
<span class="k">for</span> <span class="n">t</span> <span class="ow">in</span> <span class="n">count</span><span class="p">():</span>
<span class="c1"># 행동 선택과 수행</span>
<span class="n">action</span> <span class="o">=</span> <span class="n">select_action</span><span class="p">(</span><span class="n">state</span><span class="p">)</span>
<span class="n">_</span><span class="p">,</span> <span class="n">reward</span><span class="p">,</span> <span class="n">done</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">action</span><span class="o">.</span><span class="n">item</span><span class="p">())</span>
<span class="n">reward</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="n">reward</span><span class="p">],</span> <span class="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">)</span>
<span class="c1"># 새로운 상태 관찰</span>
<span class="n">last_screen</span> <span class="o">=</span> <span class="n">current_screen</span>
<span class="n">current_screen</span> <span class="o">=</span> <span class="n">get_screen</span><span class="p">()</span>
<span class="k">if</span> <span class="ow">not</span> <span class="n">done</span><span class="p">:</span>
<span class="n">next_state</span> <span class="o">=</span> <span class="n">current_screen</span> <span class="o">-</span> <span class="n">last_screen</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">next_state</span> <span class="o">=</span> <span class="kc">None</span>
<span class="c1"># 메모리에 변이 저장</span>
<span class="n">memory</span><span class="o">.</span><span class="n">push</span><span class="p">(</span><span class="n">state</span><span class="p">,</span> <span class="n">action</span><span class="p">,</span> <span class="n">next_state</span><span class="p">,</span> <span class="n">reward</span><span class="p">)</span>
<span class="c1"># 다음 상태로 이동</span>
<span class="n">state</span> <span class="o">=</span> <span class="n">next_state</span>
<span class="c1"># (정책 네트워크에서) 최적화 한단계 수행</span>
<span class="n">optimize_model</span><span class="p">()</span>
<span class="k">if</span> <span class="n">done</span><span class="p">:</span>
<span class="n">episode_durations</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">t</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">plot_durations</span><span class="p">()</span>
<span class="k">break</span>
<span class="c1"># 목표 네트워크 업데이트, 모든 웨이트와 바이어스 복사</span>
<span class="k">if</span> <span class="n">i_episode</span> <span class="o">%</span> <span class="n">TARGET_UPDATE</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">target_net</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">policy_net</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="s1">'Complete'</span><span class="p">)</span>
<span class="n">env</span><span class="o">.</span><span class="n">render</span><span class="p">()</span>
<span class="n">env</span><span class="o">.</span><span class="n">close</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">show</span><span class="p">()</span>
</pre></div>
</div>
<p>다음은 전체 결과 데이터 흐름을 보여주는 다이어그램입니다.</p>
<div class="figure">
<img alt="../_images/reinforcement_learning_diagram.jpg" src="../_images/reinforcement_learning_diagram.jpg" />
</div>
<p>행동은 무작위 또는 정책에 따라 선택되어, gym 환경에서 다음 단계 샘플을 가져옵니다.
결과를 재현 메모리에 저장하고 모든 반복에서 최적화 단계를 실행합니다.
최적화는 재현 메모리에서 무작위 배치를 선택하여 새 정책을 학습합니다.
“이전” target_net은 최적화에서 기대 Q 값을 계산하는 데에도 사용되고,
최신 상태를 유지하기 위해 가끔 업데이트됩니다.</p>
<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 0 minutes 0.000 seconds)</p>
<div class="sphx-glr-footer class sphx-glr-footer-example docutils container" id="sphx-glr-download-intermediate-reinforcement-q-learning-py">
<div class="sphx-glr-download docutils container">
<a class="reference download internal" download="" href="../_downloads/d080af12070c2b2dcebdb06d8bf2a119/reinforcement_q_learning.py"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Python</span> <span class="pre">source</span> <span class="pre">code:</span> <span class="pre">reinforcement_q_learning.py</span></code></a></div>
<div class="sphx-glr-download docutils container">
<a class="reference download internal" download="" href="../_downloads/d1777ec944f083ed073b3fd6a721b1d2/reinforcement_q_learning.ipynb"><code class="xref download docutils literal notranslate"><span class="pre">Download</span> <span class="pre">Jupyter</span> <span class="pre">notebook:</span> <span class="pre">reinforcement_q_learning.ipynb</span></code></a></div>
</div>
<p class="sphx-glr-signature"><a class="reference external" href="https://sphinx-gallery.readthedocs.io">Gallery generated by Sphinx-Gallery</a></p>
</div>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="mario_rl_tutorial.html" class="btn btn-neutral float-right" title="Train a Mario-playing RL Agent" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="../beginner/translation_transformer.html" class="btn btn-neutral" title="nn.Transformer와 torchtext로 언어 번역하기" 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="#">강화 학습 (DQN) 튜토리얼</a><ul>
<li><a class="reference internal" href="#replay-memory">재현 메모리(Replay Memory)</a></li>
<li><a class="reference internal" href="#id2">DQN 알고리즘</a><ul>
<li><a class="reference internal" href="#q">Q-네트워크</a></li>
<li><a class="reference internal" href="#id3">입력 추출</a></li>
</ul>
</li>
<li><a class="reference internal" href="#id4">학습</a><ul>
<li><a class="reference internal" href="#id5">하이퍼 파라미터와 유틸리티</a></li>
<li><a class="reference internal" href="#id6">학습 루프</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
});
});