-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathdqn_with_rnn_tutorial.html
1306 lines (1134 loc) · 95.6 KB
/
dqn_with_rnn_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="ko" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="ko" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta property="og:title" content="Recurrent DQN: Training recurrent policies" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/intermediate/dqn_with_rnn_tutorial.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="Author: Vincent Moens What you will learn How to incorporating an RNN in an actor in TorchRL, How to use that memory-based policy with a replay buffer and a loss module. Prerequisites PyTorch v2.0.0, gym[mujoco], tqdm. Overview: Memory-based policies are crucial not only when the observations are..." />
<meta property="og:image" content="https://tutorials.pytorch.kr/_static/logos/logo-kr-sm-dark.png" />
<meta property="og:image:alt" content="PyTorch Tutorials KR" />
<meta name="description" content="Author: Vincent Moens What you will learn How to incorporating an RNN in an actor in TorchRL, How to use that memory-based policy with a replay buffer and a loss module. Prerequisites PyTorch v2.0.0, gym[mujoco], tqdm. Overview: Memory-based policies are crucial not only when the observations are..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Recurrent DQN: Training recurrent policies — 파이토치 한국어 튜토리얼 (PyTorch tutorials in Korean)</title>
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<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/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.10/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx-design.5ea377869091fd0449014c60fc090103.min.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="../_static/css/custom.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/custom2.css" type="text/css" />
<link rel="index" title="색인" href="../genindex.html" />
<link rel="search" title="검색" href="../search.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">
2.3.1+cu121
</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" role="heading"><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" role="heading"><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" role="heading"><span class="caption-text">Introduction to PyTorch on YouTube</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt.html">PyTorch 소개 - YouTube 시리즈</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">PyTorch 소개</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Pytorch Tensor 소개</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" role="heading"><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" role="heading"><span class="caption-text">이미지/비디오</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision_tutorial.html">TorchVision Object Detection Finetuning Tutorial</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 위해 비전 트랜스포머(Vision Transformer) 모델 최적화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="tiatoolbox_tutorial.html">Whole Slide Image Classification Using PyTorch and TIAToolbox</a></li>
</ul>
<p class="caption" role="heading"><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="text_to_speech_with_torchaudio.html">Text-to-speech with Tacotron2</a></li>
<li class="toctree-l1"><a class="reference internal" href="forced_alignment_with_torchaudio_tutorial.html">wav2vec2을 이용한 강제 정렬</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">텍스트</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/bettertransformer_tutorial.html">Fast Transformer Inference with Better Transformer</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"><code class="docutils literal notranslate"><span class="pre">nn.Transformer</span></code> 와 torchtext로 언어 번역하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/torchtext_custom_dataset_tutorial.html">Preprocess custom text dataset using Torchtext</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">백엔드</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
</ul>
<p class="caption" role="heading"><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="reinforcement_ppo.html">Reinforcement Learning (PPO) with TorchRL Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="mario_rl_tutorial.html">마리오 게임 RL 에이전트로 학습하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/pendulum.html">Pendulum: Writing your environment and transforms with TorchRL</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch 모델을 프로덕션 환경에 배포하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/onnx/intro_onnx.html">Introduction to ONNX</a></li>
<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>
<li class="toctree-l1"><a class="reference internal" href="realtime_rpi.html">Raspberry Pi 4 에서 실시간 추론(Inference) (30fps!)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch 프로파일링</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch 모듈 프로파일링하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_intro_tutorial.html">Introduction to Holistic Trace Analysis</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hta_trace_diff_tutorial.html">Trace Diff using Holistic Trace Analysis</a></li>
</ul>
<p class="caption" role="heading"><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" role="heading"><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="jacobians_hessians.html">Jacobians, Hessians, hvp, vhp, and more: composing function transforms</a></li>
<li class="toctree-l1"><a class="reference internal" href="ensembling.html">모델 앙상블</a></li>
<li class="toctree-l1"><a class="reference internal" href="per_sample_grads.html">Per-sample-gradients</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" role="heading"><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>
<li class="toctree-l1"><a class="reference internal" href="../advanced/privateuseone.html">Facilitating New Backend Integration by PrivateUse1</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">모델 최적화</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch 모듈 프로파일링하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_profiler_tutorial.html">텐서보드를 이용한 파이토치 프로파일러</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Ray Tune을 사용한 하이퍼파라미터 튜닝</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">배포를 위해 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">(베타) PyTorch에서 Eager Mode를 이용한 정적 양자화</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex.html">Grokking PyTorch Intel CPU performance from first principles</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchserve_with_ipex_2.html">Grokking PyTorch Intel CPU performance from first principles (Part 2)</a></li>
<li class="toctree-l1"><a class="reference internal" href="nvfuser_intro_tutorial.html">Getting Started - Accelerate Your Scripts with nvFuser</a></li>
<li class="toctree-l1"><a class="reference internal" href="ax_multiobjective_nas_tutorial.html">Multi-Objective NAS with Ax</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_compile_tutorial.html">Introduction to <code class="docutils literal notranslate"><span class="pre">torch.compile</span></code></a></li>
<li class="toctree-l1"><a class="reference internal" href="inductor_debug_cpu.html">Inductor CPU backend debugging and profiling</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html">(Beta) Scaled Dot Product Attention (SDPA)로 고성능 트랜스포머(Transformers) 구현하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#torch-compile-sdpa"><code class="docutils literal notranslate"><span class="pre">torch.compile</span></code> 과 함께 SDPA 사용하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#sdpa-atteition-bias">SDPA를 <code class="docutils literal notranslate"><span class="pre">atteition.bias</span></code> 하위 클래스와 사용하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#id8">결론</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/knowledge_distillation_tutorial.html">Knowledge Distillation Tutorial</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">병렬 및 분산 학습</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../distributed/home.html">Distributed and Parallel Training Tutorials</a></li>
<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="../beginner/ddp_series_intro.html">Distributed Data Parallel in PyTorch - Video Tutorials</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="FSDP_tutorial.html">Getting Started with Fully Sharded Data Parallel(FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="FSDP_adavnced_tutorial.html">Advanced Model Training with Fully Sharded Data Parallel (FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="TP_tutorial.html">Large Scale Transformer model training with Tensor Parallel (TP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="process_group_cpp_extension_tutorial.html">Cpp 확장을 사용한 프로세스 그룹 백엔드 사용자 정의</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="../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" role="heading"><span class="caption-text">Edge with ExecuTorch</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/export-to-executorch-tutorial.html">Exporting to ExecuTorch Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href=" https://pytorch.org/executorch/stable/running-a-model-cpp-tutorial.html">Running an ExecuTorch Model in C++ Tutorial</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/tutorials/sdk-integration-tutorial.html">Using the ExecuTorch SDK to Profile a Model</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-ios.html">Building an ExecuTorch iOS Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/demo-apps-android.html">Building an ExecuTorch Android Demo App</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/executorch/stable/examples-end-to-end-to-lower-model-to-delegate.html">Lowering a Model as a Delegate</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">추천 시스템</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchrec_tutorial.html">TorchRec 소개</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/sharding.html">Exploring TorchRec sharding</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Multimodality</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/flava_finetuning_tutorial.html">TorchMultimodal 튜토리얼: FLAVA 미세조정</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>Recurrent DQN: Training recurrent policies</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/dqn_with_rnn_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/dqn_with_rnn_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="sphx-glr-download-link-note admonition note">
<p class="admonition-title">참고</p>
<p>Click <a class="reference internal" href="#sphx-glr-download-intermediate-dqn-with-rnn-tutorial-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="recurrent-dqn-training-recurrent-policies">
<span id="sphx-glr-intermediate-dqn-with-rnn-tutorial-py"></span><h1>Recurrent DQN: Training recurrent policies<a class="headerlink" href="#recurrent-dqn-training-recurrent-policies" title="이 제목에 대한 퍼머링크">¶</a></h1>
<p><strong>Author</strong>: <a class="reference external" href="https://github.com/vmoens">Vincent Moens</a></p>
<div class="sd-container-fluid sd-sphinx-override sd-mb-4 docutils">
<div class="sd-row sd-row-cols-2 sd-row-cols-xs-2 sd-row-cols-sm-2 sd-row-cols-md-2 sd-row-cols-lg-2 docutils">
<div class="sd-col sd-d-flex-row docutils">
<div class="sd-card sd-sphinx-override sd-w-100 sd-shadow-sm docutils">
<div class="sd-card-body docutils">
<div class="sd-card-title sd-font-weight-bold docutils">
<svg version="1.1" width="1.0em" height="1.0em" class="sd-octicon sd-octicon-mortar-board" viewBox="0 0 16 16" aria-hidden="true"><path d="M7.693 1.066a.747.747 0 0 1 .614 0l7.25 3.25a.75.75 0 0 1 0 1.368L13 6.831v2.794c0 1.024-.81 1.749-1.66 2.173-.893.447-2.075.702-3.34.702-.278 0-.55-.012-.816-.036a.75.75 0 0 1 .133-1.494c.22.02.45.03.683.03 1.082 0 2.025-.221 2.67-.543.69-.345.83-.682.83-.832V7.503L8.307 8.934a.747.747 0 0 1-.614 0L4 7.28v1.663c.296.105.575.275.812.512.438.438.688 1.059.688 1.796v3a.75.75 0 0 1-.75.75h-3a.75.75 0 0 1-.75-.75v-3c0-.737.25-1.358.688-1.796.237-.237.516-.407.812-.512V6.606L.443 5.684a.75.75 0 0 1 0-1.368ZM2.583 5 8 7.428 13.416 5 8 2.572ZM2.5 11.25v2.25H4v-2.25c0-.388-.125-.611-.25-.735a.697.697 0 0 0-.5-.203.707.707 0 0 0-.5.203c-.125.124-.25.347-.25.735Z"></path></svg> What you will learn</div>
<ul class="simple">
<li><p class="sd-card-text">How to incorporating an RNN in an actor in TorchRL</p></li>
<li><p class="sd-card-text">How to use that memory-based policy with a replay buffer and a loss module</p></li>
</ul>
</div>
</div>
</div>
<div class="sd-col sd-d-flex-row docutils">
<div class="sd-card sd-sphinx-override sd-w-100 sd-shadow-sm docutils">
<div class="sd-card-body docutils">
<div class="sd-card-title sd-font-weight-bold docutils">
<svg version="1.1" width="1.0em" height="1.0em" class="sd-octicon sd-octicon-list-unordered" viewBox="0 0 16 16" aria-hidden="true"><path d="M5.75 2.5h8.5a.75.75 0 0 1 0 1.5h-8.5a.75.75 0 0 1 0-1.5Zm0 5h8.5a.75.75 0 0 1 0 1.5h-8.5a.75.75 0 0 1 0-1.5Zm0 5h8.5a.75.75 0 0 1 0 1.5h-8.5a.75.75 0 0 1 0-1.5ZM2 14a1 1 0 1 1 0-2 1 1 0 0 1 0 2Zm1-6a1 1 0 1 1-2 0 1 1 0 0 1 2 0ZM2 4a1 1 0 1 1 0-2 1 1 0 0 1 0 2Z"></path></svg> Prerequisites</div>
<ul class="simple">
<li><p class="sd-card-text">PyTorch v2.0.0</p></li>
<li><p class="sd-card-text">gym[mujoco]</p></li>
<li><p class="sd-card-text">tqdm</p></li>
</ul>
</div>
</div>
</div>
</div>
</div>
<div class="section" id="overview">
<h2>Overview<a class="headerlink" href="#overview" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>Memory-based policies are crucial not only when the observations are partially
observable but also when the time dimension must be taken into account to
make informed decisions.</p>
<p>Recurrent neural network have long been a popular tool for memory-based
policies. The idea is to keep a recurrent state in memory between two
consecutive steps, and use this as an input to the policy along with the
current observation.</p>
<p>This tutorial shows how to incorporate an RNN in a policy using TorchRL.</p>
<p>Key learnings:</p>
<ul class="simple">
<li><p>Incorporating an RNN in an actor in TorchRL;</p></li>
<li><p>Using that memory-based policy with a replay buffer and a loss module.</p></li>
</ul>
<p>The core idea of using RNNs in TorchRL is to use TensorDict as a data carrier
for the hidden states from one step to another. We’ll build a policy that
reads the previous recurrent state from the current TensorDict, and writes the
current recurrent states in the TensorDict of the next state:</p>
<div class="figure align-default">
<img alt="Data collection with a recurrent policy" src="../_images/rollout_recurrent.png" />
</div>
<p>As this figure shows, our environment populates the TensorDict with zeroed recurrent
states which are read by the policy together with the observation to produce an
action, and recurrent states that will be used for the next step.
When the <code class="xref py py-func docutils literal notranslate"><span class="pre">step_mdp()</span></code> function is called, the recurrent states
from the next state are brought to the current TensorDict. Let’s see how this
is implemented in practice.</p>
<p>If you are running this in Google Colab, make sure you install the following dependencies:</p>
<div class="highlight-bash notranslate"><div class="highlight"><pre><span></span>!pip3<span class="w"> </span>install<span class="w"> </span>torchrl
!pip3<span class="w"> </span>install<span class="w"> </span>gym<span class="o">[</span>mujoco<span class="o">]</span>
!pip3<span class="w"> </span>install<span class="w"> </span>tqdm
</pre></div>
</div>
</div>
<div class="section" id="setup">
<h2>Setup<a class="headerlink" href="#setup" title="이 제목에 대한 퍼머링크">¶</a></h2>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">tqdm</span>
<span class="kn">from</span> <span class="nn">tensordict.nn</span> <span class="kn">import</span> <span class="n">TensorDictModule</span> <span class="k">as</span> <span class="n">Mod</span><span class="p">,</span> <span class="n">TensorDictSequential</span> <span class="k">as</span> <span class="n">Seq</span>
<span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
<span class="kn">from</span> <span class="nn">torchrl.collectors</span> <span class="kn">import</span> <span class="n">SyncDataCollector</span>
<span class="kn">from</span> <span class="nn">torchrl.data</span> <span class="kn">import</span> <span class="n">LazyMemmapStorage</span><span class="p">,</span> <span class="n">TensorDictReplayBuffer</span>
<span class="kn">from</span> <span class="nn">torchrl.envs</span> <span class="kn">import</span> <span class="p">(</span>
<span class="n">Compose</span><span class="p">,</span>
<span class="n">ExplorationType</span><span class="p">,</span>
<span class="n">GrayScale</span><span class="p">,</span>
<span class="n">InitTracker</span><span class="p">,</span>
<span class="n">ObservationNorm</span><span class="p">,</span>
<span class="n">Resize</span><span class="p">,</span>
<span class="n">RewardScaling</span><span class="p">,</span>
<span class="n">set_exploration_type</span><span class="p">,</span>
<span class="n">StepCounter</span><span class="p">,</span>
<span class="n">ToTensorImage</span><span class="p">,</span>
<span class="n">TransformedEnv</span><span class="p">,</span>
<span class="p">)</span>
<span class="kn">from</span> <span class="nn">torchrl.envs.libs.gym</span> <span class="kn">import</span> <span class="n">GymEnv</span>
<span class="kn">from</span> <span class="nn">torchrl.modules</span> <span class="kn">import</span> <span class="n">ConvNet</span><span class="p">,</span> <span class="n">EGreedyModule</span><span class="p">,</span> <span class="n">LSTMModule</span><span class="p">,</span> <span class="n">MLP</span><span class="p">,</span> <span class="n">QValueModule</span>
<span class="kn">from</span> <span class="nn">torchrl.objectives</span> <span class="kn">import</span> <span class="n">DQNLoss</span><span class="p">,</span> <span class="n">SoftUpdate</span>
<span class="n">is_fork</span> <span class="o">=</span> <span class="n">multiprocessing</span><span class="o">.</span><span class="n">get_start_method</span><span class="p">()</span> <span class="o">==</span> <span class="s2">"fork"</span>
<span class="n">device</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span><span class="mi">0</span><span class="p">)</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="ow">and</span> <span class="ow">not</span> <span class="n">is_fork</span>
<span class="k">else</span> <span class="n">torch</span><span class="o">.</span><span class="n">device</span><span class="p">(</span><span class="s2">"cpu"</span><span class="p">)</span>
<span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="environment">
<h2>Environment<a class="headerlink" href="#environment" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>As usual, the first step is to build our environment: it helps us
define the problem and build the policy network accordingly. For this tutorial,
we’ll be running a single pixel-based instance of the CartPole gym
environment with some custom transforms: turning to grayscale, resizing to
84x84, scaling down the rewards and normalizing the observations.</p>
<div class="admonition note">
<p class="admonition-title">참고</p>
<p>The <code class="xref py py-class docutils literal notranslate"><span class="pre">StepCounter</span></code> transform is accessory. Since the CartPole
task goal is to make trajectories as long as possible, counting the steps
can help us track the performance of our policy.</p>
</div>
<p>Two transforms are important for the purpose of this tutorial:</p>
<ul class="simple">
<li><p><code class="xref py py-class docutils literal notranslate"><span class="pre">InitTracker</span></code> will stamp the
calls to <code class="xref py py-meth docutils literal notranslate"><span class="pre">reset()</span></code> by adding a <code class="docutils literal notranslate"><span class="pre">"is_init"</span></code>
boolean mask in the TensorDict that will track which steps require a reset
of the RNN hidden states.</p></li>
<li><p>The <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDictPrimer</span></code> transform is a bit more
technical. It is not required to use RNN policies. However, it
instructs the environment (and subsequently the collector) that some extra
keys are to be expected. Once added, a call to <cite>env.reset()</cite> will populate
the entries indicated in the primer with zeroed tensors. Knowing that
these tensors are expected by the policy, the collector will pass them on
during collection. Eventually, we’ll be storing our hidden states in the
replay buffer, which will help us bootstrap the computation of the
RNN operations in the loss module (which would otherwise be initiated
with 0s). In summary: not including this transform will not impact hugely
the training of our policy, but it will make the recurrent keys disappear
from the collected data and the replay buffer, which will in turn lead to
a slightly less optimal training.
Fortunately, the <code class="xref py py-class docutils literal notranslate"><span class="pre">LSTMModule</span></code> we propose is
equipped with a helper method to build just that transform for us, so
we can wait until we build it!</p></li>
</ul>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">env</span> <span class="o">=</span> <span class="n">TransformedEnv</span><span class="p">(</span>
<span class="n">GymEnv</span><span class="p">(</span><span class="s2">"CartPole-v1"</span><span class="p">,</span> <span class="n">from_pixels</span><span class="o">=</span><span class="kc">True</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">Compose</span><span class="p">(</span>
<span class="n">ToTensorImage</span><span class="p">(),</span>
<span class="n">GrayScale</span><span class="p">(),</span>
<span class="n">Resize</span><span class="p">(</span><span class="mi">84</span><span class="p">,</span> <span class="mi">84</span><span class="p">),</span>
<span class="n">StepCounter</span><span class="p">(),</span>
<span class="n">InitTracker</span><span class="p">(),</span>
<span class="n">RewardScaling</span><span class="p">(</span><span class="n">loc</span><span class="o">=</span><span class="mf">0.0</span><span class="p">,</span> <span class="n">scale</span><span class="o">=</span><span class="mf">0.1</span><span class="p">),</span>
<span class="n">ObservationNorm</span><span class="p">(</span><span class="n">standard_normal</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">in_keys</span><span class="o">=</span><span class="p">[</span><span class="s2">"pixels"</span><span class="p">]),</span>
<span class="p">),</span>
<span class="p">)</span>
</pre></div>
</div>
<p>As always, we need to initialize manually our normalization constants:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">env</span><span class="o">.</span><span class="n">transform</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">init_stats</span><span class="p">(</span><span class="mi">1000</span><span class="p">,</span> <span class="n">reduce_dim</span><span class="o">=</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="mi">2</span><span class="p">],</span> <span class="n">cat_dim</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">keep_dims</span><span class="o">=</span><span class="p">[</span><span class="mi">0</span><span class="p">])</span>
<span class="n">td</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="policy">
<h2>Policy<a class="headerlink" href="#policy" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>Our policy will have 3 components: a <code class="xref py py-class docutils literal notranslate"><span class="pre">ConvNet</span></code>
backbone, an <code class="xref py py-class docutils literal notranslate"><span class="pre">LSTMModule</span></code> memory layer and a shallow
<code class="xref py py-class docutils literal notranslate"><span class="pre">MLP</span></code> block that will map the LSTM output onto the
action values.</p>
<div class="section" id="convolutional-network">
<h3>Convolutional network<a class="headerlink" href="#convolutional-network" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>We build a convolutional network flanked with a <a class="reference external" href="https://pytorch.org/docs/stable/generated/torch.nn.AdaptiveAvgPool2d.html#torch.nn.AdaptiveAvgPool2d" title="(PyTorch v2.3에서)"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.AdaptiveAvgPool2d</span></code></a>
that will squash the output in a vector of size 64. The <code class="xref py py-class docutils literal notranslate"><span class="pre">ConvNet</span></code>
can assist us with this:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">feature</span> <span class="o">=</span> <span class="n">Mod</span><span class="p">(</span>
<span class="n">ConvNet</span><span class="p">(</span>
<span class="n">num_cells</span><span class="o">=</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="mi">64</span><span class="p">],</span>
<span class="n">squeeze_output</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">aggregator_class</span><span class="o">=</span><span class="n">nn</span><span class="o">.</span><span class="n">AdaptiveAvgPool2d</span><span class="p">,</span>
<span class="n">aggregator_kwargs</span><span class="o">=</span><span class="p">{</span><span class="s2">"output_size"</span><span class="p">:</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="n">device</span><span class="o">=</span><span class="n">device</span><span class="p">,</span>
<span class="p">),</span>
<span class="n">in_keys</span><span class="o">=</span><span class="p">[</span><span class="s2">"pixels"</span><span class="p">],</span>
<span class="n">out_keys</span><span class="o">=</span><span class="p">[</span><span class="s2">"embed"</span><span class="p">],</span>
<span class="p">)</span>
</pre></div>
</div>
<p>we execute the first module on a batch of data to gather the size of the
output vector:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">n_cells</span> <span class="o">=</span> <span class="n">feature</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="s2">"embed"</span><span class="p">]</span><span class="o">.</span><span class="n">shape</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="lstm-module">
<h3>LSTM Module<a class="headerlink" href="#lstm-module" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>TorchRL provides a specialized <code class="xref py py-class docutils literal notranslate"><span class="pre">LSTMModule</span></code> class
to incorporate LSTMs in your code-base. It is a <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDictModuleBase</span></code>
subclass: as such, it has a set of <code class="docutils literal notranslate"><span class="pre">in_keys</span></code> and <code class="docutils literal notranslate"><span class="pre">out_keys</span></code> that indicate
what values should be expected to be read and written/updated during the
execution of the module. The class comes with customizable predefined
values for these attributes to facilitate its construction.</p>
<div class="admonition note">
<p class="admonition-title">참고</p>
<p><em>Usage limitations</em>: The class supports almost all LSTM features such as
dropout or multi-layered LSTMs.
However, to respect TorchRL’s conventions, this LSTM must have the <code class="docutils literal notranslate"><span class="pre">batch_first</span></code>
attribute set to <code class="docutils literal notranslate"><span class="pre">True</span></code> which is <strong>not</strong> the default in PyTorch. However,
our <code class="xref py py-class docutils literal notranslate"><span class="pre">LSTMModule</span></code> changes this default
behavior, so we’re good with a native call.</p>
<p>Also, the LSTM cannot have a <code class="docutils literal notranslate"><span class="pre">bidirectional</span></code> attribute set to <code class="docutils literal notranslate"><span class="pre">True</span></code> as
this wouldn’t be usable in online settings. In this case, the default value
is the correct one.</p>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">lstm</span> <span class="o">=</span> <span class="n">LSTMModule</span><span class="p">(</span>
<span class="n">input_size</span><span class="o">=</span><span class="n">n_cells</span><span class="p">,</span>
<span class="n">hidden_size</span><span class="o">=</span><span class="mi">128</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">in_key</span><span class="o">=</span><span class="s2">"embed"</span><span class="p">,</span>
<span class="n">out_key</span><span class="o">=</span><span class="s2">"embed"</span><span class="p">,</span>
<span class="p">)</span>
</pre></div>
</div>
<p>Let us look at the LSTM Module class, specifically its in and out_keys:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nb">print</span><span class="p">(</span><span class="s2">"in_keys"</span><span class="p">,</span> <span class="n">lstm</span><span class="o">.</span><span class="n">in_keys</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"out_keys"</span><span class="p">,</span> <span class="n">lstm</span><span class="o">.</span><span class="n">out_keys</span><span class="p">)</span>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>in_keys ['embed', 'recurrent_state_h', 'recurrent_state_c', 'is_init']
out_keys ['embed', ('next', 'recurrent_state_h'), ('next', 'recurrent_state_c')]
</pre></div>
</div>
<p>We can see that these values contain the key we indicated as the in_key (and out_key)
as well as recurrent key names. The out_keys are preceded by a 《next》 prefix
that indicates that they will need to be written in the 《next》 TensorDict.
We use this convention (which can be overridden by passing the in_keys/out_keys
arguments) to make sure that a call to <code class="xref py py-func docutils literal notranslate"><span class="pre">step_mdp()</span></code> will
move the recurrent state to the root TensorDict, making it available to the
RNN during the following call (see figure in the intro).</p>
<p>As mentioned earlier, we have one more optional transform to add to our
environment to make sure that the recurrent states are passed to the buffer.
The <code class="xref py py-meth docutils literal notranslate"><span class="pre">make_tensordict_primer()</span></code> method does
exactly that:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">env</span><span class="o">.</span><span class="n">append_transform</span><span class="p">(</span><span class="n">lstm</span><span class="o">.</span><span class="n">make_tensordict_primer</span><span class="p">())</span>
</pre></div>
</div>
<p>and that’s it! We can print the environment to check that everything looks good now
that we have added the primer:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nb">print</span><span class="p">(</span><span class="n">env</span><span class="p">)</span>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>TransformedEnv(
env=GymEnv(env=CartPole-v1, batch_size=torch.Size([]), device=cpu),
transform=Compose(
ToTensorImage(keys=['pixels']),
GrayScale(keys=['pixels']),
Resize(w=84, h=84, interpolation=InterpolationMode.BILINEAR, keys=['pixels']),
StepCounter(keys=[]),
InitTracker(keys=[]),
RewardScaling(loc=0.0000, scale=0.1000, keys=['reward']),
ObservationNorm(keys=['pixels']),
TensorDictPrimer(primers=CompositeSpec(
recurrent_state_h: UnboundedContinuousTensorSpec(
shape=torch.Size([1, 128]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous),
recurrent_state_c: UnboundedContinuousTensorSpec(
shape=torch.Size([1, 128]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous), device=cpu, shape=torch.Size([])), default_value=0.0, random=False)))
</pre></div>
</div>
</div>
<div class="section" id="mlp">
<h3>MLP<a class="headerlink" href="#mlp" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>We use a single-layer MLP to represent the action values we’ll be using for
our policy.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">mlp</span> <span class="o">=</span> <span class="n">MLP</span><span class="p">(</span>
<span class="n">out_features</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span>
<span class="n">num_cells</span><span class="o">=</span><span class="p">[</span>
<span class="mi">64</span><span class="p">,</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="p">)</span>
</pre></div>
</div>
<p>and fill the bias with zeros:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">mlp</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">bias</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">fill_</span><span class="p">(</span><span class="mf">0.0</span><span class="p">)</span>
<span class="n">mlp</span> <span class="o">=</span> <span class="n">Mod</span><span class="p">(</span><span class="n">mlp</span><span class="p">,</span> <span class="n">in_keys</span><span class="o">=</span><span class="p">[</span><span class="s2">"embed"</span><span class="p">],</span> <span class="n">out_keys</span><span class="o">=</span><span class="p">[</span><span class="s2">"action_value"</span><span class="p">])</span>
</pre></div>
</div>
</div>
<div class="section" id="using-the-q-values-to-select-an-action">
<h3>Using the Q-Values to select an action<a class="headerlink" href="#using-the-q-values-to-select-an-action" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>The last part of our policy is the Q-Value Module.
The Q-Value module <code class="xref py py-class docutils literal notranslate"><span class="pre">QValueModule</span></code>
will read the <code class="docutils literal notranslate"><span class="pre">"action_values"</span></code> key that is produced by our MLP and
from it, gather the action that has the maximum value.
The only thing we need to do is to specify the action space, which can be done
either by passing a string or an action-spec. This allows us to use
Categorical (sometimes called 《sparse》) encoding or the one-hot version of it.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">qval</span> <span class="o">=</span> <span class="n">QValueModule</span><span class="p">(</span><span class="n">action_space</span><span class="o">=</span><span class="n">env</span><span class="o">.</span><span class="n">action_spec</span><span class="p">)</span>
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">참고</p>
<p>TorchRL also provides a wrapper class <code class="xref py py-class docutils literal notranslate"><span class="pre">torchrl.modules.QValueActor</span></code> that
wraps a module in a Sequential together with a <code class="xref py py-class docutils literal notranslate"><span class="pre">QValueModule</span></code>
like we are doing explicitly here. There is little advantage to do this
and the process is less transparent, but the end results will be similar to
what we do here.</p>
</div>
<p>We can now put things together in a <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDictSequential</span></code></p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">stoch_policy</span> <span class="o">=</span> <span class="n">Seq</span><span class="p">(</span><span class="n">feature</span><span class="p">,</span> <span class="n">lstm</span><span class="p">,</span> <span class="n">mlp</span><span class="p">,</span> <span class="n">qval</span><span class="p">)</span>
</pre></div>
</div>
<p>DQN being a deterministic algorithm, exploration is a crucial part of it.
We’ll be using an <span class="math">\(\epsilon\)</span>-greedy policy with an epsilon of 0.2 decaying
progressively to 0.
This decay is achieved via a call to <code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code>
(see training loop below).</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">exploration_module</span> <span class="o">=</span> <span class="n">EGreedyModule</span><span class="p">(</span>
<span class="n">annealing_num_steps</span><span class="o">=</span><span class="mi">1_000_000</span><span class="p">,</span> <span class="n">spec</span><span class="o">=</span><span class="n">env</span><span class="o">.</span><span class="n">action_spec</span><span class="p">,</span> <span class="n">eps_init</span><span class="o">=</span><span class="mf">0.2</span>
<span class="p">)</span>
<span class="n">stoch_policy</span> <span class="o">=</span> <span class="n">Seq</span><span class="p">(</span>
<span class="n">stoch_policy</span><span class="p">,</span>
<span class="n">exploration_module</span><span class="p">,</span>
<span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="using-the-model-for-the-loss">
<h3>Using the model for the loss<a class="headerlink" href="#using-the-model-for-the-loss" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>The model as we’ve built it is well equipped to be used in sequential settings.
However, the class <a class="reference external" href="https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html#torch.nn.LSTM" title="(PyTorch v2.3에서)"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.LSTM</span></code></a> can use a cuDNN-optimized backend
to run the RNN sequence faster on GPU device. We would not want to miss
such an opportunity to speed up our training loop!
To use it, we just need to tell the LSTM module to run on 《recurrent-mode》
when used by the loss.
As we’ll usually want to have two copies of the LSTM module, we do this by
calling a <code class="xref py py-meth docutils literal notranslate"><span class="pre">set_recurrent_mode()</span></code> method that
will return a new instance of the LSTM (with shared weights) that will
assume that the input data is sequential in nature.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">policy</span> <span class="o">=</span> <span class="n">Seq</span><span class="p">(</span><span class="n">feature</span><span class="p">,</span> <span class="n">lstm</span><span class="o">.</span><span class="n">set_recurrent_mode</span><span class="p">(</span><span class="kc">True</span><span class="p">),</span> <span class="n">mlp</span><span class="p">,</span> <span class="n">qval</span><span class="p">)</span>
</pre></div>
</div>
<p>Because we still have a couple of uninitialized parameters we should
initialize them before creating an optimizer and such.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">policy</span><span class="p">(</span><span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">())</span>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>TensorDict(
fields={
action: Tensor(shape=torch.Size([2]), device=cpu, dtype=torch.int64, is_shared=False),
action_value: Tensor(shape=torch.Size([2]), device=cpu, dtype=torch.float32, is_shared=False),
chosen_action_value: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),
done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
embed: Tensor(shape=torch.Size([128]), device=cpu, dtype=torch.float32, is_shared=False),
is_init: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
next: TensorDict(
fields={
recurrent_state_c: Tensor(shape=torch.Size([1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_h: Tensor(shape=torch.Size([1, 128]), device=cpu, dtype=torch.float32, is_shared=False)},
batch_size=torch.Size([]),
device=cpu,
is_shared=False),
pixels: Tensor(shape=torch.Size([1, 84, 84]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_c: Tensor(shape=torch.Size([1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_h: Tensor(shape=torch.Size([1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
step_count: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.int64, is_shared=False),
terminated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),
truncated: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False)},
batch_size=torch.Size([]),
device=cpu,
is_shared=False)
</pre></div>
</div>
</div>
</div>
<div class="section" id="dqn-loss">
<h2>DQN Loss<a class="headerlink" href="#dqn-loss" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>Out DQN loss requires us to pass the policy and, again, the action-space.
While this may seem redundant, it is important as we want to make sure that
the <code class="xref py py-class docutils literal notranslate"><span class="pre">DQNLoss</span></code> and the <code class="xref py py-class docutils literal notranslate"><span class="pre">QValueModule</span></code>
classes are compatible, but aren’t strongly dependent on each other.</p>
<p>To use the Double-DQN, we ask for a <code class="docutils literal notranslate"><span class="pre">delay_value</span></code> argument that will
create a non-differentiable copy of the network parameters to be used
as a target network.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">loss_fn</span> <span class="o">=</span> <span class="n">DQNLoss</span><span class="p">(</span><span class="n">policy</span><span class="p">,</span> <span class="n">action_space</span><span class="o">=</span><span class="n">env</span><span class="o">.</span><span class="n">action_spec</span><span class="p">,</span> <span class="n">delay_value</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
</pre></div>
</div>
<p>Since we are using a double DQN, we need to update the target parameters.
We’ll use a <code class="xref py py-class docutils literal notranslate"><span class="pre">SoftUpdate</span></code> instance to carry out
this work.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">updater</span> <span class="o">=</span> <span class="n">SoftUpdate</span><span class="p">(</span><span class="n">loss_fn</span><span class="p">,</span> <span class="n">eps</span><span class="o">=</span><span class="mf">0.95</span><span class="p">)</span>
<span class="n">optim</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">Adam</span><span class="p">(</span><span class="n">policy</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">3e-4</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="collector-and-replay-buffer">
<h2>Collector and replay buffer<a class="headerlink" href="#collector-and-replay-buffer" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>We build the simplest data collector there is. We’ll try to train our algorithm
with a million frames, extending the buffer with 50 frames at a time. The buffer
will be designed to store 20 thousands trajectories of 50 steps each.
At each optimization step (16 per data collection), we’ll collect 4 items
from our buffer, for a total of 200 transitions.
We’ll use a <code class="xref py py-class docutils literal notranslate"><span class="pre">LazyMemmapStorage</span></code> storage to keep the data
on disk.</p>
<div class="admonition note">
<p class="admonition-title">참고</p>
<p>For the sake of efficiency, we’re only running a few thousands iterations
here. In a real setting, the total number of frames should be set to 1M.</p>
</div>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">collector</span> <span class="o">=</span> <span class="n">SyncDataCollector</span><span class="p">(</span><span class="n">env</span><span class="p">,</span> <span class="n">stoch_policy</span><span class="p">,</span> <span class="n">frames_per_batch</span><span class="o">=</span><span class="mi">50</span><span class="p">,</span> <span class="n">total_frames</span><span class="o">=</span><span class="mi">200</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">rb</span> <span class="o">=</span> <span class="n">TensorDictReplayBuffer</span><span class="p">(</span>
<span class="n">storage</span><span class="o">=</span><span class="n">LazyMemmapStorage</span><span class="p">(</span><span class="mi">20_000</span><span class="p">),</span> <span class="n">batch_size</span><span class="o">=</span><span class="mi">4</span><span class="p">,</span> <span class="n">prefetch</span><span class="o">=</span><span class="mi">10</span>
<span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="training-loop">
<h2>Training loop<a class="headerlink" href="#training-loop" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>To keep track of the progress, we will run the policy in the environment once
every 50 data collection, and plot the results after training.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">utd</span> <span class="o">=</span> <span class="mi">16</span>
<span class="n">pbar</span> <span class="o">=</span> <span class="n">tqdm</span><span class="o">.</span><span class="n">tqdm</span><span class="p">(</span><span class="n">total</span><span class="o">=</span><span class="mi">1_000_000</span><span class="p">)</span>
<span class="n">longest</span> <span class="o">=</span> <span class="mi">0</span>
<span class="n">traj_lens</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">data</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">collector</span><span class="p">):</span>
<span class="k">if</span> <span class="n">i</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span>
<span class="s2">"Let us print the first batch of data.</span><span class="se">\n</span><span class="s2">Pay attention to the key names "</span>
<span class="s2">"which will reflect what can be found in this data structure, in particular: "</span>
<span class="s2">"the output of the QValueModule (action_values, action and chosen_action_value),"</span>
<span class="s2">"the 'is_init' key that will tell us if a step is initial or not, and the "</span>
<span class="s2">"recurrent_state keys.</span><span class="se">\n</span><span class="s2">"</span><span class="p">,</span>
<span class="n">data</span><span class="p">,</span>
<span class="p">)</span>
<span class="n">pbar</span><span class="o">.</span><span class="n">update</span><span class="p">(</span><span class="n">data</span><span class="o">.</span><span class="n">numel</span><span class="p">())</span>
<span class="c1"># it is important to pass data that is not flattened</span>
<span class="n">rb</span><span class="o">.</span><span class="n">extend</span><span class="p">(</span><span class="n">data</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="o">.</span><span class="n">to_tensordict</span><span class="p">()</span><span class="o">.</span><span class="n">cpu</span><span class="p">())</span>
<span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">utd</span><span class="p">):</span>
<span class="n">s</span> <span class="o">=</span> <span class="n">rb</span><span class="o">.</span><span class="n">sample</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">non_blocking</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">loss_vals</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="n">s</span><span class="p">)</span>
<span class="n">loss_vals</span><span class="p">[</span><span class="s2">"loss"</span><span class="p">]</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optim</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="n">optim</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">longest</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="n">longest</span><span class="p">,</span> <span class="n">data</span><span class="p">[</span><span class="s2">"step_count"</span><span class="p">]</span><span class="o">.</span><span class="n">max</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">())</span>
<span class="n">pbar</span><span class="o">.</span><span class="n">set_description</span><span class="p">(</span>
<span class="sa">f</span><span class="s2">"steps: </span><span class="si">{</span><span class="n">longest</span><span class="si">}</span><span class="s2">, loss_val: </span><span class="si">{</span><span class="n">loss_vals</span><span class="p">[</span><span class="s1">'loss'</span><span class="p">]</span><span class="o">.</span><span class="n">item</span><span class="p">()</span><span class="si">:</span><span class="s2"> 4.4f</span><span class="si">}</span><span class="s2">, action_spread: </span><span class="si">{</span><span class="n">data</span><span class="p">[</span><span class="s1">'action'</span><span class="p">]</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="si">}</span><span class="s2">"</span>
<span class="p">)</span>
<span class="n">exploration_module</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">data</span><span class="o">.</span><span class="n">numel</span><span class="p">())</span>
<span class="n">updater</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="k">with</span> <span class="n">set_exploration_type</span><span class="p">(</span><span class="n">ExplorationType</span><span class="o">.</span><span class="n">MODE</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="n">rollout</span> <span class="o">=</span> <span class="n">env</span><span class="o">.</span><span class="n">rollout</span><span class="p">(</span><span class="mi">10000</span><span class="p">,</span> <span class="n">stoch_policy</span><span class="p">)</span>
<span class="n">traj_lens</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">rollout</span><span class="o">.</span><span class="n">get</span><span class="p">((</span><span class="s2">"next"</span><span class="p">,</span> <span class="s2">"step_count"</span><span class="p">))</span><span class="o">.</span><span class="n">max</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">())</span>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span> 0%| | 0/1000000 [00:00<?, ?it/s]Let us print the first batch of data.
Pay attention to the key names which will reflect what can be found in this data structure, in particular: the output of the QValueModule (action_values, action and chosen_action_value),the 'is_init' key that will tell us if a step is initial or not, and the recurrent_state keys.
TensorDict(
fields={
action: Tensor(shape=torch.Size([50, 2]), device=cpu, dtype=torch.int64, is_shared=False),
action_value: Tensor(shape=torch.Size([50, 2]), device=cpu, dtype=torch.float32, is_shared=False),
chosen_action_value: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.float32, is_shared=False),
collector: TensorDict(
fields={
traj_ids: Tensor(shape=torch.Size([50]), device=cpu, dtype=torch.int64, is_shared=False)},
batch_size=torch.Size([50]),
device=cpu,
is_shared=False),
done: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
embed: Tensor(shape=torch.Size([50, 128]), device=cpu, dtype=torch.float32, is_shared=False),
is_init: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
next: TensorDict(
fields={
done: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
is_init: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
pixels: Tensor(shape=torch.Size([50, 1, 84, 84]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_c: Tensor(shape=torch.Size([50, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_h: Tensor(shape=torch.Size([50, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
reward: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.float32, is_shared=False),
step_count: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.int64, is_shared=False),
terminated: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
truncated: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
batch_size=torch.Size([50]),
device=cpu,
is_shared=False),
pixels: Tensor(shape=torch.Size([50, 1, 84, 84]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_c: Tensor(shape=torch.Size([50, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
recurrent_state_h: Tensor(shape=torch.Size([50, 1, 128]), device=cpu, dtype=torch.float32, is_shared=False),
step_count: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.int64, is_shared=False),
terminated: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False),
truncated: Tensor(shape=torch.Size([50, 1]), device=cpu, dtype=torch.bool, is_shared=False)},
batch_size=torch.Size([50]),
device=cpu,
is_shared=False)
0%| | 50/1000000 [00:00<1:24:57, 196.16it/s]
0%| | 50/1000000 [00:13<1:24:57, 196.16it/s]
steps: 15, loss_val: 0.0001, action_spread: tensor([37, 13]): 0%| | 50/1000000 [00:18<1:24:57, 196.16it/s]
steps: 15, loss_val: 0.0001, action_spread: tensor([37, 13]): 0%| | 100/1000000 [00:18<60:48:55, 4.57it/s]
steps: 15, loss_val: 0.0002, action_spread: tensor([ 3, 47]): 0%| | 100/1000000 [00:36<60:48:55, 4.57it/s]
steps: 15, loss_val: 0.0002, action_spread: tensor([ 3, 47]): 0%| | 150/1000000 [00:36<78:45:54, 3.53it/s]
steps: 18, loss_val: 0.0002, action_spread: tensor([37, 13]): 0%| | 150/1000000 [00:54<78:45:54, 3.53it/s]
steps: 18, loss_val: 0.0002, action_spread: tensor([37, 13]): 0%| | 200/1000000 [00:54<87:14:55, 3.18it/s]
steps: 18, loss_val: 0.0003, action_spread: tensor([ 8, 42]): 0%| | 200/1000000 [01:12<87:14:55, 3.18it/s]
</pre></div>
</div>
<p>Let’s plot our results:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">if</span> <span class="n">traj_lens</span><span class="p">:</span>
<span class="kn">from</span> <span class="nn">matplotlib</span> <span class="kn">import</span> <span class="n">pyplot</span> <span class="k">as</span> <span class="n">plt</span>
<span class="n">plt</span><span class="o">.</span><span class="n">plot</span><span class="p">(</span><span class="n">traj_lens</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="s2">"Test collection"</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="s2">"Test trajectory lengths"</span><span class="p">)</span>
</pre></div>
</div>
<img src="../_images/sphx_glr_dqn_with_rnn_tutorial_001.png" srcset="../_images/sphx_glr_dqn_with_rnn_tutorial_001.png" alt="Test trajectory lengths" class = "sphx-glr-single-img"/></div>
<div class="section" id="conclusion">
<h2>Conclusion<a class="headerlink" href="#conclusion" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>We have seen how an RNN can be incorporated in a policy in TorchRL.
You should now be able:</p>
<ul class="simple">
<li><p>Create an LSTM module that acts as a <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDictModule</span></code></p></li>
<li><p>Indicate to the LSTM module that a reset is needed via an <code class="xref py py-class docutils literal notranslate"><span class="pre">InitTracker</span></code>
transform</p></li>
<li><p>Incorporate this module in a policy and in a loss module</p></li>
<li><p>Make sure that the collector is made aware of the recurrent state entries
such that they can be stored in the replay buffer along with the rest of
the data</p></li>
</ul>
</div>
<div class="section" id="further-reading">
<h2>Further Reading<a class="headerlink" href="#further-reading" title="이 제목에 대한 퍼머링크">¶</a></h2>
<ul class="simple">
<li><p>The TorchRL documentation can be found <a class="reference external" href="https://pytorch.org/rl/">here</a>.</p></li>
</ul>
<p class="sphx-glr-timing"><strong>Total running time of the script:</strong> ( 1 minutes 16.189 seconds)</p>
<div class="sphx-glr-footer sphx-glr-footer-example docutils container" id="sphx-glr-download-intermediate-dqn-with-rnn-tutorial-py">
<div class="sphx-glr-download sphx-glr-download-python docutils container">
<p><a class="reference download internal" download="" href="../_downloads/14b68b9764c79afe8ef88b11fc27bff7/dqn_with_rnn_tutorial.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">dqn_with_rnn_tutorial.py</span></code></a></p>
</div>
<div class="sphx-glr-download sphx-glr-download-jupyter docutils container">
<p><a class="reference download internal" download="" href="../_downloads/224d2179034ef4c00cd9b86f2976062a/dqn_with_rnn_tutorial.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">dqn_with_rnn_tutorial.ipynb</span></code></a></p>
</div>
</div>
<p class="sphx-glr-signature"><a class="reference external" href="https://sphinx-gallery.github.io">Gallery generated by Sphinx-Gallery</a></p>
</div>
</div>
</article>
</div>
<footer>
<hr class="community-hr hr-top" />