-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathpendulum.html
3837 lines (2696 loc) · 267 KB
/
pendulum.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="Pendulum: Writing your environment and transforms with TorchRL" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/advanced/pendulum.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="Author: Vincent Moens Creating an environment (a simulator or an interface to a physical control system) is an integrative part of reinforcement learning and control engineering. TorchRL provides a set of tools to do this in multiple contexts. This tutorial demonstrates how to use PyTorch and Tor..." />
<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 Creating an environment (a simulator or an interface to a physical control system) is an integrative part of reinforcement learning and control engineering. TorchRL provides a set of tools to do this in multiple contexts. This tutorial demonstrates how to use PyTorch and Tor..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Pendulum: Writing your environment and transforms with TorchRL — 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" />
<link rel="next" title="Flask를 사용하여 Python에서 PyTorch를 REST API로 배포하기" href="../intermediate/flask_rest_api_tutorial.html" />
<link rel="prev" title="마리오 게임 RL 에이전트로 학습하기" href="../intermediate/mario_rl_tutorial.html" />
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.kr/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.kr/get-started">시작하기</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr/">튜토리얼</a>
</li>
<li>
<a href="https://pytorch.kr/hub">허브</a>
</li>
<li>
<a href="https://discuss.pytorch.kr/">커뮤니티</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
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="../intermediate/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="../intermediate/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="../intermediate/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="../intermediate/speech_recognition_pipeline_tutorial.html">Speech Recognition with Wav2Vec2</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/text_to_speech_with_torchaudio.html">Text-to-speech with Tacotron2</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/char_rnn_classification_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 분류하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/char_rnn_generation_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 생성하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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 class="current">
<li class="toctree-l1"><a class="reference internal" href="../intermediate/reinforcement_q_learning.html">강화 학습 (DQN) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/reinforcement_ppo.html">Reinforcement Learning (PPO) with TorchRL Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/mario_rl_tutorial.html">마리오 게임 RL 에이전트로 학습하기</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">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="../intermediate/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="cpp_export.html">C++에서 TorchScript 모델 로딩하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="super_resolution_with_onnxruntime.html">(선택) PyTorch 모델을 ONNX으로 변환하고 ONNX 런타임에서 실행하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/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="../intermediate/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="../intermediate/memory_format_tutorial.html">(베타) PyTorch를 사용한 Channels Last 메모리 형식</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/forward_ad_usage.html">Forward-mode Automatic Differentiation (Beta)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/jacobians_hessians.html">Jacobians, Hessians, hvp, vhp, and more: composing function transforms</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/ensembling.html">모델 앙상블</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/per_sample_grads.html">Per-sample-gradients</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_frontend.html">PyTorch C++ 프론트엔드 사용하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch-script-parallelism.html">TorchScript의 동적 병렬 처리(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="../intermediate/custom_function_double_backward_tutorial.html">Double Backward with Custom Functions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="cpp_extension.html">Custom C++ and CUDA Extensions</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_script_custom_ops.html">Extending TorchScript with Custom C++ Operators</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_script_custom_classes.html">커스텀 C++ 클래스로 TorchScript 확장하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="dispatcher.html">Registering a Dispatched Operator in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="extend_dispatcher.html">Extending dispatcher for a new backend in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="../intermediate/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="../intermediate/parametrizations.html">Parametrizations Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/pruning_tutorial.html">가지치기 기법(Pruning) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_tutorial.html">(베타) LSTM 기반 단어 단위 언어 모델의 동적 양자화</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dynamic_quantization_bert_tutorial.html">(베타) BERT 모델 동적 양자화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="static_quantization_tutorial.html">(베타) PyTorch에서 Eager Mode를 이용한 정적 양자화</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/torchserve_with_ipex.html">Grokking PyTorch Intel CPU performance from first principles</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/nvfuser_intro_tutorial.html">Getting Started - Accelerate Your Scripts with nvFuser</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/ax_multiobjective_nas_tutorial.html">Multi-Objective NAS with Ax</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/inductor_debug_cpu.html">Inductor CPU backend debugging and profiling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/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="../intermediate/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="../intermediate/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="../intermediate/model_parallel_tutorial.html">단일 머신을 사용한 모델 병렬화 모범 사례</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/ddp_tutorial.html">분산 데이터 병렬 처리 시작하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dist_tuto.html">PyTorch로 분산 어플리케이션 개발하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/FSDP_tutorial.html">Getting Started with Fully Sharded Data Parallel(FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/FSDP_adavnced_tutorial.html">Advanced Model Training with Fully Sharded Data Parallel (FSDP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/TP_tutorial.html">Large Scale Transformer model training with Tensor Parallel (TP)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/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="../intermediate/rpc_tutorial.html">Getting Started with Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/rpc_param_server_tutorial.html">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism Using RPC</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/rpc_async_execution.html">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_ddp_tutorial.html">분산 데이터 병렬(DDP)과 분산 RPC 프레임워크 결합</a></li>
<li class="toctree-l1"><a class="reference internal" href="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="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="../intermediate/torchrec_tutorial.html">TorchRec 소개</a></li>
<li class="toctree-l1"><a class="reference internal" href="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>Pendulum: Writing your environment and transforms with TorchRL</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/advanced/pendulum.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">advanced/pendulum</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-advanced-pendulum-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="pendulum-writing-your-environment-and-transforms-with-torchrl">
<span id="sphx-glr-advanced-pendulum-py"></span><h1>Pendulum: Writing your environment and transforms with TorchRL<a class="headerlink" href="#pendulum-writing-your-environment-and-transforms-with-torchrl" title="이 제목에 대한 퍼머링크">¶</a></h1>
<p><strong>Author</strong>: <a class="reference external" href="https://github.com/vmoens">Vincent Moens</a></p>
<p>Creating an environment (a simulator or an interface to a physical control system)
is an integrative part of reinforcement learning and control engineering.</p>
<p>TorchRL provides a set of tools to do this in multiple contexts.
This tutorial demonstrates how to use PyTorch and TorchRL code a pendulum
simulator from the ground up.
It is freely inspired by the Pendulum-v1 implementation from <a class="reference external" href="https://github.com/Farama-Foundation/Gymnasium">OpenAI-Gym/Farama-Gymnasium
control library</a>.</p>
<div class="figure align-center" id="id1">
<img alt="Pendulum" src="../_images/pendulum.gif" />
<p class="caption"><span class="caption-text">Simple Pendulum</span><a class="headerlink" href="#id1" title="이 이미지에 대한 퍼머링크">¶</a></p>
</div>
<p>Key learnings:</p>
<ul>
<li><p>How to design an environment in TorchRL:
- Writing specs (input, observation and reward);
- Implementing behavior: seeding, reset and step.</p></li>
<li><p>Transforming your environment inputs and outputs, and writing your own
transforms;</p></li>
<li><p>How to use <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDict</span></code> to carry arbitrary data structures
through the <code class="docutils literal notranslate"><span class="pre">codebase</span></code>.</p>
<p>In the process, we will touch three crucial components of TorchRL:</p>
</li>
</ul>
<ul class="simple">
<li><p><a class="reference external" href="https://pytorch.org/rl/reference/envs.html">environments</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/rl/reference/envs.html#transforms">transforms</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/rl/reference/modules.html">models (policy and value function)</a></p></li>
</ul>
<p>To give a sense of what can be achieved with TorchRL’s environments, we will
be designing a <em>stateless</em> environment. While stateful environments keep track of
the latest physical state encountered and rely on this to simulate the state-to-state
transition, stateless environments expect the current state to be provided to
them at each step, along with the action undertaken. TorchRL supports both
types of environments, but stateless environments are more generic and hence
cover a broader range of features of the environment API in TorchRL.</p>
<p>Modeling stateless environments gives users full control over the input and
outputs of the simulator: one can reset an experiment at any stage or actively
modify the dynamics from the outside. However, it assumes that we have some control
over a task, which may not always be the case: solving a problem where we cannot
control the current state is more challenging but has a much wider set of applications.</p>
<p>Another advantage of stateless environments is that they can enable
batched execution of transition simulations. If the backend and the
implementation allow it, an algebraic operation can be executed seamlessly on
scalars, vectors, or tensors. This tutorial gives such examples.</p>
<p>This tutorial will be structured as follows:</p>
<ul class="simple">
<li><p>We will first get acquainted with the environment properties:
its shape (<code class="docutils literal notranslate"><span class="pre">batch_size</span></code>), its methods (mainly <code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code>,
<code class="xref py py-meth docutils literal notranslate"><span class="pre">reset()</span></code> and <code class="xref py py-meth docutils literal notranslate"><span class="pre">set_seed()</span></code>)
and finally its specs.</p></li>
<li><p>After having coded our simulator, we will demonstrate how it can be used
during training with transforms.</p></li>
<li><p>We will explore new avenues that follow from the TorchRL’s API,
including: the possibility of transforming inputs, the vectorized execution
of the simulation and the possibility of backpropagation through the
simulation graph.</p></li>
<li><p>Finally, we will train a simple policy to solve the system we implemented.</p></li>
</ul>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">collections</span> <span class="kn">import</span> <span class="n">defaultdict</span>
<span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">Optional</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">torch</span>
<span class="kn">import</span> <span class="nn">tqdm</span>
<span class="kn">from</span> <span class="nn">tensordict</span> <span class="kn">import</span> <span class="n">TensorDict</span><span class="p">,</span> <span class="n">TensorDictBase</span>
<span class="kn">from</span> <span class="nn">tensordict.nn</span> <span class="kn">import</span> <span class="n">TensorDictModule</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.data</span> <span class="kn">import</span> <span class="n">BoundedTensorSpec</span><span class="p">,</span> <span class="n">CompositeSpec</span><span class="p">,</span> <span class="n">UnboundedContinuousTensorSpec</span>
<span class="kn">from</span> <span class="nn">torchrl.envs</span> <span class="kn">import</span> <span class="p">(</span>
<span class="n">CatTensors</span><span class="p">,</span>
<span class="n">EnvBase</span><span class="p">,</span>
<span class="n">Transform</span><span class="p">,</span>
<span class="n">TransformedEnv</span><span class="p">,</span>
<span class="n">UnsqueezeTransform</span><span class="p">,</span>
<span class="p">)</span>
<span class="kn">from</span> <span class="nn">torchrl.envs.transforms.transforms</span> <span class="kn">import</span> <span class="n">_apply_to_composite</span>
<span class="kn">from</span> <span class="nn">torchrl.envs.utils</span> <span class="kn">import</span> <span class="n">check_env_specs</span><span class="p">,</span> <span class="n">step_mdp</span>
<span class="n">DEFAULT_X</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">pi</span>
<span class="n">DEFAULT_Y</span> <span class="o">=</span> <span class="mf">1.0</span>
</pre></div>
</div>
<p>There are four things you must take care of when designing a new environment
class:</p>
<ul class="simple">
<li><p><code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase._reset()</span></code>, which codes for the resetting of the simulator
at a (potentially random) initial state;</p></li>
<li><p><code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase._step()</span></code> which codes for the state transition dynamic;</p></li>
<li><p><code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase._set_seed`()</span></code> which implements the seeding mechanism;</p></li>
<li><p>the environment specs.</p></li>
</ul>
<p>Let us first describe the problem at hand: we would like to model a simple
pendulum over which we can control the torque applied on its fixed point.
Our goal is to place the pendulum in upward position (angular position at 0
by convention) and having it standing still in that position.
To design our dynamic system, we need to define two equations: the motion
equation following an action (the torque applied) and the reward equation
that will constitute our objective function.</p>
<p>For the motion equation, we will update the angular velocity following:</p>
<div class="math">
\[\dot{\theta}_{t+1} = \dot{\theta}_t + (3 * g / (2 * L) * \sin(\theta_t) + 3 / (m * L^2) * u) * dt\]</div>
<p>where <span class="math">\(\dot{\theta}\)</span> is the angular velocity in rad/sec, <span class="math">\(g\)</span> is the
gravitational force, <span class="math">\(L\)</span> is the pendulum length, <span class="math">\(m\)</span> is its mass,
<span class="math">\(\theta\)</span> is its angular position and <span class="math">\(u\)</span> is the torque. The
angular position is then updated according to</p>
<div class="math">
\[\theta_{t+1} = \theta_{t} + \dot{\theta}_{t+1} dt\]</div>
<p>We define our reward as</p>
<div class="math">
\[r = -(\theta^2 + 0.1 * \dot{\theta}^2 + 0.001 * u^2)\]</div>
<p>which will be maximized when the angle is close to 0 (pendulum in upward
position), the angular velocity is close to 0 (no motion) and the torque is
0 too.</p>
<div class="section" id="coding-the-effect-of-an-action-step">
<h2>Coding the effect of an action: <code class="xref py py-func docutils literal notranslate"><span class="pre">_step()</span></code><a class="headerlink" href="#coding-the-effect-of-an-action-step" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>The step method is the first thing to consider, as it will encode
the simulation that is of interest to us. In TorchRL, the
<code class="xref py py-class docutils literal notranslate"><span class="pre">EnvBase</span></code> class has a <code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase.step()</span></code>
method that receives a <code class="xref py py-class docutils literal notranslate"><span class="pre">tensordict.TensorDict</span></code>
instance with an <code class="docutils literal notranslate"><span class="pre">"action"</span></code> entry indicating what action is to be taken.</p>
<p>To facilitate the reading and writing from that <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> and to make sure
that the keys are consistent with what’s expected from the library, the
simulation part has been delegated to a private abstract method <code class="xref py py-meth docutils literal notranslate"><span class="pre">_step()</span></code>
which reads input data from a <code class="docutils literal notranslate"><span class="pre">tensordict</span></code>, and writes a <em>new</em> <code class="docutils literal notranslate"><span class="pre">tensordict</span></code>
with the output data.</p>
<p>The <code class="xref py py-func docutils literal notranslate"><span class="pre">_step()</span></code> method should do the following:</p>
<blockquote>
<div><ol class="arabic simple">
<li><p>Read the input keys (such as <code class="docutils literal notranslate"><span class="pre">"action"</span></code>) and execute the simulation
based on these;</p></li>
<li><p>Retrieve observations, done state and reward;</p></li>
<li><p>Write the set of observation values along with the reward and done state
at the corresponding entries in a new <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorDict</span></code>.</p></li>
</ol>
</div></blockquote>
<p>Next, the <code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code> method will merge the output
of <code class="xref py py-meth docutils literal notranslate"><span class="pre">step()</span></code> in the input <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> to enforce
input/output consistency.</p>
<p>Typically, for stateful environments, this will look like this:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </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>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="n">tensordict</span><span class="p">)</span>
<span class="go">TensorDict(</span>
<span class="go"> fields={</span>
<span class="go"> action: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),</span>
<span class="go"> done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),</span>
<span class="go"> observation: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False)},</span>
<span class="go"> batch_size=torch.Size([]),</span>
<span class="go"> device=cpu,</span>
<span class="go"> is_shared=False)</span>
<span class="gp">>>> </span><span class="n">env</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">tensordict</span><span class="p">)</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="n">tensordict</span><span class="p">)</span>
<span class="go">TensorDict(</span>
<span class="go"> fields={</span>
<span class="go"> action: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False),</span>
<span class="go"> done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),</span>
<span class="go"> next: TensorDict(</span>
<span class="go"> fields={</span>
<span class="go"> done: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.bool, is_shared=False),</span>
<span class="go"> observation: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False),</span>
<span class="go"> reward: Tensor(shape=torch.Size([1]), device=cpu, dtype=torch.float32, is_shared=False)},</span>
<span class="go"> batch_size=torch.Size([]),</span>
<span class="go"> device=cpu,</span>
<span class="go"> is_shared=False),</span>
<span class="go"> observation: Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, is_shared=False)},</span>
<span class="go"> batch_size=torch.Size([]),</span>
<span class="go"> device=cpu,</span>
<span class="go"> is_shared=False)</span>
</pre></div>
</div>
<p>Notice that the root <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> has not changed, the only modification is the
appearance of a new <code class="docutils literal notranslate"><span class="pre">"next"</span></code> entry that contains the new information.</p>
<p>In the Pendulum example, our <code class="xref py py-meth docutils literal notranslate"><span class="pre">_step()</span></code> method will read the relevant
entries from the input <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> and compute the position and velocity of
the pendulum after the force encoded by the <code class="docutils literal notranslate"><span class="pre">"action"</span></code> key has been applied
onto it. We compute the new angular position of the pendulum
<code class="docutils literal notranslate"><span class="pre">"new_th"</span></code> as the result of the previous position <code class="docutils literal notranslate"><span class="pre">"th"</span></code> plus the new
velocity <code class="docutils literal notranslate"><span class="pre">"new_thdot"</span></code> over a time interval <code class="docutils literal notranslate"><span class="pre">dt</span></code>.</p>
<p>Since our goal is to turn the pendulum up and maintain it still in that
position, our <code class="docutils literal notranslate"><span class="pre">cost</span></code> (negative reward) function is lower for positions
close to the target and low speeds.
Indeed, we want to discourage positions that are far from being 《upward》
and/or speeds that are far from 0.</p>
<p>In our example, <code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase._step()</span></code> is encoded as a static method since our
environment is stateless. In stateful settings, the <code class="docutils literal notranslate"><span class="pre">self</span></code> argument is
needed as the state needs to be read from the environment.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">_step</span><span class="p">(</span><span class="n">tensordict</span><span class="p">):</span>
<span class="n">th</span><span class="p">,</span> <span class="n">thdot</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"th"</span><span class="p">],</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"thdot"</span><span class="p">]</span> <span class="c1"># th := theta</span>
<span class="n">g_force</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"g"</span><span class="p">]</span>
<span class="n">mass</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"m"</span><span class="p">]</span>
<span class="n">length</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"l"</span><span class="p">]</span>
<span class="n">dt</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"dt"</span><span class="p">]</span>
<span class="n">u</span> <span class="o">=</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"action"</span><span class="p">]</span><span class="o">.</span><span class="n">squeeze</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
<span class="n">u</span> <span class="o">=</span> <span class="n">u</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="o">-</span><span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_torque"</span><span class="p">],</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_torque"</span><span class="p">])</span>
<span class="n">costs</span> <span class="o">=</span> <span class="n">angle_normalize</span><span class="p">(</span><span class="n">th</span><span class="p">)</span> <span class="o">**</span> <span class="mi">2</span> <span class="o">+</span> <span class="mf">0.1</span> <span class="o">*</span> <span class="n">thdot</span><span class="o">**</span><span class="mi">2</span> <span class="o">+</span> <span class="mf">0.001</span> <span class="o">*</span> <span class="p">(</span><span class="n">u</span><span class="o">**</span><span class="mi">2</span><span class="p">)</span>
<span class="n">new_thdot</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">thdot</span>
<span class="o">+</span> <span class="p">(</span><span class="mi">3</span> <span class="o">*</span> <span class="n">g_force</span> <span class="o">/</span> <span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">length</span><span class="p">)</span> <span class="o">*</span> <span class="n">th</span><span class="o">.</span><span class="n">sin</span><span class="p">()</span> <span class="o">+</span> <span class="mf">3.0</span> <span class="o">/</span> <span class="p">(</span><span class="n">mass</span> <span class="o">*</span> <span class="n">length</span><span class="o">**</span><span class="mi">2</span><span class="p">)</span> <span class="o">*</span> <span class="n">u</span><span class="p">)</span> <span class="o">*</span> <span class="n">dt</span>
<span class="p">)</span>
<span class="n">new_thdot</span> <span class="o">=</span> <span class="n">new_thdot</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span>
<span class="o">-</span><span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_speed"</span><span class="p">],</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_speed"</span><span class="p">]</span>
<span class="p">)</span>
<span class="n">new_th</span> <span class="o">=</span> <span class="n">th</span> <span class="o">+</span> <span class="n">new_thdot</span> <span class="o">*</span> <span class="n">dt</span>
<span class="n">reward</span> <span class="o">=</span> <span class="o">-</span><span class="n">costs</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">*</span><span class="n">tensordict</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">done</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">reward</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">out</span> <span class="o">=</span> <span class="n">TensorDict</span><span class="p">(</span>
<span class="p">{</span>
<span class="s2">"th"</span><span class="p">:</span> <span class="n">new_th</span><span class="p">,</span>
<span class="s2">"thdot"</span><span class="p">:</span> <span class="n">new_thdot</span><span class="p">,</span>
<span class="s2">"params"</span><span class="p">:</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">],</span>
<span class="s2">"reward"</span><span class="p">:</span> <span class="n">reward</span><span class="p">,</span>
<span class="s2">"done"</span><span class="p">:</span> <span class="n">done</span><span class="p">,</span>
<span class="p">},</span>
<span class="n">tensordict</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">out</span>
<span class="k">def</span> <span class="nf">angle_normalize</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="p">((</span><span class="n">x</span> <span class="o">+</span> <span class="n">torch</span><span class="o">.</span><span class="n">pi</span><span class="p">)</span> <span class="o">%</span> <span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">torch</span><span class="o">.</span><span class="n">pi</span><span class="p">))</span> <span class="o">-</span> <span class="n">torch</span><span class="o">.</span><span class="n">pi</span>
</pre></div>
</div>
</div>
<div class="section" id="resetting-the-simulator-reset">
<h2>Resetting the simulator: <code class="xref py py-func docutils literal notranslate"><span class="pre">_reset()</span></code><a class="headerlink" href="#resetting-the-simulator-reset" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>The second method we need to care about is the
<code class="xref py py-meth docutils literal notranslate"><span class="pre">_reset()</span></code> method. Like
<code class="xref py py-meth docutils literal notranslate"><span class="pre">_step()</span></code>, it should write the observation entries
and possibly a done state in the <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> it outputs (if the done state is
omitted, it will be filled as <code class="docutils literal notranslate"><span class="pre">False</span></code> by the parent method
<code class="xref py py-meth docutils literal notranslate"><span class="pre">reset()</span></code>). In some contexts, it is required that
the <code class="docutils literal notranslate"><span class="pre">_reset</span></code> method receives a command from the function that called
it (for example, in multi-agent settings we may want to indicate which agents need
to be reset). This is why the <code class="xref py py-meth docutils literal notranslate"><span class="pre">_reset()</span></code> method
also expects a <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> as input, albeit it may perfectly be empty or
<code class="docutils literal notranslate"><span class="pre">None</span></code>.</p>
<p>The parent <code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase.reset()</span></code> does some simple checks like the
<code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase.step()</span></code> does, such as making sure that a <code class="docutils literal notranslate"><span class="pre">"done"</span></code> state
is returned in the output <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> and that the shapes match what is
expected from the specs.</p>
<p>For us, the only important thing to consider is whether
<code class="xref py py-meth docutils literal notranslate"><span class="pre">EnvBase._reset()</span></code> contains all the expected observations. Once more,
since we are working with a stateless environment, we pass the configuration
of the pendulum in a nested <code class="docutils literal notranslate"><span class="pre">tensordict</span></code> named <code class="docutils literal notranslate"><span class="pre">"params"</span></code>.</p>
<p>In this example, we do not pass a done state as this is not mandatory
for <code class="xref py py-meth docutils literal notranslate"><span class="pre">_reset()</span></code> and our environment is non-terminating, so we always
expect it to be <code class="docutils literal notranslate"><span class="pre">False</span></code>.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">_reset</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">tensordict</span><span class="p">):</span>
<span class="k">if</span> <span class="n">tensordict</span> <span class="ow">is</span> <span class="kc">None</span> <span class="ow">or</span> <span class="n">tensordict</span><span class="o">.</span><span class="n">is_empty</span><span class="p">():</span>
<span class="c1"># if no ``tensordict`` is passed, we generate a single set of hyperparameters</span>
<span class="c1"># Otherwise, we assume that the input ``tensordict`` contains all the relevant</span>
<span class="c1"># parameters to get started.</span>
<span class="n">tensordict</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">gen_params</span><span class="p">(</span><span class="n">batch_size</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">batch_size</span><span class="p">)</span>
<span class="n">high_th</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">DEFAULT_X</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
<span class="n">high_thdot</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">DEFAULT_Y</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
<span class="n">low_th</span> <span class="o">=</span> <span class="o">-</span><span class="n">high_th</span>
<span class="n">low_thdot</span> <span class="o">=</span> <span class="o">-</span><span class="n">high_thdot</span>
<span class="c1"># for non batch-locked environments, the input ``tensordict`` shape dictates the number</span>
<span class="c1"># of simulators run simultaneously. In other contexts, the initial</span>
<span class="c1"># random state's shape will depend upon the environment batch-size instead.</span>
<span class="n">th</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="n">tensordict</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="n">generator</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rng</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
<span class="o">*</span> <span class="p">(</span><span class="n">high_th</span> <span class="o">-</span> <span class="n">low_th</span><span class="p">)</span>
<span class="o">+</span> <span class="n">low_th</span>
<span class="p">)</span>
<span class="n">thdot</span> <span class="o">=</span> <span class="p">(</span>
<span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="n">tensordict</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="n">generator</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">rng</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
<span class="o">*</span> <span class="p">(</span><span class="n">high_thdot</span> <span class="o">-</span> <span class="n">low_thdot</span><span class="p">)</span>
<span class="o">+</span> <span class="n">low_thdot</span>
<span class="p">)</span>
<span class="n">out</span> <span class="o">=</span> <span class="n">TensorDict</span><span class="p">(</span>
<span class="p">{</span>
<span class="s2">"th"</span><span class="p">:</span> <span class="n">th</span><span class="p">,</span>
<span class="s2">"thdot"</span><span class="p">:</span> <span class="n">thdot</span><span class="p">,</span>
<span class="s2">"params"</span><span class="p">:</span> <span class="n">tensordict</span><span class="p">[</span><span class="s2">"params"</span><span class="p">],</span>
<span class="p">},</span>
<span class="n">batch_size</span><span class="o">=</span><span class="n">tensordict</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">out</span>
</pre></div>
</div>
</div>
<div class="section" id="environment-metadata-env-spec">
<h2>Environment metadata: <code class="docutils literal notranslate"><span class="pre">env.*_spec</span></code><a class="headerlink" href="#environment-metadata-env-spec" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>The specs define the input and output domain of the environment.
It is important that the specs accurately define the tensors that will be
received at runtime, as they are often used to carry information about
environments in multiprocessing and distributed settings. They can also be
used to instantiate lazily defined neural networks and test scripts without
actually querying the environment (which can be costly with real-world
physical systems for instance).</p>
<p>There are four specs that we must code in our environment:</p>
<ul class="simple">
<li><p><code class="xref py py-obj docutils literal notranslate"><span class="pre">EnvBase.observation_spec</span></code>: This will be a <code class="xref py py-class docutils literal notranslate"><span class="pre">CompositeSpec</span></code>
instance where each key is an observation (a <code class="xref py py-class docutils literal notranslate"><span class="pre">CompositeSpec</span></code> can be
viewed as a dictionary of specs).</p></li>
<li><p><code class="xref py py-obj docutils literal notranslate"><span class="pre">EnvBase.action_spec</span></code>: It can be any type of spec, but it is required
that it corresponds to the <code class="docutils literal notranslate"><span class="pre">"action"</span></code> entry in the input <code class="docutils literal notranslate"><span class="pre">tensordict</span></code>;</p></li>
<li><p><code class="xref py py-obj docutils literal notranslate"><span class="pre">EnvBase.reward_spec</span></code>: provides information about the reward space;</p></li>
<li><p><code class="xref py py-obj docutils literal notranslate"><span class="pre">EnvBase.done_spec</span></code>: provides information about the space of the done
flag.</p></li>
</ul>
<p>TorchRL specs are organized in two general containers: <code class="docutils literal notranslate"><span class="pre">input_spec</span></code> which
contains the specs of the information that the step function reads (divided
between <code class="docutils literal notranslate"><span class="pre">action_spec</span></code> containing the action and <code class="docutils literal notranslate"><span class="pre">state_spec</span></code> containing
all the rest), and <code class="docutils literal notranslate"><span class="pre">output_spec</span></code> which encodes the specs that the
step outputs (<code class="docutils literal notranslate"><span class="pre">observation_spec</span></code>, <code class="docutils literal notranslate"><span class="pre">reward_spec</span></code> and <code class="docutils literal notranslate"><span class="pre">done_spec</span></code>).
In general, you should not interact directly with <code class="docutils literal notranslate"><span class="pre">output_spec</span></code> and
<code class="docutils literal notranslate"><span class="pre">input_spec</span></code> but only with their content: <code class="docutils literal notranslate"><span class="pre">observation_spec</span></code>,
<code class="docutils literal notranslate"><span class="pre">reward_spec</span></code>, <code class="docutils literal notranslate"><span class="pre">done_spec</span></code>, <code class="docutils literal notranslate"><span class="pre">action_spec</span></code> and <code class="docutils literal notranslate"><span class="pre">state_spec</span></code>.
The reason if that the specs are organized in a non-trivial way
within <code class="docutils literal notranslate"><span class="pre">output_spec</span></code> and
<code class="docutils literal notranslate"><span class="pre">input_spec</span></code> and neither of these should be directly modified.</p>
<p>In other words, the <code class="docutils literal notranslate"><span class="pre">observation_spec</span></code> and related properties are
convenient shortcuts to the content of the output and input spec containers.</p>
<p>TorchRL offers multiple <code class="xref py py-class docutils literal notranslate"><span class="pre">TensorSpec</span></code>
<a class="reference external" href="https://pytorch.org/rl/reference/data.html#tensorspec">subclasses</a> to
encode the environment’s input and output characteristics.</p>
<div class="section" id="specs-shape">
<h3>Specs shape<a class="headerlink" href="#specs-shape" title="이 제목에 대한 퍼머링크">¶</a></h3>
<p>The environment specs leading dimensions must match the
environment batch-size. This is done to enforce that every component of an
environment (including its transforms) have an accurate representation of
the expected input and output shapes. This is something that should be
accurately coded in stateful settings.</p>
<p>For non batch-locked environments, such as the one in our example (see below),
this is irrelevant as the environment batch size will most likely be empty.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">_make_spec</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">td_params</span><span class="p">):</span>
<span class="c1"># Under the hood, this will populate self.output_spec["observation"]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">observation_spec</span> <span class="o">=</span> <span class="n">CompositeSpec</span><span class="p">(</span>
<span class="n">th</span><span class="o">=</span><span class="n">BoundedTensorSpec</span><span class="p">(</span>
<span class="n">low</span><span class="o">=-</span><span class="n">torch</span><span class="o">.</span><span class="n">pi</span><span class="p">,</span>
<span class="n">high</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">pi</span><span class="p">,</span>
<span class="n">shape</span><span class="o">=</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">float32</span><span class="p">,</span>
<span class="p">),</span>
<span class="n">thdot</span><span class="o">=</span><span class="n">BoundedTensorSpec</span><span class="p">(</span>
<span class="n">low</span><span class="o">=-</span><span class="n">td_params</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_speed"</span><span class="p">],</span>
<span class="n">high</span><span class="o">=</span><span class="n">td_params</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_speed"</span><span class="p">],</span>
<span class="n">shape</span><span class="o">=</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">float32</span><span class="p">,</span>
<span class="p">),</span>
<span class="c1"># we need to add the ``params`` to the observation specs, as we want</span>
<span class="c1"># to pass it at each step during a rollout</span>
<span class="n">params</span><span class="o">=</span><span class="n">make_composite_from_td</span><span class="p">(</span><span class="n">td_params</span><span class="p">[</span><span class="s2">"params"</span><span class="p">]),</span>
<span class="n">shape</span><span class="o">=</span><span class="p">(),</span>
<span class="p">)</span>
<span class="c1"># since the environment is stateless, we expect the previous output as input.</span>
<span class="c1"># For this, ``EnvBase`` expects some state_spec to be available</span>
<span class="bp">self</span><span class="o">.</span><span class="n">state_spec</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">observation_spec</span><span class="o">.</span><span class="n">clone</span><span class="p">()</span>
<span class="c1"># action-spec will be automatically wrapped in input_spec when</span>
<span class="c1"># `self.action_spec = spec` will be called supported</span>
<span class="bp">self</span><span class="o">.</span><span class="n">action_spec</span> <span class="o">=</span> <span class="n">BoundedTensorSpec</span><span class="p">(</span>
<span class="n">low</span><span class="o">=-</span><span class="n">td_params</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_torque"</span><span class="p">],</span>
<span class="n">high</span><span class="o">=</span><span class="n">td_params</span><span class="p">[</span><span class="s2">"params"</span><span class="p">,</span> <span class="s2">"max_torque"</span><span class="p">],</span>
<span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="mi">1</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">float32</span><span class="p">,</span>
<span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">reward_spec</span> <span class="o">=</span> <span class="n">UnboundedContinuousTensorSpec</span><span class="p">(</span><span class="n">shape</span><span class="o">=</span><span class="p">(</span><span class="o">*</span><span class="n">td_params</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">make_composite_from_td</span><span class="p">(</span><span class="n">td</span><span class="p">):</span>
<span class="c1"># custom function to convert a ``tensordict`` in a similar spec structure</span>
<span class="c1"># of unbounded values.</span>
<span class="n">composite</span> <span class="o">=</span> <span class="n">CompositeSpec</span><span class="p">(</span>
<span class="p">{</span>
<span class="n">key</span><span class="p">:</span> <span class="n">make_composite_from_td</span><span class="p">(</span><span class="n">tensor</span><span class="p">)</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">tensor</span><span class="p">,</span> <span class="n">TensorDictBase</span><span class="p">)</span>
<span class="k">else</span> <span class="n">UnboundedContinuousTensorSpec</span><span class="p">(</span>
<span class="n">dtype</span><span class="o">=</span><span class="n">tensor</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">tensor</span><span class="o">.</span><span class="n">device</span><span class="p">,</span> <span class="n">shape</span><span class="o">=</span><span class="n">tensor</span><span class="o">.</span><span class="n">shape</span>
<span class="p">)</span>
<span class="k">for</span> <span class="n">key</span><span class="p">,</span> <span class="n">tensor</span> <span class="ow">in</span> <span class="n">td</span><span class="o">.</span><span class="n">items</span><span class="p">()</span>
<span class="p">},</span>
<span class="n">shape</span><span class="o">=</span><span class="n">td</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span>
<span class="p">)</span>
<span class="k">return</span> <span class="n">composite</span>
</pre></div>
</div>
</div>
</div>
<div class="section" id="reproducible-experiments-seeding">
<h2>Reproducible experiments: seeding<a class="headerlink" href="#reproducible-experiments-seeding" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>Seeding an environment is a common operation when initializing an experiment.
The only goal of <code class="xref py py-func docutils literal notranslate"><span class="pre">EnvBase._set_seed()</span></code> is to set the seed of the contained
simulator. If possible, this operation should not call <code class="docutils literal notranslate"><span class="pre">reset()</span></code> or interact
with the environment execution. The parent <code class="xref py py-func docutils literal notranslate"><span class="pre">EnvBase.set_seed()</span></code> method
incorporates a mechanism that allows seeding multiple environments with a
different pseudo-random and reproducible seed.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">_set_seed</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">seed</span><span class="p">:</span> <span class="n">Optional</span><span class="p">[</span><span class="nb">int</span><span class="p">]):</span>
<span class="n">rng</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">manual_seed</span><span class="p">(</span><span class="n">seed</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">rng</span> <span class="o">=</span> <span class="n">rng</span>
</pre></div>
</div>
</div>
<div class="section" id="wrapping-things-together-the-envbase-class">
<h2>Wrapping things together: the <code class="xref py py-class docutils literal notranslate"><span class="pre">EnvBase</span></code> class<a class="headerlink" href="#wrapping-things-together-the-envbase-class" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>We can finally put together the pieces and design our environment class.
The specs initialization needs to be performed during the environment
construction, so we must take care of calling the <code class="xref py py-func docutils literal notranslate"><span class="pre">_make_spec()</span></code> method
within <code class="xref py py-func docutils literal notranslate"><span class="pre">PendulumEnv.__init__()</span></code>.</p>
<p>We add a static method <code class="xref py py-meth docutils literal notranslate"><span class="pre">PendulumEnv.gen_params()</span></code> which deterministically
generates a set of hyperparameters to be used during execution:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">gen_params</span><span class="p">(</span><span class="n">g</span><span class="o">=</span><span class="mf">10.0</span><span class="p">,</span> <span class="n">batch_size</span><span class="o">=</span><span class="kc">None</span><span class="p">)</span> <span class="o">-></span> <span class="n">TensorDictBase</span><span class="p">:</span>
<span class="w"> </span><span class="sd">"""Returns a ``tensordict`` containing the physical parameters such as gravitational force and torque or speed limits."""</span>
<span class="k">if</span> <span class="n">batch_size</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">batch_size</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">td</span> <span class="o">=</span> <span class="n">TensorDict</span><span class="p">(</span>
<span class="p">{</span>
<span class="s2">"params"</span><span class="p">:</span> <span class="n">TensorDict</span><span class="p">(</span>
<span class="p">{</span>
<span class="s2">"max_speed"</span><span class="p">:</span> <span class="mi">8</span><span class="p">,</span>
<span class="s2">"max_torque"</span><span class="p">:</span> <span class="mf">2.0</span><span class="p">,</span>
<span class="s2">"dt"</span><span class="p">:</span> <span class="mf">0.05</span><span class="p">,</span>
<span class="s2">"g"</span><span class="p">:</span> <span class="n">g</span><span class="p">,</span>
<span class="s2">"m"</span><span class="p">:</span> <span class="mf">1.0</span><span class="p">,</span>
<span class="s2">"l"</span><span class="p">:</span> <span class="mf">1.0</span><span class="p">,</span>
<span class="p">},</span>
<span class="p">[],</span>
<span class="p">)</span>
<span class="p">},</span>
<span class="p">[],</span>
<span class="p">)</span>
<span class="k">if</span> <span class="n">batch_size</span><span class="p">:</span>
<span class="n">td</span> <span class="o">=</span> <span class="n">td</span><span class="o">.</span><span class="n">expand</span><span class="p">(</span><span class="n">batch_size</span><span class="p">)</span><span class="o">.</span><span class="n">contiguous</span><span class="p">()</span>
<span class="k">return</span> <span class="n">td</span>
</pre></div>
</div>
<p>We define the environment as non-<code class="docutils literal notranslate"><span class="pre">batch_locked</span></code> by turning the <code class="docutils literal notranslate"><span class="pre">homonymous</span></code>
attribute to <code class="docutils literal notranslate"><span class="pre">False</span></code>. This means that we will <strong>not</strong> enforce the input
<code class="docutils literal notranslate"><span class="pre">tensordict</span></code> to have a <code class="docutils literal notranslate"><span class="pre">batch-size</span></code> that matches the one of the environment.</p>
<p>The following code will just put together the pieces we have coded above.</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">PendulumEnv</span><span class="p">(</span><span class="n">EnvBase</span><span class="p">):</span>
<span class="n">metadata</span> <span class="o">=</span> <span class="p">{</span>
<span class="s2">"render_modes"</span><span class="p">:</span> <span class="p">[</span><span class="s2">"human"</span><span class="p">,</span> <span class="s2">"rgb_array"</span><span class="p">],</span>
<span class="s2">"render_fps"</span><span class="p">:</span> <span class="mi">30</span><span class="p">,</span>
<span class="p">}</span>
<span class="n">batch_locked</span> <span class="o">=</span> <span class="kc">False</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">td_params</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">seed</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cpu"</span><span class="p">):</span>
<span class="k">if</span> <span class="n">td_params</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">td_params</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">gen_params</span><span class="p">()</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</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">batch_size</span><span class="o">=</span><span class="p">[])</span>
<span class="bp">self</span><span class="o">.</span><span class="n">_make_spec</span><span class="p">(</span><span class="n">td_params</span><span class="p">)</span>
<span class="k">if</span> <span class="n">seed</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">seed</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">empty</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">int64</span><span class="p">)</span><span class="o">.</span><span class="n">random_</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">set_seed</span><span class="p">(</span><span class="n">seed</span><span class="p">)</span>
<span class="c1"># Helpers: _make_step and gen_params</span>
<span class="n">gen_params</span> <span class="o">=</span> <span class="nb">staticmethod</span><span class="p">(</span><span class="n">gen_params</span><span class="p">)</span>
<span class="n">_make_spec</span> <span class="o">=</span> <span class="n">_make_spec</span>
<span class="c1"># Mandatory methods: _step, _reset and _set_seed</span>
<span class="n">_reset</span> <span class="o">=</span> <span class="n">_reset</span>
<span class="n">_step</span> <span class="o">=</span> <span class="nb">staticmethod</span><span class="p">(</span><span class="n">_step</span><span class="p">)</span>
<span class="n">_set_seed</span> <span class="o">=</span> <span class="n">_set_seed</span>
</pre></div>
</div>
</div>
<div class="section" id="testing-our-environment">
<h2>Testing our environment<a class="headerlink" href="#testing-our-environment" title="이 제목에 대한 퍼머링크">¶</a></h2>
<p>TorchRL provides a simple function <code class="xref py py-func docutils literal notranslate"><span class="pre">check_env_specs()</span></code>
to check that a (transformed) environment has an input/output structure that
matches the one dictated by its specs.
Let us try it out:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">env</span> <span class="o">=</span> <span class="n">PendulumEnv</span><span class="p">()</span>
<span class="n">check_env_specs</span><span class="p">(</span><span class="n">env</span><span class="p">)</span>
</pre></div>
</div>
<p>We can have a look at our specs to have a visual representation of the environment
signature:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="nb">print</span><span class="p">(</span><span class="s2">"observation_spec:"</span><span class="p">,</span> <span class="n">env</span><span class="o">.</span><span class="n">observation_spec</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"state_spec:"</span><span class="p">,</span> <span class="n">env</span><span class="o">.</span><span class="n">state_spec</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"reward_spec:"</span><span class="p">,</span> <span class="n">env</span><span class="o">.</span><span class="n">reward_spec</span><span class="p">)</span>
</pre></div>
</div>
<div class="sphx-glr-script-out highlight-none notranslate"><div class="highlight"><pre><span></span>observation_spec: CompositeSpec(
th: BoundedTensorSpec(
shape=torch.Size([]),
space=ContinuousBox(
low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True),
high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)),
device=cpu,
dtype=torch.float32,
domain=continuous),
thdot: BoundedTensorSpec(
shape=torch.Size([]),
space=ContinuousBox(
low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True),
high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)),
device=cpu,
dtype=torch.float32,
domain=continuous),
params: CompositeSpec(
max_speed: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.int64,
domain=discrete),
max_torque: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous),
dt: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous),
g: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous),
m: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous),
l: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,
device=cpu,
dtype=torch.float32,
domain=continuous), device=cpu, shape=torch.Size([])), device=cpu, shape=torch.Size([]))
state_spec: CompositeSpec(
th: BoundedTensorSpec(
shape=torch.Size([]),
space=ContinuousBox(
low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True),
high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)),
device=cpu,
dtype=torch.float32,
domain=continuous),
thdot: BoundedTensorSpec(
shape=torch.Size([]),
space=ContinuousBox(
low=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True),
high=Tensor(shape=torch.Size([]), device=cpu, dtype=torch.float32, contiguous=True)),
device=cpu,
dtype=torch.float32,
domain=continuous),
params: CompositeSpec(
max_speed: UnboundedContinuousTensorSpec(
shape=torch.Size([]),
space=None,