-
Notifications
You must be signed in to change notification settings - Fork 258
/
Copy pathrpc_async_execution.html
1249 lines (1038 loc) ยท 115 KB
/
rpc_async_execution.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="Implementing Batch RPC Processing Using Asynchronous Executions" />
<meta property="og:type" content="article" />
<meta property="og:url" content="https://tutorials.pytorch.kr/intermediate/rpc_async_execution.html" />
<meta property="og:site_name" content="PyTorch Tutorials KR" />
<meta property="og:description" content="Author: Shen Li Prerequisites: PyTorch Distributed Overview, Getting started with Distributed RPC Framework, Implementing a Parameter Server using Distributed RPC Framework, RPC Asynchronous Execution Decorator. This tutorial demonstrates how to build batch-processing RPC applications with the@rp..." />
<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: Shen Li Prerequisites: PyTorch Distributed Overview, Getting started with Distributed RPC Framework, Implementing a Parameter Server using Distributed RPC Framework, RPC Asynchronous Execution Decorator. This tutorial demonstrates how to build batch-processing RPC applications with the@rp..." />
<meta property="og:ignore_canonical" content="true" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Implementing Batch RPC Processing Using Asynchronous Executions — ํ์ดํ ์น ํ๊ตญ์ด ํํ ๋ฆฌ์ผ (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="์์ธ" href="../genindex.html" />
<link rel="search" title="๊ฒ์" href="../search.html" />
<link rel="next" title="๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ(DDP)๊ณผ ๋ถ์ฐ RPC ํ๋ ์์ํฌ ๊ฒฐํฉ" href="../advanced/rpc_ddp_tutorial.html" />
<link rel="prev" title="Distributed Pipeline Parallelism Using RPC" href="dist_pipeline_parallel_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">์์ํ๊ธฐ</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr/">ํํ ๋ฆฌ์ผ</a>
</li>
<li>
<a href="https://pytorch.kr/hub">ํ๋ธ</a>
</li>
<li>
<a href="https://discuss.pytorch.kr/">์ปค๋ฎค๋ํฐ</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">ํ์ดํ ์น(PyTorch) ๋ ์ํผ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../recipes/recipes_index.html">๋ชจ๋ ๋ ์ํผ ๋ณด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../prototype/prototype_index.html">๋ชจ๋ ํ๋กํ ํ์
๋ ์ํผ ๋ณด๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ์ดํ ์น(PyTorch) ์์ํ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/intro.html">ํ์ดํ ์น(PyTorch) ๊ธฐ๋ณธ ์ตํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/quickstart_tutorial.html">๋น ๋ฅธ ์์(Quickstart)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/tensorqs_tutorial.html">ํ
์(Tensor)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/data_tutorial.html">Dataset๊ณผ DataLoader</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/transforms_tutorial.html">๋ณํ(Transform)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/buildmodel_tutorial.html">์ ๊ฒฝ๋ง ๋ชจ๋ธ ๊ตฌ์ฑํ๊ธฐ</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>๋ฅผ ์ฌ์ฉํ ์๋ ๋ฏธ๋ถ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/optimization_tutorial.html">๋ชจ๋ธ ๋งค๊ฐ๋ณ์ ์ต์ ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/saveloadrun_tutorial.html">๋ชจ๋ธ ์ ์ฅํ๊ณ ๋ถ๋ฌ์ค๊ธฐ</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 ์๊ฐ - YouTube ์๋ฆฌ์ฆ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">PyTorch ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Pytorch Tensor ์๊ฐ</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">ํ์ดํ ์น(PyTorch) ๋ฐฐ์ฐ๊ธฐ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deep_learning_60min_blitz.html">PyTorch๋ก ๋ฅ๋ฌ๋ํ๊ธฐ: 60๋ถ๋ง์ ๋์ฅ๋ด๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/pytorch_with_examples.html">์์ ๋ก ๋ฐฐ์ฐ๋ ํ์ดํ ์น(PyTorch)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/nn_tutorial.html"><cite>torch.nn</cite> ์ด <em>์ค์ ๋ก</em> ๋ฌด์์ธ๊ฐ์?</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_tutorial.html">TensorBoard๋ก ๋ชจ๋ธ, ๋ฐ์ดํฐ, ํ์ต ์๊ฐํํ๊ธฐ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">์ด๋ฏธ์ง/๋น๋์ค</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">์ปดํจํฐ ๋น์ (Vision)์ ์ํ ์ ์ดํ์ต(Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/fgsm_tutorial.html">์ ๋์ ์์ ์์ฑ(Adversarial Example Generation)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dcgan_faces_tutorial.html">DCGAN ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</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">์ค๋์ค</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">wav2vec2์ ์ด์ฉํ ๊ฐ์ ์ ๋ ฌ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">ํ
์คํธ</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">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ๋ถ๋ฅํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="char_rnn_generation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: ๋ฌธ์-๋จ์ RNN์ผ๋ก ์ด๋ฆ ์์ฑํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="seq2seq_translation_tutorial.html">๊ธฐ์ด๋ถํฐ ์์ํ๋ NLP: Sequence to Sequence ๋คํธ์ํฌ์ Attention์ ์ด์ฉํ ๋ฒ์ญ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/text_sentiment_ngrams_tutorial.html">torchtext ๋ผ์ด๋ธ๋ฌ๋ฆฌ๋ก ํ
์คํธ ๋ถ๋ฅํ๊ธฐ</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> ์ torchtext๋ก ์ธ์ด ๋ฒ์ญํ๊ธฐ</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">๋ฐฑ์๋</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">๊ฐํํ์ต</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="reinforcement_q_learning.html">๊ฐํ ํ์ต (DQN) ํํ ๋ฆฌ์ผ</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">๋ง๋ฆฌ์ค ๊ฒ์ RL ์์ด์ ํธ๋ก ํ์ตํ๊ธฐ</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 ๋ชจ๋ธ์ ํ๋ก๋์
ํ๊ฒฝ์ ๋ฐฐํฌํ๊ธฐ</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">Flask๋ฅผ ์ฌ์ฉํ์ฌ Python์์ PyTorch๋ฅผ REST API๋ก ๋ฐฐํฌํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/Intro_to_TorchScript_tutorial.html">TorchScript ์๊ฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_export.html">C++์์ TorchScript ๋ชจ๋ธ ๋ก๋ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/super_resolution_with_onnxruntime.html">(์ ํ) PyTorch ๋ชจ๋ธ์ ONNX์ผ๋ก ๋ณํํ๊ณ ONNX ๋ฐํ์์์ ์คํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="realtime_rpi.html">Raspberry Pi 4 ์์ ์ค์๊ฐ ์ถ๋ก (Inference) (30fps!)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ๋กํ์ผ๋ง</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</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">(๋ฒ ํ) FX์์ ํฉ์ฑ๊ณฑ/๋ฐฐ์น ์ ๊ทํ(Convolution/Batch Norm) ๊ฒฐํฉ๊ธฐ(Fuser) ๋ง๋ค๊ธฐ</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">ํ๋ก ํธ์๋ API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="memory_format_tutorial.html">(๋ฒ ํ) PyTorch๋ฅผ ์ฌ์ฉํ Channels Last ๋ฉ๋ชจ๋ฆฌ ํ์</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">๋ชจ๋ธ ์์๋ธ</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++ ํ๋ก ํธ์๋ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/torch-script-parallelism.html">TorchScript์ ๋์ ๋ณ๋ ฌ ์ฒ๋ฆฌ(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/cpp_autograd.html">C++ ํ๋ก ํธ์๋์ ์๋ ๋ฏธ๋ถ (autograd)</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">PyTorch ํ์ฅํ๊ธฐ</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">์ปค์คํ
C++ ํด๋์ค๋ก TorchScript ํ์ฅํ๊ธฐ</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">๋ชจ๋ธ ์ต์ ํ</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch ๋ชจ๋ ํ๋กํ์ผ๋งํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard_profiler_tutorial.html">ํ
์๋ณด๋๋ฅผ ์ด์ฉํ ํ์ดํ ์น ํ๋กํ์ผ๋ฌ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Ray Tune์ ์ฌ์ฉํ ํ์ดํผํ๋ผ๋ฏธํฐ ํ๋</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">๋ฐฐํฌ๋ฅผ ์ํด ๋น์ ํธ๋์คํฌ๋จธ(Vision Transformer) ๋ชจ๋ธ ์ต์ ํํ๊ธฐ</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">๊ฐ์ง์น๊ธฐ ๊ธฐ๋ฒ(Pruning) ํํ ๋ฆฌ์ผ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/dynamic_quantization_tutorial.html">(๋ฒ ํ) LSTM ๊ธฐ๋ฐ ๋จ์ด ๋จ์ ์ธ์ด ๋ชจ๋ธ์ ๋์ ์์ํ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_bert_tutorial.html">(๋ฒ ํ) BERT ๋ชจ๋ธ ๋์ ์์ํํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantized_transfer_learning_tutorial.html">(๋ฒ ํ) ์ปดํจํฐ ๋น์ ํํ ๋ฆฌ์ผ์ ์ํ ์์ํ๋ ์ ์ดํ์ต(Quantized Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/static_quantization_tutorial.html">(๋ฒ ํ) PyTorch์์ Eager Mode๋ฅผ ์ด์ฉํ ์ ์ ์์ํ</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)๋ก ๊ณ ์ฑ๋ฅ ํธ๋์คํฌ๋จธ(Transformers) ๊ตฌํํ๊ธฐ</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> ๊ณผ ํจ๊ป SDPA ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#sdpa-atteition-bias">SDPA๋ฅผ <code class="docutils literal notranslate"><span class="pre">atteition.bias</span></code> ํ์ ํด๋์ค์ ์ฌ์ฉํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="scaled_dot_product_attention_tutorial.html#id8">๊ฒฐ๋ก </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">๋ณ๋ ฌ ๋ฐ ๋ถ์ฐ ํ์ต</span></p>
<ul class="current">
<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">๋จ์ผ ๋จธ์ ์ ์ฌ์ฉํ ๋ชจ๋ธ ๋ณ๋ ฌํ ๋ชจ๋ฒ ์ฌ๋ก</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ ์์ํ๊ธฐ</a></li>
<li class="toctree-l1"><a class="reference internal" href="dist_tuto.html">PyTorch๋ก ๋ถ์ฐ ์ดํ๋ฆฌ์ผ์ด์
๊ฐ๋ฐํ๊ธฐ</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 ํ์ฅ์ ์ฌ์ฉํ ํ๋ก์ธ์ค ๊ทธ๋ฃน ๋ฐฑ์๋ ์ฌ์ฉ์ ์ ์</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 current"><a class="current reference internal" href="#">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/rpc_ddp_tutorial.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ(DDP)๊ณผ ๋ถ์ฐ RPC ํ๋ ์์ํฌ ๊ฒฐํฉ</a></li>
<li class="toctree-l1"><a class="reference internal" href="../advanced/ddp_pipeline.html">๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ ์ฒ๋ฆฌ์ ๋ณ๋ ฌ ์ฒ๋ฆฌ ํ์ดํ๋ผ์ธ์ ์ฌ์ฉํ ํธ๋์คํฌ๋จธ ๋ชจ๋ธ ํ์ต</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">์ถ์ฒ ์์คํ
</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchrec_tutorial.html">TorchRec ์๊ฐ</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 ํํ ๋ฆฌ์ผ: FLAVA ๋ฏธ์ธ์กฐ์ </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>Implementing Batch RPC Processing Using Asynchronous Executions</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/intermediate/rpc_async_execution.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/rpc_async_execution</div>
<div id="google-colab-link">
<img class="call-to-action-img" src="../_static/images/pytorch-colab.svg"/>
<div class="call-to-action-desktop-view">Run in Google Colab</div>
<div class="call-to-action-mobile-view">Colab</div>
</div>
<div id="download-notebook-link">
<img class="call-to-action-notebook-img" src="../_static/images/pytorch-download.svg"/>
<div class="call-to-action-desktop-view">Download Notebook</div>
<div class="call-to-action-mobile-view">Notebook</div>
</div>
<div id="github-view-link">
<img class="call-to-action-img" src="../_static/images/pytorch-github.svg"/>
<div class="call-to-action-desktop-view">View on GitHub</div>
<div class="call-to-action-mobile-view">GitHub</div>
</div>
</div>
<div class="rst-content">
<div role="main" class="main-content" itemscope="itemscope" itemtype="http://schema.org/Article">
<article itemprop="articleBody" id="pytorch-article" class="pytorch-article">
<div class="section" id="implementing-batch-rpc-processing-using-asynchronous-executions">
<h1>Implementing Batch RPC Processing Using Asynchronous Executions<a class="headerlink" href="#implementing-batch-rpc-processing-using-asynchronous-executions" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h1>
<p><strong>Author</strong>: <a class="reference external" href="https://mrshenli.github.io/">Shen Li</a></p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p><a class="reference internal" href="../_images/pencil-16.png"><img alt="edit" src="../_images/pencil-16.png" style="width: 16px; height: 16px;" /></a> View and edit this tutorial in <a class="reference external" href="https://github.com/pytorch/tutorials/blob/main/intermediate_source/rpc_async_execution.rst">github</a>.</p>
</div>
<p>Prerequisites:</p>
<ul class="simple">
<li><p><a class="reference external" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></p></li>
<li><p><a class="reference external" href="rpc_tutorial.html">Getting started with Distributed RPC Framework</a></p></li>
<li><p><a class="reference external" href="rpc_param_server_tutorial.html">Implementing a Parameter Server using Distributed RPC Framework</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">RPC Asynchronous Execution Decorator</a></p></li>
</ul>
<p>This tutorial demonstrates how to build batch-processing RPC applications with
the <a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
decorator, which helps to speed up training by reducing the number of blocked
RPC threads and consolidating CUDA operations on the callee. This shares the
same idea as <a class="reference external" href="https://pytorch.org/serve/batch_inference_with_ts.html">Batch Inference with TorchServe</a>.</p>
<div class="admonition note">
<p class="admonition-title">์ฐธ๊ณ </p>
<p>This tutorial requires PyTorch v1.6.0 or above.</p>
</div>
<div class="section" id="basics">
<h2>Basics<a class="headerlink" href="#basics" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>Previous tutorials have shown the steps to build distributed training
applications using <a class="reference external" href="https://pytorch.org/docs/stable/rpc.html">torch.distributed.rpc</a>,
but they didnโt elaborate on what happens on the callee side when processing an
RPC request. As of PyTorch v1.5, each RPC request will block one thread on the
callee to execute the function in that request until that function returns.
This works for many use cases, but there is one caveat. If the user function
blocks on IO, e.g., with nested RPC invocation, or signaling, e.g., waiting for
a different RPC request to unblock, the RPC thread on the callee will have to
idle waiting until the IO finishes or the signaling event occurs. As a result,
RPC callees are likely to use more threads than necessary. The cause of this
problem is that RPC treats user functions as black boxes, and knows very little
about what happens in the function. To allow user functions to yield and free
RPC threads, more hints need to be provided to the RPC system.</p>
<p>Since v1.6.0, PyTorch addresses this problem by introducing two new concepts:</p>
<ul class="simple">
<li><p>A <a class="reference external" href="https://pytorch.org/docs/master/futures.html">torch.futures.Future</a> type
that encapsulates an asynchronous execution, which also supports installing
callback functions.</p></li>
<li><p>An <a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
decorator that allows applications to tell the callee that the target function
will return a future and can pause and yield multiple times during execution.</p></li>
</ul>
<p>With these two tools, the application code can break a user function into
multiple smaller functions, chain them together as callbacks on <code class="docutils literal notranslate"><span class="pre">Future</span></code>
objects, and return the <code class="docutils literal notranslate"><span class="pre">Future</span></code> that contains the final result. On the callee
side, when getting the <code class="docutils literal notranslate"><span class="pre">Future</span></code> object, it installs subsequent RPC response
preparation and communication as callbacks as well, which will be triggered
when the final result is ready. In this way, the callee no longer needs to block
one thread and wait until the final return value is ready. Please refer to the
API doc of
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
for simple examples.</p>
<p>Besides reducing the number of idle threads on the callee, these tools also help
to make batch RPC processing easier and faster. The following two sections of
this tutorial demonstrate how to build distributed batch-updating parameter
server and batch-processing reinforcement learning applications using the
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
decorator.</p>
</div>
<div class="section" id="batch-updating-parameter-server">
<h2>Batch-Updating Parameter Server<a class="headerlink" href="#batch-updating-parameter-server" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>Consider a synchronized parameter server training application with one parameter
server (PS) and multiple trainers. In this application, the PS holds the
parameters and waits for all trainers to report gradients. In every iteration,
it waits until receiving gradients from all trainers and then updates all
parameters in one shot. The code below shows the implementation of the PS class.
The <code class="docutils literal notranslate"><span class="pre">update_and_fetch_model</span></code> method is decorated using
<code class="docutils literal notranslate"><span class="pre">@rpc.functions.async_execution</span></code> and will be called by trainers. Each
invocation returns a <code class="docutils literal notranslate"><span class="pre">Future</span></code> object that will be populated with the updated
model. Invocations launched by most trainers just accumulate gradients to the
<code class="docutils literal notranslate"><span class="pre">.grad</span></code> field, return immediately, and yield the RPC thread on the PS. The
last arriving trainer will trigger the optimizer step and consume all previously
reported gradients. Then it sets the <code class="docutils literal notranslate"><span class="pre">future_model</span></code> with the updated model,
which in turn notifies all previous requests from other trainers through the
<code class="docutils literal notranslate"><span class="pre">Future</span></code> object and sends out the updated model to all trainers.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">threading</span>
<span class="kn">import</span> <span class="nn">torchvision</span>
<span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">optim</span>
<span class="n">num_classes</span><span class="p">,</span> <span class="n">batch_update_size</span> <span class="o">=</span> <span class="mi">30</span><span class="p">,</span> <span class="mi">5</span>
<span class="k">class</span> <span class="nc">BatchUpdateParameterServer</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">batch_update_size</span><span class="o">=</span><span class="n">batch_update_size</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">model</span> <span class="o">=</span> <span class="n">torchvision</span><span class="o">.</span><span class="n">models</span><span class="o">.</span><span class="n">resnet50</span><span class="p">(</span><span class="n">num_classes</span><span class="o">=</span><span class="n">num_classes</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">lock</span> <span class="o">=</span> <span class="n">threading</span><span class="o">.</span><span class="n">Lock</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">future_model</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">futures</span><span class="o">.</span><span class="n">Future</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">batch_update_size</span> <span class="o">=</span> <span class="n">batch_update_size</span>
<span class="bp">self</span><span class="o">.</span><span class="n">curr_update_size</span> <span class="o">=</span> <span class="mi">0</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.001</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
<span class="k">for</span> <span class="n">p</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">():</span>
<span class="n">p</span><span class="o">.</span><span class="n">grad</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">p</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">get_model</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span>
<span class="nd">@staticmethod</span>
<span class="nd">@rpc</span><span class="o">.</span><span class="n">functions</span><span class="o">.</span><span class="n">async_execution</span>
<span class="k">def</span> <span class="nf">update_and_fetch_model</span><span class="p">(</span><span class="n">ps_rref</span><span class="p">,</span> <span class="n">grads</span><span class="p">):</span>
<span class="c1"># Using the RRef to retrieve the local PS instance</span>
<span class="bp">self</span> <span class="o">=</span> <span class="n">ps_rref</span><span class="o">.</span><span class="n">local_value</span><span class="p">()</span>
<span class="k">with</span> <span class="bp">self</span><span class="o">.</span><span class="n">lock</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">curr_update_size</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="c1"># accumulate gradients into .grad field</span>
<span class="k">for</span> <span class="n">p</span><span class="p">,</span> <span class="n">g</span> <span class="ow">in</span> <span class="nb">zip</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">grads</span><span class="p">):</span>
<span class="n">p</span><span class="o">.</span><span class="n">grad</span> <span class="o">+=</span> <span class="n">g</span>
<span class="c1"># Save the current future_model and return it to make sure the</span>
<span class="c1"># returned Future object holds the correct model even if another</span>
<span class="c1"># thread modifies future_model before this thread returns.</span>
<span class="n">fut</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">future_model</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">curr_update_size</span> <span class="o">>=</span> <span class="bp">self</span><span class="o">.</span><span class="n">batch_update_size</span><span class="p">:</span>
<span class="c1"># update the model</span>
<span class="k">for</span> <span class="n">p</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">():</span>
<span class="n">p</span><span class="o">.</span><span class="n">grad</span> <span class="o">/=</span> <span class="bp">self</span><span class="o">.</span><span class="n">batch_update_size</span>
<span class="bp">self</span><span class="o">.</span><span class="n">curr_update_size</span> <span class="o">=</span> <span class="mi">0</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="c1"># by settiing the result on the Future object, all previous</span>
<span class="c1"># requests expecting this updated model will be notified and</span>
<span class="c1"># the their responses will be sent accordingly.</span>
<span class="n">fut</span><span class="o">.</span><span class="n">set_result</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">model</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">future_model</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">futures</span><span class="o">.</span><span class="n">Future</span><span class="p">()</span>
<span class="k">return</span> <span class="n">fut</span>
</pre></div>
</div>
<p>For the trainers, they are all initialized using the same set of
parameters from the PS. In every iteration, each trainer first runs the forward
and the backward passes to generate gradients locally. Then, each trainer
reports its gradients to the PS using RPC, and fetches back the updated
parameters through the return value of the same RPC request. In the trainerโs
implementation, whether the target function is marked with
<code class="docutils literal notranslate"><span class="pre">@rpc.functions.async_execution</span></code> or not makes no difference. The
trainer simply calls <code class="docutils literal notranslate"><span class="pre">update_and_fetch_model</span></code> using <code class="docutils literal notranslate"><span class="pre">rpc_sync</span></code> which will
block on the trainer until the updated model is returned.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">batch_size</span><span class="p">,</span> <span class="n">image_w</span><span class="p">,</span> <span class="n">image_h</span> <span class="o">=</span> <span class="mi">20</span><span class="p">,</span> <span class="mi">64</span><span class="p">,</span> <span class="mi">64</span>
<span class="k">class</span> <span class="nc">Trainer</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">ps_rref</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">ps_rref</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">loss_fn</span> <span class="o">=</span> <span class="n">ps_rref</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">MSELoss</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">one_hot_indices</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">LongTensor</span><span class="p">(</span><span class="n">batch_size</span><span class="p">)</span> \
<span class="o">.</span><span class="n">random_</span><span class="p">(</span><span class="mi">0</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">)</span> \
<span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="n">batch_size</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">get_next_batch</span><span class="p">(</span><span class="bp">self</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="mi">6</span><span class="p">):</span>
<span class="n">inputs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="n">batch_size</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="n">image_w</span><span class="p">,</span> <span class="n">image_h</span><span class="p">)</span>
<span class="n">labels</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">batch_size</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">)</span> \
<span class="o">.</span><span class="n">scatter_</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">one_hot_indices</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">yield</span> <span class="n">inputs</span><span class="o">.</span><span class="n">cuda</span><span class="p">(),</span> <span class="n">labels</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">train</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="n">name</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">get_worker_info</span><span class="p">()</span><span class="o">.</span><span class="n">name</span>
<span class="c1"># get initial model parameters</span>
<span class="n">m</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">ps_rref</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">()</span><span class="o">.</span><span class="n">get_model</span><span class="p">()</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="c1"># start training</span>
<span class="k">for</span> <span class="n">inputs</span><span class="p">,</span> <span class="n">labels</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_next_batch</span><span class="p">():</span>
<span class="bp">self</span><span class="o">.</span><span class="n">loss_fn</span><span class="p">(</span><span class="n">m</span><span class="p">(</span><span class="n">inputs</span><span class="p">),</span> <span class="n">labels</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span>
<span class="bp">self</span><span class="o">.</span><span class="n">ps_rref</span><span class="o">.</span><span class="n">owner</span><span class="p">(),</span>
<span class="n">BatchUpdateParameterServer</span><span class="o">.</span><span class="n">update_and_fetch_model</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ps_rref</span><span class="p">,</span> <span class="p">[</span><span class="n">p</span><span class="o">.</span><span class="n">grad</span> <span class="k">for</span> <span class="n">p</span> <span class="ow">in</span> <span class="n">m</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">parameters</span><span class="p">()]),</span>
<span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
</pre></div>
</div>
<p>We skip the code that launches multiple processes in this tutorial and please
refer to the <a class="reference external" href="https://github.com/pytorch/examples/tree/master/distributed/rpc">examples</a>
repo for the full implementation. Note that, it is possible to implement batch
processing without the
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
decorator. However, that would require either blocking more RPC threads on
the PS or use another round of RPC to fetch updated models, where the latter
would add both more code complexity and more communication overhead.</p>
<p>This section uses a simple parameter sever training example to show how to
implement batch RPC applications using the
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
decorator. In the next section, we re-implement the reinforcement learning
example in the previous
<a class="reference external" href="https://tutorials.pytorch.kr/intermediate/rpc_tutorial.html">Getting started with Distributed RPC Framework</a>
tutorial using batch processing, and demonstrate its impact on the training
speed.</p>
</div>
<div class="section" id="batch-processing-cartpole-solver">
<h2>Batch-Processing CartPole Solver<a class="headerlink" href="#batch-processing-cartpole-solver" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<p>This section uses CartPole-v1 from <a class="reference external" href="https://gym.openai.com/">OpenAI Gym</a> as
an example to show the performance impact of batch processing RPC. Please note
that since the goal is to demonstrate the usage of
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>
instead of building the best CartPole solver or solving most different RL
problems, we use very simple policies and reward calculation strategies and
focus on the multi-observer single-agent batch RPC implementation. We use a
similar <code class="docutils literal notranslate"><span class="pre">Policy</span></code> model as the previous tutorial which is shown below. Compared
to the previous tutorial, the difference is that its constructor takes an
additional <code class="docutils literal notranslate"><span class="pre">batch</span></code> argument which controls the <code class="docutils literal notranslate"><span class="pre">dim</span></code> parameter for
<code class="docutils literal notranslate"><span class="pre">F.softmax</span></code> because with batching, the <code class="docutils literal notranslate"><span class="pre">x</span></code> argument in the <code class="docutils literal notranslate"><span class="pre">forward</span></code>
function contains states from multiple observers and hence the dimension needs
to change properly. Everything else stays intact.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">argparse</span>
<span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="n">parser</span> <span class="o">=</span> <span class="n">argparse</span><span class="o">.</span><span class="n">ArgumentParser</span><span class="p">(</span><span class="n">description</span><span class="o">=</span><span class="s1">'PyTorch RPC Batch RL example'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--gamma'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">float</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mf">1.0</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'G'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'discount factor (default: 1.0)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--seed'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">543</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'N'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'random seed (default: 543)'</span><span class="p">)</span>
<span class="n">parser</span><span class="o">.</span><span class="n">add_argument</span><span class="p">(</span><span class="s1">'--num-episode'</span><span class="p">,</span> <span class="nb">type</span><span class="o">=</span><span class="nb">int</span><span class="p">,</span> <span class="n">default</span><span class="o">=</span><span class="mi">10</span><span class="p">,</span> <span class="n">metavar</span><span class="o">=</span><span class="s1">'E'</span><span class="p">,</span>
<span class="n">help</span><span class="o">=</span><span class="s1">'number of episodes (default: 10)'</span><span class="p">)</span>
<span class="n">args</span> <span class="o">=</span> <span class="n">parser</span><span class="o">.</span><span class="n">parse_args</span><span class="p">()</span>
<span class="n">torch</span><span class="o">.</span><span class="n">manual_seed</span><span class="p">(</span><span class="n">args</span><span class="o">.</span><span class="n">seed</span><span class="p">)</span>
<span class="k">class</span> <span class="nc">Policy</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">batch</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">Policy</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">affine1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">128</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dropout</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Dropout</span><span class="p">(</span><span class="n">p</span><span class="o">=</span><span class="mf">0.6</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">affine2</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">dim</span> <span class="o">=</span> <span class="mi">2</span> <span class="k">if</span> <span class="n">batch</span> <span class="k">else</span> <span class="mi">1</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">affine1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dropout</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">action_scores</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">affine2</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">F</span><span class="o">.</span><span class="n">softmax</span><span class="p">(</span><span class="n">action_scores</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="bp">self</span><span class="o">.</span><span class="n">dim</span><span class="p">)</span>
</pre></div>
</div>
<p>The constructor of the <code class="docutils literal notranslate"><span class="pre">Observer</span></code> adjusts accordingly as well. It also takes a
<code class="docutils literal notranslate"><span class="pre">batch</span></code> argument, which governs which <code class="docutils literal notranslate"><span class="pre">Agent</span></code> function it uses to select
actions. In batch mode, it calls <code class="docutils literal notranslate"><span class="pre">select_action_batch</span></code> function on <code class="docutils literal notranslate"><span class="pre">Agent</span></code>
which will be presented shortly, and this function will be decorated with
<a class="reference external" href="https://pytorch.org/docs/master/rpc.html#torch.distributed.rpc.functions.async_execution">@rpc.functions.async_execution</a>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">gym</span>
<span class="kn">import</span> <span class="nn">torch.distributed.rpc</span> <span class="k">as</span> <span class="nn">rpc</span>
<span class="k">class</span> <span class="nc">Observer</span><span class="p">:</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">batch</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">id</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">get_worker_info</span><span class="p">()</span><span class="o">.</span><span class="n">id</span> <span class="o">-</span> <span class="mi">1</span>
<span class="bp">self</span><span class="o">.</span><span class="n">env</span> <span class="o">=</span> <span class="n">gym</span><span class="o">.</span><span class="n">make</span><span class="p">(</span><span class="s1">'CartPole-v1'</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">env</span><span class="o">.</span><span class="n">seed</span><span class="p">(</span><span class="n">args</span><span class="o">.</span><span class="n">seed</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">select_action</span> <span class="o">=</span> <span class="n">Agent</span><span class="o">.</span><span class="n">select_action_batch</span> <span class="k">if</span> <span class="n">batch</span> <span class="k">else</span> <span class="n">Agent</span><span class="o">.</span><span class="n">select_action</span>
</pre></div>
</div>
<p>Compared to the previous tutorial
<a class="reference external" href="https://tutorials.pytorch.kr/intermediate/rpc_tutorial.html">Getting started with Distributed RPC Framework</a>,
observers behave a little differently. Instead of exiting when the environment
is stopped, it always runs <code class="docutils literal notranslate"><span class="pre">n_steps</span></code> iterations in every episode. When the
environment returns, the observer simply resets the environment and start over
again. With this design, the agent will receive a fixed number of states from
every observer and hence can pack them into a fixed-size tensor. In every
step, the <code class="docutils literal notranslate"><span class="pre">Observer</span></code> uses RPC to send its state to the <code class="docutils literal notranslate"><span class="pre">Agent</span></code> and fetches
the action through the return value. At the end of every episode, it returns the
rewards of all steps to <code class="docutils literal notranslate"><span class="pre">Agent</span></code>. Note that this <code class="docutils literal notranslate"><span class="pre">run_episode</span></code> function will
be called by the <code class="docutils literal notranslate"><span class="pre">Agent</span></code> using RPC. So the <code class="docutils literal notranslate"><span class="pre">rpc_sync</span></code> call in this function
will be a nested RPC invocation. We could mark this function as <code class="docutils literal notranslate"><span class="pre">@rpc.functions.async_execution</span></code>
too to avoid blocking one thread on the <code class="docutils literal notranslate"><span class="pre">Observer</span></code>. However, as the bottleneck
is the <code class="docutils literal notranslate"><span class="pre">Agent</span></code> instead of the <code class="docutils literal notranslate"><span class="pre">Observer</span></code>, it should be OK to block one
thread on the <code class="docutils literal notranslate"><span class="pre">Observer</span></code> process.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="k">class</span> <span class="nc">Observer</span><span class="p">:</span>
<span class="o">...</span>
<span class="k">def</span> <span class="nf">run_episode</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">agent_rref</span><span class="p">,</span> <span class="n">n_steps</span><span class="p">):</span>
<span class="n">state</span><span class="p">,</span> <span class="n">ep_reward</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">(),</span> <span class="n">NUM_STEPS</span>
<span class="n">rewards</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">n_steps</span><span class="p">)</span>
<span class="n">start_step</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">step</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">n_steps</span><span class="p">):</span>
<span class="n">state</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">state</span><span class="p">)</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
<span class="c1"># send the state to the agent to get an action</span>
<span class="n">action</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">rpc_sync</span><span class="p">(</span>
<span class="n">agent_rref</span><span class="o">.</span><span class="n">owner</span><span class="p">(),</span>
<span class="bp">self</span><span class="o">.</span><span class="n">select_action</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">agent_rref</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">id</span><span class="p">,</span> <span class="n">state</span><span class="p">)</span>
<span class="p">)</span>
<span class="c1"># apply the action to the environment, and get the reward</span>
<span class="n">state</span><span class="p">,</span> <span class="n">reward</span><span class="p">,</span> <span class="n">done</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">env</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">action</span><span class="p">)</span>
<span class="n">rewards</span><span class="p">[</span><span class="n">step</span><span class="p">]</span> <span class="o">=</span> <span class="n">reward</span>
<span class="k">if</span> <span class="n">done</span> <span class="ow">or</span> <span class="n">step</span> <span class="o">+</span> <span class="mi">1</span> <span class="o">>=</span> <span class="n">n_steps</span><span class="p">:</span>
<span class="n">curr_rewards</span> <span class="o">=</span> <span class="n">rewards</span><span class="p">[</span><span class="n">start_step</span><span class="p">:(</span><span class="n">step</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)]</span>
<span class="n">R</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">curr_rewards</span><span class="o">.</span><span class="n">numel</span><span class="p">()</span> <span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="o">-</span><span class="mi">1</span><span class="p">):</span>
<span class="n">R</span> <span class="o">=</span> <span class="n">curr_rewards</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">+</span> <span class="n">args</span><span class="o">.</span><span class="n">gamma</span> <span class="o">*</span> <span class="n">R</span>
<span class="n">curr_rewards</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="n">R</span>
<span class="n">state</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">env</span><span class="o">.</span><span class="n">reset</span><span class="p">()</span>
<span class="k">if</span> <span class="n">start_step</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="n">ep_reward</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">ep_reward</span><span class="p">,</span> <span class="n">step</span> <span class="o">-</span> <span class="n">start_step</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span>
<span class="n">start_step</span> <span class="o">=</span> <span class="n">step</span> <span class="o">+</span> <span class="mi">1</span>
<span class="k">return</span> <span class="p">[</span><span class="n">rewards</span><span class="p">,</span> <span class="n">ep_reward</span><span class="p">]</span>
</pre></div>
</div>
<p>The constructor of the <code class="docutils literal notranslate"><span class="pre">Agent</span></code> also takes a <code class="docutils literal notranslate"><span class="pre">batch</span></code> argument, which controls
how action probs are batched. In batch mode, the <code class="docutils literal notranslate"><span class="pre">saved_log_probs</span></code> contains a
list of tensors, where each tensor contains action robs from all observers in
one step. Without batching, the <code class="docutils literal notranslate"><span class="pre">saved_log_probs</span></code> is a dictionary where the
key is the observer id and the value is a list of action probs for that
observer.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">threading</span>
<span class="kn">from</span> <span class="nn">torch.distributed.rpc</span> <span class="kn">import</span> <span class="n">RRef</span>
<span class="k">class</span> <span class="nc">Agent</span><span class="p">:</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">batch</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">agent_rref</span> <span class="o">=</span> <span class="n">RRef</span><span class="p">(</span><span class="bp">self</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">rewards</span> <span class="o">=</span> <span class="p">{}</span>
<span class="bp">self</span><span class="o">.</span><span class="n">policy</span> <span class="o">=</span> <span class="n">Policy</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</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="bp">self</span><span class="o">.</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">1e-2</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">running_reward</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">ob_rank</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">world_size</span><span class="p">):</span>
<span class="n">ob_info</span> <span class="o">=</span> <span class="n">rpc</span><span class="o">.</span><span class="n">get_worker_info</span><span class="p">(</span><span class="n">OBSERVER_NAME</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">ob_rank</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">rpc</span><span class="o">.</span><span class="n">remote</span><span class="p">(</span><span class="n">ob_info</span><span class="p">,</span> <span class="n">Observer</span><span class="p">,</span> <span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">batch</span><span class="p">,)))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">rewards</span><span class="p">[</span><span class="n">ob_info</span><span class="o">.</span><span class="n">id</span><span class="p">]</span> <span class="o">=</span> <span class="p">[]</span>
<span class="bp">self</span><span class="o">.</span><span class="n">states</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">),</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">batch</span> <span class="o">=</span> <span class="n">batch</span>
<span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span> <span class="o">=</span> <span class="p">[]</span> <span class="k">if</span> <span class="n">batch</span> <span class="k">else</span> <span class="p">{</span><span class="n">k</span><span class="p">:[]</span> <span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">))}</span>
<span class="bp">self</span><span class="o">.</span><span class="n">future_actions</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">futures</span><span class="o">.</span><span class="n">Future</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">lock</span> <span class="o">=</span> <span class="n">threading</span><span class="o">.</span><span class="n">Lock</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">pending_states</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">)</span>
</pre></div>
</div>
<p>The non-batching <code class="docutils literal notranslate"><span class="pre">select_acion</span></code> simply runs the state throw the policy, saves
the action prob, and returns the action to the observer right away.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">torch.distributions</span> <span class="kn">import</span> <span class="n">Categorical</span>
<span class="k">class</span> <span class="nc">Agent</span><span class="p">:</span>
<span class="o">...</span>
<span class="nd">@staticmethod</span>
<span class="k">def</span> <span class="nf">select_action</span><span class="p">(</span><span class="n">agent_rref</span><span class="p">,</span> <span class="n">ob_id</span><span class="p">,</span> <span class="n">state</span><span class="p">):</span>
<span class="bp">self</span> <span class="o">=</span> <span class="n">agent_rref</span><span class="o">.</span><span class="n">local_value</span><span class="p">()</span>
<span class="n">probs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">policy</span><span class="p">(</span><span class="n">state</span><span class="o">.</span><span class="n">cuda</span><span class="p">())</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">Categorical</span><span class="p">(</span><span class="n">probs</span><span class="p">)</span>
<span class="n">action</span> <span class="o">=</span> <span class="n">m</span><span class="o">.</span><span class="n">sample</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span><span class="p">[</span><span class="n">ob_id</span><span class="p">]</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">log_prob</span><span class="p">(</span><span class="n">action</span><span class="p">))</span>
<span class="k">return</span> <span class="n">action</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
</pre></div>
</div>
<p>With batching, the state is stored in a 2D tensor <code class="docutils literal notranslate"><span class="pre">self.states</span></code>, using the
observer id as the row id. Then, it chains a <code class="docutils literal notranslate"><span class="pre">Future</span></code> by installing a callback
function to the batch-generated <code class="docutils literal notranslate"><span class="pre">self.future_actions</span></code> <code class="docutils literal notranslate"><span class="pre">Future</span></code> object, which
will be populated with the specific row indexed using the id of that observer.
The last arriving observer runs all batched states through the policy in one
shot and set <code class="docutils literal notranslate"><span class="pre">self.future_actions</span></code> accordingly. When this occurs, all the
callback functions installed on <code class="docutils literal notranslate"><span class="pre">self.future_actions</span></code> will be triggered and
their return values will be used to populate the chained <code class="docutils literal notranslate"><span class="pre">Future</span></code> object,
which in turn notifies the <code class="docutils literal notranslate"><span class="pre">Agent</span></code> to prepare and communicate responses for
all previous RPC requests from other observers.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">Agent</span><span class="p">:</span>
<span class="o">...</span>
<span class="nd">@staticmethod</span>
<span class="nd">@rpc</span><span class="o">.</span><span class="n">functions</span><span class="o">.</span><span class="n">async_execution</span>
<span class="k">def</span> <span class="nf">select_action_batch</span><span class="p">(</span><span class="n">agent_rref</span><span class="p">,</span> <span class="n">ob_id</span><span class="p">,</span> <span class="n">state</span><span class="p">):</span>
<span class="bp">self</span> <span class="o">=</span> <span class="n">agent_rref</span><span class="o">.</span><span class="n">local_value</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">states</span><span class="p">[</span><span class="n">ob_id</span><span class="p">]</span><span class="o">.</span><span class="n">copy_</span><span class="p">(</span><span class="n">state</span><span class="p">)</span>
<span class="n">future_action</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">future_actions</span><span class="o">.</span><span class="n">then</span><span class="p">(</span>
<span class="k">lambda</span> <span class="n">future_actions</span><span class="p">:</span> <span class="n">future_actions</span><span class="o">.</span><span class="n">wait</span><span class="p">()[</span><span class="n">ob_id</span><span class="p">]</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
<span class="p">)</span>
<span class="k">with</span> <span class="bp">self</span><span class="o">.</span><span class="n">lock</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">pending_states</span> <span class="o">-=</span> <span class="mi">1</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">pending_states</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="bp">self</span><span class="o">.</span><span class="n">pending_states</span> <span class="o">=</span> <span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">)</span>
<span class="n">probs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">policy</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">states</span><span class="o">.</span><span class="n">cuda</span><span class="p">())</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">Categorical</span><span class="p">(</span><span class="n">probs</span><span class="p">)</span>
<span class="n">actions</span> <span class="o">=</span> <span class="n">m</span><span class="o">.</span><span class="n">sample</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">log_prob</span><span class="p">(</span><span class="n">actions</span><span class="p">)</span><span class="o">.</span><span class="n">t</span><span class="p">()[</span><span class="mi">0</span><span class="p">])</span>
<span class="n">future_actions</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">future_actions</span>
<span class="bp">self</span><span class="o">.</span><span class="n">future_actions</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">futures</span><span class="o">.</span><span class="n">Future</span><span class="p">()</span>
<span class="n">future_actions</span><span class="o">.</span><span class="n">set_result</span><span class="p">(</span><span class="n">actions</span><span class="o">.</span><span class="n">cpu</span><span class="p">())</span>
<span class="k">return</span> <span class="n">future_action</span>
</pre></div>
</div>
<p>Now letโs define how different RPC functions are stitched together. The <code class="docutils literal notranslate"><span class="pre">Agent</span></code>
controls the execution of every episode. It first uses <code class="docutils literal notranslate"><span class="pre">rpc_async</span></code> to kick off
the episode on all observers and block on the returned futures which will be
populated with observer rewards. Note that the code below uses the RRef helper
<code class="docutils literal notranslate"><span class="pre">ob_rref.rpc_async()</span></code> to launch the <code class="docutils literal notranslate"><span class="pre">run_episode</span></code> function on the owner
of the <code class="docutils literal notranslate"><span class="pre">ob_rref</span></code> RRef with the provided arguments.
It then converts the saved action probs and returned observer rewards into
expected data format, and launch the training step. Finally, it resets all
states and returns the reward of the current episode. This function is the entry
point to run one episode.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">Agent</span><span class="p">:</span>
<span class="o">...</span>
<span class="k">def</span> <span class="nf">run_episode</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">n_steps</span><span class="o">=</span><span class="mi">0</span><span class="p">):</span>
<span class="n">futs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">ob_rref</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">:</span>
<span class="c1"># make async RPC to kick off an episode on all observers</span>
<span class="n">futs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">ob_rref</span><span class="o">.</span><span class="n">rpc_async</span><span class="p">()</span><span class="o">.</span><span class="n">run_episode</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">agent_rref</span><span class="p">,</span> <span class="n">n_steps</span><span class="p">))</span>
<span class="c1"># wait until all obervers have finished this episode</span>
<span class="n">rets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">futures</span><span class="o">.</span><span class="n">wait_all</span><span class="p">(</span><span class="n">futs</span><span class="p">)</span>
<span class="n">rewards</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">([</span><span class="n">ret</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="k">for</span> <span class="n">ret</span> <span class="ow">in</span> <span class="n">rets</span><span class="p">])</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span><span class="o">.</span><span class="n">t</span><span class="p">()</span>
<span class="n">ep_rewards</span> <span class="o">=</span> <span class="nb">sum</span><span class="p">([</span><span class="n">ret</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="k">for</span> <span class="n">ret</span> <span class="ow">in</span> <span class="n">rets</span><span class="p">])</span> <span class="o">/</span> <span class="nb">len</span><span class="p">(</span><span class="n">rets</span><span class="p">)</span>
<span class="c1"># stack saved probs into one tensor</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">batch</span><span class="p">:</span>
<span class="n">probs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span><span class="p">)</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">probs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span><span class="p">[</span><span class="n">i</span><span class="p">])</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">rets</span><span class="p">))]</span>
<span class="n">probs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="n">probs</span><span class="p">)</span>
<span class="n">policy_loss</span> <span class="o">=</span> <span class="o">-</span><span class="n">probs</span> <span class="o">*</span> <span class="n">rewards</span> <span class="o">/</span> <span class="nb">len</span><span class="p">(</span><span class="n">rets</span><span class="p">)</span>
<span class="n">policy_loss</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="c1"># reset variables</span>
<span class="bp">self</span><span class="o">.</span><span class="n">saved_log_probs</span> <span class="o">=</span> <span class="p">[]</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">batch</span> <span class="k">else</span> <span class="p">{</span><span class="n">k</span><span class="p">:[]</span> <span class="k">for</span> <span class="n">k</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">))}</span>
<span class="bp">self</span><span class="o">.</span><span class="n">states</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">ob_rrefs</span><span class="p">),</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="c1"># calculate running rewards</span>
<span class="bp">self</span><span class="o">.</span><span class="n">running_reward</span> <span class="o">=</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">ep_rewards</span> <span class="o">+</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">running_reward</span>
<span class="k">return</span> <span class="n">ep_rewards</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">running_reward</span>
</pre></div>
</div>
<p>The rest of the code is normal processes launching and logging which are
similar to other RPC tutorials. In this tutorial, all observers passively
waiting for commands from the agent. Please refer to the
<a class="reference external" href="https://github.com/pytorch/examples/tree/master/distributed/rpc">examples</a>
repo for the full implementation.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">run_worker</span><span class="p">(</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="p">,</span> <span class="n">n_episode</span><span class="p">,</span> <span class="n">batch</span><span class="p">,</span> <span class="n">print_log</span><span class="o">=</span><span class="kc">True</span><span class="p">):</span>
<span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_ADDR'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'localhost'</span>
<span class="n">os</span><span class="o">.</span><span class="n">environ</span><span class="p">[</span><span class="s1">'MASTER_PORT'</span><span class="p">]</span> <span class="o">=</span> <span class="s1">'29500'</span>
<span class="k">if</span> <span class="n">rank</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="c1"># rank0 is the agent</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="n">AGENT_NAME</span><span class="p">,</span> <span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="n">world_size</span><span class="p">)</span>
<span class="n">agent</span> <span class="o">=</span> <span class="n">Agent</span><span class="p">(</span><span class="n">world_size</span><span class="p">,</span> <span class="n">batch</span><span class="p">)</span>
<span class="k">for</span> <span class="n">i_episode</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">n_episode</span><span class="p">):</span>
<span class="n">last_reward</span><span class="p">,</span> <span class="n">running_reward</span> <span class="o">=</span> <span class="n">agent</span><span class="o">.</span><span class="n">run_episode</span><span class="p">(</span><span class="n">n_steps</span><span class="o">=</span><span class="n">NUM_STEPS</span><span class="p">)</span>
<span class="k">if</span> <span class="n">print_log</span><span class="p">:</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'Episode </span><span class="si">{}</span><span class="se">\t</span><span class="s1">Last reward: </span><span class="si">{:.2f}</span><span class="se">\t</span><span class="s1">Average reward: </span><span class="si">{:.2f}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span>
<span class="n">i_episode</span><span class="p">,</span> <span class="n">last_reward</span><span class="p">,</span> <span class="n">running_reward</span><span class="p">))</span>
<span class="k">else</span><span class="p">:</span>
<span class="c1"># other ranks are the observer</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">init_rpc</span><span class="p">(</span><span class="n">OBSERVER_NAME</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">rank</span><span class="p">),</span> <span class="n">rank</span><span class="o">=</span><span class="n">rank</span><span class="p">,</span> <span class="n">world_size</span><span class="o">=</span><span class="n">world_size</span><span class="p">)</span>
<span class="c1"># observers passively waiting for instructions from agents</span>
<span class="n">rpc</span><span class="o">.</span><span class="n">shutdown</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">main</span><span class="p">():</span>
<span class="k">for</span> <span class="n">world_size</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">12</span><span class="p">):</span>
<span class="n">delays</span> <span class="o">=</span> <span class="p">[]</span>
<span class="k">for</span> <span class="n">batch</span> <span class="ow">in</span> <span class="p">[</span><span class="kc">True</span><span class="p">,</span> <span class="kc">False</span><span class="p">]:</span>
<span class="n">tik</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
<span class="n">mp</span><span class="o">.</span><span class="n">spawn</span><span class="p">(</span>
<span class="n">run_worker</span><span class="p">,</span>
<span class="n">args</span><span class="o">=</span><span class="p">(</span><span class="n">world_size</span><span class="p">,</span> <span class="n">args</span><span class="o">.</span><span class="n">num_episode</span><span class="p">,</span> <span class="n">batch</span><span class="p">),</span>
<span class="n">nprocs</span><span class="o">=</span><span class="n">world_size</span><span class="p">,</span>
<span class="n">join</span><span class="o">=</span><span class="kc">True</span>
<span class="p">)</span>
<span class="n">tok</span> <span class="o">=</span> <span class="n">time</span><span class="o">.</span><span class="n">time</span><span class="p">()</span>
<span class="n">delays</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">tok</span> <span class="o">-</span> <span class="n">tik</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="sa">f</span><span class="s2">"</span><span class="si">{</span><span class="n">world_size</span><span class="si">}</span><span class="s2">, </span><span class="si">{</span><span class="n">delays</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span><span class="si">}</span><span class="s2">, </span><span class="si">{</span><span class="n">delays</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="k">if</span> <span class="vm">__name__</span> <span class="o">==</span> <span class="s1">'__main__'</span><span class="p">:</span>
<span class="n">main</span><span class="p">()</span>
</pre></div>
</div>
<p>Batch RPC helps to consolidate the action inference into less CUDA operations,
and hence reduces the amortized overhead. The above <code class="docutils literal notranslate"><span class="pre">main</span></code> function runs the
same code on both batch and no-batch modes using different numbers of observers,
ranging from 1 to 10. The figure below plots the execution time of different
world sizes using default argument values. The results confirmed our expectation
that batch processing helped to speed up training.</p>
<div class="figure align-default">
<img alt="" src="../_images/batch.png" />
</div>
</div>
<div class="section" id="learn-more">
<h2>Learn More<a class="headerlink" href="#learn-more" title="์ด ์ ๋ชฉ์ ๋ํ ํผ๋จธ๋งํฌ">ยถ</a></h2>
<ul class="simple">
<li><p><a class="reference external" href="https://github.com/pytorch/examples/blob/master/distributed/rpc/batch/parameter_server.py">Batch-Updating Parameter Server Source Code</a></p></li>
<li><p><a class="reference external" href="https://github.com/pytorch/examples/blob/master/distributed/rpc/batch/reinforce.py">Batch-Processing CartPole Solver</a></p></li>
<li><p><a class="reference external" href="https://pytorch.org/docs/master/rpc.html#distributed-autograd-framework">Distributed Autograd</a></p></li>
<li><p><a class="reference external" href="dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism</a></p></li>
</ul>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="../advanced/rpc_ddp_tutorial.html" class="btn btn-neutral float-right" title="๋ถ์ฐ ๋ฐ์ดํฐ ๋ณ๋ ฌ(DDP)๊ณผ ๋ถ์ฐ RPC ํ๋ ์์ํฌ ๊ฒฐํฉ" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="dist_pipeline_parallel_tutorial.html" class="btn btn-neutral" title="Distributed Pipeline Parallelism Using RPC" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr class="community-hr hr-top" />
<div class="community-container">
<div class="community-prompt">๋ ๊ถ๊ธํ์๊ฑฐ๋ ๊ฐ์ ํ ๋ด์ฉ์ด ์์ผ์ ๊ฐ์? ์ปค๋ฎค๋ํฐ์ ์ฐธ์ฌํด๋ณด์ธ์!</div>
<div class="community-link"><a href="https://discuss.pytorch.kr/" aria-label="PyTorchKoreaCommunity">ํ๊ตญ์ด ์ปค๋ฎค๋ํฐ ๋ฐ๋ก๊ฐ๊ธฐ</a></div>
</div>
<hr class="community-hr hr-bottom"/>
<hr class="rating-hr hr-top" />
<div class="rating-container">
<div class="rating-prompt">์ด ํํ ๋ฆฌ์ผ์ด ์ด๋ ์
จ๋์? ํ๊ฐํด์ฃผ์๋ฉด ์ดํ ๊ฐ์ ์ ์ฐธ๊ณ ํ๊ฒ ์ต๋๋ค! :)</div>
<div class="stars-outer">
<i class="far fa-star" title="1 Star" data-behavior="tutorial-rating" data-count="1"></i>
<i class="far fa-star" title="2 Stars" data-behavior="tutorial-rating" data-count="2"></i>
<i class="far fa-star" title="3 Stars" data-behavior="tutorial-rating" data-count="3"></i>
<i class="far fa-star" title="4 Stars" data-behavior="tutorial-rating" data-count="4"></i>
<i class="far fa-star" title="5 Stars" data-behavior="tutorial-rating" data-count="5"></i>
</div>
</div>
<hr class="rating-hr hr-bottom"/>
<div role="contentinfo">
<p>
© Copyright 2018-2024, PyTorch & ํ์ดํ ์น ํ๊ตญ ์ฌ์ฉ์ ๋ชจ์(PyTorch Korea User Group).
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
if((window.location.href.indexOf("/prototype/")!= -1) && (window.location.href.indexOf("/prototype/prototype_index")< 1))
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-flask" aria-hidden="true"> </i> ์ด ํํ ๋ฆฌ์ผ์ ํ๋กํ ํ์
(prototype) ๊ธฐ๋ฅ๋ค์ ๋ํด์ ์ค๋ช
ํ๊ณ ์์ต๋๋ค. ํ๋กํ ํ์
๊ธฐ๋ฅ์ ์ผ๋ฐ์ ์ผ๋ก ํผ๋๋ฐฑ ๋ฐ ํ
์คํธ์ฉ์ผ๋ก, ๋ฐํ์ ํ๋๊ทธ ์์ด๋ PyPI๋ Conda๋ก ๋ฐฐํฌ๋๋ ๋ฐ์ด๋๋ฆฌ์์๋ ์ฌ์ฉํ ์ ์์ต๋๋ค.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">