-
Notifications
You must be signed in to change notification settings - Fork 257
/
Copy pathtorch_script_custom_ops.html
1657 lines (1452 loc) · 142 KB
/
torch_script_custom_ops.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Extending TorchScript with Custom C++ Operators — PyTorch Tutorials 1.10.2+cu102 documentation</title>
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="../_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/gallery.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.10.0-beta/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.13.11/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="커스텀 C++ 클래스로 TorchScript 확장하기" href="torch_script_custom_classes.html" />
<link rel="prev" title="Custom C++ and CUDA Extensions" href="cpp_extension.html" />
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/katex@0.10.0/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.kr/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.kr/get-started">시작하기</a>
</li>
<li class="active">
<a href="https://tutorials.pytorch.kr">튜토리얼</a>
</li>
<li>
<a href="https://pytorch.kr/hub">허브</a>
</li>
<li>
<a href="https://discuss.pytorch.kr">커뮤니티</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
1.10.2+cu102
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Tutorials" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 레시피</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../recipes/recipes_index.html">모든 레시피 보기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../prototype/prototype_index.html">모든 프로토타입 레시피 보기</a></li>
</ul>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 시작하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/intro.html">파이토치(PyTorch) 기본 익히기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/quickstart_tutorial.html">빠른 시작(Quickstart)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/tensorqs_tutorial.html">텐서(Tensor)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/data_tutorial.html">Dataset과 DataLoader</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/transforms_tutorial.html">변형(Transform)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/buildmodel_tutorial.html">신경망 모델 구성하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/autogradqs_tutorial.html"><code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code>를 사용한 자동 미분</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/optimization_tutorial.html">모델 매개변수 최적화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/basics/saveloadrun_tutorial.html">모델 저장하고 불러오기</a></li>
</ul>
<p class="caption"><span class="caption-text">Introduction to PyTorch on YouTube</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt.html">Introduction to PyTorch - YouTube Series</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/introyt1_tutorial.html">Introduction to PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensors_deeper_tutorial.html">Introduction to PyTorch Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/autogradyt_tutorial.html">The Fundamentals of Autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/modelsyt_tutorial.html">Building Models with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/tensorboardyt_tutorial.html">PyTorch TensorBoard Support</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/trainingyt.html">Training with PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/introyt/captumyt.html">Model Understanding with Captum</a></li>
</ul>
<p class="caption"><span class="caption-text">파이토치(PyTorch) 배우기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deep_learning_60min_blitz.html">PyTorch로 딥러닝하기: 60분만에 끝장내기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/pytorch_with_examples.html">예제로 배우는 파이토치(PyTorch)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/nn_tutorial.html"><cite>torch.nn</cite> 이 <em>실제로</em> 무엇인가요?</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/tensorboard_tutorial.html">TensorBoard로 모델, 데이터, 학습 시각화하기</a></li>
</ul>
<p class="caption"><span class="caption-text">이미지/비디오</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/torchvision_tutorial.html">TorchVision 객체 검출 미세조정(Finetuning) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transfer_learning_tutorial.html">컴퓨터 비전(Vision)을 위한 전이학습(Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/fgsm_tutorial.html">적대적 예제 생성(Adversarial Example Generation)</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dcgan_faces_tutorial.html">DCGAN 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">배포를 위한 비전 트랜스포머(Vision Transformer) 모델 최적화하기</a></li>
</ul>
<p class="caption"><span class="caption-text">오디오</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_io_tutorial.html">Audio I/O</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_resampling_tutorial.html">Audio Resampling</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_data_augmentation_tutorial.html">Audio Data Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_extractions_tutorial.html">Audio Feature Extractions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_feature_augmentation_tutorial.html">Audio Feature Augmentation</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/audio_datasets_tutorial.html">Audio Datasets</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/speech_recognition_pipeline_tutorial.html">Speech Recognition with Wav2Vec2</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/speech_command_classification_with_torchaudio_tutorial.html">Speech Command Classification with torchaudio</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/text_to_speech_with_torchaudio.html">Text-to-speech with torchaudio</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/forced_alignment_with_torchaudio_tutorial.html">Forced Alignment with Wav2Vec2</a></li>
</ul>
<p class="caption"><span class="caption-text">텍스트</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/transformer_tutorial.html">nn.Transformer 와 TorchText 로 시퀀스-투-시퀀스(Sequence-to-Sequence) 모델링하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/char_rnn_classification_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 분류하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/char_rnn_generation_tutorial.html">기초부터 시작하는 NLP: 문자-단위 RNN으로 이름 생성하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/seq2seq_translation_tutorial.html">기초부터 시작하는 NLP: Sequence to Sequence 네트워크와 Attention을 이용한 번역</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/text_sentiment_ngrams_tutorial.html">torchtext 라이브러리로 텍스트 분류하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/translation_transformer.html">nn.Transformer와 torchtext로 언어 번역하기</a></li>
</ul>
<p class="caption"><span class="caption-text">강화학습</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/reinforcement_q_learning.html">강화 학습 (DQN) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/mario_rl_tutorial.html">Train a Mario-playing RL Agent</a></li>
</ul>
<p class="caption"><span class="caption-text">PyTorch 모델을 프로덕션 환경에 배포하기</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/flask_rest_api_tutorial.html">Flask를 사용하여 Python에서 PyTorch를 REST API로 배포하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/Intro_to_TorchScript_tutorial.html">TorchScript 소개</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_export.html">C++에서 TorchScript 모델 로딩하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="super_resolution_with_onnxruntime.html">(선택) PyTorch 모델을 ONNX으로 변환하고 ONNX 런타임에서 실행하기</a></li>
</ul>
<p class="caption"><span class="caption-text">Code Transforms with FX</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/fx_conv_bn_fuser.html">(베타) FX에서 합성곱/배치 정규화(Convolution/Batch Norm) 결합기(Fuser) 만들기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/fx_profiling_tutorial.html">(beta) Building a Simple CPU Performance Profiler with FX</a></li>
</ul>
<p class="caption"><span class="caption-text">프론트엔드 API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/memory_format_tutorial.html">(베타) PyTorch를 사용한 Channels Last 메모리 형식</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/forward_ad_usage.html">Forward-mode Automatic Differentiation (Beta)</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_frontend.html">PyTorch C++ 프론트엔드 사용하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch-script-parallelism.html">TorchScript의 동적 병렬 처리(Dynamic Parallelism)</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_autograd.html">C++ 프론트엔드의 자동 미분 (autograd)</a></li>
</ul>
<p class="caption"><span class="caption-text">PyTorch 확장하기</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../intermediate/custom_function_double_backward_tutorial.html">Double Backward with Custom Functions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/custom_function_conv_bn_tutorial.html">Fusing Convolution and Batch Norm using Custom Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">Custom C++ and CUDA Extensions</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Extending TorchScript with Custom C++ Operators</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch_script_custom_classes.html">커스텀 C++ 클래스로 TorchScript 확장하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="dispatcher.html">Registering a Dispatched Operator in C++</a></li>
<li class="toctree-l1"><a class="reference internal" href="extend_dispatcher.html">Extending dispatcher for a new backend in C++</a></li>
</ul>
<p class="caption"><span class="caption-text">모델 최적화</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/profiler.html">PyTorch 모듈 프로파일링 하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/tensorboard_profiler_tutorial.html">PyTorch Profiler With TensorBoard</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/hyperparameter_tuning_tutorial.html">Hyperparameter tuning with Ray Tune</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/vt_tutorial.html">배포를 위한 비전 트랜스포머(Vision Transformer) 모델 최적화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/parametrizations.html">Parametrizations Tutorial</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/pruning_tutorial.html">가지치기 기법(Pruning) 튜토리얼</a></li>
<li class="toctree-l1"><a class="reference internal" href="dynamic_quantization_tutorial.html">(베타) LSTM 기반 단어 단위 언어 모델의 동적 양자화</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dynamic_quantization_bert_tutorial.html">(베타) BERT 모델 동적 양자화하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/quantized_transfer_learning_tutorial.html">(베타) 컴퓨터 비전 튜토리얼을 위한 양자화된 전이학습(Quantized Transfer Learning)</a></li>
<li class="toctree-l1"><a class="reference internal" href="static_quantization_tutorial.html">(beta) Static Quantization with Eager Mode in PyTorch</a></li>
</ul>
<p class="caption"><span class="caption-text">병렬 및 분산 학습</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/dist_overview.html">PyTorch Distributed Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/model_parallel_tutorial.html">단일 머신을 사용한 모델 병렬화 모범 사례</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/ddp_tutorial.html">분산 데이터 병렬 처리 시작하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dist_tuto.html">PyTorch로 분산 어플리케이션 개발하기</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/rpc_tutorial.html">Getting Started with Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/rpc_param_server_tutorial.html">Implementing a Parameter Server Using Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/dist_pipeline_parallel_tutorial.html">Distributed Pipeline Parallelism Using RPC</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/rpc_async_execution.html">Implementing Batch RPC Processing Using Asynchronous Executions</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc_ddp_tutorial.html">분산 데이터 병렬(DDP)과 분산 RPC 프레임워크 결합</a></li>
<li class="toctree-l1"><a class="reference internal" href="../intermediate/pipeline_tutorial.html">파이프라인 병렬화로 트랜스포머 모델 학습시키기</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_pipeline.html">분산 데이터 병렬 처리와 병렬 처리 파이프라인을 사용한 트랜스포머 모델 학습</a></li>
<li class="toctree-l1"><a class="reference internal" href="generic_join.html">Distributed Training with Uneven Inputs Using the Join Context Manager</a></li>
</ul>
<p class="caption"><span class="caption-text">Mobile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deeplabv3_on_ios.html">iOS에서의 이미지 분할 DeepLapV3</a></li>
<li class="toctree-l1"><a class="reference internal" href="../beginner/deeplabv3_on_android.html">안드로이드에서의 이미지 분할 DeepLapV3</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="../index.html">
Tutorials
</a> >
</li>
<li>Extending TorchScript with Custom C++ Operators</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/advanced/torch_script_custom_ops.rst.txt" rel="nofollow"><img src="../_static/images/view-page-source-icon.svg"></a>
</li>
</ul>
</div>
</div>
<div class="pytorch-shortcuts-wrapper" id="pytorch-shortcuts-wrapper">
Shortcuts
</div>
</div>
<section data-toggle="wy-nav-shift" id="pytorch-content-wrap" class="pytorch-content-wrap">
<div class="pytorch-content-left">
<div class="pytorch-call-to-action-links">
<div id="tutorial-type">advanced/torch_script_custom_ops</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="extending-torchscript-with-custom-c-operators">
<h1>Extending TorchScript with Custom C++ Operators<a class="headerlink" href="#extending-torchscript-with-custom-c-operators" title="Permalink to this headline">¶</a></h1>
<p>The PyTorch 1.0 release introduced a new programming model to PyTorch called
<a class="reference external" href="https://pytorch.org/docs/master/jit.html">TorchScript</a>. TorchScript is a
subset of the Python programming language which can be parsed, compiled and
optimized by the TorchScript compiler. Further, compiled TorchScript models have
the option of being serialized into an on-disk file format, which you can
subsequently load and run from pure C++ (as well as Python) for inference.</p>
<p>TorchScript supports a large subset of operations provided by the <code class="docutils literal notranslate"><span class="pre">torch</span></code>
package, allowing you to express many kinds of complex models purely as a series
of tensor operations from PyTorch’s “standard library”. Nevertheless, there may
be times where you find yourself in need of extending TorchScript with a custom
C++ or CUDA function. While we recommend that you only resort to this option if
your idea cannot be expressed (efficiently enough) as a simple Python function,
we do provide a very friendly and simple interface for defining custom C++ and
CUDA kernels using <a class="reference external" href="https://pytorch.org/cppdocs/#aten">ATen</a>, PyTorch’s high
performance C++ tensor library. Once bound into TorchScript, you can embed these
custom kernels (or “ops”) into your TorchScript model and execute them both in
Python and in their serialized form directly in C++.</p>
<p>The following paragraphs give an example of writing a TorchScript custom op to
call into <a class="reference external" href="https://www.opencv.org">OpenCV</a>, a computer vision library written
in C++. We will discuss how to work with tensors in C++, how to efficiently
convert them to third party tensor formats (in this case, OpenCV <code class="docutils literal notranslate"><span class="pre">Mat</span></code>), how
to register your operator with the TorchScript runtime and finally how to
compile the operator and use it in Python and C++.</p>
<div class="section" id="implementing-the-custom-operator-in-c">
<h2>Implementing the Custom Operator in C++<a class="headerlink" href="#implementing-the-custom-operator-in-c" title="Permalink to this headline">¶</a></h2>
<p>For this tutorial, we’ll be exposing the <a class="reference external" href="https://docs.opencv.org/2.4/modules/imgproc/doc/geometric_transformations.html#warpperspective">warpPerspective</a>
function, which applies a perspective transformation to an image, from OpenCV to
TorchScript as a custom operator. The first step is to write the implementation
of our custom operator in C++. Let’s call the file for this implementation
<code class="docutils literal notranslate"><span class="pre">op.cpp</span></code> and make it look like this:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">warp_perspective</span><span class="p">(</span><span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">image</span><span class="p">,</span> <span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">warp</span><span class="p">)</span> <span class="p">{</span>
<span class="c1">// BEGIN image_mat</span>
<span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">image_mat</span><span class="p">(</span><span class="cm">/*rows=*/</span><span class="n">image</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
<span class="cm">/*cols=*/</span><span class="n">image</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span>
<span class="cm">/*type=*/</span><span class="n">CV_32FC1</span><span class="p">,</span>
<span class="cm">/*data=*/</span><span class="n">image</span><span class="p">.</span><span class="n">data_ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">());</span>
<span class="c1">// END image_mat</span>
<span class="c1">// BEGIN warp_mat</span>
<span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">warp_mat</span><span class="p">(</span><span class="cm">/*rows=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
<span class="cm">/*cols=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span>
<span class="cm">/*type=*/</span><span class="n">CV_32FC1</span><span class="p">,</span>
<span class="cm">/*data=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">data_ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">());</span>
<span class="c1">// END warp_mat</span>
<span class="c1">// BEGIN output_mat</span>
<span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">output_mat</span><span class="p">;</span>
<span class="n">cv</span><span class="o">::</span><span class="n">warpPerspective</span><span class="p">(</span><span class="n">image_mat</span><span class="p">,</span> <span class="n">output_mat</span><span class="p">,</span> <span class="n">warp_mat</span><span class="p">,</span> <span class="cm">/*dsize=*/</span><span class="p">{</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">});</span>
<span class="c1">// END output_mat</span>
<span class="c1">// BEGIN output_tensor</span>
<span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">output</span> <span class="o">=</span> <span class="n">torch</span><span class="o">::</span><span class="n">from_blob</span><span class="p">(</span><span class="n">output_mat</span><span class="p">.</span><span class="n">ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">(),</span> <span class="cm">/*sizes=*/</span><span class="p">{</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">});</span>
<span class="k">return</span> <span class="n">output</span><span class="p">.</span><span class="n">clone</span><span class="p">();</span>
<span class="c1">// END output_tensor</span>
<span class="p">}</span>
</pre></div>
</div>
<p>The code for this operator is quite short. At the top of the file, we include
the OpenCV header file, <code class="docutils literal notranslate"><span class="pre">opencv2/opencv.hpp</span></code>, alongside the <code class="docutils literal notranslate"><span class="pre">torch/script.h</span></code>
header which exposes all the necessary goodies from PyTorch’s C++ API that we
need to write custom TorchScript operators. Our function <code class="docutils literal notranslate"><span class="pre">warp_perspective</span></code>
takes two arguments: an input <code class="docutils literal notranslate"><span class="pre">image</span></code> and the <code class="docutils literal notranslate"><span class="pre">warp</span></code> transformation matrix
we wish to apply to the image. The type of these inputs is <code class="docutils literal notranslate"><span class="pre">torch::Tensor</span></code>,
PyTorch’s tensor type in C++ (which is also the underlying type of all tensors
in Python). The return type of our <code class="docutils literal notranslate"><span class="pre">warp_perspective</span></code> function will also be a
<code class="docutils literal notranslate"><span class="pre">torch::Tensor</span></code>.</p>
<div class="admonition tip">
<p class="first admonition-title">Tip</p>
<p class="last">See <a class="reference external" href="https://pytorch.org/cppdocs/notes/tensor_basics.html">this note</a> for
more information about ATen, the library that provides the <code class="docutils literal notranslate"><span class="pre">Tensor</span></code> class to
PyTorch. Further, <a class="reference external" href="https://pytorch.org/cppdocs/notes/tensor_creation.html">this tutorial</a> describes how to
allocate and initialize new tensor objects in C++ (not required for this
operator).</p>
</div>
<div class="admonition attention">
<p class="first admonition-title">Attention</p>
<p class="last">The TorchScript compiler understands a fixed number of types. Only these types
can be used as arguments to your custom operator. Currently these types are:
<code class="docutils literal notranslate"><span class="pre">torch::Tensor</span></code>, <code class="docutils literal notranslate"><span class="pre">torch::Scalar</span></code>, <code class="docutils literal notranslate"><span class="pre">double</span></code>, <code class="docutils literal notranslate"><span class="pre">int64_t</span></code> and
<code class="docutils literal notranslate"><span class="pre">std::vector</span></code> s of these types. Note that <em>only</em> <code class="docutils literal notranslate"><span class="pre">double</span></code> and <em>not</em>
<code class="docutils literal notranslate"><span class="pre">float</span></code>, and <em>only</em> <code class="docutils literal notranslate"><span class="pre">int64_t</span></code> and <em>not</em> other integral types such as
<code class="docutils literal notranslate"><span class="pre">int</span></code>, <code class="docutils literal notranslate"><span class="pre">short</span></code> or <code class="docutils literal notranslate"><span class="pre">long</span></code> are supported.</p>
</div>
<p>Inside of our function, the first thing we need to do is convert our PyTorch
tensors to OpenCV matrices, as OpenCV’s <code class="docutils literal notranslate"><span class="pre">warpPerspective</span></code> expects <code class="docutils literal notranslate"><span class="pre">cv::Mat</span></code>
objects as inputs. Fortunately, there is a way to do this <strong>without copying
any</strong> data. In the first few lines,</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span> <span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">image_mat</span><span class="p">(</span><span class="cm">/*rows=*/</span><span class="n">image</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
<span class="cm">/*cols=*/</span><span class="n">image</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span>
<span class="cm">/*type=*/</span><span class="n">CV_32FC1</span><span class="p">,</span>
<span class="cm">/*data=*/</span><span class="n">image</span><span class="p">.</span><span class="n">data_ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">());</span>
</pre></div>
</div>
<p>we are calling <a class="reference external" href="https://docs.opencv.org/trunk/d3/d63/classcv_1_1Mat.html#a922de793eabcec705b3579c5f95a643e">this constructor</a>
of the OpenCV <code class="docutils literal notranslate"><span class="pre">Mat</span></code> class to convert our tensor to a <code class="docutils literal notranslate"><span class="pre">Mat</span></code> object. We pass
it the number of rows and columns of the original <code class="docutils literal notranslate"><span class="pre">image</span></code> tensor, the datatype
(which we’ll fix as <code class="docutils literal notranslate"><span class="pre">float32</span></code> for this example), and finally a raw pointer to
the underlying data – a <code class="docutils literal notranslate"><span class="pre">float*</span></code>. What is special about this constructor of
the <code class="docutils literal notranslate"><span class="pre">Mat</span></code> class is that it does not copy the input data. Instead, it will
simply reference this memory for all operations performed on the <code class="docutils literal notranslate"><span class="pre">Mat</span></code>. If an
in-place operation is performed on the <code class="docutils literal notranslate"><span class="pre">image_mat</span></code>, this will be reflected in
the original <code class="docutils literal notranslate"><span class="pre">image</span></code> tensor (and vice-versa). This allows us to call
subsequent OpenCV routines with the library’s native matrix type, even though
we’re actually storing the data in a PyTorch tensor. We repeat this procedure to
convert the <code class="docutils literal notranslate"><span class="pre">warp</span></code> PyTorch tensor to the <code class="docutils literal notranslate"><span class="pre">warp_mat</span></code> OpenCV matrix:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span> <span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">warp_mat</span><span class="p">(</span><span class="cm">/*rows=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
<span class="cm">/*cols=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">size</span><span class="p">(</span><span class="mi">1</span><span class="p">),</span>
<span class="cm">/*type=*/</span><span class="n">CV_32FC1</span><span class="p">,</span>
<span class="cm">/*data=*/</span><span class="n">warp</span><span class="p">.</span><span class="n">data_ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">());</span>
</pre></div>
</div>
<p>Next, we are ready to call the OpenCV function we were so eager to use in
TorchScript: <code class="docutils literal notranslate"><span class="pre">warpPerspective</span></code>. For this, we pass the OpenCV function the
<code class="docutils literal notranslate"><span class="pre">image_mat</span></code> and <code class="docutils literal notranslate"><span class="pre">warp_mat</span></code> matrices, as well as an empty output matrix
called <code class="docutils literal notranslate"><span class="pre">output_mat</span></code>. We also specify the size <code class="docutils literal notranslate"><span class="pre">dsize</span></code> we want the output
matrix (image) to be. It is hardcoded to <code class="docutils literal notranslate"><span class="pre">8</span> <span class="pre">x</span> <span class="pre">8</span></code> for this example:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span> <span class="n">cv</span><span class="o">::</span><span class="n">Mat</span> <span class="n">output_mat</span><span class="p">;</span>
<span class="n">cv</span><span class="o">::</span><span class="n">warpPerspective</span><span class="p">(</span><span class="n">image_mat</span><span class="p">,</span> <span class="n">output_mat</span><span class="p">,</span> <span class="n">warp_mat</span><span class="p">,</span> <span class="cm">/*dsize=*/</span><span class="p">{</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">});</span>
</pre></div>
</div>
<p>The final step in our custom operator implementation is to convert the
<code class="docutils literal notranslate"><span class="pre">output_mat</span></code> back into a PyTorch tensor, so that we can further use it in
PyTorch. This is strikingly similar to what we did earlier to convert in the
other direction. In this case, PyTorch provides a <code class="docutils literal notranslate"><span class="pre">torch::from_blob</span></code> method. A
<em>blob</em> in this case is intended to mean some opaque, flat pointer to memory that
we want to interpret as a PyTorch tensor. The call to <code class="docutils literal notranslate"><span class="pre">torch::from_blob</span></code> looks
like this:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span> <span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">output</span> <span class="o">=</span> <span class="n">torch</span><span class="o">::</span><span class="n">from_blob</span><span class="p">(</span><span class="n">output_mat</span><span class="p">.</span><span class="n">ptr</span><span class="o"><</span><span class="kt">float</span><span class="o">></span><span class="p">(),</span> <span class="cm">/*sizes=*/</span><span class="p">{</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">});</span>
<span class="k">return</span> <span class="n">output</span><span class="p">.</span><span class="n">clone</span><span class="p">();</span>
</pre></div>
</div>
<p>We use the <code class="docutils literal notranslate"><span class="pre">.ptr<float>()</span></code> method on the OpenCV <code class="docutils literal notranslate"><span class="pre">Mat</span></code> class to get a raw
pointer to the underlying data (just like <code class="docutils literal notranslate"><span class="pre">.data_ptr<float>()</span></code> for the PyTorch
tensor earlier). We also specify the output shape of the tensor, which we
hardcoded as <code class="docutils literal notranslate"><span class="pre">8</span> <span class="pre">x</span> <span class="pre">8</span></code>. The output of <code class="docutils literal notranslate"><span class="pre">torch::from_blob</span></code> is then a
<code class="docutils literal notranslate"><span class="pre">torch::Tensor</span></code>, pointing to the memory owned by the OpenCV matrix.</p>
<p>Before returning this tensor from our operator implementation, we must call
<code class="docutils literal notranslate"><span class="pre">.clone()</span></code> on the tensor to perform a memory copy of the underlying data. The
reason for this is that <code class="docutils literal notranslate"><span class="pre">torch::from_blob</span></code> returns a tensor that does not own
its data. At that point, the data is still owned by the OpenCV matrix. However,
this OpenCV matrix will go out of scope and be deallocated at the end of the
function. If we returned the <code class="docutils literal notranslate"><span class="pre">output</span></code> tensor as-is, it would point to invalid
memory by the time we use it outside the function. Calling <code class="docutils literal notranslate"><span class="pre">.clone()</span></code> returns
a new tensor with a copy of the original data that the new tensor owns itself.
It is thus safe to return to the outside world.</p>
</div>
<div class="section" id="registering-the-custom-operator-with-torchscript">
<h2>Registering the Custom Operator with TorchScript<a class="headerlink" href="#registering-the-custom-operator-with-torchscript" title="Permalink to this headline">¶</a></h2>
<p>Now that have implemented our custom operator in C++, we need to <em>register</em> it
with the TorchScript runtime and compiler. This will allow the TorchScript
compiler to resolve references to our custom operator in TorchScript code.
If you have ever used the pybind11 library, our syntax for registration
resembles the pybind11 syntax very closely. To register a single function,
we write:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="n">TORCH_LIBRARY</span><span class="p">(</span><span class="n">my_ops</span><span class="p">,</span> <span class="n">m</span><span class="p">)</span> <span class="p">{</span>
<span class="n">m</span><span class="p">.</span><span class="n">def</span><span class="p">(</span><span class="s">"warp_perspective"</span><span class="p">,</span> <span class="n">warp_perspective</span><span class="p">);</span>
<span class="p">}</span>
</pre></div>
</div>
<p>somewhere at the top level of our <code class="docutils literal notranslate"><span class="pre">op.cpp</span></code> file. The <code class="docutils literal notranslate"><span class="pre">TORCH_LIBRARY</span></code> macro
creates a function that will be called when your program starts. The name
of your library (<code class="docutils literal notranslate"><span class="pre">my_ops</span></code>) is given as the first argument (it should not
be in quotes). The second argument (<code class="docutils literal notranslate"><span class="pre">m</span></code>) defines a variable of type
<code class="docutils literal notranslate"><span class="pre">torch::Library</span></code> which is the main interface to register your operators.
The method <code class="docutils literal notranslate"><span class="pre">Library::def</span></code> actually creates an operator named <code class="docutils literal notranslate"><span class="pre">warp_perspective</span></code>,
exposing it to both Python and TorchScript. You can define as many operators
as you like by making multiple calls to <code class="docutils literal notranslate"><span class="pre">def</span></code>.</p>
<p>Behinds the scenes, the <code class="docutils literal notranslate"><span class="pre">def</span></code> function is actually doing quite a bit of work:
it is using template metaprogramming to inspect the type signature of your
function and translate it into an operator schema which specifies the operators
type within TorchScript’s type system.</p>
</div>
<div class="section" id="building-the-custom-operator">
<h2>Building the Custom Operator<a class="headerlink" href="#building-the-custom-operator" title="Permalink to this headline">¶</a></h2>
<p>Now that we have implemented our custom operator in C++ and written its
registration code, it is time to build the operator into a (shared) library that
we can load into Python for research and experimentation, or into C++ for
inference in a no-Python environment. There exist multiple ways to build our
operator, using either pure CMake, or Python alternatives like <code class="docutils literal notranslate"><span class="pre">setuptools</span></code>.
For brevity, the paragraphs below only discuss the CMake approach. The appendix
of this tutorial dives into other alternatives.</p>
<div class="section" id="environment-setup">
<h3>Environment setup<a class="headerlink" href="#environment-setup" title="Permalink to this headline">¶</a></h3>
<p>We need an installation of PyTorch and OpenCV. The easiest and most platform
independent way to get both is to via Conda:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">conda</span> <span class="n">install</span> <span class="o">-</span><span class="n">c</span> <span class="n">pytorch</span> <span class="n">pytorch</span>
<span class="n">conda</span> <span class="n">install</span> <span class="n">opencv</span>
</pre></div>
</div>
</div>
<div class="section" id="building-with-cmake">
<h3>Building with CMake<a class="headerlink" href="#building-with-cmake" title="Permalink to this headline">¶</a></h3>
<p>To build our custom operator into a shared library using the <a class="reference external" href="https://cmake.org">CMake</a> build system, we need to write a short <code class="docutils literal notranslate"><span class="pre">CMakeLists.txt</span></code>
file and place it with our previous <code class="docutils literal notranslate"><span class="pre">op.cpp</span></code> file. For this, let’s agree on a
a directory structure that looks like this:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">warp</span><span class="o">-</span><span class="n">perspective</span><span class="o">/</span>
<span class="n">op</span><span class="o">.</span><span class="n">cpp</span>
<span class="n">CMakeLists</span><span class="o">.</span><span class="n">txt</span>
</pre></div>
</div>
<p>The contents of our <code class="docutils literal notranslate"><span class="pre">CMakeLists.txt</span></code> file should then be the following:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="n">cmake_minimum_required</span><span class="p">(</span><span class="n">VERSION</span> <span class="mf">3.1</span> <span class="n">FATAL_ERROR</span><span class="p">)</span>
<span class="n">project</span><span class="p">(</span><span class="n">warp_perspective</span><span class="p">)</span>
<span class="n">find_package</span><span class="p">(</span><span class="n">Torch</span> <span class="n">REQUIRED</span><span class="p">)</span>
<span class="n">find_package</span><span class="p">(</span><span class="n">OpenCV</span> <span class="n">REQUIRED</span><span class="p">)</span>
<span class="cp"># Define our library target</span>
<span class="n">add_library</span><span class="p">(</span><span class="n">warp_perspective</span> <span class="n">SHARED</span> <span class="n">op</span><span class="p">.</span><span class="n">cpp</span><span class="p">)</span>
<span class="cp"># Enable C++14</span>
<span class="n">target_compile_features</span><span class="p">(</span><span class="n">warp_perspective</span> <span class="n">PRIVATE</span> <span class="n">cxx_std_14</span><span class="p">)</span>
<span class="cp"># Link against LibTorch</span>
<span class="n">target_link_libraries</span><span class="p">(</span><span class="n">warp_perspective</span> <span class="s">"${TORCH_LIBRARIES}"</span><span class="p">)</span>
<span class="cp"># Link against OpenCV</span>
<span class="n">target_link_libraries</span><span class="p">(</span><span class="n">warp_perspective</span> <span class="n">opencv_core</span> <span class="n">opencv_imgproc</span><span class="p">)</span>
</pre></div>
</div>
<p>To now build our operator, we can run the following commands from our
<code class="docutils literal notranslate"><span class="pre">warp_perspective</span></code> folder:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>$ mkdir build
$ <span class="nb">cd</span> build
$ cmake -DCMAKE_PREFIX_PATH<span class="o">=</span><span class="s2">"</span><span class="k">$(</span>python -c <span class="s1">'import torch.utils; print(torch.utils.cmake_prefix_path)'</span><span class="k">)</span><span class="s2">"</span> ..
-- The C compiler identification is GNU <span class="m">5</span>.4.0
-- The CXX compiler identification is GNU <span class="m">5</span>.4.0
-- Check <span class="k">for</span> working C compiler: /usr/bin/cc
-- Check <span class="k">for</span> working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - <span class="k">done</span>
-- Detecting C compile features
-- Detecting C compile features - <span class="k">done</span>
-- Check <span class="k">for</span> working CXX compiler: /usr/bin/c++
-- Check <span class="k">for</span> working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - <span class="k">done</span>
-- Detecting CXX compile features
-- Detecting CXX compile features - <span class="k">done</span>
-- Looking <span class="k">for</span> pthread.h
-- Looking <span class="k">for</span> pthread.h - found
-- Looking <span class="k">for</span> pthread_create
-- Looking <span class="k">for</span> pthread_create - not found
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthreads
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthreads - not found
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthread
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthread - found
-- Found Threads: TRUE
-- Found torch: /libtorch/lib/libtorch.so
-- Configuring <span class="k">done</span>
-- Generating <span class="k">done</span>
-- Build files have been written to: /warp_perspective/build
$ make -j
Scanning dependencies of target warp_perspective
<span class="o">[</span> <span class="m">50</span>%<span class="o">]</span> Building CXX object CMakeFiles/warp_perspective.dir/op.cpp.o
<span class="o">[</span><span class="m">100</span>%<span class="o">]</span> Linking CXX shared library libwarp_perspective.so
<span class="o">[</span><span class="m">100</span>%<span class="o">]</span> Built target warp_perspective
</pre></div>
</div>
<p>which will place a <code class="docutils literal notranslate"><span class="pre">libwarp_perspective.so</span></code> shared library file in the
<code class="docutils literal notranslate"><span class="pre">build</span></code> folder. In the <code class="docutils literal notranslate"><span class="pre">cmake</span></code> command above, we use the helper
variable <code class="docutils literal notranslate"><span class="pre">torch.utils.cmake_prefix_path</span></code> to conveniently tell us where
the cmake files for our PyTorch install are.</p>
<p>We will explore how to use and call our operator in detail further below, but to
get an early sensation of success, we can try running the following code in
Python:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">load_library</span><span class="p">(</span><span class="s2">"build/libwarp_perspective.so"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">my_ops</span><span class="o">.</span><span class="n">warp_perspective</span><span class="p">)</span>
</pre></div>
</div>
<p>If all goes well, this should print something like:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="n">my_ops</span><span class="p">::</span><span class="n">warp_perspective</span> <span class="n">of</span> <span class="n">PyCapsule</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f618fc6fa50</span><span class="o">></span>
</pre></div>
</div>
<p>which is the Python function we will later use to invoke our custom operator.</p>
</div>
</div>
<div class="section" id="using-the-torchscript-custom-operator-in-python">
<h2>Using the TorchScript Custom Operator in Python<a class="headerlink" href="#using-the-torchscript-custom-operator-in-python" title="Permalink to this headline">¶</a></h2>
<p>Once our custom operator is built into a shared library we are ready to use
this operator in our TorchScript models in Python. There are two parts to this:
first loading the operator into Python, and second using the operator in
TorchScript code.</p>
<p>You already saw how to import your operator into Python:
<code class="docutils literal notranslate"><span class="pre">torch.ops.load_library()</span></code>. This function takes the path to a shared library
containing custom operators, and loads it into the current process. Loading the
shared library will also execute the <code class="docutils literal notranslate"><span class="pre">TORCH_LIBRARY</span></code> block. This will register
our custom operator with the TorchScript compiler and allow us to use that
operator in TorchScript code.</p>
<p>You can refer to your loaded operator as <code class="docutils literal notranslate"><span class="pre">torch.ops.<namespace>.<function></span></code>,
where <code class="docutils literal notranslate"><span class="pre"><namespace></span></code> is the namespace part of your operator name, and
<code class="docutils literal notranslate"><span class="pre"><function></span></code> the function name of your operator. For the operator we wrote
above, the namespace was <code class="docutils literal notranslate"><span class="pre">my_ops</span></code> and the function name <code class="docutils literal notranslate"><span class="pre">warp_perspective</span></code>,
which means our operator is available as <code class="docutils literal notranslate"><span class="pre">torch.ops.my_ops.warp_perspective</span></code>.
While this function can be used in scripted or traced TorchScript modules, we
can also just use it in vanilla eager PyTorch and pass it regular PyTorch
tensors:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">load_library</span><span class="p">(</span><span class="s2">"build/libwarp_perspective.so"</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">my_ops</span><span class="o">.</span><span class="n">warp_perspective</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">32</span><span class="p">,</span> <span class="mi">32</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">)))</span>
</pre></div>
</div>
<p>producing:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">tensor</span><span class="p">([[</span><span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.3218</span><span class="p">,</span> <span class="mf">0.4611</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">],</span>
<span class="p">[</span><span class="mf">0.3746</span><span class="p">,</span> <span class="mf">0.0978</span><span class="p">,</span> <span class="mf">0.5005</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">,</span> <span class="mf">0.4636</span><span class="p">],</span>
<span class="p">[</span><span class="mf">0.3245</span><span class="p">,</span> <span class="mf">0.0169</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.4458</span><span class="p">,</span> <span class="mf">0.4458</span><span class="p">,</span> <span class="mf">0.4458</span><span class="p">],</span>
<span class="o">...</span><span class="p">,</span>
<span class="p">[</span><span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1692</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">],</span>
<span class="p">[</span><span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1692</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">],</span>
<span class="p">[</span><span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1862</span><span class="p">,</span> <span class="mf">0.1692</span><span class="p">,</span> <span class="o">...</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">,</span> <span class="mf">0.0000</span><span class="p">]])</span>
</pre></div>
</div>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">What happens behind the scenes is that the first time you access
<code class="docutils literal notranslate"><span class="pre">torch.ops.namespace.function</span></code> in Python, the TorchScript compiler (in C++
land) will see if a function <code class="docutils literal notranslate"><span class="pre">namespace::function</span></code> has been registered, and
if so, return a Python handle to this function that we can subsequently use to
call into our C++ operator implementation from Python. This is one noteworthy
difference between TorchScript custom operators and C++ extensions: C++
extensions are bound manually using pybind11, while TorchScript custom ops are
bound on the fly by PyTorch itself. Pybind11 gives you more flexibility with
regards to what types and classes you can bind into Python and is thus
recommended for purely eager code, but it is not supported for TorchScript
ops.</p>
</div>
<p>From here on, you can use your custom operator in scripted or traced code just
as you would other functions from the <code class="docutils literal notranslate"><span class="pre">torch</span></code> package. In fact, “standard
library” functions like <code class="docutils literal notranslate"><span class="pre">torch.matmul</span></code> go through largely the same
registration path as custom operators, which makes custom operators really
first-class citizens when it comes to how and where they can be used in
TorchScript. (One difference, however, is that standard library functions
have custom written Python argument parsing logic that differs from
<code class="docutils literal notranslate"><span class="pre">torch.ops</span></code> argument parsing.)</p>
<div class="section" id="using-the-custom-operator-with-tracing">
<h3>Using the Custom Operator with Tracing<a class="headerlink" href="#using-the-custom-operator-with-tracing" title="Permalink to this headline">¶</a></h3>
<p>Let’s start by embedding our operator in a traced function. Recall that for
tracing, we start with some vanilla Pytorch code:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">,</span> <span class="n">z</span><span class="p">):</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="o">+</span> <span class="n">torch</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">z</span><span class="p">)</span>
</pre></div>
</div>
<p>and then call <code class="docutils literal notranslate"><span class="pre">torch.jit.trace</span></code> on it. We further pass <code class="docutils literal notranslate"><span class="pre">torch.jit.trace</span></code>
some example inputs, which it will forward to our implementation to record the
sequence of operations that occur as the inputs flow through it. The result of
this is effectively a “frozen” version of the eager PyTorch program, which the
TorchScript compiler can further analyze, optimize and serialize:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">8</span><span class="p">,</span> <span class="mi">5</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">5</span><span class="p">)]</span>
<span class="n">trace</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">compute</span><span class="p">,</span> <span class="n">inputs</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">trace</span><span class="o">.</span><span class="n">graph</span><span class="p">)</span>
</pre></div>
</div>
<p>Producing:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">graph</span><span class="p">(</span><span class="o">%</span><span class="n">x</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">:</span><span class="mi">1</span><span class="p">),</span>
<span class="o">%</span><span class="n">y</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">),</span>
<span class="o">%</span><span class="n">z</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)):</span>
<span class="o">%</span><span class="mi">3</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">matmul</span><span class="p">(</span><span class="o">%</span><span class="n">x</span><span class="p">,</span> <span class="o">%</span><span class="n">y</span><span class="p">)</span> <span class="c1"># test.py:10:0</span>
<span class="o">%</span><span class="mi">4</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">relu</span><span class="p">(</span><span class="o">%</span><span class="n">z</span><span class="p">)</span> <span class="c1"># test.py:10:0</span>
<span class="o">%</span><span class="mi">5</span> <span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">1</span><span class="p">]()</span> <span class="c1"># test.py:10:0</span>
<span class="o">%</span><span class="mi">6</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">add</span><span class="p">(</span><span class="o">%</span><span class="mi">3</span><span class="p">,</span> <span class="o">%</span><span class="mi">4</span><span class="p">,</span> <span class="o">%</span><span class="mi">5</span><span class="p">)</span> <span class="c1"># test.py:10:0</span>
<span class="k">return</span> <span class="p">(</span><span class="o">%</span><span class="mi">6</span><span class="p">)</span>
</pre></div>
</div>
<p>Now, the exciting revelation is that we can simply drop our custom operator into
our PyTorch trace as if it were <code class="docutils literal notranslate"><span class="pre">torch.relu</span></code> or any other <code class="docutils literal notranslate"><span class="pre">torch</span></code> function:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">,</span> <span class="n">z</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">my_ops</span><span class="o">.</span><span class="n">warp_perspective</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">eye</span><span class="p">(</span><span class="mi">3</span><span class="p">))</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="o">+</span> <span class="n">torch</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">z</span><span class="p">)</span>
</pre></div>
</div>
<p>and then trace it as before:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">8</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">8</span><span class="p">,</span> <span class="mi">5</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">8</span><span class="p">,</span> <span class="mi">5</span><span class="p">)]</span>
<span class="n">trace</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">trace</span><span class="p">(</span><span class="n">compute</span><span class="p">,</span> <span class="n">inputs</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">trace</span><span class="o">.</span><span class="n">graph</span><span class="p">)</span>
</pre></div>
</div>
<p>Producing:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">graph</span><span class="p">(</span><span class="o">%</span><span class="n">x</span><span class="o">.</span><span class="mi">1</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">4</span><span class="p">:</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">:</span><span class="mi">1</span><span class="p">),</span>
<span class="o">%</span><span class="n">y</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">),</span>
<span class="o">%</span><span class="n">z</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)):</span>
<span class="o">%</span><span class="mi">3</span> <span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">3</span><span class="p">]()</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">4</span> <span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">6</span><span class="p">]()</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">5</span> <span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">0</span><span class="p">]()</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">6</span> <span class="p">:</span> <span class="n">Device</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="s2">"cpu"</span><span class="p">]()</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">7</span> <span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">0</span><span class="p">]()</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">8</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">3</span><span class="p">:</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">eye</span><span class="p">(</span><span class="o">%</span><span class="mi">3</span><span class="p">,</span> <span class="o">%</span><span class="mi">4</span><span class="p">,</span> <span class="o">%</span><span class="mi">5</span><span class="p">,</span> <span class="o">%</span><span class="mi">6</span><span class="p">,</span> <span class="o">%</span><span class="mi">7</span><span class="p">)</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="n">x</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">8</span><span class="p">,</span> <span class="mi">8</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">my_ops</span><span class="p">::</span><span class="n">warp_perspective</span><span class="p">(</span><span class="o">%</span><span class="n">x</span><span class="o">.</span><span class="mi">1</span><span class="p">,</span> <span class="o">%</span><span class="mi">8</span><span class="p">)</span> <span class="c1"># test.py:25:0</span>
<span class="o">%</span><span class="mi">10</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">matmul</span><span class="p">(</span><span class="o">%</span><span class="n">x</span><span class="p">,</span> <span class="o">%</span><span class="n">y</span><span class="p">)</span> <span class="c1"># test.py:26:0</span>
<span class="o">%</span><span class="mi">11</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">relu</span><span class="p">(</span><span class="o">%</span><span class="n">z</span><span class="p">)</span> <span class="c1"># test.py:26:0</span>
<span class="o">%</span><span class="mi">12</span> <span class="p">:</span> <span class="nb">int</span> <span class="o">=</span> <span class="n">prim</span><span class="p">::</span><span class="n">Constant</span><span class="p">[</span><span class="n">value</span><span class="o">=</span><span class="mi">1</span><span class="p">]()</span> <span class="c1"># test.py:26:0</span>
<span class="o">%</span><span class="mi">13</span> <span class="p">:</span> <span class="n">Float</span><span class="p">(</span><span class="mi">8</span><span class="p">:</span><span class="mi">5</span><span class="p">,</span> <span class="mi">5</span><span class="p">:</span><span class="mi">1</span><span class="p">)</span> <span class="o">=</span> <span class="n">aten</span><span class="p">::</span><span class="n">add</span><span class="p">(</span><span class="o">%</span><span class="mi">10</span><span class="p">,</span> <span class="o">%</span><span class="mi">11</span><span class="p">,</span> <span class="o">%</span><span class="mi">12</span><span class="p">)</span> <span class="c1"># test.py:26:0</span>
<span class="k">return</span> <span class="p">(</span><span class="o">%</span><span class="mi">13</span><span class="p">)</span>
</pre></div>
</div>
<p>Integrating TorchScript custom ops into traced PyTorch code is as easy as this!</p>
</div>
<div class="section" id="using-the-custom-operator-with-script">
<h3>Using the Custom Operator with Script<a class="headerlink" href="#using-the-custom-operator-with-script" title="Permalink to this headline">¶</a></h3>
<p>Besides tracing, another way to arrive at a TorchScript representation of a
PyTorch program is to directly write your code <em>in</em> TorchScript. TorchScript is
largely a subset of the Python language, with some restrictions that make it
easier for the TorchScript compiler to reason about programs. You turn your
regular PyTorch code into TorchScript by annotating it with
<code class="docutils literal notranslate"><span class="pre">@torch.jit.script</span></code> for free functions and <code class="docutils literal notranslate"><span class="pre">@torch.jit.script_method</span></code> for
methods in a class (which must also derive from <code class="docutils literal notranslate"><span class="pre">torch.jit.ScriptModule</span></code>). See
<a class="reference external" href="https://pytorch.org/docs/master/jit.html">here</a> for more details on
TorchScript annotations.</p>
<p>One particular reason to use TorchScript instead of tracing is that tracing is
unable to capture control flow in PyTorch code. As such, let us consider this
function which does use control flow:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">bool</span><span class="p">(</span><span class="n">x</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="mi">42</span><span class="p">):</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">5</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">10</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="o">+</span> <span class="n">z</span>
</pre></div>
</div>
<p>To convert this function from vanilla PyTorch to TorchScript, we annotate it
with <code class="docutils literal notranslate"><span class="pre">@torch.jit.script</span></code>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nd">@torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span>
<span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">bool</span><span class="p">(</span><span class="n">x</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="mi">42</span><span class="p">):</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">5</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">10</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="o">+</span> <span class="n">z</span>
</pre></div>
</div>
<p>This will just-in-time compile the <code class="docutils literal notranslate"><span class="pre">compute</span></code> function into a graph
representation, which we can inspect in the <code class="docutils literal notranslate"><span class="pre">compute.graph</span></code> property:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">compute</span><span class="o">.</span><span class="n">graph</span>
<span class="go">graph(%x : Dynamic</span>
<span class="go"> %y : Dynamic) {</span>
<span class="go"> %14 : int = prim::Constant[value=1]()</span>
<span class="go"> %2 : int = prim::Constant[value=0]()</span>
<span class="go"> %7 : int = prim::Constant[value=42]()</span>
<span class="go"> %z.1 : int = prim::Constant[value=5]()</span>
<span class="go"> %z.2 : int = prim::Constant[value=10]()</span>
<span class="go"> %4 : Dynamic = aten::select(%x, %2, %2)</span>
<span class="go"> %6 : Dynamic = aten::select(%4, %2, %2)</span>
<span class="go"> %8 : Dynamic = aten::eq(%6, %7)</span>
<span class="go"> %9 : bool = prim::TensorToBool(%8)</span>
<span class="go"> %z : int = prim::If(%9)</span>
<span class="go"> block0() {</span>
<span class="go"> -> (%z.1)</span>
<span class="go"> }</span>
<span class="go"> block1() {</span>
<span class="go"> -> (%z.2)</span>
<span class="go"> }</span>
<span class="go"> %13 : Dynamic = aten::matmul(%x, %y)</span>
<span class="go"> %15 : Dynamic = aten::add(%13, %z, %14)</span>
<span class="go"> return (%15);</span>
<span class="go">}</span>
</pre></div>
</div>
<p>And now, just like before, we can use our custom operator like any other
function inside of our script code:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">load_library</span><span class="p">(</span><span class="s2">"libwarp_perspective.so"</span><span class="p">)</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span>
<span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">bool</span><span class="p">(</span><span class="n">x</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="mi">42</span><span class="p">):</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">5</span>
<span class="k">else</span><span class="p">:</span>
<span class="n">z</span> <span class="o">=</span> <span class="mi">10</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">my_ops</span><span class="o">.</span><span class="n">warp_perspective</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">eye</span><span class="p">(</span><span class="mi">3</span><span class="p">))</span>
<span class="k">return</span> <span class="n">x</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">y</span><span class="p">)</span> <span class="o">+</span> <span class="n">z</span>
</pre></div>
</div>
<p>When the TorchScript compiler sees the reference to
<code class="docutils literal notranslate"><span class="pre">torch.ops.my_ops.warp_perspective</span></code>, it will find the implementation we
registered via the <code class="docutils literal notranslate"><span class="pre">TORCH_LIBRARY</span></code> function in C++, and compile it into its
graph representation:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">compute</span><span class="o">.</span><span class="n">graph</span>
<span class="go">graph(%x.1 : Dynamic</span>
<span class="go"> %y : Dynamic) {</span>
<span class="go"> %20 : int = prim::Constant[value=1]()</span>
<span class="go"> %16 : int[] = prim::Constant[value=[0, -1]]()</span>
<span class="go"> %14 : int = prim::Constant[value=6]()</span>
<span class="go"> %2 : int = prim::Constant[value=0]()</span>
<span class="go"> %7 : int = prim::Constant[value=42]()</span>
<span class="go"> %z.1 : int = prim::Constant[value=5]()</span>
<span class="go"> %z.2 : int = prim::Constant[value=10]()</span>
<span class="go"> %13 : int = prim::Constant[value=3]()</span>
<span class="go"> %4 : Dynamic = aten::select(%x.1, %2, %2)</span>
<span class="go"> %6 : Dynamic = aten::select(%4, %2, %2)</span>
<span class="go"> %8 : Dynamic = aten::eq(%6, %7)</span>
<span class="go"> %9 : bool = prim::TensorToBool(%8)</span>
<span class="go"> %z : int = prim::If(%9)</span>
<span class="go"> block0() {</span>
<span class="go"> -> (%z.1)</span>
<span class="go"> }</span>
<span class="go"> block1() {</span>
<span class="go"> -> (%z.2)</span>
<span class="go"> }</span>
<span class="go"> %17 : Dynamic = aten::eye(%13, %14, %2, %16)</span>
<span class="go"> %x : Dynamic = my_ops::warp_perspective(%x.1, %17)</span>
<span class="go"> %19 : Dynamic = aten::matmul(%x, %y)</span>
<span class="go"> %21 : Dynamic = aten::add(%19, %z, %20)</span>
<span class="go"> return (%21);</span>
<span class="go"> }</span>
</pre></div>
</div>
<p>Notice in particular the reference to <code class="docutils literal notranslate"><span class="pre">my_ops::warp_perspective</span></code> at the end of
the graph.</p>
<div class="admonition attention">
<p class="first admonition-title">Attention</p>
<p class="last">The TorchScript graph representation is still subject to change. Do not rely
on it looking like this.</p>
</div>
<p>And that’s really it when it comes to using our custom operator in Python. In
short, you import the library containing your operator(s) using
<code class="docutils literal notranslate"><span class="pre">torch.ops.load_library</span></code>, and call your custom op like any other <code class="docutils literal notranslate"><span class="pre">torch</span></code>
operator from your traced or scripted TorchScript code.</p>
</div>
</div>
<div class="section" id="using-the-torchscript-custom-operator-in-c">
<h2>Using the TorchScript Custom Operator in C++<a class="headerlink" href="#using-the-torchscript-custom-operator-in-c" title="Permalink to this headline">¶</a></h2>
<p>One useful feature of TorchScript is the ability to serialize a model into an
on-disk file. This file can be sent over the wire, stored in a file system or,
more importantly, be dynamically deserialized and executed without needing to
keep the original source code around. This is possible in Python, but also in
C++. For this, PyTorch provides <a class="reference external" href="https://pytorch.org/cppdocs/">a pure C++ API</a>
for deserializing as well as executing TorchScript models. If you haven’t yet,
please read <a class="reference external" href="https://tutorials.pytorch.kr/advanced/cpp_export.html">the tutorial on loading and running serialized TorchScript models
in C++</a>, on which the
next few paragraphs will build.</p>
<p>In short, custom operators can be executed just like regular <code class="docutils literal notranslate"><span class="pre">torch</span></code> operators
even when deserialized from a file and run in C++. The only requirement for this
is to link the custom operator shared library we built earlier with the C++
application in which we execute the model. In Python, this worked simply calling
<code class="docutils literal notranslate"><span class="pre">torch.ops.load_library</span></code>. In C++, you need to link the shared library with
your main application in whatever build system you are using. The following
example will showcase this using CMake.</p>
<div class="admonition note">
<p class="first admonition-title">Note</p>
<p class="last">Technically, you can also dynamically load the shared library into your C++
application at runtime in much the same way we did it in Python. On Linux,
<a class="reference external" href="https://tldp.org/HOWTO/Program-Library-HOWTO/dl-libraries.html">you can do this with dlopen</a>. There exist
equivalents on other platforms.</p>
</div>
<p>Building on the C++ execution tutorial linked above, let’s start with a minimal
C++ application in one file, <code class="docutils literal notranslate"><span class="pre">main.cpp</span></code> in a different folder from our
custom operator, that loads and executes a serialized TorchScript model:</p>
<div class="highlight-cpp notranslate"><div class="highlight"><pre><span></span><span class="cp">#include</span> <span class="cpf"><torch/script.h></span><span class="c1"> // One-stop header.</span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><iostream></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><memory></span><span class="cp"></span>
<span class="kt">int</span> <span class="nf">main</span><span class="p">(</span><span class="kt">int</span> <span class="n">argc</span><span class="p">,</span> <span class="k">const</span> <span class="kt">char</span><span class="o">*</span> <span class="n">argv</span><span class="p">[])</span> <span class="p">{</span>
<span class="k">if</span> <span class="p">(</span><span class="n">argc</span> <span class="o">!=</span> <span class="mi">2</span><span class="p">)</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cerr</span> <span class="o"><<</span> <span class="s">"usage: example-app <path-to-exported-script-module></span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">return</span> <span class="mi">-1</span><span class="p">;</span>
<span class="p">}</span>
<span class="c1">// Deserialize the ScriptModule from a file using torch::jit::load().</span>
<span class="n">torch</span><span class="o">::</span><span class="n">jit</span><span class="o">::</span><span class="n">script</span><span class="o">::</span><span class="n">Module</span> <span class="k">module</span> <span class="o">=</span> <span class="n">torch</span><span class="o">::</span><span class="n">jit</span><span class="o">::</span><span class="n">load</span><span class="p">(</span><span class="n">argv</span><span class="p">[</span><span class="mi">1</span><span class="p">]);</span>
<span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o"><</span><span class="n">torch</span><span class="o">::</span><span class="n">jit</span><span class="o">::</span><span class="n">IValue</span><span class="o">></span> <span class="n">inputs</span><span class="p">;</span>
<span class="n">inputs</span><span class="p">.</span><span class="n">push_back</span><span class="p">(</span><span class="n">torch</span><span class="o">::</span><span class="n">randn</span><span class="p">({</span><span class="mi">4</span><span class="p">,</span> <span class="mi">8</span><span class="p">}));</span>
<span class="n">inputs</span><span class="p">.</span><span class="n">push_back</span><span class="p">(</span><span class="n">torch</span><span class="o">::</span><span class="n">randn</span><span class="p">({</span><span class="mi">8</span><span class="p">,</span> <span class="mi">5</span><span class="p">}));</span>
<span class="n">torch</span><span class="o">::</span><span class="n">Tensor</span> <span class="n">output</span> <span class="o">=</span> <span class="k">module</span><span class="p">.</span><span class="n">forward</span><span class="p">(</span><span class="n">std</span><span class="o">::</span><span class="n">move</span><span class="p">(</span><span class="n">inputs</span><span class="p">)).</span><span class="n">toTensor</span><span class="p">();</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="n">output</span> <span class="o"><<</span> <span class="n">std</span><span class="o">::</span><span class="n">endl</span><span class="p">;</span>
<span class="p">}</span>
</pre></div>
</div>
<p>Along with a small <code class="docutils literal notranslate"><span class="pre">CMakeLists.txt</span></code> file:</p>
<div class="highlight-cmake notranslate"><div class="highlight"><pre><span></span><span class="nb">cmake_minimum_required</span><span class="p">(</span><span class="s">VERSION</span> <span class="s">3.1</span> <span class="s">FATAL_ERROR</span><span class="p">)</span>
<span class="nb">project</span><span class="p">(</span><span class="s">example_app</span><span class="p">)</span>
<span class="nb">find_package</span><span class="p">(</span><span class="s">Torch</span> <span class="s">REQUIRED</span><span class="p">)</span>
<span class="nb">add_executable</span><span class="p">(</span><span class="s">example_app</span> <span class="s">main.cpp</span><span class="p">)</span>
<span class="nb">target_link_libraries</span><span class="p">(</span><span class="s">example_app</span> <span class="s2">"${TORCH_LIBRARIES}"</span><span class="p">)</span>
<span class="nb">target_compile_features</span><span class="p">(</span><span class="s">example_app</span> <span class="s">PRIVATE</span> <span class="s">cxx_range_for</span><span class="p">)</span>
</pre></div>
</div>
<p>At this point, we should be able to build the application:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>$ mkdir build
$ <span class="nb">cd</span> build
$ cmake -DCMAKE_PREFIX_PATH<span class="o">=</span><span class="s2">"</span><span class="k">$(</span>python -c <span class="s1">'import torch.utils; print(torch.utils.cmake_prefix_path)'</span><span class="k">)</span><span class="s2">"</span> ..
-- The C compiler identification is GNU <span class="m">5</span>.4.0
-- The CXX compiler identification is GNU <span class="m">5</span>.4.0
-- Check <span class="k">for</span> working C compiler: /usr/bin/cc
-- Check <span class="k">for</span> working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - <span class="k">done</span>
-- Detecting C compile features
-- Detecting C compile features - <span class="k">done</span>
-- Check <span class="k">for</span> working CXX compiler: /usr/bin/c++
-- Check <span class="k">for</span> working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - <span class="k">done</span>
-- Detecting CXX compile features
-- Detecting CXX compile features - <span class="k">done</span>
-- Looking <span class="k">for</span> pthread.h
-- Looking <span class="k">for</span> pthread.h - found
-- Looking <span class="k">for</span> pthread_create
-- Looking <span class="k">for</span> pthread_create - not found
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthreads
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthreads - not found
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthread
-- Looking <span class="k">for</span> pthread_create <span class="k">in</span> pthread - found
-- Found Threads: TRUE
-- Found torch: /libtorch/lib/libtorch.so
-- Configuring <span class="k">done</span>
-- Generating <span class="k">done</span>
-- Build files have been written to: /example_app/build
$ make -j
Scanning dependencies of target example_app
<span class="o">[</span> <span class="m">50</span>%<span class="o">]</span> Building CXX object CMakeFiles/example_app.dir/main.cpp.o
<span class="o">[</span><span class="m">100</span>%<span class="o">]</span> Linking CXX executable example_app
<span class="o">[</span><span class="m">100</span>%<span class="o">]</span> Built target example_app
</pre></div>
</div>
<p>And run it without passing a model just yet:</p>
<div class="highlight-shell notranslate"><div class="highlight"><pre><span></span>$ ./example_app
usage: example_app <path-to-exported-script-module>
</pre></div>
</div>
<p>Next, let’s serialize the script function we wrote earlier that uses our custom
operator:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">ops</span><span class="o">.</span><span class="n">load_library</span><span class="p">(</span><span class="s2">"libwarp_perspective.so"</span><span class="p">)</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">jit</span><span class="o">.</span><span class="n">script</span>
<span class="k">def</span> <span class="nf">compute</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">y</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">bool</span><span class="p">(</span><span class="n">x</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="mi">42</span><span class="p">):</span>