forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodules.html
1444 lines (1208 loc) · 134 KB
/
modules.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 name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Modules — PyTorch 1.10.1 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/notes/modules.html"/>
<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="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="Multiprocessing best practices" href="multiprocessing.html" />
<link rel="prev" title="Features for large-scale deployments" href="large_scale_deployments.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<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/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/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.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/elastic/">
<span class="dropdown-title">TorchElastic</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</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">
<a href='https://pytorch.org/docs/versions.html'>1.10.1 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="../search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<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">Notes</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="windows.html">Windows FAQ</a></li>
</ul>
<p class="caption"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
</ul>
<p class="caption"><span class="caption-text">Python API</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="../autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="../amp.html">torch.cuda.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="../backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="../futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="../fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="../jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="../linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="../special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="../package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="../profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="../nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="../onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="../optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="../complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="../ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1"><a class="reference internal" href="../quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="../rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="../random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="../sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="../storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="../benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="../bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="../checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="../cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="../data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="../dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="../model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="../tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="../type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="../named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="../name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="../__config__.html">torch.__config__</a></li>
</ul>
<p class="caption"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="../community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/governance.html">PyTorch Governance</a></li>
<li class="toctree-l1"><a class="reference internal" href="../community/persons_of_interest.html">PyTorch Governance | Persons of Interest</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">
Docs
</a> >
</li>
<li>Modules</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/notes/modules.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="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="modules">
<span id="id1"></span><h1>Modules<a class="headerlink" href="#modules" title="Permalink to this headline">¶</a></h1>
<p>PyTorch uses modules to represent neural networks. Modules are:</p>
<ul class="simple">
<li><p><strong>Building blocks of stateful computation.</strong>
PyTorch provides a robust library of modules and makes it simple to define new custom modules, allowing for
easy construction of elaborate, multi-layer neural networks.</p></li>
<li><p><strong>Tightly integrated with PyTorch’s</strong>
<a class="reference external" href="https://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html">autograd</a>
<strong>system.</strong> Modules make it simple to specify learnable parameters for PyTorch’s Optimizers to update.</p></li>
<li><p><strong>Easy to work with and transform.</strong> Modules are straightforward to save and restore, transfer between
CPU / GPU / TPU devices, prune, quantize, and more.</p></li>
</ul>
<p>This note describes modules, and is intended for all PyTorch users. Since modules are so fundamental to PyTorch,
many topics in this note are elaborated on in other notes or tutorials, and links to many of those documents
are provided here as well.</p>
<div class="contents local topic" id="contents">
<ul class="simple">
<li><p><a class="reference internal" href="#a-simple-custom-module" id="id4">A Simple Custom Module</a></p></li>
<li><p><a class="reference internal" href="#modules-as-building-blocks" id="id5">Modules as Building Blocks</a></p></li>
<li><p><a class="reference internal" href="#neural-network-training-with-modules" id="id6">Neural Network Training with Modules</a></p></li>
<li><p><a class="reference internal" href="#module-state" id="id7">Module State</a></p></li>
<li><p><a class="reference internal" href="#module-initialization" id="id8">Module Initialization</a></p></li>
<li><p><a class="reference internal" href="#module-hooks" id="id9">Module Hooks</a></p></li>
<li><p><a class="reference internal" href="#advanced-features" id="id10">Advanced Features</a></p>
<ul>
<li><p><a class="reference internal" href="#distributed-training" id="id11">Distributed Training</a></p></li>
<li><p><a class="reference internal" href="#profiling-performance" id="id12">Profiling Performance</a></p></li>
<li><p><a class="reference internal" href="#improving-performance-with-quantization" id="id13">Improving Performance with Quantization</a></p></li>
<li><p><a class="reference internal" href="#improving-memory-usage-with-pruning" id="id14">Improving Memory Usage with Pruning</a></p></li>
<li><p><a class="reference internal" href="#deploying-with-torchscript" id="id15">Deploying with TorchScript</a></p></li>
<li><p><a class="reference internal" href="#parametrizations" id="id16">Parametrizations</a></p></li>
<li><p><a class="reference internal" href="#transforming-modules-with-fx" id="id17">Transforming Modules with FX</a></p></li>
</ul>
</li>
</ul>
</div>
<div class="section" id="a-simple-custom-module">
<h2><a class="toc-backref" href="#id4">A Simple Custom Module</a><a class="headerlink" href="#a-simple-custom-module" title="Permalink to this headline">¶</a></h2>
<p>To get started, let’s look at a simpler, custom version of PyTorch’s <a class="reference internal" href="../generated/torch.nn.Linear.html#torch.nn.Linear" title="torch.nn.Linear"><code class="xref py py-class docutils literal notranslate"><span class="pre">Linear</span></code></a> module.
This module applies an affine transformation to its input.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
<span class="k">class</span> <span class="nc">MyLinear</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">in_features</span><span class="p">,</span> <span class="n">out_features</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weight</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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="n">in_features</span><span class="p">,</span> <span class="n">out_features</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">bias</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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="n">out_features</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="k">return</span> <span class="p">(</span><span class="nb">input</span> <span class="o">@</span> <span class="bp">self</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">bias</span>
</pre></div>
</div>
<p>This simple module has the following fundamental characteristics of modules:</p>
<ul class="simple">
<li><p><strong>It inherits from the base Module class.</strong>
All modules should subclass <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module" title="torch.nn.Module"><code class="xref py py-class docutils literal notranslate"><span class="pre">Module</span></code></a> for composability with other modules.</p></li>
<li><p><strong>It defines some “state” that is used in computation.</strong>
Here, the state consists of randomly-initialized <code class="docutils literal notranslate"><span class="pre">weight</span></code> and <code class="docutils literal notranslate"><span class="pre">bias</span></code> tensors that define the affine
transformation. Because each of these is defined as a <a class="reference internal" href="../generated/torch.nn.parameter.Parameter.html#torch.nn.parameter.Parameter" title="torch.nn.parameter.Parameter"><code class="xref py py-class docutils literal notranslate"><span class="pre">Parameter</span></code></a>, they are
<em>registered</em> for the module and will automatically be tracked and returned from calls
to <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.parameters" title="torch.nn.Module.parameters"><code class="xref py py-func docutils literal notranslate"><span class="pre">parameters()</span></code></a>. Parameters can be
considered the “learnable” aspects of the module’s computation (more on this later). Note that modules
are not required to have state, and can also be stateless.</p></li>
<li><p><strong>It defines a forward() function that performs the computation.</strong> For this affine transformation module, the input
is matrix-multiplied with the <code class="docutils literal notranslate"><span class="pre">weight</span></code> parameter (using the <code class="docutils literal notranslate"><span class="pre">@</span></code> short-hand notation) and added to the <code class="docutils literal notranslate"><span class="pre">bias</span></code>
parameter to produce the output. More generally, the <code class="docutils literal notranslate"><span class="pre">forward()</span></code> implementation for a module can perform arbitrary
computation involving any number of inputs and outputs.</p></li>
</ul>
<p>This simple module demonstrates how modules package state and computation together. Instances of this module can be
constructed and called:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">m</span> <span class="o">=</span> <span class="n">MyLinear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">sample_input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">m</span><span class="p">(</span><span class="n">sample_input</span><span class="p">)</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.3037</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.0413</span><span class="p">,</span> <span class="o">-</span><span class="mf">4.2057</span><span class="p">],</span> <span class="n">grad_fn</span><span class="o">=<</span><span class="n">AddBackward0</span><span class="o">></span><span class="p">)</span>
</pre></div>
</div>
<p>Note that the module itself is callable, and that calling it invokes its <code class="docutils literal notranslate"><span class="pre">forward()</span></code> function.
This name is in reference to the concepts of “forward pass” and “backward pass”, which apply to each module.
The “forward pass” is responsible for applying the computation represented by the module
to the given input(s) (as shown in the above snippet). The “backward pass” computes gradients of
module outputs with respect to its inputs, which can be used for “training” parameters through gradient
descent methods. PyTorch’s autograd system automatically takes care of this backward pass computation, so it
is not required to manually implement a <code class="docutils literal notranslate"><span class="pre">backward()</span></code> function for each module. The process of training
module parameters through successive forward / backward passes is covered in detail in
<a class="reference internal" href="#neural-network-training-with-modules"><span class="std std-ref">Neural Network Training with Modules</span></a>.</p>
<p>The full set of parameters registered by the module can be iterated through via a call to
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.parameters" title="torch.nn.Module.parameters"><code class="xref py py-func docutils literal notranslate"><span class="pre">parameters()</span></code></a> or <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.named_parameters" title="torch.nn.Module.named_parameters"><code class="xref py py-func docutils literal notranslate"><span class="pre">named_parameters()</span></code></a>,
where the latter includes each parameter’s name:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">parameter</span> <span class="ow">in</span> <span class="n">m</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="n">parameter</span><span class="p">)</span>
<span class="p">:</span> <span class="p">(</span><span class="s1">'weight'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span> <span class="mf">1.0597</span><span class="p">,</span> <span class="mf">1.1796</span><span class="p">,</span> <span class="mf">0.8247</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.5080</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.2635</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.1045</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">0.0593</span><span class="p">,</span> <span class="mf">0.2469</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.4299</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.4926</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.5457</span><span class="p">,</span> <span class="mf">0.4793</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'bias'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([</span> <span class="mf">0.3634</span><span class="p">,</span> <span class="mf">0.2015</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.8525</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
</pre></div>
</div>
<p>In general, the parameters registered by a module are aspects of the module’s computation that should be
“learned”. A later section of this note shows how to update these parameters using one of PyTorch’s Optimizers.
Before we get to that, however, let’s first examine how modules can be composed with one another.</p>
</div>
<div class="section" id="modules-as-building-blocks">
<h2><a class="toc-backref" href="#id5">Modules as Building Blocks</a><a class="headerlink" href="#modules-as-building-blocks" title="Permalink to this headline">¶</a></h2>
<p>Modules can contain other modules, making them useful building blocks for developing more elaborate functionality.
The simplest way to do this is using the <a class="reference internal" href="../generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sequential</span></code></a> module. It allows us to chain together
multiple modules:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">net</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Sequential</span><span class="p">(</span>
<span class="n">MyLinear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">),</span>
<span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">(),</span>
<span class="n">MyLinear</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="p">)</span>
<span class="n">sample_input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">net</span><span class="p">(</span><span class="n">sample_input</span><span class="p">)</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.6749</span><span class="p">],</span> <span class="n">grad_fn</span><span class="o">=<</span><span class="n">AddBackward0</span><span class="o">></span><span class="p">)</span>
</pre></div>
</div>
<p>Note that <a class="reference internal" href="../generated/torch.nn.Sequential.html#torch.nn.Sequential" title="torch.nn.Sequential"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sequential</span></code></a> automatically feeds the output of the first <code class="docutils literal notranslate"><span class="pre">MyLinear</span></code> module as input
into the <a class="reference internal" href="../generated/torch.nn.ReLU.html#torch.nn.ReLU" title="torch.nn.ReLU"><code class="xref py py-class docutils literal notranslate"><span class="pre">ReLU</span></code></a>, and the output of that as input into the second <code class="docutils literal notranslate"><span class="pre">MyLinear</span></code> module. As
shown, it is limited to in-order chaining of modules with a single input and output.</p>
<p>In general, it is recommended to define a custom module for anything beyond the simplest use cases, as this gives
full flexibility on how submodules are used for a module’s computation.</p>
<p>For example, here’s a simple neural network implemented as a custom module:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
<span class="k">class</span> <span class="nc">Net</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l0</span> <span class="o">=</span> <span class="n">MyLinear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l1</span> <span class="o">=</span> <span class="n">MyLinear</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="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l0</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
</pre></div>
</div>
<p>This module is composed of two “children” or “submodules” (<code class="docutils literal notranslate"><span class="pre">l0</span></code> and <code class="docutils literal notranslate"><span class="pre">l1</span></code>) that define the layers of
the neural network and are utilized for computation within the module’s <code class="docutils literal notranslate"><span class="pre">forward()</span></code> method. Immediate
children of a module can be iterated through via a call to <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.children" title="torch.nn.Module.children"><code class="xref py py-func docutils literal notranslate"><span class="pre">children()</span></code></a> or
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.named_children" title="torch.nn.Module.named_children"><code class="xref py py-func docutils literal notranslate"><span class="pre">named_children()</span></code></a>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">net</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span>
<span class="k">for</span> <span class="n">child</span> <span class="ow">in</span> <span class="n">net</span><span class="o">.</span><span class="n">named_children</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="n">child</span><span class="p">)</span>
<span class="p">:</span> <span class="p">(</span><span class="s1">'l0'</span><span class="p">,</span> <span class="n">MyLinear</span><span class="p">())</span>
<span class="p">(</span><span class="s1">'l1'</span><span class="p">,</span> <span class="n">MyLinear</span><span class="p">())</span>
</pre></div>
</div>
<p>To go deeper than just the immediate children, <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.modules" title="torch.nn.Module.modules"><code class="xref py py-func docutils literal notranslate"><span class="pre">modules()</span></code></a> and
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.named_modules" title="torch.nn.Module.named_modules"><code class="xref py py-func docutils literal notranslate"><span class="pre">named_modules()</span></code></a> <em>recursively</em> iterate through a module and its child modules:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">BigNet</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">l1</span> <span class="o">=</span> <span class="n">MyLinear</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">net</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">net</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">l1</span><span class="p">(</span><span class="n">x</span><span class="p">))</span>
<span class="n">big_net</span> <span class="o">=</span> <span class="n">BigNet</span><span class="p">()</span>
<span class="k">for</span> <span class="n">module</span> <span class="ow">in</span> <span class="n">big_net</span><span class="o">.</span><span class="n">named_modules</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="n">module</span><span class="p">)</span>
<span class="p">:</span> <span class="p">(</span><span class="s1">''</span><span class="p">,</span> <span class="n">BigNet</span><span class="p">(</span>
<span class="p">(</span><span class="n">l1</span><span class="p">):</span> <span class="n">MyLinear</span><span class="p">()</span>
<span class="p">(</span><span class="n">net</span><span class="p">):</span> <span class="n">Net</span><span class="p">(</span>
<span class="p">(</span><span class="n">l0</span><span class="p">):</span> <span class="n">MyLinear</span><span class="p">()</span>
<span class="p">(</span><span class="n">l1</span><span class="p">):</span> <span class="n">MyLinear</span><span class="p">()</span>
<span class="p">)</span>
<span class="p">))</span>
<span class="p">(</span><span class="s1">'l1'</span><span class="p">,</span> <span class="n">MyLinear</span><span class="p">())</span>
<span class="p">(</span><span class="s1">'net'</span><span class="p">,</span> <span class="n">Net</span><span class="p">(</span>
<span class="p">(</span><span class="n">l0</span><span class="p">):</span> <span class="n">MyLinear</span><span class="p">()</span>
<span class="p">(</span><span class="n">l1</span><span class="p">):</span> <span class="n">MyLinear</span><span class="p">()</span>
<span class="p">))</span>
<span class="p">(</span><span class="s1">'net.l0'</span><span class="p">,</span> <span class="n">MyLinear</span><span class="p">())</span>
<span class="p">(</span><span class="s1">'net.l1'</span><span class="p">,</span> <span class="n">MyLinear</span><span class="p">())</span>
</pre></div>
</div>
<p>Sometimes, it’s necessary for a module to dynamically define submodules.
The <a class="reference internal" href="../generated/torch.nn.ModuleList.html#torch.nn.ModuleList" title="torch.nn.ModuleList"><code class="xref py py-class docutils literal notranslate"><span class="pre">ModuleList</span></code></a> and <a class="reference internal" href="../generated/torch.nn.ModuleDict.html#torch.nn.ModuleDict" title="torch.nn.ModuleDict"><code class="xref py py-class docutils literal notranslate"><span class="pre">ModuleDict</span></code></a> modules are useful here; they
register submodules from a list or dict:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">DynamicNet</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">num_layers</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linears</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ModuleList</span><span class="p">(</span>
<span class="p">[</span><span class="n">MyLinear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">num_layers</span><span class="p">)])</span>
<span class="bp">self</span><span class="o">.</span><span class="n">activations</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ModuleDict</span><span class="p">({</span>
<span class="s1">'relu'</span><span class="p">:</span> <span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">(),</span>
<span class="s1">'lrelu'</span><span class="p">:</span> <span class="n">nn</span><span class="o">.</span><span class="n">LeakyReLU</span><span class="p">()</span>
<span class="p">})</span>
<span class="bp">self</span><span class="o">.</span><span class="n">final</span> <span class="o">=</span> <span class="n">MyLinear</span><span class="p">(</span><span class="mi">4</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">,</span> <span class="n">act</span><span class="p">):</span>
<span class="k">for</span> <span class="n">linear</span> <span class="ow">in</span> <span class="bp">self</span><span class="o">.</span><span class="n">linears</span><span class="p">:</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">linear</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">activations</span><span class="p">[</span><span class="n">act</span><span class="p">](</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">final</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
<span class="n">dynamic_net</span> <span class="o">=</span> <span class="n">DynamicNet</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span>
<span class="n">sample_input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">dynamic_net</span><span class="p">(</span><span class="n">sample_input</span><span class="p">,</span> <span class="s1">'relu'</span><span class="p">)</span>
</pre></div>
</div>
<p>For any given module, its parameters consist of its direct parameters as well as the parameters of all submodules.
This means that calls to <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.parameters" title="torch.nn.Module.parameters"><code class="xref py py-func docutils literal notranslate"><span class="pre">parameters()</span></code></a> and <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.named_parameters" title="torch.nn.Module.named_parameters"><code class="xref py py-func docutils literal notranslate"><span class="pre">named_parameters()</span></code></a> will
recursively include child parameters, allowing for convenient optimization of all parameters within the network:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">parameter</span> <span class="ow">in</span> <span class="n">dynamic_net</span><span class="o">.</span><span class="n">named_parameters</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="n">parameter</span><span class="p">)</span>
<span class="p">:</span> <span class="p">(</span><span class="s1">'linears.0.weight'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span><span class="o">-</span><span class="mf">1.2051</span><span class="p">,</span> <span class="mf">0.7601</span><span class="p">,</span> <span class="mf">1.1065</span><span class="p">,</span> <span class="mf">0.1963</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">3.0592</span><span class="p">,</span> <span class="mf">0.4354</span><span class="p">,</span> <span class="mf">1.6598</span><span class="p">,</span> <span class="mf">0.9828</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.4446</span><span class="p">,</span> <span class="mf">0.4628</span><span class="p">,</span> <span class="mf">0.8774</span><span class="p">,</span> <span class="mf">1.6848</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.1222</span><span class="p">,</span> <span class="mf">1.5458</span><span class="p">,</span> <span class="mf">1.1729</span><span class="p">,</span> <span class="mf">1.4647</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'linears.0.bias'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([</span> <span class="mf">1.5310</span><span class="p">,</span> <span class="mf">1.0609</span><span class="p">,</span> <span class="o">-</span><span class="mf">2.0940</span><span class="p">,</span> <span class="mf">1.1266</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'linears.1.weight'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span> <span class="mf">2.1113</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.0623</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.0806</span><span class="p">,</span> <span class="mf">0.3508</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.0550</span><span class="p">,</span> <span class="mf">1.5317</span><span class="p">,</span> <span class="mf">1.1064</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.5562</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.4028</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.6942</span><span class="p">,</span> <span class="mf">1.5793</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.0140</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.0329</span><span class="p">,</span> <span class="mf">0.1160</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.7183</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.0434</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'linears.1.bias'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([</span> <span class="mf">0.0361</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.9768</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.3889</span><span class="p">,</span> <span class="mf">1.1613</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'linears.2.weight'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span><span class="o">-</span><span class="mf">2.6340</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.3887</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.9979</span><span class="p">,</span> <span class="mf">0.0767</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.3526</span><span class="p">,</span> <span class="mf">0.8756</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.5847</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.6016</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.3269</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.1608</span><span class="p">,</span> <span class="mf">0.2897</span><span class="p">,</span> <span class="o">-</span><span class="mf">2.0829</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">2.6338</span><span class="p">,</span> <span class="mf">0.9239</span><span class="p">,</span> <span class="mf">0.6943</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.5034</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'linears.2.bias'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([</span> <span class="mf">1.0268</span><span class="p">,</span> <span class="mf">0.4489</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.9403</span><span class="p">,</span> <span class="mf">0.1571</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'final.weight'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span> <span class="mf">0.2509</span><span class="p">],</span> <span class="p">[</span><span class="o">-</span><span class="mf">0.5052</span><span class="p">],</span> <span class="p">[</span> <span class="mf">0.3088</span><span class="p">],</span> <span class="p">[</span><span class="o">-</span><span class="mf">1.4951</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
<span class="p">(</span><span class="s1">'final.bias'</span><span class="p">,</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([</span><span class="mf">0.3381</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">))</span>
</pre></div>
</div>
<p>It’s also easy to move all parameters to a different device or change their precision using
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.to" title="torch.nn.Module.to"><code class="xref py py-func docutils literal notranslate"><span class="pre">to()</span></code></a>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Move all parameters to a CUDA device</span>
<span class="n">dynamic_net</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="o">=</span><span class="s1">'cuda'</span><span class="p">)</span>
<span class="c1"># Change precision of all parameters</span>
<span class="n">dynamic_net</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float64</span><span class="p">)</span>
<span class="n">dynamic_net</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">5</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s1">'cuda'</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float64</span><span class="p">))</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span><span class="mf">6.5166</span><span class="p">],</span> <span class="n">device</span><span class="o">=</span><span class="s1">'cuda:0'</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float64</span><span class="p">,</span> <span class="n">grad_fn</span><span class="o">=<</span><span class="n">AddBackward0</span><span class="o">></span><span class="p">)</span>
</pre></div>
</div>
<p>More generally, an arbitrary function can be applied to a module and its submodules recursively by
using the <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.apply" title="torch.nn.Module.apply"><code class="xref py py-func docutils literal notranslate"><span class="pre">apply()</span></code></a> function. For example, to apply custom initialization to parameters
of a module and its submodules:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Define a function to initialize Linear weights.</span>
<span class="c1"># Note that no_grad() is used here to avoid tracking this computation in the autograd graph.</span>
<span class="nd">@torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">init_weights</span><span class="p">(</span><span class="n">m</span><span class="p">):</span>
<span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">):</span>
<span class="n">nn</span><span class="o">.</span><span class="n">init</span><span class="o">.</span><span class="n">xavier_normal_</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
<span class="n">m</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">fill_</span><span class="p">(</span><span class="mf">0.0</span><span class="p">)</span>
<span class="c1"># Apply the function recursively on the module and its submodules.</span>
<span class="n">dynamic_net</span><span class="o">.</span><span class="n">apply</span><span class="p">(</span><span class="n">init_weights</span><span class="p">)</span>
</pre></div>
</div>
<p>These examples show how elaborate neural networks can be formed through module composition and conveniently
manipulated. To allow for quick and easy construction of neural networks with minimal boilerplate, PyTorch
provides a large library of performant modules within the <code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.nn</span></code> namespace that perform common neural
network operations like pooling, convolutions, loss functions, etc.</p>
<p>In the next section, we give a full example of training a neural network.</p>
<p>For more information, check out:</p>
<ul class="simple">
<li><p>Library of PyTorch-provided modules: <a class="reference external" href="https://pytorch.org/docs/stable/nn.html">torch.nn</a></p></li>
<li><p>Defining neural net modules: <a class="reference external" href="https://pytorch.org/tutorials/beginner/examples_nn/two_layer_net_module.html">https://pytorch.org/tutorials/beginner/examples_nn/two_layer_net_module.html</a></p></li>
</ul>
</div>
<div class="section" id="neural-network-training-with-modules">
<span id="id2"></span><h2><a class="toc-backref" href="#id6">Neural Network Training with Modules</a><a class="headerlink" href="#neural-network-training-with-modules" title="Permalink to this headline">¶</a></h2>
<p>Once a network is built, it has to be trained, and its parameters can be easily optimized with one of PyTorch’s
Optimizers from <a class="reference internal" href="../optim.html#module-torch.optim" title="torch.optim"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.optim</span></code></a>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Create the network (from previous section) and optimizer</span>
<span class="n">net</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span>
<span class="n">optimizer</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">net</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">1e-4</span><span class="p">,</span> <span class="n">weight_decay</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
<span class="c1"># Run a sample training loop that "teaches" the network</span>
<span class="c1"># to output the constant zero function</span>
<span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">10000</span><span class="p">):</span>
<span class="nb">input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">net</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">abs</span><span class="p">(</span><span class="n">output</span><span class="p">)</span>
<span class="n">net</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="c1"># After training, switch the module to eval mode to do inference, compute performance metrics, etc.</span>
<span class="c1"># (see discussion below for a description of training and evaluation modes)</span>
<span class="o">...</span>
<span class="n">net</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="o">...</span>
</pre></div>
</div>
<p>In this simplified example, the network learns to simply output zero, as any non-zero output is “penalized” according
to its absolute value by employing <a class="reference internal" href="../generated/torch.abs.html#torch.abs" title="torch.abs"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.abs()</span></code></a> as a loss function. While this is not a very interesting task, the
key parts of training are present:</p>
<ul class="simple">
<li><p>A network is created.</p></li>
<li><p>An optimizer (in this case, a stochastic gradient descent optimizer) is created, and the network’s
parameters are associated with it.</p></li>
<li><dl class="simple">
<dt>A training loop…</dt><dd><ul>
<li><p>acquires an input,</p></li>
<li><p>runs the network,</p></li>
<li><p>computes a loss,</p></li>
<li><p>zeros the network’s parameters’ gradients,</p></li>
<li><p>calls loss.backward() to update the parameters’ gradients,</p></li>
<li><p>calls optimizer.step() to apply the gradients to the parameters.</p></li>
</ul>
</dd>
</dl>
</li>
</ul>
<p>After the above snippet has been run, note that the network’s parameters have changed. In particular, examining the
value of <code class="docutils literal notranslate"><span class="pre">l1</span></code>‘s <code class="docutils literal notranslate"><span class="pre">weight</span></code> parameter shows that its values are now much closer to 0 (as may be expected):</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nb">print</span><span class="p">(</span><span class="n">net</span><span class="o">.</span><span class="n">l1</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
<span class="p">:</span> <span class="n">Parameter</span> <span class="n">containing</span><span class="p">:</span>
<span class="n">tensor</span><span class="p">([[</span><span class="o">-</span><span class="mf">0.0013</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">0.0030</span><span class="p">],</span>
<span class="p">[</span><span class="o">-</span><span class="mf">0.0008</span><span class="p">]],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
</pre></div>
</div>
<p>Note that the above process is done entirely while the network module is in “training mode”. Modules default to
training mode and can be switched between training and evaluation modes using <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.train" title="torch.nn.Module.train"><code class="xref py py-func docutils literal notranslate"><span class="pre">train()</span></code></a> and
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.eval" title="torch.nn.Module.eval"><code class="xref py py-func docutils literal notranslate"><span class="pre">eval()</span></code></a>. They can behave differently depending on which mode they are in. For example, the
<code class="xref py py-class docutils literal notranslate"><span class="pre">BatchNorm</span></code> module maintains a running mean and variance during training that are not updated
when the module is in evaluation mode. In general, modules should be in training mode during training
and only switched to evaluation mode for inference or evaluation. Below is an example of a custom module
that behaves differently between the two modes:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ModalModule</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">training</span><span class="p">:</span>
<span class="c1"># Add a constant only in training mode.</span>
<span class="k">return</span> <span class="n">x</span> <span class="o">+</span> <span class="mf">1.</span>
<span class="k">else</span><span class="p">:</span>
<span class="k">return</span> <span class="n">x</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">ModalModule</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">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'training mode output: </span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">m</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span><span class="mf">1.6614</span><span class="p">,</span> <span class="mf">1.2669</span><span class="p">,</span> <span class="mf">1.0617</span><span class="p">,</span> <span class="mf">1.6213</span><span class="p">,</span> <span class="mf">0.5481</span><span class="p">])</span>
<span class="n">m</span><span class="o">.</span><span class="n">eval</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="s1">'evaluation mode output: </span><span class="si">{}</span><span class="s1">'</span><span class="o">.</span><span class="n">format</span><span class="p">(</span><span class="n">m</span><span class="p">(</span><span class="n">x</span><span class="p">)))</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">0.6614</span><span class="p">,</span> <span class="mf">0.2669</span><span class="p">,</span> <span class="mf">0.0617</span><span class="p">,</span> <span class="mf">0.6213</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.4519</span><span class="p">])</span>
</pre></div>
</div>
<p>Training neural networks can often be tricky. For more information, check out:</p>
<ul class="simple">
<li><p>Using Optimizers: <a class="reference external" href="https://pytorch.org/tutorials/beginner/examples_nn/two_layer_net_optim.html">https://pytorch.org/tutorials/beginner/examples_nn/two_layer_net_optim.html</a>.</p></li>
<li><p>Neural network training: <a class="reference external" href="https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html">https://pytorch.org/tutorials/beginner/blitz/neural_networks_tutorial.html</a></p></li>
<li><p>Introduction to autograd: <a class="reference external" href="https://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html">https://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html</a></p></li>
</ul>
</div>
<div class="section" id="module-state">
<h2><a class="toc-backref" href="#id7">Module State</a><a class="headerlink" href="#module-state" title="Permalink to this headline">¶</a></h2>
<p>In the previous section, we demonstrated training a module’s “parameters”, or learnable aspects of computation.
Now, if we want to save the trained model to disk, we can do so by saving its <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> (i.e. “state dictionary”):</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Save the module</span>
<span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">net</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="s1">'net.pt'</span><span class="p">)</span>
<span class="o">...</span>
<span class="c1"># Load the module later on</span>
<span class="n">new_net</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span>
<span class="n">new_net</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'net.pt'</span><span class="p">))</span>
<span class="p">:</span> <span class="o"><</span><span class="n">All</span> <span class="n">keys</span> <span class="n">matched</span> <span class="n">successfully</span><span class="o">></span>
</pre></div>
</div>
<p>A module’s <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> contains state that affects its computation. This includes, but is not limited to, the
module’s parameters. For some modules, it may be useful to have state beyond parameters that affects module
computation but is not learnable. For such cases, PyTorch provides the concept of “buffers”, both “persistent”
and “non-persistent”. Following is an overview of the various types of state a module can have:</p>
<ul class="simple">
<li><p><strong>Parameters</strong>: learnable aspects of computation; contained within the <code class="docutils literal notranslate"><span class="pre">state_dict</span></code></p></li>
<li><p><strong>Buffers</strong>: non-learnable aspects of computation</p>
<ul>
<li><p><strong>Persistent</strong> buffers: contained within the <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> (i.e. serialized when saving & loading)</p></li>
<li><p><strong>Non-persistent</strong> buffers: not contained within the <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> (i.e. left out of serialization)</p></li>
</ul>
</li>
</ul>
<p>As a motivating example for the use of buffers, consider a simple module that maintains a running mean. We want
the current value of the running mean to be considered part of the module’s <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> so that it will be
restored when loading a serialized form of the module, but we don’t want it to be learnable.
This snippet shows how to use <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.register_buffer" title="torch.nn.Module.register_buffer"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_buffer()</span></code></a> to accomplish this:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">RunningMean</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">num_features</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">momentum</span> <span class="o">=</span> <span class="n">momentum</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_buffer</span><span class="p">(</span><span class="s1">'mean'</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="n">num_features</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">mean</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">momentum</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">mean</span> <span class="o">+</span> <span class="p">(</span><span class="mf">1.0</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">momentum</span><span class="p">)</span> <span class="o">*</span> <span class="n">x</span>
<span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">mean</span>
</pre></div>
</div>
<p>Now, the current value of the running mean is considered part of the module’s <code class="docutils literal notranslate"><span class="pre">state_dict</span></code>
and will be properly restored when loading the module from disk:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">m</span> <span class="o">=</span> <span class="n">RunningMean</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">10</span><span class="p">):</span>
<span class="nb">input</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">m</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">state_dict</span><span class="p">())</span>
<span class="p">:</span> <span class="n">OrderedDict</span><span class="p">([(</span><span class="s1">'mean'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">0.1041</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.1113</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.0647</span><span class="p">,</span> <span class="mf">0.1515</span><span class="p">]))]))</span>
<span class="c1"># Serialized form will contain the 'mean' tensor</span>
<span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="s1">'mean.pt'</span><span class="p">)</span>
<span class="n">m_loaded</span> <span class="o">=</span> <span class="n">RunningMean</span><span class="p">(</span><span class="mi">4</span><span class="p">)</span>
<span class="n">m_loaded</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'mean.pt'</span><span class="p">))</span>
<span class="k">assert</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">all</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">mean</span> <span class="o">==</span> <span class="n">m_loaded</span><span class="o">.</span><span class="n">mean</span><span class="p">))</span>
</pre></div>
</div>
<p>As mentioned previously, buffers can be left out of the module’s <code class="docutils literal notranslate"><span class="pre">state_dict</span></code> by marking them as non-persistent:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="bp">self</span><span class="o">.</span><span class="n">register_buffer</span><span class="p">(</span><span class="s1">'unserialized_thing'</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">5</span><span class="p">),</span> <span class="n">persistent</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
</pre></div>
</div>
<p>Both persistent and non-persistent buffers are affected by model-wide device / dtype changes applied with
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.to" title="torch.nn.Module.to"><code class="xref py py-func docutils literal notranslate"><span class="pre">to()</span></code></a>:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Moves all module parameters and buffers to the specified device / dtype</span>
<span class="n">m</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">device</span><span class="o">=</span><span class="s1">'cuda'</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float64</span><span class="p">)</span>
</pre></div>
</div>
<p>Buffers of a module can be iterated over using <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.buffers" title="torch.nn.Module.buffers"><code class="xref py py-func docutils literal notranslate"><span class="pre">buffers()</span></code></a> or
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.named_buffers" title="torch.nn.Module.named_buffers"><code class="xref py py-func docutils literal notranslate"><span class="pre">named_buffers()</span></code></a>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">buffer</span> <span class="ow">in</span> <span class="n">m</span><span class="o">.</span><span class="n">named_buffers</span><span class="p">():</span>
<span class="nb">print</span><span class="p">(</span><span class="n">buffer</span><span class="p">)</span>
</pre></div>
</div>
<p>The following class demonstrates the various ways of registering parameters and buffers within a module:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">StatefulModule</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="c1"># Setting a nn.Parameter as an attribute of the module automatically registers the tensor</span>
<span class="c1"># as a parameter of the module.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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">2</span><span class="p">))</span>
<span class="c1"># Alternative string-based way to register a parameter.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span><span class="s1">'param2'</span><span class="p">,</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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">3</span><span class="p">)))</span>
<span class="c1"># Reserves the "param3" attribute as a parameter, preventing it from being set to anything</span>
<span class="c1"># except a parameter. "None" entries like this will not be present in the module's state_dict.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_parameter</span><span class="p">(</span><span class="s1">'param3'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="c1"># Registers a list of parameters.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param_list</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ParameterList</span><span class="p">([</span><span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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">2</span><span class="p">))</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">3</span><span class="p">)])</span>
<span class="c1"># Registers a dictionary of parameters.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">param_dict</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ParameterDict</span><span class="p">({</span>
<span class="s1">'foo'</span><span class="p">:</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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">3</span><span class="p">)),</span>
<span class="s1">'bar'</span><span class="p">:</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</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="p">})</span>
<span class="c1"># Registers a persistent buffer (one that appears in the module's state_dict).</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_buffer</span><span class="p">(</span><span class="s1">'buffer1'</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="n">persistent</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="c1"># Registers a non-persistent buffer (one that does not appear in the module's state_dict).</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_buffer</span><span class="p">(</span><span class="s1">'buffer2'</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">5</span><span class="p">),</span> <span class="n">persistent</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
<span class="c1"># Reserves the "buffer3" attribute as a buffer, preventing it from being set to anything</span>
<span class="c1"># except a buffer. "None" entries like this will not be present in the module's state_dict.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">register_buffer</span><span class="p">(</span><span class="s1">'buffer3'</span><span class="p">,</span> <span class="kc">None</span><span class="p">)</span>
<span class="c1"># Adding a submodule registers its parameters as parameters of the module.</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">StatefulModule</span><span class="p">()</span>
<span class="c1"># Save and load state_dict.</span>
<span class="n">torch</span><span class="o">.</span><span class="n">save</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">state_dict</span><span class="p">(),</span> <span class="s1">'state.pt'</span><span class="p">)</span>
<span class="n">m_loaded</span> <span class="o">=</span> <span class="n">StatefulModule</span><span class="p">()</span>
<span class="n">m_loaded</span><span class="o">.</span><span class="n">load_state_dict</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="s1">'state.pt'</span><span class="p">))</span>
<span class="c1"># Note that non-persistent buffer "buffer2" and reserved attributes "param3" and "buffer3" do</span>
<span class="c1"># not appear in the state_dict.</span>
<span class="nb">print</span><span class="p">(</span><span class="n">m_loaded</span><span class="o">.</span><span class="n">state_dict</span><span class="p">())</span>
<span class="p">:</span> <span class="n">OrderedDict</span><span class="p">([(</span><span class="s1">'param1'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.0322</span><span class="p">,</span> <span class="mf">0.9066</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param2'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.4472</span><span class="p">,</span> <span class="mf">0.1409</span><span class="p">,</span> <span class="mf">0.4852</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'buffer1'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">0.6949</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.1944</span><span class="p">,</span> <span class="mf">1.2911</span><span class="p">,</span> <span class="o">-</span><span class="mf">2.1044</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param_list.0'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">0.4202</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.1953</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param_list.1'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span> <span class="mf">1.5299</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.8747</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param_list.2'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">1.6289</span><span class="p">,</span> <span class="mf">1.4898</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param_dict.bar'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.6434</span><span class="p">,</span> <span class="mf">1.5187</span><span class="p">,</span> <span class="mf">0.0346</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.4077</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'param_dict.foo'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.0845</span><span class="p">,</span> <span class="o">-</span><span class="mf">1.4324</span><span class="p">,</span> <span class="mf">0.7022</span><span class="p">])),</span>
<span class="p">(</span><span class="s1">'linear.weight'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([[</span><span class="o">-</span><span class="mf">0.3915</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.6176</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">0.6062</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.5992</span><span class="p">],</span>
<span class="p">[</span> <span class="mf">0.4452</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.2843</span><span class="p">]])),</span>
<span class="p">(</span><span class="s1">'linear.bias'</span><span class="p">,</span> <span class="n">tensor</span><span class="p">([</span><span class="o">-</span><span class="mf">0.3710</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.0795</span><span class="p">,</span> <span class="o">-</span><span class="mf">0.3947</span><span class="p">]))])</span>
</pre></div>
</div>
<p>For more information, check out:</p>
<ul class="simple">
<li><p>Saving and loading: <a class="reference external" href="https://pytorch.org/tutorials/beginner/saving_loading_models.html">https://pytorch.org/tutorials/beginner/saving_loading_models.html</a></p></li>
<li><p>Serialization semantics: <a class="reference external" href="https://pytorch.org/docs/master/notes/serialization.html">https://pytorch.org/docs/master/notes/serialization.html</a></p></li>
<li><p>What is a state dict? <a class="reference external" href="https://pytorch.org/tutorials/recipes/recipes/what_is_state_dict.html">https://pytorch.org/tutorials/recipes/recipes/what_is_state_dict.html</a></p></li>
</ul>
</div>
<div class="section" id="module-initialization">
<h2><a class="toc-backref" href="#id8">Module Initialization</a><a class="headerlink" href="#module-initialization" title="Permalink to this headline">¶</a></h2>
<p>By default, parameters and floating-point buffers for modules provided by <code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.nn</span></code> are initialized during
module instantiation as 32-bit floating point values on the CPU using an initialization scheme determined to
perform well historically for the module type. For certain use cases, it may be desired to initialize with a different
dtype, device (e.g. GPU), or initialization technique.</p>
<p>Examples:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># Initialize module directly onto GPU.</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s1">'cuda'</span><span class="p">)</span>
<span class="c1"># Initialize module with 16-bit floating point parameters.</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">5</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">half</span><span class="p">)</span>
<span class="c1"># Skip default parameter initialization and perform custom (e.g. orthogonal) initialization.</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">skip_init</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">,</span> <span class="mi">5</span><span class="p">,</span> <span class="mi">3</span><span class="p">)</span>
<span class="n">nn</span><span class="o">.</span><span class="n">init</span><span class="o">.</span><span class="n">orthogonal_</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
</pre></div>
</div>
<p>Note that the device and dtype options demonstrated above also apply to any floating-point buffers registered
for the module:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">m</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BatchNorm2d</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">half</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="n">m</span><span class="o">.</span><span class="n">running_mean</span><span class="p">)</span>
<span class="p">:</span> <span class="n">tensor</span><span class="p">([</span><span class="mf">0.</span><span class="p">,</span> <span class="mf">0.</span><span class="p">,</span> <span class="mf">0.</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float16</span><span class="p">)</span>
</pre></div>
</div>
<p>While module writers can use any device or dtype to initialize parameters in their custom modules, good practice is
to use <code class="docutils literal notranslate"><span class="pre">dtype=torch.float</span></code> and <code class="docutils literal notranslate"><span class="pre">device='cpu'</span></code> by default as well. Optionally, you can provide full flexibility
in these areas for your custom module by conforming to the convention demonstrated above that all
<code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.nn</span></code> modules follow:</p>
<ul class="simple">
<li><p>Provide a <code class="docutils literal notranslate"><span class="pre">device</span></code> constructor kwarg that applies to any parameters / buffers registered by the module.</p></li>
<li><p>Provide a <code class="docutils literal notranslate"><span class="pre">dtype</span></code> constructor kwarg that applies to any parameters / floating-point buffers registered by
the module.</p></li>
<li><p>Only use initialization functions (i.e. functions from <code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.nn.init</span></code>) on parameters and buffers within the
module’s constructor. Note that this is only required to use <a class="reference internal" href="../generated/torch.nn.utils.skip_init.html#torch.nn.utils.skip_init" title="torch.nn.utils.skip_init"><code class="xref py py-func docutils literal notranslate"><span class="pre">skip_init()</span></code></a>; see
<a class="reference external" href="https://pytorch.org/tutorials/prototype/skip_param_init.html#updating-modules-to-support-skipping-initialization">this page</a> for an explanation.</p></li>
</ul>
<p>For more information, check out:</p>
<ul class="simple">
<li><p>Skipping module parameter initialization: <a class="reference external" href="https://pytorch.org/tutorials/prototype/skip_param_init.html">https://pytorch.org/tutorials/prototype/skip_param_init.html</a></p></li>
</ul>
</div>
<div class="section" id="module-hooks">
<h2><a class="toc-backref" href="#id9">Module Hooks</a><a class="headerlink" href="#module-hooks" title="Permalink to this headline">¶</a></h2>
<p>In <a class="reference internal" href="#neural-network-training-with-modules"><span class="std std-ref">Neural Network Training with Modules</span></a>, we demonstrated the training process for a module, which iteratively
performs forward and backward passes, updating module parameters each iteration. For more control
over this process, PyTorch provides “hooks” that can perform arbitrary computation during a forward or backward
pass, even modifying how the pass is done if desired. Some useful examples for this functionality include
debugging, visualizing activations, examining gradients in-depth, etc. Hooks can be added to modules
you haven’t written yourself, meaning this functionality can be applied to third-party or PyTorch-provided modules.</p>
<p>PyTorch provides two types of hooks for modules:</p>
<ul class="simple">
<li><p><strong>Forward hooks</strong> are called during the forward pass. They can be installed for a given module with
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.register_forward_pre_hook" title="torch.nn.Module.register_forward_pre_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_forward_pre_hook()</span></code></a> and <a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.register_forward_hook" title="torch.nn.Module.register_forward_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_forward_hook()</span></code></a>.
These hooks will be called respectively just before the forward function is called and just after it is called.
Alternatively, these hooks can be installed globally for all modules with the analagous
<a class="reference internal" href="../generated/torch.nn.modules.module.register_module_forward_pre_hook.html#torch.nn.modules.module.register_module_forward_pre_hook" title="torch.nn.modules.module.register_module_forward_pre_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_module_forward_pre_hook()</span></code></a> and
<a class="reference internal" href="../generated/torch.nn.modules.module.register_module_forward_hook.html#torch.nn.modules.module.register_module_forward_hook" title="torch.nn.modules.module.register_module_forward_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_module_forward_hook()</span></code></a> functions.</p></li>
<li><p><strong>Backward hooks</strong> are called during the backward pass. They can be installed with
<a class="reference internal" href="../generated/torch.nn.Module.html#torch.nn.Module.register_full_backward_hook" title="torch.nn.Module.register_full_backward_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_full_backward_hook()</span></code></a>. These hooks will be called when the backward for this
Module has been computed and will allow the user to access the gradients for both the inputs and outputs.
Alternatively, they can be installed globally for all modules with
<a class="reference internal" href="../generated/torch.nn.modules.module.register_module_full_backward_hook.html#torch.nn.modules.module.register_module_full_backward_hook" title="torch.nn.modules.module.register_module_full_backward_hook"><code class="xref py py-func docutils literal notranslate"><span class="pre">register_module_full_backward_hook()</span></code></a>.</p></li>
</ul>
<p>All hooks allow the user to return an updated value that will be used throughout the remaining computation.
Thus, these hooks can be used to either execute arbitrary code along the regular module forward/backward or
modify some inputs/outputs without having to change the module’s <code class="docutils literal notranslate"><span class="pre">forward()</span></code> function.</p>
<p>Below is an example demonstrating usage of forward and backward hooks:</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">manual_seed</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward_pre_hook</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">inputs</span><span class="p">):</span>
<span class="c1"># Allows for examination and modification of the input before the forward pass.</span>
<span class="c1"># Note that inputs are always wrapped in a tuple.</span>
<span class="nb">input</span> <span class="o">=</span> <span class="n">inputs</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="k">return</span> <span class="nb">input</span> <span class="o">+</span> <span class="mf">1.</span>
<span class="k">def</span> <span class="nf">forward_hook</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">inputs</span><span class="p">,</span> <span class="n">output</span><span class="p">):</span>
<span class="c1"># Allows for examination of inputs / outputs and modification of the outputs</span>
<span class="c1"># after the forward pass. Note that inputs are always wrapped in a tuple while outputs</span>
<span class="c1"># are passed as-is.</span>
<span class="c1"># Residual computation a la ResNet.</span>
<span class="k">return</span> <span class="n">output</span> <span class="o">+</span> <span class="n">inputs</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="k">def</span> <span class="nf">backward_hook</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">grad_inputs</span><span class="p">,</span> <span class="n">grad_outputs</span><span class="p">):</span>
<span class="c1"># Allows for examination of grad_inputs / grad_outputs and modification of</span>
<span class="c1"># grad_inputs used in the rest of the backwards pass. Note that grad_inputs and</span>
<span class="c1"># grad_outputs are always wrapped in tuples.</span>
<span class="n">new_grad_inputs</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">ones_like</span><span class="p">(</span><span class="n">gi</span><span class="p">)</span> <span class="o">*</span> <span class="mf">42.</span> <span class="k">for</span> <span class="n">gi</span> <span class="ow">in</span> <span class="n">grad_inputs</span><span class="p">]</span>
<span class="k">return</span> <span class="n">new_grad_inputs</span>
<span class="c1"># Create sample module & input.</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">3</span><span class="p">,</span> <span class="mi">3</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">randn</span><span class="p">(</span><span class="mi">2</span><span class="p">,</span> <span class="mi">3</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>