forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathoptim.html
1328 lines (1126 loc) · 97.8 KB
/
optim.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="generator" content="Docutils 0.18.1: http://docutils.sourceforge.net/" />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>torch.optim — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/optim.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/pygments.css" type="text/css" />
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[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/sphinx-dropdown.css" type="text/css" />
<link rel="stylesheet" href="_static/panels-bootstrap.min.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="torch.optim.Optimizer.add_param_group" href="generated/torch.optim.Optimizer.add_param_group.html" />
<link rel="prev" title="POE0004:operator-supported-in-newer-opset-version" href="generated/onnx_diagnostics_rules/POE0004%3Aoperator-supported-in-newer-opset-version.html" />
<!--
Search engines should not index the master version of documentation.
Stable documentation are built without release == 'master'.
-->
<meta name="robots" content="noindex">
<!-- 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/torcharrow">
<span class="dropdown-title">torcharrow</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/data">
<span class="dropdown-title">TorchData</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/torchrec">
<span class="dropdown-title">TorchRec</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/torchx/">
<span class="dropdown-title">TorchX</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/foundation">
<span class="dropdown-title">PyTorch Foundation</span>
<p>Learn about the PyTorch foundation</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/community-stories">
<span class="dropdown-title">Community Stories</span>
<p>Learn how our community solves real, everyday machine learning problems with PyTorch.</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://pytorch.org/events">
<span class="dropdown-title">Events</span>
<p>Find events, webinars, and podcasts</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'>master (2.1.0a0+gitbe0b12e ) ▼</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>
<div>
<a style="color:#F05732" href="https://pytorch.org/docs/stable/optim.html">
You are viewing unstable developer preview docs.
Click here to view docs for latest stable release.
</a>
</div>
<p class="caption" role="heading"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/build_ci_governance.html">PyTorch Governance | Build + CI</a></li>
<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/design.html">PyTorch Design Philosophy</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance | Mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Maintainers</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Developer Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">CUDA Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.func.html">Extending torch.func with autograd.Function</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/mps.html">MPS backend</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/numerical_accuracy.html">Numerical accuracy</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">torch.compile</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="compile/index.html">torch.compile</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/get-started.html">Getting Started</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/troubleshooting.html">PyTorch 2.0 Troubleshooting</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/technical-overview.html">Technical Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/guards-overview.html">Guards Overview</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/custom-backends.html">Custom Backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="compile/deep-dive.html">TorchDynamo Deeper Dive</a></li>
<li class="toctree-l1"><a class="reference internal" href="ir.html">IRs</a></li>
</ul>
<p class="caption" role="heading"><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>
<li class="toctree-l1"><a class="reference internal" href="deploy.html">torch::deploy</a></li>
</ul>
<p class="caption" role="heading"><span class="caption-text">Python API</span></p>
<ul class="current">
<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="amp.html">torch.amp</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="library.html">torch.library</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="mps.html">torch.mps</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="fsdp.html">torch.distributed.fsdp</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="distributed.tensor.parallel.html">torch.distributed.tensor.parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.checkpoint.html">torch.distributed.checkpoint</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="_dynamo.html">torch._dynamo</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="func.html">torch.func</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="monitor.html">torch.monitor</a></li>
<li class="toctree-l1"><a class="reference internal" href="signal.html">torch.signal</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="onnx_diagnostics.html">torch.onnx diagnostics</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">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="masked.html">torch.masked</a></li>
<li class="toctree-l1"><a class="reference internal" href="nested.html">torch.nested</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="jit_utils.html">torch.utils.jit</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_mod.html">torch.__config__</a></li>
</ul>
<p class="caption" role="heading"><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/data">TorchData</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/torchrec">TorchRec</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="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/xla/">PyTorch on XLA Devices</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>torch.optim</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/optim.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">
<section id="module-torch.optim">
<span id="torch-optim"></span><h1>torch.optim<a class="headerlink" href="#module-torch.optim" title="Permalink to this heading">¶</a></h1>
<p><a class="reference internal" href="#module-torch.optim" title="torch.optim"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.optim</span></code></a> is a package implementing various optimization algorithms.
Most commonly used methods are already supported, and the interface is general
enough, so that more sophisticated ones can also be easily integrated in the
future.</p>
<section id="how-to-use-an-optimizer">
<h2>How to use an optimizer<a class="headerlink" href="#how-to-use-an-optimizer" title="Permalink to this heading">¶</a></h2>
<p>To use <a class="reference internal" href="#module-torch.optim" title="torch.optim"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.optim</span></code></a> you have to construct an optimizer object that will hold
the current state and will update the parameters based on the computed gradients.</p>
<section id="constructing-it">
<h3>Constructing it<a class="headerlink" href="#constructing-it" title="Permalink to this heading">¶</a></h3>
<p>To construct an <a class="reference internal" href="#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Optimizer</span></code></a> you have to give it an iterable containing the
parameters (all should be <code class="xref py py-class docutils literal notranslate"><span class="pre">Variable</span></code> s) to optimize. Then,
you can specify optimizer-specific options such as the learning rate, weight decay, etc.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</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="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">Adam</span><span class="p">([</span><span class="n">var1</span><span class="p">,</span> <span class="n">var2</span><span class="p">],</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.0001</span><span class="p">)</span>
</pre></div>
</div>
</section>
<section id="per-parameter-options">
<h3>Per-parameter options<a class="headerlink" href="#per-parameter-options" title="Permalink to this heading">¶</a></h3>
<p><a class="reference internal" href="#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Optimizer</span></code></a> s also support specifying per-parameter options. To do this, instead
of passing an iterable of <code class="xref py py-class docutils literal notranslate"><span class="pre">Variable</span></code> s, pass in an iterable of
<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> s. Each of them will define a separate parameter group, and should contain
a <code class="docutils literal notranslate"><span class="pre">params</span></code> key, containing a list of parameters belonging to it. Other keys
should match the keyword arguments accepted by the optimizers, and will be used
as optimization options for this group.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>You can still pass options as keyword arguments. They will be used as
defaults, in the groups that didn’t override them. This is useful when you
only want to vary a single option, while keeping all others consistent
between parameter groups.</p>
</div>
<p>For example, this is very useful when one wants to specify per-layer learning rates:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">([</span>
<span class="p">{</span><span class="s1">'params'</span><span class="p">:</span> <span class="n">model</span><span class="o">.</span><span class="n">base</span><span class="o">.</span><span class="n">parameters</span><span class="p">()},</span>
<span class="p">{</span><span class="s1">'params'</span><span class="p">:</span> <span class="n">model</span><span class="o">.</span><span class="n">classifier</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="s1">'lr'</span><span class="p">:</span> <span class="mf">1e-3</span><span class="p">}</span>
<span class="p">],</span> <span class="n">lr</span><span class="o">=</span><span class="mf">1e-2</span><span class="p">,</span> <span class="n">momentum</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
</pre></div>
</div>
<p>This means that <code class="docutils literal notranslate"><span class="pre">model.base</span></code>’s parameters will use the default learning rate of <code class="docutils literal notranslate"><span class="pre">1e-2</span></code>,
<code class="docutils literal notranslate"><span class="pre">model.classifier</span></code>’s parameters will use a learning rate of <code class="docutils literal notranslate"><span class="pre">1e-3</span></code>, and a momentum of
<code class="docutils literal notranslate"><span class="pre">0.9</span></code> will be used for all parameters.</p>
</section>
<section id="taking-an-optimization-step">
<h3>Taking an optimization step<a class="headerlink" href="#taking-an-optimization-step" title="Permalink to this heading">¶</a></h3>
<p>All optimizers implement a <a class="reference internal" href="generated/torch.optim.Optimizer.step.html#torch.optim.Optimizer.step" title="torch.optim.Optimizer.step"><code class="xref py py-func docutils literal notranslate"><span class="pre">step()</span></code></a> method, that updates the
parameters. It can be used in two ways:</p>
<section id="optimizer-step">
<h4><code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code><a class="headerlink" href="#optimizer-step" title="Permalink to this heading">¶</a></h4>
<p>This is a simplified version supported by most optimizers. The function can be
called once the gradients are computed using e.g.
<code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code>.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">dataset</span><span class="p">:</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</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">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</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>
</pre></div>
</div>
</section>
<section id="optimizer-step-closure">
<h4><code class="docutils literal notranslate"><span class="pre">optimizer.step(closure)</span></code><a class="headerlink" href="#optimizer-step-closure" title="Permalink to this heading">¶</a></h4>
<p>Some optimization algorithms such as Conjugate Gradient and LBFGS need to
reevaluate the function multiple times, so you have to pass in a closure that
allows them to recompute your model. The closure should clear the gradients,
compute the loss, and return it.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">dataset</span><span class="p">:</span>
<span class="k">def</span> <span class="nf">closure</span><span class="p">():</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</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">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</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="k">return</span> <span class="n">loss</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">closure</span><span class="p">)</span>
</pre></div>
</div>
</section>
</section>
</section>
<section id="base-class">
<span id="optimizer-algorithms"></span><h2>Base class<a class="headerlink" href="#base-class" title="Permalink to this heading">¶</a></h2>
<dl class="py class">
<dt class="sig sig-object py" id="torch.optim.Optimizer">
<em class="property"><span class="pre">class</span><span class="w"> </span></em><span class="sig-prename descclassname"><span class="pre">torch.optim.</span></span><span class="sig-name descname"><span class="pre">Optimizer</span></span><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">params</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">defaults</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/optim/optimizer.html#Optimizer"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.optim.Optimizer" title="Permalink to this definition">¶</a></dt>
<dd><p>Base class for all optimizers.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Parameters need to be specified as collections that have a deterministic
ordering that is consistent between runs. Examples of objects that don’t
satisfy those properties are sets and iterators over values of dictionaries.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters<span class="colon">:</span></dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>params</strong> (<em>iterable</em>) – an iterable of <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> s or
<a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a> s. Specifies what Tensors should be optimized.</p></li>
<li><p><strong>defaults</strong> – (dict): a dict containing default values of optimization
options (used when a parameter group doesn’t specify them).</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<table class="autosummary longtable docutils align-default">
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.Optimizer.add_param_group.html#torch.optim.Optimizer.add_param_group" title="torch.optim.Optimizer.add_param_group"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Optimizer.add_param_group</span></code></a></p></td>
<td><p>Add a param group to the <a class="reference internal" href="#torch.optim.Optimizer" title="torch.optim.Optimizer"><code class="xref py py-class docutils literal notranslate"><span class="pre">Optimizer</span></code></a> s <cite>param_groups</cite>.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.Optimizer.load_state_dict.html#torch.optim.Optimizer.load_state_dict" title="torch.optim.Optimizer.load_state_dict"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Optimizer.load_state_dict</span></code></a></p></td>
<td><p>Loads the optimizer state.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.Optimizer.state_dict.html#torch.optim.Optimizer.state_dict" title="torch.optim.Optimizer.state_dict"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Optimizer.state_dict</span></code></a></p></td>
<td><p>Returns the state of the optimizer as a <a class="reference external" href="https://docs.python.org/3/library/stdtypes.html#dict" title="(in Python v3.11)"><code class="xref py py-class docutils literal notranslate"><span class="pre">dict</span></code></a>.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.Optimizer.step.html#torch.optim.Optimizer.step" title="torch.optim.Optimizer.step"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Optimizer.step</span></code></a></p></td>
<td><p>Performs a single optimization step (parameter update).</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.Optimizer.zero_grad.html#torch.optim.Optimizer.zero_grad" title="torch.optim.Optimizer.zero_grad"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Optimizer.zero_grad</span></code></a></p></td>
<td><p>Resets the gradients of all optimized <a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.Tensor</span></code></a> s.</p></td>
</tr>
</tbody>
</table>
</section>
<section id="algorithms">
<h2>Algorithms<a class="headerlink" href="#algorithms" title="Permalink to this heading">¶</a></h2>
<table class="autosummary longtable docutils align-default">
<tbody>
<tr class="row-odd"><td><p><p id="torch.optim.Adadelta"/><a class="reference internal" href="generated/torch.optim.Adadelta.html#torch.optim.Adadelta" title="torch.optim.Adadelta"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Adadelta</span></code></a></p></td>
<td><p>Implements Adadelta algorithm.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.Adagrad"/><a class="reference internal" href="generated/torch.optim.Adagrad.html#torch.optim.Adagrad" title="torch.optim.Adagrad"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Adagrad</span></code></a></p></td>
<td><p>Implements Adagrad algorithm.</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.Adam"/><a class="reference internal" href="generated/torch.optim.Adam.html#torch.optim.Adam" title="torch.optim.Adam"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Adam</span></code></a></p></td>
<td><p>Implements Adam algorithm.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.AdamW"/><a class="reference internal" href="generated/torch.optim.AdamW.html#torch.optim.AdamW" title="torch.optim.AdamW"><code class="xref py py-obj docutils literal notranslate"><span class="pre">AdamW</span></code></a></p></td>
<td><p>Implements AdamW algorithm.</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.SparseAdam"/><a class="reference internal" href="generated/torch.optim.SparseAdam.html#torch.optim.SparseAdam" title="torch.optim.SparseAdam"><code class="xref py py-obj docutils literal notranslate"><span class="pre">SparseAdam</span></code></a></p></td>
<td><p>Implements lazy version of Adam algorithm suitable for sparse tensors.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.Adamax"/><a class="reference internal" href="generated/torch.optim.Adamax.html#torch.optim.Adamax" title="torch.optim.Adamax"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Adamax</span></code></a></p></td>
<td><p>Implements Adamax algorithm (a variant of Adam based on infinity norm).</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.ASGD"/><a class="reference internal" href="generated/torch.optim.ASGD.html#torch.optim.ASGD" title="torch.optim.ASGD"><code class="xref py py-obj docutils literal notranslate"><span class="pre">ASGD</span></code></a></p></td>
<td><p>Implements Averaged Stochastic Gradient Descent.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.LBFGS"/><a class="reference internal" href="generated/torch.optim.LBFGS.html#torch.optim.LBFGS" title="torch.optim.LBFGS"><code class="xref py py-obj docutils literal notranslate"><span class="pre">LBFGS</span></code></a></p></td>
<td><p>Implements L-BFGS algorithm, heavily inspired by <a class="reference external" href="https://www.cs.ubc.ca/~schmidtm/Software/minFunc.html">minFunc</a>.</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.NAdam"/><a class="reference internal" href="generated/torch.optim.NAdam.html#torch.optim.NAdam" title="torch.optim.NAdam"><code class="xref py py-obj docutils literal notranslate"><span class="pre">NAdam</span></code></a></p></td>
<td><p>Implements NAdam algorithm.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.RAdam"/><a class="reference internal" href="generated/torch.optim.RAdam.html#torch.optim.RAdam" title="torch.optim.RAdam"><code class="xref py py-obj docutils literal notranslate"><span class="pre">RAdam</span></code></a></p></td>
<td><p>Implements RAdam algorithm.</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.RMSprop"/><a class="reference internal" href="generated/torch.optim.RMSprop.html#torch.optim.RMSprop" title="torch.optim.RMSprop"><code class="xref py py-obj docutils literal notranslate"><span class="pre">RMSprop</span></code></a></p></td>
<td><p>Implements RMSprop algorithm.</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.Rprop"/><a class="reference internal" href="generated/torch.optim.Rprop.html#torch.optim.Rprop" title="torch.optim.Rprop"><code class="xref py py-obj docutils literal notranslate"><span class="pre">Rprop</span></code></a></p></td>
<td><p>Implements the resilient backpropagation algorithm.</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.SGD"/><a class="reference internal" href="generated/torch.optim.SGD.html#torch.optim.SGD" title="torch.optim.SGD"><code class="xref py py-obj docutils literal notranslate"><span class="pre">SGD</span></code></a></p></td>
<td><p><p></p>
</p></td>
</tr>
</tbody>
</table>
<p>Many of our algorithms have various implementations optimized for performance,
readability and/or generality, so we attempt to default to the generally fastest
implementation for the current device if no particular implementation has been
specified by the user.</p>
<p>We have 3 major categories of implementations: for-loop, foreach (multi-tensor), and
fused. The most straightforward implementations are for-loops over the parameters with
big chunks of computation. For-looping is usually slower than our foreach
implementations, which combine parameters into a multi-tensor and run the big chunks
of computation all at once, thereby saving many sequential kernel calls. A few of our
optimizers have even faster fused implementations, which fuse the big chunks of
computation into one kernel. We can think of foreach implementations as fusing
horizontally and fused implementations as fusing vertically on top of that.</p>
<p>In general, the performance ordering of the 3 implementations is fused > foreach > for-loop.
So when applicable, we default to foreach over for-loop. Applicable means the foreach
implementation is available, the user has not specified any implementation-specific kwargs
(e.g., fused, foreach, differentiable), and all tensors are native and on CUDA. Note that
while fused should be even faster than foreach, the implementations are newer and we would
like to give them more bake-in time before flipping the switch everywhere. You are welcome
to try them out though!</p>
<p>Below is a table showing the available and default implementations of each algorithm:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 25%" />
<col style="width: 25%" />
<col style="width: 25%" />
<col style="width: 25%" />
</colgroup>
<thead>
<tr class="row-odd"><th class="head"><p>Algorithm</p></th>
<th class="head"><p>Default</p></th>
<th class="head"><p>Has foreach?</p></th>
<th class="head"><p>Has fused?</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p><p id="torch.optim.Adadelta"/><a class="reference internal" href="generated/torch.optim.Adadelta.html#torch.optim.Adadelta" title="torch.optim.Adadelta"><code class="xref py py-class docutils literal notranslate"><span class="pre">Adadelta</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.Adagrad"/><a class="reference internal" href="generated/torch.optim.Adagrad.html#torch.optim.Adagrad" title="torch.optim.Adagrad"><code class="xref py py-class docutils literal notranslate"><span class="pre">Adagrad</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.Adam"/><a class="reference internal" href="generated/torch.optim.Adam.html#torch.optim.Adam" title="torch.optim.Adam"><code class="xref py py-class docutils literal notranslate"><span class="pre">Adam</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>yes</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.AdamW"/><a class="reference internal" href="generated/torch.optim.AdamW.html#torch.optim.AdamW" title="torch.optim.AdamW"><code class="xref py py-class docutils literal notranslate"><span class="pre">AdamW</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>yes</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.SparseAdam"/><a class="reference internal" href="generated/torch.optim.SparseAdam.html#torch.optim.SparseAdam" title="torch.optim.SparseAdam"><code class="xref py py-class docutils literal notranslate"><span class="pre">SparseAdam</span></code></a></p></td>
<td><p>for-loop</p></td>
<td><p>no</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.Adamax"/><a class="reference internal" href="generated/torch.optim.Adamax.html#torch.optim.Adamax" title="torch.optim.Adamax"><code class="xref py py-class docutils literal notranslate"><span class="pre">Adamax</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.ASGD"/><a class="reference internal" href="generated/torch.optim.ASGD.html#torch.optim.ASGD" title="torch.optim.ASGD"><code class="xref py py-class docutils literal notranslate"><span class="pre">ASGD</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.LBFGS"/><a class="reference internal" href="generated/torch.optim.LBFGS.html#torch.optim.LBFGS" title="torch.optim.LBFGS"><code class="xref py py-class docutils literal notranslate"><span class="pre">LBFGS</span></code></a></p></td>
<td><p>for-loop</p></td>
<td><p>no</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.NAdam"/><a class="reference internal" href="generated/torch.optim.NAdam.html#torch.optim.NAdam" title="torch.optim.NAdam"><code class="xref py py-class docutils literal notranslate"><span class="pre">NAdam</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.RAdam"/><a class="reference internal" href="generated/torch.optim.RAdam.html#torch.optim.RAdam" title="torch.optim.RAdam"><code class="xref py py-class docutils literal notranslate"><span class="pre">RAdam</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.RMSprop"/><a class="reference internal" href="generated/torch.optim.RMSprop.html#torch.optim.RMSprop" title="torch.optim.RMSprop"><code class="xref py py-class docutils literal notranslate"><span class="pre">RMSprop</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-odd"><td><p><p id="torch.optim.Rprop"/><a class="reference internal" href="generated/torch.optim.Rprop.html#torch.optim.Rprop" title="torch.optim.Rprop"><code class="xref py py-class docutils literal notranslate"><span class="pre">Rprop</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
<tr class="row-even"><td><p><p id="torch.optim.SGD"/><a class="reference internal" href="generated/torch.optim.SGD.html#torch.optim.SGD" title="torch.optim.SGD"><code class="xref py py-class docutils literal notranslate"><span class="pre">SGD</span></code></a></p></td>
<td><p>foreach</p></td>
<td><p>yes</p></td>
<td><p>no</p></td>
</tr>
</tbody>
</table>
</section>
<section id="how-to-adjust-learning-rate">
<h2>How to adjust learning rate<a class="headerlink" href="#how-to-adjust-learning-rate" title="Permalink to this heading">¶</a></h2>
<p><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.optim.lr_scheduler</span></code> provides several methods to adjust the learning
rate based on the number of epochs. <a class="reference internal" href="generated/torch.optim.lr_scheduler.ReduceLROnPlateau.html#torch.optim.lr_scheduler.ReduceLROnPlateau" title="torch.optim.lr_scheduler.ReduceLROnPlateau"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.lr_scheduler.ReduceLROnPlateau</span></code></a>
allows dynamic learning rate reducing based on some validation measurements.</p>
<p>Learning rate scheduling should be applied after optimizer’s update; e.g., you
should write your code this way:</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</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="n">scheduler</span> <span class="o">=</span> <span class="n">ExponentialLR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">gamma</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">20</span><span class="p">):</span>
<span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">dataset</span><span class="p">:</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</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">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</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="n">scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>Most learning rate schedulers can be called back-to-back (also referred to as
chaining schedulers). The result is that each scheduler is applied one after the
other on the learning rate obtained by the one preceding it.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">optimizer</span> <span class="o">=</span> <span class="n">optim</span><span class="o">.</span><span class="n">SGD</span><span class="p">(</span><span class="n">model</span><span class="o">.</span><span class="n">parameters</span><span class="p">(),</span> <span class="n">lr</span><span class="o">=</span><span class="mf">0.01</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="n">scheduler1</span> <span class="o">=</span> <span class="n">ExponentialLR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">gamma</span><span class="o">=</span><span class="mf">0.9</span><span class="p">)</span>
<span class="n">scheduler2</span> <span class="o">=</span> <span class="n">MultiStepLR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">milestones</span><span class="o">=</span><span class="p">[</span><span class="mi">30</span><span class="p">,</span><span class="mi">80</span><span class="p">],</span> <span class="n">gamma</span><span class="o">=</span><span class="mf">0.1</span><span class="p">)</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">20</span><span class="p">):</span>
<span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">dataset</span><span class="p">:</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">model</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">loss_fn</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">target</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="n">scheduler1</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="n">scheduler2</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>In many places in the documentation, we will use the following template to refer to schedulers
algorithms.</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">scheduler</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">100</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="n">train</span><span class="p">(</span><span class="o">...</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="n">validate</span><span class="p">(</span><span class="o">...</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="n">scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Prior to PyTorch 1.1.0, the learning rate scheduler was expected to be called before
the optimizer’s update; 1.1.0 changed this behavior in a BC-breaking way. If you use
the learning rate scheduler (calling <code class="docutils literal notranslate"><span class="pre">scheduler.step()</span></code>) before the optimizer’s update
(calling <code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code>), this will skip the first value of the learning rate schedule.
If you are unable to reproduce results after upgrading to PyTorch 1.1.0, please check
if you are calling <code class="docutils literal notranslate"><span class="pre">scheduler.step()</span></code> at the wrong time.</p>
</div>
<table class="autosummary longtable docutils align-default">
<tbody>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.LambdaLR.html#torch.optim.lr_scheduler.LambdaLR" title="torch.optim.lr_scheduler.LambdaLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.LambdaLR</span></code></a></p></td>
<td><p>Sets the learning rate of each parameter group to the initial lr times a given function.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.MultiplicativeLR.html#torch.optim.lr_scheduler.MultiplicativeLR" title="torch.optim.lr_scheduler.MultiplicativeLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.MultiplicativeLR</span></code></a></p></td>
<td><p>Multiply the learning rate of each parameter group by the factor given in the specified function.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.StepLR.html#torch.optim.lr_scheduler.StepLR" title="torch.optim.lr_scheduler.StepLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.StepLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group by gamma every step_size epochs.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.MultiStepLR.html#torch.optim.lr_scheduler.MultiStepLR" title="torch.optim.lr_scheduler.MultiStepLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.MultiStepLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group by gamma once the number of epoch reaches one of the milestones.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.ConstantLR.html#torch.optim.lr_scheduler.ConstantLR" title="torch.optim.lr_scheduler.ConstantLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.ConstantLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group by a small constant factor until the number of epoch reaches a pre-defined milestone: total_iters.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.LinearLR.html#torch.optim.lr_scheduler.LinearLR" title="torch.optim.lr_scheduler.LinearLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.LinearLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group by linearly changing small multiplicative factor until the number of epoch reaches a pre-defined milestone: total_iters.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.ExponentialLR.html#torch.optim.lr_scheduler.ExponentialLR" title="torch.optim.lr_scheduler.ExponentialLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.ExponentialLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group by gamma every epoch.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.PolynomialLR.html#torch.optim.lr_scheduler.PolynomialLR" title="torch.optim.lr_scheduler.PolynomialLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.PolynomialLR</span></code></a></p></td>
<td><p>Decays the learning rate of each parameter group using a polynomial function in the given total_iters.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.CosineAnnealingLR.html#torch.optim.lr_scheduler.CosineAnnealingLR" title="torch.optim.lr_scheduler.CosineAnnealingLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.CosineAnnealingLR</span></code></a></p></td>
<td><p>Set the learning rate of each parameter group using a cosine annealing schedule, where <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>η</mi><mrow><mi>m</mi><mi>a</mi><mi>x</mi></mrow></msub></mrow><annotation encoding="application/x-tex">\eta_{max}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">η</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">ma</span><span class="mord mathnormal mtight">x</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span> is set to the initial lr and <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>T</mi><mrow><mi>c</mi><mi>u</mi><mi>r</mi></mrow></msub></mrow><annotation encoding="application/x-tex">T_{cur}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8333em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">T</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">c</span><span class="mord mathnormal mtight">u</span><span class="mord mathnormal mtight" style="margin-right:0.02778em;">r</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span> is the number of epochs since the last restart in SGDR:</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.ChainedScheduler.html#torch.optim.lr_scheduler.ChainedScheduler" title="torch.optim.lr_scheduler.ChainedScheduler"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.ChainedScheduler</span></code></a></p></td>
<td><p>Chains list of learning rate schedulers.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.SequentialLR.html#torch.optim.lr_scheduler.SequentialLR" title="torch.optim.lr_scheduler.SequentialLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.SequentialLR</span></code></a></p></td>
<td><p>Receives the list of schedulers that is expected to be called sequentially during optimization process and milestone points that provides exact intervals to reflect which scheduler is supposed to be called at a given epoch.</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.ReduceLROnPlateau.html#torch.optim.lr_scheduler.ReduceLROnPlateau" title="torch.optim.lr_scheduler.ReduceLROnPlateau"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.ReduceLROnPlateau</span></code></a></p></td>
<td><p>Reduce learning rate when a metric has stopped improving.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.CyclicLR.html#torch.optim.lr_scheduler.CyclicLR" title="torch.optim.lr_scheduler.CyclicLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.CyclicLR</span></code></a></p></td>
<td><p>Sets the learning rate of each parameter group according to cyclical learning rate policy (CLR).</p></td>
</tr>
<tr class="row-even"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.OneCycleLR.html#torch.optim.lr_scheduler.OneCycleLR" title="torch.optim.lr_scheduler.OneCycleLR"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.OneCycleLR</span></code></a></p></td>
<td><p>Sets the learning rate of each parameter group according to the 1cycle learning rate policy.</p></td>
</tr>
<tr class="row-odd"><td><p><a class="reference internal" href="generated/torch.optim.lr_scheduler.CosineAnnealingWarmRestarts.html#torch.optim.lr_scheduler.CosineAnnealingWarmRestarts" title="torch.optim.lr_scheduler.CosineAnnealingWarmRestarts"><code class="xref py py-obj docutils literal notranslate"><span class="pre">lr_scheduler.CosineAnnealingWarmRestarts</span></code></a></p></td>
<td><p>Set the learning rate of each parameter group using a cosine annealing schedule, where <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>η</mi><mrow><mi>m</mi><mi>a</mi><mi>x</mi></mrow></msub></mrow><annotation encoding="application/x-tex">\eta_{max}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.625em;vertical-align:-0.1944em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.03588em;">η</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">ma</span><span class="mord mathnormal mtight">x</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span> is set to the initial lr, <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>T</mi><mrow><mi>c</mi><mi>u</mi><mi>r</mi></mrow></msub></mrow><annotation encoding="application/x-tex">T_{cur}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8333em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">T</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.1514em;"><span style="top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">c</span><span class="mord mathnormal mtight">u</span><span class="mord mathnormal mtight" style="margin-right:0.02778em;">r</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span> is the number of epochs since the last restart and <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><msub><mi>T</mi><mi>i</mi></msub></mrow><annotation encoding="application/x-tex">T_{i}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.8333em;vertical-align:-0.15em;"></span><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">T</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.3117em;"><span style="top:-2.55em;margin-left:-0.1389em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mtight"><span class="mord mathnormal mtight">i</span></span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span> is the number of epochs between two warm restarts in SGDR:</p></td>
</tr>
</tbody>
</table>
</section>
<section id="stochastic-weight-averaging">
<h2>Stochastic Weight Averaging<a class="headerlink" href="#stochastic-weight-averaging" title="Permalink to this heading">¶</a></h2>
<p><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.optim.swa_utils</span></code> implements Stochastic Weight Averaging (SWA). In particular,
<code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.swa_utils.AveragedModel</span></code> class implements SWA models,
<code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.swa_utils.SWALR</span></code> implements the SWA learning rate scheduler and
<code class="xref py py-func docutils literal notranslate"><span class="pre">torch.optim.swa_utils.update_bn()</span></code> is a utility function used to update SWA batch
normalization statistics at the end of training.</p>
<p>SWA has been proposed in <a class="reference external" href="https://arxiv.org/abs/1803.05407">Averaging Weights Leads to Wider Optima and Better Generalization</a>.</p>
<section id="constructing-averaged-models">
<h3>Constructing averaged models<a class="headerlink" href="#constructing-averaged-models" title="Permalink to this heading">¶</a></h3>
<p><cite>AveragedModel</cite> class serves to compute the weights of the SWA model. You can create an
averaged model by running:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">swa_model</span> <span class="o">=</span> <span class="n">AveragedModel</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
</pre></div>
</div>
<p>Here the model <code class="docutils literal notranslate"><span class="pre">model</span></code> can be an arbitrary <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">torch.nn.Module</span></code></a> object. <code class="docutils literal notranslate"><span class="pre">swa_model</span></code>
will keep track of the running averages of the parameters of the <code class="docutils literal notranslate"><span class="pre">model</span></code>. To update these
averages, you can use the <code class="xref py py-func docutils literal notranslate"><span class="pre">update_parameters()</span></code> function:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">swa_model</span><span class="o">.</span><span class="n">update_parameters</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
</pre></div>
</div>
</section>
<section id="swa-learning-rate-schedules">
<h3>SWA learning rate schedules<a class="headerlink" href="#swa-learning-rate-schedules" title="Permalink to this heading">¶</a></h3>
<p>Typically, in SWA the learning rate is set to a high constant value. <code class="xref py py-class docutils literal notranslate"><span class="pre">SWALR</span></code> is a
learning rate scheduler that anneals the learning rate to a fixed value, and then keeps it
constant. For example, the following code creates a scheduler that linearly anneals the
learning rate from its initial value to 0.05 in 5 epochs within each parameter group:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">swa_scheduler</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">swa_utils</span><span class="o">.</span><span class="n">SWALR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> \
<span class="gp">>>> </span> <span class="n">anneal_strategy</span><span class="o">=</span><span class="s2">"linear"</span><span class="p">,</span> <span class="n">anneal_epochs</span><span class="o">=</span><span class="mi">5</span><span class="p">,</span> <span class="n">swa_lr</span><span class="o">=</span><span class="mf">0.05</span><span class="p">)</span>
</pre></div>
</div>
<p>You can also use cosine annealing to a fixed value instead of linear annealing by setting
<code class="docutils literal notranslate"><span class="pre">anneal_strategy="cos"</span></code>.</p>
</section>
<section id="taking-care-of-batch-normalization">
<h3>Taking care of batch normalization<a class="headerlink" href="#taking-care-of-batch-normalization" title="Permalink to this heading">¶</a></h3>
<p><code class="xref py py-func docutils literal notranslate"><span class="pre">update_bn()</span></code> is a utility function that allows to compute the batchnorm statistics for the SWA model
on a given dataloader <code class="docutils literal notranslate"><span class="pre">loader</span></code> at the end of training:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">swa_utils</span><span class="o">.</span><span class="n">update_bn</span><span class="p">(</span><span class="n">loader</span><span class="p">,</span> <span class="n">swa_model</span><span class="p">)</span>
</pre></div>
</div>
<p><code class="xref py py-func docutils literal notranslate"><span class="pre">update_bn()</span></code> applies the <code class="docutils literal notranslate"><span class="pre">swa_model</span></code> to every element in the dataloader and computes the activation
statistics for each batch normalization layer in the model.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><code class="xref py py-func docutils literal notranslate"><span class="pre">update_bn()</span></code> assumes that each batch in the dataloader <code class="docutils literal notranslate"><span class="pre">loader</span></code> is either a tensors or a list of
tensors where the first element is the tensor that the network <code class="docutils literal notranslate"><span class="pre">swa_model</span></code> should be applied to.
If your dataloader has a different structure, you can update the batch normalization statistics of the
<code class="docutils literal notranslate"><span class="pre">swa_model</span></code> by doing a forward pass with the <code class="docutils literal notranslate"><span class="pre">swa_model</span></code> on each element of the dataset.</p>
</div>
</section>
<section id="custom-averaging-strategies">
<h3>Custom averaging strategies<a class="headerlink" href="#custom-averaging-strategies" title="Permalink to this heading">¶</a></h3>
<p>By default, <code class="xref py py-class docutils literal notranslate"><span class="pre">torch.optim.swa_utils.AveragedModel</span></code> computes a running equal average of
the parameters that you provide, but you can also use custom averaging functions with the
<code class="docutils literal notranslate"><span class="pre">avg_fn</span></code> parameter. In the following example <code class="docutils literal notranslate"><span class="pre">ema_model</span></code> computes an exponential moving average.</p>
<p>Example:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">ema_avg</span> <span class="o">=</span> <span class="k">lambda</span> <span class="n">averaged_model_parameter</span><span class="p">,</span> <span class="n">model_parameter</span><span class="p">,</span> <span class="n">num_averaged</span><span class="p">:</span>\
<span class="gp">>>> </span> <span class="mf">0.1</span> <span class="o">*</span> <span class="n">averaged_model_parameter</span> <span class="o">+</span> <span class="mf">0.9</span> <span class="o">*</span> <span class="n">model_parameter</span>
<span class="gp">>>> </span><span class="n">ema_model</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">swa_utils</span><span class="o">.</span><span class="n">AveragedModel</span><span class="p">(</span><span class="n">model</span><span class="p">,</span> <span class="n">avg_fn</span><span class="o">=</span><span class="n">ema_avg</span><span class="p">)</span>
</pre></div>
</div>
</section>
<section id="putting-it-all-together">
<h3>Putting it all together<a class="headerlink" href="#putting-it-all-together" title="Permalink to this heading">¶</a></h3>
<p>In the example below, <code class="docutils literal notranslate"><span class="pre">swa_model</span></code> is the SWA model that accumulates the averages of the weights.
We train the model for a total of 300 epochs and we switch to the SWA learning rate schedule
and start to collect SWA averages of the parameters at epoch 160:</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">loader</span><span class="p">,</span> <span class="n">optimizer</span><span class="p">,</span> <span class="n">model</span><span class="p">,</span> <span class="n">loss_fn</span> <span class="o">=</span> <span class="o">...</span>
<span class="gp">>>> </span><span class="n">swa_model</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">swa_utils</span><span class="o">.</span><span class="n">AveragedModel</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">scheduler</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">lr_scheduler</span><span class="o">.</span><span class="n">CosineAnnealingLR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">T_max</span><span class="o">=</span><span class="mi">300</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">swa_start</span> <span class="o">=</span> <span class="mi">160</span>
<span class="gp">>>> </span><span class="n">swa_scheduler</span> <span class="o">=</span> <span class="n">SWALR</span><span class="p">(</span><span class="n">optimizer</span><span class="p">,</span> <span class="n">swa_lr</span><span class="o">=</span><span class="mf">0.05</span><span class="p">)</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">300</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="k">for</span> <span class="nb">input</span><span class="p">,</span> <span class="n">target</span> <span class="ow">in</span> <span class="n">loader</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="n">loss_fn</span><span class="p">(</span><span class="n">model</span><span class="p">(</span><span class="nb">input</span><span class="p">),</span> <span class="n">target</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="k">if</span> <span class="n">epoch</span> <span class="o">></span> <span class="n">swa_start</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="n">swa_model</span><span class="o">.</span><span class="n">update_parameters</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="n">swa_scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="k">else</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="n">scheduler</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Update bn statistics for the swa_model at the end</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">optim</span><span class="o">.</span><span class="n">swa_utils</span><span class="o">.</span><span class="n">update_bn</span><span class="p">(</span><span class="n">loader</span><span class="p">,</span> <span class="n">swa_model</span><span class="p">)</span>
<span class="gp">>>> </span><span class="c1"># Use swa_model to make predictions on test data</span>
<span class="gp">>>> </span><span class="n">preds</span> <span class="o">=</span> <span class="n">swa_model</span><span class="p">(</span><span class="n">test_input</span><span class="p">)</span>
</pre></div>
</div>
</section>
</section>
</section>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="generated/torch.optim.Optimizer.add_param_group.html" class="btn btn-neutral float-right" title="torch.optim.Optimizer.add_param_group" accesskey="n" rel="next">Next <img src="_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="generated/onnx_diagnostics_rules/POE0004%3Aoperator-supported-in-newer-opset-version.html" class="btn btn-neutral" title="POE0004:operator-supported-in-newer-opset-version" accesskey="p" rel="prev"><img src="_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2023, PyTorch Contributors.
</p>
</div>
<div>
Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/rtfd/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>.
</div>
</footer>
</div>
<script>
var match = window.location.href.match(/\/_[a-zA-Z0-9_]*.html|_dynamo/gi);
var url = window.location.href.lastIndexOf(match[match.length-1]);
if (url)
{
var div = '<div class="admonition note"><p class="admonition-title">Note</p><p><i class="fa fa-exclamation-circle" aria-hidden="true"> </i> This page describes an internal API which is not intended to be used outside of the PyTorch codebase and can be modified or removed without notice.</p></div>'
document.getElementById("pytorch-article").insertAdjacentHTML('afterBegin', div)
}
</script>