forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathamp_examples.html
1070 lines (845 loc) · 80 KB
/
amp_examples.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Automatic Mixed Precision examples — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/notes/amp_examples.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="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/jit.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="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="Autograd mechanics" href="autograd.html" />
<link rel="prev" title="PyTorch documentation" href="../index.html" />
<script src="../_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="../_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[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">
</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>
<div class="ecosystem-dropdown">
<a id="dropdownMenuButton" data-toggle="ecosystem-dropdown">
Ecosystem
</a>
<div class="ecosystem-dropdown-menu">
<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>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class=dropdown-title>Tools & Libraries</span>
<p>Explore the ecosystem of tools and libraries</p>
</a>
</div>
</div>
</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">
<a href="https://pytorch.org/docs/stable/index.html">Docs</a>
</li>
<li>
<div class="resources-dropdown">
<a id="resourcesDropdownButton" data-toggle="resources-dropdown">
Resources
</a>
<div class="resources-dropdown-menu">
<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/features">
<span class=dropdown-title>About</span>
<p>Learn about PyTorch’s features and capabilities</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='http://pytorch.org/docs/versions.html'>1.6.0a0+f2f8027 ▼</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/notes/amp_examples.html">
You are viewing unstable developer preview docs.
Click here to view docs for latest stable release.
</a>
</div>
<p class="caption"><span class="caption-text">Notes</span></p>
<ul class="current">
<li class="toctree-l1 current"><a class="current reference internal" href="#">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="large_scale_deployments.html">Features for large-scale deployments</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 external" href="https://pytorch.org/cppdocs/">C++ API</a></li>
<li class="toctree-l1"><a class="reference internal" href="../packages.html">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="../distributed.html">torch.distributed</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="../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="../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="../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="../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="../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="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/elastic/">PyTorch Elastic (torchelastic)</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text">torchtext</a></li>
<li class="toctree-l1"><a class="reference internal" href="../torchvision/index.html">torchvision</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>Automatic Mixed Precision examples</li>
<li class="pytorch-breadcrumbs-aside">
<a href="../_sources/notes/amp_examples.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="automatic-mixed-precision-examples">
<span id="amp-examples"></span><h1>Automatic Mixed Precision examples<a class="headerlink" href="#automatic-mixed-precision-examples" title="Permalink to this headline">¶</a></h1>
<p>Ordinarily, “automatic mixed precision training” means training with
<a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.autocast</span></code></a> and <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.GradScaler</span></code></a> together.</p>
<p>Instances of <a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.autocast</span></code></a> enable autocasting for chosen regions.
Autocasting automatically chooses the precision for GPU operations to improve performance
while maintaining accuracy.</p>
<p>Instances of <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.GradScaler</span></code></a> help perform the steps of
gradient scaling conveniently. Gradient scaling improves convergence for networks with <code class="docutils literal notranslate"><span class="pre">float16</span></code>
gradients by minimizing gradient underflow, as explained <a class="reference internal" href="../amp.html#gradient-scaling"><span class="std std-ref">here</span></a>.</p>
<p><a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.autocast</span></code></a> and <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.cuda.amp.GradScaler</span></code></a> are modular.
In the samples below, each is used as its individual documentation suggests.</p>
<div class="contents local topic" id="contents">
<ul class="simple">
<li><p><a class="reference internal" href="#typical-mixed-precision-training" id="id2">Typical Mixed Precision Training</a></p></li>
<li><p><a class="reference internal" href="#working-with-unscaled-gradients" id="id3">Working with Unscaled Gradients</a></p>
<ul>
<li><p><a class="reference internal" href="#gradient-clipping" id="id4">Gradient clipping</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#working-with-scaled-gradients" id="id5">Working with Scaled Gradients</a></p>
<ul>
<li><p><a class="reference internal" href="#gradient-accumulation" id="id6">Gradient accumulation</a></p></li>
<li><p><a class="reference internal" href="#gradient-penalty" id="id7">Gradient penalty</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#working-with-multiple-models-losses-and-optimizers" id="id8">Working with Multiple Models, Losses, and Optimizers</a></p></li>
<li><p><a class="reference internal" href="#working-with-multiple-gpus" id="id9">Working with Multiple GPUs</a></p>
<ul>
<li><p><a class="reference internal" href="#dataparallel-in-a-single-process" id="id10">DataParallel in a single process</a></p></li>
<li><p><a class="reference internal" href="#distributeddataparallel-one-gpu-per-process" id="id11">DistributedDataParallel, one GPU per process</a></p></li>
<li><p><a class="reference internal" href="#distributeddataparallel-multiple-gpus-per-process" id="id12">DistributedDataParallel, multiple GPUs per process</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#autocast-and-custom-autograd-functions" id="id13">Autocast and Custom Autograd Functions</a></p>
<ul>
<li><p><a class="reference internal" href="#functions-with-multiple-inputs-or-autocastable-ops" id="id14">Functions with multiple inputs or autocastable ops</a></p></li>
<li><p><a class="reference internal" href="#functions-that-need-a-particular-dtype" id="id15">Functions that need a particular dtype</a></p></li>
</ul>
</li>
</ul>
</div>
<div class="section" id="typical-mixed-precision-training">
<h2><a class="toc-backref" href="#id2">Typical Mixed Precision Training</a><a class="headerlink" href="#typical-mixed-precision-training" title="Permalink to this headline">¶</a></h2>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="c1"># Creates model and optimizer in default precision</span>
<span class="n">model</span> <span class="o">=</span> <span class="n">Net</span><span class="p">()</span><span class="o">.</span><span class="n">cuda</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">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="o">...</span><span class="p">)</span>
<span class="c1"># Creates a GradScaler once at the beginning of training.</span>
<span class="n">scaler</span> <span class="o">=</span> <span class="n">GradScaler</span><span class="p">()</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</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">data</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="c1"># Runs the forward pass with autocasting.</span>
<span class="k">with</span> <span class="n">autocast</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="c1"># Scales loss. Calls backward() on scaled loss to create scaled gradients.</span>
<span class="c1"># Backward passes under autocast are not recommended.</span>
<span class="c1"># Backward ops run in the same precision that autocast used for corresponding forward ops.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="c1"># scaler.step() first unscales the gradients of the optimizer's assigned params.</span>
<span class="c1"># If these gradients do not contain infs or NaNs, optimizer.step() is then called,</span>
<span class="c1"># otherwise, optimizer.step() is skipped.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="c1"># Updates the scale for next iteration.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="working-with-unscaled-gradients">
<span id="id1"></span><h2><a class="toc-backref" href="#id3">Working with Unscaled Gradients</a><a class="headerlink" href="#working-with-unscaled-gradients" title="Permalink to this headline">¶</a></h2>
<p>All gradients produced by <code class="docutils literal notranslate"><span class="pre">scaler.scale(loss).backward()</span></code> are scaled. If you wish to modify or inspect
the parameters’ <code class="docutils literal notranslate"><span class="pre">.grad</span></code> attributes between <code class="docutils literal notranslate"><span class="pre">backward()</span></code> and <code class="docutils literal notranslate"><span class="pre">scaler.step(optimizer)</span></code>, you should
unscale them first. For example, gradient clipping manipulates a set of gradients such that their global norm
(see <a class="reference internal" href="../generated/torch.nn.utils.clip_grad_norm_.html#torch.nn.utils.clip_grad_norm_" title="torch.nn.utils.clip_grad_norm_"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.utils.clip_grad_norm_()</span></code></a>) or maximum magnitude (see <a class="reference internal" href="../generated/torch.nn.utils.clip_grad_value_.html#torch.nn.utils.clip_grad_value_" title="torch.nn.utils.clip_grad_value_"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.nn.utils.clip_grad_value_()</span></code></a>)
is <span class="math"><span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mo><</mo><mo>=</mo></mrow><annotation encoding="application/x-tex"><=</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:0.5782em;vertical-align:-0.0391em;"></span><span class="mrel"><</span></span><span class="base"><span class="strut" style="height:0.36687em;vertical-align:0em;"></span><span class="mrel">=</span></span></span></span>
</span> some user-imposed threshold. If you attempted to clip <em>without</em> unscaling, the gradients’ norm/maximum
magnitude would also be scaled, so your requested threshold (which was meant to be the threshold for <em>unscaled</em>
gradients) would be invalid.</p>
<p><code class="docutils literal notranslate"><span class="pre">scaler.unscale_(optimizer)</span></code> unscales gradients held by <code class="docutils literal notranslate"><span class="pre">optimizer</span></code>’s assigned parameters.
If your model or models contain other parameters that were assigned to another optimizer
(say <code class="docutils literal notranslate"><span class="pre">optimizer2</span></code>), you may call <code class="docutils literal notranslate"><span class="pre">scaler.unscale_(optimizer2)</span></code> separately to unscale those
parameters’ gradients as well.</p>
<div class="section" id="gradient-clipping">
<h3><a class="toc-backref" href="#id4">Gradient clipping</a><a class="headerlink" href="#gradient-clipping" title="Permalink to this headline">¶</a></h3>
<p>Calling <code class="docutils literal notranslate"><span class="pre">scaler.unscale_(optimizer)</span></code> before clipping enables you to clip unscaled gradients as usual:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">scaler</span> <span class="o">=</span> <span class="n">GradScaler</span><span class="p">()</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</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">data</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="k">with</span> <span class="n">autocast</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">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="c1"># Unscales the gradients of optimizer's assigned params in-place</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">unscale_</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="c1"># Since the gradients of optimizer's assigned params are unscaled, clips as usual:</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">clip_grad_norm_</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">max_norm</span><span class="p">)</span>
<span class="c1"># optimizer's gradients are already unscaled, so scaler.step does not unscale them,</span>
<span class="c1"># although it still skips optimizer.step() if the gradients contain infs or NaNs.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="c1"># Updates the scale for next iteration.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
</pre></div>
</div>
<p><code class="docutils literal notranslate"><span class="pre">scaler</span></code> records that <code class="docutils literal notranslate"><span class="pre">scaler.unscale_(optimizer)</span></code> was already called for this optimizer
this iteration, so <code class="docutils literal notranslate"><span class="pre">scaler.step(optimizer)</span></code> knows not to redundantly unscale gradients before
(internally) calling <code class="docutils literal notranslate"><span class="pre">optimizer.step()</span></code>.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_</span></code></a> should only be called once per optimizer per <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a> call,
and only after all gradients for that optimizer’s assigned parameters have been accumulated.
Calling <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_</span></code></a> twice for a given optimizer between each <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a> triggers a RuntimeError.</p>
</div>
</div>
</div>
<div class="section" id="working-with-scaled-gradients">
<h2><a class="toc-backref" href="#id5">Working with Scaled Gradients</a><a class="headerlink" href="#working-with-scaled-gradients" title="Permalink to this headline">¶</a></h2>
<div class="section" id="gradient-accumulation">
<h3><a class="toc-backref" href="#id6">Gradient accumulation</a><a class="headerlink" href="#gradient-accumulation" title="Permalink to this headline">¶</a></h3>
<p>Gradient accumulation adds gradients over an effective batch of size <code class="docutils literal notranslate"><span class="pre">batch_per_iter</span> <span class="pre">*</span> <span class="pre">iters_to_accumulate</span></code>
(<code class="docutils literal notranslate"><span class="pre">*</span> <span class="pre">num_procs</span></code> if distributed). The scale should be calibrated for the effective batch, which means inf/NaN checking,
step skipping if inf/NaN grads are found, and scale updates should occur at effective-batch granularity.
Also, grads should remain scaled, and the scale factor should remain constant, while grads for a given effective
batch are accumulated. If grads are unscaled (or the scale factor changes) before accumulation is complete,
the next backward pass will add scaled grads to unscaled grads (or grads scaled by a different factor)
after which it’s impossible to recover the accumulated unscaled grads <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a> must apply.</p>
<p>Therefore, if you want to <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_</span></code></a> grads (e.g., to allow clipping unscaled grads),
call <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">unscale_</span></code></a> just before <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a>, after all (scaled) grads for the upcoming
<a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a> have been accumulated. Also, only call <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.update" title="torch.cuda.amp.GradScaler.update"><code class="xref py py-meth docutils literal notranslate"><span class="pre">update</span></code></a> at the end of iterations
where you called <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">step</span></code></a> for a full effective batch:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">scaler</span> <span class="o">=</span> <span class="n">GradScaler</span><span class="p">()</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</span><span class="p">:</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</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="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">data</span><span class="p">):</span>
<span class="k">with</span> <span class="n">autocast</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">loss</span> <span class="o">/</span> <span class="n">iters_to_accumulate</span>
<span class="c1"># Accumulates scaled gradients.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="k">if</span> <span class="p">(</span><span class="n">i</span> <span class="o">+</span> <span class="mi">1</span><span class="p">)</span> <span class="o">%</span> <span class="n">iters_to_accumulate</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
<span class="c1"># may unscale_ here if desired (e.g., to allow clipping unscaled gradients)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="gradient-penalty">
<h3><a class="toc-backref" href="#id7">Gradient penalty</a><a class="headerlink" href="#gradient-penalty" title="Permalink to this headline">¶</a></h3>
<p>A gradient penalty implementation commonly creates gradients using
<a class="reference internal" href="../autograd.html#torch.autograd.grad" title="torch.autograd.grad"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.autograd.grad()</span></code></a>, combines them to create the penalty value,
and adds the penalty value to the loss.</p>
<p>Here’s an ordinary example of an L2 penalty without gradient scaling or autocasting:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</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">data</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="c1"># Creates gradients</span>
<span class="n">grad_params</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">grad</span><span class="p">(</span><span class="n">loss</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">create_graph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="c1"># Computes the penalty term and adds it to the loss</span>
<span class="n">grad_norm</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">grad</span> <span class="ow">in</span> <span class="n">grad_params</span><span class="p">:</span>
<span class="n">grad_norm</span> <span class="o">+=</span> <span class="n">grad</span><span class="o">.</span><span class="n">pow</span><span class="p">(</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span>
<span class="n">grad_norm</span> <span class="o">=</span> <span class="n">grad_norm</span><span class="o">.</span><span class="n">sqrt</span><span class="p">()</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">loss</span> <span class="o">+</span> <span class="n">grad_norm</span>
<span class="n">loss</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="c1"># clip gradients here, if desired</span>
<span class="n">optimizer</span><span class="o">.</span><span class="n">step</span><span class="p">()</span>
</pre></div>
</div>
<p>To implement a gradient penalty <em>with</em> gradient scaling, the loss passed to
<a class="reference internal" href="../autograd.html#torch.autograd.grad" title="torch.autograd.grad"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.autograd.grad()</span></code></a> should be scaled. The resulting gradients
will therefore be scaled, and should be unscaled before being combined to create the
penalty value.</p>
<p>Also, the penalty term computation is part of the forward pass, and therefore should be
inside an <a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a> context.</p>
<p>Here’s how that looks for the same L2 penalty:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">scaler</span> <span class="o">=</span> <span class="n">GradScaler</span><span class="p">()</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</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">data</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="k">with</span> <span class="n">autocast</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="c1"># Scales the loss for autograd.grad's backward pass, resulting in scaled grad_params</span>
<span class="n">scaled_grad_params</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">grad</span><span class="p">(</span><span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</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">create_graph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="c1"># Creates unscaled grad_params before computing the penalty. scaled_grad_params are</span>
<span class="c1"># not owned by any optimizer, so ordinary division is used instead of scaler.unscale_:</span>
<span class="n">inv_scale</span> <span class="o">=</span> <span class="mf">1.</span><span class="o">/</span><span class="n">scaler</span><span class="o">.</span><span class="n">get_scale</span><span class="p">()</span>
<span class="n">grad_params</span> <span class="o">=</span> <span class="p">[</span><span class="n">p</span> <span class="o">*</span> <span class="n">inv_scale</span> <span class="k">for</span> <span class="n">p</span> <span class="ow">in</span> <span class="n">scaled_grad_params</span><span class="p">]</span>
<span class="c1"># Computes the penalty term and adds it to the loss</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">grad_norm</span> <span class="o">=</span> <span class="mi">0</span>
<span class="k">for</span> <span class="n">grad</span> <span class="ow">in</span> <span class="n">grad_params</span><span class="p">:</span>
<span class="n">grad_norm</span> <span class="o">+=</span> <span class="n">grad</span><span class="o">.</span><span class="n">pow</span><span class="p">(</span><span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span>
<span class="n">grad_norm</span> <span class="o">=</span> <span class="n">grad_norm</span><span class="o">.</span><span class="n">sqrt</span><span class="p">()</span>
<span class="n">loss</span> <span class="o">=</span> <span class="n">loss</span> <span class="o">+</span> <span class="n">grad_norm</span>
<span class="c1"># Applies scaling to the backward call as usual.</span>
<span class="c1"># Accumulates leaf gradients that are correctly scaled.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="c1"># may unscale_ here if desired (e.g., to allow clipping unscaled gradients)</span>
<span class="c1"># step() and update() proceed as usual.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
</pre></div>
</div>
</div>
</div>
<div class="section" id="working-with-multiple-models-losses-and-optimizers">
<h2><a class="toc-backref" href="#id8">Working with Multiple Models, Losses, and Optimizers</a><a class="headerlink" href="#working-with-multiple-models-losses-and-optimizers" title="Permalink to this headline">¶</a></h2>
<p>If your network has multiple losses, you must call <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.scale" title="torch.cuda.amp.GradScaler.scale"><code class="xref py py-meth docutils literal notranslate"><span class="pre">scaler.scale</span></code></a> on each of them individually.
If your network has multiple optimizers, you may call <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.unscale_" title="torch.cuda.amp.GradScaler.unscale_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">scaler.unscale_</span></code></a> on any of them individually,
and you must call <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.step" title="torch.cuda.amp.GradScaler.step"><code class="xref py py-meth docutils literal notranslate"><span class="pre">scaler.step</span></code></a> on each of them individually.</p>
<p>However, <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler.update" title="torch.cuda.amp.GradScaler.update"><code class="xref py py-meth docutils literal notranslate"><span class="pre">scaler.update</span></code></a> should only be called once,
after all optimizers used this iteration have been stepped:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">scaler</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">amp</span><span class="o">.</span><span class="n">GradScaler</span><span class="p">()</span>
<span class="k">for</span> <span class="n">epoch</span> <span class="ow">in</span> <span class="n">epochs</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">data</span><span class="p">:</span>
<span class="n">optimizer0</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="n">optimizer1</span><span class="o">.</span><span class="n">zero_grad</span><span class="p">()</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">output0</span> <span class="o">=</span> <span class="n">model0</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">output1</span> <span class="o">=</span> <span class="n">model1</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="n">loss0</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="mi">2</span> <span class="o">*</span> <span class="n">output0</span> <span class="o">+</span> <span class="mi">3</span> <span class="o">*</span> <span class="n">output1</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span>
<span class="n">loss1</span> <span class="o">=</span> <span class="n">loss_fn</span><span class="p">(</span><span class="mi">3</span> <span class="o">*</span> <span class="n">output0</span> <span class="o">-</span> <span class="mi">5</span> <span class="o">*</span> <span class="n">output1</span><span class="p">,</span> <span class="n">target</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss0</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">(</span><span class="n">retain_graph</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">scale</span><span class="p">(</span><span class="n">loss1</span><span class="p">)</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="c1"># You can choose which optimizers receive explicit unscaling, if you</span>
<span class="c1"># want to inspect or modify the gradients of the params they own.</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">unscale_</span><span class="p">(</span><span class="n">optimizer0</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer0</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">step</span><span class="p">(</span><span class="n">optimizer1</span><span class="p">)</span>
<span class="n">scaler</span><span class="o">.</span><span class="n">update</span><span class="p">()</span>
</pre></div>
</div>
<p>Each optimizer checks its gradients for infs/NaNs and makes an independent decision
whether or not to skip the step. This may result in one optimizer skipping the step
while the other one does not. Since step skipping occurs rarely (every several hundred iterations)
this should not impede convergence. If you observe poor convergence after adding gradient scaling
to a multiple-optimizer model, please report a bug.</p>
</div>
<div class="section" id="working-with-multiple-gpus">
<span id="amp-multigpu"></span><h2><a class="toc-backref" href="#id9">Working with Multiple GPUs</a><a class="headerlink" href="#working-with-multiple-gpus" title="Permalink to this headline">¶</a></h2>
<p>The issues described here only affect <a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a>. <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradScaler</span></code></a>’s usage is unchanged.</p>
<div class="section" id="dataparallel-in-a-single-process">
<span id="amp-dataparallel"></span><h3><a class="toc-backref" href="#id10">DataParallel in a single process</a><a class="headerlink" href="#dataparallel-in-a-single-process" title="Permalink to this headline">¶</a></h3>
<p><a class="reference internal" href="../generated/torch.nn.DataParallel.html#torch.nn.DataParallel" title="torch.nn.DataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.DataParallel</span></code></a> spawns threads to run the forward pass on each device.
The autocast state is thread local, so the following will not work:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">model</span> <span class="o">=</span> <span class="n">MyModel</span><span class="p">()</span>
<span class="n">dp_model</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">DataParallel</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="c1"># Sets autocast in the main thread</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="c1"># dp_model's internal threads won't autocast. The main thread's autocast state has no effect.</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">dp_model</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="c1"># loss_fn still autocasts, but it's too late...</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>
</pre></div>
</div>
<p>The fix is simple. Enable autocast as part of <code class="docutils literal notranslate"><span class="pre">MyModel.forward</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">MyModel</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="o">...</span>
<span class="nd">@autocast</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="o">...</span>
<span class="c1"># Alternatively</span>
<span class="n">MyModel</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="o">...</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">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="o">...</span>
</pre></div>
</div>
<p>The following now autocasts in <code class="docutils literal notranslate"><span class="pre">dp_model</span></code>’s threads (which execute <code class="docutils literal notranslate"><span class="pre">forward</span></code>) and the main thread
(which executes <code class="docutils literal notranslate"><span class="pre">loss_fn</span></code>):</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">model</span> <span class="o">=</span> <span class="n">MyModel</span><span class="p">()</span>
<span class="n">dp_model</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">DataParallel</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">dp_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>
</pre></div>
</div>
</div>
<div class="section" id="distributeddataparallel-one-gpu-per-process">
<h3><a class="toc-backref" href="#id11">DistributedDataParallel, one GPU per process</a><a class="headerlink" href="#distributeddataparallel-one-gpu-per-process" title="Permalink to this headline">¶</a></h3>
<p><a class="reference internal" href="../generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel</span></code></a>’s documentation recommends one GPU per process for best
performance. In this case, <code class="docutils literal notranslate"><span class="pre">DistributedDataParallel</span></code> does not spawn threads internally,
so usages of <a class="reference internal" href="../amp.html#torch.cuda.amp.autocast" title="torch.cuda.amp.autocast"><code class="xref py py-class docutils literal notranslate"><span class="pre">autocast</span></code></a> and <a class="reference internal" href="../amp.html#torch.cuda.amp.GradScaler" title="torch.cuda.amp.GradScaler"><code class="xref py py-class docutils literal notranslate"><span class="pre">GradScaler</span></code></a> are not affected.</p>
</div>
<div class="section" id="distributeddataparallel-multiple-gpus-per-process">
<h3><a class="toc-backref" href="#id12">DistributedDataParallel, multiple GPUs per process</a><a class="headerlink" href="#distributeddataparallel-multiple-gpus-per-process" title="Permalink to this headline">¶</a></h3>
<p>Here <a class="reference internal" href="../generated/torch.nn.parallel.DistributedDataParallel.html#torch.nn.parallel.DistributedDataParallel" title="torch.nn.parallel.DistributedDataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.parallel.DistributedDataParallel</span></code></a> may spawn a side thread to run the forward pass on each
device, like <a class="reference internal" href="../generated/torch.nn.DataParallel.html#torch.nn.DataParallel" title="torch.nn.DataParallel"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.DataParallel</span></code></a>. <a class="reference internal" href="#amp-dataparallel"><span class="std std-ref">The fix is the same</span></a>:
apply autocast as part of your model’s <code class="docutils literal notranslate"><span class="pre">forward</span></code> method to ensure it’s enabled in side threads.</p>
</div>
</div>
<div class="section" id="autocast-and-custom-autograd-functions">
<span id="amp-custom-examples"></span><h2><a class="toc-backref" href="#id13">Autocast and Custom Autograd Functions</a><a class="headerlink" href="#autocast-and-custom-autograd-functions" title="Permalink to this headline">¶</a></h2>
<p>If your network uses <a class="reference internal" href="extending.html#extending-autograd"><span class="std std-ref">custom autograd functions</span></a>
(subclasses of <a class="reference internal" href="../autograd.html#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.autograd.Function</span></code></a>), changes are required for
autocast compatibility if any function</p>
<ul class="simple">
<li><p>takes multiple floating-point Tensor inputs,</p></li>
<li><p>wraps any autocastable op (see the <a class="reference internal" href="../amp.html#autocast-op-reference"><span class="std std-ref">Autocast Op Reference</span></a>), or</p></li>
<li><p>requires a particular <code class="docutils literal notranslate"><span class="pre">dtype</span></code> (for example, if it wraps
<a class="reference external" href="https://pytorch.org/tutorials/advanced/cpp_extension.html">CUDA extensions</a>
that were only compiled for <code class="docutils literal notranslate"><span class="pre">dtype</span></code>).</p></li>
</ul>
<p>In all cases, if you’re importing the function and can’t alter its definition, a safe fallback
is to disable autocast and force execution in <code class="docutils literal notranslate"><span class="pre">float32</span></code> ( or <code class="docutils literal notranslate"><span class="pre">dtype</span></code>) at any points of use where errors occur:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="o">...</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">(</span><span class="n">enabled</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">imported_function</span><span class="p">(</span><span class="n">input1</span><span class="o">.</span><span class="n">float</span><span class="p">(),</span> <span class="n">input2</span><span class="o">.</span><span class="n">float</span><span class="p">())</span>
</pre></div>
</div>
<p>If you’re the function’s author (or can alter its definition) a better solution is to use the
<a class="reference internal" href="../amp.html#torch.cuda.amp.custom_fwd" title="torch.cuda.amp.custom_fwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.cuda.amp.custom_fwd()</span></code></a> and <a class="reference internal" href="../amp.html#torch.cuda.amp.custom_bwd" title="torch.cuda.amp.custom_bwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.cuda.amp.custom_bwd()</span></code></a> decorators as shown in
the relevant case below.</p>
<div class="section" id="functions-with-multiple-inputs-or-autocastable-ops">
<h3><a class="toc-backref" href="#id14">Functions with multiple inputs or autocastable ops</a><a class="headerlink" href="#functions-with-multiple-inputs-or-autocastable-ops" title="Permalink to this headline">¶</a></h3>
<p>Apply <a class="reference internal" href="../amp.html#torch.cuda.amp.custom_fwd" title="torch.cuda.amp.custom_fwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">custom_fwd()</span></code></a> and <a class="reference internal" href="../amp.html#torch.cuda.amp.custom_bwd" title="torch.cuda.amp.custom_bwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">custom_bwd()</span></code></a> (with no arguments) to <code class="docutils literal notranslate"><span class="pre">forward</span></code> and <code class="docutils literal notranslate"><span class="pre">backward</span></code>
respectively. These ensure <code class="docutils literal notranslate"><span class="pre">forward</span></code> executes with the current autocast state and <code class="docutils literal notranslate"><span class="pre">backward</span></code>
executes with the same autocast state as <code class="docutils literal notranslate"><span class="pre">forward</span></code> (which can prevent type mismatch errors):</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">MyMM</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">Function</span><span class="p">):</span>
<span class="nd">@staticmethod</span>
<span class="nd">@custom_fwd</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">):</span>
<span class="n">ctx</span><span class="o">.</span><span class="n">save_for_backward</span><span class="p">(</span><span class="n">a</span><span class="p">,</span> <span class="n">b</span><span class="p">)</span>
<span class="k">return</span> <span class="n">a</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">b</span><span class="p">)</span>
<span class="nd">@staticmethod</span>
<span class="nd">@custom_bwd</span>
<span class="k">def</span> <span class="nf">backward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="n">grad</span><span class="p">):</span>
<span class="n">a</span><span class="p">,</span> <span class="n">b</span> <span class="o">=</span> <span class="n">ctx</span><span class="o">.</span><span class="n">saved_tensors</span>
<span class="k">return</span> <span class="n">grad</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">b</span><span class="o">.</span><span class="n">t</span><span class="p">()),</span> <span class="n">a</span><span class="o">.</span><span class="n">t</span><span class="p">()</span><span class="o">.</span><span class="n">mm</span><span class="p">(</span><span class="n">grad</span><span class="p">)</span>
</pre></div>
</div>
<p>Now <code class="docutils literal notranslate"><span class="pre">MyMM</span></code> can be invoked anywhere, without disabling autocast or manually casting inputs:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">mymm</span> <span class="o">=</span> <span class="n">MyMM</span><span class="o">.</span><span class="n">apply</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">mymm</span><span class="p">(</span><span class="n">input1</span><span class="p">,</span> <span class="n">input2</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="functions-that-need-a-particular-dtype">
<h3><a class="toc-backref" href="#id15">Functions that need a particular dtype</a><a class="headerlink" href="#functions-that-need-a-particular-dtype" title="Permalink to this headline">¶</a></h3>
<p>Apply <a class="reference internal" href="../amp.html#torch.cuda.amp.custom_fwd" title="torch.cuda.amp.custom_fwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">custom_fwd(cast_inputs=torch.float32)</span></code></a> to <code class="docutils literal notranslate"><span class="pre">forward</span></code>
and <a class="reference internal" href="../amp.html#torch.cuda.amp.custom_bwd" title="torch.cuda.amp.custom_bwd"><code class="xref py py-func docutils literal notranslate"><span class="pre">custom_bwd</span></code></a> (with no arguments) to <code class="docutils literal notranslate"><span class="pre">backward</span></code>.
With <code class="docutils literal notranslate"><span class="pre">cast_inputs=torch.float32</span></code>, these disable autocast in <code class="docutils literal notranslate"><span class="pre">forward</span></code> and <code class="docutils literal notranslate"><span class="pre">backward</span></code>,
and <code class="docutils literal notranslate"><span class="pre">forward</span></code> casts incoming floating-point CUDA Tensors to <code class="docutils literal notranslate"><span class="pre">float32</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">MyFloat32Func</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">Function</span><span class="p">):</span>
<span class="nd">@staticmethod</span>
<span class="nd">@custom_fwd</span><span class="p">(</span><span class="n">cast_inputs</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="nb">input</span><span class="p">):</span>
<span class="n">ctx</span><span class="o">.</span><span class="n">save_for_backward</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="o">...</span>
<span class="k">return</span> <span class="n">fwd_output</span>
<span class="nd">@staticmethod</span>
<span class="nd">@custom_bwd</span>
<span class="k">def</span> <span class="nf">backward</span><span class="p">(</span><span class="n">ctx</span><span class="p">,</span> <span class="n">grad</span><span class="p">):</span>
<span class="o">...</span>
</pre></div>
</div>
<p>Now <code class="docutils literal notranslate"><span class="pre">MyFloat32Func</span></code> can be invoked anywhere, without disabling autocast or manually casting inputs:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">func</span> <span class="o">=</span> <span class="n">MyFloat32Func</span><span class="o">.</span><span class="n">apply</span>
<span class="k">with</span> <span class="n">autocast</span><span class="p">():</span>
<span class="c1"># func will run in float32, regardless of the surrounding autocast state</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">func</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
</pre></div>
</div>
</div>
</div>
</div>
</article>
</div>
<footer>
<div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
<a href="autograd.html" class="btn btn-neutral float-right" title="Autograd mechanics" accesskey="n" rel="next">Next <img src="../_static/images/chevron-right-orange.svg" class="next-page"></a>
<a href="../index.html" class="btn btn-neutral" title="PyTorch documentation" accesskey="p" rel="prev"><img src="../_static/images/chevron-right-orange.svg" class="previous-page"> Previous</a>
</div>
<hr>
<div role="contentinfo">
<p>
© Copyright 2019, Torch 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>
</div>
<div class="pytorch-content-right" id="pytorch-content-right">
<div class="pytorch-right-menu" id="pytorch-right-menu">
<div class="pytorch-side-scroll" id="pytorch-side-scroll-right">
<ul>
<li><a class="reference internal" href="#">Automatic Mixed Precision examples</a><ul>
<li><a class="reference internal" href="#typical-mixed-precision-training">Typical Mixed Precision Training</a></li>
<li><a class="reference internal" href="#working-with-unscaled-gradients">Working with Unscaled Gradients</a><ul>
<li><a class="reference internal" href="#gradient-clipping">Gradient clipping</a></li>
</ul>
</li>
<li><a class="reference internal" href="#working-with-scaled-gradients">Working with Scaled Gradients</a><ul>
<li><a class="reference internal" href="#gradient-accumulation">Gradient accumulation</a></li>
<li><a class="reference internal" href="#gradient-penalty">Gradient penalty</a></li>
</ul>
</li>
<li><a class="reference internal" href="#working-with-multiple-models-losses-and-optimizers">Working with Multiple Models, Losses, and Optimizers</a></li>
<li><a class="reference internal" href="#working-with-multiple-gpus">Working with Multiple GPUs</a><ul>
<li><a class="reference internal" href="#dataparallel-in-a-single-process">DataParallel in a single process</a></li>
<li><a class="reference internal" href="#distributeddataparallel-one-gpu-per-process">DistributedDataParallel, one GPU per process</a></li>
<li><a class="reference internal" href="#distributeddataparallel-multiple-gpus-per-process">DistributedDataParallel, multiple GPUs per process</a></li>
</ul>
</li>
<li><a class="reference internal" href="#autocast-and-custom-autograd-functions">Autocast and Custom Autograd Functions</a><ul>
<li><a class="reference internal" href="#functions-with-multiple-inputs-or-autocastable-ops">Functions with multiple inputs or autocastable ops</a></li>
<li><a class="reference internal" href="#functions-that-need-a-particular-dtype">Functions that need a particular dtype</a></li>
</ul>
</li>
</ul>
</li>
</ul>
</div>
</div>
</div>
</section>
</div>
<script type="text/javascript" id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/language_data.js"></script>
<script type="text/javascript" src="../_static/js/vendor/popper.min.js"></script>
<script type="text/javascript" src="../_static/js/vendor/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/list.js/1.5.0/list.min.js"></script>
<script type="text/javascript" src="../_static/js/theme.js"></script>
<script type="text/javascript">
jQuery(function () {
SphinxRtdTheme.Navigation.enable(true);
});
</script>
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','https://www.google-analytics.com/analytics.js','ga');
ga('create', 'UA-90545585-1', 'auto');
ga('send', 'pageview');
</script>
<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>
<img height="1" width="1" style="border-style:none;" alt="" src="https://www.googleadservices.com/pagead/conversion/795629140/?label=txkmCPmdtosBENSssfsC&guid=ON&script=0"/>
<!-- Begin Footer -->
<div class="container-fluid docs-tutorials-resources" id="docs-tutorials-resources">
<div class="container">
<div class="row">
<div class="col-md-4 text-center">
<h2>Docs</h2>
<p>Access comprehensive developer documentation for PyTorch</p>
<a class="with-right-arrow" href="https://pytorch.org/docs/stable/index.html">View Docs</a>
</div>
<div class="col-md-4 text-center">
<h2>Tutorials</h2>
<p>Get in-depth tutorials for beginners and advanced developers</p>
<a class="with-right-arrow" href="https://pytorch.org/tutorials">View Tutorials</a>
</div>
<div class="col-md-4 text-center">
<h2>Resources</h2>
<p>Find development resources and get your questions answered</p>
<a class="with-right-arrow" href="https://pytorch.org/resources">View Resources</a>
</div>
</div>
</div>
</div>
<footer class="site-footer">
<div class="container footer-container">
<div class="footer-logo-wrapper">
<a href="https://pytorch.org/" class="footer-logo"></a>
</div>
<div class="footer-links-wrapper">
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/">PyTorch</a></li>
<li><a href="https://pytorch.org/get-started">Get Started</a></li>
<li><a href="https://pytorch.org/features">Features</a></li>
<li><a href="https://pytorch.org/ecosystem">Ecosystem</a></li>
<li><a href="https://pytorch.org/blog/">Blog</a></li>
<li><a href="https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md">Contributing</a></li>
</ul>
</div>
<div class="footer-links-col">
<ul>
<li class="list-title"><a href="https://pytorch.org/resources">Resources</a></li>
<li><a href="https://pytorch.org/tutorials">Tutorials</a></li>
<li><a href="https://pytorch.org/docs/stable/index.html">Docs</a></li>
<li><a href="https://discuss.pytorch.org" target="_blank">Discuss</a></li>
<li><a href="https://github.com/pytorch/pytorch/issues" target="_blank">Github Issues</a></li>
<li><a href="https://pytorch.org/assets/brand-guidelines/PyTorch-Brand-Guidelines.pdf" target="_blank">Brand Guidelines</a></li>
</ul>
</div>
<div class="footer-links-col follow-us-col">
<ul>
<li class="list-title">Stay Connected</li>
<li>
<div id="mc_embed_signup">
<form
action="https://twitter.us14.list-manage.com/subscribe/post?u=75419c71fe0a935e53dfa4a3f&id=91d0dccd39"
method="post"
id="mc-embedded-subscribe-form"
name="mc-embedded-subscribe-form"
class="email-subscribe-form validate"
target="_blank"
novalidate>
<div id="mc_embed_signup_scroll" class="email-subscribe-form-fields-wrapper">
<div class="mc-field-group">
<label for="mce-EMAIL" style="display:none;">Email Address</label>
<input type="email" value="" name="EMAIL" class="required email" id="mce-EMAIL" placeholder="Email Address">
</div>
<div id="mce-responses" class="clear">
<div class="response" id="mce-error-response" style="display:none"></div>
<div class="response" id="mce-success-response" style="display:none"></div>
</div> <!-- real people should not fill this in and expect good things - do not remove this or risk form bot signups-->
<div style="position: absolute; left: -5000px;" aria-hidden="true"><input type="text" name="b_75419c71fe0a935e53dfa4a3f_91d0dccd39" tabindex="-1" value=""></div>
<div class="clear">
<input type="submit" value="" name="subscribe" id="mc-embedded-subscribe" class="button email-subscribe-button">
</div>
</div>
</form>
</div>
</li>
</ul>
<div class="footer-social-icons">
<a href="https://www.facebook.com/pytorch" target="_blank" class="facebook"></a>
<a href="https://twitter.com/pytorch" target="_blank" class="twitter"></a>
<a href="https://www.youtube.com/pytorch" target="_blank" class="youtube"></a>
</div>
</div>
</div>
</div>
</footer>
<div class="cookie-banner-wrapper">
<div class="container">
<p class="gdpr-notice">To analyze traffic and optimize your experience, we serve cookies on this site. By clicking or navigating, you agree to allow our usage of cookies. As the current maintainers of this site, Facebook’s Cookies Policy applies. Learn more, including about available controls: <a href="https://www.facebook.com/policies/cookies/">Cookies Policy</a>.</p>
<img class="close-button" src="../_static/images/pytorch-x.svg">
</div>
</div>
<!-- End Footer -->
<!-- Begin Mobile Menu -->
<div class="mobile-main-menu">
<div class="container-fluid">
<div class="container">
<div class="mobile-main-menu-header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<a class="main-menu-close-button" href="#" data-behavior="close-mobile-menu"></a>
</div>
</div>
</div>