forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathautograd.html
1554 lines (1330 loc) · 118 KB
/
autograd.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
<meta name="robots" content="noindex">
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Automatic differentiation package - torch.autograd — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/autograd.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="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="torch.cuda" href="cuda.html" />
<link rel="prev" title="Tensor Attributes" href="tensor_attributes.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'>1.4.0 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">Notes</span></p>
<ul>
<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/distributed_autograd.html">Distributed Autograd Design</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/faq.html">Frequently Asked Questions</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/multiprocessing.html">Multiprocessing best practices</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/rref.html">Remote Reference Protocol</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>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">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 external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
</ul>
<p class="caption"><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 current"><a class="current reference internal" href="#">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="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">torchvision Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="torchvision/index.html">torchvision</a></li>
</ul>
<p class="caption"><span class="caption-text">torchaudio Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio">torchaudio</a></li>
</ul>
<p class="caption"><span class="caption-text">torchtext Reference</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text">torchtext</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 differentiation package - torch.autograd</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/autograd.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="module-torch.autograd">
<span id="automatic-differentiation-package-torch-autograd"></span><h1>Automatic differentiation package - torch.autograd<a class="headerlink" href="#module-torch.autograd" title="Permalink to this headline">¶</a></h1>
<p><code class="docutils literal notranslate"><span class="pre">torch.autograd</span></code> provides classes and functions implementing automatic
differentiation of arbitrary scalar valued functions. It requires minimal
changes to the existing code - you only need to declare <code class="xref py py-class docutils literal notranslate"><span class="pre">Tensor</span></code> s
for which gradients should be computed with the <code class="docutils literal notranslate"><span class="pre">requires_grad=True</span></code> keyword.</p>
<dl class="function">
<dt id="torch.autograd.backward">
<code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">backward</code><span class="sig-paren">(</span><em class="sig-param">tensors</em>, <em class="sig-param">grad_tensors=None</em>, <em class="sig-param">retain_graph=None</em>, <em class="sig-param">create_graph=False</em>, <em class="sig-param">grad_variables=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd.html#backward"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.backward" title="Permalink to this definition">¶</a></dt>
<dd><p>Computes the sum of gradients of given tensors w.r.t. graph leaves.</p>
<p>The graph is differentiated using the chain rule. If any of <code class="docutils literal notranslate"><span class="pre">tensors</span></code>
are non-scalar (i.e. their data has more than one element) and require
gradient, then the Jacobian-vector product would be computed, in this
case the function additionally requires specifying <code class="docutils literal notranslate"><span class="pre">grad_tensors</span></code>.
It should be a sequence of matching length, that contains the “vector”
in the Jacobian-vector product, usually the gradient of the differentiated
function w.r.t. corresponding tensors (<code class="docutils literal notranslate"><span class="pre">None</span></code> is an acceptable value for
all tensors that don’t need gradient tensors).</p>
<p>This function accumulates gradients in the leaves - you might need to zero
them before calling it.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>tensors</strong> (<em>sequence of Tensor</em>) – Tensors of which the derivative will be
computed.</p></li>
<li><p><strong>grad_tensors</strong> (<em>sequence of</em><em> (</em><a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a><em> or </em><em>None</em><em>)</em>) – The “vector” in the Jacobian-vector
product, usually gradients w.r.t. each element of corresponding tensors.
None values can be specified for scalar Tensors or ones that don’t require
grad. If a None value would be acceptable for all grad_tensors, then this
argument is optional.</p></li>
<li><p><strong>retain_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">False</span></code>, the graph used to compute the grad
will be freed. Note that in nearly all cases setting this option to <code class="docutils literal notranslate"><span class="pre">True</span></code>
is not needed and often can be worked around in a much more efficient
way. Defaults to the value of <code class="docutils literal notranslate"><span class="pre">create_graph</span></code>.</p></li>
<li><p><strong>create_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, graph of the derivative will
be constructed, allowing to compute higher order derivative products.
Defaults to <code class="docutils literal notranslate"><span class="pre">False</span></code>.</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.autograd.grad">
<code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">grad</code><span class="sig-paren">(</span><em class="sig-param">outputs</em>, <em class="sig-param">inputs</em>, <em class="sig-param">grad_outputs=None</em>, <em class="sig-param">retain_graph=None</em>, <em class="sig-param">create_graph=False</em>, <em class="sig-param">only_inputs=True</em>, <em class="sig-param">allow_unused=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd.html#grad"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.grad" title="Permalink to this definition">¶</a></dt>
<dd><p>Computes and returns the sum of gradients of outputs w.r.t. the inputs.</p>
<p><code class="docutils literal notranslate"><span class="pre">grad_outputs</span></code> should be a sequence of length matching <code class="docutils literal notranslate"><span class="pre">output</span></code>
containing the “vector” in Jacobian-vector product, usually the pre-computed
gradients w.r.t. each of the outputs. If an output doesn’t require_grad,
then the gradient can be <code class="docutils literal notranslate"><span class="pre">None</span></code>).</p>
<p>If <code class="docutils literal notranslate"><span class="pre">only_inputs</span></code> is <code class="docutils literal notranslate"><span class="pre">True</span></code>, the function will only return a list of gradients
w.r.t the specified inputs. If it’s <code class="docutils literal notranslate"><span class="pre">False</span></code>, then gradient w.r.t. all remaining
leaves will still be computed, and will be accumulated into their <code class="docutils literal notranslate"><span class="pre">.grad</span></code>
attribute.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>outputs</strong> (<em>sequence of Tensor</em>) – outputs of the differentiated function.</p></li>
<li><p><strong>inputs</strong> (<em>sequence of Tensor</em>) – Inputs w.r.t. which the gradient will be
returned (and not accumulated into <code class="docutils literal notranslate"><span class="pre">.grad</span></code>).</p></li>
<li><p><strong>grad_outputs</strong> (<em>sequence of Tensor</em>) – The “vector” in the Jacobian-vector product.
Usually gradients w.r.t. each output. None values can be specified for scalar
Tensors or ones that don’t require grad. If a None value would be acceptable
for all grad_tensors, then this argument is optional. Default: None.</p></li>
<li><p><strong>retain_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">False</span></code>, the graph used to compute the grad
will be freed. Note that in nearly all cases setting this option to <code class="docutils literal notranslate"><span class="pre">True</span></code>
is not needed and often can be worked around in a much more efficient
way. Defaults to the value of <code class="docutils literal notranslate"><span class="pre">create_graph</span></code>.</p></li>
<li><p><strong>create_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, graph of the derivative will
be constructed, allowing to compute higher order derivative products.
Default: <code class="docutils literal notranslate"><span class="pre">False</span></code>.</p></li>
<li><p><strong>allow_unused</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">False</span></code>, specifying inputs that were not
used when computing outputs (and therefore their grad is always zero)
is an error. Defaults to <code class="docutils literal notranslate"><span class="pre">False</span></code>.</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<div class="section" id="locally-disabling-gradient-computation">
<span id="locally-disable-grad"></span><h2>Locally disabling gradient computation<a class="headerlink" href="#locally-disabling-gradient-computation" title="Permalink to this headline">¶</a></h2>
<dl class="class">
<dt id="torch.autograd.no_grad">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">no_grad</code><a class="reference internal" href="_modules/torch/autograd/grad_mode.html#no_grad"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.no_grad" title="Permalink to this definition">¶</a></dt>
<dd><p>Context-manager that disabled gradient calculation.</p>
<p>Disabling gradient calculation is useful for inference, when you are sure
that you will not call <code class="xref py py-meth docutils literal notranslate"><span class="pre">Tensor.backward()</span></code>. It will reduce memory
consumption for computations that would otherwise have <cite>requires_grad=True</cite>.</p>
<p>In this mode, the result of every computation will have
<cite>requires_grad=False</cite>, even when the inputs have <cite>requires_grad=True</cite>.</p>
<p>This mode has no effect when using <a class="reference internal" href="#torch.autograd.enable_grad" title="torch.autograd.enable_grad"><code class="xref py py-class docutils literal notranslate"><span class="pre">enable_grad</span></code></a> context manager .</p>
<p>This context manager is thread local; it will not affect computation
in other threads.</p>
<p>Also functions as a decorator.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">1</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="gp">... </span> <span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">False</span>
<span class="gp">>>> </span><span class="nd">@torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">()</span>
<span class="gp">... </span><span class="k">def</span> <span class="nf">doubler</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">return</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">z</span> <span class="o">=</span> <span class="n">doubler</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">z</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">False</span>
</pre></div>
</div>
</dd></dl>
<dl class="class">
<dt id="torch.autograd.enable_grad">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">enable_grad</code><a class="reference internal" href="_modules/torch/autograd/grad_mode.html#enable_grad"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.enable_grad" title="Permalink to this definition">¶</a></dt>
<dd><p>Context-manager that enables gradient calculation.</p>
<p>Enables gradient calculation, if it has been disabled via <a class="reference internal" href="#torch.autograd.no_grad" title="torch.autograd.no_grad"><code class="xref py py-class docutils literal notranslate"><span class="pre">no_grad</span></code></a>
or <a class="reference internal" href="#torch.autograd.set_grad_enabled" title="torch.autograd.set_grad_enabled"><code class="xref py py-class docutils literal notranslate"><span class="pre">set_grad_enabled</span></code></a>.</p>
<p>This context manager is thread local; it will not affect computation
in other threads.</p>
<p>Also functions as a decorator.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">1</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="gp">... </span> <span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">enable_grad</span><span class="p">():</span>
<span class="gp">... </span> <span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">backward</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">x</span><span class="o">.</span><span class="n">grad</span>
<span class="gp">>>> </span><span class="nd">@torch</span><span class="o">.</span><span class="n">enable_grad</span><span class="p">()</span>
<span class="gp">... </span><span class="k">def</span> <span class="nf">doubler</span><span class="p">(</span><span class="n">x</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">return</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">():</span>
<span class="gp">... </span> <span class="n">z</span> <span class="o">=</span> <span class="n">doubler</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">z</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">True</span>
</pre></div>
</div>
</dd></dl>
<dl class="class">
<dt id="torch.autograd.set_grad_enabled">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">set_grad_enabled</code><span class="sig-paren">(</span><em class="sig-param">mode</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/grad_mode.html#set_grad_enabled"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.set_grad_enabled" title="Permalink to this definition">¶</a></dt>
<dd><p>Context-manager that sets gradient calculation to on or off.</p>
<p><code class="docutils literal notranslate"><span class="pre">set_grad_enabled</span></code> will enable or disable grads based on its argument <code class="xref py py-attr docutils literal notranslate"><span class="pre">mode</span></code>.
It can be used as a context-manager or as a function.</p>
<p>When using <a class="reference internal" href="#torch.autograd.enable_grad" title="torch.autograd.enable_grad"><code class="xref py py-class docutils literal notranslate"><span class="pre">enable_grad</span></code></a> context manager, <code class="xref py py-class docutils literal notranslate"><span class="pre">set_grad_enabled(False)</span></code>
has no effect.</p>
<p>This context manager is thread local; it will not affect computation
in other threads.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>mode</strong> (<em>bool</em>) – Flag whether to enable grad (<code class="docutils literal notranslate"><span class="pre">True</span></code>), or disable
(<code class="docutils literal notranslate"><span class="pre">False</span></code>). This can be used to conditionally enable
gradients.</p>
</dd>
</dl>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mi">1</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">is_train</span> <span class="o">=</span> <span class="kc">False</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">set_grad_enabled</span><span class="p">(</span><span class="n">is_train</span><span class="p">):</span>
<span class="gp">... </span> <span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">False</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">set_grad_enabled</span><span class="p">(</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">torch</span><span class="o">.</span><span class="n">set_grad_enabled</span><span class="p">(</span><span class="kc">False</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">*</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">y</span><span class="o">.</span><span class="n">requires_grad</span>
<span class="go">False</span>
</pre></div>
</div>
</dd></dl>
</div>
<div class="section" id="in-place-operations-on-tensors">
<h2>In-place operations on Tensors<a class="headerlink" href="#in-place-operations-on-tensors" title="Permalink to this headline">¶</a></h2>
<p>Supporting in-place operations in autograd is a hard matter, and we discourage
their use in most cases. Autograd’s aggressive buffer freeing and reuse makes
it very efficient and there are very few occasions when in-place operations
actually lower memory usage by any significant amount. Unless you’re operating
under heavy memory pressure, you might never need to use them.</p>
<div class="section" id="in-place-correctness-checks">
<h3>In-place correctness checks<a class="headerlink" href="#in-place-correctness-checks" title="Permalink to this headline">¶</a></h3>
<p>All <code class="xref py py-class docutils literal notranslate"><span class="pre">Tensor</span></code> s keep track of in-place operations applied to them, and
if the implementation detects that a tensor was saved for backward in one of
the functions, but it was modified in-place afterwards, an error will be raised
once backward pass is started. This ensures that if you’re using in-place
functions and not seeing any errors, you can be sure that the computed
gradients are correct.</p>
</div>
</div>
<div class="section" id="variable-deprecated">
<h2>Variable (deprecated)<a class="headerlink" href="#variable-deprecated" title="Permalink to this headline">¶</a></h2>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>The Variable API has been deprecated: Variables are no longer necessary to
use autograd with tensors. Autograd automatically supports Tensors with
<code class="docutils literal notranslate"><span class="pre">requires_grad</span></code> set to <code class="docutils literal notranslate"><span class="pre">True</span></code>. Below please find a quick guide on what
has changed:</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">Variable(tensor)</span></code> and <code class="docutils literal notranslate"><span class="pre">Variable(tensor,</span> <span class="pre">requires_grad)</span></code> still work as expected,
but they return Tensors instead of Variables.</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">var.data</span></code> is the same thing as <code class="docutils literal notranslate"><span class="pre">tensor.data</span></code>.</p></li>
<li><p>Methods such as <code class="docutils literal notranslate"><span class="pre">var.backward(),</span> <span class="pre">var.detach(),</span> <span class="pre">var.register_hook()</span></code> now work on tensors
with the same method names.</p></li>
</ul>
<p>In addition, one can now create tensors with <code class="docutils literal notranslate"><span class="pre">requires_grad=True</span></code> using factory
methods such as <a class="reference internal" href="torch.html#torch.randn" title="torch.randn"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.randn()</span></code></a>, <a class="reference internal" href="torch.html#torch.zeros" title="torch.zeros"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.zeros()</span></code></a>, <a class="reference internal" href="torch.html#torch.ones" title="torch.ones"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.ones()</span></code></a>, and others
like the following:</p>
<p><code class="docutils literal notranslate"><span class="pre">autograd_tensor</span> <span class="pre">=</span> <span class="pre">torch.randn((2,</span> <span class="pre">3,</span> <span class="pre">4),</span> <span class="pre">requires_grad=True)</span></code></p>
</div>
</div>
<div class="section" id="tensor-autograd-functions">
<h2>Tensor autograd functions<a class="headerlink" href="#tensor-autograd-functions" title="Permalink to this headline">¶</a></h2>
<dl class="class">
<dt>
<em class="property">class </em><code class="sig-prename descclassname">torch.</code><code class="sig-name descname">Tensor</code></dt>
<dd><dl class="attribute">
<dt id="torch.Tensor.grad">
<code class="sig-name descname">grad</code><a class="headerlink" href="#torch.Tensor.grad" title="Permalink to this definition">¶</a></dt>
<dd><p>This attribute is <code class="docutils literal notranslate"><span class="pre">None</span></code> by default and becomes a Tensor the first time a call to
<a class="reference internal" href="#torch.Tensor.backward" title="torch.Tensor.backward"><code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code></a> computes gradients for <code class="docutils literal notranslate"><span class="pre">self</span></code>.
The attribute will then contain the gradients computed and future calls to
<a class="reference internal" href="#torch.Tensor.backward" title="torch.Tensor.backward"><code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code></a> will accumulate (add) gradients into it.</p>
</dd></dl>
<dl class="attribute">
<dt id="torch.Tensor.requires_grad">
<code class="sig-name descname">requires_grad</code><a class="headerlink" href="#torch.Tensor.requires_grad" title="Permalink to this definition">¶</a></dt>
<dd><p>Is <code class="docutils literal notranslate"><span class="pre">True</span></code> if gradients need to be computed for this Tensor, <code class="docutils literal notranslate"><span class="pre">False</span></code> otherwise.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The fact that gradients need to be computed for a Tensor do not mean that the <a class="reference internal" href="#torch.Tensor.grad" title="torch.Tensor.grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">grad</span></code></a>
attribute will be populated, see <a class="reference internal" href="#torch.Tensor.is_leaf" title="torch.Tensor.is_leaf"><code class="xref py py-attr docutils literal notranslate"><span class="pre">is_leaf</span></code></a> for more details.</p>
</div>
</dd></dl>
<dl class="attribute">
<dt id="torch.Tensor.is_leaf">
<code class="sig-name descname">is_leaf</code><a class="headerlink" href="#torch.Tensor.is_leaf" title="Permalink to this definition">¶</a></dt>
<dd><p>All Tensors that have <a class="reference internal" href="#torch.Tensor.requires_grad" title="torch.Tensor.requires_grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">requires_grad</span></code></a> which is <code class="docutils literal notranslate"><span class="pre">False</span></code> will be leaf Tensors by convention.</p>
<p>For Tensors that have <a class="reference internal" href="#torch.Tensor.requires_grad" title="torch.Tensor.requires_grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">requires_grad</span></code></a> which is <code class="docutils literal notranslate"><span class="pre">True</span></code>, they will be leaf Tensors if they were
created by the user. This means that they are not the result of an operation and so
<code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_fn</span></code> is None.</p>
<p>Only leaf Tensors will have their <a class="reference internal" href="#torch.Tensor.grad" title="torch.Tensor.grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">grad</span></code></a> populated during a call to <a class="reference internal" href="#torch.Tensor.backward" title="torch.Tensor.backward"><code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code></a>.
To get <a class="reference internal" href="#torch.Tensor.grad" title="torch.Tensor.grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">grad</span></code></a> populated for non-leaf Tensors, you can use <a class="reference internal" href="#torch.Tensor.retain_grad" title="torch.Tensor.retain_grad"><code class="xref py py-func docutils literal notranslate"><span class="pre">retain_grad()</span></code></a>.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">a</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">True</span>
<span class="gp">>>> </span><span class="n">b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">b</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">False</span>
<span class="go"># b was created by the operation that cast a cpu Tensor into a cuda Tensor</span>
<span class="gp">>>> </span><span class="n">c</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span> <span class="o">+</span> <span class="mi">2</span>
<span class="gp">>>> </span><span class="n">c</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">False</span>
<span class="go"># c was created by the addition operation</span>
<span class="gp">>>> </span><span class="n">d</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">d</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">True</span>
<span class="go"># d does not require gradients and so has no operation creating it (that is tracked by the autograd engine)</span>
<span class="gp">>>> </span><span class="n">e</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">)</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span><span class="o">.</span><span class="n">requires_grad_</span><span class="p">()</span>
<span class="gp">>>> </span><span class="n">e</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">True</span>
<span class="go"># e requires gradients and has no operations creating it</span>
<span class="gp">>>> </span><span class="n">f</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">rand</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="s2">"cuda"</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">f</span><span class="o">.</span><span class="n">is_leaf</span>
<span class="go">True</span>
<span class="go"># f requires grad, has no operation creating it</span>
</pre></div>
</div>
</dd></dl>
<dl class="method">
<dt id="torch.Tensor.backward">
<code class="sig-name descname">backward</code><span class="sig-paren">(</span><em class="sig-param">gradient=None</em>, <em class="sig-param">retain_graph=None</em>, <em class="sig-param">create_graph=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/tensor.html#Tensor.backward"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.Tensor.backward" title="Permalink to this definition">¶</a></dt>
<dd><p>Computes the gradient of current tensor w.r.t. graph leaves.</p>
<p>The graph is differentiated using the chain rule. If the tensor is
non-scalar (i.e. its data has more than one element) and requires
gradient, the function additionally requires specifying <code class="docutils literal notranslate"><span class="pre">gradient</span></code>.
It should be a tensor of matching type and location, that contains
the gradient of the differentiated function w.r.t. <code class="docutils literal notranslate"><span class="pre">self</span></code>.</p>
<p>This function accumulates gradients in the leaves - you might need to
zero them before calling it.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>gradient</strong> (<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a><em> or </em><em>None</em>) – Gradient w.r.t. the
tensor. If it is a tensor, it will be automatically converted
to a Tensor that does not require grad unless <code class="docutils literal notranslate"><span class="pre">create_graph</span></code> is True.
None values can be specified for scalar Tensors or ones that
don’t require grad. If a None value would be acceptable then
this argument is optional.</p></li>
<li><p><strong>retain_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">False</span></code>, the graph used to compute
the grads will be freed. Note that in nearly all cases setting
this option to True is not needed and often can be worked around
in a much more efficient way. Defaults to the value of
<code class="docutils literal notranslate"><span class="pre">create_graph</span></code>.</p></li>
<li><p><strong>create_graph</strong> (<em>bool</em><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, graph of the derivative will
be constructed, allowing to compute higher order derivative
products. Defaults to <code class="docutils literal notranslate"><span class="pre">False</span></code>.</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="method">
<dt id="torch.Tensor.detach">
<code class="sig-name descname">detach</code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="headerlink" href="#torch.Tensor.detach" title="Permalink to this definition">¶</a></dt>
<dd><p>Returns a new Tensor, detached from the current graph.</p>
<p>The result will never require gradient.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Returned Tensor shares the same storage with the original one.
In-place modifications on either of them will be seen, and may trigger
errors in correctness checks.
IMPORTANT NOTE: Previously, in-place size / stride / storage changes
(such as <cite>resize_</cite> / <cite>resize_as_</cite> / <cite>set_</cite> / <cite>transpose_</cite>) to the returned tensor
also update the original tensor. Now, these in-place changes will not update the
original tensor anymore, and will instead trigger an error.
For sparse tensors:
In-place indices / values changes (such as <cite>zero_</cite> / <cite>copy_</cite> / <cite>add_</cite>) to the
returned tensor will not update the original tensor anymore, and will instead
trigger an error.</p>
</div>
</dd></dl>
<dl class="method">
<dt id="torch.Tensor.detach_">
<code class="sig-name descname">detach_</code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="headerlink" href="#torch.Tensor.detach_" title="Permalink to this definition">¶</a></dt>
<dd><p>Detaches the Tensor from the graph that created it, making it a leaf.
Views cannot be detached in-place.</p>
</dd></dl>
<dl class="method">
<dt id="torch.Tensor.register_hook">
<code class="sig-name descname">register_hook</code><span class="sig-paren">(</span><em class="sig-param">hook</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/tensor.html#Tensor.register_hook"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.Tensor.register_hook" title="Permalink to this definition">¶</a></dt>
<dd><p>Registers a backward hook.</p>
<p>The hook will be called every time a gradient with respect to the
Tensor is computed. The hook should have the following signature:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">hook</span><span class="p">(</span><span class="n">grad</span><span class="p">)</span> <span class="o">-></span> <span class="n">Tensor</span> <span class="ow">or</span> <span class="kc">None</span>
</pre></div>
</div>
<p>The hook should not modify its argument, but it can optionally return
a new gradient which will be used in place of <a class="reference internal" href="#torch.Tensor.grad" title="torch.Tensor.grad"><code class="xref py py-attr docutils literal notranslate"><span class="pre">grad</span></code></a>.</p>
<p>This function returns a handle with a method <code class="docutils literal notranslate"><span class="pre">handle.remove()</span></code>
that removes the hook from the module.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">v</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mf">0.</span><span class="p">,</span> <span class="mf">0.</span><span class="p">,</span> <span class="mf">0.</span><span class="p">],</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">h</span> <span class="o">=</span> <span class="n">v</span><span class="o">.</span><span class="n">register_hook</span><span class="p">(</span><span class="k">lambda</span> <span class="n">grad</span><span class="p">:</span> <span class="n">grad</span> <span class="o">*</span> <span class="mi">2</span><span class="p">)</span> <span class="c1"># double the gradient</span>
<span class="gp">>>> </span><span class="n">v</span><span class="o">.</span><span class="n">backward</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">([</span><span class="mf">1.</span><span class="p">,</span> <span class="mf">2.</span><span class="p">,</span> <span class="mf">3.</span><span class="p">]))</span>
<span class="gp">>>> </span><span class="n">v</span><span class="o">.</span><span class="n">grad</span>
<span class="go"> 2</span>
<span class="go"> 4</span>
<span class="go"> 6</span>
<span class="go">[torch.FloatTensor of size (3,)]</span>
<span class="gp">>>> </span><span class="n">h</span><span class="o">.</span><span class="n">remove</span><span class="p">()</span> <span class="c1"># removes the hook</span>
</pre></div>
</div>
</dd></dl>
<dl class="method">
<dt id="torch.Tensor.retain_grad">
<code class="sig-name descname">retain_grad</code><span class="sig-paren">(</span><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/tensor.html#Tensor.retain_grad"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.Tensor.retain_grad" title="Permalink to this definition">¶</a></dt>
<dd><p>Enables .grad attribute for non-leaf Tensors.</p>
</dd></dl>
</dd></dl>
</div>
<div class="section" id="function">
<h2><span class="hidden-section">Function</span><a class="headerlink" href="#function" title="Permalink to this headline">¶</a></h2>
<dl class="class">
<dt id="torch.autograd.Function">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">Function</code><a class="reference internal" href="_modules/torch/autograd/function.html#Function"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.Function" title="Permalink to this definition">¶</a></dt>
<dd><p>Records operation history and defines formulas for differentiating ops.</p>
<p>Every operation performed on <code class="xref py py-class docutils literal notranslate"><span class="pre">Tensor</span></code> s creates a new function
object, that performs the computation, and records that it happened.
The history is retained in the form of a DAG of functions, with edges
denoting data dependencies (<code class="docutils literal notranslate"><span class="pre">input</span> <span class="pre"><-</span> <span class="pre">output</span></code>). Then, when backward is
called, the graph is processed in the topological ordering, by calling
<a class="reference internal" href="#torch.autograd.backward" title="torch.autograd.backward"><code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code></a> methods of each <a class="reference internal" href="#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">Function</span></code></a> object, and passing
returned gradients on to next <a class="reference internal" href="#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">Function</span></code></a> s.</p>
<p>Normally, the only way users interact with functions is by creating
subclasses and defining new operations. This is a recommended way of
extending torch.autograd.</p>
<p>Each function object is meant to be used only once (in the forward pass).</p>
<p>Examples:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="k">class</span> <span class="nc">Exp</span><span class="p">(</span><span class="n">Function</span><span class="p">):</span>
<span class="go">>>></span>
<span class="gp">>>> </span> <span class="nd">@staticmethod</span>
<span class="gp">>>> </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">i</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="n">result</span> <span class="o">=</span> <span class="n">i</span><span class="o">.</span><span class="n">exp</span><span class="p">()</span>
<span class="gp">>>> </span> <span class="n">ctx</span><span class="o">.</span><span class="n">save_for_backward</span><span class="p">(</span><span class="n">result</span><span class="p">)</span>
<span class="gp">>>> </span> <span class="k">return</span> <span class="n">result</span>
<span class="go">>>></span>
<span class="gp">>>> </span> <span class="nd">@staticmethod</span>
<span class="gp">>>> </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_output</span><span class="p">):</span>
<span class="gp">>>> </span> <span class="n">result</span><span class="p">,</span> <span class="o">=</span> <span class="n">ctx</span><span class="o">.</span><span class="n">saved_tensors</span>
<span class="gp">>>> </span> <span class="k">return</span> <span class="n">grad_output</span> <span class="o">*</span> <span class="n">result</span>
</pre></div>
</div>
<dl class="method">
<dt id="torch.autograd.Function.backward">
<em class="property">static </em><code class="sig-name descname">backward</code><span class="sig-paren">(</span><em class="sig-param">ctx</em>, <em class="sig-param">*grad_outputs</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/function.html#Function.backward"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.Function.backward" title="Permalink to this definition">¶</a></dt>
<dd><p>Defines a formula for differentiating the operation.</p>
<p>This function is to be overridden by all subclasses.</p>
<p>It must accept a context <code class="xref py py-attr docutils literal notranslate"><span class="pre">ctx</span></code> as the first argument, followed by
as many outputs did <a class="reference internal" href="#torch.autograd.Function.forward" title="torch.autograd.Function.forward"><code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code></a> return, and it should return as many
tensors, as there were inputs to <a class="reference internal" href="#torch.autograd.Function.forward" title="torch.autograd.Function.forward"><code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code></a>. Each argument is the
gradient w.r.t the given output, and each returned value should be the
gradient w.r.t. the corresponding input.</p>
<p>The context can be used to retrieve tensors saved during the forward
pass. It also has an attribute <code class="xref py py-attr docutils literal notranslate"><span class="pre">ctx.needs_input_grad</span></code> as a tuple
of booleans representing whether each input needs gradient. E.g.,
<a class="reference internal" href="#torch.autograd.backward" title="torch.autograd.backward"><code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code></a> will have <code class="docutils literal notranslate"><span class="pre">ctx.needs_input_grad[0]</span> <span class="pre">=</span> <span class="pre">True</span></code> if the
first input to <a class="reference internal" href="#torch.autograd.Function.forward" title="torch.autograd.Function.forward"><code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code></a> needs gradient computated w.r.t. the
output.</p>
</dd></dl>
<dl class="method">
<dt id="torch.autograd.Function.forward">
<em class="property">static </em><code class="sig-name descname">forward</code><span class="sig-paren">(</span><em class="sig-param">ctx</em>, <em class="sig-param">*args</em>, <em class="sig-param">**kwargs</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/function.html#Function.forward"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.Function.forward" title="Permalink to this definition">¶</a></dt>
<dd><p>Performs the operation.</p>
<p>This function is to be overridden by all subclasses.</p>
<p>It must accept a context ctx as the first argument, followed by any
number of arguments (tensors or other types).</p>
<p>The context can be used to store tensors that can be then retrieved
during the backward pass.</p>
</dd></dl>
</dd></dl>
</div>
<div class="section" id="context-method-mixins">
<h2>Context method mixins<a class="headerlink" href="#context-method-mixins" title="Permalink to this headline">¶</a></h2>
<p>When creating a new <a class="reference internal" href="#torch.autograd.Function" title="torch.autograd.Function"><code class="xref py py-class docutils literal notranslate"><span class="pre">Function</span></code></a>, the following methods are available to <cite>ctx</cite>.</p>
<dl class="class">
<dt id="torch.autograd.function._ContextMethodMixin">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.function.</code><code class="sig-name descname">_ContextMethodMixin</code><a class="reference internal" href="_modules/torch/autograd/function.html#_ContextMethodMixin"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.function._ContextMethodMixin" title="Permalink to this definition">¶</a></dt>
<dd><dl class="method">
<dt id="torch.autograd.function._ContextMethodMixin.mark_dirty">
<code class="sig-name descname">mark_dirty</code><span class="sig-paren">(</span><em class="sig-param">*args</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/function.html#_ContextMethodMixin.mark_dirty"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.function._ContextMethodMixin.mark_dirty" title="Permalink to this definition">¶</a></dt>
<dd><p>Marks given tensors as modified in an in-place operation.</p>
<p><strong>This should be called at most once, only from inside the</strong>
<code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code> <strong>method, and all arguments should be inputs.</strong></p>
<p>Every tensor that’s been modified in-place in a call to <code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code>
should be given to this function, to ensure correctness of our checks.
It doesn’t matter whether the function is called before or after
modification.</p>
</dd></dl>
<dl class="method">
<dt id="torch.autograd.function._ContextMethodMixin.mark_non_differentiable">
<code class="sig-name descname">mark_non_differentiable</code><span class="sig-paren">(</span><em class="sig-param">*args</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/function.html#_ContextMethodMixin.mark_non_differentiable"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.function._ContextMethodMixin.mark_non_differentiable" title="Permalink to this definition">¶</a></dt>
<dd><p>Marks outputs as non-differentiable.</p>
<p><strong>This should be called at most once, only from inside the</strong>
<code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code> <strong>method, and all arguments should be outputs.</strong></p>
<p>This will mark outputs as not requiring gradients, increasing the
efficiency of backward computation. You still need to accept a gradient
for each output in <code class="xref py py-meth docutils literal notranslate"><span class="pre">backward()</span></code>, but it’s always going to
be a zero tensor with the same shape as the shape of a corresponding
output.</p>
<p>This is used e.g. for indices returned from a max <code class="xref py py-class docutils literal notranslate"><span class="pre">Function</span></code>.</p>
</dd></dl>
<dl class="method">
<dt id="torch.autograd.function._ContextMethodMixin.save_for_backward">
<code class="sig-name descname">save_for_backward</code><span class="sig-paren">(</span><em class="sig-param">*tensors</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/function.html#_ContextMethodMixin.save_for_backward"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.function._ContextMethodMixin.save_for_backward" title="Permalink to this definition">¶</a></dt>
<dd><p>Saves given tensors for a future call to <code class="xref py py-func docutils literal notranslate"><span class="pre">backward()</span></code>.</p>
<p><strong>This should be called at most once, and only from inside the</strong>
<code class="xref py py-func docutils literal notranslate"><span class="pre">forward()</span></code> <strong>method.</strong></p>
<p>Later, saved tensors can be accessed through the <code class="xref py py-attr docutils literal notranslate"><span class="pre">saved_tensors</span></code>
attribute. Before returning them to the user, a check is made to ensure
they weren’t used in any in-place operation that modified their content.</p>
<p>Arguments can also be <code class="docutils literal notranslate"><span class="pre">None</span></code>.</p>
</dd></dl>
</dd></dl>
</div>
<div class="section" id="numerical-gradient-checking">
<span id="grad-check"></span><h2>Numerical gradient checking<a class="headerlink" href="#numerical-gradient-checking" title="Permalink to this headline">¶</a></h2>
<dl class="function">
<dt id="torch.autograd.gradcheck">
<code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">gradcheck</code><span class="sig-paren">(</span><em class="sig-param">func</em>, <em class="sig-param">inputs</em>, <em class="sig-param">eps=1e-06</em>, <em class="sig-param">atol=1e-05</em>, <em class="sig-param">rtol=0.001</em>, <em class="sig-param">raise_exception=True</em>, <em class="sig-param">check_sparse_nnz=False</em>, <em class="sig-param">nondet_tol=0.0</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/gradcheck.html#gradcheck"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.gradcheck" title="Permalink to this definition">¶</a></dt>
<dd><p>Check gradients computed via small finite differences against analytical
gradients w.r.t. tensors in <code class="xref py py-attr docutils literal notranslate"><span class="pre">inputs</span></code> that are of floating point type
and with <code class="docutils literal notranslate"><span class="pre">requires_grad=True</span></code>.</p>
<p>The check between numerical and analytical gradients uses <a class="reference internal" href="torch.html#torch.allclose" title="torch.allclose"><code class="xref py py-func docutils literal notranslate"><span class="pre">allclose()</span></code></a>.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The default values are designed for <code class="xref py py-attr docutils literal notranslate"><span class="pre">input</span></code> of double precision.
This check will likely fail if <code class="xref py py-attr docutils literal notranslate"><span class="pre">input</span></code> is of less precision, e.g.,
<code class="docutils literal notranslate"><span class="pre">FloatTensor</span></code>.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If any checked tensor in <code class="xref py py-attr docutils literal notranslate"><span class="pre">input</span></code> has overlapping memory, i.e.,
different indices pointing to the same memory address (e.g., from
<code class="xref py py-func docutils literal notranslate"><span class="pre">torch.expand()</span></code>), this check will likely fail because the numerical
gradients computed by point perturbation at such indices will change
values at all other indices that share the same memory address.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>func</strong> (<em>function</em>) – a Python function that takes Tensor inputs and returns
a Tensor or a tuple of Tensors</p></li>
<li><p><strong>inputs</strong> (<em>tuple of Tensor</em><em> or </em><a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a>) – inputs to the function</p></li>
<li><p><strong>eps</strong> (<em>python:float</em><em>, </em><em>optional</em>) – perturbation for finite differences</p></li>
<li><p><strong>atol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – absolute tolerance</p></li>
<li><p><strong>rtol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – relative tolerance</p></li>
<li><p><strong>raise_exception</strong> (<em>bool</em><em>, </em><em>optional</em>) – indicating whether to raise an exception if
the check fails. The exception gives more information about the
exact nature of the failure. This is helpful when debugging gradchecks.</p></li>
<li><p><strong>check_sparse_nnz</strong> (<em>bool</em><em>, </em><em>optional</em>) – if True, gradcheck allows for SparseTensor input,
and for any SparseTensor at input, gradcheck will perform check at nnz positions only.</p></li>
<li><p><strong>nondet_tol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – tolerance for non-determinism. When running
identical inputs through the differentiation, the results must either match
exactly (default, 0.0) or be within this tolerance.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>True if all differences satisfy allclose condition</p>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.autograd.gradgradcheck">
<code class="sig-prename descclassname">torch.autograd.</code><code class="sig-name descname">gradgradcheck</code><span class="sig-paren">(</span><em class="sig-param">func</em>, <em class="sig-param">inputs</em>, <em class="sig-param">grad_outputs=None</em>, <em class="sig-param">eps=1e-06</em>, <em class="sig-param">atol=1e-05</em>, <em class="sig-param">rtol=0.001</em>, <em class="sig-param">gen_non_contig_grad_outputs=False</em>, <em class="sig-param">raise_exception=True</em>, <em class="sig-param">nondet_tol=0.0</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/gradcheck.html#gradgradcheck"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.gradgradcheck" title="Permalink to this definition">¶</a></dt>
<dd><p>Check gradients of gradients computed via small finite differences
against analytical gradients w.r.t. tensors in <code class="xref py py-attr docutils literal notranslate"><span class="pre">inputs</span></code> and
<code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_outputs</span></code> that are of floating point type and with
<code class="docutils literal notranslate"><span class="pre">requires_grad=True</span></code>.</p>
<p>This function checks that backpropagating through the gradients computed
to the given <code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_outputs</span></code> are correct.</p>
<p>The check between numerical and analytical gradients uses <a class="reference internal" href="torch.html#torch.allclose" title="torch.allclose"><code class="xref py py-func docutils literal notranslate"><span class="pre">allclose()</span></code></a>.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The default values are designed for <code class="xref py py-attr docutils literal notranslate"><span class="pre">input</span></code> and
<code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_outputs</span></code> of double precision. This check will likely fail if
they are of less precision, e.g., <code class="docutils literal notranslate"><span class="pre">FloatTensor</span></code>.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If any checked tensor in <code class="xref py py-attr docutils literal notranslate"><span class="pre">input</span></code> and <code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_outputs</span></code> has
overlapping memory, i.e., different indices pointing to the same memory
address (e.g., from <code class="xref py py-func docutils literal notranslate"><span class="pre">torch.expand()</span></code>), this check will likely fail
because the numerical gradients computed by point perturbation at such
indices will change values at all other indices that share the same
memory address.</p>
</div>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>func</strong> (<em>function</em>) – a Python function that takes Tensor inputs and returns
a Tensor or a tuple of Tensors</p></li>
<li><p><strong>inputs</strong> (<em>tuple of Tensor</em><em> or </em><a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a>) – inputs to the function</p></li>
<li><p><strong>grad_outputs</strong> (<em>tuple of Tensor</em><em> or </em><a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a><em>, </em><em>optional</em>) – The gradients with
respect to the function’s outputs.</p></li>
<li><p><strong>eps</strong> (<em>python:float</em><em>, </em><em>optional</em>) – perturbation for finite differences</p></li>
<li><p><strong>atol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – absolute tolerance</p></li>
<li><p><strong>rtol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – relative tolerance</p></li>
<li><p><strong>gen_non_contig_grad_outputs</strong> (<em>bool</em><em>, </em><em>optional</em>) – if <code class="xref py py-attr docutils literal notranslate"><span class="pre">grad_outputs</span></code> is
<code class="docutils literal notranslate"><span class="pre">None</span></code> and <code class="xref py py-attr docutils literal notranslate"><span class="pre">gen_non_contig_grad_outputs</span></code> is <code class="docutils literal notranslate"><span class="pre">True</span></code>, the
randomly generated gradient outputs are made to be noncontiguous</p></li>
<li><p><strong>raise_exception</strong> (<em>bool</em><em>, </em><em>optional</em>) – indicating whether to raise an exception if
the check fails. The exception gives more information about the
exact nature of the failure. This is helpful when debugging gradchecks.</p></li>
<li><p><strong>nondet_tol</strong> (<em>python:float</em><em>, </em><em>optional</em>) – tolerance for non-determinism. When running
identical inputs through the differentiation, the results must either match
exactly (default, 0.0) or be within this tolerance. Note that a small amount
of nondeterminism in the gradient will lead to larger inaccuracies in
the second derivative.</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>True if all differences satisfy allclose condition</p>
</dd>
</dl>
</dd></dl>
</div>
<div class="section" id="profiler">
<h2>Profiler<a class="headerlink" href="#profiler" title="Permalink to this headline">¶</a></h2>
<p>Autograd includes a profiler that lets you inspect the cost of different
operators inside your model - both on the CPU and GPU. There are two modes
implemented at the moment - CPU-only using <a class="reference internal" href="#torch.autograd.profiler.profile" title="torch.autograd.profiler.profile"><code class="xref py py-class docutils literal notranslate"><span class="pre">profile</span></code></a>.
and nvprof based (registers both CPU and GPU activity) using
<a class="reference internal" href="#torch.autograd.profiler.emit_nvtx" title="torch.autograd.profiler.emit_nvtx"><code class="xref py py-class docutils literal notranslate"><span class="pre">emit_nvtx</span></code></a>.</p>
<dl class="class">
<dt id="torch.autograd.profiler.profile">
<em class="property">class </em><code class="sig-prename descclassname">torch.autograd.profiler.</code><code class="sig-name descname">profile</code><span class="sig-paren">(</span><em class="sig-param">enabled=True</em>, <em class="sig-param">use_cuda=False</em>, <em class="sig-param">record_shapes=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/profiler.html#profile"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.profiler.profile" title="Permalink to this definition">¶</a></dt>
<dd><p>Context manager that manages autograd profiler state and holds a summary of results.
Under the hood it just records events of functions being executed in C++ and
exposes those events to Python. You can wrap any code into it and it will
only report runtime of PyTorch functions.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>enabled</strong> (<em>bool</em><em>, </em><em>optional</em>) – Setting this to False makes this context manager a no-op.
Default: <code class="docutils literal notranslate"><span class="pre">True</span></code>.</p></li>
<li><p><strong>use_cuda</strong> (<em>bool</em><em>, </em><em>optional</em>) – Enables timing of CUDA events as well using the cudaEvent API.
Adds approximately 4us of overhead to each tensor operation.
Default: <code class="docutils literal notranslate"><span class="pre">False</span></code></p></li>
<li><p><strong>record_shapes</strong> (<em>bool</em><em>, </em><em>optional</em>) – If shapes recording is set, information
about input dimensions will be collected. This allows one to see which
dimensions have been used under the hood and further group by them
using prof.key_averages(group_by_input_shape=True). Please note that
shape recording might skew your profiling data. It is recommended to
use separate runs with and without shape recording to validate the timing.
Most likely the skew will be negligible for bottom most events (in a case
of nested function calls). But for higher level functions the total
self cpu time might be artificially increased because of the shape
collection.</p></li>
</ul>
</dd>
</dl>
<p class="rubric">Example</p>
<div class="doctest highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">((</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="n">requires_grad</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="gp">>>> </span><span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">autograd</span><span class="o">.</span><span class="n">profiler</span><span class="o">.</span><span class="n">profile</span><span class="p">()</span> <span class="k">as</span> <span class="n">prof</span><span class="p">:</span>
<span class="gp">>>> </span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="mi">100</span><span class="p">):</span> <span class="c1"># any normal python code, really!</span>
<span class="gp">>>> </span> <span class="n">y</span> <span class="o">=</span> <span class="n">x</span> <span class="o">**</span> <span class="mi">2</span>
<span class="go">>> y.backward()</span>
<span class="gp">>>> </span><span class="c1"># NOTE: some columns were removed for brevity</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="n">prof</span><span class="o">.</span><span class="n">key_averages</span><span class="p">()</span><span class="o">.</span><span class="n">table</span><span class="p">(</span><span class="n">sort_by</span><span class="o">=</span><span class="s2">"self_cpu_time_total"</span><span class="p">))</span>
<span class="go">----------------------------------- --------------- --------------- ---------------</span>
<span class="go">Name Self CPU total CPU time avg Number of Calls</span>
<span class="go">----------------------------------- --------------- --------------- ---------------</span>
<span class="go">mul 32.048ms 32.048ms 200</span>
<span class="go">pow 27.041ms 27.041ms 200</span>
<span class="go">PowBackward0 9.727ms 55.483ms 100</span>
<span class="go">torch::autograd::AccumulateGrad 9.148ms 9.148ms 100</span>
<span class="go">torch::autograd::GraphRoot 691.816us 691.816us 100</span>
<span class="go">----------------------------------- --------------- --------------- ---------------</span>
</pre></div>
</div>
<dl class="method">
<dt id="torch.autograd.profiler.profile.export_chrome_trace">
<code class="sig-name descname">export_chrome_trace</code><span class="sig-paren">(</span><em class="sig-param">path</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/profiler.html#profile.export_chrome_trace"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.profiler.profile.export_chrome_trace" title="Permalink to this definition">¶</a></dt>
<dd><p>Exports an EventList as a Chrome tracing tools file.</p>
<p>The checkpoint can be later loaded and inspected under <code class="docutils literal notranslate"><span class="pre">chrome://tracing</span></code> URL.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>path</strong> (<em>str</em>) – Path where the trace will be written.</p>
</dd>
</dl>
</dd></dl>
<dl class="method">
<dt id="torch.autograd.profiler.profile.key_averages">
<code class="sig-name descname">key_averages</code><span class="sig-paren">(</span><em class="sig-param">group_by_input_shape=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/autograd/profiler.html#profile.key_averages"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.autograd.profiler.profile.key_averages" title="Permalink to this definition">¶</a></dt>
<dd><p>Averages all function events over their keys.</p>
<p>@param group_by_input_shapes The key would become
(event name, input dimensions) rather than just event name.
This is useful to see which dimensionality contributes to the runtime
the most and may help with dimension specific optimizations or