forked from pytorch/pytorch.github.io
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdata.html
1577 lines (1359 loc) · 141 KB
/
data.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>torch.utils.data — PyTorch 1.10.1 documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/data.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="torch.utils.dlpack" href="dlpack.html" />
<link rel="prev" title="torch.utils.cpp_extension" href="cpp_extension.html" />
<!-- Google Analytics -->
<script async src="https://www.googletagmanager.com/gtag/js?id=UA-117752657-2"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'UA-117752657-2');
</script>
<!-- End Google Analytics -->
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="stylesheet" href="https://use.fontawesome.com/releases/v5.15.2/css/all.css" integrity="sha384-vSIIfh2YWi9wW0r9iZe7RJPrKwp6bG+s9QZMoITbCckVJqGCCRhc+ccxNcdpHuYu" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<a href="https://pytorch.org/ecosystem">Ecosystem</a>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active docs-active">
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-orange-arrow">
Docs
</a>
<div class="resources-dropdown-menu">
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/docs/stable/index.html">
<span class="dropdown-title">PyTorch</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/audio/stable/index.html">
<span class="dropdown-title">torchaudio</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/text/stable/index.html">
<span class="dropdown-title">torchtext</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/vision/stable/index.html">
<span class="dropdown-title">torchvision</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/elastic/">
<span class="dropdown-title">TorchElastic</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/serve/">
<span class="dropdown-title">TorchServe</span>
<p></p>
</a>
<a class="doc-dropdown-option nav-dropdown-item" href="https://pytorch.org/xla">
<span class="dropdown-title">PyTorch on XLA Devices</span>
<p></p>
</a>
</div>
</li>
<li>
<div id="resourcesDropdownButton" data-toggle="resources-dropdown" class="resources-dropdown">
<a class="resource-option with-down-arrow">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class="dropdown-title">About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/#community-module">
<span class="dropdown-title">Community</span>
<p>Join the PyTorch developer community to contribute, learn, and get your questions answered.</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/resources">
<span class="dropdown-title">Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://discuss.pytorch.org/" target="_blank">
<span class="dropdown-title">Forums</span>
<p>A place to discuss PyTorch code, issues, install, research</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/hub">
<span class="dropdown-title">Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">GitHub</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='https://pytorch.org/docs/versions.html'>1.10.1 ▼</a>
</div>
<div role="search">
<form id="rtd-search-form" class="wy-form" action="search.html" method="get">
<input type="text" name="q" placeholder="Search Docs" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</form>
</div>
</div>
<p class="caption"><span class="caption-text">Notes</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="notes/amp_examples.html">Automatic Mixed Precision examples</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/autograd.html">Autograd mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/broadcasting.html">Broadcasting semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cpu_threading_torchscript_inference.html">CPU threading and TorchScript inference</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/cuda.html">CUDA semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/ddp.html">Distributed Data Parallel</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/extending.html">Extending PyTorch</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/faq.html">Frequently Asked Questions</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/gradcheck.html">Gradcheck mechanics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/hip.html">HIP (ROCm) semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/modules.html">Modules</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/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/serialization.html">Serialization semantics</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/windows.html">Windows FAQ</a></li>
</ul>
<p class="caption"><span class="caption-text">Language Bindings</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="cpp_index.html">C++</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/javadoc/">Javadoc</a></li>
</ul>
<p class="caption"><span class="caption-text">Python API</span></p>
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="torch.html">torch</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.html">torch.nn</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.functional.html">torch.nn.functional</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensors.html">torch.Tensor</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_attributes.html">Tensor Attributes</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensor_view.html">Tensor Views</a></li>
<li class="toctree-l1"><a class="reference internal" href="autograd.html">torch.autograd</a></li>
<li class="toctree-l1"><a class="reference internal" href="cuda.html">torch.cuda</a></li>
<li class="toctree-l1"><a class="reference internal" href="amp.html">torch.cuda.amp</a></li>
<li class="toctree-l1"><a class="reference internal" href="backends.html">torch.backends</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.algorithms.join.html">torch.distributed.algorithms.join</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.elastic.html">torch.distributed.elastic</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributed.optim.html">torch.distributed.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="fft.html">torch.fft</a></li>
<li class="toctree-l1"><a class="reference internal" href="futures.html">torch.futures</a></li>
<li class="toctree-l1"><a class="reference internal" href="fx.html">torch.fx</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="linalg.html">torch.linalg</a></li>
<li class="toctree-l1"><a class="reference internal" href="special.html">torch.special</a></li>
<li class="toctree-l1"><a class="reference internal" href="torch.overrides.html">torch.overrides</a></li>
<li class="toctree-l1"><a class="reference internal" href="package.html">torch.package</a></li>
<li class="toctree-l1"><a class="reference internal" href="profiler.html">torch.profiler</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1"><a class="reference internal" href="complex_numbers.html">Complex Numbers</a></li>
<li class="toctree-l1"><a class="reference internal" href="ddp_comm_hooks.html">DDP Communication Hooks</a></li>
<li class="toctree-l1"><a class="reference internal" href="pipeline.html">Pipeline Parallelism</a></li>
<li class="toctree-l1"><a class="reference internal" href="quantization.html">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="testing.html">torch.testing</a></li>
<li class="toctree-l1"><a class="reference internal" href="benchmark_utils.html">torch.utils.benchmark</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="mobile_optimizer.html">torch.utils.mobile_optimizer</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="__config__.html">torch.__config__</a></li>
</ul>
<p class="caption"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio/stable">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text/stable">torchtext</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/vision/stable">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/serve">TorchServe</a></li>
<li class="toctree-l1"><a class="reference external" href="http://pytorch.org/xla/">PyTorch on XLA Devices</a></li>
</ul>
<p class="caption"><span class="caption-text">Community</span></p>
<ul>
<li class="toctree-l1"><a class="reference internal" href="community/contribution_guide.html">PyTorch Contribution Guide</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/governance.html">PyTorch Governance</a></li>
<li class="toctree-l1"><a class="reference internal" href="community/persons_of_interest.html">PyTorch Governance | Persons of Interest</a></li>
</ul>
</div>
</div>
</nav>
<div class="pytorch-container">
<div class="pytorch-page-level-bar" id="pytorch-page-level-bar">
<div class="pytorch-breadcrumbs-wrapper">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="pytorch-breadcrumbs">
<li>
<a href="index.html">
Docs
</a> >
</li>
<li>torch.utils.data</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/data.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.utils.data">
<span id="torch-utils-data"></span><h1>torch.utils.data<a class="headerlink" href="#module-torch.utils.data" title="Permalink to this headline">¶</a></h1>
<p>At the heart of PyTorch data loading utility is the <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.utils.data.DataLoader</span></code></a>
class. It represents a Python iterable over a dataset, with support for</p>
<ul class="simple">
<li><p><a class="reference internal" href="#dataset-types">map-style and iterable-style datasets</a>,</p></li>
<li><p><a class="reference internal" href="#data-loading-order-and-sampler">customizing data loading order</a>,</p></li>
<li><p><a class="reference internal" href="#loading-batched-and-non-batched-data">automatic batching</a>,</p></li>
<li><p><a class="reference internal" href="#single-and-multi-process-data-loading">single- and multi-process data loading</a>,</p></li>
<li><p><a class="reference internal" href="#memory-pinning">automatic memory pinning</a>.</p></li>
</ul>
<p>These options are configured by the constructor arguments of a
<a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>, which has signature:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">DataLoader</span><span class="p">(</span><span class="n">dataset</span><span class="p">,</span> <span class="n">batch_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">shuffle</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span> <span class="n">sampler</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">batch_sampler</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">collate_fn</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span>
<span class="n">pin_memory</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span> <span class="n">drop_last</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span> <span class="n">timeout</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span>
<span class="n">worker_init_fn</span><span class="o">=</span><span class="kc">None</span><span class="p">,</span> <span class="o">*</span><span class="p">,</span> <span class="n">prefetch_factor</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span>
<span class="n">persistent_workers</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
</pre></div>
</div>
<p>The sections below describe in details the effects and usages of these options.</p>
<div class="section" id="dataset-types">
<h2>Dataset Types<a class="headerlink" href="#dataset-types" title="Permalink to this headline">¶</a></h2>
<p>The most important argument of <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>
constructor is <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code>, which indicates a dataset object to load data
from. PyTorch supports two different types of datasets:</p>
<ul class="simple">
<li><p><a class="reference internal" href="#map-style-datasets">map-style datasets</a>,</p></li>
<li><p><a class="reference internal" href="#iterable-style-datasets">iterable-style datasets</a>.</p></li>
</ul>
<div class="section" id="map-style-datasets">
<h3>Map-style datasets<a class="headerlink" href="#map-style-datasets" title="Permalink to this headline">¶</a></h3>
<p>A map-style dataset is one that implements the <code class="xref py py-meth docutils literal notranslate"><span class="pre">__getitem__()</span></code> and
<code class="xref py py-meth docutils literal notranslate"><span class="pre">__len__()</span></code> protocols, and represents a map from (possibly non-integral)
indices/keys to data samples.</p>
<p>For example, such a dataset, when accessed with <code class="docutils literal notranslate"><span class="pre">dataset[idx]</span></code>, could read
the <code class="docutils literal notranslate"><span class="pre">idx</span></code>-th image and its corresponding label from a folder on the disk.</p>
<p>See <a class="reference internal" href="#torch.utils.data.Dataset" title="torch.utils.data.Dataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">Dataset</span></code></a> for more details.</p>
</div>
<div class="section" id="iterable-style-datasets">
<h3>Iterable-style datasets<a class="headerlink" href="#iterable-style-datasets" title="Permalink to this headline">¶</a></h3>
<p>An iterable-style dataset is an instance of a subclass of <a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a>
that implements the <code class="xref py py-meth docutils literal notranslate"><span class="pre">__iter__()</span></code> protocol, and represents an iterable over
data samples. This type of datasets is particularly suitable for cases where
random reads are expensive or even improbable, and where the batch size depends
on the fetched data.</p>
<p>For example, such a dataset, when called <code class="docutils literal notranslate"><span class="pre">iter(dataset)</span></code>, could return a
stream of data reading from a database, a remote server, or even logs generated
in real time.</p>
<p>See <a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> for more details.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>When using an <a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> with
<a class="reference internal" href="#multi-process-data-loading">multi-process data loading</a>. The same
dataset object is replicated on each worker process, and thus the
replicas must be configured differently to avoid duplicated data. See
<a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> documentations for how to
achieve this.</p>
</div>
</div>
</div>
<div class="section" id="data-loading-order-and-sampler">
<h2>Data Loading Order and <a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sampler</span></code></a><a class="headerlink" href="#data-loading-order-and-sampler" title="Permalink to this headline">¶</a></h2>
<p>For <a class="reference internal" href="#iterable-style-datasets">iterable-style datasets</a>, data loading order
is entirely controlled by the user-defined iterable. This allows easier
implementations of chunk-reading and dynamic batch size (e.g., by yielding a
batched sample at each time).</p>
<p>The rest of this section concerns the case with
<a class="reference internal" href="#map-style-datasets">map-style datasets</a>. <a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.utils.data.Sampler</span></code></a>
classes are used to specify the sequence of indices/keys used in data loading.
They represent iterable objects over the indices to datasets. E.g., in the
common case with stochastic gradient decent (SGD), a
<a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sampler</span></code></a> could randomly permute a list of indices
and yield each one at a time, or yield a small number of them for mini-batch
SGD.</p>
<p>A sequential or shuffled sampler will be automatically constructed based on the <code class="xref py py-attr docutils literal notranslate"><span class="pre">shuffle</span></code> argument to a <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>.
Alternatively, users may use the <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code> argument to specify a
custom <a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sampler</span></code></a> object that at each time yields
the next index/key to fetch.</p>
<p>A custom <a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sampler</span></code></a> that yields a list of batch
indices at a time can be passed as the <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code> argument.
Automatic batching can also be enabled via <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code> and
<code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code> arguments. See
<a class="reference internal" href="#loading-batched-and-non-batched-data">the next section</a> for more details
on this.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Neither <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code> nor <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code> is compatible with
iterable-style datasets, since such datasets have no notion of a key or an
index.</p>
</div>
</div>
<div class="section" id="loading-batched-and-non-batched-data">
<h2>Loading Batched and Non-Batched Data<a class="headerlink" href="#loading-batched-and-non-batched-data" title="Permalink to this headline">¶</a></h2>
<p><a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> supports automatically collating
individual fetched data samples into batches via arguments
<code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code>, and <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code>.</p>
<div class="section" id="automatic-batching-default">
<h3>Automatic batching (default)<a class="headerlink" href="#automatic-batching-default" title="Permalink to this headline">¶</a></h3>
<p>This is the most common case, and corresponds to fetching a minibatch of
data and collating them into batched samples, i.e., containing Tensors with
one dimension being the batch dimension (usually the first).</p>
<p>When <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code> (default <code class="docutils literal notranslate"><span class="pre">1</span></code>) is not <code class="docutils literal notranslate"><span class="pre">None</span></code>, the data loader yields
batched samples instead of individual samples. <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code> and
<code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code> arguments are used to specify how the data loader obtains
batches of dataset keys. For map-style datasets, users can alternatively
specify <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code>, which yields a list of keys at a time.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code> and <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code> arguments essentially are used
to construct a <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code> from <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code>. For map-style
datasets, the <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code> is either provided by user or constructed
based on the <code class="xref py py-attr docutils literal notranslate"><span class="pre">shuffle</span></code> argument. For iterable-style datasets, the
<code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code> is a dummy infinite one. See
<a class="reference internal" href="#data-loading-order-and-sampler">this section</a> on more details on
samplers.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>When fetching from
<a class="reference internal" href="#iterable-style-datasets">iterable-style datasets</a> with
<a class="reference internal" href="#multi-process-data-loading">multi-processing</a>, the <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code>
argument drops the last non-full batch of each worker’s dataset replica.</p>
</div>
<p>After fetching a list of samples using the indices from sampler, the function
passed as the <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> argument is used to collate lists of samples
into batches.</p>
<p>In this case, loading from a map-style dataset is roughly equivalent with:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">indices</span> <span class="ow">in</span> <span class="n">batch_sampler</span><span class="p">:</span>
<span class="k">yield</span> <span class="n">collate_fn</span><span class="p">([</span><span class="n">dataset</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="n">indices</span><span class="p">])</span>
</pre></div>
</div>
<p>and loading from an iterable-style dataset is roughly equivalent with:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">dataset_iter</span> <span class="o">=</span> <span class="nb">iter</span><span class="p">(</span><span class="n">dataset</span><span class="p">)</span>
<span class="k">for</span> <span class="n">indices</span> <span class="ow">in</span> <span class="n">batch_sampler</span><span class="p">:</span>
<span class="k">yield</span> <span class="n">collate_fn</span><span class="p">([</span><span class="nb">next</span><span class="p">(</span><span class="n">dataset_iter</span><span class="p">)</span> <span class="k">for</span> <span class="n">_</span> <span class="ow">in</span> <span class="n">indices</span><span class="p">])</span>
</pre></div>
</div>
<p>A custom <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> can be used to customize collation, e.g., padding
sequential data to max length of a batch. See
<a class="reference internal" href="#dataloader-collate-fn">this section</a> on more about <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code>.</p>
</div>
<div class="section" id="disable-automatic-batching">
<h3>Disable automatic batching<a class="headerlink" href="#disable-automatic-batching" title="Permalink to this headline">¶</a></h3>
<p>In certain cases, users may want to handle batching manually in dataset code,
or simply load individual samples. For example, it could be cheaper to directly
load batched data (e.g., bulk reads from a database or reading continuous
chunks of memory), or the batch size is data dependent, or the program is
designed to work on individual samples. Under these scenarios, it’s likely
better to not use automatic batching (where <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> is used to
collate the samples), but let the data loader directly return each member of
the <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> object.</p>
<p>When both <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code> and <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code> are <code class="docutils literal notranslate"><span class="pre">None</span></code> (default
value for <code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_sampler</span></code> is already <code class="docutils literal notranslate"><span class="pre">None</span></code>), automatic batching is
disabled. Each sample obtained from the <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> is processed with the
function passed as the <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> argument.</p>
<p><strong>When automatic batching is disabled</strong>, the default <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> simply
converts NumPy arrays into PyTorch Tensors, and keeps everything else untouched.</p>
<p>In this case, loading from a map-style dataset is roughly equivalent with:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">index</span> <span class="ow">in</span> <span class="n">sampler</span><span class="p">:</span>
<span class="k">yield</span> <span class="n">collate_fn</span><span class="p">(</span><span class="n">dataset</span><span class="p">[</span><span class="n">index</span><span class="p">])</span>
</pre></div>
</div>
<p>and loading from an iterable-style dataset is roughly equivalent with:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">for</span> <span class="n">data</span> <span class="ow">in</span> <span class="nb">iter</span><span class="p">(</span><span class="n">dataset</span><span class="p">):</span>
<span class="k">yield</span> <span class="n">collate_fn</span><span class="p">(</span><span class="n">data</span><span class="p">)</span>
</pre></div>
</div>
<p>See <a class="reference internal" href="#dataloader-collate-fn">this section</a> on more about <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code>.</p>
</div>
<div class="section" id="working-with-collate-fn">
<span id="dataloader-collate-fn"></span><h3>Working with <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code><a class="headerlink" href="#working-with-collate-fn" title="Permalink to this headline">¶</a></h3>
<p>The use of <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> is slightly different when automatic batching is
enabled or disabled.</p>
<p><strong>When automatic batching is disabled</strong>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> is called with
each individual data sample, and the output is yielded from the data loader
iterator. In this case, the default <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> simply converts NumPy
arrays in PyTorch tensors.</p>
<p><strong>When automatic batching is enabled</strong>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> is called with a list
of data samples at each time. It is expected to collate the input samples into
a batch for yielding from the data loader iterator. The rest of this section
describes behavior of the default <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> in this case.</p>
<p>For instance, if each data sample consists of a 3-channel image and an integral
class label, i.e., each element of the dataset returns a tuple
<code class="docutils literal notranslate"><span class="pre">(image,</span> <span class="pre">class_index)</span></code>, the default <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> collates a list of
such tuples into a single tuple of a batched image tensor and a batched class
label Tensor. In particular, the default <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> has the following
properties:</p>
<ul class="simple">
<li><p>It always prepends a new dimension as the batch dimension.</p></li>
<li><p>It automatically converts NumPy arrays and Python numerical values into
PyTorch Tensors.</p></li>
<li><p>It preserves the data structure, e.g., if each sample is a dictionary, it
outputs a dictionary with the same set of keys but batched Tensors as values
(or lists if the values can not be converted into Tensors). Same
for <code class="docutils literal notranslate"><span class="pre">list</span></code> s, <code class="docutils literal notranslate"><span class="pre">tuple</span></code> s, <code class="docutils literal notranslate"><span class="pre">namedtuple</span></code> s, etc.</p></li>
</ul>
<p>Users may use customized <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> to achieve custom batching, e.g.,
collating along a dimension other than the first, padding sequences of
various lengths, or adding support for custom data types.</p>
</div>
</div>
<div class="section" id="single-and-multi-process-data-loading">
<h2>Single- and Multi-process Data Loading<a class="headerlink" href="#single-and-multi-process-data-loading" title="Permalink to this headline">¶</a></h2>
<p>A <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> uses single-process data loading by
default.</p>
<p>Within a Python process, the
<a class="reference external" href="https://wiki.python.org/moin/GlobalInterpreterLock">Global Interpreter Lock (GIL)</a>
prevents true fully parallelizing Python code across threads. To avoid blocking
computation code with data loading, PyTorch provides an easy switch to perform
multi-process data loading by simply setting the argument <code class="xref py py-attr docutils literal notranslate"><span class="pre">num_workers</span></code>
to a positive integer.</p>
<div class="section" id="single-process-data-loading-default">
<h3>Single-process data loading (default)<a class="headerlink" href="#single-process-data-loading-default" title="Permalink to this headline">¶</a></h3>
<p>In this mode, data fetching is done in the same process a
<a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> is initialized. Therefore, data loading
may block computing. However, this mode may be preferred when resource(s) used
for sharing data among processes (e.g., shared memory, file descriptors) is
limited, or when the entire dataset is small and can be loaded entirely in
memory. Additionally, single-process loading often shows more readable error
traces and thus is useful for debugging.</p>
</div>
<div class="section" id="multi-process-data-loading">
<h3>Multi-process data loading<a class="headerlink" href="#multi-process-data-loading" title="Permalink to this headline">¶</a></h3>
<p>Setting the argument <code class="xref py py-attr docutils literal notranslate"><span class="pre">num_workers</span></code> as a positive integer will
turn on multi-process data loading with the specified number of loader worker
processes.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>After several iterations, the loader worker processes will consume
the same amount of CPU memory as the parent process for all Python
objects in the parent process which are accessed from the worker
processes. This can be problematic if the Dataset contains a lot of
data (e.g., you are loading a very large list of filenames at Dataset
construction time) and/or you are using a lot of workers (overall
memory usage is <code class="docutils literal notranslate"><span class="pre">number</span> <span class="pre">of</span> <span class="pre">workers</span> <span class="pre">*</span> <span class="pre">size</span> <span class="pre">of</span> <span class="pre">parent</span> <span class="pre">process</span></code>). The
simplest workaround is to replace Python objects with non-refcounted
representations such as Pandas, Numpy or PyArrow objects. Check out
<a class="reference external" href="https://github.com/pytorch/pytorch/issues/13246#issuecomment-905703662">issue #13246</a>
for more details on why this occurs and example code for how to
workaround these problems.</p>
</div>
<p>In this mode, each time an iterator of a <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>
is created (e.g., when you call <code class="docutils literal notranslate"><span class="pre">enumerate(dataloader)</span></code>), <code class="xref py py-attr docutils literal notranslate"><span class="pre">num_workers</span></code>
worker processes are created. At this point, the <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code>,
<code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code>, and <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code> are passed to each
worker, where they are used to initialize, and fetch data. This means that
dataset access together with its internal IO, transforms
(including <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code>) runs in the worker process.</p>
<p><a class="reference internal" href="#torch.utils.data.get_worker_info" title="torch.utils.data.get_worker_info"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.utils.data.get_worker_info()</span></code></a> returns various useful information
in a worker process (including the worker id, dataset replica, initial seed,
etc.), and returns <code class="docutils literal notranslate"><span class="pre">None</span></code> in main process. Users may use this function in
dataset code and/or <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code> to individually configure each
dataset replica, and to determine whether the code is running in a worker
process. For example, this can be particularly helpful in sharding the dataset.</p>
<p>For map-style datasets, the main process generates the indices using
<code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code> and sends them to the workers. So any shuffle randomization is
done in the main process which guides loading by assigning indices to load.</p>
<p>For iterable-style datasets, since each worker process gets a replica of the
<code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> object, naive multi-process loading will often result in
duplicated data. Using <a class="reference internal" href="#torch.utils.data.get_worker_info" title="torch.utils.data.get_worker_info"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.utils.data.get_worker_info()</span></code></a> and/or
<code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code>, users may configure each replica independently. (See
<a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> documentations for how to achieve
this. ) For similar reasons, in multi-process loading, the <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code>
argument drops the last non-full batch of each worker’s iterable-style dataset
replica.</p>
<p>Workers are shut down once the end of the iteration is reached, or when the
iterator becomes garbage collected.</p>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>It is generally not recommended to return CUDA tensors in multi-process
loading because of many subtleties in using CUDA and sharing CUDA tensors in
multiprocessing (see <a class="reference internal" href="notes/multiprocessing.html#multiprocessing-cuda-note"><span class="std std-ref">CUDA in multiprocessing</span></a>). Instead, we recommend
using <a class="reference internal" href="#memory-pinning">automatic memory pinning</a> (i.e., setting
<code class="xref py py-attr docutils literal notranslate"><span class="pre">pin_memory=True</span></code>), which enables fast data transfer to CUDA-enabled
GPUs.</p>
</div>
<div class="section" id="platform-specific-behaviors">
<h4>Platform-specific behaviors<a class="headerlink" href="#platform-specific-behaviors" title="Permalink to this headline">¶</a></h4>
<p>Since workers rely on Python <a class="reference external" href="https://docs.python.org/3/library/multiprocessing.html#module-multiprocessing" title="(in Python v3.10)"><code class="xref py py-mod docutils literal notranslate"><span class="pre">multiprocessing</span></code></a>, worker launch behavior is
different on Windows compared to Unix.</p>
<ul class="simple">
<li><p>On Unix, <code class="xref py py-func docutils literal notranslate"><span class="pre">fork()</span></code> is the default <a class="reference external" href="https://docs.python.org/3/library/multiprocessing.html#module-multiprocessing" title="(in Python v3.10)"><code class="xref py py-mod docutils literal notranslate"><span class="pre">multiprocessing</span></code></a> start method.
Using <code class="xref py py-func docutils literal notranslate"><span class="pre">fork()</span></code>, child workers typically can access the <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> and
Python argument functions directly through the cloned address space.</p></li>
<li><p>On Windows or MacOS, <code class="xref py py-func docutils literal notranslate"><span class="pre">spawn()</span></code> is the default <a class="reference external" href="https://docs.python.org/3/library/multiprocessing.html#module-multiprocessing" title="(in Python v3.10)"><code class="xref py py-mod docutils literal notranslate"><span class="pre">multiprocessing</span></code></a> start method.
Using <code class="xref py py-func docutils literal notranslate"><span class="pre">spawn()</span></code>, another interpreter is launched which runs your main script,
followed by the internal worker function that receives the <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code>,
<code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> and other arguments through <a class="reference external" href="https://docs.python.org/3/library/pickle.html#module-pickle" title="(in Python v3.10)"><code class="xref py py-mod docutils literal notranslate"><span class="pre">pickle</span></code></a> serialization.</p></li>
</ul>
<p>This separate serialization means that you should take two steps to ensure you
are compatible with Windows while using multi-process data loading:</p>
<ul class="simple">
<li><p>Wrap most of you main script’s code within <code class="docutils literal notranslate"><span class="pre">if</span> <span class="pre">__name__</span> <span class="pre">==</span> <span class="pre">'__main__':</span></code> block,
to make sure it doesn’t run again (most likely generating error) when each worker
process is launched. You can place your dataset and <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>
instance creation logic here, as it doesn’t need to be re-executed in workers.</p></li>
<li><p>Make sure that any custom <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code>
or <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> code is declared as top level definitions, outside of the
<code class="docutils literal notranslate"><span class="pre">__main__</span></code> check. This ensures that they are available in worker processes.
(this is needed since functions are pickled as references only, not <code class="docutils literal notranslate"><span class="pre">bytecode</span></code>.)</p></li>
</ul>
</div>
<div class="section" id="randomness-in-multi-process-data-loading">
<span id="data-loading-randomness"></span><h4>Randomness in multi-process data loading<a class="headerlink" href="#randomness-in-multi-process-data-loading" title="Permalink to this headline">¶</a></h4>
<p>By default, each worker will have its PyTorch seed set to <code class="docutils literal notranslate"><span class="pre">base_seed</span> <span class="pre">+</span> <span class="pre">worker_id</span></code>,
where <code class="docutils literal notranslate"><span class="pre">base_seed</span></code> is a long generated by main process using its RNG (thereby,
consuming a RNG state mandatorily) or a specified <code class="xref py py-attr docutils literal notranslate"><span class="pre">generator</span></code>. However, seeds for other
libraries may be duplicated upon initializing workers, causing each worker to return
identical random numbers. (See <a class="reference internal" href="notes/faq.html#dataloader-workers-random-seed"><span class="std std-ref">this section</span></a> in FAQ.).</p>
<p>In <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code>, you may access the PyTorch seed set for each worker
with either <a class="reference internal" href="#torch.utils.data.get_worker_info" title="torch.utils.data.get_worker_info"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.utils.data.get_worker_info().seed</span></code></a>
or <a class="reference internal" href="generated/torch.initial_seed.html#torch.initial_seed" title="torch.initial_seed"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.initial_seed()</span></code></a>, and use it to seed other libraries before data
loading.</p>
</div>
</div>
</div>
<div class="section" id="memory-pinning">
<h2>Memory Pinning<a class="headerlink" href="#memory-pinning" title="Permalink to this headline">¶</a></h2>
<p>Host to GPU copies are much faster when they originate from pinned (page-locked)
memory. See <a class="reference internal" href="notes/cuda.html#cuda-memory-pinning"><span class="std std-ref">Use pinned memory buffers</span></a> for more details on when and how to use
pinned memory generally.</p>
<p>For data loading, passing <code class="xref py py-attr docutils literal notranslate"><span class="pre">pin_memory=True</span></code> to a
<a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> will automatically put the fetched data
Tensors in pinned memory, and thus enables faster data transfer to CUDA-enabled
GPUs.</p>
<p>The default memory pinning logic only recognizes Tensors and maps and iterables
containing Tensors. By default, if the pinning logic sees a batch that is a
custom type (which will occur if you have a <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> that returns a
custom batch type), or if each element of your batch is a custom type, the
pinning logic will not recognize them, and it will return that batch (or those
elements) without pinning the memory. To enable memory pinning for custom
batch or data type(s), define a <code class="xref py py-meth docutils literal notranslate"><span class="pre">pin_memory()</span></code> method on your custom
type(s).</p>
<p>See the example below.</p>
<p>Example:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">SimpleCustomBatch</span><span class="p">:</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">data</span><span class="p">):</span>
<span class="n">transposed_data</span> <span class="o">=</span> <span class="nb">list</span><span class="p">(</span><span class="nb">zip</span><span class="p">(</span><span class="o">*</span><span class="n">data</span><span class="p">))</span>
<span class="bp">self</span><span class="o">.</span><span class="n">inp</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="n">transposed_data</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="mi">0</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">tgt</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">(</span><span class="n">transposed_data</span><span class="p">[</span><span class="mi">1</span><span class="p">],</span> <span class="mi">0</span><span class="p">)</span>
<span class="c1"># custom memory pinning method on custom type</span>
<span class="k">def</span> <span class="nf">pin_memory</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="bp">self</span><span class="o">.</span><span class="n">inp</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">inp</span><span class="o">.</span><span class="n">pin_memory</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">tgt</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">tgt</span><span class="o">.</span><span class="n">pin_memory</span><span class="p">()</span>
<span class="k">return</span> <span class="bp">self</span>
<span class="k">def</span> <span class="nf">collate_wrapper</span><span class="p">(</span><span class="n">batch</span><span class="p">):</span>
<span class="k">return</span> <span class="n">SimpleCustomBatch</span><span class="p">(</span><span class="n">batch</span><span class="p">)</span>
<span class="n">inps</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">10</span> <span class="o">*</span> <span class="mi">5</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span>
<span class="n">tgts</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="mi">10</span> <span class="o">*</span> <span class="mi">5</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">float32</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="mi">5</span><span class="p">)</span>
<span class="n">dataset</span> <span class="o">=</span> <span class="n">TensorDataset</span><span class="p">(</span><span class="n">inps</span><span class="p">,</span> <span class="n">tgts</span><span class="p">)</span>
<span class="n">loader</span> <span class="o">=</span> <span class="n">DataLoader</span><span class="p">(</span><span class="n">dataset</span><span class="p">,</span> <span class="n">batch_size</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span> <span class="n">collate_fn</span><span class="o">=</span><span class="n">collate_wrapper</span><span class="p">,</span>
<span class="n">pin_memory</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">for</span> <span class="n">batch_ndx</span><span class="p">,</span> <span class="n">sample</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">loader</span><span class="p">):</span>
<span class="nb">print</span><span class="p">(</span><span class="n">sample</span><span class="o">.</span><span class="n">inp</span><span class="o">.</span><span class="n">is_pinned</span><span class="p">())</span>
<span class="nb">print</span><span class="p">(</span><span class="n">sample</span><span class="o">.</span><span class="n">tgt</span><span class="o">.</span><span class="n">is_pinned</span><span class="p">())</span>
</pre></div>
</div>
<dl class="py class">
<dt id="torch.utils.data.DataLoader">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">DataLoader</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">dataset</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">batch_size</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">1</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">shuffle</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">sampler</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">batch_sampler</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">num_workers</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">collate_fn</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">pin_memory</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">drop_last</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">timeout</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">0</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">worker_init_fn</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">multiprocessing_context</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">generator</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">None</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">*</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">prefetch_factor</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">2</span></span></em>, <em class="sig-param"><span class="n"><span class="pre">persistent_workers</span></span><span class="o"><span class="pre">=</span></span><span class="default_value"><span class="pre">False</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataloader.html#DataLoader"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.DataLoader" title="Permalink to this definition">¶</a></dt>
<dd><p>Data loader. Combines a dataset and a sampler, and provides an iterable over
the given dataset.</p>
<p>The <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> supports both map-style and
iterable-style datasets with single- or multi-process loading, customizing
loading order and optional automatic batching (collation) and memory pinning.</p>
<p>See <a class="reference internal" href="#module-torch.utils.data" title="torch.utils.data"><code class="xref py py-mod docutils literal notranslate"><span class="pre">torch.utils.data</span></code></a> documentation page for more details.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>dataset</strong> (<a class="reference internal" href="#torch.utils.data.Dataset" title="torch.utils.data.Dataset"><em>Dataset</em></a>) – dataset from which to load the data.</p></li>
<li><p><strong>batch_size</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a><em>, </em><em>optional</em>) – how many samples per batch to load
(default: <code class="docutils literal notranslate"><span class="pre">1</span></code>).</p></li>
<li><p><strong>shuffle</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – set to <code class="docutils literal notranslate"><span class="pre">True</span></code> to have the data reshuffled
at every epoch (default: <code class="docutils literal notranslate"><span class="pre">False</span></code>).</p></li>
<li><p><strong>sampler</strong> (<a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><em>Sampler</em></a><em> or </em><em>Iterable</em><em>, </em><em>optional</em>) – defines the strategy to draw
samples from the dataset. Can be any <code class="docutils literal notranslate"><span class="pre">Iterable</span></code> with <code class="docutils literal notranslate"><span class="pre">__len__</span></code>
implemented. If specified, <code class="xref py py-attr docutils literal notranslate"><span class="pre">shuffle</span></code> must not be specified.</p></li>
<li><p><strong>batch_sampler</strong> (<a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><em>Sampler</em></a><em> or </em><em>Iterable</em><em>, </em><em>optional</em>) – like <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code>, but
returns a batch of indices at a time. Mutually exclusive with
<code class="xref py py-attr docutils literal notranslate"><span class="pre">batch_size</span></code>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">shuffle</span></code>, <code class="xref py py-attr docutils literal notranslate"><span class="pre">sampler</span></code>,
and <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code>.</p></li>
<li><p><strong>num_workers</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a><em>, </em><em>optional</em>) – how many subprocesses to use for data
loading. <code class="docutils literal notranslate"><span class="pre">0</span></code> means that the data will be loaded in the main process.
(default: <code class="docutils literal notranslate"><span class="pre">0</span></code>)</p></li>
<li><p><strong>collate_fn</strong> (<em>callable</em><em>, </em><em>optional</em>) – merges a list of samples to form a
mini-batch of Tensor(s). Used when using batched loading from a
map-style dataset.</p></li>
<li><p><strong>pin_memory</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, the data loader will copy Tensors
into CUDA pinned memory before returning them. If your data elements
are a custom type, or your <code class="xref py py-attr docutils literal notranslate"><span class="pre">collate_fn</span></code> returns a batch that is a custom type,
see the example below.</p></li>
<li><p><strong>drop_last</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – set to <code class="docutils literal notranslate"><span class="pre">True</span></code> to drop the last incomplete batch,
if the dataset size is not divisible by the batch size. If <code class="docutils literal notranslate"><span class="pre">False</span></code> and
the size of dataset is not divisible by the batch size, then the last batch
will be smaller. (default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
<li><p><strong>timeout</strong> (<em>numeric</em><em>, </em><em>optional</em>) – if positive, the timeout value for collecting a batch
from workers. Should always be non-negative. (default: <code class="docutils literal notranslate"><span class="pre">0</span></code>)</p></li>
<li><p><strong>worker_init_fn</strong> (<em>callable</em><em>, </em><em>optional</em>) – If not <code class="docutils literal notranslate"><span class="pre">None</span></code>, this will be called on each
worker subprocess with the worker id (an int in <code class="docutils literal notranslate"><span class="pre">[0,</span> <span class="pre">num_workers</span> <span class="pre">-</span> <span class="pre">1]</span></code>) as
input, after seeding and before data loading. (default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>generator</strong> (<a class="reference internal" href="generated/torch.Generator.html#torch.Generator" title="torch.Generator"><em>torch.Generator</em></a><em>, </em><em>optional</em>) – If not <code class="docutils literal notranslate"><span class="pre">None</span></code>, this RNG will be used
by RandomSampler to generate random indexes and multiprocessing to generate
<cite>base_seed</cite> for workers. (default: <code class="docutils literal notranslate"><span class="pre">None</span></code>)</p></li>
<li><p><strong>prefetch_factor</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#int" title="(in Python v3.10)"><em>int</em></a><em>, </em><em>optional</em><em>, </em><em>keyword-only arg</em>) – Number of samples loaded
in advance by each worker. <code class="docutils literal notranslate"><span class="pre">2</span></code> means there will be a total of
2 * num_workers samples prefetched across all workers. (default: <code class="docutils literal notranslate"><span class="pre">2</span></code>)</p></li>
<li><p><strong>persistent_workers</strong> (<a class="reference external" href="https://docs.python.org/3/library/functions.html#bool" title="(in Python v3.10)"><em>bool</em></a><em>, </em><em>optional</em>) – If <code class="docutils literal notranslate"><span class="pre">True</span></code>, the data loader will not shutdown
the worker processes after a dataset has been consumed once. This allows to
maintain the workers <cite>Dataset</cite> instances alive. (default: <code class="docutils literal notranslate"><span class="pre">False</span></code>)</p></li>
</ul>
</dd>
</dl>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>If the <code class="docutils literal notranslate"><span class="pre">spawn</span></code> start method is used, <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code>
cannot be an unpicklable object, e.g., a lambda function. See
<a class="reference internal" href="notes/multiprocessing.html#multiprocessing-best-practices"><span class="std std-ref">Multiprocessing best practices</span></a> on more details related
to multiprocessing in PyTorch.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p><code class="docutils literal notranslate"><span class="pre">len(dataloader)</span></code> heuristic is based on the length of the sampler used.
When <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> is an <a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a>,
it instead returns an estimate based on <code class="docutils literal notranslate"><span class="pre">len(dataset)</span> <span class="pre">/</span> <span class="pre">batch_size</span></code>, with proper
rounding depending on <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code>, regardless of multi-process loading
configurations. This represents the best guess PyTorch can make because PyTorch
trusts user <code class="xref py py-attr docutils literal notranslate"><span class="pre">dataset</span></code> code in correctly handling multi-process
loading to avoid duplicate data.</p>
<p>However, if sharding results in multiple workers having incomplete last batches,
this estimate can still be inaccurate, because (1) an otherwise complete batch can
be broken into multiple ones and (2) more than one batch worth of samples can be
dropped when <code class="xref py py-attr docutils literal notranslate"><span class="pre">drop_last</span></code> is set. Unfortunately, PyTorch can not detect such
cases in general.</p>
<p>See <a class="reference internal" href="#dataset-types">Dataset Types</a> for more details on these two types of datasets and how
<a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> interacts with
<a class="reference internal" href="#multi-process-data-loading">Multi-process data loading</a>.</p>
</div>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>See <a class="reference internal" href="notes/randomness.html#reproducibility"><span class="std std-ref">Reproducibility</span></a>, and <a class="reference internal" href="notes/faq.html#dataloader-workers-random-seed"><span class="std std-ref">My data loader workers return identical random numbers</span></a>, and
<a class="reference internal" href="#data-loading-randomness"><span class="std std-ref">Randomness in multi-process data loading</span></a> notes for random seed related questions.</p>
</div>
</dd></dl>
<dl class="py class">
<dt id="torch.utils.data.Dataset">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">Dataset</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwds</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataset.html#Dataset"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.Dataset" title="Permalink to this definition">¶</a></dt>
<dd><p>An abstract class representing a <a class="reference internal" href="#torch.utils.data.Dataset" title="torch.utils.data.Dataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">Dataset</span></code></a>.</p>
<p>All datasets that represent a map from keys to data samples should subclass
it. All subclasses should overwrite <code class="xref py py-meth docutils literal notranslate"><span class="pre">__getitem__()</span></code>, supporting fetching a
data sample for a given key. Subclasses could also optionally overwrite
<code class="xref py py-meth docutils literal notranslate"><span class="pre">__len__()</span></code>, which is expected to return the size of the dataset by many
<a class="reference internal" href="#torch.utils.data.Sampler" title="torch.utils.data.Sampler"><code class="xref py py-class docutils literal notranslate"><span class="pre">Sampler</span></code></a> implementations and the default options
of <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> by default constructs a index
sampler that yields integral indices. To make it work with a map-style
dataset with non-integral indices/keys, a custom sampler must be provided.</p>
</div>
</dd></dl>
<dl class="py class">
<dt id="torch.utils.data.IterableDataset">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">IterableDataset</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">args</span></span></em>, <em class="sig-param"><span class="o"><span class="pre">**</span></span><span class="n"><span class="pre">kwds</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataset.html#IterableDataset"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.IterableDataset" title="Permalink to this definition">¶</a></dt>
<dd><p>An iterable Dataset.</p>
<p>All datasets that represent an iterable of data samples should subclass it.
Such form of datasets is particularly useful when data come from a stream.</p>
<p>All subclasses should overwrite <code class="xref py py-meth docutils literal notranslate"><span class="pre">__iter__()</span></code>, which would return an
iterator of samples in this dataset.</p>
<p>When a subclass is used with <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>, each
item in the dataset will be yielded from the <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a>
iterator. When <code class="xref py py-attr docutils literal notranslate"><span class="pre">num_workers</span> <span class="pre">></span> <span class="pre">0</span></code>, each worker process will have a
different copy of the dataset object, so it is often desired to configure
each copy independently to avoid having duplicate data returned from the
workers. <a class="reference internal" href="#torch.utils.data.get_worker_info" title="torch.utils.data.get_worker_info"><code class="xref py py-func docutils literal notranslate"><span class="pre">get_worker_info()</span></code></a>, when called in a worker
process, returns information about the worker. It can be used in either the
dataset’s <code class="xref py py-meth docutils literal notranslate"><span class="pre">__iter__()</span></code> method or the <a class="reference internal" href="#torch.utils.data.DataLoader" title="torch.utils.data.DataLoader"><code class="xref py py-class docutils literal notranslate"><span class="pre">DataLoader</span></code></a> ‘s
<code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code> option to modify each copy’s behavior.</p>
<p>Example 1: splitting workload across all workers in <code class="xref py py-meth docutils literal notranslate"><span class="pre">__iter__()</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="k">class</span> <span class="nc">MyIterableDataset</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">IterableDataset</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">start</span><span class="p">,</span> <span class="n">end</span><span class="p">):</span>
<span class="gp">... </span> <span class="nb">super</span><span class="p">(</span><span class="n">MyIterableDataset</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="gp">... </span> <span class="k">assert</span> <span class="n">end</span> <span class="o">></span> <span class="n">start</span><span class="p">,</span> <span class="s2">"this example code only works with end >= start"</span>
<span class="gp">... </span> <span class="bp">self</span><span class="o">.</span><span class="n">start</span> <span class="o">=</span> <span class="n">start</span>
<span class="gp">... </span> <span class="bp">self</span><span class="o">.</span><span class="n">end</span> <span class="o">=</span> <span class="n">end</span>
<span class="gp">...</span>
<span class="gp">... </span> <span class="k">def</span> <span class="fm">__iter__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="gp">... </span> <span class="n">worker_info</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">get_worker_info</span><span class="p">()</span>
<span class="gp">... </span> <span class="k">if</span> <span class="n">worker_info</span> <span class="ow">is</span> <span class="kc">None</span><span class="p">:</span> <span class="c1"># single-process data loading, return the full iterator</span>
<span class="gp">... </span> <span class="n">iter_start</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">start</span>
<span class="gp">... </span> <span class="n">iter_end</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">end</span>
<span class="gp">... </span> <span class="k">else</span><span class="p">:</span> <span class="c1"># in a worker process</span>
<span class="gp">... </span> <span class="c1"># split workload</span>
<span class="gp">... </span> <span class="n">per_worker</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">math</span><span class="o">.</span><span class="n">ceil</span><span class="p">((</span><span class="bp">self</span><span class="o">.</span><span class="n">end</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">start</span><span class="p">)</span> <span class="o">/</span> <span class="nb">float</span><span class="p">(</span><span class="n">worker_info</span><span class="o">.</span><span class="n">num_workers</span><span class="p">)))</span>
<span class="gp">... </span> <span class="n">worker_id</span> <span class="o">=</span> <span class="n">worker_info</span><span class="o">.</span><span class="n">id</span>
<span class="gp">... </span> <span class="n">iter_start</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">start</span> <span class="o">+</span> <span class="n">worker_id</span> <span class="o">*</span> <span class="n">per_worker</span>
<span class="gp">... </span> <span class="n">iter_end</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">iter_start</span> <span class="o">+</span> <span class="n">per_worker</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">end</span><span class="p">)</span>
<span class="gp">... </span> <span class="k">return</span> <span class="nb">iter</span><span class="p">(</span><span class="nb">range</span><span class="p">(</span><span class="n">iter_start</span><span class="p">,</span> <span class="n">iter_end</span><span class="p">))</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="c1"># should give same set of data as range(3, 7), i.e., [3, 4, 5, 6].</span>
<span class="gp">>>> </span><span class="n">ds</span> <span class="o">=</span> <span class="n">MyIterableDataset</span><span class="p">(</span><span class="n">start</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">end</span><span class="o">=</span><span class="mi">7</span><span class="p">)</span>
<span class="gp">>>> </span><span class="c1"># Single-process loading</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">0</span><span class="p">)))</span>
<span class="go">[3, 4, 5, 6]</span>
<span class="gp">>>> </span><span class="c1"># Mult-process loading with two worker processes</span>
<span class="gp">>>> </span><span class="c1"># Worker 0 fetched [3, 4]. Worker 1 fetched [5, 6].</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">2</span><span class="p">)))</span>
<span class="go">[3, 5, 4, 6]</span>
<span class="gp">>>> </span><span class="c1"># With even more workers</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">20</span><span class="p">)))</span>
<span class="go">[3, 4, 5, 6]</span>
</pre></div>
</div>
<p>Example 2: splitting workload across all workers using <code class="xref py py-attr docutils literal notranslate"><span class="pre">worker_init_fn</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="k">class</span> <span class="nc">MyIterableDataset</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">IterableDataset</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">start</span><span class="p">,</span> <span class="n">end</span><span class="p">):</span>
<span class="gp">... </span> <span class="nb">super</span><span class="p">(</span><span class="n">MyIterableDataset</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="gp">... </span> <span class="k">assert</span> <span class="n">end</span> <span class="o">></span> <span class="n">start</span><span class="p">,</span> <span class="s2">"this example code only works with end >= start"</span>
<span class="gp">... </span> <span class="bp">self</span><span class="o">.</span><span class="n">start</span> <span class="o">=</span> <span class="n">start</span>
<span class="gp">... </span> <span class="bp">self</span><span class="o">.</span><span class="n">end</span> <span class="o">=</span> <span class="n">end</span>
<span class="gp">...</span>
<span class="gp">... </span> <span class="k">def</span> <span class="fm">__iter__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="gp">... </span> <span class="k">return</span> <span class="nb">iter</span><span class="p">(</span><span class="nb">range</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">start</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">end</span><span class="p">))</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="c1"># should give same set of data as range(3, 7), i.e., [3, 4, 5, 6].</span>
<span class="gp">>>> </span><span class="n">ds</span> <span class="o">=</span> <span class="n">MyIterableDataset</span><span class="p">(</span><span class="n">start</span><span class="o">=</span><span class="mi">3</span><span class="p">,</span> <span class="n">end</span><span class="o">=</span><span class="mi">7</span><span class="p">)</span>
<span class="gp">>>> </span><span class="c1"># Single-process loading</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">0</span><span class="p">)))</span>
<span class="go">[3, 4, 5, 6]</span>
<span class="go">>>></span>
<span class="gp">>>> </span><span class="c1"># Directly doing multi-process loading yields duplicate data</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">2</span><span class="p">)))</span>
<span class="go">[3, 3, 4, 4, 5, 5, 6, 6]</span>
<span class="gp">>>> </span><span class="c1"># Define a `worker_init_fn` that configures each dataset copy differently</span>
<span class="gp">>>> </span><span class="k">def</span> <span class="nf">worker_init_fn</span><span class="p">(</span><span class="n">worker_id</span><span class="p">):</span>
<span class="gp">... </span> <span class="n">worker_info</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">get_worker_info</span><span class="p">()</span>
<span class="gp">... </span> <span class="n">dataset</span> <span class="o">=</span> <span class="n">worker_info</span><span class="o">.</span><span class="n">dataset</span> <span class="c1"># the dataset copy in this worker process</span>
<span class="gp">... </span> <span class="n">overall_start</span> <span class="o">=</span> <span class="n">dataset</span><span class="o">.</span><span class="n">start</span>
<span class="gp">... </span> <span class="n">overall_end</span> <span class="o">=</span> <span class="n">dataset</span><span class="o">.</span><span class="n">end</span>
<span class="gp">... </span> <span class="c1"># configure the dataset to only process the split workload</span>
<span class="gp">... </span> <span class="n">per_worker</span> <span class="o">=</span> <span class="nb">int</span><span class="p">(</span><span class="n">math</span><span class="o">.</span><span class="n">ceil</span><span class="p">((</span><span class="n">overall_end</span> <span class="o">-</span> <span class="n">overall_start</span><span class="p">)</span> <span class="o">/</span> <span class="nb">float</span><span class="p">(</span><span class="n">worker_info</span><span class="o">.</span><span class="n">num_workers</span><span class="p">)))</span>
<span class="gp">... </span> <span class="n">worker_id</span> <span class="o">=</span> <span class="n">worker_info</span><span class="o">.</span><span class="n">id</span>
<span class="gp">... </span> <span class="n">dataset</span><span class="o">.</span><span class="n">start</span> <span class="o">=</span> <span class="n">overall_start</span> <span class="o">+</span> <span class="n">worker_id</span> <span class="o">*</span> <span class="n">per_worker</span>
<span class="gp">... </span> <span class="n">dataset</span><span class="o">.</span><span class="n">end</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="n">dataset</span><span class="o">.</span><span class="n">start</span> <span class="o">+</span> <span class="n">per_worker</span><span class="p">,</span> <span class="n">overall_end</span><span class="p">)</span>
<span class="gp">...</span>
<span class="gp">>>> </span><span class="c1"># Mult-process loading with the custom `worker_init_fn`</span>
<span class="gp">>>> </span><span class="c1"># Worker 0 fetched [3, 4]. Worker 1 fetched [5, 6].</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">2</span><span class="p">,</span> <span class="n">worker_init_fn</span><span class="o">=</span><span class="n">worker_init_fn</span><span class="p">)))</span>
<span class="go">[3, 5, 4, 6]</span>
<span class="gp">>>> </span><span class="c1"># With even more workers</span>
<span class="gp">>>> </span><span class="nb">print</span><span class="p">(</span><span class="nb">list</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">ds</span><span class="p">,</span> <span class="n">num_workers</span><span class="o">=</span><span class="mi">20</span><span class="p">,</span> <span class="n">worker_init_fn</span><span class="o">=</span><span class="n">worker_init_fn</span><span class="p">)))</span>
<span class="go">[3, 4, 5, 6]</span>
</pre></div>
</div>
</dd></dl>
<dl class="py class">
<dt id="torch.utils.data.TensorDataset">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">TensorDataset</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="o"><span class="pre">*</span></span><span class="n"><span class="pre">tensors</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataset.html#TensorDataset"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.TensorDataset" title="Permalink to this definition">¶</a></dt>
<dd><p>Dataset wrapping tensors.</p>
<p>Each sample will be retrieved by indexing tensors along the first dimension.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>*tensors</strong> (<a class="reference internal" href="tensors.html#torch.Tensor" title="torch.Tensor"><em>Tensor</em></a>) – tensors that have the same size of the first dimension.</p>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt id="torch.utils.data.ConcatDataset">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">ConcatDataset</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">datasets</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataset.html#ConcatDataset"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.ConcatDataset" title="Permalink to this definition">¶</a></dt>
<dd><p>Dataset as a concatenation of multiple datasets.</p>
<p>This class is useful to assemble different existing datasets.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>datasets</strong> (<em>sequence</em>) – List of datasets to be concatenated</p>
</dd>
</dl>
</dd></dl>
<dl class="py class">
<dt id="torch.utils.data.ChainDataset">
<em class="property"><span class="pre">class</span> </em><code class="sig-prename descclassname"><span class="pre">torch.utils.data.</span></code><code class="sig-name descname"><span class="pre">ChainDataset</span></code><span class="sig-paren">(</span><em class="sig-param"><span class="n"><span class="pre">datasets</span></span></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/utils/data/dataset.html#ChainDataset"><span class="viewcode-link"><span class="pre">[source]</span></span></a><a class="headerlink" href="#torch.utils.data.ChainDataset" title="Permalink to this definition">¶</a></dt>
<dd><p>Dataset for chaining multiple <a class="reference internal" href="#torch.utils.data.IterableDataset" title="torch.utils.data.IterableDataset"><code class="xref py py-class docutils literal notranslate"><span class="pre">IterableDataset</span></code></a> s.</p>
<p>This class is useful to assemble different existing dataset streams. The
chaining operation is done on-the-fly, so concatenating large-scale
datasets with this class will be efficient.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><p><strong>datasets</strong> (<em>iterable of IterableDataset</em>) – datasets to be chained together</p>
</dd>
</dl>