-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathquantization.html
2922 lines (2620 loc) · 322 KB
/
quantization.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>Quantization — PyTorch master documentation</title>
<link rel="canonical" href="https://pytorch.org/docs/stable/quantization.html"/>
<link rel="stylesheet" href="_static/css/theme.css" type="text/css" />
<!-- <link rel="stylesheet" href="_static/pygments.css" type="text/css" /> -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/css/jit.css" type="text/css" />
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/katex.min.css" type="text/css" />
<link rel="stylesheet" href="_static/katex-math.css" type="text/css" />
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Distributed RPC Framework" href="rpc/index.html" />
<link rel="prev" title="torch.optim" href="optim.html" />
<script src="_static/js/modernizr.min.js"></script>
<!-- Preload the theme fonts -->
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-book.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-Medium.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/FreightSans/freight-sans-medium-italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="_static/fonts/IBMPlexMono/IBMPlexMono-SemiBold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<!-- Preload the katex fonts -->
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Math-Italic.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Main-Bold.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size1-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size4-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size2-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Size3-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
<link rel="preload" href="https://cdn.jsdelivr.net/npm/[email protected]/dist/fonts/KaTeX_Caligraphic-Regular.woff2" as="font" type="font/woff2" crossorigin="anonymous">
</head>
<div class="container-fluid header-holder tutorials-header" id="header-holder">
<div class="container">
<div class="header-container">
<a class="header-logo" href="https://pytorch.org/" aria-label="PyTorch"></a>
<div class="main-menu">
<ul>
<li>
<a href="https://pytorch.org/get-started">Get Started</a>
</li>
<li>
<div class="ecosystem-dropdown">
<a id="dropdownMenuButton" data-toggle="ecosystem-dropdown">
Ecosystem
</a>
<div class="ecosystem-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/hub"">
<span class=dropdown-title>Models (Beta)</span>
<p>Discover, publish, and reuse pre-trained models</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/ecosystem">
<span class=dropdown-title>Tools & Libraries</span>
<p>Explore the ecosystem of tools and libraries</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://pytorch.org/mobile">Mobile</a>
</li>
<li>
<a href="https://pytorch.org/blog/">Blog</a>
</li>
<li>
<a href="https://pytorch.org/tutorials">Tutorials</a>
</li>
<li class="active">
<a href="https://pytorch.org/docs/stable/index.html">Docs</a>
</li>
<li>
<div class="resources-dropdown">
<a id="resourcesDropdownButton" data-toggle="resources-dropdown">
Resources
</a>
<div class="resources-dropdown-menu">
<a class="nav-dropdown-item" href="https://pytorch.org/resources"">
<span class=dropdown-title>Developer Resources</span>
<p>Find resources and get questions answered</p>
</a>
<a class="nav-dropdown-item" href="https://pytorch.org/features">
<span class=dropdown-title>About</span>
<p>Learn about PyTorch’s features and capabilities</p>
</a>
</div>
</div>
</li>
<li>
<a href="https://github.com/pytorch/pytorch">Github</a>
</li>
</ul>
</div>
<a class="main-menu-open-button" href="#" data-behavior="open-mobile-menu"></a>
</div>
</div>
</div>
<body class="pytorch-body">
<div class="table-of-contents-link-wrapper">
<span>Table of Contents</span>
<a href="#" class="toggle-table-of-contents" data-behavior="toggle-table-of-contents"></a>
</div>
<nav data-toggle="wy-nav-shift" class="pytorch-left-menu" id="pytorch-left-menu">
<div class="pytorch-side-scroll">
<div class="pytorch-menu pytorch-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
<div class="pytorch-left-menu-search">
<div class="version">
<a href='http://pytorch.org/docs/versions.html'>1.5.1 ▼</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/large_scale_deployments.html">Features for large-scale deployments</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/multiprocessing.html">Multiprocessing best practices</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/randomness.html">Reproducibility</a></li>
<li class="toctree-l1"><a class="reference internal" href="notes/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="distributed.html">torch.distributed</a></li>
<li class="toctree-l1"><a class="reference internal" href="distributions.html">torch.distributions</a></li>
<li class="toctree-l1"><a class="reference internal" href="hub.html">torch.hub</a></li>
<li class="toctree-l1"><a class="reference internal" href="jit.html">torch.jit</a></li>
<li class="toctree-l1"><a class="reference internal" href="nn.init.html">torch.nn.init</a></li>
<li class="toctree-l1"><a class="reference internal" href="onnx.html">torch.onnx</a></li>
<li class="toctree-l1"><a class="reference internal" href="optim.html">torch.optim</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">Quantization</a></li>
<li class="toctree-l1"><a class="reference internal" href="rpc/index.html">Distributed RPC Framework</a></li>
<li class="toctree-l1"><a class="reference internal" href="random.html">torch.random</a></li>
<li class="toctree-l1"><a class="reference internal" href="sparse.html">torch.sparse</a></li>
<li class="toctree-l1"><a class="reference internal" href="storage.html">torch.Storage</a></li>
<li class="toctree-l1"><a class="reference internal" href="bottleneck.html">torch.utils.bottleneck</a></li>
<li class="toctree-l1"><a class="reference internal" href="checkpoint.html">torch.utils.checkpoint</a></li>
<li class="toctree-l1"><a class="reference internal" href="cpp_extension.html">torch.utils.cpp_extension</a></li>
<li class="toctree-l1"><a class="reference internal" href="data.html">torch.utils.data</a></li>
<li class="toctree-l1"><a class="reference internal" href="dlpack.html">torch.utils.dlpack</a></li>
<li class="toctree-l1"><a class="reference internal" href="model_zoo.html">torch.utils.model_zoo</a></li>
<li class="toctree-l1"><a class="reference internal" href="tensorboard.html">torch.utils.tensorboard</a></li>
<li class="toctree-l1"><a class="reference internal" href="type_info.html">Type Info</a></li>
<li class="toctree-l1"><a class="reference internal" href="named_tensor.html">Named Tensors</a></li>
<li class="toctree-l1"><a class="reference internal" href="name_inference.html">Named Tensors operator coverage</a></li>
<li class="toctree-l1"><a class="reference internal" href="__config__.html">torch.__config__</a></li>
</ul>
<p class="caption"><span class="caption-text">Libraries</span></p>
<ul>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/audio">torchaudio</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/text">torchtext</a></li>
<li class="toctree-l1"><a class="reference internal" href="torchvision/index.html">torchvision</a></li>
<li class="toctree-l1"><a class="reference external" href="https://pytorch.org/elastic/">TorchElastic</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>Quantization</li>
<li class="pytorch-breadcrumbs-aside">
<a href="_sources/quantization.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="quantization">
<span id="quantization-doc"></span><h1>Quantization<a class="headerlink" href="#quantization" title="Permalink to this headline">¶</a></h1>
<div class="admonition warning">
<p class="admonition-title">Warning</p>
<p>Quantization is experimental and subject to change.</p>
</div>
<div class="section" id="introduction-to-quantization">
<h2>Introduction to Quantization<a class="headerlink" href="#introduction-to-quantization" title="Permalink to this headline">¶</a></h2>
<p>Quantization refers to techniques for performing computations and storing tensors at lower bitwidths than
floating point precision. A quantized model executes some or all of the operations on tensors with
integers rather than floating point values. This allows for a more
compact model representation and the use of high performance vectorized
operations on many hardware platforms. PyTorch supports INT8
quantization compared to typical FP32 models allowing for a 4x reduction in the model size and
a 4x reduction in memory bandwidth requirements. Hardware support for INT8 computations
is typically 2 to 4 times faster compared to FP32 compute. Quantization is primarily a technique
to speed up inference and only the forward pass is supported for quantized operators.</p>
<p>PyTorch supports multiple approaches to quantizing a deep learning model. In most cases the model is trained
in FP32 and then the model is converted to INT8. In addition, PyTorch also supports quantization aware
training, which models quantization errors in both the forward and backward passes using fake-quantization
modules. Note that the entire computation is carried out in floating point. At the end of quantization aware
training, PyTorch provides conversion functions to convert the trained model into lower precision.</p>
<p>At lower level, PyTorch provides a way to represent quantized tensors and
perform operations with them. They can be used to directly construct models that
perform all or part of the computation in lower precision. Higher-level APIs are
provided that incorporate typical workflows of converting FP32 model to lower
precision with minimal accuracy loss.</p>
<p>Today, PyTorch supports the following backends for running quantized operators efficiently:</p>
<ul class="simple">
<li><p>x86 CPUs with AVX2 support or higher (without AVX2 some operations have inefficient implementations)</p></li>
<li><p>ARM CPUs (typically found in mobile/embedded devices)</p></li>
</ul>
<p>The corresponding implementation is chosen automatically based on the PyTorch build mode.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>PyTorch 1.3 doesn’t provide quantized operator implementations on CUDA yet - this is direction of future work.
Move the model to CPU in order to test the quantized functionality.</p>
<p>Quantization-aware training (through <a class="reference internal" href="#torch.quantization.FakeQuantize" title="torch.quantization.FakeQuantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">FakeQuantize</span></code></a>) supports both CPU and CUDA.</p>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>When preparing a quantized model, it is necessary to ensure that qconfig and the engine used for quantized computations match
the backend on which the model will be executed. Quantization currently supports two backends: fbgemm (for use on x86,
<a class="reference external" href="https://github.com/pytorch/FBGEMM">https://github.com/pytorch/FBGEMM</a>) and qnnpack (for use on the ARM QNNPACK library <a class="reference external" href="https://github.com/pytorch/QNNPACK">https://github.com/pytorch/QNNPACK</a>).
For example, if you are interested in quantizing a model to run on ARM, it is recommended to set the qconfig by calling:</p>
<p><code class="docutils literal notranslate"><span class="pre">qconfig</span> <span class="pre">=</span> <span class="pre">torch.quantization.get_default_qconfig('qnnpack')</span></code></p>
<p>for post training quantization and</p>
<p><code class="docutils literal notranslate"><span class="pre">qconfig</span> <span class="pre">=</span> <span class="pre">torch.quantization.get_default_qat_qconfig('qnnpack')</span></code></p>
<p>for quantization aware training.</p>
<p>In addition, the torch.backends.quantized.engine parameter should be set to match the backend. For using qnnpack for inference, the
backend is set to qnnpack as follows</p>
<p><code class="docutils literal notranslate"><span class="pre">torch.backends.quantized.engine</span> <span class="pre">=</span> <span class="pre">'qnnpack'</span></code></p>
</div>
</div>
<div class="section" id="quantized-tensors">
<h2>Quantized Tensors<a class="headerlink" href="#quantized-tensors" title="Permalink to this headline">¶</a></h2>
<p>PyTorch supports both per tensor and per channel asymmetric linear
quantization. Per tensor means that all the values within the tensor are
scaled the same way. Per channel means that for each dimension, typically
the channel dimension of a tensor, the values
in the tensor are scaled and offset by a different value (effectively
the scale and offset become vectors). This allows for lesser error in converting tensors
to quantized values.</p>
<p>The mapping is performed by converting the floating point tensors using</p>
<a class="reference internal image-reference" href="_images/math-quantizer-equation.png"><img alt="_images/math-quantizer-equation.png" src="_images/math-quantizer-equation.png" style="width: 40%;" /></a>
<p>Note that, we ensure that zero in floating point is represented with no error after quantization,
thereby ensuring that operations like padding do not cause additional quantization error.</p>
<p>In order to do quantization in PyTorch, we need to be able to represent
quantized data in Tensors. A Quantized Tensor allows for storing
quantized data (represented as int8/uint8/int32) along with quantization
parameters like scale and zero_point. Quantized Tensors allow for many
useful operations making quantized arithmetic easy, in addition to
allowing for serialization of data in a quantized format.</p>
</div>
<div class="section" id="operation-coverage">
<h2>Operation coverage<a class="headerlink" href="#operation-coverage" title="Permalink to this headline">¶</a></h2>
<p>Quantized Tensors support a limited subset of data manipulation methods of the regular
full-precision tensor. (see list below)</p>
<p>For NN operators included in PyTorch, we restrict support to:</p>
<blockquote>
<div><ol class="arabic simple">
<li><p>8 bit weights (data_type = qint8)</p></li>
<li><p>8 bit activations (data_type = quint8)</p></li>
</ol>
</div></blockquote>
<p>Note that operator implementations currently only
support per channel quantization for weights of the <strong>conv</strong> and <strong>linear</strong>
operators. Furthermore the minimum and the maximum of the input data is
mapped linearly to the minimum and the maximum of the quantized data
type such that zero is represented with no quantization error.</p>
<p>Additional data types and quantization schemes can be implemented through
the <a class="reference external" href="https://pytorch.org/tutorials/advanced/torch_script_custom_ops.html">custom operator mechanism</a>.</p>
<p>Many operations for quantized tensors are available under the same API as full
float version in <code class="docutils literal notranslate"><span class="pre">torch</span></code> or <code class="docutils literal notranslate"><span class="pre">torch.nn</span></code>. Quantized version of NN modules that
perform re-quantization are available in <code class="docutils literal notranslate"><span class="pre">torch.nn.quantized</span></code>. Those
operations explicitly take output quantization parameters (scale and zero_point) in
the operation signature.</p>
<p>In addition, we also support fused versions corresponding to common fusion patterns that impact quantization at:
torch.nn.intrinsic.quantized.</p>
<p>For quantization aware training, we support modules prepared for quantization aware training at
torch.nn.qat and torch.nn.intrinsic.qat</p>
<p>Current quantized operation list is sufficient to cover typical CNN and RNN
models:</p>
<div class="section" id="quantized-torch-tensor-operations">
<h3>Quantized <code class="docutils literal notranslate"><span class="pre">torch.Tensor</span></code> operations<a class="headerlink" href="#quantized-torch-tensor-operations" title="Permalink to this headline">¶</a></h3>
<p>Operations that are available from the <code class="docutils literal notranslate"><span class="pre">torch</span></code> namespace or as methods on Tensor for quantized tensors:</p>
<ul class="simple">
<li><p><a class="reference internal" href="torch.html#torch.quantize_per_tensor" title="torch.quantize_per_tensor"><code class="xref py py-func docutils literal notranslate"><span class="pre">quantize_per_tensor()</span></code></a> - Convert float tensor to quantized tensor with per-tensor scale and zero point</p></li>
<li><p><a class="reference internal" href="torch.html#torch.quantize_per_channel" title="torch.quantize_per_channel"><code class="xref py py-func docutils literal notranslate"><span class="pre">quantize_per_channel()</span></code></a> - Convert float tensor to quantized tensor with per-channel scale and zero point</p></li>
<li><p>View-based operations like <a class="reference internal" href="tensors.html#torch.Tensor.view" title="torch.Tensor.view"><code class="xref py py-meth docutils literal notranslate"><span class="pre">view()</span></code></a>, <a class="reference internal" href="tensors.html#torch.Tensor.as_strided" title="torch.Tensor.as_strided"><code class="xref py py-meth docutils literal notranslate"><span class="pre">as_strided()</span></code></a>, <a class="reference internal" href="tensors.html#torch.Tensor.expand" title="torch.Tensor.expand"><code class="xref py py-meth docutils literal notranslate"><span class="pre">expand()</span></code></a>, <a class="reference internal" href="tensors.html#torch.Tensor.flatten" title="torch.Tensor.flatten"><code class="xref py py-meth docutils literal notranslate"><span class="pre">flatten()</span></code></a>, <a class="reference internal" href="tensors.html#torch.Tensor.select" title="torch.Tensor.select"><code class="xref py py-meth docutils literal notranslate"><span class="pre">select()</span></code></a>, python-style indexing, etc - work as on regular tensor (if quantization is not per-channel)</p></li>
<li><dl class="simple">
<dt>Comparators</dt><dd><ul>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.ne" title="torch.Tensor.ne"><code class="xref py py-meth docutils literal notranslate"><span class="pre">ne()</span></code></a> — Not equal</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.eq" title="torch.Tensor.eq"><code class="xref py py-meth docutils literal notranslate"><span class="pre">eq()</span></code></a> — Equal</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.ge" title="torch.Tensor.ge"><code class="xref py py-meth docutils literal notranslate"><span class="pre">ge()</span></code></a> — Greater or equal</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.le" title="torch.Tensor.le"><code class="xref py py-meth docutils literal notranslate"><span class="pre">le()</span></code></a> — Less or equal</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.gt" title="torch.Tensor.gt"><code class="xref py py-meth docutils literal notranslate"><span class="pre">gt()</span></code></a> — Greater</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.lt" title="torch.Tensor.lt"><code class="xref py py-meth docutils literal notranslate"><span class="pre">lt()</span></code></a> — Less</p></li>
</ul>
</dd>
</dl>
</li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.copy_" title="torch.Tensor.copy_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">copy_()</span></code></a> — Copies src to self in-place</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.clone" title="torch.Tensor.clone"><code class="xref py py-meth docutils literal notranslate"><span class="pre">clone()</span></code></a> — Returns a deep copy of the passed-in tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.dequantize" title="torch.Tensor.dequantize"><code class="xref py py-meth docutils literal notranslate"><span class="pre">dequantize()</span></code></a> — Convert quantized tensor to float tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.equal" title="torch.Tensor.equal"><code class="xref py py-meth docutils literal notranslate"><span class="pre">equal()</span></code></a> — Compares two tensors, returns true if quantization parameters and all integer elements are the same</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.int_repr" title="torch.Tensor.int_repr"><code class="xref py py-meth docutils literal notranslate"><span class="pre">int_repr()</span></code></a> — Prints the underlying integer representation of the quantized tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.max" title="torch.Tensor.max"><code class="xref py py-meth docutils literal notranslate"><span class="pre">max()</span></code></a> — Returns the maximum value of the tensor (reduction only)</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.mean" title="torch.Tensor.mean"><code class="xref py py-meth docutils literal notranslate"><span class="pre">mean()</span></code></a> — Mean function. Supported variants: reduction, dim, out</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.min" title="torch.Tensor.min"><code class="xref py py-meth docutils literal notranslate"><span class="pre">min()</span></code></a> — Returns the minimum value of the tensor (reduction only)</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.q_scale" title="torch.Tensor.q_scale"><code class="xref py py-meth docutils literal notranslate"><span class="pre">q_scale()</span></code></a> — Returns the scale of the per-tensor quantized tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.q_zero_point" title="torch.Tensor.q_zero_point"><code class="xref py py-meth docutils literal notranslate"><span class="pre">q_zero_point()</span></code></a> — Returns the zero_point of the per-tensor quantized zero point</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.q_per_channel_scales" title="torch.Tensor.q_per_channel_scales"><code class="xref py py-meth docutils literal notranslate"><span class="pre">q_per_channel_scales()</span></code></a> — Returns the scales of the per-channel quantized tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.q_per_channel_zero_points" title="torch.Tensor.q_per_channel_zero_points"><code class="xref py py-meth docutils literal notranslate"><span class="pre">q_per_channel_zero_points()</span></code></a> — Returns the zero points of the per-channel quantized tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.q_per_channel_axis" title="torch.Tensor.q_per_channel_axis"><code class="xref py py-meth docutils literal notranslate"><span class="pre">q_per_channel_axis()</span></code></a> — Returns the channel axis of the per-channel quantized tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.resize_" title="torch.Tensor.resize_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">resize_()</span></code></a> — In-place resize</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.sort" title="torch.Tensor.sort"><code class="xref py py-meth docutils literal notranslate"><span class="pre">sort()</span></code></a> — Sorts the tensor</p></li>
<li><p><a class="reference internal" href="tensors.html#torch.Tensor.topk" title="torch.Tensor.topk"><code class="xref py py-meth docutils literal notranslate"><span class="pre">topk()</span></code></a> — Returns k largest values of a tensor</p></li>
</ul>
</div>
<div class="section" id="torch-nn-functional">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.functional</span></code><a class="headerlink" href="#torch-nn-functional" title="Permalink to this headline">¶</a></h3>
<p>Basic activations are supported.</p>
<ul class="simple">
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.relu" title="torch.nn.functional.relu"><code class="xref py py-meth docutils literal notranslate"><span class="pre">relu()</span></code></a> — Rectified linear unit (copy)</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.relu_" title="torch.nn.functional.relu_"><code class="xref py py-meth docutils literal notranslate"><span class="pre">relu_()</span></code></a> — Rectified linear unit (inplace)</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.max_pool2d" title="torch.nn.functional.max_pool2d"><code class="xref py py-meth docutils literal notranslate"><span class="pre">max_pool2d()</span></code></a> - Maximum pooling</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.adaptive_avg_pool2d" title="torch.nn.functional.adaptive_avg_pool2d"><code class="xref py py-meth docutils literal notranslate"><span class="pre">adaptive_avg_pool2d()</span></code></a> - Adaptive average pooling</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.avg_pool2d" title="torch.nn.functional.avg_pool2d"><code class="xref py py-meth docutils literal notranslate"><span class="pre">avg_pool2d()</span></code></a> - Average pooling</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.interpolate" title="torch.nn.functional.interpolate"><code class="xref py py-meth docutils literal notranslate"><span class="pre">interpolate()</span></code></a> - Interpolation</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.upsample" title="torch.nn.functional.upsample"><code class="xref py py-meth docutils literal notranslate"><span class="pre">upsample()</span></code></a> - Upsampling</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.upsample_bilinear" title="torch.nn.functional.upsample_bilinear"><code class="xref py py-meth docutils literal notranslate"><span class="pre">upsample_bilinear()</span></code></a> - Bilinear Upsampling</p></li>
<li><p><a class="reference internal" href="nn.functional.html#torch.nn.functional.upsample_nearest" title="torch.nn.functional.upsample_nearest"><code class="xref py py-meth docutils literal notranslate"><span class="pre">upsample_nearest()</span></code></a> - Upsampling Nearest</p></li>
</ul>
</div>
<div class="section" id="torch-nn-intrinsic">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.intrinsic</span></code><a class="headerlink" href="#torch-nn-intrinsic" title="Permalink to this headline">¶</a></h3>
<p>Fused modules are provided for common patterns in CNNs. Combining several operations together (like convolution and relu) allows for better quantization accuracy</p>
<ul class="simple">
<li><dl class="simple">
<dt><code class="docutils literal notranslate"><span class="pre">torch.nn.intrinsic</span></code> — float versions of the modules, can be swapped with quantized version 1 to 1</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.ConvBn2d" title="torch.nn.intrinsic.ConvBn2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvBn2d</span></code></a> — Conv2d + BatchNorm</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.ConvBnReLU2d" title="torch.nn.intrinsic.ConvBnReLU2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvBnReLU2d</span></code></a> — Conv2d + BatchNorm + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.ConvReLU2d" title="torch.nn.intrinsic.ConvReLU2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvReLU2d</span></code></a> — Conv2d + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.ConvReLU3d" title="torch.nn.intrinsic.ConvReLU3d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvReLU3d</span></code></a> — Conv3d + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.LinearReLU" title="torch.nn.intrinsic.LinearReLU"><code class="xref py py-class docutils literal notranslate"><span class="pre">LinearReLU</span></code></a> — Linear + ReLU</p></li>
</ul>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt><code class="docutils literal notranslate"><span class="pre">torch.nn.intrinsic.qat</span></code> — versions of layers for quantization-aware training</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.qat.ConvBn2d" title="torch.nn.intrinsic.qat.ConvBn2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvBn2d</span></code></a> — Conv2d + BatchNorm</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.qat.ConvBnReLU2d" title="torch.nn.intrinsic.qat.ConvBnReLU2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvBnReLU2d</span></code></a> — Conv2d + BatchNorm + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.qat.ConvReLU2d" title="torch.nn.intrinsic.qat.ConvReLU2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvReLU2d</span></code></a> — Conv2d + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.qat.LinearReLU" title="torch.nn.intrinsic.qat.LinearReLU"><code class="xref py py-class docutils literal notranslate"><span class="pre">LinearReLU</span></code></a> — Linear + ReLU</p></li>
</ul>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt><code class="docutils literal notranslate"><span class="pre">torch.nn.intrinsic.quantized</span></code> — quantized version of fused layers for inference (no BatchNorm variants as it’s usually folded into convolution for inference)</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.quantized.LinearReLU" title="torch.nn.intrinsic.quantized.LinearReLU"><code class="xref py py-class docutils literal notranslate"><span class="pre">LinearReLU</span></code></a> — Linear + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.quantized.ConvReLU2d" title="torch.nn.intrinsic.quantized.ConvReLU2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvReLU2d</span></code></a> — 2D Convolution + ReLU</p></li>
<li><p><a class="reference internal" href="#torch.nn.intrinsic.quantized.ConvReLU3d" title="torch.nn.intrinsic.quantized.ConvReLU3d"><code class="xref py py-class docutils literal notranslate"><span class="pre">ConvReLU3d</span></code></a> — 3D Convolution + ReLU</p></li>
</ul>
</dd>
</dl>
</li>
</ul>
</div>
<div class="section" id="torch-nn-qat">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.qat</span></code><a class="headerlink" href="#torch-nn-qat" title="Permalink to this headline">¶</a></h3>
<p>Layers for the quantization-aware training</p>
<ul class="simple">
<li><p><a class="reference internal" href="#torch.nn.qat.Linear" title="torch.nn.qat.Linear"><code class="xref py py-class docutils literal notranslate"><span class="pre">Linear</span></code></a> — Linear (fully-connected) layer</p></li>
<li><p><a class="reference internal" href="#torch.nn.qat.Conv2d" title="torch.nn.qat.Conv2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">Conv2d</span></code></a> — 2D convolution</p></li>
</ul>
</div>
<div class="section" id="torch-quantization">
<h3><code class="docutils literal notranslate"><span class="pre">torch.quantization</span></code><a class="headerlink" href="#torch-quantization" title="Permalink to this headline">¶</a></h3>
<ul class="simple">
<li><dl class="simple">
<dt>Functions for quantization</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.quantization.add_observer_" title="torch.quantization.add_observer_"><code class="xref py py-func docutils literal notranslate"><span class="pre">add_observer_()</span></code></a> — Adds observer for the leaf modules (if quantization configuration is provided)</p></li>
<li><p><a class="reference internal" href="#torch.quantization.add_quant_dequant" title="torch.quantization.add_quant_dequant"><code class="xref py py-func docutils literal notranslate"><span class="pre">add_quant_dequant()</span></code></a>— Wraps the leaf child module using <a class="reference internal" href="#torch.quantization.QuantWrapper" title="torch.quantization.QuantWrapper"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantWrapper</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.quantization.convert" title="torch.quantization.convert"><code class="xref py py-func docutils literal notranslate"><span class="pre">convert()</span></code></a> — Converts float module with observers into its quantized counterpart. Must have quantization configuration</p></li>
<li><p><a class="reference internal" href="#torch.quantization.get_observer_dict" title="torch.quantization.get_observer_dict"><code class="xref py py-func docutils literal notranslate"><span class="pre">get_observer_dict()</span></code></a> — Traverses the module children and collects all observers into a <code class="docutils literal notranslate"><span class="pre">dict</span></code></p></li>
<li><p><a class="reference internal" href="#torch.quantization.prepare" title="torch.quantization.prepare"><code class="xref py py-func docutils literal notranslate"><span class="pre">prepare()</span></code></a> — Prepares a copy of a model for quantization</p></li>
<li><p><a class="reference internal" href="#torch.quantization.prepare_qat" title="torch.quantization.prepare_qat"><code class="xref py py-func docutils literal notranslate"><span class="pre">prepare_qat()</span></code></a> — Prepares a copy of a model for quantization aware training</p></li>
<li><p><a class="reference internal" href="#torch.quantization.propagate_qconfig_" title="torch.quantization.propagate_qconfig_"><code class="xref py py-func docutils literal notranslate"><span class="pre">propagate_qconfig_()</span></code></a> — Propagates quantization configurations through the module hierarchy and assign them to each leaf module</p></li>
<li><p><a class="reference internal" href="#torch.quantization.quantize" title="torch.quantization.quantize"><code class="xref py py-func docutils literal notranslate"><span class="pre">quantize()</span></code></a> — Converts a float module to quantized version</p></li>
<li><p><a class="reference internal" href="#torch.quantization.quantize_dynamic" title="torch.quantization.quantize_dynamic"><code class="xref py py-func docutils literal notranslate"><span class="pre">quantize_dynamic()</span></code></a> — Converts a float module to dynamically quantized version</p></li>
<li><p><a class="reference internal" href="#torch.quantization.quantize_qat" title="torch.quantization.quantize_qat"><code class="xref py py-func docutils literal notranslate"><span class="pre">quantize_qat()</span></code></a>— Converts a float module to quantized version used in quantization aware training</p></li>
<li><p><a class="reference internal" href="#torch.quantization.swap_module" title="torch.quantization.swap_module"><code class="xref py py-func docutils literal notranslate"><span class="pre">swap_module()</span></code></a> — Swaps the module with its quantized counterpart (if quantizable and if it has an observer)</p></li>
</ul>
</dd>
</dl>
</li>
<li><p><a class="reference internal" href="#torch.quantization.default_eval_fn" title="torch.quantization.default_eval_fn"><code class="xref py py-func docutils literal notranslate"><span class="pre">default_eval_fn()</span></code></a> — Default evaluation function used by the <a class="reference internal" href="#torch.quantization.quantize" title="torch.quantization.quantize"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.quantization.quantize()</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.quantization.fuse_modules" title="torch.quantization.fuse_modules"><code class="xref py py-func docutils literal notranslate"><span class="pre">fuse_modules()</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.quantization.FakeQuantize" title="torch.quantization.FakeQuantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">FakeQuantize</span></code></a> — Module for simulating the quantization/dequantization at training time</p></li>
<li><dl class="simple">
<dt>Default Observers. The rest of observers are available from <code class="docutils literal notranslate"><span class="pre">torch.quantization.observer</span></code></dt><dd><ul>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">default_observer</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">MinMaxObserver.with_args(reduce_range=True)</span></code></p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">default_weight_observer</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">MinMaxObserver.with_args(dtype=torch.qint8,</span> <span class="pre">qscheme=torch.per_tensor_symmetric)</span></code></p></li>
<li><p><code class="xref py py-class docutils literal notranslate"><span class="pre">Observer</span></code> — Abstract base class for observers</p></li>
</ul>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt>Quantization configurations</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.quantization.QConfig" title="torch.quantization.QConfig"><code class="xref py py-class docutils literal notranslate"><span class="pre">QConfig</span></code></a> — Quantization configuration class</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">default_qconfig</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">QConfig(activation=default_observer,</span> <span class="pre">weight=default_weight_observer)</span></code> (See <code class="xref py py-class docutils literal notranslate"><span class="pre">QConfig</span></code>)</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">default_qat_qconfig</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">QConfig(activation=default_fake_quant,</span> <span class="pre">weight=default_weight_fake_quant)</span></code> (See <code class="xref py py-class docutils literal notranslate"><span class="pre">QConfig</span></code>)</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">default_dynamic_qconfig</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">QConfigDynamic(weight=default_weight_observer)</span></code> (See <code class="xref py py-class docutils literal notranslate"><span class="pre">QConfigDynamic</span></code>)</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">float16_dynamic_qconfig</span></code> — Same as <code class="docutils literal notranslate"><span class="pre">QConfigDynamic(weight=NoopObserver.with_args(dtype=torch.float16))</span></code> (See <code class="xref py py-class docutils literal notranslate"><span class="pre">QConfigDynamic</span></code>)</p></li>
</ul>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt>Stubs</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.quantization.DeQuantStub" title="torch.quantization.DeQuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">DeQuantStub</span></code></a> - placeholder module for dequantize() operation in float-valued models</p></li>
<li><p><a class="reference internal" href="#torch.quantization.QuantStub" title="torch.quantization.QuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantStub</span></code></a> - placeholder module for quantize() operation in float-valued models</p></li>
<li><p><a class="reference internal" href="#torch.quantization.QuantWrapper" title="torch.quantization.QuantWrapper"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantWrapper</span></code></a> — wraps the module to be quantized. Inserts the <a class="reference internal" href="#torch.quantization.QuantStub" title="torch.quantization.QuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantStub</span></code></a> and <a class="reference internal" href="#torch.quantization.DeQuantStub" title="torch.quantization.DeQuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">DeQuantStub</span></code></a></p></li>
</ul>
</dd>
</dl>
</li>
</ul>
<p>Observers for computing the quantization parameters</p>
<ul class="simple">
<li><p><a class="reference internal" href="#torch.quantization.MinMaxObserver" title="torch.quantization.MinMaxObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">MinMaxObserver</span></code></a> — Derives the quantization parameters from the running minimum and maximum of the observed tensor inputs (per tensor variant)</p></li>
<li><p><a class="reference internal" href="#torch.quantization.MovingAverageMinMaxObserver" title="torch.quantization.MovingAverageMinMaxObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">MovingAverageMinMaxObserver</span></code></a> — Derives the quantization parameters from the running averages of the minimums and maximums of the observed tensor inputs (per tensor variant)</p></li>
<li><p><a class="reference internal" href="#torch.quantization.PerChannelMinMaxObserver" title="torch.quantization.PerChannelMinMaxObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">PerChannelMinMaxObserver</span></code></a>— Derives the quantization parameters from the running minimum and maximum of the observed tensor inputs (per channel variant)</p></li>
<li><p><a class="reference internal" href="#torch.quantization.MovingAveragePerChannelMinMaxObserver" title="torch.quantization.MovingAveragePerChannelMinMaxObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">MovingAveragePerChannelMinMaxObserver</span></code></a> — Derives the quantization parameters from the running averages of the minimums and maximums of the observed tensor inputs (per channel variant)</p></li>
<li><p><a class="reference internal" href="#torch.quantization.HistogramObserver" title="torch.quantization.HistogramObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistogramObserver</span></code></a> — Derives the quantization parameters by creating a histogram of running minimums and maximums.</p></li>
<li><dl class="simple">
<dt>Observers that do not compute the quantization parameters:</dt><dd><ul>
<li><p><a class="reference internal" href="#torch.quantization.RecordingObserver" title="torch.quantization.RecordingObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">RecordingObserver</span></code></a> — Records all incoming tensors. Used for debugging only.</p></li>
<li><p><a class="reference internal" href="#torch.quantization.NoopObserver" title="torch.quantization.NoopObserver"><code class="xref py py-class docutils literal notranslate"><span class="pre">NoopObserver</span></code></a> — Pass-through observer. Used for situation when there are no quantization parameters (i.e. quantization to <code class="docutils literal notranslate"><span class="pre">float16</span></code>)</p></li>
</ul>
</dd>
</dl>
</li>
</ul>
</div>
<div class="section" id="torch-nn-quantized">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.quantized</span></code><a class="headerlink" href="#torch-nn-quantized" title="Permalink to this headline">¶</a></h3>
<p>Quantized version of standard NN layers.</p>
<ul class="simple">
<li><p><a class="reference internal" href="#torch.nn.quantized.Quantize" title="torch.nn.quantized.Quantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">Quantize</span></code></a> — Quantization layer, used to automatically replace <a class="reference internal" href="#torch.quantization.QuantStub" title="torch.quantization.QuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantStub</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.DeQuantize" title="torch.nn.quantized.DeQuantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">DeQuantize</span></code></a> — Dequantization layer, used to replace <a class="reference internal" href="#torch.quantization.DeQuantStub" title="torch.quantization.DeQuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">DeQuantStub</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.FloatFunctional" title="torch.nn.quantized.FloatFunctional"><code class="xref py py-class docutils literal notranslate"><span class="pre">FloatFunctional</span></code></a> — Wrapper class to make stateless float operations stateful so that they can be replaced with quantized versions</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.QFunctional" title="torch.nn.quantized.QFunctional"><code class="xref py py-class docutils literal notranslate"><span class="pre">QFunctional</span></code></a> — Wrapper class for quantized versions of stateless operations like <code class="docutils literal notranslate"><span class="pre">torch.add</span></code></p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.Conv2d" title="torch.nn.quantized.Conv2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">Conv2d</span></code></a> — 2D convolution</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.Conv3d" title="torch.nn.quantized.Conv3d"><code class="xref py py-class docutils literal notranslate"><span class="pre">Conv3d</span></code></a> — 3D convolution</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.Linear" title="torch.nn.quantized.Linear"><code class="xref py py-class docutils literal notranslate"><span class="pre">Linear</span></code></a> — Linear (fully-connected) layer</p></li>
<li><p><a class="reference internal" href="nn.html#torch.nn.MaxPool2d" title="torch.nn.MaxPool2d"><code class="xref py py-class docutils literal notranslate"><span class="pre">MaxPool2d</span></code></a> — 2D max pooling</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.ReLU" title="torch.nn.quantized.ReLU"><code class="xref py py-class docutils literal notranslate"><span class="pre">ReLU</span></code></a> — Rectified linear unit</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.ReLU6" title="torch.nn.quantized.ReLU6"><code class="xref py py-class docutils literal notranslate"><span class="pre">ReLU6</span></code></a> — Rectified linear unit with cut-off at quantized representation of 6</p></li>
</ul>
</div>
<div class="section" id="torch-nn-quantized-dynamic">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.quantized.dynamic</span></code><a class="headerlink" href="#torch-nn-quantized-dynamic" title="Permalink to this headline">¶</a></h3>
<p>Layers used in dynamically quantized models (i.e. quantized only on weights)</p>
<ul class="simple">
<li><p><a class="reference internal" href="#torch.nn.quantized.dynamic.Linear" title="torch.nn.quantized.dynamic.Linear"><code class="xref py py-class docutils literal notranslate"><span class="pre">Linear</span></code></a> — Linear (fully-connected) layer</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.dynamic.LSTM" title="torch.nn.quantized.dynamic.LSTM"><code class="xref py py-class docutils literal notranslate"><span class="pre">LSTM</span></code></a> — Long-Short Term Memory RNN module</p></li>
</ul>
</div>
<div class="section" id="torch-nn-quantized-functional">
<h3><code class="docutils literal notranslate"><span class="pre">torch.nn.quantized.functional</span></code><a class="headerlink" href="#torch-nn-quantized-functional" title="Permalink to this headline">¶</a></h3>
<p>Functional versions of quantized NN layers (many of them accept explicit quantization output parameters)</p>
<ul class="simple">
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.adaptive_avg_pool2d" title="torch.nn.quantized.functional.adaptive_avg_pool2d"><code class="xref py py-func docutils literal notranslate"><span class="pre">adaptive_avg_pool2d()</span></code></a> — 2D adaptive average pooling</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.avg_pool2d" title="torch.nn.quantized.functional.avg_pool2d"><code class="xref py py-func docutils literal notranslate"><span class="pre">avg_pool2d()</span></code></a> — 2D average pooling</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.conv2d" title="torch.nn.quantized.functional.conv2d"><code class="xref py py-func docutils literal notranslate"><span class="pre">conv2d()</span></code></a> — 2D convolution</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.conv3d" title="torch.nn.quantized.functional.conv3d"><code class="xref py py-func docutils literal notranslate"><span class="pre">conv3d()</span></code></a> — 3D convolution</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.interpolate" title="torch.nn.quantized.functional.interpolate"><code class="xref py py-func docutils literal notranslate"><span class="pre">interpolate()</span></code></a> — Down-/up- sampler</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.linear" title="torch.nn.quantized.functional.linear"><code class="xref py py-func docutils literal notranslate"><span class="pre">linear()</span></code></a> — Linear (fully-connected) op</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.max_pool2d" title="torch.nn.quantized.functional.max_pool2d"><code class="xref py py-func docutils literal notranslate"><span class="pre">max_pool2d()</span></code></a> — 2D max pooling</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.relu" title="torch.nn.quantized.functional.relu"><code class="xref py py-func docutils literal notranslate"><span class="pre">relu()</span></code></a> — Rectified linear unit</p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.upsample" title="torch.nn.quantized.functional.upsample"><code class="xref py py-func docutils literal notranslate"><span class="pre">upsample()</span></code></a> — Upsampler. Will be deprecated in favor of <a class="reference internal" href="#torch.nn.quantized.functional.interpolate" title="torch.nn.quantized.functional.interpolate"><code class="xref py py-func docutils literal notranslate"><span class="pre">interpolate()</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.upsample_bilinear" title="torch.nn.quantized.functional.upsample_bilinear"><code class="xref py py-func docutils literal notranslate"><span class="pre">upsample_bilinear()</span></code></a> — Bilenear upsampler. Will be deprecated in favor of <a class="reference internal" href="#torch.nn.quantized.functional.interpolate" title="torch.nn.quantized.functional.interpolate"><code class="xref py py-func docutils literal notranslate"><span class="pre">interpolate()</span></code></a></p></li>
<li><p><a class="reference internal" href="#torch.nn.quantized.functional.upsample_nearest" title="torch.nn.quantized.functional.upsample_nearest"><code class="xref py py-func docutils literal notranslate"><span class="pre">upsample_nearest()</span></code></a> — Nearest neighbor upsampler. Will be deprecated in favor of <a class="reference internal" href="#torch.nn.quantized.functional.interpolate" title="torch.nn.quantized.functional.interpolate"><code class="xref py py-func docutils literal notranslate"><span class="pre">interpolate()</span></code></a></p></li>
</ul>
</div>
<div class="section" id="quantized-dtypes-and-quantization-schemes">
<h3>Quantized dtypes and quantization schemes<a class="headerlink" href="#quantized-dtypes-and-quantization-schemes" title="Permalink to this headline">¶</a></h3>
<ul class="simple">
<li><dl class="simple">
<dt><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.qscheme</span></code> — Type to describe the quantization scheme of a tensor. Supported types:</dt><dd><ul>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.per_tensor_affine</span></code> — per tensor, asymmetric</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.per_channel_affine</span></code> — per channel, asymmetric</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.per_tensor_symmetric</span></code> — per tensor, symmetric</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.per_channel_symmetric</span></code> — per tensor, symmetric</p></li>
</ul>
</dd>
</dl>
</li>
<li><dl class="simple">
<dt><code class="docutils literal notranslate"><span class="pre">torch.dtype</span></code> — Type to describe the data. Supported types:</dt><dd><ul>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.quint8</span></code> — 8-bit unsigned integer</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.qint8</span></code> — 8-bit signed integer</p></li>
<li><p><code class="xref py py-attr docutils literal notranslate"><span class="pre">torch.qint32</span></code> — 32-bit signed integer</p></li>
</ul>
</dd>
</dl>
</li>
</ul>
</div>
</div>
<div class="section" id="quantization-workflows">
<h2>Quantization Workflows<a class="headerlink" href="#quantization-workflows" title="Permalink to this headline">¶</a></h2>
<p>PyTorch provides three approaches to quantize models.</p>
<ol class="arabic">
<li><p>Post Training Dynamic Quantization: This is the simplest to apply form of
quantization where the weights are quantized ahead of time but the
activations are dynamically quantized during inference. This is used
for situations where the model execution time is dominated by loading
weights from memory rather than computing the matrix multiplications.
This is true for for LSTM and Transformer type models with small
batch size. Applying dynamic quantization to a whole model can be
done with a single call to <a class="reference internal" href="#torch.quantization.quantize_dynamic" title="torch.quantization.quantize_dynamic"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.quantization.quantize_dynamic()</span></code></a>.
See the <a class="reference external" href="https://pytorch.org/tutorials/#quantization-experimental">quantization tutorials</a></p></li>
<li><p>Post Training Static Quantization: This is the most commonly used form of
quantization where the weights are quantized ahead of time and the
scale factor and bias for the activation tensors is pre-computed
based on observing the behavior of the model during a calibration
process. Post Training Quantization is typically when both memory bandwidth and compute
savings are important with CNNs being a typical use case.
The general process for doing post training quantization is:</p>
<ol class="arabic simple">
<li><p>Prepare the model:
a. Specify where the activations are quantized and dequantized explicitly by adding QuantStub and DeQuantStub modules.
b. Ensure that modules are not reused.
c. Convert any operations that require requantization into modules</p></li>
<li><p>Fuse operations like conv + relu or conv+batchnorm + relu together to improve both model accuracy and performance.</p></li>
<li><p>Specify the configuration of the quantization methods ‘97 such as
selecting symmetric or asymmetric quantization and MinMax or
L2Norm calibration techniques.</p></li>
<li><p>Use the <a class="reference internal" href="#torch.quantization.prepare" title="torch.quantization.prepare"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.quantization.prepare()</span></code></a> to insert modules
that will observe activation tensors during calibration</p></li>
<li><p>Calibrate the model by running inference against a calibration
dataset</p></li>
<li><p>Finally, convert the model itself with the
torch.quantization.convert() method. This does several things: it
quantizes the weights, computes and stores the scale and bias
value to be used each activation tensor, and replaces key
operators quantized implementations.</p></li>
</ol>
<p>See the <a class="reference external" href="https://pytorch.org/tutorials/#quantization-experimental">quantization tutorials</a></p>
</li>
<li><p>Quantization Aware Training: In the rare cases where post training
quantization does not provide adequate accuracy training can be done
with simulated quantization using the <a class="reference internal" href="#torch.quantization.FakeQuantize" title="torch.quantization.FakeQuantize"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.quantization.FakeQuantize</span></code></a>. Computations
will take place in FP32 but with values clamped and rounded to
simulate the effects of INT8 quantization. The sequence of steps is
very similar.</p>
<ol class="arabic simple">
<li><p>Steps (1) and (2) are identical.</p></li>
</ol>
<ol class="arabic simple" start="3">
<li><p>Specify the configuration of the fake quantization methods ‘97 such as
selecting symmetric or asymmetric quantization and MinMax or Moving Average
or L2Norm calibration techniques.</p></li>
<li><p>Use the <a class="reference internal" href="#torch.quantization.prepare_qat" title="torch.quantization.prepare_qat"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.quantization.prepare_qat()</span></code></a> to insert modules
that will simulate quantization during training.</p></li>
<li><p>Train or fine tune the model.</p></li>
<li><p>Identical to step (6) for post training quantization</p></li>
</ol>
<p>See the <a class="reference external" href="https://pytorch.org/tutorials/#quantization-experimental">quantization tutorials</a></p>
</li>
</ol>
<p>While default implementations of observers to select the scale factor and bias
based on observed tensor data are provided, developers can provide their own
quantization functions. Quantization can be applied selectively to different
parts of the model or configured differently for different parts of the model.</p>
<p>We also provide support for per channel quantization for <strong>conv2d()</strong>,
<strong>conv3d()</strong> and <strong>linear()</strong></p>
<p>Quantization workflows work by adding (e.g. adding observers as
<code class="docutils literal notranslate"><span class="pre">.observer</span></code> submodule) or replacing (e.g. converting <code class="docutils literal notranslate"><span class="pre">nn.Conv2d</span></code> to
<code class="docutils literal notranslate"><span class="pre">nn.quantized.Conv2d</span></code>) submodules in the model’s module hierarchy. It
means that the model stays a regular <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>-based instance throughout the
process and thus can work with the rest of PyTorch APIs.</p>
</div>
<div class="section" id="model-preparation-for-quantization">
<h2>Model Preparation for Quantization<a class="headerlink" href="#model-preparation-for-quantization" title="Permalink to this headline">¶</a></h2>
<p>It is necessary to currently make some modifications to the model definition
prior to quantization. This is because currently quantization works on a module
by module basis. Specifically, for all quantization techniques, the user needs to:</p>
<ol class="arabic simple">
<li><p>Convert any operations that require output requantization (and thus have additional parameters) from functionals to module form.</p></li>
<li><p>Specify which parts of the model need to be quantized either by assigning <code class="docutils literal notranslate"><span class="pre">`.qconfig</span></code> attributes on submodules or by specifying <code class="docutils literal notranslate"><span class="pre">qconfig_dict</span></code></p></li>
</ol>
<p>For static quantization techniques which quantize activations, the user needs to do the following in addition:</p>
<ol class="arabic simple">
<li><p>Specify where activations are quantized and de-quantized. This is done using <a class="reference internal" href="#torch.quantization.QuantStub" title="torch.quantization.QuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">QuantStub</span></code></a> and <a class="reference internal" href="#torch.quantization.DeQuantStub" title="torch.quantization.DeQuantStub"><code class="xref py py-class docutils literal notranslate"><span class="pre">DeQuantStub</span></code></a> modules.</p></li>
<li><p>Use <a class="reference internal" href="#torch.nn.quantized.FloatFunctional" title="torch.nn.quantized.FloatFunctional"><code class="xref py py-class docutils literal notranslate"><span class="pre">torch.nn.quantized.FloatFunctional</span></code></a> to wrap tensor operations that require special handling for quantization into modules. Examples
are operations like <code class="docutils literal notranslate"><span class="pre">add</span></code> and <code class="docutils literal notranslate"><span class="pre">cat</span></code> which require special handling to determine output quantization parameters.</p></li>
<li><p>Fuse modules: combine operations/modules into a single module to obtain higher accuracy and performance. This is done using the
<a class="reference internal" href="#torch.quantization.fuse_modules" title="torch.quantization.fuse_modules"><code class="xref py py-func docutils literal notranslate"><span class="pre">torch.quantization.fuse_modules()</span></code></a> API, which takes in lists of modules to be fused. We currently support the following fusions:
[Conv, Relu], [Conv, BatchNorm], [Conv, BatchNorm, Relu], [Linear, Relu]</p></li>
</ol>
</div>
<div class="section" id="id1">
<h2>torch.quantization<a class="headerlink" href="#id1" title="Permalink to this headline">¶</a></h2>
<span class="target" id="module-torch.quantization"></span><p>This module implements the functions you call
directly to convert your model from FP32 to quantized form. For
example the <a class="reference internal" href="#torch.quantization.prepare" title="torch.quantization.prepare"><code class="xref py py-func docutils literal notranslate"><span class="pre">prepare()</span></code></a> is used in post training
quantization to prepares your model for the calibration step and
<a class="reference internal" href="#torch.quantization.convert" title="torch.quantization.convert"><code class="xref py py-func docutils literal notranslate"><span class="pre">convert()</span></code></a> actually converts the weights to int8 and
replaces the operations with their quantized counterparts. There are
other helper functions for things like quantizing the input to your
model and performing critical fusions like conv+relu.</p>
<div class="section" id="top-level-quantization-apis">
<h3>Top-level quantization APIs<a class="headerlink" href="#top-level-quantization-apis" title="Permalink to this headline">¶</a></h3>
<dl class="function">
<dt id="torch.quantization.quantize">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">quantize</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">run_fn</em>, <em class="sig-param">run_args</em>, <em class="sig-param">mapping=None</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#quantize"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.quantize" title="Permalink to this definition">¶</a></dt>
<dd><p>Converts a float model to quantized model.</p>
<p>First it will prepare the model for calibration or training, then it calls
<cite>run_fn</cite> which will run the calibration step or training step,
after that we will call <cite>convert</cite> which will convert the model to a
quantized model.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> – input model</p></li>
<li><p><strong>run_fn</strong> – a function for evaluating the prepared model, can be a
function that simply runs the prepared model or a training loop</p></li>
<li><p><strong>run_args</strong> – positional arguments for <cite>run_fn</cite></p></li>
<li><p><strong>inplace</strong> – carry out model transformations in-place, the original module is mutated</p></li>
<li><p><strong>mapping</strong> – correspondence between original module types and quantized counterparts</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Quantized model.</p>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.quantization.quantize_dynamic">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">quantize_dynamic</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">qconfig_spec=None</em>, <em class="sig-param">dtype=torch.qint8</em>, <em class="sig-param">mapping=None</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#quantize_dynamic"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.quantize_dynamic" title="Permalink to this definition">¶</a></dt>
<dd><p>Converts a float model to dynamic (i.e. weights-only) quantized model.</p>
<p>Replaces specified modules with dynamic weight-only quantized versions and output the quantized model.</p>
<p>For simplest usage provide <cite>dtype</cite> argument that can be float16 or qint8. Weight-only quantization
by default is performed for layers with large weights size - i.e. Linear and RNN variants.</p>
<p>Fine grained control is possible with <cite>qconfig</cite> and <cite>mapping</cite> that act similarly to <cite>quantize()</cite>.
If <cite>qconfig</cite> is provided, the <cite>dtype</cite> argument is ignored.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> – input model</p></li>
<li><p><strong>qconfig_spec</strong> – <p>Either:</p>
<ul>
<li><p>A dictionary that maps from name or type of submodule to quantization
configuration, qconfig applies to all submodules of a given
module unless qconfig for the submodules are specified (when the
submodule already has qconfig attribute). Entries in the dictionary
need to be QConfigDynamic instances.</p></li>
<li><p>A set of types and/or submodule names to apply dynamic quantization to,
in which case the <cite>dtype</cite> argument is used to specifiy the bit-width</p></li>
</ul>
</p></li>
<li><p><strong>inplace</strong> – carry out model transformations in-place, the original module is mutated</p></li>
<li><p><strong>mapping</strong> – maps type of a submodule to a type of corresponding dynamically quantized version
with which the submodule needs to be replaced</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.quantization.quantize_qat">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">quantize_qat</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">run_fn</em>, <em class="sig-param">run_args</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#quantize_qat"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.quantize_qat" title="Permalink to this definition">¶</a></dt>
<dd><p>Do quantization aware training and output a quantized model</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> – input model</p></li>
<li><p><strong>run_fn</strong> – a function for evaluating the prepared model, can be a
function that simply runs the prepared model or a training
loop</p></li>
<li><p><strong>run_args</strong> – positional arguments for <cite>run_fn</cite></p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>Quantized model.</p>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.quantization.prepare">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">prepare</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#prepare"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.prepare" title="Permalink to this definition">¶</a></dt>
<dd><p>Prepares a copy of the model for quantization calibration or quantization-aware training.</p>
<p>Quantization configuration should be assigned preemptively
to individual submodules in <cite>.qconfig</cite> attribute.</p>
<p>The model will be attached with observer or fake quant modules, and qconfig
will be propagated.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> – input model to be modified in-place</p></li>
<li><p><strong>inplace</strong> – carry out model transformations in-place, the original module is mutated</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.quantization.prepare_qat">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">prepare_qat</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">mapping=None</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#prepare_qat"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.prepare_qat" title="Permalink to this definition">¶</a></dt>
<dd><p>Prepares a copy of the model for quantization calibration or
quantization-aware training and convers it to quantized version.</p>
<p>Quantization configuration should be assigned preemptively
to individual submodules in <cite>.qconfig</cite> attribute.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> – input model to be modified in-place</p></li>
<li><p><strong>mapping</strong> – dictionary that maps float modules to quantized modules to be
replaced.</p></li>
<li><p><strong>inplace</strong> – carry out model transformations in-place, the original module
is mutated</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="function">
<dt id="torch.quantization.convert">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">convert</code><span class="sig-paren">(</span><em class="sig-param">module</em>, <em class="sig-param">mapping=None</em>, <em class="sig-param">inplace=False</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/quantize.html#convert"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.convert" title="Permalink to this definition">¶</a></dt>
<dd><p>Converts the float module with observers (where we can get quantization
parameters) to a quantized module.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>module</strong> – calibrated module with observers</p></li>
<li><p><strong>mapping</strong> – a dictionary that maps from float module type to quantized
module type, can be overwrritten to allow swapping user defined
Modules</p></li>
<li><p><strong>inplace</strong> – carry out model transformations in-place, the original module
is mutated</p></li>
</ul>
</dd>
</dl>
</dd></dl>
<dl class="class">
<dt id="torch.quantization.QConfig">
<em class="property">class </em><code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">QConfig</code><a class="reference internal" href="_modules/torch/quantization/qconfig.html#QConfig"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.QConfig" title="Permalink to this definition">¶</a></dt>
<dd><p>Describes how to quantize a layer or a part of the network by providing
settings (observer classes) for activations and weights respectively.</p>
<p>Note that QConfig needs to contain observer <strong>classes</strong> (like MinMaxObserver) or a callable that returns
instances on invocation, not the concrete observer instances themselves.
Quantization preparation function will instantiate observers multiple times for each of the layers.</p>
<p>Observer classes have usually reasonable default arguments, but they can be overwritten with <cite>with_args</cite>
method (that behaves like functools.partial):</p>
<blockquote>
<div><p>my_qconfig = QConfig(activation=MinMaxObserver.with_args(dtype=torch.qint8),
weight=default_observer.with_args(dtype=torch.qint8))</p>
</div></blockquote>
</dd></dl>
<dl class="class">
<dt id="torch.quantization.QConfigDynamic">
<em class="property">class </em><code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">QConfigDynamic</code><a class="reference internal" href="_modules/torch/quantization/qconfig.html#QConfigDynamic"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.QConfigDynamic" title="Permalink to this definition">¶</a></dt>
<dd><p>Describes how to dynamically quantize a layer or a part of the network by providing
settings (observer classe) for weights.</p>
<p>It’s like QConfig, but for dynamic quantization.</p>
<p>Note that QConfigDynamic needs to contain observer <strong>classes</strong> (like MinMaxObserver) or a callable that returns
instances on invocation, not the concrete observer instances themselves.
Quantization function will instantiate observers multiple times for each of the layers.</p>
<p>Observer classes have usually reasonable default arguments, but they can be overwritten with <cite>with_args</cite>
method (that behaves like functools.partial):</p>
<blockquote>
<div><p>my_qconfig = QConfigDynamic(weight=default_observer.with_args(dtype=torch.qint8))</p>
</div></blockquote>
</dd></dl>
</div>
<div class="section" id="preparing-model-for-quantization">
<h3>Preparing model for quantization<a class="headerlink" href="#preparing-model-for-quantization" title="Permalink to this headline">¶</a></h3>
<dl class="function">
<dt id="torch.quantization.fuse_modules">
<code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">fuse_modules</code><span class="sig-paren">(</span><em class="sig-param">model</em>, <em class="sig-param">modules_to_fuse</em>, <em class="sig-param">inplace=False</em>, <em class="sig-param">fuser_func=<function fuse_known_modules></em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/fuse_modules.html#fuse_modules"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.fuse_modules" title="Permalink to this definition">¶</a></dt>
<dd><p>Fuses a list of modules into a single module</p>
<p>Fuses only the following sequence of modules:</p>
<ul class="simple">
<li><p>conv, bn</p></li>
<li><p>conv, bn, relu</p></li>
<li><p>conv, relu</p></li>
<li><p>linear, relu</p></li>
</ul>
<p>All other sequences are left unchanged.
For these sequences, replaces the first item in the list
with the fused module, replacing the rest of the modules
with identity.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>
<dd class="field-odd"><ul class="simple">
<li><p><strong>model</strong> – Model containing the modules to be fused</p></li>
<li><p><strong>modules_to_fuse</strong> – list of list of module names to fuse. Can also be a list
of strings if there is only a single list of modules to fuse.</p></li>
<li><p><strong>inplace</strong> – bool specifying if fusion happens in place on the model, by default
a new model is returned</p></li>
<li><p><strong>fuser_func</strong> – Function that takes in a list of modules and outputs a list of fused modules
of the same length. For example,
fuser_func([convModule, BNModule]) returns the list [ConvBNModule, nn.Identity()]
Defaults to torch.quantization.fuse_known_modules</p></li>
</ul>
</dd>
<dt class="field-even">Returns</dt>
<dd class="field-even"><p>model with fused modules. A new copy is created if inplace=True.</p>
</dd>
</dl>
<p>Examples:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">myModel</span><span class="p">()</span>
<span class="gp">>>> </span><span class="c1"># m is a module containing the sub-modules below</span>
<span class="gp">>>> </span><span class="n">modules_to_fuse</span> <span class="o">=</span> <span class="p">[</span> <span class="p">[</span><span class="s1">'conv1'</span><span class="p">,</span> <span class="s1">'bn1'</span><span class="p">,</span> <span class="s1">'relu1'</span><span class="p">],</span> <span class="p">[</span><span class="s1">'submodule.conv'</span><span class="p">,</span> <span class="s1">'submodule.relu'</span><span class="p">]]</span>
<span class="gp">>>> </span><span class="n">fused_m</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">fuse_modules</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">modules_to_fuse</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output</span> <span class="o">=</span> <span class="n">fused_m</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">m</span> <span class="o">=</span> <span class="n">myModel</span><span class="p">()</span>
<span class="gp">>>> </span><span class="c1"># Alternately provide a single list of modules to fuse</span>
<span class="gp">>>> </span><span class="n">modules_to_fuse</span> <span class="o">=</span> <span class="p">[</span><span class="s1">'conv1'</span><span class="p">,</span> <span class="s1">'bn1'</span><span class="p">,</span> <span class="s1">'relu1'</span><span class="p">]</span>
<span class="gp">>>> </span><span class="n">fused_m</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">quantization</span><span class="o">.</span><span class="n">fuse_modules</span><span class="p">(</span><span class="n">m</span><span class="p">,</span> <span class="n">modules_to_fuse</span><span class="p">)</span>
<span class="gp">>>> </span><span class="n">output</span> <span class="o">=</span> <span class="n">fused_m</span><span class="p">(</span><span class="nb">input</span><span class="p">)</span>
</pre></div>
</div>
</dd></dl>
<dl class="class">
<dt id="torch.quantization.QuantStub">
<em class="property">class </em><code class="sig-prename descclassname">torch.quantization.</code><code class="sig-name descname">QuantStub</code><span class="sig-paren">(</span><em class="sig-param">qconfig=None</em><span class="sig-paren">)</span><a class="reference internal" href="_modules/torch/quantization/stubs.html#QuantStub"><span class="viewcode-link">[source]</span></a><a class="headerlink" href="#torch.quantization.QuantStub" title="Permalink to this definition">¶</a></dt>
<dd><p>Quantize stub module, before calibration, this is same as an observer,
it will be swapped as <cite>nnq.Quantize</cite> in <cite>convert</cite>.</p>
<dl class="field-list simple">
<dt class="field-odd">Parameters</dt>