Register
Login
Resources
Docs Blog Datasets Glossary Case Studies Tutorials & Webinars
Product
Data Engine LLMs Platform Enterprise
Pricing Explore
Connect to our Discord channel
Resources  
Docs Blog Datasets Glossary Case Studies Tutorials & Webinars
Explore Pricing
Product  
Data Engine LLMs Platform Enterprise
Connect to Dagshub Discord channel
Register
Login

Deci-AI
/
super-gradients
connected to https://github.com/Deci-AI/super-gradients.git

Watch 4
Issues 62 Pull Requests 44 Discussions 0

yolox_loss.html 184 KB
Permalink History Raw

You have to be logged in to leave a comment. Sign In
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
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120 
  1. <!DOCTYPE html>
    2. 

  2. <html class="writer-html5" lang="en" >
    3. 

  3. <head>
    4. 

  4.   <meta charset="utf-8" />
    5. 

  5.   <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    6. 

  6.   <title>super_gradients.training.losses.yolox_loss &mdash; SuperGradients 3.0.3 documentation</title>
    7. 

  7.       <link rel="stylesheet" href="../../../../_static/pygments.css" type="text/css" />
    8. 

  8.       <link rel="stylesheet" href="../../../../_static/css/theme.css" type="text/css" />
    9. 

  9.       <link rel="stylesheet" href="../../../../_static/graphviz.css" type="text/css" />
    10. 

  10.       <link rel="stylesheet" href="../../../../_static/custom.css" type="text/css" />
    11. 

  11.   <!--[if lt IE 9]>
    12. 

  12.     <script src="../../../../_static/js/html5shiv.min.js"></script>
    13. 

  13.   <![endif]-->
    14. 

  14.   
    15. 

  15.         <script data-url_root="../../../../" id="documentation_options" src="../../../../_static/documentation_options.js"></script>
    16. 

  16.         <script src="../../../../_static/jquery.js"></script>
    17. 

  17.         <script src="../../../../_static/underscore.js"></script>
    18. 

  18.         <script src="../../../../_static/_sphinx_javascript_frameworks_compat.js"></script>
    19. 

  19.         <script src="../../../../_static/doctools.js"></script>
    20. 

  20.         <script src="../../../../_static/sphinx_highlight.js"></script>
    21. 

  21.     <script src="../../../../_static/js/theme.js"></script>
    22. 

  22.     <link rel="index" title="Index" href="../../../../genindex.html" />
    23. 

  23.     <link rel="search" title="Search" href="../../../../search.html" /> 
    24. 

  24. </head>
    25. 

  25. 

  26. <body class="wy-body-for-nav"> 
    27. 

  27.   <div class="wy-grid-for-nav">
    28. 

  28.     <nav data-toggle="wy-nav-shift" class="wy-nav-side">
    29. 

  29.       <div class="wy-side-scroll">
    30. 

  30.         <div class="wy-side-nav-search" >
    31. 

  31.             <a href="../../../../index.html" class="icon icon-home"> SuperGradients
    32. 

  32.           </a>
    33. 

  33. <div role="search">
    34. 

  34.   <form id="rtd-search-form" class="wy-form" action="../../../../search.html" method="get">
    35. 

  35.     <input type="text" name="q" placeholder="Search docs" />
    36. 

  36.     <input type="hidden" name="check_keywords" value="yes" />
    37. 

  37.     <input type="hidden" name="area" value="default" />
    38. 

  38.   </form>
    39. 

  39. </div>
    40. 

  40.         </div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
    41. 

  41.               <p class="caption" role="heading"><span class="caption-text">Welcome To SuperGradients</span></p>
    42. 

  42. <ul>
    43. 

  43. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html">Version 3 is out! Notebooks have been updated!</a></li>
    44. 

  44. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#build-with-supergradients">Build with SuperGradients</a></li>
    45. 

  45. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#quick-installation">Quick Installation</a></li>
    46. 

  46. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#what-s-new">What’s New</a></li>
    47. 

  47. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#coming-soon">Coming soon</a></li>
    48. 

  48. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#table-of-content">Table of Content</a></li>
    49. 

  49. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#getting-started">Getting Started</a></li>
    50. 

  50. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#advanced-features">Advanced Features</a></li>
    51. 

  51. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#installation-methods">Installation Methods</a></li>
    52. 

  52. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#implemented-model-architectures">Implemented Model Architectures</a></li>
    53. 

  53. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#documentation">Documentation</a></li>
    54. 

  54. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#contributing">Contributing</a></li>
    55. 

  55. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#citation">Citation</a></li>
    56. 

  56. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#community">Community</a></li>
    57. 

  57. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#license">License</a></li>
    58. 

  58. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#deci-platform">Deci Platform</a></li>
    59. 

  59. </ul>
    60. 

  60. <p class="caption" role="heading"><span class="caption-text">Technical Documentation</span></p>
    61. 

  61. <ul>
    62. 

  62. <li class="toctree-l1"><a class="reference internal" href="../../../../super_gradients.common.html">Common package</a></li>
    63. 

  63. <li class="toctree-l1"><a class="reference internal" href="../../../../super_gradients.training.html">Training package</a></li>
    64. 

  64. </ul>
    65. 

  65. 

  66.         </div>
    67. 

  67.       </div>
    68. 

  68.     </nav>
    69. 

  69. 

  70.     <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
    71. 

  71.           <i data-toggle="wy-nav-top" class="fa fa-bars"></i>
    72. 

  72.           <a href="../../../../index.html">SuperGradients</a>
    73. 

  73.       </nav>
    74. 

  74. 

  75.       <div class="wy-nav-content">
    76. 

  76.         <div class="rst-content">
    77. 

  77.           <div role="navigation" aria-label="Page navigation">
    78. 

  78.   <ul class="wy-breadcrumbs">
    79. 

  79.       <li><a href="../../../../index.html" class="icon icon-home"></a> &raquo;</li>
    80. 

  80.           <li><a href="../../../index.html">Module code</a> &raquo;</li>
    81. 

  81.       <li>super_gradients.training.losses.yolox_loss</li>
    82. 

  82.       <li class="wy-breadcrumbs-aside">
    83. 

  83.       </li>
    84. 

  84.   </ul>
    85. 

  85.   <hr/>
    86. 

  86. </div>
    87. 

  87.           <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
    88. 

  88.            <div itemprop="articleBody">
    89. 

  89.              
    90. 

  90.   <h1>Source code for super_gradients.training.losses.yolox_loss</h1><div class="highlight"><pre>
    91. 

  91. <span></span><span class="sd">&quot;&quot;&quot;</span>
    92. 

  92. <span class="sd">Based on https://github.com/Megvii-BaseDetection/YOLOX (Apache-2.0 license)</span>
    93. 

  93. 

  94. <span class="sd">&quot;&quot;&quot;</span>
    95. 

  95. 

  96. <span class="kn">import</span> <span class="nn">logging</span>
    97. 

  97. <span class="kn">from</span> <span class="nn">typing</span> <span class="kn">import</span> <span class="n">List</span><span class="p">,</span> <span class="n">Tuple</span><span class="p">,</span> <span class="n">Union</span>
    98. 

  98. 

  99. <span class="kn">import</span> <span class="nn">torch</span>
    100. 

  100. <span class="kn">import</span> <span class="nn">torch.distributed</span> <span class="k">as</span> <span class="nn">dist</span>
    101. 

  101. <span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
    102. 

  102. <span class="kn">from</span> <span class="nn">torch.nn.modules.loss</span> <span class="kn">import</span> <span class="n">_Loss</span>
    103. 

  103. <span class="kn">import</span> <span class="nn">torch.nn.functional</span> <span class="k">as</span> <span class="nn">F</span>
    104. 

  104. 

  105. <span class="kn">from</span> <span class="nn">super_gradients.common.abstractions.abstract_logger</span> <span class="kn">import</span> <span class="n">get_logger</span>
    106. 

  106. <span class="kn">from</span> <span class="nn">super_gradients.training.utils</span> <span class="kn">import</span> <span class="n">torch_version_is_greater_or_equal</span>
    107. 

  107. <span class="kn">from</span> <span class="nn">super_gradients.training.utils.detection_utils</span> <span class="kn">import</span> <span class="n">calculate_bbox_iou_matrix</span>
    108. 

  108. 

  109. <span class="n">logger</span> <span class="o">=</span> <span class="n">get_logger</span><span class="p">(</span><span class="vm">__name__</span><span class="p">)</span>
    110. 

  110. 

  111. 

  112. <span class="k">class</span> <span class="nc">IOUloss</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
    113. 

  113.     <span class="sd">&quot;&quot;&quot;</span>
    114. 

  114. <span class="sd">    IoU loss with the following supported loss types:</span>
    115. 

  115. <span class="sd">    Attributes:</span>
    116. 

  116. <span class="sd">        reduction: str: One of [&quot;mean&quot;, &quot;sum&quot;, &quot;none&quot;] reduction to apply to the computed loss (Default=&quot;none&quot;)</span>
    117. 

  117. <span class="sd">        loss_type: str: One of [&quot;iou&quot;, &quot;giou&quot;] where:</span>
    118. 

  118. <span class="sd">            * &#39;iou&#39; for</span>
    119. 

  119. <span class="sd">                (1 - iou^2)</span>
    120. 

  120. <span class="sd">            * &#39;giou&#39; according to &quot;Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression&quot;</span>
    121. 

  121. <span class="sd">                (1 - giou), where giou = iou - (cover_box - union_box)/cover_box</span>
    122. 

  122. <span class="sd">    &quot;&quot;&quot;</span>
    123. 

  123. 

  124.     <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">reduction</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s2">&quot;none&quot;</span><span class="p">,</span> <span class="n">loss_type</span><span class="p">:</span> <span class="nb">str</span> <span class="o">=</span> <span class="s2">&quot;iou&quot;</span><span class="p">):</span>
    125. 

  125.         <span class="nb">super</span><span class="p">(</span><span class="n">IOUloss</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
    126. 

  126.         <span class="bp">self</span><span class="o">.</span><span class="n">_validate_args</span><span class="p">(</span><span class="n">loss_type</span><span class="p">,</span> <span class="n">reduction</span><span class="p">)</span>
    127. 

  127.         <span class="bp">self</span><span class="o">.</span><span class="n">reduction</span> <span class="o">=</span> <span class="n">reduction</span>
    128. 

  128.         <span class="bp">self</span><span class="o">.</span><span class="n">loss_type</span> <span class="o">=</span> <span class="n">loss_type</span>
    129. 

  129. 

  130.     <span class="nd">@staticmethod</span>
    131. 

  131.     <span class="k">def</span> <span class="nf">_validate_args</span><span class="p">(</span><span class="n">loss_type</span><span class="p">,</span> <span class="n">reduction</span><span class="p">):</span>
    132. 

  132.         <span class="n">supported_losses</span> <span class="o">=</span> <span class="p">[</span><span class="s2">&quot;iou&quot;</span><span class="p">,</span> <span class="s2">&quot;giou&quot;</span><span class="p">]</span>
    133. 

  133.         <span class="n">supported_reductions</span> <span class="o">=</span> <span class="p">[</span><span class="s2">&quot;mean&quot;</span><span class="p">,</span> <span class="s2">&quot;sum&quot;</span><span class="p">,</span> <span class="s2">&quot;none&quot;</span><span class="p">]</span>
    134. 

  134.         <span class="k">if</span> <span class="n">loss_type</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">supported_losses</span><span class="p">:</span>
    135. 

  135.             <span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s2">&quot;Illegal loss_type value: &quot;</span> <span class="o">+</span> <span class="n">loss_type</span> <span class="o">+</span> <span class="s2">&quot;, expected one of: &quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">supported_losses</span><span class="p">))</span>
    136. 

  136.         <span class="k">if</span> <span class="n">reduction</span> <span class="ow">not</span> <span class="ow">in</span> <span class="n">supported_reductions</span><span class="p">:</span>
    137. 

  137.             <span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span><span class="s2">&quot;Illegal reduction value: &quot;</span> <span class="o">+</span> <span class="n">reduction</span> <span class="o">+</span> <span class="s2">&quot;, expected one of: &quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">supported_reductions</span><span class="p">))</span>
    138. 

  138. 

  139.     <span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">pred</span><span class="p">,</span> <span class="n">target</span><span class="p">):</span>
    140. 

  140.         <span class="k">assert</span> <span class="n">pred</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="n">target</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
    141. 

  141. 

  142.         <span class="n">pred</span> <span class="o">=</span> <span class="n">pred</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
    143. 

  143.         <span class="n">target</span> <span class="o">=</span> <span class="n">target</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)</span>
    144. 

  144.         <span class="n">tl</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">((</span><span class="n">pred</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">pred</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span> <span class="p">(</span><span class="n">target</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">))</span>
    145. 

  145.         <span class="n">br</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">min</span><span class="p">((</span><span class="n">pred</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">pred</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span> <span class="p">(</span><span class="n">target</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">))</span>
    146. 

  146. 

  147.         <span class="n">area_p</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">pred</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:],</span> <span class="mi">1</span><span class="p">)</span>
    148. 

  148.         <span class="n">area_g</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:],</span> <span class="mi">1</span><span class="p">)</span>
    149. 

  149. 

  150.         <span class="n">en</span> <span class="o">=</span> <span class="p">(</span><span class="n">tl</span> <span class="o">&lt;</span> <span class="n">br</span><span class="p">)</span><span class="o">.</span><span class="n">type</span><span class="p">(</span><span class="n">tl</span><span class="o">.</span><span class="n">type</span><span class="p">())</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    151. 

  151.         <span class="n">area_i</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">br</span> <span class="o">-</span> <span class="n">tl</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="o">*</span> <span class="n">en</span>
    152. 

  152.         <span class="n">area_u</span> <span class="o">=</span> <span class="n">area_p</span> <span class="o">+</span> <span class="n">area_g</span> <span class="o">-</span> <span class="n">area_i</span>
    153. 

  153.         <span class="n">iou</span> <span class="o">=</span> <span class="p">(</span><span class="n">area_i</span><span class="p">)</span> <span class="o">/</span> <span class="p">(</span><span class="n">area_u</span> <span class="o">+</span> <span class="mf">1e-16</span><span class="p">)</span>
    154. 

  154. 

  155.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">loss_type</span> <span class="o">==</span> <span class="s2">&quot;iou&quot;</span><span class="p">:</span>
    156. 

  156.             <span class="n">loss</span> <span class="o">=</span> <span class="mi">1</span> <span class="o">-</span> <span class="n">iou</span><span class="o">**</span><span class="mi">2</span>
    157. 

  157.         <span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">loss_type</span> <span class="o">==</span> <span class="s2">&quot;giou&quot;</span><span class="p">:</span>
    158. 

  158.             <span class="n">c_tl</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">min</span><span class="p">((</span><span class="n">pred</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">pred</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span> <span class="p">(</span><span class="n">target</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">))</span>
    159. 

  159.             <span class="n">c_br</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">((</span><span class="n">pred</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">pred</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span> <span class="p">(</span><span class="n">target</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">))</span>
    160. 

  160.             <span class="n">area_c</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">c_br</span> <span class="o">-</span> <span class="n">c_tl</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    161. 

  161.             <span class="n">giou</span> <span class="o">=</span> <span class="n">iou</span> <span class="o">-</span> <span class="p">(</span><span class="n">area_c</span> <span class="o">-</span> <span class="n">area_u</span><span class="p">)</span> <span class="o">/</span> <span class="n">area_c</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="mf">1e-16</span><span class="p">)</span>
    162. 

  162.             <span class="n">loss</span> <span class="o">=</span> <span class="mi">1</span> <span class="o">-</span> <span class="n">giou</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="nb">min</span><span class="o">=-</span><span class="mf">1.0</span><span class="p">,</span> <span class="nb">max</span><span class="o">=</span><span class="mf">1.0</span><span class="p">)</span>
    163. 

  163. 

  164.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">reduction</span> <span class="o">==</span> <span class="s2">&quot;mean&quot;</span><span class="p">:</span>
    165. 

  165.             <span class="n">loss</span> <span class="o">=</span> <span class="n">loss</span><span class="o">.</span><span class="n">mean</span><span class="p">()</span>
    166. 

  166.         <span class="k">elif</span> <span class="bp">self</span><span class="o">.</span><span class="n">reduction</span> <span class="o">==</span> <span class="s2">&quot;sum&quot;</span><span class="p">:</span>
    167. 

  167.             <span class="n">loss</span> <span class="o">=</span> <span class="n">loss</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span>
    168. 

  168. 

  169.         <span class="k">return</span> <span class="n">loss</span>
    170. 

  170. 

  171. 

  172. <div class="viewcode-block" id="YoloXDetectionLoss"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss">[docs]</a><span class="k">class</span> <span class="nc">YoloXDetectionLoss</span><span class="p">(</span><span class="n">_Loss</span><span class="p">):</span>
    173. 

  173.     <span class="sd">&quot;&quot;&quot;</span>
    174. 

  174. <span class="sd">    Calculate YOLOX loss:</span>
    175. 

  175. <span class="sd">    L = L_objectivness + L_iou + L_classification + 1[use_l1]*L_l1</span>
    176. 

  176. 

  177. <span class="sd">    where:</span>
    178. 

  178. <span class="sd">        * L_iou, L_classification and L_l1 are calculated only between cells and targets that suit them;</span>
    179. 

  179. <span class="sd">        * L_objectivness is calculated for all cells.</span>
    180. 

  180. 

  181. <span class="sd">        L_classification:</span>
    182. 

  182. <span class="sd">            for cells that have suitable ground truths in their grid locations add BCEs</span>
    183. 

  183. <span class="sd">            to force a prediction of IoU with a GT in a multi-label way</span>
    184. 

  184. <span class="sd">            Coef: 1.</span>
    185. 

  185. <span class="sd">        L_iou:</span>
    186. 

  186. <span class="sd">            for cells that have suitable ground truths in their grid locations</span>
    187. 

  187. <span class="sd">            add (1 - IoU^2), IoU between a predicted box and each GT box, force maximum IoU</span>
    188. 

  188. <span class="sd">            Coef: 5.</span>
    189. 

  189. <span class="sd">        L_l1:</span>
    190. 

  190. <span class="sd">            for cells that have suitable ground truths in their grid locations</span>
    191. 

  191. <span class="sd">            l1 distance between the logits and GTs in “logits” format (the inverse of “logits to predictions” ops)</span>
    192. 

  192. <span class="sd">            Coef: 1[use_l1]</span>
    193. 

  193. <span class="sd">        L_objectness:</span>
    194. 

  194. <span class="sd">            for each cell add BCE with a label of 1 if there is GT assigned to the cell</span>
    195. 

  195. <span class="sd">            Coef: 1</span>
    196. 

  196. 

  197. <span class="sd">    Attributes:</span>
    198. 

  198. <span class="sd">        strides: list: List of Yolo levels output grid sizes (i.e [8, 16, 32]).</span>
    199. 

  199. <span class="sd">        num_classes: int: Number of classes.</span>
    200. 

  200. <span class="sd">        use_l1: bool: Controls the L_l1 Coef as discussed above (default=False).</span>
    201. 

  201. <span class="sd">        center_sampling_radius: float: Sampling radius used for center sampling when creating the fg mask (default=2.5).</span>
    202. 

  202. <span class="sd">        iou_type: str: Iou loss type, one of [&quot;iou&quot;,&quot;giou&quot;] (deafult=&quot;iou&quot;).</span>
    203. 

  203. 

  204. 

  205. <span class="sd">    &quot;&quot;&quot;</span>
    206. 

  206. 

  207.     <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">strides</span><span class="p">:</span> <span class="nb">list</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">:</span> <span class="nb">int</span><span class="p">,</span> <span class="n">use_l1</span><span class="p">:</span> <span class="nb">bool</span> <span class="o">=</span> <span class="kc">False</span><span class="p">,</span> <span class="n">center_sampling_radius</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">2.5</span><span class="p">,</span> <span class="n">iou_type</span><span class="o">=</span><span class="s2">&quot;iou&quot;</span><span class="p">):</span>
    208. 

  208.         <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
    209. 

  209.         <span class="bp">self</span><span class="o">.</span><span class="n">grids</span> <span class="o">=</span> <span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mi">1</span><span class="p">)]</span> <span class="o">*</span> <span class="nb">len</span><span class="p">(</span><span class="n">strides</span><span class="p">)</span>
    210. 

  210.         <span class="bp">self</span><span class="o">.</span><span class="n">strides</span> <span class="o">=</span> <span class="n">strides</span>
    211. 

  211.         <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span> <span class="o">=</span> <span class="n">num_classes</span>
    212. 

  212. 

  213.         <span class="bp">self</span><span class="o">.</span><span class="n">center_sampling_radius</span> <span class="o">=</span> <span class="n">center_sampling_radius</span>
    214. 

  214.         <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span> <span class="o">=</span> <span class="n">use_l1</span>
    215. 

  215.         <span class="bp">self</span><span class="o">.</span><span class="n">l1_loss</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">L1Loss</span><span class="p">(</span><span class="n">reduction</span><span class="o">=</span><span class="s2">&quot;none&quot;</span><span class="p">)</span>
    216. 

  216.         <span class="bp">self</span><span class="o">.</span><span class="n">bcewithlog_loss</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">BCEWithLogitsLoss</span><span class="p">(</span><span class="n">reduction</span><span class="o">=</span><span class="s2">&quot;none&quot;</span><span class="p">)</span>
    217. 

  217.         <span class="bp">self</span><span class="o">.</span><span class="n">iou_loss</span> <span class="o">=</span> <span class="n">IOUloss</span><span class="p">(</span><span class="n">reduction</span><span class="o">=</span><span class="s2">&quot;none&quot;</span><span class="p">,</span> <span class="n">loss_type</span><span class="o">=</span><span class="n">iou_type</span><span class="p">)</span>
    218. 

  218. 

  219.     <span class="nd">@property</span>
    220. 

  220.     <span class="k">def</span> <span class="nf">component_names</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
    221. 

  221.         <span class="sd">&quot;&quot;&quot;</span>
    222. 

  222. <span class="sd">        Component names for logging during training.</span>
    223. 

  223. <span class="sd">        These correspond to 2nd item in the tuple returned in self.forward(...).</span>
    224. 

  224. <span class="sd">        See super_gradients.Trainer.train() docs for more info.</span>
    225. 

  225. <span class="sd">        &quot;&quot;&quot;</span>
    226. 

  226.         <span class="k">return</span> <span class="p">[</span><span class="s2">&quot;iou&quot;</span><span class="p">,</span> <span class="s2">&quot;obj&quot;</span><span class="p">,</span> <span class="s2">&quot;cls&quot;</span><span class="p">,</span> <span class="s2">&quot;l1&quot;</span><span class="p">,</span> <span class="s2">&quot;num_fg&quot;</span><span class="p">,</span> <span class="s2">&quot;Loss&quot;</span><span class="p">]</span>
    227. 

  227. 

  228. <div class="viewcode-block" id="YoloXDetectionLoss.forward"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.forward">[docs]</a>    <span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">model_output</span><span class="p">:</span> <span class="n">Union</span><span class="p">[</span><span class="nb">list</span><span class="p">,</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">List</span><span class="p">]],</span> <span class="n">targets</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">):</span>
    229. 

  229.         <span class="sd">&quot;&quot;&quot;</span>
    230. 

  230. <span class="sd">        :param model_output: Union[list, Tuple[torch.Tensor, List]]:</span>
    231. 

  231. <span class="sd">             When list-</span>
    232. 

  232. <span class="sd">              output from all Yolo levels, each of shape [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    233. 

  233. <span class="sd">             And when tuple- the second item is the described list (first item is discarded)</span>
    234. 

  234. 

  235. <span class="sd">        :param targets: torch.Tensor: Num_targets x (4 + 2)], values on dim 1 are: image id in a batch, class, box x y w h</span>
    236. 

  236. 

  237. <span class="sd">        :return: loss, all losses separately in a detached tensor</span>
    238. 

  238. <span class="sd">        &quot;&quot;&quot;</span>
    239. 

  239.         <span class="k">if</span> <span class="nb">isinstance</span><span class="p">(</span><span class="n">model_output</span><span class="p">,</span> <span class="nb">tuple</span><span class="p">)</span> <span class="ow">and</span> <span class="nb">len</span><span class="p">(</span><span class="n">model_output</span><span class="p">)</span> <span class="o">==</span> <span class="mi">2</span><span class="p">:</span>
    240. 

  240.             <span class="c1"># in test/eval mode the Yolo model outputs a tuple where the second item is the raw predictions</span>
    241. 

  241.             <span class="n">_</span><span class="p">,</span> <span class="n">predictions</span> <span class="o">=</span> <span class="n">model_output</span>
    242. 

  242.         <span class="k">else</span><span class="p">:</span>
    243. 

  243.             <span class="n">predictions</span> <span class="o">=</span> <span class="n">model_output</span>
    244. 

  244. 

  245.         <span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_loss</span><span class="p">(</span><span class="n">predictions</span><span class="p">,</span> <span class="n">targets</span><span class="p">)</span></div>
    246. 

  246. 

  247.     <span class="nd">@staticmethod</span>
    248. 

  248.     <span class="k">def</span> <span class="nf">_make_grid</span><span class="p">(</span><span class="n">nx</span><span class="o">=</span><span class="mi">20</span><span class="p">,</span> <span class="n">ny</span><span class="o">=</span><span class="mi">20</span><span class="p">):</span>
    249. 

  249.         <span class="sd">&quot;&quot;&quot;</span>
    250. 

  250. <span class="sd">        Creates a tensor of xy coordinates of size (1,1,nx,ny,2)</span>
    251. 

  251. 

  252. <span class="sd">        :param nx: int: cells along x axis (default=20)</span>
    253. 

  253. <span class="sd">        :param ny: int: cells along the y axis (default=20)</span>
    254. 

  254. <span class="sd">        :return: torch.tensor of xy coordinates of size (1,1,nx,ny,2)</span>
    255. 

  255. <span class="sd">        &quot;&quot;&quot;</span>
    256. 

  256.         <span class="k">if</span> <span class="n">torch_version_is_greater_or_equal</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">10</span><span class="p">):</span>
    257. 

  257.             <span class="c1"># https://github.com/pytorch/pytorch/issues/50276</span>
    258. 

  258.             <span class="n">yv</span><span class="p">,</span> <span class="n">xv</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">meshgrid</span><span class="p">([</span><span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="n">ny</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="n">nx</span><span class="p">)],</span> <span class="n">indexing</span><span class="o">=</span><span class="s2">&quot;ij&quot;</span><span class="p">)</span>
    259. 

  259.         <span class="k">else</span><span class="p">:</span>
    260. 

  260.             <span class="n">yv</span><span class="p">,</span> <span class="n">xv</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">meshgrid</span><span class="p">([</span><span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="n">ny</span><span class="p">),</span> <span class="n">torch</span><span class="o">.</span><span class="n">arange</span><span class="p">(</span><span class="n">nx</span><span class="p">)])</span>
    261. 

  261.         <span class="k">return</span> <span class="n">torch</span><span class="o">.</span><span class="n">stack</span><span class="p">((</span><span class="n">xv</span><span class="p">,</span> <span class="n">yv</span><span class="p">),</span> <span class="mi">2</span><span class="p">)</span><span class="o">.</span><span class="n">view</span><span class="p">((</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="n">ny</span><span class="p">,</span> <span class="n">nx</span><span class="p">,</span> <span class="mi">2</span><span class="p">))</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    262. 

  262. 

  263.     <span class="k">def</span> <span class="nf">_compute_loss</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">predictions</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">],</span> <span class="n">targets</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]:</span>
    264. 

  264.         <span class="sd">&quot;&quot;&quot;</span>
    265. 

  265. <span class="sd">        :param predictions:     output from all Yolo levels, each of shape</span>
    266. 

  266. <span class="sd">                                [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    267. 

  267. <span class="sd">        :param targets:         [Num_targets x (4 + 2)], values on dim 1 are: image id in a batch, class, box x y w h</span>
    268. 

  268. 

  269. <span class="sd">        :return:                loss, all losses separately in a detached tensor</span>
    270. 

  270. <span class="sd">        &quot;&quot;&quot;</span>
    271. 

  271.         <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">transformed_outputs</span><span class="p">,</span> <span class="n">raw_outputs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">prepare_predictions</span><span class="p">(</span><span class="n">predictions</span><span class="p">)</span>
    272. 

  272. 

  273.         <span class="n">bbox_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="p">:</span><span class="mi">4</span><span class="p">]</span>  <span class="c1"># [batch, n_anchors_all, 4]</span>
    274. 

  274.         <span class="n">obj_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">]</span>  <span class="c1"># [batch, n_anchors_all, 1]</span>
    275. 

  275.         <span class="n">cls_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">5</span><span class="p">:]</span>  <span class="c1"># [batch, n_anchors_all, n_cls]</span>
    276. 

  276. 

  277.         <span class="c1"># calculate targets</span>
    278. 

  278.         <span class="n">total_num_anchors</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span>
    279. 

  279.         <span class="n">cls_targets</span> <span class="o">=</span> <span class="p">[]</span>
    280. 

  280.         <span class="n">reg_targets</span> <span class="o">=</span> <span class="p">[]</span>
    281. 

  281.         <span class="n">l1_targets</span> <span class="o">=</span> <span class="p">[]</span>
    282. 

  282.         <span class="n">obj_targets</span> <span class="o">=</span> <span class="p">[]</span>
    283. 

  283.         <span class="n">fg_masks</span> <span class="o">=</span> <span class="p">[]</span>
    284. 

  284. 

  285.         <span class="n">num_fg</span><span class="p">,</span> <span class="n">num_gts</span> <span class="o">=</span> <span class="mf">0.0</span><span class="p">,</span> <span class="mf">0.0</span>
    286. 

  286. 

  287.         <span class="k">for</span> <span class="n">image_idx</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]):</span>
    288. 

  288.             <span class="n">labels_im</span> <span class="o">=</span> <span class="n">targets</span><span class="p">[</span><span class="n">targets</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="n">image_idx</span><span class="p">]</span>
    289. 

  289.             <span class="n">num_gt</span> <span class="o">=</span> <span class="n">labels_im</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
    290. 

  290.             <span class="n">num_gts</span> <span class="o">+=</span> <span class="n">num_gt</span>
    291. 

  291.             <span class="k">if</span> <span class="n">num_gt</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
    292. 

  292.                 <span class="n">cls_target</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</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">num_classes</span><span class="p">))</span>
    293. 

  293.                 <span class="n">reg_target</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="mi">0</span><span class="p">,</span> <span class="mi">4</span><span class="p">))</span>
    294. 

  294.                 <span class="n">l1_target</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="mi">0</span><span class="p">,</span> <span class="mi">4</span><span class="p">))</span>
    295. 

  295.                 <span class="n">obj_target</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="n">total_num_anchors</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span>
    296. 

  296.                 <span class="n">fg_mask</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">(</span><span class="n">total_num_anchors</span><span class="p">)</span><span class="o">.</span><span class="n">bool</span><span class="p">()</span>
    297. 

  297.             <span class="k">else</span><span class="p">:</span>
    298. 

  298.                 <span class="c1"># GT boxes to image coordinates</span>
    299. 

  299.                 <span class="n">gt_bboxes_per_image</span> <span class="o">=</span> <span class="n">labels_im</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:</span><span class="mi">6</span><span class="p">]</span><span class="o">.</span><span class="n">clone</span><span class="p">()</span>
    300. 

  300.                 <span class="n">gt_classes</span> <span class="o">=</span> <span class="n">labels_im</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span>
    301. 

  301.                 <span class="n">bboxes_preds_per_image</span> <span class="o">=</span> <span class="n">bbox_preds</span><span class="p">[</span><span class="n">image_idx</span><span class="p">]</span>
    302. 

  302. 

  303.                 <span class="k">try</span><span class="p">:</span>
    304. 

  304.                     <span class="c1"># assign cells to ground truths, at most one GT per cell</span>
    305. 

  305.                     <span class="n">gt_matched_classes</span><span class="p">,</span> <span class="n">fg_mask</span><span class="p">,</span> <span class="n">pred_ious_this_matching</span><span class="p">,</span> <span class="n">matched_gt_inds</span><span class="p">,</span> <span class="n">num_fg_img</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_assignments</span><span class="p">(</span>
    306. 

  306.                         <span class="n">image_idx</span><span class="p">,</span>
    307. 

  307.                         <span class="n">num_gt</span><span class="p">,</span>
    308. 

  308.                         <span class="n">total_num_anchors</span><span class="p">,</span>
    309. 

  309.                         <span class="n">gt_bboxes_per_image</span><span class="p">,</span>
    310. 

  310.                         <span class="n">gt_classes</span><span class="p">,</span>
    311. 

  311.                         <span class="n">bboxes_preds_per_image</span><span class="p">,</span>
    312. 

  312.                         <span class="n">expanded_strides</span><span class="p">,</span>
    313. 

  313.                         <span class="n">x_shifts</span><span class="p">,</span>
    314. 

  314.                         <span class="n">y_shifts</span><span class="p">,</span>
    315. 

  315.                         <span class="n">cls_preds</span><span class="p">,</span>
    316. 

  316.                         <span class="n">obj_preds</span><span class="p">,</span>
    317. 

  317.                     <span class="p">)</span>
    318. 

  318. 

  319.                 <span class="c1"># TODO: CHECK IF ERROR IS CUDA OUT OF MEMORY</span>
    320. 

  320.                 <span class="k">except</span> <span class="ne">RuntimeError</span><span class="p">:</span>
    321. 

  321.                     <span class="n">logging</span><span class="o">.</span><span class="n">error</span><span class="p">(</span>
    322. 

  322.                         <span class="s2">&quot;OOM RuntimeError is raised due to the huge memory cost during label assignment. </span><span class="se">\</span>
    323. 

  323. <span class="s2">                                   CPU mode is applied in this batch. If you want to avoid this issue, </span><span class="se">\</span>
    324. 

  324. <span class="s2">                                   try to reduce the batch size or image size.&quot;</span>
    325. 

  325.                     <span class="p">)</span>
    326. 

  326.                     <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">empty_cache</span><span class="p">()</span>
    327. 

  327.                     <span class="n">gt_matched_classes</span><span class="p">,</span> <span class="n">fg_mask</span><span class="p">,</span> <span class="n">pred_ious_this_matching</span><span class="p">,</span> <span class="n">matched_gt_inds</span><span class="p">,</span> <span class="n">num_fg_img</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_assignments</span><span class="p">(</span>
    328. 

  328.                         <span class="n">image_idx</span><span class="p">,</span>
    329. 

  329.                         <span class="n">num_gt</span><span class="p">,</span>
    330. 

  330.                         <span class="n">total_num_anchors</span><span class="p">,</span>
    331. 

  331.                         <span class="n">gt_bboxes_per_image</span><span class="p">,</span>
    332. 

  332.                         <span class="n">gt_classes</span><span class="p">,</span>
    333. 

  333.                         <span class="n">bboxes_preds_per_image</span><span class="p">,</span>
    334. 

  334.                         <span class="n">expanded_strides</span><span class="p">,</span>
    335. 

  335.                         <span class="n">x_shifts</span><span class="p">,</span>
    336. 

  336.                         <span class="n">y_shifts</span><span class="p">,</span>
    337. 

  337.                         <span class="n">cls_preds</span><span class="p">,</span>
    338. 

  338.                         <span class="n">obj_preds</span><span class="p">,</span>
    339. 

  339.                         <span class="s2">&quot;cpu&quot;</span><span class="p">,</span>
    340. 

  340.                     <span class="p">)</span>
    341. 

  341. 

  342.                 <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">empty_cache</span><span class="p">()</span>
    343. 

  343.                 <span class="n">num_fg</span> <span class="o">+=</span> <span class="n">num_fg_img</span>
    344. 

  344. 

  345.                 <span class="n">cls_target</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">one_hot</span><span class="p">(</span><span class="n">gt_matched_classes</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span><span class="p">)</span> <span class="o">*</span> <span class="n">pred_ious_this_matching</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
    346. 

  346.                 <span class="n">obj_target</span> <span class="o">=</span> <span class="n">fg_mask</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
    347. 

  347.                 <span class="n">reg_target</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image</span><span class="p">[</span><span class="n">matched_gt_inds</span><span class="p">]</span>
    348. 

  348.                 <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    349. 

  349.                     <span class="n">l1_target</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_l1_target</span><span class="p">(</span>
    350. 

  350.                         <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="n">num_fg_img</span><span class="p">,</span> <span class="mi">4</span><span class="p">)),</span>
    351. 

  351.                         <span class="n">gt_bboxes_per_image</span><span class="p">[</span><span class="n">matched_gt_inds</span><span class="p">],</span>
    352. 

  352.                         <span class="n">expanded_strides</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="n">fg_mask</span><span class="p">],</span>
    353. 

  353.                         <span class="n">x_shifts</span><span class="o">=</span><span class="n">x_shifts</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="n">fg_mask</span><span class="p">],</span>
    354. 

  354.                         <span class="n">y_shifts</span><span class="o">=</span><span class="n">y_shifts</span><span class="p">[</span><span class="mi">0</span><span class="p">][</span><span class="n">fg_mask</span><span class="p">],</span>
    355. 

  355.                     <span class="p">)</span>
    356. 

  356. 

  357.             <span class="c1"># collect targets for all loss terms over the whole batch</span>
    358. 

  358.             <span class="n">cls_targets</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">cls_target</span><span class="p">)</span>
    359. 

  359.             <span class="n">reg_targets</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">reg_target</span><span class="p">)</span>
    360. 

  360.             <span class="n">obj_targets</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">obj_target</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">dtype</span><span class="p">))</span>
    361. 

  361.             <span class="n">fg_masks</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">fg_mask</span><span class="p">)</span>
    362. 

  362.             <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    363. 

  363.                 <span class="n">l1_targets</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">l1_target</span><span class="p">)</span>
    364. 

  364. 

  365.         <span class="c1"># concat all targets over the batch (get rid of batch dim)</span>
    366. 

  366.         <span class="n">cls_targets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">cls_targets</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
    367. 

  367.         <span class="n">reg_targets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">reg_targets</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
    368. 

  368.         <span class="n">obj_targets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">obj_targets</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
    369. 

  369.         <span class="n">fg_masks</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">fg_masks</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
    370. 

  370.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    371. 

  371.             <span class="n">l1_targets</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">l1_targets</span><span class="p">,</span> <span class="mi">0</span><span class="p">)</span>
    372. 

  372. 

  373.         <span class="n">num_fg</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="n">num_fg</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    374. 

  374.         <span class="c1"># loss terms divided by the total number of foregrounds</span>
    375. 

  375.         <span class="n">loss_iou</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">iou_loss</span><span class="p">(</span><span class="n">bbox_preds</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)[</span><span class="n">fg_masks</span><span class="p">],</span> <span class="n">reg_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    376. 

  376.         <span class="n">loss_obj</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bcewithlog_loss</span><span class="p">(</span><span class="n">obj_preds</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">),</span> <span class="n">obj_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    377. 

  377.         <span class="n">loss_cls</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bcewithlog_loss</span><span class="p">(</span><span class="n">cls_preds</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span><span class="p">)[</span><span class="n">fg_masks</span><span class="p">],</span> <span class="n">cls_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    378. 

  378.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    379. 

  379.             <span class="n">loss_l1</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l1_loss</span><span class="p">(</span><span class="n">raw_outputs</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">,</span> <span class="mi">4</span><span class="p">)[</span><span class="n">fg_masks</span><span class="p">],</span> <span class="n">l1_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    380. 

  380.         <span class="k">else</span><span class="p">:</span>
    381. 

  381.             <span class="n">loss_l1</span> <span class="o">=</span> <span class="mf">0.0</span>
    382. 

  382. 

  383.         <span class="n">reg_weight</span> <span class="o">=</span> <span class="mf">5.0</span>
    384. 

  384.         <span class="n">loss</span> <span class="o">=</span> <span class="n">reg_weight</span> <span class="o">*</span> <span class="n">loss_iou</span> <span class="o">+</span> <span class="n">loss_obj</span> <span class="o">+</span> <span class="n">loss_cls</span> <span class="o">+</span> <span class="n">loss_l1</span>
    385. 

  385. 

  386.         <span class="k">return</span> <span class="p">(</span>
    387. 

  387.             <span class="n">loss</span><span class="p">,</span>
    388. 

  388.             <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span>
    389. 

  389.                 <span class="p">(</span>
    390. 

  390.                     <span class="n">loss_iou</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    391. 

  391.                     <span class="n">loss_obj</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    392. 

  392.                     <span class="n">loss_cls</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    393. 

  393.                     <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="n">loss_l1</span><span class="p">)</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">loss</span><span class="o">.</span><span class="n">device</span><span class="p">),</span>
    394. 

  394.                     <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="n">num_fg</span> <span class="o">/</span> <span class="nb">max</span><span class="p">(</span><span class="n">num_gts</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">loss</span><span class="o">.</span><span class="n">device</span><span class="p">),</span>
    395. 

  395.                     <span class="n">loss</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    396. 

  396.                 <span class="p">)</span>
    397. 

  397.             <span class="p">)</span><span class="o">.</span><span class="n">detach</span><span class="p">(),</span>
    398. 

  398.         <span class="p">)</span>
    399. 

  399. 

  400. <div class="viewcode-block" id="YoloXDetectionLoss.prepare_predictions"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.prepare_predictions">[docs]</a>    <span class="k">def</span> <span class="nf">prepare_predictions</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">predictions</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">])</span> <span class="o">-&gt;</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]:</span>
    401. 

  401.         <span class="sd">&quot;&quot;&quot;</span>
    402. 

  402. <span class="sd">        Convert raw outputs of the network into a format that merges outputs from all levels</span>
    403. 

  403. <span class="sd">        :param predictions:     output from all Yolo levels, each of shape</span>
    404. 

  404. <span class="sd">                                [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    405. 

  405. <span class="sd">        :return:    5 tensors representing predictions:</span>
    406. 

  406. <span class="sd">                        * x_shifts: shape [1 x * num_cells x 1],</span>
    407. 

  407. <span class="sd">                          where num_cells = grid1X * grid1Y + grid2X * grid2Y + grid3X * grid3Y,</span>
    408. 

  408. <span class="sd">                          x coordinate on the grid cell the prediction is coming from</span>
    409. 

  409. <span class="sd">                        * y_shifts: shape [1 x  num_cells x 1],</span>
    410. 

  410. <span class="sd">                          y coordinate on the grid cell the prediction is coming from</span>
    411. 

  411. <span class="sd">                        * expanded_strides: shape [1 x num_cells x 1],</span>
    412. 

  412. <span class="sd">                          stride of the output grid the prediction is coming from</span>
    413. 

  413. <span class="sd">                        * transformed_outputs: shape [batch_size x num_cells x (num_classes + 5)],</span>
    414. 

  414. <span class="sd">                          predictions with boxes in real coordinates and logprobabilities</span>
    415. 

  415. <span class="sd">                        * raw_outputs: shape [batch_size x num_cells x (num_classes + 5)],</span>
    416. 

  416. <span class="sd">                          raw predictions with boxes and confidences as logits</span>
    417. 

  417. 

  418. <span class="sd">        &quot;&quot;&quot;</span>
    419. 

  419.         <span class="n">raw_outputs</span> <span class="o">=</span> <span class="p">[]</span>
    420. 

  420.         <span class="n">transformed_outputs</span> <span class="o">=</span> <span class="p">[]</span>
    421. 

  421.         <span class="n">x_shifts</span> <span class="o">=</span> <span class="p">[]</span>
    422. 

  422.         <span class="n">y_shifts</span> <span class="o">=</span> <span class="p">[]</span>
    423. 

  423.         <span class="n">expanded_strides</span> <span class="o">=</span> <span class="p">[]</span>
    424. 

  424.         <span class="k">for</span> <span class="n">k</span><span class="p">,</span> <span class="n">output</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">predictions</span><span class="p">):</span>
    425. 

  425.             <span class="n">batch_size</span><span class="p">,</span> <span class="n">num_anchors</span><span class="p">,</span> <span class="n">h</span><span class="p">,</span> <span class="n">w</span><span class="p">,</span> <span class="n">num_outputs</span> <span class="o">=</span> <span class="n">output</span><span class="o">.</span><span class="n">shape</span>
    426. 

  426. 

  427.             <span class="c1"># IN FIRST PASS CREATE GRIDS ACCORDING TO OUTPUT SHAPE (BATCH,1,IMAGE_H/STRIDE,IMAGE_2/STRIDE,NUM_CLASSES+5)</span>
    428. 

  428.             <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">grids</span><span class="p">[</span><span class="n">k</span><span class="p">]</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">2</span><span class="p">:</span><span class="mi">4</span><span class="p">]</span> <span class="o">!=</span> <span class="n">output</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">2</span><span class="p">:</span><span class="mi">4</span><span class="p">]:</span>
    429. 

  429.                 <span class="bp">self</span><span class="o">.</span><span class="n">grids</span><span class="p">[</span><span class="n">k</span><span class="p">]</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_make_grid</span><span class="p">(</span><span class="n">w</span><span class="p">,</span> <span class="n">h</span><span class="p">)</span><span class="o">.</span><span class="n">type_as</span><span class="p">(</span><span class="n">output</span><span class="p">)</span>
    430. 

  430. 

  431.             <span class="c1"># e.g. [batch_size, 1, 28, 28, 85] -&gt; [batch_size, 784, 85]</span>
    432. 

  432.             <span class="n">output_raveled</span> <span class="o">=</span> <span class="n">output</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="n">batch_size</span><span class="p">,</span> <span class="n">num_anchors</span> <span class="o">*</span> <span class="n">h</span> <span class="o">*</span> <span class="n">w</span><span class="p">,</span> <span class="n">num_outputs</span><span class="p">)</span>
    433. 

  433.             <span class="c1"># e.g [1, 784, 2]</span>
    434. 

  434.             <span class="n">grid_raveled</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">grids</span><span class="p">[</span><span class="n">k</span><span class="p">]</span><span class="o">.</span><span class="n">view</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">num_anchors</span> <span class="o">*</span> <span class="n">h</span> <span class="o">*</span> <span class="n">w</span><span class="p">,</span> <span class="mi">2</span><span class="p">)</span>
    435. 

  435.             <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    436. 

  436.                 <span class="c1"># e.g [1, 784, 4]</span>
    437. 

  437.                 <span class="n">raw_outputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">output_raveled</span><span class="p">[:,</span> <span class="p">:,</span> <span class="p">:</span><span class="mi">4</span><span class="p">]</span><span class="o">.</span><span class="n">clone</span><span class="p">())</span>
    438. 

  438. 

  439.             <span class="c1"># box logits to coordinates</span>
    440. 

  440.             <span class="n">centers</span> <span class="o">=</span> <span class="p">(</span><span class="n">output_raveled</span><span class="p">[</span><span class="o">...</span><span class="p">,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">grid_raveled</span><span class="p">)</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">strides</span><span class="p">[</span><span class="n">k</span><span class="p">]</span>
    441. 

  441.             <span class="n">wh</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">exp</span><span class="p">(</span><span class="n">output_raveled</span><span class="p">[</span><span class="o">...</span><span class="p">,</span> <span class="mi">2</span><span class="p">:</span><span class="mi">4</span><span class="p">])</span> <span class="o">*</span> <span class="bp">self</span><span class="o">.</span><span class="n">strides</span><span class="p">[</span><span class="n">k</span><span class="p">]</span>
    442. 

  442.             <span class="n">classes</span> <span class="o">=</span> <span class="n">output_raveled</span><span class="p">[</span><span class="o">...</span><span class="p">,</span> <span class="mi">4</span><span class="p">:]</span>
    443. 

  443.             <span class="n">output_raveled</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">([</span><span class="n">centers</span><span class="p">,</span> <span class="n">wh</span><span class="p">,</span> <span class="n">classes</span><span class="p">],</span> <span class="o">-</span><span class="mi">1</span><span class="p">)</span>
    444. 

  444. 

  445.             <span class="c1"># outputs with boxes in real coordinates, probs as logits</span>
    446. 

  446.             <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">output_raveled</span><span class="p">)</span>
    447. 

  447.             <span class="c1"># x cell coordinates of all 784 predictions, 0, 0, 0, ..., 1, 1, 1, ...</span>
    448. 

  448.             <span class="n">x_shifts</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">grid_raveled</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">0</span><span class="p">])</span>
    449. 

  449.             <span class="c1"># y cell coordinates of all 784 predictions, 0, 1, 2, ..., 0, 1, 2, ...</span>
    450. 

  450.             <span class="n">y_shifts</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">grid_raveled</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">1</span><span class="p">])</span>
    451. 

  451.             <span class="c1"># e.g. [1, 784, stride of this level (one of [8, 16, 32])]</span>
    452. 

  452.             <span class="n">expanded_strides</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">grid_raveled</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">])</span><span class="o">.</span><span class="n">fill_</span><span class="p">(</span><span class="bp">self</span><span class="o">.</span><span class="n">strides</span><span class="p">[</span><span class="n">k</span><span class="p">])</span><span class="o">.</span><span class="n">type_as</span><span class="p">(</span><span class="n">output</span><span class="p">))</span>
    453. 

  453. 

  454.         <span class="c1"># all 4 below have shapes of [batch_size , num_cells, num_values_pre_cell]</span>
    455. 

  455.         <span class="c1"># where num_anchors * num_cells is e.g. 1 * (28 * 28 + 14 * 14 + 17 * 17)</span>
    456. 

  456.         <span class="n">transformed_outputs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    457. 

  457.         <span class="n">x_shifts</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">x_shifts</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    458. 

  458.         <span class="n">y_shifts</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">y_shifts</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    459. 

  459.         <span class="n">expanded_strides</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">expanded_strides</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    460. 

  460.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    461. 

  461.             <span class="n">raw_outputs</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span><span class="n">raw_outputs</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    462. 

  462. 

  463.         <span class="k">return</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">transformed_outputs</span><span class="p">,</span> <span class="n">raw_outputs</span></div>
    464. 

  464. 

  465. <div class="viewcode-block" id="YoloXDetectionLoss.get_l1_target"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.get_l1_target">[docs]</a>    <span class="k">def</span> <span class="nf">get_l1_target</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">l1_target</span><span class="p">,</span> <span class="n">gt</span><span class="p">,</span> <span class="n">stride</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">eps</span><span class="o">=</span><span class="mf">1e-8</span><span class="p">):</span>
    466. 

  466.         <span class="sd">&quot;&quot;&quot;</span>
    467. 

  467. <span class="sd">        :param l1_target:   tensor of zeros of shape [Num_cell_gt_pairs x 4]</span>
    468. 

  468. <span class="sd">        :param gt:          targets in coordinates [Num_cell_gt_pairs x (4 + 1 + num_classes)]</span>
    469. 

  469. 

  470. <span class="sd">        :return:            targets in the format corresponding to logits</span>
    471. 

  471. <span class="sd">        &quot;&quot;&quot;</span>
    472. 

  472.         <span class="n">l1_target</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span> <span class="o">=</span> <span class="n">gt</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span> <span class="o">/</span> <span class="n">stride</span> <span class="o">-</span> <span class="n">x_shifts</span>
    473. 

  473.         <span class="n">l1_target</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span> <span class="o">=</span> <span class="n">gt</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span> <span class="o">/</span> <span class="n">stride</span> <span class="o">-</span> <span class="n">y_shifts</span>
    474. 

  474.         <span class="n">l1_target</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">gt</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">]</span> <span class="o">/</span> <span class="n">stride</span> <span class="o">+</span> <span class="n">eps</span><span class="p">)</span>
    475. 

  475.         <span class="n">l1_target</span><span class="p">[:,</span> <span class="mi">3</span><span class="p">]</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">gt</span><span class="p">[:,</span> <span class="mi">3</span><span class="p">]</span> <span class="o">/</span> <span class="n">stride</span> <span class="o">+</span> <span class="n">eps</span><span class="p">)</span>
    476. 

  476.         <span class="k">return</span> <span class="n">l1_target</span></div>
    477. 

  477. 

  478. <div class="viewcode-block" id="YoloXDetectionLoss.get_assignments"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.get_assignments">[docs]</a>    <span class="nd">@torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">()</span>
    479. 

  479.     <span class="k">def</span> <span class="nf">get_assignments</span><span class="p">(</span>
    480. 

  480.         <span class="bp">self</span><span class="p">,</span>
    481. 

  481.         <span class="n">image_idx</span><span class="p">,</span>
    482. 

  482.         <span class="n">num_gt</span><span class="p">,</span>
    483. 

  483.         <span class="n">total_num_anchors</span><span class="p">,</span>
    484. 

  484.         <span class="n">gt_bboxes_per_image</span><span class="p">,</span>
    485. 

  485.         <span class="n">gt_classes</span><span class="p">,</span>
    486. 

  486.         <span class="n">bboxes_preds_per_image</span><span class="p">,</span>
    487. 

  487.         <span class="n">expanded_strides</span><span class="p">,</span>
    488. 

  488.         <span class="n">x_shifts</span><span class="p">,</span>
    489. 

  489.         <span class="n">y_shifts</span><span class="p">,</span>
    490. 

  490.         <span class="n">cls_preds</span><span class="p">,</span>
    491. 

  491.         <span class="n">obj_preds</span><span class="p">,</span>
    492. 

  492.         <span class="n">mode</span><span class="o">=</span><span class="s2">&quot;gpu&quot;</span><span class="p">,</span>
    493. 

  493.         <span class="n">ious_loss_cost_coeff</span><span class="o">=</span><span class="mf">3.0</span><span class="p">,</span>
    494. 

  494.         <span class="n">outside_boxes_and_center_cost_coeff</span><span class="o">=</span><span class="mf">100000.0</span><span class="p">,</span>
    495. 

  495.     <span class="p">):</span>
    496. 

  496.         <span class="sd">&quot;&quot;&quot;</span>
    497. 

  497. <span class="sd">        Match cells to ground truth:</span>
    498. 

  498. <span class="sd">            * at most 1 GT per cell</span>
    499. 

  499. <span class="sd">            * dynamic number of cells per GT</span>
    500. 

  500. 

  501. <span class="sd">        :param outside_boxes_and_center_cost_coeff: float: Cost coefficiant of cells the radius and bbox of gts in dynamic</span>
    502. 

  502. <span class="sd">         matching (default=100000).</span>
    503. 

  503. <span class="sd">        :param ious_loss_cost_coeff: float: Cost coefficiant for iou loss in dynamic matching (default=3).</span>
    504. 

  504. <span class="sd">        :param image_idx: int: Image index in batch.</span>
    505. 

  505. <span class="sd">        :param num_gt: int: Number of ground trunth targets in the image.</span>
    506. 

  506. <span class="sd">        :param total_num_anchors: int: Total number of possible bboxes = sum of all grid cells.</span>
    507. 

  507. <span class="sd">        :param gt_bboxes_per_image: torch.Tensor: Tensor of gt bboxes for  the image, shape: (num_gt, 4).</span>
    508. 

  508. <span class="sd">        :param gt_classes: torch.Tesnor: Tensor of the classes in the image, shape: (num_preds,4).</span>
    509. 

  509. <span class="sd">        :param bboxes_preds_per_image: Tensor of the classes in the image, shape: (num_preds).</span>
    510. 

  510. <span class="sd">        :param expanded_strides: torch.Tensor: Stride of the output grid the prediction is coming from,</span>
    511. 

  511. <span class="sd">            shape (1 x num_cells x 1).</span>
    512. 

  512. <span class="sd">        :param x_shifts: torch.Tensor: X&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    513. 

  513. <span class="sd">        :param y_shifts: torch.Tensor: Y&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    514. 

  514. <span class="sd">        :param cls_preds: torch.Tensor: Class predictions in all cells, shape (batch_size, num_cells).</span>
    515. 

  515. <span class="sd">        :param obj_preds: torch.Tensor: Objectness predictions in all cells, shape (batch_size, num_cells).</span>
    516. 

  516. <span class="sd">        :param mode: str: One of [&quot;gpu&quot;,&quot;cpu&quot;], Controls the device the assignment operation should be taken place on (deafult=&quot;gpu&quot;)</span>
    517. 

  517. 

  518. <span class="sd">        &quot;&quot;&quot;</span>
    519. 

  519.         <span class="k">if</span> <span class="n">mode</span> <span class="o">==</span> <span class="s2">&quot;cpu&quot;</span><span class="p">:</span>
    520. 

  520.             <span class="nb">print</span><span class="p">(</span><span class="s2">&quot;------------CPU Mode for This Batch-------------&quot;</span><span class="p">)</span>
    521. 

  521.             <span class="n">gt_bboxes_per_image</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    522. 

  522.             <span class="n">bboxes_preds_per_image</span> <span class="o">=</span> <span class="n">bboxes_preds_per_image</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    523. 

  523.             <span class="n">gt_classes</span> <span class="o">=</span> <span class="n">gt_classes</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    524. 

  524.             <span class="n">expanded_strides</span> <span class="o">=</span> <span class="n">expanded_strides</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    525. 

  525.             <span class="n">x_shifts</span> <span class="o">=</span> <span class="n">x_shifts</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
    526. 

  526.             <span class="n">y_shifts</span> <span class="o">=</span> <span class="n">y_shifts</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
    527. 

  527. 

  528.         <span class="c1"># create a mask for foreground cells</span>
    529. 

  529.         <span class="n">fg_mask</span><span class="p">,</span> <span class="n">is_in_boxes_and_center</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_in_boxes_info</span><span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">)</span>
    530. 

  530. 

  531.         <span class="n">bboxes_preds_per_image</span> <span class="o">=</span> <span class="n">bboxes_preds_per_image</span><span class="p">[</span><span class="n">fg_mask</span><span class="p">]</span>
    532. 

  532.         <span class="n">cls_preds_</span> <span class="o">=</span> <span class="n">cls_preds</span><span class="p">[</span><span class="n">image_idx</span><span class="p">][</span><span class="n">fg_mask</span><span class="p">]</span>
    533. 

  533.         <span class="n">obj_preds_</span> <span class="o">=</span> <span class="n">obj_preds</span><span class="p">[</span><span class="n">image_idx</span><span class="p">][</span><span class="n">fg_mask</span><span class="p">]</span>
    534. 

  534.         <span class="n">num_in_boxes_anchor</span> <span class="o">=</span> <span class="n">bboxes_preds_per_image</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
    535. 

  535. 

  536.         <span class="k">if</span> <span class="n">mode</span> <span class="o">==</span> <span class="s2">&quot;cpu&quot;</span><span class="p">:</span>
    537. 

  537.             <span class="n">gt_bboxes_per_image</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
    538. 

  538.             <span class="n">bboxes_preds_per_image</span> <span class="o">=</span> <span class="n">bboxes_preds_per_image</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
    539. 

  539. 

  540.         <span class="c1"># calculate cost between all foregrounds and all ground truths (used only for matching)</span>
    541. 

  541.         <span class="n">pair_wise_ious</span> <span class="o">=</span> <span class="n">calculate_bbox_iou_matrix</span><span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">,</span> <span class="n">bboxes_preds_per_image</span><span class="p">,</span> <span class="n">x1y1x2y2</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
    542. 

  542.         <span class="n">gt_cls_per_image</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">one_hot</span><span class="p">(</span><span class="n">gt_classes</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span><span class="p">)</span>
    543. 

  543.         <span class="n">gt_cls_per_image</span> <span class="o">=</span> <span class="n">gt_cls_per_image</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">num_in_boxes_anchor</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    544. 

  544.         <span class="n">pair_wise_ious_loss</span> <span class="o">=</span> <span class="o">-</span><span class="n">torch</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">pair_wise_ious</span> <span class="o">+</span> <span class="mf">1e-8</span><span class="p">)</span>
    545. 

  545. 

  546.         <span class="k">if</span> <span class="n">mode</span> <span class="o">==</span> <span class="s2">&quot;cpu&quot;</span><span class="p">:</span>
    547. 

  547.             <span class="n">cls_preds_</span><span class="p">,</span> <span class="n">obj_preds_</span> <span class="o">=</span> <span class="n">cls_preds_</span><span class="o">.</span><span class="n">cpu</span><span class="p">(),</span> <span class="n">obj_preds_</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span>
    548. 

  548. 

  549.         <span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">amp</span><span class="o">.</span><span class="n">autocast</span><span class="p">(</span><span class="n">enabled</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
    550. 

  550.             <span class="n">cls_preds_</span> <span class="o">=</span> <span class="n">cls_preds_</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="n">num_gt</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">sigmoid_</span><span class="p">()</span> <span class="o">*</span> <span class="n">obj_preds_</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="n">num_gt</span><span class="p">,</span> <span class="mi">1</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">sigmoid_</span><span class="p">()</span>
    551. 

  551.             <span class="n">pair_wise_cls_loss</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">binary_cross_entropy</span><span class="p">(</span><span class="n">cls_preds_</span><span class="o">.</span><span class="n">sqrt_</span><span class="p">(),</span> <span class="n">gt_cls_per_image</span><span class="p">,</span> <span class="n">reduction</span><span class="o">=</span><span class="s2">&quot;none&quot;</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
    552. 

  552.         <span class="k">del</span> <span class="n">cls_preds_</span>
    553. 

  553. 

  554.         <span class="n">cost</span> <span class="o">=</span> <span class="n">pair_wise_cls_loss</span> <span class="o">+</span> <span class="n">ious_loss_cost_coeff</span> <span class="o">*</span> <span class="n">pair_wise_ious_loss</span> <span class="o">+</span> <span class="n">outside_boxes_and_center_cost_coeff</span> <span class="o">*</span> <span class="p">(</span><span class="o">~</span><span class="n">is_in_boxes_and_center</span><span class="p">)</span>
    555. 

  555. 

  556.         <span class="c1"># further filter foregrounds: create pairs between cells and ground truth, based on cost and IoUs</span>
    557. 

  557.         <span class="n">num_fg</span><span class="p">,</span> <span class="n">gt_matched_classes</span><span class="p">,</span> <span class="n">pred_ious_this_matching</span><span class="p">,</span> <span class="n">matched_gt_inds</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">dynamic_k_matching</span><span class="p">(</span><span class="n">cost</span><span class="p">,</span> <span class="n">pair_wise_ious</span><span class="p">,</span> <span class="n">gt_classes</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">,</span> <span class="n">fg_mask</span><span class="p">)</span>
    558. 

  558.         <span class="c1"># discard tensors related to cost</span>
    559. 

  559.         <span class="k">del</span> <span class="n">pair_wise_cls_loss</span><span class="p">,</span> <span class="n">cost</span><span class="p">,</span> <span class="n">pair_wise_ious</span><span class="p">,</span> <span class="n">pair_wise_ious_loss</span>
    560. 

  560. 

  561.         <span class="k">if</span> <span class="n">mode</span> <span class="o">==</span> <span class="s2">&quot;cpu&quot;</span><span class="p">:</span>
    562. 

  562.             <span class="n">gt_matched_classes</span> <span class="o">=</span> <span class="n">gt_matched_classes</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
    563. 

  563.             <span class="n">fg_mask</span> <span class="o">=</span> <span class="n">fg_mask</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
    564. 

  564.             <span class="n">pred_ious_this_matching</span> <span class="o">=</span> <span class="n">pred_ious_this_matching</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
    565. 

  565.             <span class="n">matched_gt_inds</span> <span class="o">=</span> <span class="n">matched_gt_inds</span><span class="o">.</span><span class="n">cuda</span><span class="p">()</span>
    566. 

  566. 

  567.         <span class="k">return</span> <span class="n">gt_matched_classes</span><span class="p">,</span> <span class="n">fg_mask</span><span class="p">,</span> <span class="n">pred_ious_this_matching</span><span class="p">,</span> <span class="n">matched_gt_inds</span><span class="p">,</span> <span class="n">num_fg</span></div>
    568. 

  568. 

  569. <div class="viewcode-block" id="YoloXDetectionLoss.get_in_boxes_info"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.get_in_boxes_info">[docs]</a>    <span class="k">def</span> <span class="nf">get_in_boxes_info</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">gt_bboxes_per_image</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">):</span>
    570. 

  570.         <span class="sd">&quot;&quot;&quot;</span>
    571. 

  571. <span class="sd">        Create a mask for all cells, mask in only foreground: cells that have a center located:</span>
    572. 

  572. <span class="sd">            * withing a GT box;</span>
    573. 

  573. <span class="sd">            OR</span>
    574. 

  574. <span class="sd">            * within a fixed radius around a GT box (center sampling);</span>
    575. 

  575. 

  576. <span class="sd">        :param num_gt: int: Number of ground trunth targets in the image.</span>
    577. 

  577. <span class="sd">        :param total_num_anchors: int: Sum of all grid cells.</span>
    578. 

  578. <span class="sd">        :param gt_bboxes_per_image: torch.Tensor: Tensor of gt bboxes for  the image, shape: (num_gt, 4).</span>
    579. 

  579. <span class="sd">        :param expanded_strides: torch.Tensor: Stride of the output grid the prediction is coming from,</span>
    580. 

  580. <span class="sd">            shape (1 x num_cells x 1).</span>
    581. 

  581. <span class="sd">        :param x_shifts: torch.Tensor: X&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    582. 

  582. <span class="sd">        :param y_shifts: torch.Tensor: Y&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    583. 

  583. 

  584. <span class="sd">        :return is_in_boxes_anchor, is_in_boxes_and_center</span>
    585. 

  585. <span class="sd">            where:</span>
    586. 

  586. <span class="sd">             - is_in_boxes_anchor masks the cells that their cell center is  inside a gt bbox and within</span>
    587. 

  587. <span class="sd">                self.center_sampling_radius cells away, without reduction (i.e shape=(num_gts, num_fgs))</span>
    588. 

  588. <span class="sd">             - is_in_boxes_and_center masks the cells that their center is either inside a gt bbox or within</span>
    589. 

  589. <span class="sd">                self.center_sampling_radius cells away, shape (num_fgs)</span>
    590. 

  590. <span class="sd">        &quot;&quot;&quot;</span>
    591. 

  591. 

  592.         <span class="n">expanded_strides_per_image</span> <span class="o">=</span> <span class="n">expanded_strides</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
    593. 

  593. 

  594.         <span class="c1"># cell coordinates, shape [n_predictions] -&gt; repeated to [n_gts, n_predictions]</span>
    595. 

  595.         <span class="n">x_shifts_per_image</span> <span class="o">=</span> <span class="n">x_shifts</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span>
    596. 

  596.         <span class="n">y_shifts_per_image</span> <span class="o">=</span> <span class="n">y_shifts</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span>
    597. 

  597.         <span class="n">x_centers_per_image</span> <span class="o">=</span> <span class="p">(</span><span class="n">x_shifts_per_image</span> <span class="o">+</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="p">)</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="n">num_gt</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    598. 

  598.         <span class="n">y_centers_per_image</span> <span class="o">=</span> <span class="p">(</span><span class="n">y_shifts_per_image</span> <span class="o">+</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="p">)</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="n">num_gt</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span>
    599. 

  599. 

  600.         <span class="c1"># FIND CELL CENTERS THAT ARE WITHIN GROUND TRUTH BOXES</span>
    601. 

  601. 

  602.         <span class="c1"># ground truth boxes, shape [n_gts] -&gt; repeated to [n_gts, n_predictions]</span>
    603. 

  603.         <span class="c1"># from (c1, c2, w, h) to left, right, top, bottom</span>
    604. 

  604.         <span class="n">gt_bboxes_per_image_l</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span> <span class="o">-</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span>
    605. 

  605.         <span class="n">gt_bboxes_per_image_r</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span> <span class="o">+</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span>
    606. 

  606.         <span class="n">gt_bboxes_per_image_t</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span> <span class="o">-</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">3</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span>
    607. 

  607.         <span class="n">gt_bboxes_per_image_b</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span> <span class="o">+</span> <span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">3</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span>
    608. 

  608. 

  609.         <span class="c1"># check which cell centers lay within the ground truth boxes</span>
    610. 

  610.         <span class="n">b_l</span> <span class="o">=</span> <span class="n">x_centers_per_image</span> <span class="o">-</span> <span class="n">gt_bboxes_per_image_l</span>  <span class="c1"># x - l &gt; 0 when l is on the lest from x</span>
    611. 

  611.         <span class="n">b_r</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image_r</span> <span class="o">-</span> <span class="n">x_centers_per_image</span>
    612. 

  612.         <span class="n">b_t</span> <span class="o">=</span> <span class="n">y_centers_per_image</span> <span class="o">-</span> <span class="n">gt_bboxes_per_image_t</span>
    613. 

  613.         <span class="n">b_b</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image_b</span> <span class="o">-</span> <span class="n">y_centers_per_image</span>
    614. 

  614.         <span class="n">bbox_deltas</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">b_l</span><span class="p">,</span> <span class="n">b_t</span><span class="p">,</span> <span class="n">b_r</span><span class="p">,</span> <span class="n">b_b</span><span class="p">],</span> <span class="mi">2</span><span class="p">)</span>  <span class="c1"># shape [n_gts, n_predictions]</span>
    615. 

  615. 

  616.         <span class="c1"># to claim that a cell center is inside a gt box all 4 differences calculated above should be positive</span>
    617. 

  617.         <span class="n">is_in_boxes</span> <span class="o">=</span> <span class="n">bbox_deltas</span><span class="o">.</span><span class="n">min</span><span class="p">(</span><span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">values</span> <span class="o">&gt;</span> <span class="mf">0.0</span>  <span class="c1"># shape [n_gts, n_predictions]</span>
    618. 

  618.         <span class="n">is_in_boxes_all</span> <span class="o">=</span> <span class="n">is_in_boxes</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span> <span class="o">&gt;</span> <span class="mi">0</span>  <span class="c1"># shape [n_predictions], whether a cell is inside at least one gt</span>
    619. 

  619. 

  620.         <span class="c1"># FIND CELL CENTERS THAT ARE WITHIN +- self.center_sampling_radius CELLS FROM GROUND TRUTH BOXES CENTERS</span>
    621. 

  621. 

  622.         <span class="c1"># define fake boxes: instead of ground truth boxes step +- self.center_sampling_radius from their centers</span>
    623. 

  623.         <span class="n">gt_bboxes_per_image_l</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span>
    624. 

  624.             <span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span>
    625. 

  625.         <span class="p">)</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">center_sampling_radius</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    626. 

  626.         <span class="n">gt_bboxes_per_image_r</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span>
    627. 

  627.             <span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span>
    628. 

  628.         <span class="p">)</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">center_sampling_radius</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    629. 

  629.         <span class="n">gt_bboxes_per_image_t</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span>
    630. 

  630.             <span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span>
    631. 

  631.         <span class="p">)</span> <span class="o">-</span> <span class="bp">self</span><span class="o">.</span><span class="n">center_sampling_radius</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    632. 

  632.         <span class="n">gt_bboxes_per_image_b</span> <span class="o">=</span> <span class="p">(</span><span class="n">gt_bboxes_per_image</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">repeat</span><span class="p">(</span>
    633. 

  633.             <span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span>
    634. 

  634.         <span class="p">)</span> <span class="o">+</span> <span class="bp">self</span><span class="o">.</span><span class="n">center_sampling_radius</span> <span class="o">*</span> <span class="n">expanded_strides_per_image</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    635. 

  635. 

  636.         <span class="n">c_l</span> <span class="o">=</span> <span class="n">x_centers_per_image</span> <span class="o">-</span> <span class="n">gt_bboxes_per_image_l</span>
    637. 

  637.         <span class="n">c_r</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image_r</span> <span class="o">-</span> <span class="n">x_centers_per_image</span>
    638. 

  638.         <span class="n">c_t</span> <span class="o">=</span> <span class="n">y_centers_per_image</span> <span class="o">-</span> <span class="n">gt_bboxes_per_image_t</span>
    639. 

  639.         <span class="n">c_b</span> <span class="o">=</span> <span class="n">gt_bboxes_per_image_b</span> <span class="o">-</span> <span class="n">y_centers_per_image</span>
    640. 

  640.         <span class="n">center_deltas</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">c_l</span><span class="p">,</span> <span class="n">c_t</span><span class="p">,</span> <span class="n">c_r</span><span class="p">,</span> <span class="n">c_b</span><span class="p">],</span> <span class="mi">2</span><span class="p">)</span>
    641. 

  641.         <span class="n">is_in_centers</span> <span class="o">=</span> <span class="n">center_deltas</span><span class="o">.</span><span class="n">min</span><span class="p">(</span><span class="n">dim</span><span class="o">=-</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">values</span> <span class="o">&gt;</span> <span class="mf">0.0</span>
    642. 

  642.         <span class="n">is_in_centers_all</span> <span class="o">=</span> <span class="n">is_in_centers</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span> <span class="o">&gt;</span> <span class="mi">0</span>
    643. 

  643. 

  644.         <span class="c1"># in boxes OR in centers</span>
    645. 

  645.         <span class="n">is_in_boxes_anchor</span> <span class="o">=</span> <span class="n">is_in_boxes_all</span> <span class="o">|</span> <span class="n">is_in_centers_all</span>
    646. 

  646. 

  647.         <span class="c1"># in boxes AND in centers, preserving a shape [num_GTs x num_FGs]</span>
    648. 

  648.         <span class="n">is_in_boxes_and_center</span> <span class="o">=</span> <span class="n">is_in_boxes</span><span class="p">[:,</span> <span class="n">is_in_boxes_anchor</span><span class="p">]</span> <span class="o">&amp;</span> <span class="n">is_in_centers</span><span class="p">[:,</span> <span class="n">is_in_boxes_anchor</span><span class="p">]</span>
    649. 

  649.         <span class="k">return</span> <span class="n">is_in_boxes_anchor</span><span class="p">,</span> <span class="n">is_in_boxes_and_center</span></div>
    650. 

  650. 

  651. <div class="viewcode-block" id="YoloXDetectionLoss.dynamic_k_matching"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXDetectionLoss.dynamic_k_matching">[docs]</a>    <span class="k">def</span> <span class="nf">dynamic_k_matching</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">cost</span><span class="p">,</span> <span class="n">pair_wise_ious</span><span class="p">,</span> <span class="n">gt_classes</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">,</span> <span class="n">fg_mask</span><span class="p">):</span>
    652. 

  652.         <span class="sd">&quot;&quot;&quot;</span>
    653. 

  653. <span class="sd">        :param cost:            pairwise cost, [num_FGs x num_GTs]</span>
    654. 

  654. <span class="sd">        :param pair_wise_ious:  pairwise IoUs, [num_FGs x num_GTs]</span>
    655. 

  655. <span class="sd">        :param gt_classes:      class of each GT</span>
    656. 

  656. <span class="sd">        :param num_gt:          number of GTs</span>
    657. 

  657. 

  658. <span class="sd">        :return num_fg, (number of foregrounds)</span>
    659. 

  659. <span class="sd">                gt_matched_classes, (the classes that have been matched with fgs)</span>
    660. 

  660. <span class="sd">                pred_ious_this_matching</span>
    661. 

  661. <span class="sd">                matched_gt_inds</span>
    662. 

  662. <span class="sd">        &quot;&quot;&quot;</span>
    663. 

  663.         <span class="c1"># create a matrix with shape [num_GTs x num_FGs]</span>
    664. 

  664.         <span class="n">matching_matrix</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">cost</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">uint8</span><span class="p">)</span>
    665. 

  665. 

  666.         <span class="c1"># for each GT get a dynamic k of foregrounds with a minimum cost: k = int(sum[top 10 IoUs])</span>
    667. 

  667.         <span class="n">ious_in_boxes_matrix</span> <span class="o">=</span> <span class="n">pair_wise_ious</span>
    668. 

  668.         <span class="n">n_candidate_k</span> <span class="o">=</span> <span class="nb">min</span><span class="p">(</span><span class="mi">10</span><span class="p">,</span> <span class="n">ious_in_boxes_matrix</span><span class="o">.</span><span class="n">size</span><span class="p">(</span><span class="mi">1</span><span class="p">))</span>
    669. 

  669.         <span class="n">topk_ious</span><span class="p">,</span> <span class="n">_</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">topk</span><span class="p">(</span><span class="n">ious_in_boxes_matrix</span><span class="p">,</span> <span class="n">n_candidate_k</span><span class="p">,</span> <span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    670. 

  670.         <span class="n">dynamic_ks</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="n">topk_ious</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">int</span><span class="p">(),</span> <span class="nb">min</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    671. 

  671.         <span class="n">dynamic_ks</span> <span class="o">=</span> <span class="n">dynamic_ks</span><span class="o">.</span><span class="n">tolist</span><span class="p">()</span>
    672. 

  672.         <span class="k">for</span> <span class="n">gt_idx</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="n">num_gt</span><span class="p">):</span>
    673. 

  673.             <span class="k">try</span><span class="p">:</span>
    674. 

  674.                 <span class="n">_</span><span class="p">,</span> <span class="n">pos_idx</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">topk</span><span class="p">(</span><span class="n">cost</span><span class="p">[</span><span class="n">gt_idx</span><span class="p">],</span> <span class="n">k</span><span class="o">=</span><span class="n">dynamic_ks</span><span class="p">[</span><span class="n">gt_idx</span><span class="p">],</span> <span class="n">largest</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
    675. 

  675.             <span class="k">except</span> <span class="ne">Exception</span><span class="p">:</span>
    676. 

  676.                 <span class="n">logger</span><span class="o">.</span><span class="n">warning</span><span class="p">(</span><span class="s2">&quot;cost[gt_idx]: &quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">cost</span><span class="p">[</span><span class="n">gt_idx</span><span class="p">])</span> <span class="o">+</span> <span class="s2">&quot; dynamic_ks[gt_idx]L &quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">dynamic_ks</span><span class="p">[</span><span class="n">gt_idx</span><span class="p">]))</span>
    677. 

  677.             <span class="n">matching_matrix</span><span class="p">[</span><span class="n">gt_idx</span><span class="p">][</span><span class="n">pos_idx</span><span class="p">]</span> <span class="o">=</span> <span class="mi">1</span>
    678. 

  678. 

  679.         <span class="k">del</span> <span class="n">topk_ious</span><span class="p">,</span> <span class="n">dynamic_ks</span><span class="p">,</span> <span class="n">pos_idx</span>
    680. 

  680. 

  681.         <span class="c1"># leave at most one GT per foreground, chose the one with the smallest cost</span>
    682. 

  682.         <span class="n">anchor_matching_gt</span> <span class="o">=</span> <span class="n">matching_matrix</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    683. 

  683.         <span class="k">if</span> <span class="p">(</span><span class="n">anchor_matching_gt</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">&gt;</span> <span class="mi">0</span><span class="p">:</span>
    684. 

  684.             <span class="n">_</span><span class="p">,</span> <span class="n">cost_argmin</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">min</span><span class="p">(</span><span class="n">cost</span><span class="p">[:,</span> <span class="n">anchor_matching_gt</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">],</span> <span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)</span>
    685. 

  685.             <span class="n">matching_matrix</span><span class="p">[:,</span> <span class="n">anchor_matching_gt</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">]</span> <span class="o">*=</span> <span class="mi">0</span>
    686. 

  686.             <span class="n">matching_matrix</span><span class="p">[</span><span class="n">cost_argmin</span><span class="p">,</span> <span class="n">anchor_matching_gt</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">]</span> <span class="o">=</span> <span class="mi">1</span>
    687. 

  687. 

  688.         <span class="n">fg_mask_inboxes</span> <span class="o">=</span> <span class="n">matching_matrix</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span> <span class="o">&gt;</span> <span class="mi">0</span>
    689. 

  689.         <span class="n">num_fg</span> <span class="o">=</span> <span class="n">fg_mask_inboxes</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span><span class="o">.</span><span class="n">item</span><span class="p">()</span>
    690. 

  690. 

  691.         <span class="n">fg_mask</span><span class="p">[</span><span class="n">fg_mask</span><span class="o">.</span><span class="n">clone</span><span class="p">()]</span> <span class="o">=</span> <span class="n">fg_mask_inboxes</span>
    692. 

  692. 

  693.         <span class="n">matched_gt_inds</span> <span class="o">=</span> <span class="n">matching_matrix</span><span class="p">[:,</span> <span class="n">fg_mask_inboxes</span><span class="p">]</span><span class="o">.</span><span class="n">argmax</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span>
    694. 

  694.         <span class="n">gt_matched_classes</span> <span class="o">=</span> <span class="n">gt_classes</span><span class="p">[</span><span class="n">matched_gt_inds</span><span class="p">]</span>
    695. 

  695. 

  696.         <span class="n">pred_ious_this_matching</span> <span class="o">=</span> <span class="p">(</span><span class="n">matching_matrix</span> <span class="o">*</span> <span class="n">pair_wise_ious</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="mi">0</span><span class="p">)[</span><span class="n">fg_mask_inboxes</span><span class="p">]</span>
    697. 

  697.         <span class="k">return</span> <span class="n">num_fg</span><span class="p">,</span> <span class="n">gt_matched_classes</span><span class="p">,</span> <span class="n">pred_ious_this_matching</span><span class="p">,</span> <span class="n">matched_gt_inds</span></div></div>
    698. 

  698. 

  699. 

  700. <div class="viewcode-block" id="YoloXFastDetectionLoss"><a class="viewcode-back" href="../../../../super_gradients.training.html#super_gradients.training.losses.YoloXFastDetectionLoss">[docs]</a><span class="k">class</span> <span class="nc">YoloXFastDetectionLoss</span><span class="p">(</span><span class="n">YoloXDetectionLoss</span><span class="p">):</span>
    701. 

  701.     <span class="sd">&quot;&quot;&quot;</span>
    702. 

  702. <span class="sd">    A completely new implementation of YOLOX loss.</span>
    703. 

  703. <span class="sd">    This is NOT an equivalent implementation to the regular yolox loss.</span>
    704. 

  704. 

  705. <span class="sd">    * Completely avoids using loops compared to the nested loops in the original implementation.</span>
    706. 

  706. <span class="sd">        As a result runs much faster (speedup depends on the type of GPUs, their count, the batch size, etc.).</span>
    707. 

  707. <span class="sd">    * Tensors format is very different the original implementation.</span>
    708. 

  708. <span class="sd">        Tensors contain image ids, ground truth ids and anchor ids as values to support variable length data.</span>
    709. 

  709. <span class="sd">    * There are differences in terms of the algorithm itself:</span>
    710. 

  710. <span class="sd">    1. When computing a dynamic k for a ground truth,</span>
    711. 

  711. <span class="sd">        in the original implementation they consider the sum of top 10 predictions sorted by ious among the initial</span>
    712. 

  712. <span class="sd">        foregrounds of any ground truth in the image,</span>
    713. 

  713. <span class="sd">        while in our implementation we consider only the initial foreground of that particular ground truth.</span>
    714. 

  714. <span class="sd">        To compensate for that difference we introduce the dynamic_ks_bias hyperparamter which makes the dynamic ks larger.</span>
    715. 

  715. <span class="sd">    2. When computing the k matched detections for a ground truth,</span>
    716. 

  716. <span class="sd">        in the original implementation they consider the initial foregrounds of any ground truth in the image as candidates,</span>
    717. 

  717. <span class="sd">        while in our implementation we consider only the initial foreground of that particular ground truth as candidates.</span>
    718. 

  718. <span class="sd">        We believe that this difference is minor.</span>
    719. 

  719. 

  720. <span class="sd">    :param dynamic_ks_bias: hyperparameter to compensate for the discrepancies between the regular loss and this loss.</span>
    721. 

  721. <span class="sd">    :param sync_num_fgs:    sync num of fgs.</span>
    722. 

  722. <span class="sd">                            Can be used for DDP training.</span>
    723. 

  723. <span class="sd">    :param obj_loss_fix:    devide by total of num anchors instead num of matching fgs.</span>
    724. 

  724. <span class="sd">                            Can be used for objectness loss.</span>
    725. 

  725. <span class="sd">    &quot;&quot;&quot;</span>
    726. 

  726. 

  727.     <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span>
    728. 

  728.         <span class="bp">self</span><span class="p">,</span> <span class="n">strides</span><span class="p">,</span> <span class="n">num_classes</span><span class="p">,</span> <span class="n">use_l1</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span> <span class="n">center_sampling_radius</span><span class="o">=</span><span class="mf">2.5</span><span class="p">,</span> <span class="n">iou_type</span><span class="o">=</span><span class="s2">&quot;iou&quot;</span><span class="p">,</span> <span class="n">dynamic_ks_bias</span><span class="o">=</span><span class="mf">1.1</span><span class="p">,</span> <span class="n">sync_num_fgs</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span> <span class="n">obj_loss_fix</span><span class="o">=</span><span class="kc">False</span>
    729. 

  729.     <span class="p">):</span>
    730. 

  730.         <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">(</span><span class="n">strides</span><span class="o">=</span><span class="n">strides</span><span class="p">,</span> <span class="n">num_classes</span><span class="o">=</span><span class="n">num_classes</span><span class="p">,</span> <span class="n">use_l1</span><span class="o">=</span><span class="n">use_l1</span><span class="p">,</span> <span class="n">center_sampling_radius</span><span class="o">=</span><span class="n">center_sampling_radius</span><span class="p">,</span> <span class="n">iou_type</span><span class="o">=</span><span class="n">iou_type</span><span class="p">)</span>
    731. 

  731. 

  732.         <span class="bp">self</span><span class="o">.</span><span class="n">dynamic_ks_bias</span> <span class="o">=</span> <span class="n">dynamic_ks_bias</span>
    733. 

  733.         <span class="bp">self</span><span class="o">.</span><span class="n">sync_num_fgs</span> <span class="o">=</span> <span class="n">sync_num_fgs</span>
    734. 

  734.         <span class="bp">self</span><span class="o">.</span><span class="n">obj_loss_fix</span> <span class="o">=</span> <span class="n">obj_loss_fix</span>
    735. 

  735. 

  736.     <span class="k">def</span> <span class="nf">_compute_loss</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">predictions</span><span class="p">:</span> <span class="n">List</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">],</span> <span class="n">targets</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]:</span>
    737. 

  737.         <span class="sd">&quot;&quot;&quot;</span>
    738. 

  738. <span class="sd">        L = L_objectness + L_iou + L_classification + 1[no_aug_epoch]*L_l1</span>
    739. 

  739. <span class="sd">        where:</span>
    740. 

  740. <span class="sd">            * L_iou, L_classification and L_l1 are calculated only between cells and targets that suit them;</span>
    741. 

  741. <span class="sd">            * L_objectness is calculated for all cells.</span>
    742. 

  742. 

  743. <span class="sd">        L_classification:</span>
    744. 

  744. <span class="sd">            for cells that have suitable ground truths in their grid locations add BCEs</span>
    745. 

  745. <span class="sd">            to force a prediction of IoU with a GT in a multi-label way</span>
    746. 

  746. <span class="sd">            Coef: 1.</span>
    747. 

  747. <span class="sd">        L_iou:</span>
    748. 

  748. <span class="sd">            for cells that have suitable ground truths in their grid locations</span>
    749. 

  749. <span class="sd">            add (1 - IoU^2), IoU between a predicted box and each GT box, force maximum IoU</span>
    750. 

  750. <span class="sd">            Coef: 1.</span>
    751. 

  751. <span class="sd">        L_l1:</span>
    752. 

  752. <span class="sd">            for cells that have suitable ground truths in their grid locations</span>
    753. 

  753. <span class="sd">            l1 distance between the logits and GTs in “logits” format (the inverse of “logits to predictions” ops)</span>
    754. 

  754. <span class="sd">            Coef: 1[no_aug_epoch]</span>
    755. 

  755. <span class="sd">        L_objectness:</span>
    756. 

  756. <span class="sd">            for each cell add BCE with a label of 1 if there is GT assigned to the cell</span>
    757. 

  757. <span class="sd">            Coef: 5</span>
    758. 

  758. 

  759. <span class="sd">        :param predictions:     output from all Yolo levels, each of shape</span>
    760. 

  760. <span class="sd">                                [Batch x Num_Anchors x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    761. 

  761. <span class="sd">        :param targets:         [Num_targets x (4 + 2)], values on dim 1 are: image id in a batch, class, box x y w h</span>
    762. 

  762. 

  763. <span class="sd">        :return:                loss, all losses separately in a detached tensor</span>
    764. 

  764. <span class="sd">        &quot;&quot;&quot;</span>
    765. 

  765.         <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">transformed_outputs</span><span class="p">,</span> <span class="n">raw_outputs</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">prepare_predictions</span><span class="p">(</span><span class="n">predictions</span><span class="p">)</span>
    766. 

  766. 

  767.         <span class="n">bbox_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="p">:</span><span class="mi">4</span><span class="p">]</span>  <span class="c1"># [batch, n_anchors_all, 4]</span>
    768. 

  768.         <span class="n">obj_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">4</span><span class="p">:</span><span class="mi">5</span><span class="p">]</span>  <span class="c1"># [batch, n_anchors_all, 1]</span>
    769. 

  769.         <span class="n">cls_preds</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="p">[:,</span> <span class="p">:,</span> <span class="mi">5</span><span class="p">:]</span>  <span class="c1"># [batch, n_anchors_all, n_cls]</span>
    770. 

  770. 

  771.         <span class="c1"># assign cells to ground truths, at most one GT per cell</span>
    772. 

  772.         <span class="n">matched_fg_ids</span><span class="p">,</span> <span class="n">matched_gt_classes</span><span class="p">,</span> <span class="n">matched_gt_ids</span><span class="p">,</span> <span class="n">matched_img_ids</span><span class="p">,</span> <span class="n">matched_ious</span><span class="p">,</span> <span class="n">flattened_gts</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_matching</span><span class="p">(</span>
    773. 

  773.             <span class="n">bbox_preds</span><span class="p">,</span> <span class="n">cls_preds</span><span class="p">,</span> <span class="n">obj_preds</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">,</span> <span class="n">targets</span>
    774. 

  774.         <span class="p">)</span>
    775. 

  775. 

  776.         <span class="n">num_gts</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="n">flattened_gts</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
    777. 

  777.         <span class="n">num_fg</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="n">matched_gt_ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
    778. 

  778.         <span class="n">total_num_anchors</span> <span class="o">=</span> <span class="nb">max</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">*</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
    779. 

  779. 

  780.         <span class="n">cls_targets</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">one_hot</span><span class="p">(</span><span class="n">matched_gt_classes</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">),</span> <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span><span class="p">)</span> <span class="o">*</span> <span class="n">matched_ious</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    781. 

  781.         <span class="n">obj_targets</span> <span class="o">=</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]))</span>
    782. 

  782.         <span class="n">obj_targets</span><span class="p">[</span><span class="n">matched_img_ids</span><span class="p">,</span> <span class="n">matched_fg_ids</span><span class="p">]</span> <span class="o">=</span> <span class="mi">1</span>
    783. 

  783.         <span class="n">reg_targets</span> <span class="o">=</span> <span class="n">flattened_gts</span><span class="p">[</span><span class="n">matched_gt_ids</span><span class="p">][:,</span> <span class="mi">1</span><span class="p">:]</span>
    784. 

  784.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    785. 

  785.             <span class="n">l1_targets</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">get_l1_target</span><span class="p">(</span>
    786. 

  786.                 <span class="n">transformed_outputs</span><span class="o">.</span><span class="n">new_zeros</span><span class="p">((</span><span class="n">num_fg</span><span class="p">,</span> <span class="mi">4</span><span class="p">)),</span>
    787. 

  787.                 <span class="n">flattened_gts</span><span class="p">[</span><span class="n">matched_gt_ids</span><span class="p">][:,</span> <span class="mi">1</span><span class="p">:],</span>
    788. 

  788.                 <span class="n">expanded_strides</span><span class="o">.</span><span class="n">squeeze</span><span class="p">()[</span><span class="n">matched_fg_ids</span><span class="p">],</span>
    789. 

  789.                 <span class="n">x_shifts</span><span class="o">=</span><span class="n">x_shifts</span><span class="o">.</span><span class="n">squeeze</span><span class="p">()[</span><span class="n">matched_fg_ids</span><span class="p">],</span>
    790. 

  790.                 <span class="n">y_shifts</span><span class="o">=</span><span class="n">y_shifts</span><span class="o">.</span><span class="n">squeeze</span><span class="p">()[</span><span class="n">matched_fg_ids</span><span class="p">],</span>
    791. 

  791.             <span class="p">)</span>
    792. 

  792.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">sync_num_fgs</span> <span class="ow">and</span> <span class="n">dist</span><span class="o">.</span><span class="n">group</span><span class="o">.</span><span class="n">WORLD</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
    793. 

  793.             <span class="n">num_fg</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">scalar_tensor</span><span class="p">(</span><span class="n">num_fg</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">matched_gt_ids</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    794. 

  794.             <span class="n">dist</span><span class="o">.</span><span class="n">all_reduce</span><span class="p">(</span><span class="n">num_fg</span><span class="p">,</span> <span class="n">op</span><span class="o">=</span><span class="n">torch</span><span class="o">.</span><span class="n">_C</span><span class="o">.</span><span class="n">_distributed_c10d</span><span class="o">.</span><span class="n">ReduceOp</span><span class="o">.</span><span class="n">AVG</span><span class="p">)</span>
    795. 

  795. 

  796.         <span class="n">loss_iou</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">iou_loss</span><span class="p">(</span><span class="n">bbox_preds</span><span class="p">[</span><span class="n">matched_img_ids</span><span class="p">,</span> <span class="n">matched_fg_ids</span><span class="p">],</span> <span class="n">reg_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    797. 

  797.         <span class="n">loss_obj</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bcewithlog_loss</span><span class="p">(</span><span class="n">obj_preds</span><span class="o">.</span><span class="n">squeeze</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">),</span> <span class="n">obj_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="p">(</span><span class="n">total_num_anchors</span> <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">obj_loss_fix</span> <span class="k">else</span> <span class="n">num_fg</span><span class="p">)</span>
    798. 

  798.         <span class="n">loss_cls</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">bcewithlog_loss</span><span class="p">(</span><span class="n">cls_preds</span><span class="p">[</span><span class="n">matched_img_ids</span><span class="p">,</span> <span class="n">matched_fg_ids</span><span class="p">],</span> <span class="n">cls_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    799. 

  799.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    800. 

  800.             <span class="n">loss_l1</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">l1_loss</span><span class="p">(</span><span class="n">raw_outputs</span><span class="p">[</span><span class="n">matched_img_ids</span><span class="p">,</span> <span class="n">matched_fg_ids</span><span class="p">],</span> <span class="n">l1_targets</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">()</span> <span class="o">/</span> <span class="n">num_fg</span>
    801. 

  801.         <span class="k">else</span><span class="p">:</span>
    802. 

  802.             <span class="n">loss_l1</span> <span class="o">=</span> <span class="mf">0.0</span>
    803. 

  803. 

  804.         <span class="n">reg_weight</span> <span class="o">=</span> <span class="mf">5.0</span>
    805. 

  805.         <span class="n">loss</span> <span class="o">=</span> <span class="n">reg_weight</span> <span class="o">*</span> <span class="n">loss_iou</span> <span class="o">+</span> <span class="n">loss_obj</span> <span class="o">+</span> <span class="n">loss_cls</span> <span class="o">+</span> <span class="n">loss_l1</span>
    806. 

  806. 

  807.         <span class="k">return</span> <span class="p">(</span>
    808. 

  808.             <span class="n">loss</span><span class="p">,</span>
    809. 

  809.             <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">(</span>
    810. 

  810.                 <span class="p">(</span>
    811. 

  811.                     <span class="n">loss_iou</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    812. 

  812.                     <span class="n">loss_obj</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    813. 

  813.                     <span class="n">loss_cls</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    814. 

  814.                     <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="n">loss_l1</span><span class="p">)</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">device</span><span class="p">),</span>
    815. 

  815.                     <span class="n">torch</span><span class="o">.</span><span class="n">tensor</span><span class="p">(</span><span class="n">num_fg</span> <span class="o">/</span> <span class="nb">max</span><span class="p">(</span><span class="n">num_gts</span><span class="p">,</span> <span class="mi">1</span><span class="p">))</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">)</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">transformed_outputs</span><span class="o">.</span><span class="n">device</span><span class="p">),</span>
    816. 

  816.                     <span class="n">loss</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">0</span><span class="p">),</span>
    817. 

  817.                 <span class="p">)</span>
    818. 

  818.             <span class="p">)</span><span class="o">.</span><span class="n">detach</span><span class="p">(),</span>
    819. 

  819.         <span class="p">)</span>
    820. 

  820. 

  821.     <span class="k">def</span> <span class="nf">_get_initial_matching</span><span class="p">(</span>
    822. 

  822.         <span class="bp">self</span><span class="p">,</span> <span class="n">gt_bboxes</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">expanded_strides</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span>
    823. 

  823.     <span class="p">)</span> <span class="o">-&gt;</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]:</span>
    824. 

  824.         <span class="sd">&quot;&quot;&quot;</span>
    825. 

  825. <span class="sd">        Get candidates using a mask for all cells.</span>
    826. 

  826. <span class="sd">        Mask in only foreground cells that have a center located:</span>
    827. 

  827. <span class="sd">            * withing a GT box (param: is_in_boxes);</span>
    828. 

  828. <span class="sd">            OR</span>
    829. 

  829. <span class="sd">            * within a fixed radius around a GT box (center sampling) (param: is_in_centers);</span>
    830. 

  830. 

  831. <span class="sd">        return:</span>
    832. 

  832. <span class="sd">            initial_matching: get a list of candidates pairs of (gt box id, anchor box id) based on cell = is_in_boxes | is_in_centers.</span>
    833. 

  833. <span class="sd">                              shape: [num_candidates, 2]</span>
    834. 

  834. <span class="sd">            strong candidate mask: get a list whether a candidate is a strong one or not.</span>
    835. 

  835. <span class="sd">                                   strong candidate is a cell from is_in_boxes &amp; is_in_centers.</span>
    836. 

  836. <span class="sd">                                   shape: [num_candidates].</span>
    837. 

  837. <span class="sd">        &quot;&quot;&quot;</span>
    838. 

  838.         <span class="n">cell_x_centers</span> <span class="o">=</span> <span class="p">(</span><span class="n">x_shifts</span> <span class="o">+</span> <span class="mf">0.5</span><span class="p">)</span> <span class="o">*</span> <span class="n">expanded_strides</span>
    839. 

  839.         <span class="n">cell_y_centers</span> <span class="o">=</span> <span class="p">(</span><span class="n">y_shifts</span> <span class="o">+</span> <span class="mf">0.5</span><span class="p">)</span> <span class="o">*</span> <span class="n">expanded_strides</span>
    840. 

  840. 

  841.         <span class="n">gt_bboxes_x_centers</span> <span class="o">=</span> <span class="n">gt_bboxes</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
    842. 

  842.         <span class="n">gt_bboxes_y_centers</span> <span class="o">=</span> <span class="n">gt_bboxes</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">]</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
    843. 

  843. 

  844.         <span class="n">gt_bboxes_half_w</span> <span class="o">=</span> <span class="p">(</span><span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
    845. 

  845.         <span class="n">gt_bboxes_half_h</span> <span class="o">=</span> <span class="p">(</span><span class="mf">0.5</span> <span class="o">*</span> <span class="n">gt_bboxes</span><span class="p">[:,</span> <span class="mi">3</span><span class="p">])</span><span class="o">.</span><span class="n">unsqueeze</span><span class="p">(</span><span class="mi">1</span><span class="p">)</span>
    846. 

  846. 

  847.         <span class="n">is_in_boxes</span> <span class="o">=</span> <span class="p">(</span>
    848. 

  848.             <span class="p">(</span><span class="n">cell_x_centers</span> <span class="o">&gt;</span> <span class="n">gt_bboxes_x_centers</span> <span class="o">-</span> <span class="n">gt_bboxes_half_w</span><span class="p">)</span>
    849. 

  849.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">gt_bboxes_x_centers</span> <span class="o">+</span> <span class="n">gt_bboxes_half_w</span> <span class="o">&gt;</span> <span class="n">cell_x_centers</span><span class="p">)</span>
    850. 

  850.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">cell_y_centers</span> <span class="o">&gt;</span> <span class="n">gt_bboxes_y_centers</span> <span class="o">-</span> <span class="n">gt_bboxes_half_h</span><span class="p">)</span>
    851. 

  851.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">gt_bboxes_y_centers</span> <span class="o">+</span> <span class="n">gt_bboxes_half_h</span> <span class="o">&gt;</span> <span class="n">cell_y_centers</span><span class="p">)</span>
    852. 

  852.         <span class="p">)</span>
    853. 

  853. 

  854.         <span class="n">radius_shifts</span> <span class="o">=</span> <span class="mf">2.5</span> <span class="o">*</span> <span class="n">expanded_strides</span>
    855. 

  855. 

  856.         <span class="n">is_in_centers</span> <span class="o">=</span> <span class="p">(</span>
    857. 

  857.             <span class="p">(</span><span class="n">cell_x_centers</span> <span class="o">+</span> <span class="n">radius_shifts</span> <span class="o">&gt;</span> <span class="n">gt_bboxes_x_centers</span><span class="p">)</span>
    858. 

  858.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">gt_bboxes_x_centers</span> <span class="o">&gt;</span> <span class="n">cell_x_centers</span> <span class="o">-</span> <span class="n">radius_shifts</span><span class="p">)</span>
    859. 

  859.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">cell_y_centers</span> <span class="o">+</span> <span class="n">radius_shifts</span> <span class="o">&gt;</span> <span class="n">gt_bboxes_y_centers</span><span class="p">)</span>
    860. 

  860.             <span class="o">&amp;</span> <span class="p">(</span><span class="n">gt_bboxes_y_centers</span> <span class="o">&gt;</span> <span class="n">cell_y_centers</span> <span class="o">-</span> <span class="n">radius_shifts</span><span class="p">)</span>
    861. 

  861.         <span class="p">)</span>
    862. 

  862. 

  863.         <span class="n">initial_mask</span> <span class="o">=</span> <span class="n">is_in_boxes</span> <span class="o">|</span> <span class="n">is_in_centers</span>
    864. 

  864.         <span class="n">initial_matching</span> <span class="o">=</span> <span class="n">initial_mask</span><span class="o">.</span><span class="n">nonzero</span><span class="p">()</span>
    865. 

  865.         <span class="n">strong_candidate_mask</span> <span class="o">=</span> <span class="p">(</span><span class="n">is_in_boxes</span> <span class="o">&amp;</span> <span class="n">is_in_centers</span><span class="p">)[</span><span class="n">initial_mask</span><span class="p">]</span>
    866. 

  866. 

  867.         <span class="k">return</span> <span class="n">initial_matching</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">],</span> <span class="n">initial_matching</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">],</span> <span class="n">strong_candidate_mask</span>
    868. 

  868. 

  869.     <span class="nd">@torch</span><span class="o">.</span><span class="n">no_grad</span><span class="p">()</span>
    870. 

  870.     <span class="k">def</span> <span class="nf">_compute_matching</span><span class="p">(</span>
    871. 

  871.         <span class="bp">self</span><span class="p">,</span>
    872. 

  872.         <span class="n">bbox_preds</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    873. 

  873.         <span class="n">cls_preds</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    874. 

  874.         <span class="n">obj_preds</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    875. 

  875.         <span class="n">expanded_strides</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    876. 

  876.         <span class="n">x_shifts</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    877. 

  877.         <span class="n">y_shifts</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    878. 

  878.         <span class="n">labels</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    879. 

  879.         <span class="n">ious_loss_cost_coeff</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">3.0</span><span class="p">,</span>
    880. 

  880.         <span class="n">outside_boxes_and_center_cost_coeff</span><span class="p">:</span> <span class="nb">float</span> <span class="o">=</span> <span class="mf">100000.0</span><span class="p">,</span>
    881. 

  881.     <span class="p">)</span> <span class="o">-&gt;</span> <span class="n">Tuple</span><span class="p">[</span><span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">]:</span>
    882. 

  882.         <span class="sd">&quot;&quot;&quot;</span>
    883. 

  883. <span class="sd">        Match cells to ground truth:</span>
    884. 

  884. <span class="sd">            * at most 1 GT per cell</span>
    885. 

  885. <span class="sd">            * dynamic number of cells per GT</span>
    886. 

  886. 

  887. <span class="sd">        :param bbox_preds: predictions of bounding boxes. shape [batch, n_anchors_all, 4]</span>
    888. 

  888. <span class="sd">        :param cls_preds:  predictions of class.          shape [batch, n_anchors_all, n_cls]</span>
    889. 

  889. <span class="sd">        :param obj_preds:  predictions for objectness.    shape [batch, n_anchors_all, 1]</span>
    890. 

  890. <span class="sd">        :param expanded_strides:  stride of the output grid the prediction is coming from. shape [1, n_anchors_all]</span>
    891. 

  891. <span class="sd">        :param x_shifts: x coordinate on the grid cell the prediction is coming from.      shape [1, n_anchors_all]</span>
    892. 

  892. <span class="sd">        :param y_shifts: y coordinate on the grid cell the prediction is coming from.      shape [1, n_anchors_all]</span>
    893. 

  893. <span class="sd">        :param labels:   labels for each grid cell.  shape [n_anchors_all, (4 + 2)]</span>
    894. 

  894. <span class="sd">        :return: candidate_fg_ids       shape [num_fg]</span>
    895. 

  895. <span class="sd">                 candidate_gt_classes   shape [num_fg]</span>
    896. 

  896. <span class="sd">                 candidate_gt_ids       shape [num_fg]</span>
    897. 

  897. <span class="sd">                 candidate_img_ids      shape [num_fg]</span>
    898. 

  898. <span class="sd">                 candidate_ious         shape [num_fg]</span>
    899. 

  899. <span class="sd">                 flattened_gts          shape [num_gts, 5]</span>
    900. 

  900. <span class="sd">        &quot;&quot;&quot;</span>
    901. 

  901. 

  902.         <span class="n">flattened_gts</span><span class="p">,</span> <span class="n">gt_id_to_img_id</span> <span class="o">=</span> <span class="n">labels</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">:],</span> <span class="n">labels</span><span class="p">[:,</span> <span class="mi">0</span><span class="p">]</span><span class="o">.</span><span class="n">type</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">)</span>
    903. 

  903. 

  904.         <span class="c1"># COMPUTE CANDIDATES</span>
    905. 

  905.         <span class="n">candidate_gt_ids</span><span class="p">,</span> <span class="n">candidate_fg_ids</span><span class="p">,</span> <span class="n">strong_candidate_mask</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_get_initial_matching</span><span class="p">(</span><span class="n">flattened_gts</span><span class="p">[:,</span> <span class="mi">1</span><span class="p">:],</span> <span class="n">expanded_strides</span><span class="p">,</span> <span class="n">x_shifts</span><span class="p">,</span> <span class="n">y_shifts</span><span class="p">)</span>
    906. 

  906.         <span class="n">candidate_img_ids</span> <span class="o">=</span> <span class="n">gt_id_to_img_id</span><span class="p">[</span><span class="n">candidate_gt_ids</span><span class="p">]</span>
    907. 

  907.         <span class="n">candidate_gts_bbox</span> <span class="o">=</span> <span class="n">flattened_gts</span><span class="p">[</span><span class="n">candidate_gt_ids</span><span class="p">,</span> <span class="mi">1</span><span class="p">:]</span>
    908. 

  908.         <span class="n">candidate_det_bbox</span> <span class="o">=</span> <span class="n">bbox_preds</span><span class="p">[</span><span class="n">candidate_img_ids</span><span class="p">,</span> <span class="n">candidate_fg_ids</span><span class="p">]</span>
    909. 

  909. 

  910.         <span class="c1"># COMPUTE DYNAMIC KS</span>
    911. 

  911.         <span class="n">candidate_ious</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_calculate_pairwise_bbox_iou</span><span class="p">(</span><span class="n">candidate_gts_bbox</span><span class="p">,</span> <span class="n">candidate_det_bbox</span><span class="p">,</span> <span class="n">xyxy</span><span class="o">=</span><span class="kc">False</span><span class="p">)</span>
    912. 

  912.         <span class="n">dynamic_ks</span><span class="p">,</span> <span class="n">matching_index_to_dynamic_k_index</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_dynamic_ks</span><span class="p">(</span><span class="n">candidate_gt_ids</span><span class="p">,</span> <span class="n">candidate_ious</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">dynamic_ks_bias</span><span class="p">)</span>
    913. 

  913.         <span class="k">del</span> <span class="n">candidate_gts_bbox</span><span class="p">,</span> <span class="n">candidate_det_bbox</span>
    914. 

  914. 

  915.         <span class="c1"># ORDER CANDIDATES BY COST</span>
    916. 

  916.         <span class="n">candidate_gt_classes</span> <span class="o">=</span> <span class="n">flattened_gts</span><span class="p">[</span><span class="n">candidate_gt_ids</span><span class="p">,</span> <span class="mi">0</span><span class="p">]</span>
    917. 

  917.         <span class="n">cost_order</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_cost_order</span><span class="p">(</span>
    918. 

  918.             <span class="bp">self</span><span class="o">.</span><span class="n">num_classes</span><span class="p">,</span>
    919. 

  919.             <span class="n">candidate_img_ids</span><span class="p">,</span>
    920. 

  920.             <span class="n">candidate_gt_classes</span><span class="p">,</span>
    921. 

  921.             <span class="n">candidate_fg_ids</span><span class="p">,</span>
    922. 

  922.             <span class="n">candidate_ious</span><span class="p">,</span>
    923. 

  923.             <span class="n">cls_preds</span><span class="p">,</span>
    924. 

  924.             <span class="n">obj_preds</span><span class="p">,</span>
    925. 

  925.             <span class="n">strong_candidate_mask</span><span class="p">,</span>
    926. 

  926.             <span class="n">ious_loss_cost_coeff</span><span class="p">,</span>
    927. 

  927.             <span class="n">outside_boxes_and_center_cost_coeff</span><span class="p">,</span>
    928. 

  928.         <span class="p">)</span>
    929. 

  929. 

  930.         <span class="n">candidate_gt_ids</span> <span class="o">=</span> <span class="n">candidate_gt_ids</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    931. 

  931.         <span class="n">candidate_gt_classes</span> <span class="o">=</span> <span class="n">candidate_gt_classes</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    932. 

  932.         <span class="n">candidate_img_ids</span> <span class="o">=</span> <span class="n">candidate_img_ids</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    933. 

  933.         <span class="n">candidate_fg_ids</span> <span class="o">=</span> <span class="n">candidate_fg_ids</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    934. 

  934.         <span class="n">candidate_ious</span> <span class="o">=</span> <span class="n">candidate_ious</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    935. 

  935.         <span class="n">matching_index_to_dynamic_k_index</span> <span class="o">=</span> <span class="n">matching_index_to_dynamic_k_index</span><span class="p">[</span><span class="n">cost_order</span><span class="p">]</span>
    936. 

  936.         <span class="k">del</span> <span class="n">cost_order</span>
    937. 

  937. 

  938.         <span class="c1"># FILTER MATCHING TO LOWEST K COST MATCHES PER GT</span>
    939. 

  939.         <span class="n">ranks</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_ranks</span><span class="p">(</span><span class="n">candidate_gt_ids</span><span class="p">)</span>
    940. 

  940.         <span class="n">corresponding_dynamic_ks</span> <span class="o">=</span> <span class="n">dynamic_ks</span><span class="p">[</span><span class="n">matching_index_to_dynamic_k_index</span><span class="p">]</span>
    941. 

  941.         <span class="n">topk_mask</span> <span class="o">=</span> <span class="n">ranks</span> <span class="o">&lt;</span> <span class="n">corresponding_dynamic_ks</span>
    942. 

  942. 

  943.         <span class="n">candidate_gt_ids</span> <span class="o">=</span> <span class="n">candidate_gt_ids</span><span class="p">[</span><span class="n">topk_mask</span><span class="p">]</span>
    944. 

  944.         <span class="n">candidate_gt_classes</span> <span class="o">=</span> <span class="n">candidate_gt_classes</span><span class="p">[</span><span class="n">topk_mask</span><span class="p">]</span>
    945. 

  945.         <span class="n">candidate_img_ids</span> <span class="o">=</span> <span class="n">candidate_img_ids</span><span class="p">[</span><span class="n">topk_mask</span><span class="p">]</span>
    946. 

  946.         <span class="n">candidate_fg_ids</span> <span class="o">=</span> <span class="n">candidate_fg_ids</span><span class="p">[</span><span class="n">topk_mask</span><span class="p">]</span>
    947. 

  947.         <span class="n">candidate_ious</span> <span class="o">=</span> <span class="n">candidate_ious</span><span class="p">[</span><span class="n">topk_mask</span><span class="p">]</span>
    948. 

  948.         <span class="k">del</span> <span class="n">ranks</span><span class="p">,</span> <span class="n">topk_mask</span><span class="p">,</span> <span class="n">dynamic_ks</span><span class="p">,</span> <span class="n">matching_index_to_dynamic_k_index</span><span class="p">,</span> <span class="n">corresponding_dynamic_ks</span>
    949. 

  949. 

  950.         <span class="c1"># FILTER MATCHING TO AT MOST 1 MATCH FOR DET BY TAKING THE LOWEST COST MATCH</span>
    951. 

  951.         <span class="n">candidate_img_and_fg_ids_combined</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_combine_candidates_img_id_fg_id</span><span class="p">(</span><span class="n">candidate_img_ids</span><span class="p">,</span> <span class="n">candidate_fg_ids</span><span class="p">)</span>
    952. 

  952.         <span class="n">top1_mask</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">_compute_is_first_mask</span><span class="p">(</span><span class="n">candidate_img_and_fg_ids_combined</span><span class="p">)</span>
    953. 

  953.         <span class="n">candidate_gt_ids</span> <span class="o">=</span> <span class="n">candidate_gt_ids</span><span class="p">[</span><span class="n">top1_mask</span><span class="p">]</span>
    954. 

  954.         <span class="n">candidate_gt_classes</span> <span class="o">=</span> <span class="n">candidate_gt_classes</span><span class="p">[</span><span class="n">top1_mask</span><span class="p">]</span>
    955. 

  955.         <span class="n">candidate_fg_ids</span> <span class="o">=</span> <span class="n">candidate_fg_ids</span><span class="p">[</span><span class="n">top1_mask</span><span class="p">]</span>
    956. 

  956.         <span class="n">candidate_img_ids</span> <span class="o">=</span> <span class="n">candidate_img_ids</span><span class="p">[</span><span class="n">top1_mask</span><span class="p">]</span>
    957. 

  957.         <span class="n">candidate_ious</span> <span class="o">=</span> <span class="n">candidate_ious</span><span class="p">[</span><span class="n">top1_mask</span><span class="p">]</span>
    958. 

  958. 

  959.         <span class="k">return</span> <span class="n">candidate_fg_ids</span><span class="p">,</span> <span class="n">candidate_gt_classes</span><span class="p">,</span> <span class="n">candidate_gt_ids</span><span class="p">,</span> <span class="n">candidate_img_ids</span><span class="p">,</span> <span class="n">candidate_ious</span><span class="p">,</span> <span class="n">flattened_gts</span>
    960. 

  960. 

  961.     <span class="k">def</span> <span class="nf">_combine_candidates_img_id_fg_id</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">candidate_img_ids</span><span class="p">,</span> <span class="n">candidate_anchor_ids</span><span class="p">):</span>
    962. 

  962.         <span class="sd">&quot;&quot;&quot;</span>
    963. 

  963. <span class="sd">        Create one dim tensor with unique pairs of img_id and fg_id.</span>
    964. 

  964. <span class="sd">        e.g: candidate_img_ids = [0,1,0,0]</span>
    965. 

  965. <span class="sd">             candidate_fg_ids = [0,0,0,1]</span>
    966. 

  966. <span class="sd">             result = [0,1,0,2]</span>
    967. 

  967. <span class="sd">        &quot;&quot;&quot;</span>
    968. 

  968.         <span class="n">candidate_img_and_fg_ids_combined</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">candidate_img_ids</span><span class="p">,</span> <span class="n">candidate_anchor_ids</span><span class="p">),</span> <span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span><span class="o">.</span><span class="n">unique</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">return_inverse</span><span class="o">=</span><span class="kc">True</span><span class="p">)[</span><span class="mi">1</span><span class="p">]</span>
    969. 

  969.         <span class="k">return</span> <span class="n">candidate_img_and_fg_ids_combined</span>
    970. 

  970. 

  971.     <span class="k">def</span> <span class="nf">_compute_dynamic_ks</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">ids</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">ious</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">dynamic_ks_bias</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
    972. 

  972.         <span class="sd">&quot;&quot;&quot;</span>
    973. 

  973. <span class="sd">        :param ids:                 ids of GTs, shape: [num_candidates]</span>
    974. 

  974. <span class="sd">        :param ious:                pairwise IoUs, shape: [num_candidates]</span>
    975. 

  975. <span class="sd">        :param dynamic_ks_bias:     multiply the resulted k to compensate the regular loss</span>
    976. 

  976. <span class="sd">        &quot;&quot;&quot;</span>
    977. 

  977.         <span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">1</span><span class="p">,</span> <span class="s2">&quot;ids must be of shape [num_candidates]&quot;</span>
    978. 

  978.         <span class="k">assert</span> <span class="nb">len</span><span class="p">(</span><span class="n">ious</span><span class="o">.</span><span class="n">shape</span><span class="p">)</span> <span class="o">==</span> <span class="mi">1</span><span class="p">,</span> <span class="s2">&quot;ious must be of shape [num_candidates]&quot;</span>
    979. 

  979.         <span class="k">assert</span> <span class="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">==</span> <span class="n">ious</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="s2">&quot;num of ids.shape[0] must be the same as num of ious.shape[0]&quot;</span>
    980. 

  980.         <span class="c1"># sort ious and ids by ious</span>
    981. 

  981.         <span class="n">ious</span><span class="p">,</span> <span class="n">ious_argsort</span> <span class="o">=</span> <span class="n">ious</span><span class="o">.</span><span class="n">sort</span><span class="p">(</span><span class="n">descending</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    982. 

  982.         <span class="n">ids</span> <span class="o">=</span> <span class="n">ids</span><span class="p">[</span><span class="n">ious_argsort</span><span class="p">]</span>
    983. 

  983. 

  984.         <span class="c1"># stable sort indices, so that ious are first sorted by id and second by value</span>
    985. 

  985.         <span class="n">ids</span><span class="p">,</span> <span class="n">ids_argsort</span> <span class="o">=</span> <span class="n">ids</span><span class="o">.</span><span class="n">sort</span><span class="p">(</span><span class="n">stable</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    986. 

  986.         <span class="n">ious</span> <span class="o">=</span> <span class="n">ious</span><span class="p">[</span><span class="n">ids_argsort</span><span class="p">]</span>
    987. 

  987. 

  988.         <span class="n">unique_ids</span><span class="p">,</span> <span class="n">ids_index_to_unique_ids_index</span> <span class="o">=</span> <span class="n">ids</span><span class="o">.</span><span class="n">unique_consecutive</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">,</span> <span class="n">return_inverse</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    989. 

  989.         <span class="n">num_unique_ids</span> <span class="o">=</span> <span class="n">unique_ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
    990. 

  990. 

  991.         <span class="k">if</span> <span class="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">&gt;</span> <span class="mi">10</span><span class="p">:</span>
    992. 

  992.             <span class="n">is_in_top_10</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">((</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">((</span><span class="mi">10</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">bool</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids</span><span class="o">.</span><span class="n">device</span><span class="p">),</span> <span class="n">ids</span><span class="p">[</span><span class="mi">10</span><span class="p">:]</span> <span class="o">!=</span> <span class="n">ids</span><span class="p">[:</span><span class="o">-</span><span class="mi">10</span><span class="p">]))</span>
    993. 

  993.         <span class="k">else</span><span class="p">:</span>
    994. 

  994.             <span class="n">is_in_top_10</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones_like</span><span class="p">(</span><span class="n">ids</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">bool</span><span class="p">)</span>
    995. 

  995. 

  996.         <span class="n">dynamic_ks</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">((</span><span class="n">num_unique_ids</span><span class="p">,),</span> <span class="n">dtype</span><span class="o">=</span><span class="n">ious</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ious</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    997. 

  997.         <span class="n">dynamic_ks</span><span class="o">.</span><span class="n">index_put_</span><span class="p">((</span><span class="n">ids_index_to_unique_ids_index</span><span class="p">,),</span> <span class="n">is_in_top_10</span> <span class="o">*</span> <span class="n">ious</span><span class="p">,</span> <span class="n">accumulate</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    998. 

  998.         <span class="k">if</span> <span class="n">dynamic_ks_bias</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
    999. 

  999.             <span class="n">dynamic_ks</span> <span class="o">*=</span> <span class="n">dynamic_ks_bias</span>
    1000. 

  1000.         <span class="n">dynamic_ks</span> <span class="o">=</span> <span class="n">dynamic_ks</span><span class="o">.</span><span class="n">long</span><span class="p">()</span><span class="o">.</span><span class="n">clamp</span><span class="p">(</span><span class="nb">min</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    1001. 

  1001. 

  1002.         <span class="n">all_argsort</span> <span class="o">=</span> <span class="n">ious_argsort</span><span class="p">[</span><span class="n">ids_argsort</span><span class="p">]</span>
    1003. 

  1003.         <span class="n">inverse_all_argsort</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">ious_argsort</span><span class="p">)</span>
    1004. 

  1004.         <span class="n">inverse_all_argsort</span><span class="p">[</span><span class="n">all_argsort</span><span class="p">]</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="n">all_argsort</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">all_argsort</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">all_argsort</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    1005. 

  1005. 

  1006.         <span class="k">return</span> <span class="n">dynamic_ks</span><span class="p">,</span> <span class="n">ids_index_to_unique_ids_index</span><span class="p">[</span><span class="n">inverse_all_argsort</span><span class="p">]</span>
    1007. 

  1007. 

  1008.     <span class="k">def</span> <span class="nf">_compute_cost_order</span><span class="p">(</span>
    1009. 

  1009.         <span class="bp">self</span><span class="p">,</span>
    1010. 

  1010.         <span class="n">num_classes</span><span class="p">,</span>
    1011. 

  1011.         <span class="n">candidate_gt_img_ids</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1012. 

  1012.         <span class="n">candidate_gt_classes</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1013. 

  1013.         <span class="n">candidate_anchor_ids</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1014. 

  1014.         <span class="n">candidate_ious</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1015. 

  1015.         <span class="n">cls_preds</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1016. 

  1016.         <span class="n">obj_preds</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1017. 

  1017.         <span class="n">strong_candidate_mask</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span>
    1018. 

  1018.         <span class="n">ious_loss_cost_coeff</span><span class="p">:</span> <span class="nb">float</span><span class="p">,</span>
    1019. 

  1019.         <span class="n">outside_boxes_and_center_cost_coeff</span><span class="p">:</span> <span class="nb">float</span><span class="p">,</span>
    1020. 

  1020.     <span class="p">)</span> <span class="o">-&gt;</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
    1021. 

  1021.         <span class="n">gt_cls_per_image</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">one_hot</span><span class="p">(</span><span class="n">candidate_gt_classes</span><span class="o">.</span><span class="n">to</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">int64</span><span class="p">),</span> <span class="n">num_classes</span><span class="p">)</span><span class="o">.</span><span class="n">float</span><span class="p">()</span>
    1022. 

  1022.         <span class="k">with</span> <span class="n">torch</span><span class="o">.</span><span class="n">cuda</span><span class="o">.</span><span class="n">amp</span><span class="o">.</span><span class="n">autocast</span><span class="p">(</span><span class="n">enabled</span><span class="o">=</span><span class="kc">False</span><span class="p">):</span>
    1023. 

  1023.             <span class="n">cls_preds_</span> <span class="o">=</span> <span class="p">(</span>
    1024. 

  1024.                 <span class="n">cls_preds</span><span class="p">[</span><span class="n">candidate_gt_img_ids</span><span class="p">,</span> <span class="n">candidate_anchor_ids</span><span class="p">]</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">sigmoid_</span><span class="p">()</span>
    1025. 

  1025.                 <span class="o">*</span> <span class="n">obj_preds</span><span class="p">[</span><span class="n">candidate_gt_img_ids</span><span class="p">,</span> <span class="n">candidate_anchor_ids</span><span class="p">]</span><span class="o">.</span><span class="n">float</span><span class="p">()</span><span class="o">.</span><span class="n">sigmoid_</span><span class="p">()</span>
    1026. 

  1026.             <span class="p">)</span>
    1027. 

  1027.             <span class="n">pair_wise_cls_cost</span> <span class="o">=</span> <span class="n">F</span><span class="o">.</span><span class="n">binary_cross_entropy</span><span class="p">(</span><span class="n">cls_preds_</span><span class="o">.</span><span class="n">sqrt_</span><span class="p">(),</span> <span class="n">gt_cls_per_image</span><span class="p">,</span> <span class="n">reduction</span><span class="o">=</span><span class="s2">&quot;none&quot;</span><span class="p">)</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="o">-</span><span class="mi">1</span><span class="p">)</span>
    1028. 

  1028. 

  1029.         <span class="n">ious_cost</span> <span class="o">=</span> <span class="o">-</span><span class="n">torch</span><span class="o">.</span><span class="n">log</span><span class="p">(</span><span class="n">candidate_ious</span> <span class="o">+</span> <span class="mf">1e-8</span><span class="p">)</span>
    1030. 

  1030.         <span class="n">cost</span> <span class="o">=</span> <span class="n">pair_wise_cls_cost</span> <span class="o">+</span> <span class="n">ious_loss_cost_coeff</span> <span class="o">*</span> <span class="n">ious_cost</span> <span class="o">+</span> <span class="n">outside_boxes_and_center_cost_coeff</span> <span class="o">*</span> <span class="n">strong_candidate_mask</span><span class="o">.</span><span class="n">logical_not</span><span class="p">()</span>
    1031. 

  1031.         <span class="k">return</span> <span class="n">cost</span><span class="o">.</span><span class="n">argsort</span><span class="p">()</span>
    1032. 

  1032. 

  1033.     <span class="k">def</span> <span class="nf">_calculate_pairwise_bbox_iou</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">bboxes_a</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">bboxes_b</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">xyxy</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
    1034. 

  1034.         <span class="k">if</span> <span class="n">bboxes_a</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">!=</span> <span class="mi">4</span> <span class="ow">or</span> <span class="n">bboxes_b</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">!=</span> <span class="mi">4</span><span class="p">:</span>
    1035. 

  1035.             <span class="k">raise</span> <span class="ne">IndexError</span>
    1036. 

  1036. 

  1037.         <span class="k">if</span> <span class="n">xyxy</span><span class="p">:</span>
    1038. 

  1038.             <span class="n">tl</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">],</span> <span class="n">bboxes_b</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">])</span>
    1039. 

  1039.             <span class="n">br</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">min</span><span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:],</span> <span class="n">bboxes_b</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:])</span>
    1040. 

  1040.             <span class="n">area_a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">-</span> <span class="n">bboxes_a</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
    1041. 

  1041.             <span class="n">area_b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">bboxes_b</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">-</span> <span class="n">bboxes_b</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">],</span> <span class="mi">1</span><span class="p">)</span>
    1042. 

  1042.         <span class="k">else</span><span class="p">:</span>
    1043. 

  1043.             <span class="n">tl</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">max</span><span class="p">(</span>
    1044. 

  1044.                 <span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">bboxes_a</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span>
    1045. 

  1045.                 <span class="p">(</span><span class="n">bboxes_b</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">-</span> <span class="n">bboxes_b</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span>
    1046. 

  1046.             <span class="p">)</span>
    1047. 

  1047.             <span class="n">br</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">min</span><span class="p">(</span>
    1048. 

  1048.                 <span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">bboxes_a</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span>
    1049. 

  1049.                 <span class="p">(</span><span class="n">bboxes_b</span><span class="p">[:,</span> <span class="p">:</span><span class="mi">2</span><span class="p">]</span> <span class="o">+</span> <span class="n">bboxes_b</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:]</span> <span class="o">/</span> <span class="mi">2</span><span class="p">),</span>
    1050. 

  1050.             <span class="p">)</span>
    1051. 

  1051. 

  1052.             <span class="n">area_a</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">bboxes_a</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:],</span> <span class="mi">1</span><span class="p">)</span>
    1053. 

  1053.             <span class="n">area_b</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">bboxes_b</span><span class="p">[:,</span> <span class="mi">2</span><span class="p">:],</span> <span class="mi">1</span><span class="p">)</span>
    1054. 

  1054.         <span class="n">en</span> <span class="o">=</span> <span class="p">(</span><span class="n">tl</span> <span class="o">&lt;</span> <span class="n">br</span><span class="p">)</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
    1055. 

  1055.         <span class="n">area_i</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">prod</span><span class="p">(</span><span class="n">br</span> <span class="o">-</span> <span class="n">tl</span><span class="p">,</span> <span class="mi">1</span><span class="p">)</span> <span class="o">*</span> <span class="n">en</span>
    1056. 

  1056.         <span class="k">return</span> <span class="n">area_i</span> <span class="o">/</span> <span class="p">(</span><span class="n">area_a</span> <span class="o">+</span> <span class="n">area_b</span> <span class="o">-</span> <span class="n">area_i</span><span class="p">)</span>
    1057. 

  1057. 

  1058.     <span class="k">def</span> <span class="nf">_compute_ranks</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">ids</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
    1059. 

  1059.         <span class="n">ids</span><span class="p">,</span> <span class="n">ids_argsort</span> <span class="o">=</span> <span class="n">ids</span><span class="o">.</span><span class="n">sort</span><span class="p">(</span><span class="n">stable</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    1060. 

  1060. 

  1061.         <span class="k">if</span> <span class="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
    1062. 

  1062.             <span class="n">is_not_first</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">((</span><span class="n">torch</span><span class="o">.</span><span class="n">zeros</span><span class="p">((</span><span class="mi">1</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">bool</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids</span><span class="o">.</span><span class="n">device</span><span class="p">),</span> <span class="n">ids</span><span class="p">[</span><span class="mi">1</span><span class="p">:]</span> <span class="o">==</span> <span class="n">ids</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">]))</span>
    1063. 

  1063.         <span class="k">else</span><span class="p">:</span>
    1064. 

  1064.             <span class="n">is_not_first</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">ids</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">bool</span><span class="p">)</span>
    1065. 

  1065. 

  1066.         <span class="n">subtract</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="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    1067. 

  1067.         <span class="n">subtract</span><span class="p">[</span><span class="n">is_not_first</span><span class="p">]</span> <span class="o">=</span> <span class="mi">0</span>
    1068. 

  1068.         <span class="n">subtract</span> <span class="o">=</span> <span class="n">subtract</span><span class="o">.</span><span class="n">cummax</span><span class="p">(</span><span class="n">dim</span><span class="o">=</span><span class="mi">0</span><span class="p">)[</span><span class="mi">0</span><span class="p">]</span>
    1069. 

  1069.         <span class="n">rank</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="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids</span><span class="o">.</span><span class="n">device</span><span class="p">)</span> <span class="o">-</span> <span class="n">subtract</span>
    1070. 

  1070. 

  1071.         <span class="n">inverse_argsort</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">ids_argsort</span><span class="p">)</span>
    1072. 

  1072.         <span class="n">inverse_argsort</span><span class="p">[</span><span class="n">ids_argsort</span><span class="p">]</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="n">ids_argsort</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    1073. 

  1073. 

  1074.         <span class="k">return</span> <span class="n">rank</span><span class="p">[</span><span class="n">inverse_argsort</span><span class="p">]</span>
    1075. 

  1075. 

  1076.     <span class="k">def</span> <span class="nf">_compute_is_first_mask</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">ids</span><span class="p">:</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-&gt;</span> <span class="n">torch</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
    1077. 

  1077.         <span class="sd">&quot;&quot;&quot;</span>
    1078. 

  1078. <span class="sd">        Filter fg that matches two gts.</span>
    1079. 

  1079. <span class="sd">        &quot;&quot;&quot;</span>
    1080. 

  1080.         <span class="n">ids</span><span class="p">,</span> <span class="n">ids_argsort</span> <span class="o">=</span> <span class="n">ids</span><span class="o">.</span><span class="n">sort</span><span class="p">(</span><span class="n">stable</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
    1081. 

  1081. 

  1082.         <span class="k">if</span> <span class="n">ids</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span> <span class="o">&gt;</span> <span class="mi">1</span><span class="p">:</span>
    1083. 

  1083.             <span class="n">is_first</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">((</span><span class="n">torch</span><span class="o">.</span><span class="n">ones</span><span class="p">((</span><span class="mi">1</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">bool</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids</span><span class="o">.</span><span class="n">device</span><span class="p">),</span> <span class="n">ids</span><span class="p">[</span><span class="mi">1</span><span class="p">:]</span> <span class="o">!=</span> <span class="n">ids</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">]))</span>
    1084. 

  1084.         <span class="k">else</span><span class="p">:</span>
    1085. 

  1085.             <span class="n">is_first</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">ones_like</span><span class="p">(</span><span class="n">ids</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">bool</span><span class="p">)</span>
    1086. 

  1086. 

  1087.         <span class="n">inverse_argsort</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">zeros_like</span><span class="p">(</span><span class="n">ids_argsort</span><span class="p">)</span>
    1088. 

  1088.         <span class="n">inverse_argsort</span><span class="p">[</span><span class="n">ids_argsort</span><span class="p">]</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="n">ids_argsort</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">dtype</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">dtype</span><span class="p">,</span> <span class="n">device</span><span class="o">=</span><span class="n">ids_argsort</span><span class="o">.</span><span class="n">device</span><span class="p">)</span>
    1089. 

  1089. 

  1090.         <span class="k">return</span> <span class="n">is_first</span><span class="p">[</span><span class="n">inverse_argsort</span><span class="p">]</span></div>
    1091. 

  1091. </pre></div>
    1092. 

  1092. 

  1093.            </div>
    1094. 

  1094.           </div>
    1095. 

  1095.           <footer>
    1096. 

  1096. 

  1097.   <hr/>
    1098. 

  1098. 

  1099.   <div role="contentinfo">
    1100. 

  1100.     <p>&#169; Copyright 2021, SuperGradients team.</p>
    1101. 

  1101.   </div>
    1102. 

  1102. 

  1103.   Built with <a href="https://www.sphinx-doc.org/">Sphinx</a> using a
    1104. 

  1104.     <a href="https://github.com/readthedocs/sphinx_rtd_theme">theme</a>
    1105. 

  1105.     provided by <a href="https://readthedocs.org">Read the Docs</a>.
    1106. 

  1106.    
    1107. 

  1107. 

  1108. </footer>
    1109. 

  1109.         </div>
    1110. 

  1110.       </div>
    1111. 

  1111.     </section>
    1112. 

  1112.   </div>
    1113. 

  1113.   <script>
    1114. 

  1114.       jQuery(function () {
    1115. 

  1115.           SphinxRtdTheme.Navigation.enable(true);
    1116. 

  1116.       });
    1117. 

  1117.   </script> 
    1118. 

  1118. 

  1119. </body>
    1120. 

  1120. </html>
    1121.
Prev
Next
Tip!

Press p or to see the previous file or, n or to see the next file

Comments

Write Preview
Loading...
Integrate AWS S3
Use S3 remote
Select bucket
Access key
Finish

Use AWS S3 as storage!

Browsing data directories saved to S3 is possible with DAGsHub. Let's configure your repository to easily display your data in the context of any commit!

Specify your S3 bucket

Access key (If needed)

Congratulations!

super-gradients is now integrated with AWS S3!

Delete Storage Key

Are you sure you want to delete this access key?

No
Yes
Integrate Google Cloud Storage
Use Google Storage
Select bucket
Upload key
Finish

Use Google Cloud Storage!

Browsing data directories saved to Google Cloud Storage is possible with DAGsHub. Let's configure your repository to easily display your data in the context of any commit!

Specify your Google Storage bucket

Service Account Key

Congratulations!

super-gradients is now integrated with Google Cloud Storage!

Delete Storage Key

Are you sure you want to delete this access key?

No
Yes
Integrate Azure Cloud Storage
Use Azure Storage
Select bucket
Set key
Finish

Use Azure Cloud Storage!

Browsing data directories saved to Azure Cloud Storage is possible with DAGsHub. Let's configure your repository to easily display your data in the context of any commit!

Specify your Azure Storage bucket

Access key (If needed)

Congratulations!

super-gradients is now integrated with Azure Cloud Storage!

Delete Storage Key

Are you sure you want to delete this access key?

No
Yes
Integrate S3 compatible storage
Use S3 like remote
Select bucket
Access key
Finish

Use any S3 compatible storage!

Browsing data directories saved to S3 compatible storage is possible with DAGsHub. Let's configure your repository to easily display your data in the context of any commit!

Specify your S3 bucket

Bucket name cannot be the same as the repository name. Please change one of them.

Access key (If needed)

Congratulations!

super-gradients is now integrated with your S3 compatible storage!

Delete Storage Key

Are you sure you want to delete this access key?

No
Yes