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 109 KB
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 
  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 1.0 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.   <!--[if lt IE 9]>
    11. 

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

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

  13.   
    14. 

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

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

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

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

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

  19.     <link rel="index" title="Index" href="../../../../genindex.html" />
    20. 

  20.     <link rel="search" title="Search" href="../../../../search.html" /> 
    21. 

  21. </head>
    22. 

  22. 

  23. <body class="wy-body-for-nav"> 
    24. 

  24.   <div class="wy-grid-for-nav">
    25. 

  25.     <nav data-toggle="wy-nav-shift" class="wy-nav-side">
    26. 

  26.       <div class="wy-side-scroll">
    27. 

  27.         <div class="wy-side-nav-search" >
    28. 

  28.             <a href="../../../../index.html" class="icon icon-home"> SuperGradients
    29. 

  29.           </a>
    30. 

  30. <div role="search">
    31. 

  31.   <form id="rtd-search-form" class="wy-form" action="../../../../search.html" method="get">
    32. 

  32.     <input type="text" name="q" placeholder="Search docs" />
    33. 

  33.     <input type="hidden" name="check_keywords" value="yes" />
    34. 

  34.     <input type="hidden" name="area" value="default" />
    35. 

  35.   </form>
    36. 

  36. </div>
    37. 

  37.         </div><div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="Navigation menu">
    38. 

  38.               <p class="caption"><span class="caption-text">Welcome To SuperGradients</span></p>
    39. 

  39. <ul>
    40. 

  40. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html">Fill our 4-question quick survey! We will raffle free SuperGradients swag between those who will participate -&gt; Fill Survey</a></li>
    41. 

  41. <li class="toctree-l1"><a class="reference internal" href="../../../../welcome.html#supergradients">SuperGradients</a></li>
    42. 

  42. </ul>
    43. 

  43. <p class="caption"><span class="caption-text">Technical Documentation</span></p>
    44. 

  44. <ul>
    45. 

  45. <li class="toctree-l1"><a class="reference internal" href="../../../../super_gradients.common.html">Common package</a></li>
    46. 

  46. <li class="toctree-l1"><a class="reference internal" href="../../../../super_gradients.training.html">Training package</a></li>
    47. 

  47. </ul>
    48. 

  48. <p class="caption"><span class="caption-text">User Guide</span></p>
    49. 

  49. <ul>
    50. 

  50. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html">What is SuperGradients?</a></li>
    51. 

  51. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#introducing-the-supergradients-library">Introducing the SuperGradients library</a></li>
    52. 

  52. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#installation">Installation</a></li>
    53. 

  53. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#integrating-your-training-code-complete-walkthrough">Integrating Your Training Code - Complete Walkthrough</a></li>
    54. 

  54. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#training-parameters">Training Parameters</a></li>
    55. 

  55. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#logs-and-checkpoints">Logs and Checkpoints</a></li>
    56. 

  56. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#dataset-parameters">Dataset Parameters</a></li>
    57. 

  57. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#network-architectures">Network Architectures</a></li>
    58. 

  58. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#pretrained-models">Pretrained Models</a></li>
    59. 

  59. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#how-to-reproduce-our-training-recipes">How To Reproduce Our Training Recipes</a></li>
    60. 

  60. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#professional-tools-integration">Professional Tools Integration</a></li>
    61. 

  61. <li class="toctree-l1"><a class="reference internal" href="../../../../user_guide.html#supergradients-faq">SuperGradients FAQ</a></li>
    62. 

  62. </ul>
    63. 

  63. 

  64.         </div>
    65. 

  65.       </div>
    66. 

  66.     </nav>
    67. 

  67. 

  68.     <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap"><nav class="wy-nav-top" aria-label="Mobile navigation menu" >
    69. 

  69.           <i data-toggle="wy-nav-top" class="fa fa-bars"></i>
    70. 

  70.           <a href="../../../../index.html">SuperGradients</a>
    71. 

  71.       </nav>
    72. 

  72. 

  73.       <div class="wy-nav-content">
    74. 

  74.         <div class="rst-content">
    75. 

  75.           <div role="navigation" aria-label="Page navigation">
    76. 

  76.   <ul class="wy-breadcrumbs">
    77. 

  77.       <li><a href="../../../../index.html" class="icon icon-home"></a> &raquo;</li>
    78. 

  78.           <li><a href="../../../index.html">Module code</a> &raquo;</li>
    79. 

  79.       <li>super_gradients.training.losses.yolox_loss</li>
    80. 

  80.       <li class="wy-breadcrumbs-aside">
    81. 

  81.       </li>
    82. 

  82.   </ul>
    83. 

  83.   <hr/>
    84. 

  84. </div>
    85. 

  85.           <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
    86. 

  86.            <div itemprop="articleBody">
    87. 

  87.              
    88. 

  88.   <h1>Source code for super_gradients.training.losses.yolox_loss</h1><div class="highlight"><pre>
    89. 

  89. <span></span><span class="sd">&quot;&quot;&quot;</span>
    90. 

  90. <span class="sd">Based on https://github.com/Megvii-BaseDetection/YOLOX (Apache-2.0 license)</span>
    91. 

  91. 

  92. <span class="sd">&quot;&quot;&quot;</span>
    93. 

  93. 

  94. <span class="kn">import</span> <span class="nn">logging</span>
    95. 

  95. <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>
    96. 

  96. 

  97. <span class="kn">import</span> <span class="nn">torch</span>
    98. 

  98. <span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">nn</span>
    99. 

  99. <span class="kn">from</span> <span class="nn">torch.nn.modules.loss</span> <span class="kn">import</span> <span class="n">_Loss</span>
    100. 

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

  101. <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>
    102. 

  102. <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>
    103. 

  103. 

  104. <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>
    105. 

  105. 

  106. 

  107. <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>
    108. 

  108.     <span class="sd">&quot;&quot;&quot;</span>
    109. 

  109. <span class="sd">    IoU loss with the following supported loss types:</span>
    110. 

  110. <span class="sd">    Attributes:</span>
    111. 

  111. <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>
    112. 

  112. <span class="sd">        loss_type: str: One of [&quot;iou&quot;, &quot;giou&quot;] where:</span>
    113. 

  113. <span class="sd">            * &#39;iou&#39; for</span>
    114. 

  114. <span class="sd">                (1 - iou^2)</span>
    115. 

  115. <span class="sd">            * &#39;giou&#39; according to &quot;Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression&quot;</span>
    116. 

  116. <span class="sd">                (1 - giou), where giou = iou - (cover_box - union_box)/cover_box</span>
    117. 

  117. <span class="sd">    &quot;&quot;&quot;</span>
    118. 

  118. 

  119.     <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>
    120. 

  120.         <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>
    121. 

  121.         <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>
    122. 

  122.         <span class="bp">self</span><span class="o">.</span><span class="n">reduction</span> <span class="o">=</span> <span class="n">reduction</span>
    123. 

  123.         <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>
    124. 

  124. 

  125.     <span class="nd">@staticmethod</span>
    126. 

  126.     <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>
    127. 

  127.         <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>
    128. 

  128.         <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>
    129. 

  129.         <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>
    130. 

  130.             <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="s1">&#39;, expected one of: &#39;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">supported_losses</span><span class="p">))</span>
    131. 

  131.         <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>
    132. 

  132.             <span class="k">raise</span> <span class="ne">ValueError</span><span class="p">(</span>
    133. 

  133.                 <span class="s2">&quot;Illegal reduction value: &quot;</span> <span class="o">+</span> <span class="n">reduction</span> <span class="o">+</span> <span class="s1">&#39;, expected one of: &#39;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">supported_reductions</span><span class="p">))</span>
    134. 

  134. 

  135.     <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>
    136. 

  136.         <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>
    137. 

  137. 

  138.         <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>
    139. 

  139.         <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>
    140. 

  140.         <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>
    141. 

  141.         <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>
    142. 

  142. 

  143.         <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>
    144. 

  144.         <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>
    145. 

  145. 

  146.         <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>
    147. 

  147.         <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>
    148. 

  148.         <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>
    149. 

  149.         <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>
    150. 

  150. 

  151.         <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>
    152. 

  152.             <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>
    153. 

  153.         <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>
    154. 

  154.             <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>
    155. 

  155.             <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>
    156. 

  156.             <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>
    157. 

  157.             <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>
    158. 

  158.             <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>
    159. 

  159. 

  160.         <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>
    161. 

  161.             <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>
    162. 

  162.         <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>
    163. 

  163.             <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>
    164. 

  164. 

  165.         <span class="k">return</span> <span class="n">loss</span>
    166. 

  166. 

  167. 

  168. <div class="viewcode-block" id="YoloXDetectionLoss"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    169. 

  169.     <span class="sd">&quot;&quot;&quot;</span>
    170. 

  170. <span class="sd">    Calculate YOLOX loss:</span>
    171. 

  171. <span class="sd">    L = L_objectivness + L_iou + L_classification + 1[use_l1]*L_l1</span>
    172. 

  172. 

  173. <span class="sd">    where:</span>
    174. 

  174. <span class="sd">        * L_iou, L_classification and L_l1 are calculated only between cells and targets that suit them;</span>
    175. 

  175. <span class="sd">        * L_objectivness is calculated for all cells.</span>
    176. 

  176. 

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

  178. <span class="sd">            for cells that have suitable ground truths in their grid locations add BCEs</span>
    179. 

  179. <span class="sd">            to force a prediction of IoU with a GT in a multi-label way</span>
    180. 

  180. <span class="sd">            Coef: 1.</span>
    181. 

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

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

  183. <span class="sd">            add (1 - IoU^2), IoU between a predicted box and each GT box, force maximum IoU</span>
    184. 

  184. <span class="sd">            Coef: 5.</span>
    185. 

  185. <span class="sd">        L_l1:</span>
    186. 

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

  187. <span class="sd">            l1 distance between the logits and GTs in “logits” format (the inverse of “logits to predictions” ops)</span>
    188. 

  188. <span class="sd">            Coef: 1[use_l1]</span>
    189. 

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

  190. <span class="sd">            for each cell add BCE with a label of 1 if there is GT assigned to the cell</span>
    191. 

  191. <span class="sd">            Coef: 1</span>
    192. 

  192. 

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

  194. <span class="sd">        strides: list: List of Yolo levels output grid sizes (i.e [8, 16, 32]).</span>
    195. 

  195. <span class="sd">        num_classes: int: Number of classes.</span>
    196. 

  196. <span class="sd">        use_l1: bool: Controls the L_l1 Coef as discussed above (default=False).</span>
    197. 

  197. <span class="sd">        center_sampling_radius: float: Sampling radius used for center sampling when creating the fg mask (default=2.5).</span>
    198. 

  198. <span class="sd">        iou_type: str: Iou loss type, one of [&quot;iou&quot;,&quot;giou&quot;] (deafult=&quot;iou&quot;).</span>
    199. 

  199. 

  200. 

  201. <span class="sd">    &quot;&quot;&quot;</span>
    202. 

  202. 

  203.     <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>
    204. 

  204.                  <span class="n">iou_type</span><span class="o">=</span><span class="s1">&#39;iou&#39;</span><span class="p">):</span>
    205. 

  205.         <span class="nb">super</span><span class="p">()</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
    206. 

  206.         <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>
    207. 

  207.         <span class="bp">self</span><span class="o">.</span><span class="n">strides</span> <span class="o">=</span> <span class="n">strides</span>
    208. 

  208.         <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>
    209. 

  209. 

  210.         <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>
    211. 

  211.         <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>
    212. 

  212.         <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>
    213. 

  213.         <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>
    214. 

  214.         <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>
    215. 

  215. 

  216. <div class="viewcode-block" id="YoloXDetectionLoss.forward"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    217. 

  217.         <span class="sd">&quot;&quot;&quot;</span>
    218. 

  218. <span class="sd">        :param model_output: Union[list, Tuple[torch.Tensor, List]]:</span>
    219. 

  219. <span class="sd">             When list-</span>
    220. 

  220. <span class="sd">              output from all Yolo levels, each of shape [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    221. 

  221. <span class="sd">             And when tuple- the second item is the described list (first item is discarded)</span>
    222. 

  222. 

  223. <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>
    224. 

  224. 

  225. <span class="sd">        :return: loss, all losses separately in a detached tensor</span>
    226. 

  226. <span class="sd">        &quot;&quot;&quot;</span>
    227. 

  227.         <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>
    228. 

  228.             <span class="c1"># in test/eval mode the Yolo model outputs a tuple where the second item is the raw predictions</span>
    229. 

  229.             <span class="n">_</span><span class="p">,</span> <span class="n">predictions</span> <span class="o">=</span> <span class="n">model_output</span>
    230. 

  230.         <span class="k">else</span><span class="p">:</span>
    231. 

  231.             <span class="n">predictions</span> <span class="o">=</span> <span class="n">model_output</span>
    232. 

  232. 

  233.         <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>
    234. 

  234. 

  235.     <span class="nd">@staticmethod</span>
    236. 

  236.     <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>
    237. 

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

  238. <span class="sd">        Creates a tensor of xy coordinates of size (1,1,nx,ny,2)</span>
    239. 

  239. 

  240. <span class="sd">        :param nx: int: cells along x axis (default=20)</span>
    241. 

  241. <span class="sd">        :param ny: int: cells along the y axis (default=20)</span>
    242. 

  242. <span class="sd">        :return: torch.tensor of xy coordinates of size (1,1,nx,ny,2)</span>
    243. 

  243. <span class="sd">        &quot;&quot;&quot;</span>
    244. 

  244.         <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>
    245. 

  245.         <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>
    246. 

  246. 

  247.     <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> \
    248. 

  248.             <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>
    249. 

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

  250. <span class="sd">        :param predictions:     output from all Yolo levels, each of shape</span>
    251. 

  251. <span class="sd">                                [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    252. 

  252. <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>
    253. 

  253. 

  254. <span class="sd">        :return:                loss, all losses separately in a detached tensor</span>
    255. 

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

  256.         <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>
    257. 

  257. 

  258.         <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>
    259. 

  259.         <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>
    260. 

  260.         <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>
    261. 

  261. 

  262.         <span class="c1"># calculate targets</span>
    263. 

  263.         <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>
    264. 

  264.         <span class="n">cls_targets</span> <span class="o">=</span> <span class="p">[]</span>
    265. 

  265.         <span class="n">reg_targets</span> <span class="o">=</span> <span class="p">[]</span>
    266. 

  266.         <span class="n">l1_targets</span> <span class="o">=</span> <span class="p">[]</span>
    267. 

  267.         <span class="n">obj_targets</span> <span class="o">=</span> <span class="p">[]</span>
    268. 

  268.         <span class="n">fg_masks</span> <span class="o">=</span> <span class="p">[]</span>
    269. 

  269. 

  270.         <span class="n">num_fg</span><span class="p">,</span> <span class="n">num_gts</span> <span class="o">=</span> <span class="mf">0.</span><span class="p">,</span> <span class="mf">0.</span>
    271. 

  271. 

  272.         <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>
    273. 

  273.             <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>
    274. 

  274.             <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>
    275. 

  275.             <span class="n">num_gts</span> <span class="o">+=</span> <span class="n">num_gt</span>
    276. 

  276.             <span class="k">if</span> <span class="n">num_gt</span> <span class="o">==</span> <span class="mi">0</span><span class="p">:</span>
    277. 

  277.                 <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>
    278. 

  278.                 <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>
    279. 

  279.                 <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>
    280. 

  280.                 <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>
    281. 

  281.                 <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>
    282. 

  282.             <span class="k">else</span><span class="p">:</span>
    283. 

  283.                 <span class="c1"># GT boxes to image coordinates</span>
    284. 

  284.                 <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>
    285. 

  285.                 <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>
    286. 

  286.                 <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>
    287. 

  287. 

  288.                 <span class="k">try</span><span class="p">:</span>
    289. 

  289.                     <span class="c1"># assign cells to ground truths, at most one GT per cell</span>
    290. 

  290.                     <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> \
    291. 

  291.                         <span class="bp">self</span><span class="o">.</span><span class="n">get_assignments</span><span class="p">(</span><span class="n">image_idx</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">,</span> <span class="n">gt_bboxes_per_image</span><span class="p">,</span>
    292. 

  292.                                              <span class="n">gt_classes</span><span class="p">,</span> <span class="n">bboxes_preds_per_image</span><span class="p">,</span>
    293. 

  293.                                              <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">cls_preds</span><span class="p">,</span> <span class="n">obj_preds</span><span class="p">)</span>
    294. 

  294. 

  295.                 <span class="c1"># TODO: CHECK IF ERROR IS CUDA OUT OF MEMORY</span>
    296. 

  296.                 <span class="k">except</span> <span class="ne">RuntimeError</span><span class="p">:</span>
    297. 

  297.                     <span class="n">logging</span><span class="o">.</span><span class="n">error</span><span class="p">(</span><span class="s2">&quot;OOM RuntimeError is raised due to the huge memory cost during label assignment. </span><span class="se">\</span>
    298. 

  298. <span class="s2">                                   CPU mode is applied in this batch. If you want to avoid this issue, </span><span class="se">\</span>
    299. 

  299. <span class="s2">                                   try to reduce the batch size or image size.&quot;</span><span class="p">)</span>
    300. 

  300.                     <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>
    301. 

  301.                     <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> \
    302. 

  302.                         <span class="bp">self</span><span class="o">.</span><span class="n">get_assignments</span><span class="p">(</span><span class="n">image_idx</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">,</span> <span class="n">gt_bboxes_per_image</span><span class="p">,</span>
    303. 

  303.                                              <span class="n">gt_classes</span><span class="p">,</span> <span class="n">bboxes_preds_per_image</span><span class="p">,</span>
    304. 

  304.                                              <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">cls_preds</span><span class="p">,</span> <span class="n">obj_preds</span><span class="p">,</span> <span class="s1">&#39;cpu&#39;</span><span class="p">)</span>
    305. 

  305. 

  306.                 <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>
    307. 

  307.                 <span class="n">num_fg</span> <span class="o">+=</span> <span class="n">num_fg_img</span>
    308. 

  308. 

  309.                 <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>
    310. 

  310.                 <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>
    311. 

  311.                 <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>
    312. 

  312.                 <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    313. 

  313.                     <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><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>
    314. 

  314.                                                    <span class="n">gt_bboxes_per_image</span><span class="p">[</span><span class="n">matched_gt_inds</span><span class="p">],</span> <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>
    315. 

  315.                                                    <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> <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>
    316. 

  316. 

  317.             <span class="c1"># collect targets for all loss terms over the whole batch</span>
    318. 

  318.             <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>
    319. 

  319.             <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>
    320. 

  320.             <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>
    321. 

  321.             <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>
    322. 

  322.             <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    323. 

  323.                 <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>
    324. 

  324. 

  325.         <span class="c1"># concat all targets over the batch (get rid of batch dim)</span>
    326. 

  326.         <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>
    327. 

  327.         <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>
    328. 

  328.         <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>
    329. 

  329.         <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>
    330. 

  330.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    331. 

  331.             <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>
    332. 

  332. 

  333.         <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>
    334. 

  334.         <span class="c1"># loss terms divided by the total number of foregrounds</span>
    335. 

  335.         <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>
    336. 

  336.         <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>
    337. 

  337.         <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>
    338. 

  338.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    339. 

  339.             <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>
    340. 

  340.         <span class="k">else</span><span class="p">:</span>
    341. 

  341.             <span class="n">loss_l1</span> <span class="o">=</span> <span class="mf">0.0</span>
    342. 

  342. 

  343.         <span class="n">reg_weight</span> <span class="o">=</span> <span class="mf">5.0</span>
    344. 

  344.         <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>
    345. 

  345. 

  346.         <span class="k">return</span> <span class="n">loss</span><span class="p">,</span> <span class="n">torch</span><span class="o">.</span><span class="n">cat</span><span class="p">((</span><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> <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> <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>
    347. 

  347.                                 <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>
    348. 

  348.                                 <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>
    349. 

  349.                                 <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><span class="o">.</span><span class="n">detach</span><span class="p">()</span>
    350. 

  350. 

  351. <div class="viewcode-block" id="YoloXDetectionLoss.prepare_predictions"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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> \
    352. 

  352.             <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>
    353. 

  353.         <span class="sd">&quot;&quot;&quot;</span>
    354. 

  354. <span class="sd">        Convert raw outputs of the network into a format that merges outputs from all levels</span>
    355. 

  355. <span class="sd">        :param predictions:     output from all Yolo levels, each of shape</span>
    356. 

  356. <span class="sd">                                [Batch x 1 x GridSizeY x GridSizeX x (4 + 1 + Num_classes)]</span>
    357. 

  357. <span class="sd">        :return:    5 tensors representing predictions:</span>
    358. 

  358. <span class="sd">                        * x_shifts: shape [1 x * num_cells x 1],</span>
    359. 

  359. <span class="sd">                          where num_cells = grid1X * grid1Y + grid2X * grid2Y + grid3X * grid3Y,</span>
    360. 

  360. <span class="sd">                          x coordinate on the grid cell the prediction is coming from</span>
    361. 

  361. <span class="sd">                        * y_shifts: shape [1 x  num_cells x 1],</span>
    362. 

  362. <span class="sd">                          y coordinate on the grid cell the prediction is coming from</span>
    363. 

  363. <span class="sd">                        * expanded_strides: shape [1 x num_cells x 1],</span>
    364. 

  364. <span class="sd">                          stride of the output grid the prediction is coming from</span>
    365. 

  365. <span class="sd">                        * transformed_outputs: shape [batch_size x num_cells x (num_classes + 5)],</span>
    366. 

  366. <span class="sd">                          predictions with boxes in real coordinates and logprobabilities</span>
    367. 

  367. <span class="sd">                        * raw_outputs: shape [batch_size x num_cells x (num_classes + 5)],</span>
    368. 

  368. <span class="sd">                          raw predictions with boxes and confidences as logits</span>
    369. 

  369. 

  370. <span class="sd">        &quot;&quot;&quot;</span>
    371. 

  371.         <span class="n">raw_outputs</span> <span class="o">=</span> <span class="p">[]</span>
    372. 

  372.         <span class="n">transformed_outputs</span> <span class="o">=</span> <span class="p">[]</span>
    373. 

  373.         <span class="n">x_shifts</span> <span class="o">=</span> <span class="p">[]</span>
    374. 

  374.         <span class="n">y_shifts</span> <span class="o">=</span> <span class="p">[]</span>
    375. 

  375.         <span class="n">expanded_strides</span> <span class="o">=</span> <span class="p">[]</span>
    376. 

  376.         <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>
    377. 

  377.             <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>
    378. 

  378. 

  379.             <span class="c1"># IN FIRST PASS CREATE GRIDS ACCORDING TO OUTPUT SHAPE (BATCH,1,IMAGE_H/STRIDE,IMAGE_2/STRIDE,NUM_CLASSES+5)</span>
    380. 

  380.             <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>
    381. 

  381.                 <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>
    382. 

  382. 

  383.             <span class="c1"># e.g. [batch_size, 1, 28, 28, 85] -&gt; [batch_size, 784, 85]</span>
    384. 

  384.             <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>
    385. 

  385.             <span class="c1"># e.g [1, 784, 2]</span>
    386. 

  386.             <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>
    387. 

  387.             <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    388. 

  388.                 <span class="c1"># e.g [1, 784, 4]</span>
    389. 

  389.                 <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>
    390. 

  390. 

  391.             <span class="c1"># box logits to coordinates</span>
    392. 

  392.             <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>
    393. 

  393.             <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>
    394. 

  394.             <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>
    395. 

  395.             <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>
    396. 

  396. 

  397.             <span class="c1"># outputs with boxes in real coordinates, probs as logits</span>
    398. 

  398.             <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>
    399. 

  399.             <span class="c1"># x cell coordinates of all 784 predictions, 0, 0, 0, ..., 1, 1, 1, ...</span>
    400. 

  400.             <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>
    401. 

  401.             <span class="c1"># y cell coordinates of all 784 predictions, 0, 1, 2, ..., 0, 1, 2, ...</span>
    402. 

  402.             <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>
    403. 

  403.             <span class="c1"># e.g. [1, 784, stride of this level (one of [8, 16, 32])]</span>
    404. 

  404.             <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>
    405. 

  405. 

  406.         <span class="c1"># all 4 below have shapes of [batch_size , num_cells, num_values_pre_cell]</span>
    407. 

  407.         <span class="c1"># where num_anchors * num_cells is e.g. 1 * (28 * 28 + 14 * 14 + 17 * 17)</span>
    408. 

  408.         <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>
    409. 

  409.         <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>
    410. 

  410.         <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>
    411. 

  411.         <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>
    412. 

  412.         <span class="k">if</span> <span class="bp">self</span><span class="o">.</span><span class="n">use_l1</span><span class="p">:</span>
    413. 

  413.             <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>
    414. 

  414. 

  415.         <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>
    416. 

  416. 

  417. <div class="viewcode-block" id="YoloXDetectionLoss.get_l1_target"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    418. 

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

  419. <span class="sd">        :param l1_target:   tensor of zeros of shape [Num_cell_gt_pairs x 4]</span>
    420. 

  420. <span class="sd">        :param gt:          targets in coordinates [Num_cell_gt_pairs x (4 + 1 + num_classes)]</span>
    421. 

  421. 

  422. <span class="sd">        :return:            targets in the format corresponding to logits</span>
    423. 

  423. <span class="sd">        &quot;&quot;&quot;</span>
    424. 

  424.         <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>
    425. 

  425.         <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>
    426. 

  426.         <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>
    427. 

  427.         <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>
    428. 

  428.         <span class="k">return</span> <span class="n">l1_target</span></div>
    429. 

  429. 

  430. <div class="viewcode-block" id="YoloXDetectionLoss.get_assignments"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    431. 

  431.     <span class="k">def</span> <span class="nf">get_assignments</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">image_idx</span><span class="p">,</span> <span class="n">num_gt</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">,</span> <span class="n">gt_bboxes_per_image</span><span class="p">,</span> <span class="n">gt_classes</span><span class="p">,</span>
    432. 

  432.                         <span class="n">bboxes_preds_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">cls_preds</span><span class="p">,</span>
    433. 

  433.                         <span class="n">obj_preds</span><span class="p">,</span> <span class="n">mode</span><span class="o">=</span><span class="s2">&quot;gpu&quot;</span><span class="p">,</span> <span class="n">ious_loss_cost_coeff</span><span class="o">=</span><span class="mf">3.0</span><span class="p">,</span> <span class="n">outside_boxes_and_center_cost_coeff</span><span class="o">=</span><span class="mf">100000.0</span><span class="p">):</span>
    434. 

  434.         <span class="sd">&quot;&quot;&quot;</span>
    435. 

  435. <span class="sd">        Match cells to ground truth:</span>
    436. 

  436. <span class="sd">            * at most 1 GT per cell</span>
    437. 

  437. <span class="sd">            * dynamic number of cells per GT</span>
    438. 

  438. 

  439. <span class="sd">        :param outside_boxes_and_center_cost_coeff: float: Cost coefficiant of cells the radius and bbox of gts in dynamic</span>
    440. 

  440. <span class="sd">         matching (default=100000).</span>
    441. 

  441. <span class="sd">        :param ious_loss_cost_coeff: float: Cost coefficiant for iou loss in dynamic matching (default=3).</span>
    442. 

  442. <span class="sd">        :param image_idx: int: Image index in batch.</span>
    443. 

  443. <span class="sd">        :param num_gt: int: Number of ground trunth targets in the image.</span>
    444. 

  444. <span class="sd">        :param total_num_anchors: int: Total number of possible bboxes = sum of all grid cells.</span>
    445. 

  445. <span class="sd">        :param gt_bboxes_per_image: torch.Tensor: Tensor of gt bboxes for  the image, shape: (num_gt, 4).</span>
    446. 

  446. <span class="sd">        :param gt_classes: torch.Tesnor: Tensor of the classes in the image, shape: (num_preds,4).</span>
    447. 

  447. <span class="sd">        :param bboxes_preds_per_image: Tensor of the classes in the image, shape: (num_preds).</span>
    448. 

  448. <span class="sd">        :param expanded_strides: torch.Tensor: Stride of the output grid the prediction is coming from,</span>
    449. 

  449. <span class="sd">            shape (1 x num_cells x 1).</span>
    450. 

  450. <span class="sd">        :param x_shifts: torch.Tensor: X&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    451. 

  451. <span class="sd">        :param y_shifts: torch.Tensor: Y&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    452. 

  452. <span class="sd">        :param cls_preds: torch.Tensor: Class predictions in all cells, shape (batch_size, num_cells).</span>
    453. 

  453. <span class="sd">        :param obj_preds: torch.Tensor: Objectness predictions in all cells, shape (batch_size, num_cells).</span>
    454. 

  454. <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>
    455. 

  455. 

  456. <span class="sd">        &quot;&quot;&quot;</span>
    457. 

  457.         <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>
    458. 

  458.             <span class="nb">print</span><span class="p">(</span><span class="s2">&quot;------------CPU Mode for This Batch-------------&quot;</span><span class="p">)</span>
    459. 

  459.             <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>
    460. 

  460.             <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>
    461. 

  461.             <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>
    462. 

  462.             <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>
    463. 

  463.             <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>
    464. 

  464.             <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>
    465. 

  465. 

  466.         <span class="c1"># create a mask for foreground cells</span>
    467. 

  467.         <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>
    468. 

  468.                                                                  <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>
    469. 

  469. 

  470.         <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>
    471. 

  471.         <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>
    472. 

  472.         <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>
    473. 

  473.         <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>
    474. 

  474. 

  475.         <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>
    476. 

  476.             <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>
    477. 

  477.             <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>
    478. 

  478. 

  479.         <span class="c1"># calculate cost between all foregrounds and all ground truths (used only for matching)</span>
    480. 

  480.         <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>
    481. 

  481.         <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>
    482. 

  482.         <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>
    483. 

  483.         <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>
    484. 

  484. 

  485.         <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>
    486. 

  486.             <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>
    487. 

  487. 

  488.         <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>
    489. 

  489.             <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>
    490. 

  490.             <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>
    491. 

  491.         <span class="k">del</span> <span class="n">cls_preds_</span>
    492. 

  492. 

  493.         <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>
    494. 

  494.             <span class="o">~</span><span class="n">is_in_boxes_and_center</span><span class="p">)</span>
    495. 

  495. 

  496.         <span class="c1"># further filter foregrounds: create pairs between cells and ground truth, based on cost and IoUs</span>
    497. 

  497.         <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> \
    498. 

  498.             <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>
    499. 

  499.         <span class="c1"># discard tensors related to cost</span>
    500. 

  500.         <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>
    501. 

  501. 

  502.         <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>
    503. 

  503.             <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>
    504. 

  504.             <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>
    505. 

  505.             <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>
    506. 

  506.             <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>
    507. 

  507. 

  508.         <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>
    509. 

  509. 

  510. <div class="viewcode-block" id="YoloXDetectionLoss.get_in_boxes_info"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    511. 

  511.         <span class="sd">&quot;&quot;&quot;</span>
    512. 

  512. <span class="sd">        Create a mask for all cells, mask in only foreground: cells that have a center located:</span>
    513. 

  513. <span class="sd">            * withing a GT box;</span>
    514. 

  514. <span class="sd">            OR</span>
    515. 

  515. <span class="sd">            * within a fixed radius around a GT box (center sampling);</span>
    516. 

  516. 

  517. <span class="sd">        :param num_gt: int: Number of ground trunth targets in the image.</span>
    518. 

  518. <span class="sd">        :param total_num_anchors: int: Sum of all grid cells.</span>
    519. 

  519. <span class="sd">        :param gt_bboxes_per_image: torch.Tensor: Tensor of gt bboxes for  the image, shape: (num_gt, 4).</span>
    520. 

  520. <span class="sd">        :param expanded_strides: torch.Tensor: Stride of the output grid the prediction is coming from,</span>
    521. 

  521. <span class="sd">            shape (1 x num_cells x 1).</span>
    522. 

  522. <span class="sd">        :param x_shifts: torch.Tensor: X&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    523. 

  523. <span class="sd">        :param y_shifts: torch.Tensor: Y&#39;s in cell coordinates, shape (1,num_cells,1).</span>
    524. 

  524. 

  525. <span class="sd">        :return is_in_boxes_anchor, is_in_boxes_and_center</span>
    526. 

  526. <span class="sd">            where:</span>
    527. 

  527. <span class="sd">             - is_in_boxes_anchor masks the cells that their cell center is  inside a gt bbox and within</span>
    528. 

  528. <span class="sd">                self.center_sampling_radius cells away, without reduction (i.e shape=(num_gts, num_fgs))</span>
    529. 

  529. <span class="sd">             - is_in_boxes_and_center masks the cells that their center is either inside a gt bbox or within</span>
    530. 

  530. <span class="sd">                self.center_sampling_radius cells away, shape (num_fgs)</span>
    531. 

  531. <span class="sd">        &quot;&quot;&quot;</span>
    532. 

  532. 

  533.         <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>
    534. 

  534. 

  535.         <span class="c1"># cell coordinates, shape [n_predictions] -&gt; repeated to [n_gts, n_predictions]</span>
    536. 

  536.         <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>
    537. 

  537.         <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>
    538. 

  538.         <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>
    539. 

  539.         <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>
    540. 

  540. 

  541.         <span class="c1"># FIND CELL CENTERS THAT ARE WITHIN GROUND TRUTH BOXES</span>
    542. 

  542. 

  543.         <span class="c1"># ground truth boxes, shape [n_gts] -&gt; repeated to [n_gts, n_predictions]</span>
    544. 

  544.         <span class="c1"># from (c1, c2, w, h) to left, right, top, bottom</span>
    545. 

  545.         <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>
    546. 

  546.         <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>
    547. 

  547.         <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>
    548. 

  548.         <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>
    549. 

  549. 

  550.         <span class="c1"># check which cell centers lay within the ground truth boxes</span>
    551. 

  551.         <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>
    552. 

  552.         <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>
    553. 

  553.         <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>
    554. 

  554.         <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>
    555. 

  555.         <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>
    556. 

  556. 

  557.         <span class="c1"># to claim that a cell center is inside a gt box all 4 differences calculated above should be positive</span>
    558. 

  558.         <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>
    559. 

  559.         <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>
    560. 

  560. 

  561.         <span class="c1"># FIND CELL CENTERS THAT ARE WITHIN +- self.center_sampling_radius CELLS FROM GROUND TRUTH BOXES CENTERS</span>
    562. 

  562. 

  563.         <span class="c1"># define fake boxes: instead of ground truth boxes step +- self.center_sampling_radius from their centers</span>
    564. 

  564.         <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><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span> <span class="o">-</span>
    565. 

  565.                                  <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>
    566. 

  566.         <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><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span> <span class="o">+</span>
    567. 

  567.                                  <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>
    568. 

  568.         <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><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span> <span class="o">-</span>
    569. 

  569.                                  <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>
    570. 

  570.         <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><span class="mi">1</span><span class="p">,</span> <span class="n">total_num_anchors</span><span class="p">)</span> <span class="o">+</span>
    571. 

  571.                                  <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>
    572. 

  572. 

  573.         <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>
    574. 

  574.         <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>
    575. 

  575.         <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>
    576. 

  576.         <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>
    577. 

  577.         <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>
    578. 

  578.         <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>
    579. 

  579.         <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>
    580. 

  580. 

  581.         <span class="c1"># in boxes OR in centers</span>
    582. 

  582.         <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>
    583. 

  583. 

  584.         <span class="c1"># in boxes AND in centers, preserving a shape [num_GTs x num_FGs]</span>
    585. 

  585.         <span class="n">is_in_boxes_and_center</span> <span class="o">=</span> <span class="p">(</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>
    586. 

  586.         <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>
    587. 

  587. 

  588. <div class="viewcode-block" id="YoloXDetectionLoss.dynamic_k_matching"><a class="viewcode-back" href="../../../../super_gradients.training.losses.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>
    589. 

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

  590. <span class="sd">        :param cost:            pairwise cost, [num_FGs x num_GTs]</span>
    591. 

  591. <span class="sd">        :param pair_wise_ious:  pairwise IoUs, [num_FGs x num_GTs]</span>
    592. 

  592. <span class="sd">        :param gt_classes:      class of each GT</span>
    593. 

  593. <span class="sd">        :param num_gt:          number of GTs</span>
    594. 

  594. 

  595. <span class="sd">        :return num_fg, (number of foregrounds)</span>
    596. 

  596. <span class="sd">                gt_matched_classes, (the classes that have been matched with fgs)</span>
    597. 

  597. <span class="sd">                pred_ious_this_matching</span>
    598. 

  598. <span class="sd">                matched_gt_inds</span>
    599. 

  599. <span class="sd">        &quot;&quot;&quot;</span>
    600. 

  600.         <span class="c1"># create a matrix with shape [num_GTs x num_FGs]</span>
    601. 

  601.         <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>
    602. 

  602. 

  603.         <span class="c1"># for each GT get a dynamic k of foregrounds with a minimum cost: k = int(sum[top 10 IoUs])</span>
    604. 

  604.         <span class="n">ious_in_boxes_matrix</span> <span class="o">=</span> <span class="n">pair_wise_ious</span>
    605. 

  605.         <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>
    606. 

  606.         <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>
    607. 

  607.         <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>
    608. 

  608.         <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>
    609. 

  609.         <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>
    610. 

  610.             <span class="k">try</span><span class="p">:</span>
    611. 

  611.                 <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>
    612. 

  612.             <span class="k">except</span> <span class="ne">Exception</span><span class="p">:</span>
    613. 

  613.                 <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>
    614. 

  614.             <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>
    615. 

  615. 

  616.         <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>
    617. 

  617. 

  618.         <span class="c1"># leave at most one GT per foreground, chose the one with the smallest cost</span>
    619. 

  619.         <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>
    620. 

  620.         <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>
    621. 

  621.             <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>
    622. 

  622.             <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>
    623. 

  623.             <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>
    624. 

  624. 

  625.         <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>
    626. 

  626.         <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>
    627. 

  627. 

  628.         <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>
    629. 

  629. 

  630.         <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>
    631. 

  631.         <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>
    632. 

  632. 

  633.         <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>
    634. 

  634.         <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>
    635. 

  635. </pre></div>
    636. 

  636. 

  637.            </div>
    638. 

  638.           </div>
    639. 

  639.           <footer>
    640. 

  640. 

  641.   <hr/>
    642. 

  642. 

  643.   <div role="contentinfo">
    644. 

  644.     <p>&#169; Copyright 2021, SuperGradients team.</p>
    645. 

  645.   </div>
    646. 

  646. 

  647.   Built with <a href="https://www.sphinx-doc.org/">Sphinx</a> using a
    648. 

  648.     <a href="https://github.com/readthedocs/sphinx_rtd_theme">theme</a>
    649. 

  649.     provided by <a href="https://readthedocs.org">Read the Docs</a>.
    650. 

  650.    
    651. 

  651. 

  652. </footer>
    653. 

  653.         </div>
    654. 

  654.       </div>
    655. 

  655.     </section>
    656. 

  656.   </div>
    657. 

  657.   <script>
    658. 

  658.       jQuery(function () {
    659. 

  659.           SphinxRtdTheme.Navigation.enable(true);
    660. 

  660.       });
    661. 

  661.   </script> 
    662. 

  662. 

  663. </body>
    664. 

  664. </html>
    665.
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