-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
1809 lines (608 loc) · 106 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html class="theme-next gemini use-motion" lang="">
<head>
<meta charset="UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=edge" />
<meta name="viewport" content="width=device-width, initial-scale=1, maximum-scale=1"/>
<meta name="theme-color" content="#222">
<meta http-equiv="Cache-Control" content="no-transform" />
<meta http-equiv="Cache-Control" content="no-siteapp" />
<link href="/lib/fancybox/source/jquery.fancybox.css?v=2.1.5" rel="stylesheet" type="text/css" />
<link href="/lib/font-awesome/css/font-awesome.min.css?v=4.6.2" rel="stylesheet" type="text/css" />
<link href="/css/main.css?v=5.1.4" rel="stylesheet" type="text/css" />
<link rel="apple-touch-icon" sizes="180x180" href="/images/apple-touch-icon-next.png?v=5.1.4">
<link rel="icon" type="image/png" sizes="32x32" href="/images/favicon-32x32-next.png?v=5.1.4">
<link rel="icon" type="image/png" sizes="16x16" href="/images/favicon-16x16-next.png?v=5.1.4">
<link rel="mask-icon" href="/images/logo.svg?v=5.1.4" color="#222">
<meta name="keywords" content="Hexo, NexT" />
<meta name="description" content="Why always me - Mario Balotelli">
<meta property="og:type" content="website">
<meta property="og:title" content="chenghaz Official 1.0">
<meta property="og:url" content="http://chenghaz.github.io/index.html">
<meta property="og:site_name" content="chenghaz Official 1.0">
<meta property="og:description" content="Why always me - Mario Balotelli">
<meta property="og:locale" content="default">
<meta name="twitter:card" content="summary">
<meta name="twitter:title" content="chenghaz Official 1.0">
<meta name="twitter:description" content="Why always me - Mario Balotelli">
<script type="text/javascript" id="hexo.configurations">
var NexT = window.NexT || {};
var CONFIG = {
root: '/',
scheme: 'Gemini',
version: '5.1.4',
sidebar: {"position":"left","display":"post","offset":12,"b2t":false,"scrollpercent":false,"onmobile":false},
fancybox: true,
tabs: true,
motion: {"enable":true,"async":false,"transition":{"post_block":"fadeIn","post_header":"slideDownIn","post_body":"slideDownIn","coll_header":"slideLeftIn","sidebar":"slideUpIn"}},
duoshuo: {
userId: '0',
author: 'Author'
},
algolia: {
applicationID: '',
apiKey: '',
indexName: '',
hits: {"per_page":10},
labels: {"input_placeholder":"Search for Posts","hits_empty":"We didn't find any results for the search: ${query}","hits_stats":"${hits} results found in ${time} ms"}
}
};
</script>
<link rel="canonical" href="http://chenghaz.github.io/"/>
<title>chenghaz Official 1.0</title>
</head>
<body itemscope itemtype="http://schema.org/WebPage" lang="default">
<div class="container sidebar-position-left
page-home">
<div class="headband"></div>
<header id="header" class="header" itemscope itemtype="http://schema.org/WPHeader">
<div class="header-inner"><div class="site-brand-wrapper">
<div class="site-meta ">
<div class="custom-logo-site-title">
<a href="/" class="brand" rel="start">
<span class="logo-line-before"><i></i></span>
<span class="site-title">chenghaz Official 1.0</span>
<span class="logo-line-after"><i></i></span>
</a>
</div>
<p class="site-subtitle"></p>
</div>
<div class="site-nav-toggle">
<button>
<span class="btn-bar"></span>
<span class="btn-bar"></span>
<span class="btn-bar"></span>
</button>
</div>
</div>
<nav class="site-nav">
<ul id="menu" class="menu">
<li class="menu-item menu-item-home">
<a href="/" rel="section">
<i class="menu-item-icon fa fa-fw fa-home"></i> <br />
Home
</a>
</li>
<li class="menu-item menu-item-about">
<a href="/about/" rel="section">
<i class="menu-item-icon fa fa-fw fa-user"></i> <br />
About
</a>
</li>
<li class="menu-item menu-item-archives">
<a href="/archives/" rel="section">
<i class="menu-item-icon fa fa-fw fa-archive"></i> <br />
Archives
</a>
</li>
</ul>
</nav>
</div>
</header>
<main id="main" class="main">
<div class="main-inner">
<div class="content-wrap">
<div id="content" class="content">
<section id="posts" class="posts-expand">
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2019/12/26/book-list/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2019/12/26/book-list/" itemprop="url">书单</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2019-12-26T15:06:24-08:00">
2019-12-26
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h4 id="理财，投资"><a href="#理财，投资" class="headerlink" title="理财，投资"></a>理财，投资</h4><h5 id="给投资新手的极简股票课"><a href="#给投资新手的极简股票课" class="headerlink" title="给投资新手的极简股票课"></a>给投资新手的极简股票课</h5>
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</div>
</article>
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2018/12/06/semantic-segmantation-3/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2018/12/06/semantic-segmantation-3/" itemprop="url">semantic-segmantation-3</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2018-12-06T11:21:49-08:00">
2018-12-06
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h2 id="Pyramid-Scene-Parsing-Network-PSPNet"><a href="#Pyramid-Scene-Parsing-Network-PSPNet" class="headerlink" title="Pyramid Scene Parsing Network (PSPNet)"></a>Pyramid Scene Parsing Network (PSPNet)</h2><h3 id="Several-failure-cases-for-FCN"><a href="#Several-failure-cases-for-FCN" class="headerlink" title="Several failure cases for FCN"></a>Several failure cases for FCN</h3><ol>
<li>For semantic segmentation, it is important for a network to understand the co-occurrent visual patterns. For example, a signal light is likely appear on a road rather than flight.</li>
<li>The label for several classes could be confusing. e.g. field and earth.</li>
<li>Several small-size things, like streetlight and signboard, are hard to find while they may be of great importance. Contrarily, big objects or stuff may exceed the receptive field of FCN and thus cause discontinuous prediction.</li>
</ol>
<h3 id="Pyramid-Pooling"><a href="#Pyramid-Pooling" class="headerlink" title="Pyramid Pooling"></a>Pyramid Pooling</h3><p>The empirical receptive field of CNN is much smaller than the theoretical one especially on high-level layers. Global average pooling is a good baseline model as the global contextual prior, which is commonly used in image classification tasks. So, the global context information along with sub-region context is helpful in semantic segmentation task.</p>
<p>In this paper, the authors proposed a hierarchical global prior, containing information with different scales and varying among different sub-regions.<br><img src="/2018/12/06/semantic-segmantation-3/1.png"></p>
<p>PSPNet provides an effective global contextual prior for pixel-level scene parsing. For authors setting, pyramid pooling module is a four-level one with bin sizes of 1×1, 2×2, 3×3 and 6×6 respectively.</p>
<h2 id="ICNet-for-Real-Time-Semantic-Segmentation-on-High-Resolution-Images"><a href="#ICNet-for-Real-Time-Semantic-Segmentation-on-High-Resolution-Images" class="headerlink" title="ICNet for Real-Time Semantic Segmentation on High-Resolution Images"></a>ICNet for Real-Time Semantic Segmentation on High-Resolution Images</h2><p>The target for this paper is to build a practically fast semantic segmentation system with decent prediction accuracy.</p>
<h3 id="Image-Cascade-Network-ICNet"><a href="#Image-Cascade-Network-ICNet" class="headerlink" title="Image Cascade Network (ICNet)"></a>Image Cascade Network (ICNet)</h3><p>It takes cascade image inputs (i.e., low-, medium- and high resolution images), adopts cascade feature fusion and is trained with cascade label guidance.</p>
<p>A 1/4 sized image is fed into PSPNet with downsampling rate 8, resulting in a 1/32-resolution feature map. To get high quality segmentation, medium and high resolution branches (middle and bottom parts in Fig. 2) help recover and refine the coarse prediction.</p>
<p>Light weighted CNNs (green dotted box) are adopted in higher resolution branches; different-branch output feature maps are fused by cascade-feature-fusion unit and trained with cascade label guidance.</p>
<p>Weights and Computation (in 17 layers) can be shared between low- and medium-branches.<br><img src="/2018/12/06/semantic-segmantation-3/2.png"></p>
<h3 id="Cascade-Feature-Fusion"><a href="#Cascade-Feature-Fusion" class="headerlink" title="Cascade Feature Fusion"></a>Cascade Feature Fusion</h3><img src="/2018/12/06/semantic-segmantation-3/3.png">
<p>F2 is with doubled spatial size of F1.<br>We first apply upsampling rate 2 on F1 through bilinear interpolation, yielding the same spatial size as F2. Then a dilated convolution layer with kernel size C3 × 3 × 3 and dilation 2 is applied to refine the upsampled features.</p>
<p>For feature F2, a projection convolution with kernel size C3 ×1×1 is utilized to project F2 so that it has the same number of channels as the output of F1. Then two batch normalization layers are used to normalize these two processed</p>
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</div>
</article>
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2018/10/17/semantic-segmantation-2/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2018/10/17/semantic-segmantation-2/" itemprop="url">Semantic Segmentation Overview (part2)</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2018-10-17T11:43:36-07:00">
2018-10-17
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h2 id="Rethinking-Atrous-Convolution-for-Semantic-Image-Segmentation-DeepLabv3"><a href="#Rethinking-Atrous-Convolution-for-Semantic-Image-Segmentation-DeepLabv3" class="headerlink" title="Rethinking Atrous Convolution for Semantic Image Segmentation (DeepLabv3)"></a>Rethinking Atrous Convolution for Semantic Image Segmentation (DeepLabv3)</h2><h3 id="Proposed-Model"><a href="#Proposed-Model" class="headerlink" title="Proposed Model"></a>Proposed Model</h3><p>Applied atrous convolution as the feature extraction method. And employ in cascade or parallel. One advantage of atrous convolution is that it can keep the dimension of each feature map and allow us to extract denser feature responses.</p>
<p>Note that <em>output_stride</em> is the ratio of input image spatial resolution to final output resolution.</p>
<h4 id="Going-Deeper-with-Atrous-Convolution"><a href="#Going-Deeper-with-Atrous-Convolution" class="headerlink" title="Going Deeper with Atrous Convolution"></a>Going Deeper with Atrous Convolution</h4><p>Cascade version of atrous convolution with ResNet architecture. The motivation of this model is that the introduced striding makes it easy to capture long range information in the deeper blocks.<br><img src="/2018/10/17/semantic-segmantation-2/1.png"><br>Also, the author introduced Multi-grid method, which employ a hierarchy of grids of different sizes.<br><img src="/2018/10/17/semantic-segmantation-2/2.png" title="Multigrid Neural Architectures"></p>
<h4 id="Atrous-Spatial-Pyramid-Pooling"><a href="#Atrous-Spatial-Pyramid-Pooling" class="headerlink" title="Atrous Spatial Pyramid Pooling"></a>Atrous Spatial Pyramid Pooling</h4><p>Atrous Spatial Pyrimid Pooling is applied on top of the feature map, with four parallel atrous convolutions with different atrous rates.</p>
<p>Final ASPP: One 1x1 convolution and three 3x3 convolutions with rates(6, 12, 18) when <em>output_stride</em> is 16. Then concat all branches results and pass through another 1x1 convolution before the final 1x1 convolution which generates the final logits. </p>
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</div>
</article>
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2018/10/15/semantic-segmantation/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2018/10/15/semantic-segmantation/" itemprop="url">Semantic Segmentation Overview (part1)</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2018-10-15T14:49:13-07:00">
2018-10-15
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h2 id="Encoder-Decoder-with-Atrous-Separable-Convolution-for-Semantic-Image-Segmentation"><a href="#Encoder-Decoder-with-Atrous-Separable-Convolution-for-Semantic-Image-Segmentation" class="headerlink" title="Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation"></a>Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation</h2><p>(The model also named DeepLab v3+)</p>
<h3 id="Related-Work"><a href="#Related-Work" class="headerlink" title="Related Work"></a>Related Work</h3><img src="/2018/10/15/semantic-segmantation/1.png">
<ol>
<li><strong>Spatial pyramid pooling</strong>:<br>Models, such as PSPNet or DeepLab, perform spatial pyramid pooling at several grid scales (including image level pooling) or apply several parallel atrous convolution with different rates (called Atrous Spatial Pyramid Pooling, or ASPP).</li>
<li><strong>Encoder-decoder</strong>: Typically, the encoder-decoder networks contain (1) an encoder module that gradually reduces the feature maps and captures higher semantic information, and (2) a decoder module that gradually recovers the spatial information.</li>
</ol>
<h3 id="Proposed-Model"><a href="#Proposed-Model" class="headerlink" title="Proposed Model"></a>Proposed Model</h3><img src="/2018/10/15/semantic-segmantation/2.png">
<p>Encoder-Decoder with Atrous convolution</p>
<h4 id="Atrous-Convolution"><a href="#Atrous-Convolution" class="headerlink" title="Atrous Convolution"></a>Atrous Convolution</h4><img src="/2018/10/15/semantic-segmantation/3.gif" title="https://towardsdatascience.com/types-of-convolutions-in-deep-learning-717013397f4d">
<p>Atrous convolution allows us to explicitly control the resolution of features computed by deep convolutional neural networks and adjust filter’s field-of-view in order to capture multi-scale information, generalizes standard convolution operation.</p>
<h4 id="Depthwise-separable-convolution"><a href="#Depthwise-separable-convolution" class="headerlink" title="Depthwise separable convolution"></a>Depthwise separable convolution</h4><p>Like MobileNet or MobileNetV2</p>
<h4 id="Encoder-DeepLab-v3"><a href="#Encoder-DeepLab-v3" class="headerlink" title="Encoder: DeepLab v3"></a>Encoder: DeepLab v3</h4><p>Use the last feature map before logits in the original DeepLabv3 as the encoder output in our proposed encoder-decoder structure. Also, author mentioned a output stride as a Hyperparameter, which means the size relationship between original image and output feature map.</p>
<h4 id="Decoder"><a href="#Decoder" class="headerlink" title="Decoder"></a>Decoder</h4><p>The naive bilinear upsample cannot recover semantic details of an image. The author mentioned a method that combine bilinear upsample and convolution together.<br>The decoder work flow for DeepLab v3+ is: upsample -> concat selected low level feature -> convolution -> upsample -> result</p>
<h4 id="Xception"><a href="#Xception" class="headerlink" title="Xception"></a>Xception</h4><p>One of the backbone model.</p>
<h2 id="PARSENET-LOOKING-WIDER-TO-SEE-BETTER"><a href="#PARSENET-LOOKING-WIDER-TO-SEE-BETTER" class="headerlink" title="PARSENET: LOOKING WIDER TO SEE BETTER"></a>PARSENET: LOOKING WIDER TO SEE BETTER</h2><ul>
<li>FCN disregards global information about an image, thus ignoring potentially useful scene-level semantic context</li>
<li>CRF(conditional random field) & DNN require a fair amount of experience in managing the idiosyncrasies of training methodology and parameters.</li>
</ul>
<h3 id="Model-ParseNet"><a href="#Model-ParseNet" class="headerlink" title="Model - ParseNet"></a>Model - ParseNet</h3><h4 id="Global-context"><a href="#Global-context" class="headerlink" title="Global context"></a>Global context</h4><p>Context is known to be very useful for improving performance on detection and segmentation tasks using deep learning. However, in practice, for FCN, the empirical size of the receptive fields is much smaller, and is not enough to capture the global context. The author proposed a global average pooling and pool the context features from the last layer or any layer if that is desired. Then concat global feature to feature map to combine the global context to model.</p>
<h4 id="Early-Fusion-and-Late-Fusion"><a href="#Early-Fusion-and-Late-Fusion" class="headerlink" title="Early Fusion and Late Fusion"></a>Early Fusion and Late Fusion</h4><ul>
<li>Early fusion: unpool (replicate) global feature to the same size as of local feature map spatially and then concatenate them.</li>
<li>Late fusion: each feature is used to learn its own classifier, followed by merging the two predictions into a single classification score.</li>
</ul>
<h4 id="l2-normalization-layer"><a href="#l2-normalization-layer" class="headerlink" title="l2 normalization layer"></a>l2 normalization layer</h4><p>Naively concatenating features leads to poor performance as the ”larger” features dominate the ”smaller” ones. The author applied L2-norm and learned the scale parameter for each channel before using the feature for classification, which leads to more stable training. The author mentioned l2-norm on each pixel on feature map instead of whole feature map.</p>
<h2 id="TBD"><a href="#TBD" class="headerlink" title="TBD"></a>TBD</h2><ol>
<li>DeepLab v3</li>
<li>Xception</li>
</ol>
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</div>
</article>
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2018/09/09/tf-object-detection-2/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2018/09/09/tf-object-detection-2/" itemprop="url">Reading note for tensorflow/models/research/object_detection (part2)</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2018-09-09T22:55:40-07:00">
2018-09-09
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h2 id="Reading-note-for-tensorflow-models-research-object-detection-part2"><a href="#Reading-note-for-tensorflow-models-research-object-detection-part2" class="headerlink" title="Reading note for tensorflow/models/research/object_detection (part2)"></a>Reading note for tensorflow/models/research/object_detection (part2)</h2><h3 id="Path-for-training-start-from-model-main-py"><a href="#Path-for-training-start-from-model-main-py" class="headerlink" title="Path for training: start from model_main.py"></a>Path for training: start from model_main.py</h3><p>First, we need to specify a config file: <a href="https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/configuring_jobs.md" target="_blank" rel="noopener">official document</a>.<br>It contains several parts: model, train_config, train_input_reader, eval_config, eval_input_reader. Now we move to main function.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br></pre></td><td class="code"><pre><span class="line"><span class="comment"># Config file we specified before</span></span><br><span class="line">flags.DEFINE_string(<span class="string">'pipeline_config_path'</span>, <span class="keyword">None</span>, <span class="string">'Path to pipeline config file.'</span>)</span><br><span class="line">...</span><br><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">main</span><span class="params">(unused_argv)</span>:</span></span><br><span class="line"> ...</span><br><span class="line"> train_and_eval_dict = model_lib.create_estimator_and_inputs(</span><br><span class="line"> run_config=config,</span><br><span class="line"> hparams=model_hparams.create_hparams(FLAGS.hparams_overrides),</span><br><span class="line"> pipeline_config_path=FLAGS.pipeline_config_path,</span><br><span class="line"> train_steps=FLAGS.num_train_steps,</span><br><span class="line"> eval_steps=FLAGS.num_eval_steps)</span><br></pre></td></tr></table></figure></p>
<h3 id="File-model-lib-py-gt-help-functions-for-model-main-py"><a href="#File-model-lib-py-gt-help-functions-for-model-main-py" class="headerlink" title="File model_lib.py -> help functions for model_main.py"></a>File model_lib.py -> help functions for model_main.py</h3><p>jump to <em>model_lib.create_estimator_and_inputs</em> function<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">53</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">create_estimator_and_inputs</span><span class="params">(run_config,</span></span></span><br><span class="line"><span class="function"><span class="params"> hparams,</span></span></span><br><span class="line"><span class="function"><span class="params"> pipeline_config_path,</span></span></span><br><span class="line"><span class="function"><span class="params"> train_steps=None,</span></span></span><br><span class="line"><span class="function"><span class="params"> eval_steps=None,</span></span></span><br><span class="line"><span class="function"><span class="params"> model_fn_creator=create_model_fn,</span></span></span><br><span class="line"><span class="function"><span class="params"> use_tpu_estimator=False,</span></span></span><br><span class="line"><span class="function"><span class="params"> use_tpu=False,</span></span></span><br><span class="line"><span class="function"><span class="params"> num_shards=<span class="number">1</span>,</span></span></span><br><span class="line"><span class="function"><span class="params"> params=None,</span></span></span><br><span class="line"><span class="function"><span class="params"> **kwargs)</span>:</span></span><br><span class="line"> <span class="comment"># Map methods</span></span><br><span class="line"> <span class="comment"># 'get_configs_from_pipeline_file': config_util.get_configs_from_pipeline_file</span></span><br><span class="line"> <span class="comment"># This function is just read pipeline file into protobuf</span></span><br><span class="line"> get_configs_from_pipeline_file = MODEL_BUILD_UTIL_MAP[</span><br><span class="line"> <span class="string">'get_configs_from_pipeline_file'</span>]</span><br><span class="line"> <span class="comment"># 'merge_external_params_with_configs': config_util.merge_external_params_with_configs</span></span><br><span class="line"> merge_external_params_with_configs = MODEL_BUILD_UTIL_MAP[</span><br><span class="line"> <span class="string">'merge_external_params_with_configs'</span>]</span><br><span class="line"> <span class="comment"># 'create_pipeline_proto_from_configs': config_util.create_pipeline_proto_from_configs</span></span><br><span class="line"> create_pipeline_proto_from_configs = MODEL_BUILD_UTIL_MAP[</span><br><span class="line"> <span class="string">'create_pipeline_proto_from_configs'</span>]</span><br><span class="line"> <span class="comment">####################################################################</span></span><br><span class="line"> <span class="comment"># -> jump to config_util.create_pipeline_proto_from_configs function</span></span><br><span class="line"> <span class="comment">####################################################################</span></span><br><span class="line"> pipeline_config = pipeline_pb2.TrainEvalPipelineConfig()</span><br><span class="line"> pipeline_config.model.CopyFrom(configs[<span class="string">"model"</span>])</span><br><span class="line"> pipeline_config.train_config.CopyFrom(configs[<span class="string">"train_config"</span>])</span><br><span class="line"> pipeline_config.train_input_reader.CopyFrom(configs[<span class="string">"train_input_config"</span>])</span><br><span class="line"> pipeline_config.eval_config.CopyFrom(configs[<span class="string">"eval_config"</span>])</span><br><span class="line"> pipeline_config.eval_input_reader.CopyFrom(configs[<span class="string">"eval_input_config"</span>])</span><br><span class="line"> <span class="comment">####################################################################</span></span><br><span class="line"> <span class="comment"># -> back to create_estimator_and_inputs function</span></span><br><span class="line"> <span class="comment">####################################################################</span></span><br><span class="line"> <span class="comment"># Map methods</span></span><br><span class="line"> <span class="comment"># 'create_train_input_fn': inputs.create_train_input_fn,</span></span><br><span class="line"> create_train_input_fn = MODEL_BUILD_UTIL_MAP[<span class="string">'create_train_input_fn'</span>]</span><br><span class="line"> <span class="comment"># 'create_eval_input_fn': inputs.create_eval_input_fn,</span></span><br><span class="line"> create_eval_input_fn = MODEL_BUILD_UTIL_MAP[<span class="string">'create_eval_input_fn'</span>]</span><br><span class="line"> <span class="comment"># 'create_predict_input_fn': inputs.create_predict_input_fn,</span></span><br><span class="line"> create_predict_input_fn = MODEL_BUILD_UTIL_MAP[<span class="string">'create_predict_input_fn'</span>]</span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># assign variables & make configs</span></span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># model_builder.build is IMPORTANT, IMPORTANT, IMPORTANT</span></span><br><span class="line"> detection_model_fn = functools.partial(</span><br><span class="line"> model_builder.build, model_config=model_config)</span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># Here create_train_input_fn function returns a</span></span><br><span class="line"> <span class="comment"># tf.data.Dataset that holds (features, labels) tuple</span></span><br><span class="line"> train_input_fn = create_train_input_fn(</span><br><span class="line"> train_config=train_config,</span><br><span class="line"> train_input_config=train_input_config,</span><br><span class="line"> model_config=model_config)</span><br><span class="line"> ...</span><br><span class="line"> model_fn = model_fn_creator(detection_model_fn, configs, hparams, use_tpu)</span><br><span class="line"> ...</span><br></pre></td></tr></table></figure></p>
<h3 id="File-model-builder-py-gt-build-DetectionModel"><a href="#File-model-builder-py-gt-build-DetectionModel" class="headerlink" title="File model_builder.py -> build DetectionModel"></a>File model_builder.py -> build DetectionModel</h3><p>Start from build function<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">build</span><span class="params">(model_config, is_training, add_summaries=True,</span></span></span><br><span class="line"><span class="function"><span class="params"> add_background_class=True)</span>:</span></span><br><span class="line"> <span class="comment"># Here, we use faster_rcnn_meta_arch as our model</span></span><br><span class="line"> <span class="keyword">if</span> <span class="keyword">not</span> isinstance(model_config, model_pb2.DetectionModel):</span><br><span class="line"> <span class="keyword">raise</span> ValueError(<span class="string">'model_config not of type model_pb2.DetectionModel.'</span>)</span><br><span class="line"> meta_architecture = model_config.WhichOneof(<span class="string">'model'</span>)</span><br><span class="line"> <span class="keyword">if</span> meta_architecture == <span class="string">'ssd'</span>:</span><br><span class="line"> <span class="keyword">return</span> _build_ssd_model(model_config.ssd, is_training, add_summaries,</span><br><span class="line"> add_background_class)</span><br><span class="line"> <span class="keyword">if</span> meta_architecture == <span class="string">'faster_rcnn'</span>:</span><br><span class="line"> <span class="keyword">return</span> _build_faster_rcnn_model(model_config.faster_rcnn, is_training,</span><br><span class="line"> add_summaries)</span><br><span class="line"> <span class="keyword">raise</span> ValueError(<span class="string">'Unknown meta architecture: {}'</span>.format(meta_architecture))</span><br></pre></td></tr></table></figure></p>
<p>_build_faster_rcnn_model - <strong>Important function</strong><br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br><span class="line">47</span><br><span class="line">48</span><br><span class="line">49</span><br><span class="line">50</span><br><span class="line">51</span><br><span class="line">52</span><br><span class="line">53</span><br><span class="line">54</span><br><span class="line">55</span><br><span class="line">56</span><br><span class="line">57</span><br><span class="line">58</span><br><span class="line">59</span><br><span class="line">60</span><br><span class="line">61</span><br><span class="line">62</span><br><span class="line">63</span><br><span class="line">64</span><br><span class="line">65</span><br><span class="line">66</span><br><span class="line">67</span><br><span class="line">68</span><br><span class="line">69</span><br><span class="line">70</span><br><span class="line">71</span><br><span class="line">72</span><br><span class="line">73</span><br><span class="line">74</span><br><span class="line">75</span><br><span class="line">76</span><br><span class="line">77</span><br><span class="line">78</span><br><span class="line">79</span><br><span class="line">80</span><br><span class="line">81</span><br><span class="line">82</span><br><span class="line">83</span><br><span class="line">84</span><br><span class="line">85</span><br><span class="line">86</span><br><span class="line">87</span><br><span class="line">88</span><br><span class="line">89</span><br><span class="line">90</span><br><span class="line">91</span><br><span class="line">92</span><br><span class="line">93</span><br><span class="line">94</span><br><span class="line">95</span><br><span class="line">96</span><br><span class="line">97</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_build_faster_rcnn_model</span><span class="params">(frcnn_config, is_training, add_summaries)</span>:</span></span><br><span class="line"> num_classes = frcnn_config.num_classes</span><br><span class="line"> image_resizer_fn = image_resizer_builder.build(frcnn_config.image_resizer)</span><br><span class="line"></span><br><span class="line"> <span class="comment"># Assign a feature extraction method for detection model. e.g. ResNet101</span></span><br><span class="line"> <span class="comment"># Resnet: nets.resnet_v1</span></span><br><span class="line"> feature_extractor = _build_faster_rcnn_feature_extractor(</span><br><span class="line"> frcnn_config.feature_extractor, is_training,</span><br><span class="line"> frcnn_config.inplace_batchnorm_update)</span><br><span class="line"></span><br><span class="line"> number_of_stages = frcnn_config.number_of_stages</span><br><span class="line"> first_stage_anchor_generator = anchor_generator_builder.build(</span><br><span class="line"> frcnn_config.first_stage_anchor_generator)</span><br><span class="line"></span><br><span class="line"> <span class="comment"># Several hyperparameters for first stage </span></span><br><span class="line"> first_stage_target_assigner = target_assigner.create_target_assigner(</span><br><span class="line"> <span class="string">'FasterRCNN'</span>,</span><br><span class="line"> <span class="string">'proposal'</span>,</span><br><span class="line"> use_matmul_gather=frcnn_config.use_matmul_gather_in_matcher)</span><br><span class="line"> first_stage_atrous_rate = frcnn_config.first_stage_atrous_rate</span><br><span class="line"> first_stage_box_predictor_arg_scope_fn = hyperparams_builder.build(</span><br><span class="line"> frcnn_config.first_stage_box_predictor_conv_hyperparams, is_training)</span><br><span class="line"> first_stage_box_predictor_kernel_size = (</span><br><span class="line"> frcnn_config.first_stage_box_predictor_kernel_size)</span><br><span class="line"> first_stage_box_predictor_depth = frcnn_config.first_stage_box_predictor_depth</span><br><span class="line"> first_stage_minibatch_size = frcnn_config.first_stage_minibatch_size</span><br><span class="line"> first_stage_sampler = sampler.BalancedPositiveNegativeSampler(</span><br><span class="line"> positive_fraction=frcnn_config.first_stage_positive_balance_fraction,</span><br><span class="line"> is_static=frcnn_config.use_static_balanced_label_sampler)</span><br><span class="line"> first_stage_nms_score_threshold = frcnn_config.first_stage_nms_score_threshold</span><br><span class="line"> first_stage_nms_iou_threshold = frcnn_config.first_stage_nms_iou_threshold</span><br><span class="line"> first_stage_max_proposals = frcnn_config.first_stage_max_proposals</span><br><span class="line"> first_stage_loc_loss_weight = (</span><br><span class="line"> frcnn_config.first_stage_localization_loss_weight)</span><br><span class="line"> first_stage_obj_loss_weight = frcnn_config.first_stage_objectness_loss_weight</span><br><span class="line"></span><br><span class="line"> initial_crop_size = frcnn_config.initial_crop_size</span><br><span class="line"> maxpool_kernel_size = frcnn_config.maxpool_kernel_size</span><br><span class="line"> maxpool_stride = frcnn_config.maxpool_stride</span><br><span class="line"></span><br><span class="line"> <span class="comment"># Several hyperparameters for second stage</span></span><br><span class="line"> second_stage_target_assigner = target_assigner.create_target_assigner(</span><br><span class="line"> <span class="string">'FasterRCNN'</span>,</span><br><span class="line"> <span class="string">'detection'</span>,</span><br><span class="line"> use_matmul_gather=frcnn_config.use_matmul_gather_in_matcher)</span><br><span class="line"> second_stage_box_predictor = box_predictor_builder.build(</span><br><span class="line"> hyperparams_builder.build,</span><br><span class="line"> frcnn_config.second_stage_box_predictor,</span><br><span class="line"> is_training=is_training,</span><br><span class="line"> num_classes=num_classes)</span><br><span class="line"> second_stage_batch_size = frcnn_config.second_stage_batch_size</span><br><span class="line"> second_stage_sampler = sampler.BalancedPositiveNegativeSampler(</span><br><span class="line"> positive_fraction=frcnn_config.second_stage_balance_fraction,</span><br><span class="line"> is_static=frcnn_config.use_static_balanced_label_sampler)</span><br><span class="line"> <span class="comment"># NMS for second stage is different API from first stage</span></span><br><span class="line"> <span class="comment"># But they point to the same function. post_processing.batch_multiclass_non_max_suppression</span></span><br><span class="line"> (second_stage_non_max_suppression_fn, second_stage_score_conversion_fn</span><br><span class="line"> ) = post_processing_builder.build(frcnn_config.second_stage_post_processing)</span><br><span class="line"> second_stage_localization_loss_weight = (</span><br><span class="line"> frcnn_config.second_stage_localization_loss_weight)</span><br><span class="line"> second_stage_classification_loss = (</span><br><span class="line"> losses_builder.build_faster_rcnn_classification_loss(</span><br><span class="line"> frcnn_config.second_stage_classification_loss))</span><br><span class="line"> second_stage_classification_loss_weight = (</span><br><span class="line"> frcnn_config.second_stage_classification_loss_weight)</span><br><span class="line"> second_stage_mask_prediction_loss_weight = (</span><br><span class="line"> frcnn_config.second_stage_mask_prediction_loss_weight)</span><br><span class="line"></span><br><span class="line"> hard_example_miner = <span class="keyword">None</span></span><br><span class="line"> <span class="keyword">if</span> frcnn_config.HasField(<span class="string">'hard_example_miner'</span>):</span><br><span class="line"> hard_example_miner = losses_builder.build_hard_example_miner(</span><br><span class="line"> frcnn_config.hard_example_miner,</span><br><span class="line"> second_stage_classification_loss_weight,</span><br><span class="line"> second_stage_localization_loss_weight)</span><br><span class="line"></span><br><span class="line"> use_matmul_crop_and_resize = (frcnn_config.use_matmul_crop_and_resize)</span><br><span class="line"> clip_anchors_to_image = (</span><br><span class="line"> frcnn_config.clip_anchors_to_image)</span><br><span class="line"></span><br><span class="line"> common_kwargs = {</span><br><span class="line"> ...</span><br><span class="line"> }</span><br><span class="line"></span><br><span class="line"> <span class="keyword">if</span> isinstance(second_stage_box_predictor,</span><br><span class="line"> rfcn_box_predictor.RfcnBoxPredictor):</span><br><span class="line"> <span class="keyword">return</span> rfcn_meta_arch.RFCNMetaArch(</span><br><span class="line"> second_stage_rfcn_box_predictor=second_stage_box_predictor,</span><br><span class="line"> **common_kwargs)</span><br><span class="line"> <span class="keyword">else</span>:</span><br><span class="line"> <span class="keyword">return</span> faster_rcnn_meta_arch.FasterRCNNMetaArch(</span><br><span class="line"> initial_crop_size=initial_crop_size,</span><br><span class="line"> maxpool_kernel_size=maxpool_kernel_size,</span><br><span class="line"> maxpool_stride=maxpool_stride,</span><br><span class="line"> second_stage_mask_rcnn_box_predictor=second_stage_box_predictor,</span><br><span class="line"> second_stage_mask_prediction_loss_weight=(</span><br><span class="line"> second_stage_mask_prediction_loss_weight),</span><br><span class="line"> **common_kwargs)</span><br></pre></td></tr></table></figure></p>
<p>The faster_rcnn_meta_arch is returned by this function</p>
<h3 id="Back-to-model-main-py"><a href="#Back-to-model-main-py" class="headerlink" title="Back to model_main.py"></a>Back to model_main.py</h3><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br></pre></td><td class="code"><pre><span class="line">...</span><br><span class="line"><span class="keyword">else</span>:</span><br><span class="line"> train_spec, eval_specs = model_lib.create_train_and_eval_specs(</span><br><span class="line"> train_input_fn,</span><br><span class="line"> eval_input_fn,</span><br><span class="line"> eval_on_train_input_fn,</span><br><span class="line"> predict_input_fn,</span><br><span class="line"> train_steps,</span><br><span class="line"> eval_steps,</span><br><span class="line"> eval_on_train_data=<span class="keyword">False</span>)</span><br><span class="line"><span class="comment">#############################################################</span></span><br><span class="line"><span class="comment"># -> jump to model.lib.create_train_and_eval_specs</span></span><br><span class="line"><span class="comment">#############################################################</span></span><br><span class="line">...</span><br><span class="line">train_spec = tf.estimator.TrainSpec(</span><br><span class="line"> input_fn=train_input_fn, max_steps=train_steps)</span><br><span class="line"></span><br><span class="line">eval_specs = [</span><br><span class="line"> tf.estimator.EvalSpec(</span><br><span class="line"> name=eval_spec_name,</span><br><span class="line"> input_fn=eval_input_fn,</span><br><span class="line"> steps=eval_steps,</span><br><span class="line"> exporters=exporter)</span><br><span class="line">]</span><br><span class="line">...</span><br></pre></td></tr></table></figure>
</div>
<footer class="post-footer">
<div class="post-eof"></div>
</footer>
</div>
</article>
<article class="post post-type-normal" itemscope itemtype="http://schema.org/Article">
<div class="post-block">
<link itemprop="mainEntityOfPage" href="http://chenghaz.github.io/2018/09/06/tf-object-detection-1/">
<span hidden itemprop="author" itemscope itemtype="http://schema.org/Person">
<meta itemprop="name" content="chenghaz">
<meta itemprop="description" content="">
<meta itemprop="image" content="/uploads/avatar.png">
</span>
<span hidden itemprop="publisher" itemscope itemtype="http://schema.org/Organization">
<meta itemprop="name" content="chenghaz Official 1.0">
</span>
<header class="post-header">
<h1 class="post-title" itemprop="name headline">
<a class="post-title-link" href="/2018/09/06/tf-object-detection-1/" itemprop="url">Reading note for tensorflow/models/research/object_detection (part1)</a></h1>
<div class="post-meta">
<span class="post-time">
<span class="post-meta-item-icon">
<i class="fa fa-calendar-o"></i>
</span>
<span class="post-meta-item-text">Posted on</span>
<time title="Post created" itemprop="dateCreated datePublished" datetime="2018-09-06T18:38:46-07:00">
2018-09-06
</time>
</span>
</div>
</header>
<div class="post-body" itemprop="articleBody">
<h3 id="Start-from-Faster-RCNN"><a href="#Start-from-Faster-RCNN" class="headerlink" title="Start from Faster-RCNN"></a>Start from Faster-RCNN</h3><h4 id="File-meta-architectures-faster-rcnn-meta-arch-py"><a href="#File-meta-architectures-faster-rcnn-meta-arch-py" class="headerlink" title="File: meta_architectures/faster_rcnn_meta_arch.py"></a>File: meta_architectures/faster_rcnn_meta_arch.py</h4><p>For FasterRCNNMetaArch class, first Step is fully understand all Args meaning<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br></pre></td><td class="code"><pre><span class="line"><span class="class"><span class="keyword">class</span> <span class="title">FasterRCNNMetaArch</span><span class="params">(model.DetectionModel)</span></span></span><br></pre></td></tr></table></figure></p>
<p>init: The init function is to assign a lot of hyperparameters or functions<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">__init__</span><span class="params">(...)</span>:</span></span><br><span class="line"> self._first_stage_box_predictor =</span><br><span class="line"> box_predictor_builder.build_convolutional_box_predictor</span><br><span class="line"> self._second_stage_nms_fn = second_stage_non_max_suppression_fn</span><br></pre></td></tr></table></figure></p>
<p>preprocess: Notice that the base function of preprocess function is at “core/model.py”. The main purpose of this function is to do input preprocessing: resizing and scaling<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">preprocess</span><span class="params">(self, inputs)</span>:</span></span><br><span class="line"> <span class="keyword">if</span> inputs.dtype <span class="keyword">is</span> <span class="keyword">not</span> tf.float32:</span><br><span class="line"> <span class="keyword">raise</span> ValueError(<span class="string">'`preprocess` expects a tf.float32 tensor'</span>)</span><br><span class="line"> <span class="keyword">with</span> tf.name_scope(<span class="string">'Preprocessor'</span>):</span><br><span class="line"> outputs = shape_utils.static_or_dynamic_map_fn(</span><br><span class="line"> self._image_resizer_fn,</span><br><span class="line"> elems=inputs,</span><br><span class="line"> dtype=[tf.float32, tf.int32],</span><br><span class="line"> parallel_iterations=self._parallel_iterations)</span><br><span class="line"> resized_inputs = outputs[<span class="number">0</span>]</span><br><span class="line"> true_image_shapes = outputs[<span class="number">1</span>]</span><br><span class="line"> <span class="keyword">return</span> (self._feature_extractor.preprocess(resized_inputs),</span><br><span class="line"> true_image_shapes)</span><br></pre></td></tr></table></figure></p>
<p>predict: Predicts unpostprocessed tensors from input tensor. From the comment, the authors mentioned “to yield “raw” un-postprocessed predictions. e.g. RPN predictions”<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">predict</span><span class="params">(self, preprocessed_inputs, true_image_shapes)</span>:</span></span><br><span class="line"> <span class="comment"># --- The explanation of these two functions is below ---</span></span><br><span class="line"> (rpn_box_predictor_features, rpn_features_to_crop, anchors_boxlist,</span><br><span class="line"> image_shape) = self._extract_rpn_feature_maps(preprocessed_inputs)</span><br><span class="line"> (rpn_box_encodings, rpn_objectness_predictions_with_background</span><br><span class="line"> ) = self._predict_rpn_proposals(rpn_box_predictor_features) </span><br><span class="line"> <span class="comment"># The Faster R-CNN paper recommends pruning anchors that venture outside</span></span><br><span class="line"> <span class="comment"># the image window at training time and clipping at inference time.</span></span><br><span class="line"> clip_window = tf.to_float(tf.stack([<span class="number">0</span>, <span class="number">0</span>, image_shape[<span class="number">1</span>], image_shape[<span class="number">2</span>]]))</span><br><span class="line"> <span class="keyword">if</span> self._is_training:</span><br><span class="line"> ...</span><br><span class="line"> ...</span><br><span class="line"> ...</span><br><span class="line"> <span class="keyword">if</span> self._number_of_stages >= <span class="number">2</span>:</span><br><span class="line"> <span class="comment"># --- The explanation of _predict_second_stage function is below ---</span></span><br><span class="line"> prediction_dict.update(self._predict_second_stage(</span><br><span class="line"> rpn_box_encodings,</span><br><span class="line"> rpn_objectness_predictions_with_background,</span><br><span class="line"> rpn_features_to_crop,</span><br><span class="line"> self._anchors.get(), image_shape, true_image_shapes))</span><br><span class="line"></span><br><span class="line"> <span class="keyword">if</span> self._number_of_stages == <span class="number">3</span>:</span><br><span class="line"> <span class="comment"># --- _predict_third_stage predicts non-box non-class outputs ---</span></span><br><span class="line"> <span class="comment"># e.g. mask prediction</span></span><br><span class="line"> prediction_dict = self._predict_third_stage(</span><br><span class="line"> prediction_dict, true_image_shapes)</span><br><span class="line"></span><br><span class="line"> <span class="keyword">return</span> ...</span><br></pre></td></tr></table></figure></p>
<p>_extract_rpn_feature_maps: Just extract RPN features. Notice that we have two features maps here: <em>rpn_box_predictor_features and rpn_features_to_crop</em>.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_extract_rpn_feature_maps</span><span class="params">(self, preprocessed_inputs)</span>:</span></span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># rpn_features_to_crop is like the feature before RPN</span></span><br><span class="line"> rpn_features_to_crop, _ = self._feature_extractor.extract_proposal_features(</span><br><span class="line"> preprocessed_inputs, scope=self.first_stage_feature_extractor_scope)</span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># Then feed rpn_features_to_crop to conv2d to get rpn_box_predictor_features</span></span><br><span class="line"> <span class="keyword">with</span> slim.arg_scope(self._first_stage_box_predictor_arg_scope_fn()):</span><br><span class="line"> kernel_size = self._first_stage_box_predictor_kernel_size</span><br><span class="line"> rpn_box_predictor_features = slim.conv2d(</span><br><span class="line"> rpn_features_to_crop,</span><br><span class="line"> self._first_stage_box_predictor_depth,</span><br><span class="line"> kernel_size=[kernel_size, kernel_size],</span><br><span class="line"> rate=self._first_stage_atrous_rate,</span><br><span class="line"> activation_fn=tf.nn.relu6)</span><br><span class="line"> ...</span><br></pre></td></tr></table></figure></p>
<p>_predict_rpn_proposals: Box predictors after RPN feature map to predict proposals. The input here is <em>rpn_box_predictor_features</em> and output is <em>box_encoding: 3-D tensor of float - predict boxes</em> and <em>object predictions: 3-D tensors - class prediction</em><br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_predict_rpn_proposals</span><span class="params">(self, rpn_box_predictor_features)</span>:</span></span><br><span class="line"> num_anchors_per_location = (</span><br><span class="line"> self._first_stage_anchor_generator.num_anchors_per_location())</span><br><span class="line"> <span class="keyword">if</span> len(num_anchors_per_location) != <span class="number">1</span>:</span><br><span class="line"> <span class="keyword">raise</span> RuntimeError(<span class="string">'anchor_generator is expected to generate anchors '</span></span><br><span class="line"> <span class="string">'corresponding to a single feature map.'</span>)</span><br><span class="line"> <span class="keyword">if</span> self._first_stage_box_predictor.is_keras_model:</span><br><span class="line"> box_predictions = self._first_stage_box_predictor(</span><br><span class="line"> [rpn_box_predictor_features])</span><br><span class="line"> <span class="keyword">else</span>:</span><br><span class="line"> <span class="comment"># The _first_stage_box_predictor is assigned at __init__</span></span><br><span class="line"> box_predictions = self._first_stage_box_predictor.predict(</span><br><span class="line"> [rpn_box_predictor_features],</span><br><span class="line"> num_anchors_per_location,</span><br><span class="line"> scope=self.first_stage_box_predictor_scope)</span><br><span class="line"> <span class="comment"># box infomation and class information</span></span><br><span class="line"> box_encodings = tf.concat(</span><br><span class="line"> box_predictions[box_predictor.BOX_ENCODINGS], axis=<span class="number">1</span>)</span><br><span class="line"> objectness_predictions_with_background = tf.concat(</span><br><span class="line"> box_predictions[box_predictor.CLASS_PREDICTIONS_WITH_BACKGROUND],</span><br><span class="line"> axis=<span class="number">1</span>)</span><br><span class="line"> <span class="keyword">return</span> (tf.squeeze(box_encodings, axis=<span class="number">2</span>),</span><br><span class="line"> objectness_predictions_with_background)</span><br></pre></td></tr></table></figure></p>
<p>_predict_second_stage: Predicts the output tensor from second stage of Faster R-CNN. This function is similar to function <em>predict</em>.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br><span class="line">44</span><br><span class="line">45</span><br><span class="line">46</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_predict_second_stage</span><span class="params">(self, rpn_box_encodings,</span></span></span><br><span class="line"><span class="function"><span class="params"> rpn_objectness_predictions_with_background,</span></span></span><br><span class="line"><span class="function"><span class="params"> rpn_features_to_crop,</span></span></span><br><span class="line"><span class="function"><span class="params"> anchors,</span></span></span><br><span class="line"><span class="function"><span class="params"> image_shape,</span></span></span><br><span class="line"><span class="function"><span class="params"> true_image_shapes)</span>:</span></span><br><span class="line"> image_shape_2d = self._image_batch_shape_2d(image_shape)</span><br><span class="line"> <span class="comment"># --- _postprocess_rpn contains first_stage_nms ---</span></span><br><span class="line"> <span class="comment"># --- explanation below ---</span></span><br><span class="line"> proposal_boxes_normalized, _, num_proposals = self._postprocess_rpn(</span><br><span class="line"> rpn_box_encodings, rpn_objectness_predictions_with_background,</span><br><span class="line"> anchors, image_shape_2d, true_image_shapes)</span><br><span class="line"> <span class="comment"># Resize the first stage RPN feature then maxpool</span></span><br><span class="line"> <span class="comment"># RoI pooling function</span></span><br><span class="line"> flattened_proposal_feature_maps = (</span><br><span class="line"> self._compute_second_stage_input_feature_maps(</span><br><span class="line"> rpn_features_to_crop, proposal_boxes_normalized))</span><br><span class="line"> <span class="comment"># Extract second stage box classifier feature</span></span><br><span class="line"> <span class="comment"># Similar to _extract_rpn_feature_maps</span></span><br><span class="line"> box_classifier_features = (</span><br><span class="line"> self._feature_extractor.extract_box_classifier_features(</span><br><span class="line"> flattened_proposal_feature_maps,</span><br><span class="line"> scope=self.second_stage_feature_extractor_scope))</span><br><span class="line"> <span class="comment"># Second Stage box predictor</span></span><br><span class="line"> <span class="keyword">if</span> self._mask_rcnn_box_predictor.is_keras_model:</span><br><span class="line"> box_predictions = self._mask_rcnn_box_predictor(</span><br><span class="line"> [box_classifier_features],</span><br><span class="line"> prediction_stage=<span class="number">2</span>)</span><br><span class="line"> <span class="keyword">else</span>:</span><br><span class="line"> <span class="comment"># ._mask_rcnn_box_predictor is second_stage_mask_rcnn_box_predictor</span></span><br><span class="line"> <span class="comment"># at __init__. Then, from faster_rcnn_meta_arch_test_lib</span></span><br><span class="line"> <span class="comment"># second_stage_mask_rcnn_box_predictor is box_predictor *core/box_predictor.py*</span></span><br><span class="line"> box_predictions = self._mask_rcnn_box_predictor.predict(</span><br><span class="line"> [box_classifier_features],</span><br><span class="line"> num_predictions_per_location=[<span class="number">1</span>],</span><br><span class="line"> scope=self.second_stage_box_predictor_scope,</span><br><span class="line"> prediction_stage=<span class="number">2</span>)</span><br><span class="line"> <span class="comment"># box information and class information</span></span><br><span class="line"> refined_box_encodings = tf.squeeze(</span><br><span class="line"> box_predictions[box_predictor.BOX_ENCODINGS],</span><br><span class="line"> axis=<span class="number">1</span>, name=<span class="string">'all_refined_box_encodings'</span>)</span><br><span class="line"> class_predictions_with_background = tf.squeeze(</span><br><span class="line"> box_predictions[box_predictor.CLASS_PREDICTIONS_WITH_BACKGROUND],</span><br><span class="line"> axis=<span class="number">1</span>, name=<span class="string">'all_class_predictions_with_background'</span>)</span><br><span class="line"> ...</span><br><span class="line"> <span class="keyword">return</span> ...</span><br></pre></td></tr></table></figure></p>
<p>_postprocess_rpn: convert RPN to proposals, perform non max suppression(NMS). This function contains many tensorflow shape API. Here, I just select the functional part.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_postprocess_rpn</span><span class="params">(self,</span></span></span><br><span class="line"><span class="function"><span class="params"> rpn_box_encodings_batch,</span></span></span><br><span class="line"><span class="function"><span class="params"> rpn_objectness_predictions_with_background_batch,</span></span></span><br><span class="line"><span class="function"><span class="params"> anchors,</span></span></span><br><span class="line"><span class="function"><span class="params"> image_shapes,</span></span></span><br><span class="line"><span class="function"><span class="params"> true_image_shapes)</span>:</span></span><br><span class="line"> <span class="comment"># tf.expand_dims tf.tile tf.squeeze tf.nn.softmax ...</span></span><br><span class="line"> ...</span><br><span class="line"> (proposal_boxes, proposal_scores, _, _, _,</span><br><span class="line"> <span class="comment"># The NMS here is post_processing.batch_multiclass_non_max_suppression</span></span><br><span class="line"> num_proposals) = post_processing.batch_multiclass_non_max_suppression(</span><br><span class="line"> tf.expand_dims(proposal_boxes, axis=<span class="number">2</span>),</span><br><span class="line"> tf.expand_dims(rpn_objectness_softmax_without_background,</span><br><span class="line"> axis=<span class="number">2</span>),</span><br><span class="line"> self._first_stage_nms_score_threshold,</span><br><span class="line"> self._first_stage_nms_iou_threshold,</span><br><span class="line"> self._first_stage_max_proposals,</span><br><span class="line"> self._first_stage_max_proposals,</span><br><span class="line"> clip_window=clip_window)</span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># normalize proposal boxes</span></span><br><span class="line"> <span class="function"><span class="keyword">def</span> <span class="title">normalize_boxes</span><span class="params">(args)</span>:</span></span><br><span class="line"> ...</span><br><span class="line"> ...</span><br></pre></td></tr></table></figure></p>
<p>Now, we jump into postprocess function</p>
<p>postprocess: Convert prediction tensors to final detections. Input for this function is the output for predict function.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br><span class="line">27</span><br><span class="line">28</span><br><span class="line">29</span><br><span class="line">30</span><br><span class="line">31</span><br><span class="line">32</span><br><span class="line">33</span><br><span class="line">34</span><br><span class="line">35</span><br><span class="line">36</span><br><span class="line">37</span><br><span class="line">38</span><br><span class="line">39</span><br><span class="line">40</span><br><span class="line">41</span><br><span class="line">42</span><br><span class="line">43</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">postprocess</span><span class="params">(self, prediction_dict, true_image_shapes)</span>:</span></span><br><span class="line"> <span class="comment"># FirstStage -> _postprocess_rpn -> explained before</span></span><br><span class="line"> <span class="comment"># SecondStage -> _postprocess_box_classifier -> explained below</span></span><br><span class="line"> <span class="comment"># ThirdStage -> already postprocessed, just simple trainsfer</span></span><br><span class="line"> <span class="keyword">with</span> tf.name_scope(<span class="string">'FirstStagePostprocessor'</span>):</span><br><span class="line"> <span class="keyword">if</span> self._number_of_stages == <span class="number">1</span>:</span><br><span class="line"> proposal_boxes, proposal_scores, num_proposals = self._postprocess_rpn(</span><br><span class="line"> prediction_dict[<span class="string">'rpn_box_encodings'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'rpn_objectness_predictions_with_background'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'anchors'</span>],</span><br><span class="line"> true_image_shapes,</span><br><span class="line"> true_image_shapes)</span><br><span class="line"> <span class="keyword">return</span> {</span><br><span class="line"> fields.DetectionResultFields.detection_boxes: proposal_boxes,</span><br><span class="line"> fields.DetectionResultFields.detection_scores: proposal_scores,</span><br><span class="line"> fields.DetectionResultFields.num_detections:</span><br><span class="line"> tf.to_float(num_proposals),</span><br><span class="line"> }</span><br><span class="line"></span><br><span class="line"> <span class="comment"># TODO(jrru): Remove mask_predictions from _post_process_box_classifier.</span></span><br><span class="line"> <span class="keyword">with</span> tf.name_scope(<span class="string">'SecondStagePostprocessor'</span>):</span><br><span class="line"> <span class="keyword">if</span> (self._number_of_stages == <span class="number">2</span> <span class="keyword">or</span></span><br><span class="line"> (self._number_of_stages == <span class="number">3</span> <span class="keyword">and</span> self._is_training)):</span><br><span class="line"> mask_predictions = prediction_dict.get(box_predictor.MASK_PREDICTIONS)</span><br><span class="line"> detections_dict = self._postprocess_box_classifier(</span><br><span class="line"> prediction_dict[<span class="string">'refined_box_encodings'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'class_predictions_with_background'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'proposal_boxes'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'num_proposals'</span>],</span><br><span class="line"> true_image_shapes,</span><br><span class="line"> mask_predictions=mask_predictions)</span><br><span class="line"> <span class="keyword">return</span> detections_dict</span><br><span class="line"></span><br><span class="line"> <span class="keyword">if</span> self._number_of_stages == <span class="number">3</span>:</span><br><span class="line"> <span class="comment"># Post processing is already performed in 3rd stage. We need to transfer</span></span><br><span class="line"> <span class="comment"># postprocessed tensors from `prediction_dict` to `detections_dict`.</span></span><br><span class="line"> detections_dict = {}</span><br><span class="line"> <span class="keyword">for</span> key <span class="keyword">in</span> prediction_dict:</span><br><span class="line"> <span class="keyword">if</span> key == fields.DetectionResultFields.detection_masks:</span><br><span class="line"> detections_dict[key] = tf.sigmoid(prediction_dict[key])</span><br><span class="line"> <span class="keyword">elif</span> <span class="string">'detection'</span> <span class="keyword">in</span> key:</span><br><span class="line"> detections_dict[key] = prediction_dict[key]</span><br><span class="line"> <span class="keyword">return</span> detections_dict</span><br></pre></td></tr></table></figure></p>
<p>_postprocess_box_classifier: Convert second stage box classifier to result. Similar to first stage operations.<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">_postprocess_box_classifier</span><span class="params">(self,</span></span></span><br><span class="line"><span class="function"><span class="params"> refined_box_encodings,</span></span></span><br><span class="line"><span class="function"><span class="params"> class_predictions_with_background,</span></span></span><br><span class="line"><span class="function"><span class="params"> proposal_boxes,</span></span></span><br><span class="line"><span class="function"><span class="params"> num_proposals,</span></span></span><br><span class="line"><span class="function"><span class="params"> image_shapes,</span></span></span><br><span class="line"><span class="function"><span class="params"> mask_predictions=None)</span>:</span></span><br><span class="line"> ...</span><br><span class="line"> <span class="comment"># _second_stage_nms_fn here is initialized at __init__</span></span><br><span class="line"> (nmsed_boxes, nmsed_scores, nmsed_classes, nmsed_masks, _,</span><br><span class="line"> num_detections) = self._second_stage_nms_fn(</span><br><span class="line"> refined_decoded_boxes_batch,</span><br><span class="line"> class_predictions_batch,</span><br><span class="line"> clip_window=clip_window,</span><br><span class="line"> change_coordinate_frame=<span class="keyword">True</span>,</span><br><span class="line"> num_valid_boxes=num_proposals,</span><br><span class="line"> masks=mask_predictions_batch)</span><br><span class="line"> ...</span><br></pre></td></tr></table></figure></p>
<p>Final part: Loss</p>
<p>loss: depends on stage number. First stage loss: self._loss_rpn, After first stage: self._loss_box_classifier<br><figure class="highlight python"><table><tr><td class="gutter"><pre><span class="line">1</span><br><span class="line">2</span><br><span class="line">3</span><br><span class="line">4</span><br><span class="line">5</span><br><span class="line">6</span><br><span class="line">7</span><br><span class="line">8</span><br><span class="line">9</span><br><span class="line">10</span><br><span class="line">11</span><br><span class="line">12</span><br><span class="line">13</span><br><span class="line">14</span><br><span class="line">15</span><br><span class="line">16</span><br><span class="line">17</span><br><span class="line">18</span><br><span class="line">19</span><br><span class="line">20</span><br><span class="line">21</span><br><span class="line">22</span><br><span class="line">23</span><br><span class="line">24</span><br><span class="line">25</span><br><span class="line">26</span><br></pre></td><td class="code"><pre><span class="line"><span class="function"><span class="keyword">def</span> <span class="title">loss</span><span class="params">(self, prediction_dict, true_image_shapes, scope=None)</span>:</span></span><br><span class="line"> <span class="keyword">with</span> tf.name_scope(scope, <span class="string">'Loss'</span>, prediction_dict.values()):</span><br><span class="line"> (groundtruth_boxlists, groundtruth_classes_with_background_list,</span><br><span class="line"> groundtruth_masks_list, groundtruth_weights_list</span><br><span class="line"> ) = self._format_groundtruth_data(true_image_shapes)</span><br><span class="line"> loss_dict = self._loss_rpn(</span><br><span class="line"> prediction_dict[<span class="string">'rpn_box_encodings'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'rpn_objectness_predictions_with_background'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'anchors'</span>], groundtruth_boxlists,</span><br><span class="line"> groundtruth_classes_with_background_list, groundtruth_weights_list)</span><br><span class="line"> <span class="comment"># For stage 2 or 3</span></span><br><span class="line"> <span class="keyword">if</span> self._number_of_stages > <span class="number">1</span>:</span><br><span class="line"> loss_dict.update(</span><br><span class="line"> self._loss_box_classifier(</span><br><span class="line"> prediction_dict[<span class="string">'refined_box_encodings'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'class_predictions_with_background'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'proposal_boxes'</span>],</span><br><span class="line"> prediction_dict[<span class="string">'num_proposals'</span>],</span><br><span class="line"> groundtruth_boxlists,</span><br><span class="line"> groundtruth_classes_with_background_list,</span><br><span class="line"> groundtruth_weights_list,</span><br><span class="line"> prediction_dict[<span class="string">'image_shape'</span>],</span><br><span class="line"> prediction_dict.get(<span class="string">'mask_predictions'</span>),</span><br><span class="line"> groundtruth_masks_list,</span><br><span class="line"> ))</span><br><span class="line"> <span class="keyword">return</span> loss_dict</span><br></pre></td></tr></table></figure></p>
</div>
<footer class="post-footer">