This repository has been archived by the owner on May 3, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 4
/
index.html
746 lines (662 loc) · 34.4 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
<?xml version="1.0" encoding="utf-8"?>
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
<html xmlns="http://www.w3.org/1999/xhtml" lang="en" xml:lang="en">
<head>
<!-- 2018-08-27 Mon 15:50 -->
<meta http-equiv="Content-Type" content="text/html;charset=utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>Reimplementing a deep network architecture for single-image rain removal</title>
<meta name="generator" content="Org mode" />
<meta name="author" content="Jonathan Jin" />
<style type="text/css">
<!--/*--><![CDATA[/*><!--*/
.title { text-align: center;
margin-bottom: .2em; }
.subtitle { text-align: center;
font-size: medium;
font-weight: bold;
margin-top:0; }
.todo { font-family: monospace; color: red; }
.done { font-family: monospace; color: green; }
.priority { font-family: monospace; color: orange; }
.tag { background-color: #eee; font-family: monospace;
padding: 2px; font-size: 80%; font-weight: normal; }
.timestamp { color: #bebebe; }
.timestamp-kwd { color: #5f9ea0; }
.org-right { margin-left: auto; margin-right: 0px; text-align: right; }
.org-left { margin-left: 0px; margin-right: auto; text-align: left; }
.org-center { margin-left: auto; margin-right: auto; text-align: center; }
.underline { text-decoration: underline; }
#postamble p, #preamble p { font-size: 90%; margin: .2em; }
p.verse { margin-left: 3%; }
pre {
border: 1px solid #ccc;
box-shadow: 3px 3px 3px #eee;
padding: 8pt;
font-family: monospace;
overflow: auto;
margin: 1.2em;
}
pre.src {
position: relative;
overflow: visible;
padding-top: 1.2em;
}
pre.src:before {
display: none;
position: absolute;
background-color: white;
top: -10px;
right: 10px;
padding: 3px;
border: 1px solid black;
}
pre.src:hover:before { display: inline;}
/* Languages per Org manual */
pre.src-asymptote:before { content: 'Asymptote'; }
pre.src-awk:before { content: 'Awk'; }
pre.src-C:before { content: 'C'; }
/* pre.src-C++ doesn't work in CSS */
pre.src-clojure:before { content: 'Clojure'; }
pre.src-css:before { content: 'CSS'; }
pre.src-D:before { content: 'D'; }
pre.src-ditaa:before { content: 'ditaa'; }
pre.src-dot:before { content: 'Graphviz'; }
pre.src-calc:before { content: 'Emacs Calc'; }
pre.src-emacs-lisp:before { content: 'Emacs Lisp'; }
pre.src-fortran:before { content: 'Fortran'; }
pre.src-gnuplot:before { content: 'gnuplot'; }
pre.src-haskell:before { content: 'Haskell'; }
pre.src-hledger:before { content: 'hledger'; }
pre.src-java:before { content: 'Java'; }
pre.src-js:before { content: 'Javascript'; }
pre.src-latex:before { content: 'LaTeX'; }
pre.src-ledger:before { content: 'Ledger'; }
pre.src-lisp:before { content: 'Lisp'; }
pre.src-lilypond:before { content: 'Lilypond'; }
pre.src-lua:before { content: 'Lua'; }
pre.src-matlab:before { content: 'MATLAB'; }
pre.src-mscgen:before { content: 'Mscgen'; }
pre.src-ocaml:before { content: 'Objective Caml'; }
pre.src-octave:before { content: 'Octave'; }
pre.src-org:before { content: 'Org mode'; }
pre.src-oz:before { content: 'OZ'; }
pre.src-plantuml:before { content: 'Plantuml'; }
pre.src-processing:before { content: 'Processing.js'; }
pre.src-python:before { content: 'Python'; }
pre.src-R:before { content: 'R'; }
pre.src-ruby:before { content: 'Ruby'; }
pre.src-sass:before { content: 'Sass'; }
pre.src-scheme:before { content: 'Scheme'; }
pre.src-screen:before { content: 'Gnu Screen'; }
pre.src-sed:before { content: 'Sed'; }
pre.src-sh:before { content: 'shell'; }
pre.src-sql:before { content: 'SQL'; }
pre.src-sqlite:before { content: 'SQLite'; }
/* additional languages in org.el's org-babel-load-languages alist */
pre.src-forth:before { content: 'Forth'; }
pre.src-io:before { content: 'IO'; }
pre.src-J:before { content: 'J'; }
pre.src-makefile:before { content: 'Makefile'; }
pre.src-maxima:before { content: 'Maxima'; }
pre.src-perl:before { content: 'Perl'; }
pre.src-picolisp:before { content: 'Pico Lisp'; }
pre.src-scala:before { content: 'Scala'; }
pre.src-shell:before { content: 'Shell Script'; }
pre.src-ebnf2ps:before { content: 'ebfn2ps'; }
/* additional language identifiers per "defun org-babel-execute"
in ob-*.el */
pre.src-cpp:before { content: 'C++'; }
pre.src-abc:before { content: 'ABC'; }
pre.src-coq:before { content: 'Coq'; }
pre.src-groovy:before { content: 'Groovy'; }
/* additional language identifiers from org-babel-shell-names in
ob-shell.el: ob-shell is the only babel language using a lambda to put
the execution function name together. */
pre.src-bash:before { content: 'bash'; }
pre.src-csh:before { content: 'csh'; }
pre.src-ash:before { content: 'ash'; }
pre.src-dash:before { content: 'dash'; }
pre.src-ksh:before { content: 'ksh'; }
pre.src-mksh:before { content: 'mksh'; }
pre.src-posh:before { content: 'posh'; }
/* Additional Emacs modes also supported by the LaTeX listings package */
pre.src-ada:before { content: 'Ada'; }
pre.src-asm:before { content: 'Assembler'; }
pre.src-caml:before { content: 'Caml'; }
pre.src-delphi:before { content: 'Delphi'; }
pre.src-html:before { content: 'HTML'; }
pre.src-idl:before { content: 'IDL'; }
pre.src-mercury:before { content: 'Mercury'; }
pre.src-metapost:before { content: 'MetaPost'; }
pre.src-modula-2:before { content: 'Modula-2'; }
pre.src-pascal:before { content: 'Pascal'; }
pre.src-ps:before { content: 'PostScript'; }
pre.src-prolog:before { content: 'Prolog'; }
pre.src-simula:before { content: 'Simula'; }
pre.src-tcl:before { content: 'tcl'; }
pre.src-tex:before { content: 'TeX'; }
pre.src-plain-tex:before { content: 'Plain TeX'; }
pre.src-verilog:before { content: 'Verilog'; }
pre.src-vhdl:before { content: 'VHDL'; }
pre.src-xml:before { content: 'XML'; }
pre.src-nxml:before { content: 'XML'; }
/* add a generic configuration mode; LaTeX export needs an additional
(add-to-list 'org-latex-listings-langs '(conf " ")) in .emacs */
pre.src-conf:before { content: 'Configuration File'; }
table { border-collapse:collapse; }
caption.t-above { caption-side: top; }
caption.t-bottom { caption-side: bottom; }
td, th { vertical-align:top; }
th.org-right { text-align: center; }
th.org-left { text-align: center; }
th.org-center { text-align: center; }
td.org-right { text-align: right; }
td.org-left { text-align: left; }
td.org-center { text-align: center; }
dt { font-weight: bold; }
.footpara { display: inline; }
.footdef { margin-bottom: 1em; }
.figure { padding: 1em; }
.figure p { text-align: center; }
.equation-container {
display: table;
text-align: center;
width: 100%;
}
.equation {
vertical-align: middle;
}
.equation-label {
display: table-cell;
text-align: right;
vertical-align: middle;
}
.inlinetask {
padding: 10px;
border: 2px solid gray;
margin: 10px;
background: #ffffcc;
}
#org-div-home-and-up
{ text-align: right; font-size: 70%; white-space: nowrap; }
textarea { overflow-x: auto; }
.linenr { font-size: smaller }
.code-highlighted { background-color: #ffff00; }
.org-info-js_info-navigation { border-style: none; }
#org-info-js_console-label
{ font-size: 10px; font-weight: bold; white-space: nowrap; }
.org-info-js_search-highlight
{ background-color: #ffff00; color: #000000; font-weight: bold; }
.org-svg { width: 90%; }
/*]]>*/-->
</style>
<link rel="stylesheet" type="text/css" href="https://www.pirilampo.org/styles/readtheorg/css/htmlize.css"/>
<link rel="stylesheet" type="text/css" href="https://www.pirilampo.org/styles/readtheorg/css/readtheorg.css"/>
<script src="https://ajax.googleapis.com/ajax/libs/jquery/2.1.3/jquery.min.js"></script>
<script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.4/js/bootstrap.min.js"></script>
<script type="text/javascript" src="https://www.pirilampo.org/styles/lib/js/jquery.stickytableheaders.min.js"></script>
<script type="text/javascript" src="https://www.pirilampo.org/styles/readtheorg/js/readtheorg.js"></script>
<script type="text/javascript">
/*
@licstart The following is the entire license notice for the
JavaScript code in this tag.
Copyright (C) 2012-2018 Free Software Foundation, Inc.
The JavaScript code in this tag is free software: you can
redistribute it and/or modify it under the terms of the GNU
General Public License (GNU GPL) as published by the Free Software
Foundation, either version 3 of the License, or (at your option)
any later version. The code is distributed WITHOUT ANY WARRANTY;
without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU GPL for more details.
As additional permission under GNU GPL version 3 section 7, you
may distribute non-source (e.g., minimized or compacted) forms of
that code without the copy of the GNU GPL normally required by
section 4, provided you include this license notice and a URL
through which recipients can access the Corresponding Source.
@licend The above is the entire license notice
for the JavaScript code in this tag.
*/
<!--/*--><![CDATA[/*><!--*/
function CodeHighlightOn(elem, id)
{
var target = document.getElementById(id);
if(null != target) {
elem.cacheClassElem = elem.className;
elem.cacheClassTarget = target.className;
target.className = "code-highlighted";
elem.className = "code-highlighted";
}
}
function CodeHighlightOff(elem, id)
{
var target = document.getElementById(id);
if(elem.cacheClassElem)
elem.className = elem.cacheClassElem;
if(elem.cacheClassTarget)
target.className = elem.cacheClassTarget;
}
/*]]>*///-->
</script>
</head>
<body>
<div id="content">
<h1 class="title">Reimplementing a deep network architecture for single-image rain removal</h1>
<div id="table-of-contents">
<h2>Table of Contents</h2>
<div id="text-table-of-contents">
<ul>
<li><a href="#orgf3e772d">1. Abstract</a></li>
<li><a href="#org1ab0537">2. <span class="todo TODO">TODO</span> <code>[1/2]</code> Development</a>
<ul>
<li><a href="#org7b06f69">2.1. <span class="done DONE">DONE</span> Data importing and preparation</a></li>
<li><a href="#org92ee195">2.2. <span class="todo TODO">TODO</span> Model Training</a>
<ul>
<li><a href="#orgfd5227e">2.2.1. Dependencies</a></li>
<li><a href="#org0f16ec8">2.2.2. Execution Environment</a></li>
<li><a href="#org05af57a">2.2.3. Defining the model</a></li>
<li><a href="#org3811aa2">2.2.4. The Rest of the Owl</a></li>
</ul>
</li>
</ul>
</li>
<li><a href="#org214f152">3. References</a></li>
<li><a href="#orge90441f">4. Appendix</a>
<ul>
<li><a href="#org2d05ee5">4.1. The Dataset</a></li>
<li><a href="#org671e1a3">4.2. The Report</a></li>
<li><a href="#orgd541356">4.3. The Site</a></li>
</ul>
</li>
</ul>
</div>
</div>
<p>
<b>Note</b>: This project is a <b>work-in-progress</b>. As such, there are bound to be some rough edges and/or missing functionality here and there. For now, I encourage readers and followers to take what’s currently here at face value and follow along for updates as the project progresses. Thanks for the interest! It means a lot to me.
</p>
<div id="outline-container-orgf3e772d" class="outline-2">
<h2 id="orgf3e772d"><span class="section-number-2">1</span> Abstract</h2>
<div class="outline-text-2" id="text-1">
<p>
We present a re-implementation of the <a href="https://arxiv.org/pdf/1609.02087v2.pdf">DerainNet method for single-image rain
removal</a> in the <a href="https://www-cs-faculty.stanford.edu/~knuth/lp.html">literate programming</a> style. Our approach here deviates slightly
from the authors’ implementation in a variety of ways, most notably of which
is the use of <a href="https://www.tensorflow.org">TensorFlow</a> for all stages in the workflow: data preparation;
model definition; model training; and so on.
</p>
<p>
For our purposes, we will use a mirror of the original authors’ dataset hosted
at <a href="https://github.com/jinnovation/rainy-image-dataset">github.com/jinnovation/rainy-image-dataset</a>.
</p>
</div>
</div>
<div id="outline-container-org1ab0537" class="outline-2">
<h2 id="org1ab0537"><span class="section-number-2">2</span> <span class="todo TODO">TODO</span> <code>[1/2]</code> Development</h2>
<div class="outline-text-2" id="text-2">
<p>
We’ll split the development process roughly into the following stages:
</p>
<ul class="org-ul">
<li>Data importing/preparation;</li>
<li>Model training.</li>
</ul>
</div>
<div id="outline-container-org7b06f69" class="outline-3">
<h3 id="org7b06f69"><span class="section-number-3">2.1</span> <span class="done DONE">DONE</span> Data importing and preparation</h3>
<div class="outline-text-3" id="text-2-1">
<p>
For starters, we’ll need to implement the gateway into our training and
evaluation data. We’ll do so using the <code>tf.data.Dataset</code> API; see <a href="https://www.tensorflow.org/programmers_guide/datasets">TensorFlow
> Develop > Programmer’s Guide > Importing Data</a> for more details.
</p>
<p>
These contents will comprise our <code>rainy_image_input</code> module.
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">import</span> glob
<span style="color: #81A1C1;">import</span> os
<span style="color: #81A1C1;">import</span> tensorflow <span style="color: #81A1C1;">as</span> tf
</pre>
</div>
<p>
The dataset directory is organized into two folders:
</p>
<ul class="org-ul">
<li><code>ground truth</code> images; and</li>
<li><code>rainy</code> images.</li>
</ul>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #D8DEE9;">GROUND_TRUTH_DIR</span> = <span style="color: #A3BE8C;">"ground truth"</span>
<span style="color: #D8DEE9;">RAINY_IMAGE_DIR</span> = <span style="color: #A3BE8C;">"rainy image"</span>
</pre>
</div>
<p>
These correspond to our model’s expected outputs, as well as the inputs,
respectively. Likewise, each “ground truth” image corresponds to one or more
“rainy” images according to some “index” value – <code>ground truth/1.jpg</code> to
<code>rainy image/1_{1,2,3}.jpg</code>, <code>ground truth/100.jpg</code> to <code>rainy
image/100_{1,2,3,4,5}</code>, and so on.
</p>
<p>
With this topology in mind, we can define a function to retrieve all index
values within the data directory:
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">_get_indices</span>(data_dir):
<span style="color: #616e88;">"""Get indices for input/output association.</span>
<span style="color: #616e88;"> Args:</span>
<span style="color: #616e88;"> data_dir: Path to the data directory.</span>
<span style="color: #616e88;"> Returns:</span>
<span style="color: #616e88;"> indices: List of numerical index values.</span>
<span style="color: #616e88;"> Raises:</span>
<span style="color: #616e88;"> ValueError: If no data_dir or no ground-truth dir.</span>
<span style="color: #616e88;"> """</span>
<span style="color: #81A1C1;">if</span> <span style="color: #81A1C1;">not</span> tf.gfile.Exists(os.path.join(data_dir, GROUND_TRUTH_DIR)):
<span style="color: #81A1C1;">raise</span> <span style="color: #8FBCBB;">ValueError</span>(<span style="color: #A3BE8C;">"Failed to find ground-truth directory."</span>)
<span style="color: #81A1C1;">return</span> [
os.path.splitext(os.path.basename(f))[0]
<span style="color: #81A1C1;">for</span> f <span style="color: #81A1C1;">in</span> glob.glob(os.path.join(data_dir, GROUND_TRUTH_DIR, <span style="color: #A3BE8C;">"*.jpg"</span>))
]
</pre>
</div>
<p>
Likewise, we can define functions to retrieve the filenames of all input and
output files corresponding to a given set of indices.
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">_get_input_files</span>(data_dir, indices):
<span style="color: #616e88;">"""Get input files from indices.</span>
<span style="color: #616e88;"> Args:</span>
<span style="color: #616e88;"> data_dir: Path to the data directory.</span>
<span style="color: #616e88;"> indices: List of numerical index values.</span>
<span style="color: #616e88;"> Returns:</span>
<span style="color: #616e88;"> Dictionary, keyed by index value, valued by string lists containing</span>
<span style="color: #616e88;"> one or more filenames.</span>
<span style="color: #616e88;"> Raises:</span>
<span style="color: #616e88;"> ValueError: If no rainy-image dir.</span>
<span style="color: #616e88;"> """</span>
<span style="color: #D8DEE9;">directory</span> = os.path.join(data_dir, RAINY_IMAGE_DIR)
<span style="color: #81A1C1;">if</span> <span style="color: #81A1C1;">not</span> tf.gfile.Exists(directory):
<span style="color: #81A1C1;">raise</span> <span style="color: #8FBCBB;">ValueError</span>(<span style="color: #A3BE8C;">"Failed to find rainy-image directory."</span>)
<span style="color: #81A1C1;">return</span> {
i: glob.glob(os.path.join(directory, <span style="color: #A3BE8C;">"{}_[0-9]*.jpg"</span>.<span style="color: #81A1C1;">format</span>(i)))
<span style="color: #81A1C1;">for</span> i <span style="color: #81A1C1;">in</span> indices
}
</pre>
</div>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">_get_output_files</span>(data_dir, indices):
<span style="color: #616e88;">"""Get output files from indices.</span>
<span style="color: #616e88;"> Args:</span>
<span style="color: #616e88;"> data_dir: Path to the data directory.</span>
<span style="color: #616e88;"> indices: List of numerical index values.</span>
<span style="color: #616e88;"> Returns:</span>
<span style="color: #616e88;"> outputs: Dictionary, keyed by index value, valued by stsring lists</span>
<span style="color: #616e88;"> containing one or more filenames.</span>
<span style="color: #616e88;"> Raises:</span>
<span style="color: #616e88;"> ValueError: If no ground-truth dir.</span>
<span style="color: #616e88;"> """</span>
<span style="color: #D8DEE9;">directory</span> = os.path.join(data_dir, GROUND_TRUTH_DIR)
<span style="color: #81A1C1;">if</span> <span style="color: #81A1C1;">not</span> tf.gfile.Exists(directory):
<span style="color: #81A1C1;">raise</span> <span style="color: #8FBCBB;">ValueError</span>(<span style="color: #A3BE8C;">"Failed to find ground-truth directory."</span>)
<span style="color: #81A1C1;">return</span> {
i: os.path.join(directory, <span style="color: #A3BE8C;">"{}.jpg"</span>.<span style="color: #81A1C1;">format</span>(i)) <span style="color: #81A1C1;">for</span> i <span style="color: #81A1C1;">in</span> indices
}
</pre>
</div>
<p>
Now we arrive at our first deviation from the author’s implementation. The
authors, in their paper, don’t seem to perform any standardized cropping on
their images, relying on the inputs to all be identical in resolution or, at
most, flipped, e.g. <code>384x512</code> and <code>512x384</code>. Here, to ease integration with
TensorFlow, we settle on a common length to truncate inputs and outputs by
along both dimensions.
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #D8DEE9;">IMAGE_SIZE</span> = 384
</pre>
</div>
<p>
Finally, we have our canonical dataset-creation interface, concluding the
first part of our project.
</p>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">dataset</span>(data_dir, indices=<span style="color: #81A1C1;">None</span>):
<span style="color: #616e88;">"""Construct dataset for rainy-image evaluation.</span>
<span style="color: #616e88;"> Args:</span>
<span style="color: #616e88;"> data_dir: Path to the data directory.</span>
<span style="color: #616e88;"> indices: The input-output pairings to return. If None (the default), uses</span>
<span style="color: #616e88;"> indices present in the data directory.</span>
<span style="color: #616e88;"> Returns:</span>
<span style="color: #616e88;"> dataset: Dataset of input-output images.</span>
<span style="color: #616e88;"> """</span>
<span style="color: #81A1C1;">if</span> <span style="color: #81A1C1;">not</span> indices:
<span style="color: #D8DEE9;">indices</span> = _get_indices(data_dir)
<span style="color: #D8DEE9;">fs_in</span> = _get_input_files(data_dir, indices)
<span style="color: #D8DEE9;">fs_out</span> = _get_output_files(data_dir, indices)
<span style="color: #D8DEE9;">ins</span> = [
fname <span style="color: #81A1C1;">for</span> k, v <span style="color: #81A1C1;">in</span> <span style="color: #81A1C1;">iter</span>(<span style="color: #81A1C1;">sorted</span>(fs_in.items()))
<span style="color: #81A1C1;">for</span> fname <span style="color: #81A1C1;">in</span> v <span style="color: #81A1C1;">if</span> k <span style="color: #81A1C1;">in</span> indices
]
<span style="color: #D8DEE9;">outs</span> = [v <span style="color: #81A1C1;">for</span> sublist <span style="color: #81A1C1;">in</span> [
[fname] * <span style="color: #81A1C1;">len</span>(fs_in[k])
<span style="color: #81A1C1;">for</span> k, fname <span style="color: #81A1C1;">in</span> <span style="color: #81A1C1;">iter</span>(<span style="color: #81A1C1;">sorted</span>(fs_out.items()))
<span style="color: #81A1C1;">if</span> k <span style="color: #81A1C1;">in</span> indices
] <span style="color: #81A1C1;">for</span> v <span style="color: #81A1C1;">in</span> sublist]
<span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">_parse_function</span>(fname_in, fname_out):
<span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">_decode_resize</span>(fname):
<span style="color: #D8DEE9;">f</span> = tf.read_file(fname)
<span style="color: #D8DEE9;">contents</span> = tf.image.decode_jpeg(f)
<span style="color: #D8DEE9;">resized</span> = tf.image.resize_image_with_crop_or_pad(
contents, IMAGE_SIZE, IMAGE_SIZE,
)
<span style="color: #D8DEE9;">casted</span> = tf.cast(resized, tf.float32)
<span style="color: #81A1C1;">return</span> casted
<span style="color: #81A1C1;">return</span> _decode_resize(fname_in), _decode_resize(fname_out)
<span style="color: #D8DEE9;">dataset</span> = tf.data.Dataset.from_tensor_slices(
(tf.constant(ins), tf.constant(outs)),
).<span style="color: #81A1C1;">map</span>(_parse_function)
<span style="color: #81A1C1;">return</span> dataset
</pre>
</div>
</div>
</div>
<div id="outline-container-org92ee195" class="outline-3">
<h3 id="org92ee195"><span class="section-number-3">2.2</span> <span class="todo TODO">TODO</span> Model Training</h3>
<div class="outline-text-3" id="text-2-2">
</div>
<div id="outline-container-orgfd5227e" class="outline-4">
<h4 id="orgfd5227e"><span class="section-number-4">2.2.1</span> Dependencies</h4>
<div class="outline-text-4" id="text-2-2-1">
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">import</span> tensorflow <span style="color: #81A1C1;">as</span> tf
<span style="color: #81A1C1;">import</span> logging
<span style="color: #81A1C1;">from</span> rainy_image_input <span style="color: #81A1C1;">import</span> dataset, IMAGE_SIZE
</pre>
</div>
</div>
</div>
<div id="outline-container-org0f16ec8" class="outline-4">
<h4 id="org0f16ec8"><span class="section-number-4">2.2.2</span> Execution Environment</h4>
<div class="outline-text-4" id="text-2-2-2">
<div class="org-src-container">
<pre class="src src-python">tf.app.flags.DEFINE_string(<span style="color: #A3BE8C;">"checkpoint_dir"</span>, <span style="color: #A3BE8C;">"/tmp/derain-checkpoint"</span>,
<span style="color: #A3BE8C;">"""Directory to write event logs and checkpointing</span>
<span style="color: #A3BE8C;"> to."""</span>)
tf.app.flags.DEFINE_string(<span style="color: #A3BE8C;">"data_dir"</span>,
<span style="color: #A3BE8C;">"/tmp/derain_data"</span>,
<span style="color: #A3BE8C;">"""Path to the derain data directory."""</span>)
tf.app.flags.DEFINE_integer(<span style="color: #A3BE8C;">"batch_size"</span>,
128,
<span style="color: #A3BE8C;">"""Number of images to process in a batch."""</span>)
tf.app.flags.DEFINE_integer(<span style="color: #A3BE8C;">"max_steps"</span>,
1000000,
<span style="color: #A3BE8C;">"""Number of training batches to run."""</span>)
<span style="color: #D8DEE9;">LEVEL</span> = tf.logging.DEBUG
<span style="color: #D8DEE9;">FLAGS</span> = tf.app.flags.FLAGS
<span style="color: #D8DEE9;">LOG</span> = logging.getLogger(<span style="color: #A3BE8C;">"derain-train"</span>)
</pre>
</div>
</div>
</div>
<div id="outline-container-org05af57a" class="outline-4">
<h4 id="org05af57a"><span class="section-number-4">2.2.3</span> Defining the model</h4>
<div class="outline-text-4" id="text-2-2-3">
<div class="org-src-container">
<pre class="src src-python"><span style="color: #D8DEE9;">MODEL_DEFAULT_PARAMS</span> = {
<span style="color: #A3BE8C;">"learn_rate"</span>: 0.01,
}
<span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">model_fn</span>(features, labels, mode, params):
<span style="color: #D8DEE9;">inputs</span> = features
<span style="color: #D8DEE9;">expected_outputs</span> = labels
tf.summary.image(<span style="color: #A3BE8C;">"inputs"</span>, inputs)
tf.summary.image(<span style="color: #A3BE8C;">"expected_outputs"</span>, expected_outputs)
<span style="color: #D8DEE9;">global_step</span> = tf.train.get_or_create_global_step()
<span style="color: #D8DEE9;">params</span> = {**MODEL_DEFAULT_PARAMS, **params}
<span style="color: #D8DEE9;">l</span> = tf.keras.layers
<span style="color: #D8DEE9;">model</span> = tf.keras.Sequential([
l.Conv2D(
3,
(16, 16),
input_shape=(IMAGE_SIZE, IMAGE_SIZE, 3),
use_bias=<span style="color: #81A1C1;">True</span>,
activation=tf.nn.tanh,
padding=<span style="color: #A3BE8C;">"same"</span>,
)
<span style="color: #616e88;"># </span><span style="color: #616e88;"># 512 kernels of 16x16x3</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">l.Conv2D(</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">512,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">(16, 16),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">input_shape=(IMAGE_SIZE, IMAGE_SIZE, 3),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">use_bias=True,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">activation=tf.nn.tanh,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">padding="same",</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;"># 512 kernels of 1x1x512</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">l.Conv2D(</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">512,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">(1, 1),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">use_bias=True,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">activation=tf.nn.tanh,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;"># 3 kernels of 8x8x512 (one for each color channel)</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">l.Conv2D(</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">3,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">(8, 8),</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">use_bias=True,</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">padding="same",</span>
<span style="color: #616e88;"># </span><span style="color: #616e88;">),</span>
])
<span style="color: #616e88;"># </span><span style="color: #616e88;">TODO: handle each of ModeKeys.{EVAL,TRAIN,PREDICT}</span>
<span style="color: #81A1C1;">if</span> mode == tf.estimator.ModeKeys.TRAIN:
<span style="color: #D8DEE9;">predictions</span> = model(inputs, training=<span style="color: #81A1C1;">True</span>)
<span style="color: #D8DEE9;">norm</span> = tf.norm(expected_outputs - predictions, <span style="color: #81A1C1;">ord</span>=<span style="color: #A3BE8C;">"fro"</span>, axis=[-2, -1])
<span style="color: #D8DEE9;">loss</span> = tf.reduce_mean(norm, 1)
<span style="color: #616e88;"># </span><span style="color: #616e88;">tf.summary.scalar("loss", loss)</span>
<span style="color: #D8DEE9;">optimizer</span> = tf.train.GradientDescentOptimizer(params[<span style="color: #A3BE8C;">"learn_rate"</span>])
<span style="color: #D8DEE9;">train_op</span> = optimizer.minimize(loss, global_step=global_step)
<span style="color: #81A1C1;">return</span> tf.estimator.EstimatorSpec(
mode,
loss=loss,
train_op=train_op,
)
<span style="color: #81A1C1;">raise</span> <span style="color: #8FBCBB;">NotImplementedError</span>
</pre>
</div>
</div>
</div>
<div id="outline-container-org3811aa2" class="outline-4">
<h4 id="org3811aa2"><span class="section-number-4">2.2.4</span> The Rest of the Owl</h4>
<div class="outline-text-4" id="text-2-2-4">
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">train</span>():
<span style="color: #D8DEE9;">regressor</span> = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=FLAGS.checkpoint_dir,
<span style="color: #616e88;"># </span><span style="color: #616e88;">TODO</span>
config=<span style="color: #81A1C1;">None</span>,
params={},
)
regressor.train(
input_fn=<span style="color: #81A1C1;">lambda</span>: dataset(FLAGS.data_dir, <span style="color: #81A1C1;">range</span>(1, 20)).batch(1),
max_steps=FLAGS.max_steps,
)
<span style="color: #81A1C1;">def</span> <span style="color: #88C0D0;">main</span>(argv=<span style="color: #81A1C1;">None</span>):
<span style="color: #81A1C1;">if</span> tf.gfile.Exists(FLAGS.checkpoint_dir):
LOG.debug(<span style="color: #A3BE8C;">"Emptying checkpoint dir"</span>)
tf.gfile.DeleteRecursively(FLAGS.checkpoint_dir)
LOG.debug(<span style="color: #A3BE8C;">"Creating checkpoint dir"</span>)
tf.gfile.MakeDirs(FLAGS.checkpoint_dir)
train()
</pre>
</div>
<div class="org-src-container">
<pre class="src src-python"><span style="color: #81A1C1;">if</span> <span style="color: #81A1C1;">__name__</span> == <span style="color: #A3BE8C;">"__main__"</span>:
tf.logging.set_verbosity(LEVEL)
tf.app.run(main)
</pre>
</div>
</div>
</div>
</div>
</div>
<div id="outline-container-org214f152" class="outline-2">
<h2 id="org214f152"><span class="section-number-2">3</span> References</h2>
<div class="outline-text-2" id="text-3">
<ul class="org-ul">
<li><a href="http://smartdsp.xmu.edu.cn/derainNet.html">http://smartdsp.xmu.edu.cn/derainNet.html</a></li>
<li>X. Fu, J. Huang, D. Zeng, Y. Huang, X. Ding and J. Paisley. ¡°Removing Rain
from Single Images via a Deep Detail Network¡±, CVPR, 2017.</li>
<li>X. Fu, J. Huang, X. Ding, Y. Liao and J. Paisley. ¡°Clearing the Skies: A
deep network architecture for single-image rain removal¡±, IEEE Transactions
on Image Processing, vol. 26, no. 6, pp. 2944-2956, 2017.</li>
</ul>
</div>
</div>
<div id="outline-container-orge90441f" class="outline-2">
<h2 id="orge90441f"><span class="section-number-2">4</span> Appendix</h2>
<div class="outline-text-2" id="text-4">
</div>
<div id="outline-container-org2d05ee5" class="outline-3">
<h3 id="org2d05ee5"><span class="section-number-3">4.1</span> The Dataset</h3>
<div class="outline-text-3" id="text-4-1">
<p>
The authors’ rainy image dataset can be found <a href="http://smartdsp.xmu.edu.cn/memberpdf/fuxueyang/cvpr2017/rainy_image_dataset.zip">here</a>. Unfortunately, that page
was, at the start of this project, unreliable at best; it is now, as of
04/28/2018, entirely unavailable. As such, at the start of this project, I
took the liberty of cloning the authors’ dataset; it is available at
<a href="https://github.com/jinnovation/rainy-image-dataset">github.com/jinnovation/rainy-image-dataset</a>.
</p>
</div>
</div>
<div id="outline-container-org671e1a3" class="outline-3">
<h3 id="org671e1a3"><span class="section-number-3">4.2</span> The Report</h3>
<div class="outline-text-3" id="text-4-2">
<p>
This article is written with Emacs and <a href="https://orgmode.org/manual/HTML-export.html">Org mode</a> in the <a href="https://www-cs-faculty.stanford.edu/~knuth/lp.html">literate programming</a>
style.
</p>
</div>
</div>
<div id="outline-container-orgd541356" class="outline-3">
<h3 id="orgd541356"><span class="section-number-3">4.3</span> The Site</h3>
<div class="outline-text-3" id="text-4-3">
<p>
The source for this write-up, as well as all tangled source files, is hosted
at <a href="https://github.com/jinnovation/derain-net">jinnovation/derain-net</a>. This write-up is generated using Org-mode’s <a href="https://orgmode.org/manual/HTML-export.html">HTML
export functionality</a>, as well as the <a href="https://github.com/fniessen/org-html-themes">ReadTheOrg theme</a>. All resulting source
files were tangled directly from the write-up document.
</p>
</div>
</div>
</div>
</div>
<div id="postamble" class="status">
<p class="date">Date: 2018-04-28 Sat 00:00</p>
<p class="author">Author: Jonathan Jin</p>
<p class="date">Created: 2018-08-27 Mon 15:50</p>
<p class="validation"><a href="http://validator.w3.org/check?uri=referer">Validate</a></p>
</div>
</body>
</html>