-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathreport.html
551 lines (478 loc) · 22.6 KB
/
report.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
<!DOCTYPE html><html><head>
<title>report</title>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<link rel="stylesheet" href="file:////Users/kobebruce/.atom/packages/markdown-preview-enhanced/node_modules/@shd101wyy/mume/dependencies/katex/katex.min.css">
<style>
/**
* prism.js Github theme based on GitHub's theme.
* @author Sam Clarke
*/
code[class*="language-"],
pre[class*="language-"] {
color: #333;
background: none;
font-family: Consolas, "Liberation Mono", Menlo, Courier, monospace;
text-align: left;
white-space: pre;
word-spacing: normal;
word-break: normal;
word-wrap: normal;
line-height: 1.4;
-moz-tab-size: 8;
-o-tab-size: 8;
tab-size: 8;
-webkit-hyphens: none;
-moz-hyphens: none;
-ms-hyphens: none;
hyphens: none;
}
/* Code blocks */
pre[class*="language-"] {
padding: .8em;
overflow: auto;
/* border: 1px solid #ddd; */
border-radius: 3px;
/* background: #fff; */
background: #f5f5f5;
}
/* Inline code */
:not(pre) > code[class*="language-"] {
padding: .1em;
border-radius: .3em;
white-space: normal;
background: #f5f5f5;
}
.token.comment,
.token.blockquote {
color: #969896;
}
.token.cdata {
color: #183691;
}
.token.doctype,
.token.punctuation,
.token.variable,
.token.macro.property {
color: #333;
}
.token.operator,
.token.important,
.token.keyword,
.token.rule,
.token.builtin {
color: #a71d5d;
}
.token.string,
.token.url,
.token.regex,
.token.attr-value {
color: #183691;
}
.token.property,
.token.number,
.token.boolean,
.token.entity,
.token.atrule,
.token.constant,
.token.symbol,
.token.command,
.token.code {
color: #0086b3;
}
.token.tag,
.token.selector,
.token.prolog {
color: #63a35c;
}
.token.function,
.token.namespace,
.token.pseudo-element,
.token.class,
.token.class-name,
.token.pseudo-class,
.token.id,
.token.url-reference .token.variable,
.token.attr-name {
color: #795da3;
}
.token.entity {
cursor: help;
}
.token.title,
.token.title .token.punctuation {
font-weight: bold;
color: #1d3e81;
}
.token.list {
color: #ed6a43;
}
.token.inserted {
background-color: #eaffea;
color: #55a532;
}
.token.deleted {
background-color: #ffecec;
color: #bd2c00;
}
.token.bold {
font-weight: bold;
}
.token.italic {
font-style: italic;
}
/* JSON */
.language-json .token.property {
color: #183691;
}
.language-markup .token.tag .token.punctuation {
color: #333;
}
/* CSS */
code.language-css,
.language-css .token.function {
color: #0086b3;
}
/* YAML */
.language-yaml .token.atrule {
color: #63a35c;
}
code.language-yaml {
color: #183691;
}
/* Ruby */
.language-ruby .token.function {
color: #333;
}
/* Markdown */
.language-markdown .token.url {
color: #795da3;
}
/* Makefile */
.language-makefile .token.symbol {
color: #795da3;
}
.language-makefile .token.variable {
color: #183691;
}
.language-makefile .token.builtin {
color: #0086b3;
}
/* Bash */
.language-bash .token.keyword {
color: #0086b3;
}
/* highlight */
pre[data-line] {
position: relative;
padding: 1em 0 1em 3em;
}
pre[data-line] .line-highlight-wrapper {
position: absolute;
top: 0;
left: 0;
background-color: transparent;
display: block;
width: 100%;
}
pre[data-line] .line-highlight {
position: absolute;
left: 0;
right: 0;
padding: inherit 0;
margin-top: 1em;
background: hsla(24, 20%, 50%,.08);
background: linear-gradient(to right, hsla(24, 20%, 50%,.1) 70%, hsla(24, 20%, 50%,0));
pointer-events: none;
line-height: inherit;
white-space: pre;
}
pre[data-line] .line-highlight:before,
pre[data-line] .line-highlight[data-end]:after {
content: attr(data-start);
position: absolute;
top: .4em;
left: .6em;
min-width: 1em;
padding: 0 .5em;
background-color: hsla(24, 20%, 50%,.4);
color: hsl(24, 20%, 95%);
font: bold 65%/1.5 sans-serif;
text-align: center;
vertical-align: .3em;
border-radius: 999px;
text-shadow: none;
box-shadow: 0 1px white;
}
pre[data-line] .line-highlight[data-end]:after {
content: attr(data-end);
top: auto;
bottom: .4em;
}html body{font-family:"Helvetica Neue",Helvetica,"Segoe UI",Arial,freesans,sans-serif;font-size:16px;line-height:1.6;color:#333;background-color:#fff;overflow:initial;box-sizing:border-box;word-wrap:break-word}html body>:first-child{margin-top:0}html body h1,html body h2,html body h3,html body h4,html body h5,html body h6{line-height:1.2;margin-top:1em;margin-bottom:16px;color:#000}html body h1{font-size:2.25em;font-weight:300;padding-bottom:.3em}html body h2{font-size:1.75em;font-weight:400;padding-bottom:.3em}html body h3{font-size:1.5em;font-weight:500}html body h4{font-size:1.25em;font-weight:600}html body h5{font-size:1.1em;font-weight:600}html body h6{font-size:1em;font-weight:600}html body h1,html body h2,html body h3,html body h4,html body h5{font-weight:600}html body h5{font-size:1em}html body h6{color:#5c5c5c}html body strong{color:#000}html body del{color:#5c5c5c}html body a:not([href]){color:inherit;text-decoration:none}html body a{color:#08c;text-decoration:none}html body a:hover{color:#00a3f5;text-decoration:none}html body img{max-width:100%}html body>p{margin-top:0;margin-bottom:16px;word-wrap:break-word}html body>ul,html body>ol{margin-bottom:16px}html body ul,html body ol{padding-left:2em}html body ul.no-list,html body ol.no-list{padding:0;list-style-type:none}html body ul ul,html body ul ol,html body ol ol,html body ol ul{margin-top:0;margin-bottom:0}html body li{margin-bottom:0}html body li.task-list-item{list-style:none}html body li>p{margin-top:0;margin-bottom:0}html body .task-list-item-checkbox{margin:0 .2em .25em -1.8em;vertical-align:middle}html body .task-list-item-checkbox:hover{cursor:pointer}html body blockquote{margin:16px 0;font-size:inherit;padding:0 15px;color:#5c5c5c;background-color:#f0f0f0;border-left:4px solid #d6d6d6}html body blockquote>:first-child{margin-top:0}html body blockquote>:last-child{margin-bottom:0}html body hr{height:4px;margin:32px 0;background-color:#d6d6d6;border:0 none}html body table{margin:10px 0 15px 0;border-collapse:collapse;border-spacing:0;display:block;width:100%;overflow:auto;word-break:normal;word-break:keep-all}html body table th{font-weight:bold;color:#000}html body table td,html body table th{border:1px solid #d6d6d6;padding:6px 13px}html body dl{padding:0}html body dl dt{padding:0;margin-top:16px;font-size:1em;font-style:italic;font-weight:bold}html body dl dd{padding:0 16px;margin-bottom:16px}html body code{font-family:Menlo,Monaco,Consolas,'Courier New',monospace;font-size:.85em !important;color:#000;background-color:#f0f0f0;border-radius:3px;padding:.2em 0}html body code::before,html body code::after{letter-spacing:-0.2em;content:"\00a0"}html body pre>code{padding:0;margin:0;font-size:.85em !important;word-break:normal;white-space:pre;background:transparent;border:0}html body .highlight{margin-bottom:16px}html body .highlight pre,html body pre{padding:1em;overflow:auto;font-size:.85em !important;line-height:1.45;border:#d6d6d6;border-radius:3px}html body .highlight pre{margin-bottom:0;word-break:normal}html body pre code,html body pre tt{display:inline;max-width:initial;padding:0;margin:0;overflow:initial;line-height:inherit;word-wrap:normal;background-color:transparent;border:0}html body pre code:before,html body pre tt:before,html body pre code:after,html body pre tt:after{content:normal}html body p,html body blockquote,html body ul,html body ol,html body dl,html body pre{margin-top:0;margin-bottom:16px}html body kbd{color:#000;border:1px solid #d6d6d6;border-bottom:2px solid #c7c7c7;padding:2px 4px;background-color:#f0f0f0;border-radius:3px}@media print{html body{background-color:#fff}html body h1,html body h2,html body h3,html body h4,html body h5,html body h6{color:#000;page-break-after:avoid}html body blockquote{color:#5c5c5c}html body pre{page-break-inside:avoid}html body table{display:table}html body img{display:block;max-width:100%;max-height:100%}html body pre,html body code{word-wrap:break-word;white-space:pre}}.markdown-preview{width:100%;height:100%;box-sizing:border-box}.markdown-preview .pagebreak,.markdown-preview .newpage{page-break-before:always}.markdown-preview pre.line-numbers{position:relative;padding-left:3.8em;counter-reset:linenumber}.markdown-preview pre.line-numbers>code{position:relative}.markdown-preview pre.line-numbers .line-numbers-rows{position:absolute;pointer-events:none;top:1em;font-size:100%;left:0;width:3em;letter-spacing:-1px;border-right:1px solid #999;-webkit-user-select:none;-moz-user-select:none;-ms-user-select:none;user-select:none}.markdown-preview pre.line-numbers .line-numbers-rows>span{pointer-events:none;display:block;counter-increment:linenumber}.markdown-preview pre.line-numbers .line-numbers-rows>span:before{content:counter(linenumber);color:#999;display:block;padding-right:.8em;text-align:right}.markdown-preview .mathjax-exps .MathJax_Display{text-align:center !important}.markdown-preview:not([for="preview"]) .code-chunk .btn-group{display:none}.markdown-preview:not([for="preview"]) .code-chunk .status{display:none}.markdown-preview:not([for="preview"]) .code-chunk .output-div{margin-bottom:16px}.scrollbar-style::-webkit-scrollbar{width:8px}.scrollbar-style::-webkit-scrollbar-track{border-radius:10px;background-color:transparent}.scrollbar-style::-webkit-scrollbar-thumb{border-radius:5px;background-color:rgba(150,150,150,0.66);border:4px solid rgba(150,150,150,0.66);background-clip:content-box}html body[for="html-export"]:not([data-presentation-mode]){position:relative;width:100%;height:100%;top:0;left:0;margin:0;padding:0;overflow:auto}html body[for="html-export"]:not([data-presentation-mode]) .markdown-preview{position:relative;top:0}@media screen and (min-width:914px){html body[for="html-export"]:not([data-presentation-mode]) .markdown-preview{padding:2em calc(50% - 457px + 2em)}}@media screen and (max-width:914px){html body[for="html-export"]:not([data-presentation-mode]) .markdown-preview{padding:2em}}@media screen and (max-width:450px){html body[for="html-export"]:not([data-presentation-mode]) .markdown-preview{font-size:14px !important;padding:1em}}@media print{html body[for="html-export"]:not([data-presentation-mode]) #sidebar-toc-btn{display:none}}html body[for="html-export"]:not([data-presentation-mode]) #sidebar-toc-btn{position:fixed;bottom:8px;left:8px;font-size:28px;cursor:pointer;color:inherit;z-index:99;width:32px;text-align:center;opacity:.4}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] #sidebar-toc-btn{opacity:1}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc{position:fixed;top:0;left:0;width:300px;height:100%;padding:32px 0 48px 0;font-size:14px;box-shadow:0 0 4px rgba(150,150,150,0.33);box-sizing:border-box;overflow:auto;background-color:inherit}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc::-webkit-scrollbar{width:8px}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc::-webkit-scrollbar-track{border-radius:10px;background-color:transparent}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc::-webkit-scrollbar-thumb{border-radius:5px;background-color:rgba(150,150,150,0.66);border:4px solid rgba(150,150,150,0.66);background-clip:content-box}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc a{text-decoration:none}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc ul{padding:0 1.6em;margin-top:.8em}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc li{margin-bottom:.8em}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .md-sidebar-toc ul{list-style-type:none}html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .markdown-preview{left:300px;width:calc(100% - 300px);padding:2em calc(50% - 457px - 150px);margin:0;box-sizing:border-box}@media screen and (max-width:1274px){html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .markdown-preview{padding:2em}}@media screen and (max-width:450px){html body[for="html-export"]:not([data-presentation-mode])[html-show-sidebar-toc] .markdown-preview{width:100%}}html body[for="html-export"]:not([data-presentation-mode]):not([html-show-sidebar-toc]) .markdown-preview{left:50%;transform:translateX(-50%)}html body[for="html-export"]:not([data-presentation-mode]):not([html-show-sidebar-toc]) .md-sidebar-toc{display:none}
/* Please visit the URL below for more information: */
/* https://shd101wyy.github.io/markdown-preview-enhanced/#/customize-css */
</style>
</head>
<body for="html-export">
<div class="mume markdown-preview ">
<h1 class="mume-header" id="report">Report</h1>
<p>(Bruce) Wai-Ting Li</p>
<h3 class="mume-header" id="table-of-content">Table of Content</h3>
<ol>
<li><a href="#clf-model">Classification Model</a></li>
<li><a href="#visualization">Network Visualization</a></li>
<li><a href="#semisupervised">Semi-Supervised Learning</a></li>
</ol>
<hr>
<h2 class="mume-header undefined" id="clf-model">1: Classification Model</h2>
<ul>
<li>Approach: multi-class classification</li>
<li>Model candidates
<ul>
<li>ResNet50</li>
<li>ResNeXt50</li>
<li>ResNeSt50</li>
</ul>
</li>
<li>Parameters & hyperparameters
<ul>
<li>optimizer: Adam</li>
<li>epochs: 100</li>
<li>learning_rate: 1e-3</li>
<li>momentum: 0.9</li>
<li>batch_size: 111</li>
<li>weight_decay: 1e-5</li>
</ul>
</li>
<li>Prevent overfitting: save model checkpoint only if the validation F1 score improves.</li>
</ul>
<p>Note</p>
<ul>
<li>Due to computational constraint, I did not perform hyperparameter tuning. The choice of hyperparameters were adapted from the <a href="https://www.kaggle.com/hmchuong/pytorch-baseline-model">Pytorch Baseline model on Kaggle</a>.</li>
<li><code>resnest50</code> + <code>batch_size=128</code> will run out of cuda memory in colab, so I used <code>batch_size=111</code> for all models for fair comparison</li>
</ul>
<h3 class="mume-header" id="a-data-splitting-model-performance-on-test-set">a: Data Splitting & Model Performance on Test Set</h3>
<p>The original labeled training data is split into 60/10/30 train/dev/test. I randomly performed the splitting and manually assessed the distribution of each class to keep the distribution as similar as possible.</p>
<center><img src="./media/data_split.jpg" width="500" style="background-color:white;"><br></center>
<h4 class="mume-header" id="f1-score-on-test-set">F1 Score on Test Set</h4>
<table>
<thead>
<tr>
<th>Type</th>
<th style="text-align:center">ResNet50</th>
<th style="text-align:center">ResNeXt50</th>
<th style="text-align:center">ResNeSt50</th>
</tr>
</thead>
<tbody>
<tr>
<td>Opacity</td>
<td style="text-align:center">0.9055</td>
<td style="text-align:center">0.9258</td>
<td style="text-align:center">0.9544</td>
</tr>
<tr>
<td>Diabetic retinopathy</td>
<td style="text-align:center">0.9070</td>
<td style="text-align:center">0.9310</td>
<td style="text-align:center">0.9515</td>
</tr>
<tr>
<td>Glaucoma</td>
<td style="text-align:center">0.8718</td>
<td style="text-align:center">0.9030</td>
<td style="text-align:center">0.9329</td>
</tr>
<tr>
<td>Macular edema</td>
<td style="text-align:center">0.8439</td>
<td style="text-align:center">0.9169</td>
<td style="text-align:center">0.9167</td>
</tr>
<tr>
<td>Macular degeneration</td>
<td style="text-align:center">0.9091</td>
<td style="text-align:center">0.8870</td>
<td style="text-align:center">0.9483</td>
</tr>
<tr>
<td>Retinal vascular occlusion</td>
<td style="text-align:center">0.8705</td>
<td style="text-align:center">0.8594</td>
<td style="text-align:center">0.9446</td>
</tr>
<tr>
<td>Normal</td>
<td style="text-align:center">0.9595</td>
<td style="text-align:center">0.9653</td>
<td style="text-align:center">0.9844</td>
</tr>
<tr>
<td><strong>Average</strong></td>
<td style="text-align:center"><strong>0.8953</strong></td>
<td style="text-align:center"><strong>0.9126</strong></td>
<td style="text-align:center"><strong>0.9476</strong></td>
</tr>
</tbody>
</table>
<h3 class="mume-header" id="b-data-augmentation">b: Data Augmentation</h3>
<ul>
<li>Random horizontal flip with <code>p=0.5</code></li>
<li>Random vertical flip with <code>p=0.5</code></li>
<li>Random rotation between [-90, +90]</li>
<li>Normalization using ImageNet's mean and std for each of the RGB channel
<ul>
<li>Mean: [0.485, 0.456, 0.406]</li>
<li>Std: [0.229, 0.224, 0.225]</li>
</ul>
</li>
<li>Changing brightness, contrast, saturation, hue worsened performance (It's possible that the parameters were not set properly)</li>
</ul>
<p>I applied random flipping and rotation as the retinal images are mostly squared and it's possible for regions contributed to diseases to occur at anywhere in the image.</p>
<h3 class="mume-header" id="c-plot-of-training-validation-loss-f1">c: Plot of Training-Validation Loss & F1</h3>
<center><img src="./media/train-dev-plot.jpg" width="700" style="background-color:white;"><br>The solid lines represent the validation score and the dotted line represent the training score. The markers on the plot of f1 score represent the best score achieved.</center>
<h4 class="mume-header" id="best-f1-score-on-dev-set">Best F1 Score on Dev Set</h4>
<table>
<thead>
<tr>
<th>Type</th>
<th style="text-align:center">ResNet50</th>
<th style="text-align:center">ResNeXt50</th>
<th style="text-align:center">ResNeSt50</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Average</strong></td>
<td style="text-align:center">0.9264</td>
<td style="text-align:center">0.9531</td>
<td style="text-align:center">0.9625</td>
</tr>
</tbody>
</table>
<hr>
<h2 class="mume-header undefined" id="visualization">2: Network Visualization</h2>
<p>The following images have only one class.</p>
<h3 class="mume-header" id="saliency-map">Saliency Map</h3>
<center><img src="./media/saliency-diabetic retinopathy.jpg" width="700" style="background-color:white;"><br></center>
<center><img src="./media/saliency-glaucoma.jpg" width="700" style="background-color:white;"><br></center>
<p>The saliency map tells us the degree to which each pixel in the image affects the classification score for that image.</p>
<h3 class="mume-header" id="grad-cam">Grad-CAM</h3>
<p>The target layer is the last Convolution layer of the last Bottleneck layer.</p>
<center><img src="./media/gradcam-diabetic retinopathy.jpg" width="700" style="background-color:white;"><br></center>
<center><img src="./media/gradcam-glaucoma.jpg" width="700" style="background-color:white;"><br></center>
<p>Based on the visualization of Grad-CAM, we may tell that some of the black dots may contribute to diabetic retinopathy where the visualization shows red. And we may also tell that the crack-like patterns may contribute to glaucoma.</p>
<hr>
<h2 class="mume-header undefined" id="semisupervised">3: Semi-Supervised Learning</h2>
<ul>
<li>Approach: Scikit-learn Label Propagation with <code>rbf</code> kernel</li>
<li>Data: the 60% labeled data (train set) + unlabeled data</li>
<li>Best gamma (based on f1 score of the prediction on the labeled train set):
<ul>
<li>ResNet50: 0.5 (avg. f1: 0.9927)</li>
<li>ResNeXt50: 0.83 (avg. f1: 0.9795)</li>
<li>ResNeSt50: 0.07 (avg. f1: 0.9987)</li>
</ul>
</li>
</ul>
<p>Retraining Model using unlabeled data</p>
<ul>
<li>Since I noticed if took longer for models to converge with addition of unlabeled data, I retrained all models for 200 epochs while keeping all other hyperparameters the same.</li>
</ul>
<h3 class="mume-header" id="model-performance-on-test-set">Model Performance on Test Set</h3>
<table>
<thead>
<tr>
<th>Type</th>
<th style="text-align:center">ResNet50</th>
<th style="text-align:center">ResNeXt50</th>
<th style="text-align:center">ResNeSt50</th>
</tr>
</thead>
<tbody>
<tr>
<td>Opacity</td>
<td style="text-align:center">0.9266</td>
<td style="text-align:center">0.9411</td>
<td style="text-align:center">0.9526</td>
</tr>
<tr>
<td>Diabetic retinopathy</td>
<td style="text-align:center">0.9254</td>
<td style="text-align:center">0.9292</td>
<td style="text-align:center">0.9227</td>
</tr>
<tr>
<td>Glaucoma</td>
<td style="text-align:center">0.8812</td>
<td style="text-align:center">0.9154</td>
<td style="text-align:center">0.9297</td>
</tr>
<tr>
<td>Macular edema</td>
<td style="text-align:center">0.8757</td>
<td style="text-align:center">0.9097</td>
<td style="text-align:center">0.9159</td>
</tr>
<tr>
<td>Macular degeneration</td>
<td style="text-align:center">0.9058</td>
<td style="text-align:center">0.8903</td>
<td style="text-align:center">0.9358</td>
</tr>
<tr>
<td>Retinal vascular occlusion</td>
<td style="text-align:center">0.9058</td>
<td style="text-align:center">0.9158</td>
<td style="text-align:center">0.9154</td>
</tr>
<tr>
<td>Normal</td>
<td style="text-align:center">0.9563</td>
<td style="text-align:center">0.9718</td>
<td style="text-align:center">0.9745</td>
</tr>
<tr>
<td><strong>Average</strong></td>
<td style="text-align:center"><strong>0.9110</strong></td>
<td style="text-align:center"><strong>0.9247</strong></td>
<td style="text-align:center"><strong>0.9352</strong></td>
</tr>
<tr>
<td><strong>Part 1 Average</strong></td>
<td style="text-align:center"><strong>0.8953</strong></td>
<td style="text-align:center"><strong>0.9126</strong></td>
<td style="text-align:center"><strong>0.9476</strong></td>
</tr>
</tbody>
</table>
<p>Based on the results, the average f1 score for ResNet50 improved from 0.8953 to 0.9110 and the score for ResNeXt50 improved from 0.9126 to 0.9247, however, the score for ResNeSt50 decreased from 0.9476 to 0.9352.</p>
<h3 class="mume-header" id="plot-of-training-validation-loss-f1">Plot of Training-Validation Loss & F1</h3>
<center><img src="./media/train-dev-plot-semi.jpg" width="700" style="background-color:white;"><br>The solid lines represent the validation score and the dotted line represent the training score. The markers on the plot of f1 score represent the best score achieved.</center>
<h4 class="mume-header" id="best-f1-score-on-dev-set-1">Best F1 Score on Dev Set</h4>
<table>
<thead>
<tr>
<th>Type</th>
<th style="text-align:center">ResNet50</th>
<th style="text-align:center">ResNeXt50</th>
<th style="text-align:center">ResNeSt50</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Average</strong></td>
<td style="text-align:center">0.9235</td>
<td style="text-align:center">0.9472</td>
<td style="text-align:center">0.9461</td>
</tr>
<tr>
<td><strong>Part 1 Average</strong></td>
<td style="text-align:center">0.9264</td>
<td style="text-align:center">0.9531</td>
<td style="text-align:center">0.9625</td>
</tr>
</tbody>
</table>
</div>
</body></html>