-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathindex.html
381 lines (358 loc) · 12.5 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
<!DOCTYPE html>
<html>
<head>
<title>auto_grader</title>
<link rel="stylesheet" href="global.css">
</head>
<header style="position:fixed;">
<div class="nav navs navs2">
<a href="#upcoming"> Upcoming</a>
<a href="#updates">Updates</a>
<a href="https://github.com/ceyxasm/auto_grader" target="_blank">Github</a>
<p class="name"> <strong>Automatic Grading of Handwritten Answer Sheets</strong> </p>
</div>
<br><br><br>
</header>
<body>
<br><br><br><br>
<!--Upcoming work-->
<div id="upcoming"><br><br><br></div>
<div class="card upcoming" >
<h2>Upcoming work</h2>
<ul>
<li>Implementing a better algorithm for finding the bounding box of the answer sheet</li>
<li>Attn/CTC/RCNN/Resnet [READ]</li>
<li> Tryout TrOCR</li>
<li>Word level detection in answer sheet
<ul>
<li> Tryout WordDetetctorNN</li>
<li> Tryout CRAFT</li>
<li> Tryout DBNet</li>
</ul>
</li>
<li><s>Understanding source code</s></li>
<li><s>Test out ResNet+Attn.</s></li>
<li>Writer verification task</li>
<li>Applying data augmentation before training</li>
<li><s>Check and compare results for train with urdu special chars</s></li>
<li><s>Applying TPS transformation or not [READ]</s></li>
<li>Designing a distance function given 2 words and whether they both are to be considered equivalent; given the threshold</li>
<li>Exploring transformers</li>
</ul>
</div>
<!--List of updates-->
<div id="updates"><br><br><br></div>
<div class="card updates" >
<h2> Updates </h2>
<ul>
<div class="update">
<li> <strong>14-Jan-2023: <br>Devnagri training 2.0</strong> <sup style="color:red">*new</sup> </li>
<ol>
<p>model details: Transformation None Resnet BiLSTM Attn <br>
augmentation: None <br>
character set= 'ऀँंः ऄअआइईउऊऋऌऍऎएऐऑऒओऔक खगघङचछजझञटठडढणतथदधनऩ पफबभमयरऱलळऴवशषसहऺऻ़ ऽािीुूृॄॅॆेैॉॊोौ्ॏ ॐ॒॑॓॔ॕॖॗग़ज़ड़ढ़फ़य़ॠॡॢॣ।॥०१२३४५६७८९॰ॱॲॳॴॵॶॷॸॹॺॻॼॽॾॿ'</p>
<table>
<tr>
<th> Iter count</th>
<th> Current Validation Accuracy</th>
<th> Best Validation Accuracy</th>
<th> Train Loss</th>
<th> Val Loss</th>
</tr>
<tr>
<td> 10k </td>
<td> 61.34 </td>
<td> 64.22 </td>
<td> 0.016 </td>
<td> 0.1.10 </td>
</tr>
<tr>
<td> 16k </td>
<td> 68.49 </td>
<td> 68.91 </td>
<td> 0.006 </td>
<td> 0.92 </td>
</tr>
<tr>
<td> 22k </td>
<td> 69.87 </td>
<td> 69.87 </td>
<td> 0.002 </td>
<td> 0.964 </td>
</tr>
<tr>
<td> 1e5 </td>
<td> 68.37 </td>
<td> 69.87 </td>
<td> 0.00008 </td>
<td> 1.08</td>
</tr>
<tr>
<td> 2e5 </td>
<td> 68.28 </td>
<td> 69.87 </td>
<td> 0.00001 </td>
<td> 1.25 </td>
</tr>
<tr>
<td> 3e5 </td>
<td> 68.32 </td>
<td> 69.87 </td>
<td> 0.00001 </td>
<td> 1.31 </td>
</tr>
</table>
<br>
<strong>Inference: </strong> <p>Model peaks at 69.87% validation accuracy @22k iters.
<br> Model took 1.52 more time (22 hours) to train 3e5 epochs as compared to RCNN-CTC
<br> But RCNN-CTC gave peak of 61.48 after 17 hours training
<br> Resnet-Attn reached 69% after 1.5 hours of training
</p>
</ol>
<br><br>
</div>
<div class="update">
<li> <strong>28-Dec-2022: <br> Updates</strong> <sup style="color:red">*new</sup> </li>
<ul> Literature review on Augmentation.</ul>
<ul> Read about ResNeT+Attn.</ul>
<ul> Read about TPS transformation.</ul>
<ul> Augmentor, Autography & Straug</ul>
<br><br>
</div>
<div class="update">
<li> <strong>22-Dec-2022: <br>Urdu Training init</strong> <sup style="color:red">*new</sup> </li>
<ol>
<p>model details: Transformation None RCNN BiLSTM CTC <br>
augmentation: None <br>
character set= 'آأابپتٹثجچحخدڈذرڑزژسشصضطظعفقکگلمنںوؤہۂۃھءیئےۓ'</p>
<table>
<tr>
<th> Iter count</th>
<th> Current Validation Accuracy</th>
<th> Best Validation Accuracy</th>
<th> Train Loss</th>
<th> Val Loss</th>
</tr>
<tr>
<td> 1e5 </td>
<td> 74.808 </td>
<td> 75.471 </td>
<td> 0.00056 </td>
<td> 0.709 </td>
</tr>
<tr>
<td> 2e5 </td>
<td> 75.839 </td>
<td> 76.412 </td>
<td> 0.0005 </td>
<td> 0.722 </td>
</tr>
<tr>
<td> 3e5 </td>
<td> 74.505 </td>
<td> 76.797 </td>
<td> 0.0008 </td>
<td> 0.764 </td>
</tr>
</table>
<br>
<strong>Inference: </strong> <p>Model touches 76% accuracy at 1.08e5 iters. <br>
Model touches 70% accuracy at 24k iters. <br>
</p>
</ol>
<br><br>
</div>
<div class="update">
<li> <strong>08-Nov-2022: <br>Devnagri Training finished</strong> </li>
<ol>
<p>model details: Transformation None RCNN BiLSTM CTC <br>
augmentation: None <br>
character set= 'ऀँंः ऄअआइईउऊऋऌऍऎएऐऑऒओऔक खगघङचछजझञटठडढणतथदधनऩ पफबभमयरऱलळऴवशषसहऺऻ़ ऽािीुूृॄॅॆेैॉॊोौ्ॏ ॐ॒॑॓॔ॕॖॗग़ज़ड़ढ़फ़य़ॠॡॢॣ।॥०१२३४५६७८९॰ॱॲॳॴॵॶॷॸॹॺॻॼॽॾॿ'</p>
<table>
<tr>
<th> Iter count</th>
<th> Current Validation Accuracy</th>
<th> Best Validation Accuracy</th>
<th> Train Loss</th>
<th> Val Loss</th>
</tr>
<tr>
<td> 22000 </td>
<td> 60.261 </td>
<td> 60.261 </td>
<td> 0.00202 </td>
<td> 0.90251 </td>
</tr>
<tr>
<td> 2.4e5 </td>
<td> 61.308 </td>
<td> 61.308 </td>
<td> 0.00001 </td>
<td> 1.12084 </td>
</tr>
<tr>
<td> 3e5 </td>
<td> 61.103 </td>
<td> 61.308 </td>
<td> 0 </td>
<td> 1.18215 </td>
</tr>
<tr>
<td> 5e5 </td>
<td> 61.300 </td>
<td> 61.481 </td>
<td> 0 </td>
<td> 1.2352 </td>
</tr>
<tr>
<td> 7.5e5 </td>
<td> 60.977 </td>
<td> 61.528 </td>
<td> 0 </td>
<td> 1.26202 </td>
</tr>
<tr>
<td> 10e5 </td>
<td> 61.056 </td>
<td> 61.528 </td>
<td> 0 </td>
<td> 1.29381 </td>
</tr>
</table>
<br>
<strong>Inference: </strong> <p>Model peaks at 7.4e5 iters with 61.528% acc. <br>
However a practically same accuracy of 61.308% is reached in 2.4e5 iters <br>
Even a sub-optimal accuracy of 60.281% can be reached with mere 22k iters.
55% was reached in 10k iters</p>
<p>Following are some results from the best_accuracy model</p>
<table class="table_img">
<tr>
<th> Image name</th>
<th> Ground Truth</th>
<th> Prediction</th>
</tr>
<tr>
<td> test/14.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/14.jpg" /> </td>
<td> कंपनी </td>
</tr>
<tr>
<td> test/29.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/29.jpg" /> </td>
<td> सपर्याप्त </td>
</tr>
<tr>
<td> test/100.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/100.jpg" /> </td>
<td> आँगलाइख </td>
</tr>
<tr>
<td> test/350.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/350.jpg" /> </td>
<td> ९ </td>
</tr>
<tr>
<td> test/549.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/549.jpg" /> </td>
<td> शशणशणर्य </td>
</tr>
<tr>
<td> test/610.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/610.jpg" /> </td>
<td> अभियान </td>
</tr>
<tr>
<td> test/1112.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/1112.jpg" /> </td>
<td> फ़यदा </td>
</tr>
<tr>
<td> test/1455.jpg </td>
<td> <img style="height:80px" src="./archive/nov_8/1455.jpg" /> </td>
<td> पेट्रोलिरम </td>
</tr>
</table>
<strong>Analysis: </strong>
<p>For lots of instances we are having exact matches. In places where there are few characters mismatched;<br>
they too can be evaluated to true, as given the context by सपर्याप्त, the examinee can only mean अपर्याप्त. <br>
A function to calculate the distance between such words and deciding the threshold is the need </p>
</ol>
<br><br>
</div>
<div class="update">
<li> <strong>28-Oct-2022:</strong></li>
<ol>
<li>Training of devnagri failing [RESOLVED](issue: GPU had run out of memory)</li>
<li>Training of bengali completed. Model trained for 1.5e6 epochs. The results on validation data is summarized below</li>
<p>model details: Transformation None RCNN BiLSTM CTC <br>
augmentations: None <br>
character set= '-।ঁংঃ অআইঈউঊঋএঐওঔকখগ ঘঙচছজঝঞটঠডঢণত থদধনপফবভমযরলশষস হ়ঽািীুূৃৄেৈোৌ্ ৎৗড়ঢ়য়ৠৢৣ০১২৩৪৫৬৭৮৯ৰৱ৲৳৴৵৶৷৹৺৻'</p>
<table>
<tr>
<th> Iter count</th>
<th> Current Validation Accuracy</th>
<th> Best Validation Accuracy</th>
<th> Train Loss</th>
<th> Val Loss</th>
</tr>
<tr>
<td> 14000 </td>
<td> 60.525 </td>
<td> 60.525 </td>
<td> 0.01673 </td>
<td> 0.7135 </td>
</tr>
<tr>
<td> 38000 </td>
<td> 66.203 </td>
<td> 66.203 </td>
<td> 0.00015 </td>
<td> 0.85342 </td>
</tr>
<tr>
<td> 3e5 </td>
<td> 66.566 </td>
<td> 66.598 </td>
<td> 0 </td>
<td> 0.99272 </td>
</tr>
<tr>
<td> 5.7e5 </td>
<td> 66.808 </td>
<td> 66.808 </td>
<td> 0 </td>
<td> 1.02564 </td>
</tr>
<tr>
<td> 6e5 </td>
<td> 66.586 </td>
<td> 66.808 </td>
<td> 0 </td>
<td> 1.03218 </td>
</tr>
<tr>
<td> 9e5 </td>
<td> 66.559 </td>
<td> 66.808 </td>
<td> 0 </td>
<td> 1.05261 </td>
</tr>
<tr>
<td> 15e5 </td>
<td> 66.628 </td>
<td> 66.808 </td>
<td> 0 </td>
<td> 1.08701 </td>
</tr>
<p1>62% was reached in 22k iters and 64% was reached in 28k iters</p1>
</table>
<br>
<strong>Inference: </strong> <p>Model peaks at 5.7e5 iters with 66.808% acc. <br>
But even a sub-optimal accuracy of 64% can be reached with mere 28k iters.</p>
</ol>
</div>
<div class="update"><li> <strong>01-Oct-2022:</strong> Init </li></div>
</ul>
</div>
</body>
</html>