-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathemotion.html
746 lines (654 loc) · 23.6 KB
/
emotion.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
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1">
<title>ConvNetJS MNIST demo</title>
<meta name="description" content="">
<meta name="author" content="">
<style>
.layer {
border: 1px solid #999;
margin-bottom: 5px;
text-align: left;
padding: 10px;
}
.layer_act {
width: 450px;
float: right;
}
.ltconv {
background-color: #FDD;
}
.ltrelu {
background-color: #FDF;
}
.ltpool {
background-color: #DDF;
}
.ltsoftmax {
background-color: #FFD;
}
.ltfc {
background-color: #DFF;
}
.ltlrn {
background-color: #DFD;
}
.ltdropout {
background-color: #AAA;
}
.ltitle {
color: #333;
font-size: 18px;
}
.actmap {
margin: 1px;
}
#trainstats {
text-align: left;
}
.clear {
clear: both;
}
#wrap {
width: 800px;
margin-left: auto;
margin-right: auto;
}
h1 {
font-size: 16px;
color: #333;
background-color: #DDD;
border-bottom: 1px #999 solid;
text-align: center;
}
.secpart {
width: 400px;
float: left;
}
#lossgraph {
/*border: 1px solid #F0F;*/
width: 100%;
}
.probsdiv canvas {
float: left;
}
.probsdiv {
height: 60px;
width: 180px;
display: inline-block;
font-size: 12px;
box-shadow: 0px 0px 2px 2px #EEE;
margin: 5px;
padding: 5px;
color: black;
}
.pp {
margin: 1px;
padding: 1px;
}
#testset_acc {
margin-bottom: 200px;
}
body {
font-family: Arial, "Helvetica Neue", Helvetica, sans-serif;
}
</style>
<script src="convnetjs/demo/jquery-1.8.3.min.js"></script>
<script src="convnetjs/build/vis.js"></script>
<script src="convnetjs/build/util.js"></script>
<script src="convnetjs/build/convnet.js"></script>
<script src="data/processed/ckp_scaled_features.js"></script>
<script src="data/processed/ckp_test_scaled_features.js"></script>
<script src="data/processed/ck+_labels.js"></script>
<script src="data/processed/ck+_test_labels.js"></script>
<script src="data/processed/jaffe_features.js"></script>
<script src="data/processed/jaffe_test_features.js"></script>
<script src="data/processed/jaffe_labels.js"></script>
<script src="data/processed/jaffe_test_labels.js"></script>
<script>
var layer_defs, net, trainer;
var t = "\n\
// this gets executed on startup\n\
var layer_defs = [];\n\
// input layer of size 1x1x71 (all volumes are 3D)\n\
layer_defs.push({type:'input', out_sx:2, out_sy:1, out_depth:71});\n\
// some fully connected layers\n\
// layer_defs.push({type:'fc', num_neurons:100, activation:'relu'});\n\
layer_defs.push({type:'fc', num_neurons:700, activation:'relu'});\n\
// layer_defs.push({type:'fc', num_neurons:100, activation:'relu'});\n\
// a softmax classifier predicting probabilities for six classes\n\
// layer_defs.push({type:'softmax', num_classes:6});\n\
layer_defs.push({type:'svm', num_classes:6});\n\
\n\
// create a net out of it\n\
net = new convnetjs.Net();\n\
net.makeLayers(layer_defs);\n\
\n\
trainer = new convnetjs.SGDTrainer(net, {method:'adadelta', batch_size:20, l2_decay:0.001});\n\
";
// ------------------------
// BEGIN MNIST SPECIFIC STUFF
// ------------------------
classes_txt = ['0','1','2','3','4','5', '6'];
var image_width = 256;
var image_height = 256;
var counter = 0;
var data, labels;
var tdata, tlabels;
function set_up_data() {
// ckp_scaled_features is an object of arrays of points
// ckp_test_scaled_features is an object of arrays of points
// ckp_labels is a array of labels
// ckp_test_labels is a array of labels
data = [];
sizes = [];
mins = [];
labels = [];
for (var key in ckp_scaled_features) {
data.push(ckp_scaled_features[key]);
labels.push(ckp_labels[key]);
width = [];
height = [];
for (var i = 0; i < ckp_scaled_features[key].length; i++) {
width.push(ckp_scaled_features[key][i][0]);
height.push(ckp_scaled_features[key][i][1]);
}
min = [];
min[0] = Math.min.apply(Math, width);
min[1] = Math.min.apply(Math, height);
mins.push(min);
size = [];
size[0] = Math.max.apply(Math, width) - min[0];
size[1] = Math.max.apply(Math, height) - min[1];
sizes.push(size);
}
// for (var key2 in jaffe_features) {
// data.push(jaffe_features[key2]);
// labels.push(jaffe_labels[key2]);
// width = [];
// height = [];
// for (var i = 0; i < jaffe_features[key2].length; i++) {
// width.push(jaffe_features[key2][i][0]);
// height.push(jaffe_features[key2][i][1]);
// }
// min = [];
// min[0] = Math.min.apply(Math, width);
// min[1] = Math.min.apply(Math, height);
// mins.push(min);
// size = [];
// size[0] = Math.max.apply(Math, width) - min[0];
// size[1] = Math.max.apply(Math, height) - min[1];
// sizes.push(size);
// }
tdata = [];
tlabels = [];
tsizes = [];
tmins = [];
for (var key in ckp_test_scaled_features) {
tdata.push(ckp_test_scaled_features[key]);
tlabels.push(ckp_test_labels[key]);
width = [];
height = [];
for (var i = 0; i < ckp_test_scaled_features[key].length; i++) {
width.push(ckp_test_scaled_features[key][i][0]);
height.push(ckp_test_scaled_features[key][i][1]);
}
tmin = [];
tmin[0] = Math.min.apply(Math, width);
tmin[1] = Math.min.apply(Math, height);
tmins.push(tmin);
tsize = [];
tsize[0] = Math.max.apply(Math, width) - tmin[0];
tsize[1] = Math.max.apply(Math, height) - tmin[1];
tsizes.push(tsize);
}
// for (var key2 in jaffe_test_features) {
// tdata.push(jaffe_test_features[key2]);
// tlabels.push(jaffe_test_labels[key2]);
// width = [];
// height = [];
// for (var i = 0; i < jaffe_test_features[key2].length; i++) {
// width.push(jaffe_test_features[key2][i][0]);
// height.push(jaffe_test_features[key2][i][1]);
// }
// tmin = [];
// tmin[0] = Math.min.apply(Math, width);
// tmin[1] = Math.min.apply(Math, height);
// tmins.push(tmin);
// tsize = [];
// tsize[0] = Math.max.apply(Math, width) - tmin[0];
// tsize[1] = Math.max.apply(Math, height) - tmin[1];
// tsizes.push(tsize);
// }
}
set_up_data();
var use_validation_data = true;
var temp_arr;
var sample_training_instance = function() {
// Randomly pick a image to train
var k = Math.floor(Math.random()*Object.keys(data).length)
var x = new convnetjs.Vol(2, 1, 71, 0.0);
for (var ix = 0; ix < 71; ix++) {
x.set(0, 0, ix, (data[k][ix][0] - mins[k][0]) / sizes[k][0]);
x.set(1, 0, ix, (data[k][ix][1] - mins[k][1]) / sizes[k][1]);
}
var isValidationData = use_validation_data && k%10===0 ? true : false;
if (counter < 1) {
console.log(mins[k]);
console.log(sizes[k]);
temp_arr = data[k];
temp_x = x;
console.log(data[k]);
console.log(x);
counter++;
}
return {x:x, label:labels[k], isValidationData:isValidationData};
}
// sample a random testing instance
var sample_test_instance = function() {
// Randomly pick a image to train
var k = Math.floor(Math.random()*Object.keys(tdata).length)
var x = new convnetjs.Vol(2, 1, 71, 0.0);
for (var ix = 0; ix < 71; ix++) {
x.set(0, 0, ix, (tdata[k][ix][0] - tmins[k][0]) / tsizes[k][0]);
x.set(1, 0, ix, (tdata[k][ix][1] - tmins[k][1]) / tsizes[k][1]);
}
console.log("Testing on image " + k);
var isValidationData = use_validation_data && k%10===0 ? true : false;
return {x:x, label:tlabels[k], isValidationData:isValidationData};
}
// int main
$(window).load(function() {
$("#newnet").val(t);
eval($("#newnet").val());
update_net_param_display();
start_fun();
});
var start_fun = function() {
setInterval(load_and_step, 0); // lets go!
}
// ------------------------
// END MNIST SPECIFIC STUFF
// ------------------------
var maxmin = cnnutil.maxmin;
var f2t = cnnutil.f2t;
// elt is the element to add all the canvas activation drawings into
// A is the Vol() to use
// scale is a multiplier to make the visualizations larger. Make higher for larger pictures
// if grads is true then gradients are used instead
var draw_activations = function(elt, A, scale, grads) {
var s = scale || 2; // scale
var draw_grads = false;
if(typeof(grads) !== 'undefined') draw_grads = grads;
// get max and min activation to scale the maps automatically
var w = draw_grads ? A.dw : A.w;
var mm = maxmin(w);
// create the canvas elements, draw and add to DOM
for(var d=0;d<A.depth;d++) {
var canv = document.createElement('canvas');
canv.className = 'actmap';
var W = A.sx * s;
var H = A.sy * s;
canv.width = W;
canv.height = H;
var ctx = canv.getContext('2d');
var g = ctx.createImageData(W, H);
for(var x=0;x<A.sx;x++) {
for(var y=0;y<A.sy;y++) {
if(draw_grads) {
var dval = Math.floor((A.get_grad(x,y,d)-mm.minv)/mm.dv*255);
} else {
var dval = Math.floor((A.get(x,y,d)-mm.minv)/mm.dv*255);
}
for(var dx=0;dx<s;dx++) {
for(var dy=0;dy<s;dy++) {
var pp = ((W * (y*s+dy)) + (dx + x*s)) * 4;
for(var i=0;i<3;i++) { g.data[pp + i] = dval; } // rgb
g.data[pp+3] = 255; // alpha channel
}
}
}
}
ctx.putImageData(g, 0, 0);
elt.appendChild(canv);
}
}
var visualize_activations = function(net, elt) {
// clear the element
elt.innerHTML = "";
// show activations in each layer
var N = net.layers.length;
for(var i=0;i<N;i++) {
var L = net.layers[i];
var layer_div = document.createElement('div');
// visualize activations
var activations_div = document.createElement('div');
activations_div.appendChild(document.createTextNode('Activations:'));
activations_div.appendChild(document.createElement('br'));
activations_div.className = 'layer_act';
var scale = 2;
if(L.layer_type==='softmax' || L.layer_type==='fc') scale = 10; // for softmax
draw_activations(activations_div, L.out_act, scale);
// visualize data gradients
if(L.layer_type !== 'softmax') {
var grad_div = document.createElement('div');
grad_div.appendChild(document.createTextNode('Data Gradients:'));
grad_div.appendChild(document.createElement('br'));
grad_div.className = 'layer_grad';
var scale = 2;
if(L.layer_type==='softmax' || L.layer_type==='fc') scale = 10; // for softmax
draw_activations(grad_div, L.out_act, scale, true);
activations_div.appendChild(grad_div);
}
// visualize filters if they are of reasonable size
if(L.layer_type === 'conv') {
var filters_div = document.createElement('div');
if(L.filters[0].sx>3) {
// actual weights
filters_div.appendChild(document.createTextNode('Weights:'));
filters_div.appendChild(document.createElement('br'));
for(var j=0;j<L.filters.length;j++) {
filters_div.appendChild(document.createTextNode('('));
draw_activations(filters_div, L.filters[j], 2);
filters_div.appendChild(document.createTextNode(')'));
}
// gradients
filters_div.appendChild(document.createElement('br'));
filters_div.appendChild(document.createTextNode('Weight Gradients:'));
filters_div.appendChild(document.createElement('br'));
for(var j=0;j<L.filters.length;j++) {
filters_div.appendChild(document.createTextNode('('));
draw_activations(filters_div, L.filters[j], 2, true);
filters_div.appendChild(document.createTextNode(')'));
}
} else {
filters_div.appendChild(document.createTextNode('Weights hidden, too small'));
}
activations_div.appendChild(filters_div);
}
layer_div.appendChild(activations_div);
// print some stats on left of the layer
layer_div.className = 'layer ' + 'lt' + L.layer_type;
var title_div = document.createElement('div');
title_div.className = 'ltitle'
var t = L.layer_type + ' (' + L.out_sx + 'x' + L.out_sy + 'x' + L.out_depth + ')';
title_div.appendChild(document.createTextNode(t));
layer_div.appendChild(title_div);
if(L.layer_type==='conv') {
var t = 'filter size ' + L.filters[0].sx + 'x' + L.filters[0].sy + 'x' + L.filters[0].depth + ', stride ' + L.stride;
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
}
if(L.layer_type==='pool') {
var t = 'pooling size ' + L.sx + 'x' + L.sy + ', stride ' + L.stride;
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
}
// find min, max activations and display them
var mma = maxmin(L.out_act.w);
var t = 'max activation: ' + f2t(mma.maxv) + ', min: ' + f2t(mma.minv);
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
var mma = maxmin(L.out_act.dw);
var t = 'max gradient: ' + f2t(mma.maxv) + ', min: ' + f2t(mma.minv);
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
// number of parameters
if(L.layer_type==='conv') {
var tot_params = L.sx*L.sy*L.in_depth*L.filters.length + L.filters.length;
var t = 'parameters: ' + L.filters.length + 'x' + L.sx + 'x' + L.sy + 'x' + L.in_depth + '+' + L.filters.length + ' = ' + tot_params;
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
}
if(L.layer_type==='fc') {
var tot_params = L.num_inputs*L.filters.length + L.filters.length;
var t = 'parameters: ' + L.filters.length + 'x' + L.num_inputs + '+' + L.filters.length + ' = ' + tot_params;
layer_div.appendChild(document.createTextNode(t));
layer_div.appendChild(document.createElement('br'));
}
// css madness needed here...
var clear = document.createElement('div');
clear.className = 'clear';
layer_div.appendChild(clear);
elt.appendChild(layer_div);
}
}
// loads a training image and trains on it with the network
var paused = false;
var load_and_step = function() {
if(paused) return;
var sample = sample_training_instance();
step(sample); // process this image
}
// evaluate current network on test set
var test_predict = function() {
var num_classes = net.layers[net.layers.length-1].out_depth;
document.getElementById('testset_acc').innerHTML = '';
// grab a random test image
for(num=0;num<50;num++) {
var sample = sample_test_instance();
var y = sample.label; // ground truth label
// forward prop it through the network
var aavg = new convnetjs.Vol(2,1,num_classes,0.0);
// ensures we always have a list, regardless if above returns single item or list
var xs = [].concat(sample.x);
var n = xs.length;
for(var i=0;i<n;i++) {
var a = net.forward(xs[i]);
aavg.addFrom(a);
}
var preds = [];
for(var k=0;k<aavg.w.length;k++) { preds.push({k:k,p:aavg.w[k]}); }
preds.sort(function(a,b){return a.p<b.p ? 1:-1;});
var div = document.createElement('div');
div.className = 'testdiv';
// draw the image into a canvas
draw_activations(div, xs[0], 2); // draw Vol into canv
// add predictions
var probsdiv = document.createElement('div');
div.className = 'probsdiv';
var t = '';
for(var k=0;k<3;k++) {
var col = preds[k].k===y ? 'rgb(85,187,85)' : 'rgb(187,85,85)';
t += '<div class=\"pp\" style=\"width:' + Math.floor(preds[k].p/n*100) + 'px; margin-left: 60px; background-color:' + col + ';\">' + classes_txt[preds[k].k] + '</div>'
}
probsdiv.innerHTML = t;
div.appendChild(probsdiv);
// add it into DOM
$("#testset_acc").append(div).fadeIn(1000);
}
}
var lossGraph = new cnnvis.Graph();
var xLossWindow = new cnnutil.Window(100);
var wLossWindow = new cnnutil.Window(100);
var trainAccWindow = new cnnutil.Window(100);
var valAccWindow = new cnnutil.Window(100);
var step_num = 0;
var step = function(sample) {
var x = sample.x;
var y = sample.label;
if(sample.isValidationData) {
// use x to build our estimate of validation error
net.forward(x);
var yhat = net.getPrediction();
var val_acc = yhat === y ? 1.0 : 0.0;
valAccWindow.add(val_acc);
return; // get out
}
// train on it with network
var stats = trainer.train(x, y);
var lossx = stats.cost_loss;
var lossw = stats.l2_decay_loss;
// keep track of stats such as the average training error and loss
var yhat = net.getPrediction();
var train_acc = yhat === y ? 1.0 : 0.0;
xLossWindow.add(lossx);
wLossWindow.add(lossw);
trainAccWindow.add(train_acc);
// visualize training status
var train_elt = document.getElementById("trainstats");
train_elt.innerHTML = '';
var t = 'Forward time per example: ' + stats.fwd_time + 'ms';
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'Backprop time per example: ' + stats.bwd_time + 'ms';
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'Classification loss: ' + f2t(xLossWindow.get_average());
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'L2 Weight decay loss: ' + f2t(wLossWindow.get_average());
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'Training accuracy: ' + f2t(trainAccWindow.get_average());
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'Validation accuracy: ' + f2t(valAccWindow.get_average());
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
var t = 'Examples seen: ' + step_num;
train_elt.appendChild(document.createTextNode(t));
train_elt.appendChild(document.createElement('br'));
// visualize activations
if(step_num % 100 === 0) {
var vis_elt = document.getElementById("visnet");
visualize_activations(net, vis_elt);
}
// log progress to graph, (full loss)
if(step_num % 200 === 0) {
var xa = xLossWindow.get_average();
var xw = wLossWindow.get_average();
if(xa >= 0 && xw >= 0) { // if they are -1 it means not enough data was accumulated yet for estimates
lossGraph.add(step_num, xa + xw);
lossGraph.drawSelf(document.getElementById("lossgraph"));
}
}
// run prediction on test set
if(step_num % 1000 === 0) {
test_predict();
}
step_num++;
}
// user settings
var change_lr = function() {
trainer.learning_rate = parseFloat(document.getElementById("lr_input").value);
update_net_param_display();
}
var change_momentum = function() {
trainer.momentum = parseFloat(document.getElementById("momentum_input").value);
update_net_param_display();
}
var change_batch_size = function() {
trainer.batch_size = parseFloat(document.getElementById("batch_size_input").value);
update_net_param_display();
}
var change_decay = function() {
trainer.l2_decay = parseFloat(document.getElementById("decay_input").value);
update_net_param_display();
}
var update_net_param_display = function() {
document.getElementById('lr_input').value = trainer.learning_rate;
document.getElementById('momentum_input').value = trainer.momentum;
document.getElementById('batch_size_input').value = trainer.batch_size;
document.getElementById('decay_input').value = trainer.l2_decay;
}
var toggle_pause = function() {
paused = !paused;
var btn = document.getElementById('buttontp');
if(paused) { btn.value = 'resume' }
else { btn.value = 'pause'; }
}
var dump_json = function() {
document.getElementById("dumpjson").value = JSON.stringify(net.toJSON());
}
var clear_graph = function() {
lossGraph = new cnnvis.Graph(); // reinit graph too
}
var reset_all = function() {
update_net_param_display();
// reinit windows that keep track of val/train accuracies
xLossWindow.reset();
wLossWindow.reset();
trainAccWindow.reset();
valAccWindow.reset();
lossGraph = new cnnvis.Graph(); // reinit graph too
step_num = 0;
}
var load_from_json = function() {
var jsonString = document.getElementById("dumpjson").value;
var json = JSON.parse(jsonString);
net = new convnetjs.Net();
net.fromJSON(json);
reset_all();
}
var change_net = function() {
eval($("#newnet").val());
reset_all();
}
</script>
</head>
<body>
<div id="wrap">
<h2 style="text-align: center;"><a href="http://cs.stanford.edu/people/karpathy/convnetjs/">ConvNetJS</a> MNIST demo</h2>
<h1>Description</h1>
<p>
This demo trains a Convolutional Neural Network on the <a href="http://yann.lecun.com/exdb/mnist/">MNIST digits dataset</a> in your browser, with nothing but Javascript. The dataset is fairly easy and one should expect to get somewhere around 99% accuracy within few minutes. I used <a href="mnist_parse.zip">this python script</a> to parse the <a href="http://deeplearning.net/tutorial/gettingstarted.html">original files</a> into batches of images that can be easily loaded into page DOM with img tags.
</p>
<p>
This network takes a 28x28 MNIST image and crops a random 24x24 window before training on it (this technique is called data augmentation and improves generalization). Similarly to do prediction, 4 random crops are sampled and the probabilities across all crops are averaged to produce final predictions. The network runs at about 5ms for both forward and backward pass on my reasonably decent Ubuntu+Chrome machine.
</p>
<p>
By default, in this demo we're using Adadelta which is one of per-parameter adaptive step size methods, so we don't have to worry about changing learning rates or momentum over time. However, I still included the text fields for changing these if you'd like to play around with SGD+Momentum trainer.
</p>
<p>Report questions/bugs/suggestions to <a href="https://twitter.com/karpathy">@karpathy</a>.</p>
<h1>Training Stats</h1>
<div class="divsec" style="270px;">
<div class="secpart">
<input id="buttontp" type="submit" value="pause" onclick="toggle_pause();" style="width: 100px; height:30px; background-color: #FCC;"/>
<div id="trainstats"></div>
<div id="controls">
Learning rate: <input name="lri" type="text" maxlength="20" id="lr_input"/>
<input id="buttonlr" type="submit" value="change" onclick="change_lr();"/>
<br />
Momentum: <input name="momi" type="text" maxlength="20" id="momentum_input"/>
<input id="buttonmom" type="submit" value="change" onclick="change_momentum();"/>
<br />
Batch size: <input name="bsi" type="text" maxlength="20" id="batch_size_input"/>
<input id="buttonbs" type="submit" value="change" onclick="change_batch_size();"/>
<br />
Weight decay: <input name="wdi" type="text" maxlength="20" id="decay_input"/>
<input id="buttonwd" type="submit" value="change" onclick="change_decay();"/>
</div>
<input id="buttondj" type="submit" value="save network snapshot as JSON" onclick="dump_json();"/><br />
<input id="buttonlfj" type="submit" value="init network from JSON snapshot" onclick="load_from_json();"/><br />
<textarea id="dumpjson"></textarea>
</div>
<div class="secpart">
<div>
Loss:<br />
<canvas id="lossgraph">
</canvas>
<br />
<input id="buttoncg" type="submit" value="clear graph" onclick="clear_graph();"/>
</div>
</div>
<div style="clear:both;"></div>
</div>
<h1>Instantiate a Network and Trainer</h1>
<div>
<textarea id="newnet" style="width:100%; height:200px;"></textarea><br />
<input id="buttonnn" type="submit" value="change network" onclick="change_net();" style="width:200px;height:30px;"/>
</div>
<div class="divsec">
<h1>Network Visualization</h1>
<div id="visnet"></div>
</div>
<div class="divsec">
<h1>Example predictions on Test set</h1>
<div id="testset_acc"></div>
</div>
</div>
</body>
</html>