Repository navigation
Expand file tree
/
Copy pathfd_ebm.html
More file actions
1132 lines (971 loc) · 37.5 KB
/
Copy pathfd_ebm.html
File metadata and controls
1132 lines (971 loc) · 37.5 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Diffusion Score Matching - Fixed</title>
<style>
:root {
--bg: #f8f9fa;
--panel-bg: #ffffff;
--text: #2c3e50;
--border: #e9ecef;
--accent: #6200ea;
--accent-sec: #00bfa5;
}
body {
margin: 0;
padding: 0;
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
background-color: var(--bg);
color: var(--text);
display: flex;
height: 100vh;
overflow: hidden;
}
#sidebar {
width: 360px;
background: var(--panel-bg);
border-right: 1px solid var(--border);
padding: 20px;
display: flex;
flex-direction: column;
gap: 12px;
overflow-y: auto;
box-shadow: 2px 0 10px rgba(0,0,0,0.03);
z-index: 10;
}
h1, h3 { margin: 0 0 5px 0; }
h1 { font-size: 1.1rem; color: #1a1a1a; font-weight: 700; }
h3 { font-size: 0.75rem; text-transform: uppercase; letter-spacing: 0.5px; color: #6c757d; margin-top: 10px; border-bottom: 1px solid #eee; padding-bottom: 5px;}
p { font-size: 0.8rem; color: #666; margin: 0; line-height: 1.4; }
.control-group { display: flex; flex-direction: column; gap: 4px; }
label { font-size: 0.75rem; font-weight: 600; display: flex; justify-content: space-between; align-items: center;}
input[type=range] { width: 100%; margin: 6px 0;}
.button-row { display: flex; gap: 8px; }
button {
flex: 1;
padding: 8px;
border: 1px solid transparent;
border-radius: 6px;
cursor: pointer;
font-weight: 600;
font-size: 0.8rem;
transition: all 0.2s;
background-color: #f1f3f5;
color: #495057;
}
button:hover { opacity: 0.9; transform: translateY(-1px); border-color: #dee2e6; }
button:active { transform: translateY(0); }
button:disabled { opacity: 0.5; cursor: not-allowed; transform: none; }
.btn-primary { background-color: var(--accent); color: white; }
.btn-danger { background-color: #fa5252; color: white; }
.btn-success { background-color: #40c057; color: white; }
.mode-toggle {
display: flex;
background: #f1f3f5;
border-radius: 6px;
padding: 3px;
}
.mode-toggle button {
background: transparent;
color: #868e96;
box-shadow: none;
font-size: 0.75rem;
border: none;
padding: 6px;
}
.mode-toggle button.active {
background: white;
color: var(--accent);
box-shadow: 0 1px 3px rgba(0,0,0,0.1);
}
.mode-toggle button.active.direct { color: var(--accent-sec); }
#main {
flex: 1;
position: relative;
display: flex;
justify-content: center;
align-items: center;
background-color: #f4f4f4;
user-select: none;
}
canvas {
background: white;
box-shadow: 0 0 30px rgba(0,0,0,0.05);
border-radius: 8px;
cursor: crosshair;
}
#loss-container {
margin-top: 5px;
border: 1px solid #e9ecef;
border-radius: 4px;
background: #fff;
padding: 5px;
height: 70px;
}
#loss-canvas { width: 100%; height: 100%; display: block; }
#stats { font-family: 'SF Mono', 'Menlo', monospace; font-size: 0.7rem; color: #555; margin-top: 2px; }
.legend {
position: absolute;
top: 20px;
left: 20px;
background: rgba(255, 255, 255, 0.95);
padding: 12px;
border: 1px solid #e9ecef;
border-radius: 8px;
pointer-events: none;
font-size: 0.75rem;
box-shadow: 0 4px 12px rgba(0,0,0,0.08);
}
.legend-item { display: flex; align-items: center; gap: 8px; margin-bottom: 5px; }
.dot { width: 8px; height: 8px; border-radius: 50%; }
.line { width: 20px; height: 2px; }
#viz-controls {
position: absolute;
bottom: 30px;
left: 50%;
transform: translateX(-50%);
background: rgba(255,255,255,0.98);
padding: 10px 20px;
border-radius: 30px;
box-shadow: 0 4px 20px rgba(0,0,0,0.15);
display: flex;
flex-direction: column;
gap: 10px;
align-items: center;
z-index: 100;
border: 1px solid rgba(0,0,0,0.05);
min-width: 240px;
}
.viz-row { display: flex; gap: 10px; align-items: center; width: 100%; justify-content: space-between; }
.viz-hidden { display: none !important; }
</style>
</head>
<body>
<div id="sidebar">
<div>
<h1>Diffusion Score Matching</h1>
<p><strong>Setup:</strong> VP-SDE.<br>
t=0 (Data) ← Diffusion → t=1 (Noise)</p>
</div>
<h3>Data Distribution (t=0)</h3>
<div class="button-row">
<button id="btn-clear-target" style="background: #ff922b; color: white;">Clear / Draw</button>
<button id="btn-reset-target">Spiral Preset</button>
</div>
<p style="margin-top:4px; font-style:italic;" id="draw-hint">Current: Spiral</p>
<h3>Model Formulation</h3>
<div class="mode-toggle">
<button id="mode-direct" onclick="setMode('direct')">Direct Score</button>
<button id="mode-energy" class="active" onclick="setMode('energy')">Energy (FD)</button>
</div>
<div id="desc-direct" style="display:none; margin-top:8px;">
<p><strong>Direct Score Matching</strong><br>
Net outputs vector $\mathbf{s}$.<br>
Loss weighted by $\sigma_t^2$ (predicting $\epsilon$).</p>
</div>
<div id="desc-energy" style="margin-top:8px;">
<p><strong>Energy-Based Model</strong><br>
Net outputs scalar $E$. Score $-\nabla E$.<br>
FD Estimator with $\sigma_t^2$ variance reduction.</p>
</div>
<h3>Hyperparameters</h3>
<div class="control-group">
<label>Width: <span id="val-width">64</span></label>
<input type="range" id="inp-width" min="32" max="128" step="16" value="64">
</div>
<div id="controls-ebm" class="control-group">
<label>FD Epsilon ($\epsilon$): <span id="val-eps">0.01</span></label>
<input type="range" id="inp-eps" min="-3" max="-1" step="0.1" value="-2">
</div>
<div class="control-group">
<label>Learning Rate: <span id="val-lr">0.001</span></label>
<input type="range" id="inp-lr" min="-4.5" max="-2" step="0.1" value="-3">
</div>
<button id="btn-reset-model" style="margin-top:5px;">Re-Initialize Weights</button>
<h3>Training Status</h3>
<div class="button-row">
<button id="btn-train" class="btn-primary">Train</button>
<button id="btn-stop" class="btn-danger" disabled>Stop</button>
</div>
<div id="stats">Step: 0 | Loss: 0.000</div>
<div id="loss-container">
<canvas id="loss-canvas" width="300" height="70"></canvas>
</div>
<h3>Generation</h3>
<div class="control-group">
<label>Count: <span id="val-gen-count">200</span></label>
<input type="range" id="inp-gen-count" min="50" max="500" step="50" value="200">
</div>
<div class="button-row">
<button id="btn-gen" class="btn-success">Generate</button>
<button id="btn-clear-gen">Clear</button>
</div>
<div style="margin-top: 15px; border-top: 1px solid #eee; padding-top: 10px;">
<button id="btn-toggle-viz" style="width:100%; background:#343a40; color:white;">Toggle Energy Viz</button>
</div>
<div class="control-group" style="margin-top:5px;">
<label><input type="checkbox" id="chk-show-true" checked> Show Data (t=0)</label>
</div>
</div>
<div id="main">
<canvas id="canvas" width="800" height="800"></canvas>
<div class="legend">
<div class="legend-item"><div class="dot" style="background:#ffcc00"></div> Data (t=0)</div>
<div class="legend-item"><div class="dot" style="background:#40c057"></div> Generated</div>
<div class="legend-item"><div class="line" style="background:#ff8888"></div> Trajectory</div>
</div>
<div id="viz-controls" class="viz-hidden">
<div class="viz-row">
<span style="font-weight:bold; font-size:0.85rem; color:#555;">Viz Time $t$:</span>
<input type="range" id="inp-viz-time" min="0" max="1" step="0.01" value="0.0" style="flex:1;">
<span id="val-viz-time" style="width:30px; font-family:monospace;">0.00</span>
</div>
<div class="viz-row">
<span style="font-weight:bold; font-size:0.85rem; color:#555;">Grid:</span>
<input type="range" id="inp-viz-res" min="30" max="150" step="10" value="40" style="flex:1;">
<span id="val-viz-res" style="width:30px; font-family:monospace;">40</span>
</div>
<div class="viz-row" style="background: #f1f3f5; padding: 4px; border-radius: 4px; justify-content:center;">
<button id="btn-viz-prob" class="active" style="flex:1; background:white; color:#333; box-shadow:0 1px 2px rgba(0,0,0,0.1);">Probability</button>
<button id="btn-viz-energy" style="flex:1; background:transparent; color:#666;">Energy</button>
</div>
</div>
</div>
<script>
/*
TINY DEEP LEARNING LIBRARY
*/
class Tensor {
constructor(rows, cols, data = null) {
this.rows = rows;
this.cols = cols;
this.length = rows * cols;
this.data = data ? data : new Float32Array(this.length);
this.grad = new Float32Array(this.length);
this.requiresGrad = false;
}
static randn(rows, cols, scale = 1.0) {
const t = new Tensor(rows, cols);
for(let i=0; i<t.length; i++) {
const u = 1 - Math.random();
const v = Math.random();
const z = Math.sqrt(-2.0 * Math.log(u)) * Math.cos(2.0 * Math.PI * v);
t.data[i] = z * scale;
}
return t;
}
static zeros(rows, cols) { return new Tensor(rows, cols); }
zeroGrad() { this.grad.fill(0); }
}
function matmul(A, B) {
const C = new Tensor(A.rows, B.cols);
for (let i = 0; i < A.rows; i++) {
for (let k = 0; k < A.cols; k++) {
const aVal = A.data[i * A.cols + k];
if (aVal === 0) continue;
for (let j = 0; j < B.cols; j++) {
C.data[i * B.cols + j] += aVal * B.data[k * B.cols + j];
}
}
}
return C;
}
class Linear {
constructor(inDim, outDim) {
const scale = Math.sqrt(2.0 / (inDim + outDim));
this.W = Tensor.randn(inDim, outDim, scale);
this.b = Tensor.zeros(1, outDim);
this.inDim = inDim;
this.outDim = outDim;
this.input = null;
}
forward(input) {
this.input = input;
const output = matmul(input, this.W);
for (let i = 0; i < output.rows; i++) {
for (let j = 0; j < output.cols; j++) {
output.data[i * output.cols + j] += this.b.data[j];
}
}
return output;
}
backward(gradOutput) {
for (let i = 0; i < gradOutput.rows; i++) {
for (let j = 0; j < gradOutput.cols; j++) {
const g = gradOutput.grad[i * gradOutput.cols + j];
this.b.grad[j] += g;
}
}
const dInput = new Tensor(this.input.rows, this.input.cols);
for (let i = 0; i < this.input.rows; i++) {
for (let j = 0; j < this.outDim; j++) {
const g = gradOutput.grad[i * this.outDim + j];
const inOffset = i * this.inDim;
for (let k = 0; k < this.inDim; k++) {
this.W.grad[k * this.outDim + j] += this.input.data[inOffset + k] * g;
}
for (let k = 0; k < this.inDim; k++) {
dInput.grad[inOffset + k] += g * this.W.data[k * this.outDim + j];
}
}
}
return dInput;
}
params() { return [this.W, this.b]; }
}
class Swish {
constructor() { this.input = null; }
forward(input) {
this.input = input;
const out = new Tensor(input.rows, input.cols);
for(let i=0; i<input.length; i++) {
const x = input.data[i];
out.data[i] = x / (1.0 + Math.exp(-x));
}
return out;
}
backward(gradOutput) {
if (!this.input) return null;
const dInput = new Tensor(this.input.rows, this.input.cols);
for(let i=0; i<this.input.length; i++) {
const x = this.input.data[i];
const sig = 1.0 / (1.0 + Math.exp(-x));
const swish = x * sig;
dInput.grad[i] = gradOutput.grad[i] * (swish + sig * (1 - swish));
}
return dInput;
}
}
class MLP {
constructor(layersConfig) {
this.layers = [];
for(let i=0; i<layersConfig.length - 1; i++) {
this.layers.push(new Linear(layersConfig[i], layersConfig[i+1]));
if (i < layersConfig.length - 2) {
this.layers.push(new Swish());
}
}
}
forward(x) {
let out = x;
for(let layer of this.layers) {
out = layer.forward(out);
}
return out;
}
backward(lossGrad) {
let grad = lossGrad;
for(let i=this.layers.length-1; i>=0; i--) {
grad = this.layers[i].backward(grad);
if (!grad) return null;
}
return grad;
}
zeroGrads() {
for(let layer of this.layers) {
if(layer.params) {
for(let p of layer.params()) p.zeroGrad();
}
}
}
params() {
let p = [];
for(let layer of this.layers) {
if(layer.params) p.push(...layer.params());
}
return p;
}
}
class Adam {
constructor(params, lr=0.001, beta1=0.9, beta2=0.999) {
this.params = params;
this.lr = lr;
this.beta1 = beta1;
this.beta2 = beta2;
this.t = 0;
this.m = params.map(p => new Float32Array(p.length));
this.v = params.map(p => new Float32Array(p.length));
}
step() {
this.t++;
for(let i=0; i<this.params.length; i++) {
const p = this.params[i];
const grad = p.grad;
const m = this.m[i];
const v = this.v[i];
for(let j=0; j<p.length; j++) {
let g = grad[j];
// Clipping
if (g > 5.0) g = 5.0; if (g < -5.0) g = -5.0;
m[j] = this.beta1 * m[j] + (1 - this.beta1) * g;
v[j] = this.beta2 * v[j] + (1 - this.beta2) * g * g;
const mHat = m[j] / (1 - Math.pow(this.beta1, this.t));
const vHat = v[j] / (1 - Math.pow(this.beta2, this.t));
p.data[j] -= this.lr * mHat / (Math.sqrt(vHat) + 1e-8);
grad[j] = 0;
}
}
}
}
/*
APPLICATION LOGIC
*/
// State
let model, optimizer;
let mode = 'energy';
let isTraining = false;
let stepCount = 0;
let lossHistory = [];
let generatedSamples = [];
let animationId;
let vizMode = false;
let vizTime = 0.0;
let vizType = 'probability';
let vizRes = 40; // Default resolution
// Config
const BATCH_SIZE = 64;
const CANVAS_SIZE = 800;
const VIEW_SCALE = 120;
const CENTER_X = CANVAS_SIZE/2;
const CENTER_Y = CANVAS_SIZE/2;
// DOM
const canvas = document.getElementById('canvas');
const ctx = canvas.getContext('2d');
const lossCanvas = document.getElementById('loss-canvas');
const lossCtx = lossCanvas.getContext('2d');
const elStep = document.getElementById('stats');
// Inputs
const inpWidth = document.getElementById('inp-width');
const inpLr = document.getElementById('inp-lr');
const inpEps = document.getElementById('inp-eps');
const inpGenCount = document.getElementById('inp-gen-count');
const inpVizTime = document.getElementById('inp-viz-time');
const inpVizRes = document.getElementById('inp-viz-res');
// Buttons
const btnTrain = document.getElementById('btn-train');
const btnStop = document.getElementById('btn-stop');
const btnGen = document.getElementById('btn-gen');
const btnClearGen = document.getElementById('btn-clear-gen');
const btnReset = document.getElementById('btn-reset-model');
const btnToggleViz = document.getElementById('btn-toggle-viz');
const btnClearTarget = document.getElementById('btn-clear-target');
const btnResetTarget = document.getElementById('btn-reset-target');
const btnVizProb = document.getElementById('btn-viz-prob');
const btnVizEnergy = document.getElementById('btn-viz-energy');
// Data State
let customTargetPoints = [];
let useCustomTarget = false;
// Utils
const worldToScreen = (x, y) => [CENTER_X + x * VIEW_SCALE, CENTER_Y - y * VIEW_SCALE];
const screenToWorld = (sx, sy) => [(sx - CENTER_X) / VIEW_SCALE, (CENTER_Y - sy) / VIEW_SCALE];
// --- Data Distribution (t=0) ---
function sampleData(n) {
const data = new Float32Array(n * 2);
if (useCustomTarget && customTargetPoints.length > 0) {
const numPoints = customTargetPoints.length / 2;
for(let i=0; i<n; i++) {
const idx = Math.floor(Math.random() * numPoints);
data[i*2] = customTargetPoints[idx*2] + (Math.random()-0.5)*0.08;
data[i*2+1] = customTargetPoints[idx*2+1] + (Math.random()-0.5)*0.08;
}
} else {
// Spiral
for(let i=0; i<n; i++) {
const rNorm = Math.sqrt(Math.random());
const theta = rNorm * 3 * Math.PI;
const r = 0.2 + 2.5 * (theta / (3*Math.PI));
const x = -r * Math.cos(theta + 1);
const y = -r * Math.sin(theta + 1);
data[i*2] = x + (Math.random()-0.5)*0.08;
data[i*2+1] = y + (Math.random()-0.5)*0.08;
}
}
return data;
}
// --- Painting ---
function addSprayPoints(ex, ey) {
const rect = canvas.getBoundingClientRect();
const sx = (ex - rect.left) * (canvas.width / rect.width);
const sy = (ey - rect.top) * (canvas.height / rect.height);
const [wx, wy] = screenToWorld(sx, sy);
const spread = 0.15;
for(let k=0; k<2; k++) {
const r = Math.sqrt(Math.random()) * spread;
const theta = Math.random() * 2 * Math.PI;
customTargetPoints.push(wx + r * Math.cos(theta));
customTargetPoints.push(wy + r * Math.sin(theta));
}
}
let isDrawing = false;
canvas.addEventListener('mousedown', (e) => { if(useCustomTarget) { isDrawing=true; addSprayPoints(e.clientX, e.clientY); }});
canvas.addEventListener('mousemove', (e) => { if(isDrawing) addSprayPoints(e.clientX, e.clientY); });
window.addEventListener('mouseup', () => isDrawing = false);
btnClearTarget.addEventListener('click', () => {
useCustomTarget = true;
customTargetPoints = [];
document.getElementById('draw-hint').innerText = "Mode: Custom. Draw on canvas.";
});
btnResetTarget.addEventListener('click', () => {
useCustomTarget = false;
document.getElementById('draw-hint').innerText = "Mode: Spiral Preset";
});
// --- Diffusion Schedule (VP) ---
function getAlphaSigma(t) {
// Cosine Schedule: t=0 (Data), t=1 (Noise)
const arg = t * Math.PI / 2.0;
const alpha = Math.cos(arg);
const sigma = Math.sin(arg);
return { a: alpha, s: sigma };
}
// --- Model Init ---
function setMode(m) {
mode = m;
document.getElementById('mode-direct').classList.toggle('active', m==='direct');
document.getElementById('mode-energy').classList.toggle('active', m==='energy');
document.getElementById('desc-direct').style.display = m==='direct'?'block':'none';
document.getElementById('desc-energy').style.display = m==='energy'?'block':'none';
document.getElementById('controls-ebm').style.opacity = m==='energy'?1:0.5;
document.getElementById('inp-eps').disabled = (m!=='energy');
document.documentElement.style.setProperty('--accent', m==='energy' ? '#6200ea' : '#00bfa5');
document.documentElement.style.setProperty('--accent-sec', m==='energy' ? '#00bfa5' : '#6200ea');
initModel();
}
function initModel() {
const width = parseInt(inpWidth.value);
const lr = Math.pow(10, parseFloat(inpLr.value));
const outDim = (mode === 'direct') ? 2 : 1;
let config = [3, width, width, width, outDim];
model = new MLP(config);
optimizer = new Adam(model.params(), lr);
stepCount = 0;
lossHistory = [];
lossCtx.clearRect(0,0, lossCanvas.width, lossCanvas.height);
updateStats(0, 0);
}
// --- Training Steps ---
function trainStep() {
const x0_flat = sampleData(BATCH_SIZE);
const x0 = new Tensor(BATCH_SIZE, 2, x0_flat);
const noise = Tensor.randn(BATCH_SIZE, 2);
const xt = new Tensor(BATCH_SIZE, 3);
const weights = new Float32Array(BATCH_SIZE);
for(let i=0; i<BATCH_SIZE; i++) {
// Sample t avoiding exact 0
const t = Math.random() * 0.999 + 0.001;
const {a, s} = getAlphaSigma(t);
const x0x = x0.data[i*2]; const x0y = x0.data[i*2+1];
const nx = noise.data[i*2]; const ny = noise.data[i*2+1];
const xtx = a * x0x + s * nx;
const xty = a * x0y + s * ny;
xt.data[i*3+0] = xtx;
xt.data[i*3+1] = xty;
xt.data[i*3+2] = t;
// Weight by sigma^2 (like DDPM predicting epsilon)
weights[i] = s * s;
}
if (mode === 'direct') {
// Direct Score Matching
xt.requiresGrad = true;
const pred = model.forward(xt);
let lossSum = 0;
const gradOut = new Tensor(BATCH_SIZE, 2);
// FIX: loop over BATCH_SIZE, not pred.length
for(let i=0; i<BATCH_SIZE; i++) {
// Loss = || sigma * pred + noise ||^2
const w = weights[i];
const sigma = Math.sqrt(w);
const nx = noise.data[i*2];
const ny = noise.data[i*2+1];
const px = pred.data[i*2];
const py = pred.data[i*2+1];
const termX = sigma * px + nx;
const termY = sigma * py + ny;
lossSum += (termX*termX + termY*termY);
gradOut.grad[i*2] = (termX * sigma) / BATCH_SIZE;
gradOut.grad[i*2+1] = (termY * sigma) / BATCH_SIZE;
}
model.zeroGrads();
model.backward(gradOut);
optimizer.step();
recordLoss(lossSum / (2*BATCH_SIZE));
} else {
// Energy FD Estimator
xt.requiresGrad = true;
const E = model.forward(xt);
model.zeroGrads();
const dE = new Tensor(BATCH_SIZE, 1);
dE.grad.fill(1.0);
const dXt = model.backward(dE);
// r = -grad E + noise/sigma
const r = new Float32Array(BATCH_SIZE * 2);
for(let i=0; i<BATCH_SIZE; i++) {
const gx = dXt.grad[i*3+0]; const gy = dXt.grad[i*3+1];
const sigma = Math.sqrt(weights[i]);
const sx = -gx; const sy = -gy;
const tx = -noise.data[i*2] / sigma;
const ty = -noise.data[i*2+1] / sigma;
r[i*2+0] = sx - tx;
r[i*2+1] = sy - ty;
}
const epsBase = Math.pow(10, parseFloat(inpEps.value));
const xt_pos = new Tensor(BATCH_SIZE, 3);
const xt_neg = new Tensor(BATCH_SIZE, 3);
for(let i=0; i<BATCH_SIZE; i++) {
const rx = r[i*2+0];
const ry = r[i*2+1];
// Adaptive Epsilon step based on residual magnitude
const rMag = Math.sqrt(rx*rx + ry*ry) + 1e-6;
const eps = epsBase / Math.max(1.0, rMag);
xt_pos.data[i*3+0] = xt.data[i*3+0] + eps * rx;
xt_pos.data[i*3+1] = xt.data[i*3+1] + eps * ry;
xt_pos.data[i*3+2] = xt.data[i*3+2];
xt_neg.data[i*3+0] = xt.data[i*3+0] - eps * rx;
xt_neg.data[i*3+1] = xt.data[i*3+1] - eps * ry;
xt_neg.data[i*3+2] = xt.data[i*3+2];
}
model.zeroGrads();
// FD Update: grad = - sigma^2 * (E+ - E-) / (2eps)
// Optimizer subtracts grad.
const baseScale = 1.0 / (2.0 * epsBase * BATCH_SIZE);
const dE_pos = new Tensor(BATCH_SIZE, 1);
const dE_neg = new Tensor(BATCH_SIZE, 1);
for(let i=0; i<BATCH_SIZE; i++) {
const rx = r[i*2+0]; const ry = r[i*2+1];
const rMag = Math.sqrt(rx*rx + ry*ry) + 1e-6;
const eps = epsBase / Math.max(1.0, rMag);
// Re-scale for adaptive epsilon used above
const s = (1.0 / (2.0 * eps * BATCH_SIZE)) * weights[i];
dE_pos.grad[i] = -s;
dE_neg.grad[i] = +s;
}
model.forward(xt_pos); model.backward(dE_pos);
model.forward(xt_neg); model.backward(dE_neg);
optimizer.step();
// Weighted Residual Loss
let weightedLoss = 0;
for(let i=0; i<BATCH_SIZE; i++) {
const r2 = r[i*2]*r[i*2] + r[i*2+1]*r[i*2+1];
weightedLoss += weights[i] * r2;
}
recordLoss(weightedLoss / (2*BATCH_SIZE));
}
}
function recordLoss(l) {
stepCount++;
lossHistory.push(l);
if(lossHistory.length > 300) lossHistory.shift();
if(stepCount % 10 === 0) updateStats(stepCount, l);
}
function updateStats(step, loss) {
elStep.innerText = `Step: ${step} | Loss: ${loss.toFixed(5)}`;
drawLossChart();
}
function drawLossChart() {
const w = lossCanvas.width;
const h = lossCanvas.height;
lossCtx.clearRect(0, 0, w, h);
if(lossHistory.length < 2) return;
lossCtx.beginPath();
lossCtx.strokeStyle = '#adb5bd';
lossCtx.lineWidth = 1.5;
let sorted = [...lossHistory].sort((a,b)=>a-b);
let max = sorted[Math.floor(sorted.length * 0.95)] * 1.2;
if(max < 0.01) max = 0.01;
for(let i=0; i<lossHistory.length; i++) {
const x = (i / (lossHistory.length - 1)) * w;
const val = Math.min(lossHistory[i], max);
const y = h - (val / max) * h * 0.9;
if(i===0) lossCtx.moveTo(x, y); else lossCtx.lineTo(x, y);
}
lossCtx.stroke();
}
// --- Inference ---
function generateSamples() {
const n = parseInt(inpGenCount.value);
const steps = 60;
const dt = 1.0 / steps;
// Start at Noise (t=0.99) to avoid alpha=0 singularity at t=1
const t_start = 0.99;
const currentX = new Float32Array(n * 2);
for(let i=0; i<n; i++) {
const noise = Tensor.randn(1,2).data;
currentX[i*2] = noise[0];
currentX[i*2+1] = noise[1];
}
const trajs = new Array(Math.min(n, 50)).fill(0).map(() => []);
for(let s=0; s<steps; s++) {
const t = t_start - s * dt;
if (t <= 0) break;
const t_next = Math.max(1e-4, t - dt);
const {a: alpha_curr, s: sigma_curr} = getAlphaSigma(t);
const {a: alpha_next, s: sigma_next} = getAlphaSigma(t_next);
const input = new Tensor(n, 3);
for(let i=0; i<n; i++) {
input.data[i*3+0] = currentX[i*2];
input.data[i*3+1] = currentX[i*2+1];
input.data[i*3+2] = t;
}
let scoresX, scoresY;
if (mode === 'direct') {
const out = model.forward(input);
scoresX = new Float32Array(n);
scoresY = new Float32Array(n);
for(let i=0; i<n; i++) {
scoresX[i] = out.data[i*2];
scoresY[i] = out.data[i*2+1];
}
} else {
input.requiresGrad = true;
model.zeroGrads();
model.forward(input);
const dOut = new Tensor(n, 1);
dOut.grad.fill(1.0);
const dIn = model.backward(dOut);
scoresX = new Float32Array(n);
scoresY = new Float32Array(n);
for(let i=0; i<n; i++) {
scoresX[i] = -dIn.grad[i*3+0];
scoresY[i] = -dIn.grad[i*3+1];
}
}
// Ancestral Sampler for VP-SDE
for(let i=0; i<n; i++) {
const sx = scoresX[i];
const sy = scoresY[i];
// eps_pred = -sigma * score
const sigma2 = sigma_curr * sigma_curr;
// Predicted x0
const predX0_x = (currentX[i*2] + sigma2 * sx) / alpha_curr;
const predX0_y = (currentX[i*2+1] + sigma2 * sy) / alpha_curr;
// Next mean (ODE Step)
const eps_pred_x = -sigma_curr * sx;
const eps_pred_y = -sigma_curr * sy;
currentX[i*2] = alpha_next * predX0_x + sigma_next * eps_pred_x;
currentX[i*2+1] = alpha_next * predX0_y + sigma_next * eps_pred_y;
if (i < trajs.length) trajs[i].push({x: currentX[i*2], y: currentX[i*2+1]});
}
}
generatedSamples = [];
for(let i=0; i<n; i++) {
generatedSamples.push({
x: currentX[i*2],
y: currentX[i*2+1],
traj: (i<trajs.length) ? trajs[i] : null
});
}
}
// --- Visualization ---
function drawViz() {
if(!vizMode) return;
const t = vizTime;
const resolution = vizRes; // Use selected resolution
const step = CANVAS_SIZE / resolution;
const inputs = new Tensor(resolution * resolution, 3);
let idx = 0;
for(let py=0; py<resolution; py++) {
for(let px=0; px<resolution; px++) {
const sx = px * step + step/2;
const sy = py * step + step/2;
const [wx, wy] = screenToWorld(sx, sy);
inputs.data[idx*3+0] = wx;
inputs.data[idx*3+1] = wy;
inputs.data[idx*3+2] = t;
idx++;
}
}
let values;
if(mode === 'energy') {
const out = model.forward(inputs);
values = out.data; // Raw Energy
} else {
const out = model.forward(inputs);
values = new Float32Array(resolution*resolution);
for(let i=0; i<resolution*resolution; i++) {
const vx = out.data[i*2];
const vy = out.data[i*2+1];
values[i] = -Math.sqrt(vx*vx + vy*vy); // Proxy for Energy
}
}
// Process values for Display
let displayValues = new Float32Array(values.length);
let minVal = Infinity, maxVal = -Infinity;
if (vizType === 'probability') {
// Probability Mode: P(x) = exp( -(E - E_min) )
// 1. Find Min Energy
let eMin = Infinity;
for(let v of values) if(v < eMin) eMin = v;
// 2. Normalize and Exponentiate
for(let i=0; i<values.length; i++) {
// E - Emin >= 0. exp is in [0, 1]
displayValues[i] = Math.exp(-(values[i] - eMin));
}
// Bounds are naturally 0 to 1
minVal = 0; maxVal = 1;
} else {
// Raw Energy Mode
for(let i=0; i<values.length; i++) {
displayValues[i] = values[i];
if(values[i] < minVal) minVal = values[i];
if(values[i] > maxVal) maxVal = values[i];
}
}
const range = maxVal - minVal || 1;
idx = 0;
for(let py=0; py<resolution; py++) {
for(let px=0; px<resolution; px++) {
const val = displayValues[idx++];
const nVal = (val - minVal) / range;
let r, b;
if (vizType === 'probability') {
// Prob: 0 (Low/Blue) -> 1 (High/Red)
r = Math.floor(nVal * 255);
b = Math.floor((1-nVal) * 255);
} else {
// Energy: Low (Blue) -> High (Red)
r = Math.floor(nVal * 255);
b = Math.floor((1-nVal) * 255);
}
ctx.fillStyle = `rgba(${r}, 0, ${b}, 0.5)`;
ctx.fillRect(px*step, py*step, step, step);
}
}
}
function draw() {
ctx.clearRect(0, 0, canvas.width, canvas.height);
if (vizMode) drawViz();
// Grid
ctx.strokeStyle = '#f1f3f5';
ctx.lineWidth = 1;
ctx.beginPath();
for(let i=0; i<=CANVAS_SIZE; i+=VIEW_SCALE) {
ctx.moveTo(i, 0); ctx.lineTo(i, CANVAS_SIZE);
ctx.moveTo(0, i); ctx.lineTo(CANVAS_SIZE, i);
}
ctx.stroke();
ctx.strokeStyle = '#dee2e6';
ctx.lineWidth = 2;
ctx.beginPath();
ctx.moveTo(CENTER_X, 0); ctx.lineTo(CENTER_X, CANVAS_SIZE);
ctx.moveTo(0, CENTER_Y); ctx.lineTo(CANVAS_SIZE, CENTER_Y);
ctx.stroke();