Spaces:
Running
Running
File size: 10,967 Bytes
d8760c5 |
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 |
window.state = window.state || {
scoreSteps: 101,
nParams: 11,
nRandLines: 50,
nMaxRand: 0,
nBatches: 4,
learningRate: 22,
}
window.pointData = window.pointData || d3.range(100).map(i => {
var color = i % 2 ? 0 : 1
var color0 = color
var color1 = color
var σ = .1
var μ = .2
if (color){
var x = d3.randomNormal(1 - μ, σ)()
var y = d3.randomNormal(1 - μ, σ*1)()
} else {
var x = d3.randomNormal(μ, σ)()
var y = d3.randomNormal(μ, σ*1)()
y = d3.clamp(0, y, .4)
}
x = d3.clamp(.03, x, .97)
y = d3.clamp(.03, y, .97)
var bucketX = x*(state.nParams - 1)
if (i == 51){
x = .25
y = .55
color = 0
color0 = 0
color1 = 1
}
return {i, x, y, bucketX, color, color0, color1}
})
var updateAllFns = []
var updateAll = () => updateAllFns.forEach(fn => fn())
var updateCircleFns = []
var updateCircle = (d) => updateCircleFns.forEach(fn => fn(d))
var sel = d3.select('.epoch-graph').html('')
.st({marginTop: 30})
.at({role: 'graphics-document', 'aria-label': `Grid of charts showing a simple 2d classifer being trained over four epochs. Changing a single outlier point from red to blue makes a big difference in the final model.`})
var dbSel = d3.select('.decision-boundry').html('').append('div')
.at({role: 'graphics-document', 'aria-label': `Slides to control the level clipping and noise applied the gradient at each step. Increasing the noise enough makes the decision boundries for the models trained on the red and blue outliers overlap.`})
var colorTypes = [{key: 'color1'}, {key: 'color0'}]
sel.appendMany('div', colorTypes)
.each(drawColorType)
drawBatch(
dbSel.append('div').parent().append('div'),
3,
colorTypes[0],
colorTypes[1]
)
function drawColorType(ct){
function calcBatches(){
var buckets = d3.nestBy(pointData, d => Math.floor(d.bucketX))
buckets = _.sortBy(buckets, d => +d.key)
pointData.forEach(d => {
d.bucketX = d.x*(state.nParams - 1)
})
buckets.forEach((bucket, i) => {
bucket.i = i
bucket.x = +bucket.key
bucket.pointData = pointData.filter(d => Math.abs(d.bucketX - bucket.key) < 1)
bucket.scores = d3.range(state.scoreSteps).map(i => {
var y = i/(state.scoreSteps - 1)
var pad = 0
var score = d3.sum(bucket.pointData, (d, i) => {
// return d[ct.key] == 0 ? d.y < y - pad : d.y > y + pad
var dif = 1 - Math.abs(d.bucketX - bucket.x)
dif = Math.min(dif, .5)
if (d[ct.key] == 0){
return d.y < y - pad ? dif : -dif
} else {
return d.y > y + pad ? dif : -dif
}
})
return {y, i, score}
})
bucket.best = _.maxBy(bucket.scores, d => d.score)
bucket.scores.forEach(score => {
var nextScoreIndex = score.i
var charge = 0
for (var j = 0; j < state.learningRate; j++){
var dif = bucket.best.score - bucket.scores[nextScoreIndex]?.score
charge += dif || 5
if (bucket.scores[nextScoreIndex | 0].score == bucket.best.score){
j = state.learningRate
} else if (charge > 2) {
nextScoreIndex += nextScoreIndex < bucket.best.i ? 1 : -1
charge = 0
}
}
score.nextScoreIndex = nextScoreIndex
})
bucket.x = (bucket.i +.5)/(state.nParams - 1)
})
var rng = new alea(ct.key)
// random lines x batches x buckets
var randLines = d3.range(state.nRandLines).map(() => {
return [buckets.map(d => Math.floor(d.x*state.scoreSteps))]
})
function calcNextBatch(){
randLines.forEach(line => {
var next = _.last(line).map((scoreIndex, i) => {
var randInt = Math.round((rng() - .5)*state.nMaxRand)
return d3.clamp(
0,
buckets[i].scores[scoreIndex | 0].nextScoreIndex + randInt,
state.scoreSteps - 1)
})
line.push(next)
})
}
d3.range(state.nBatches - 1).forEach(calcNextBatch)
ct.buckets = buckets
ct.randLines = randLines
}
calcBatches()
var sel = d3.select(this)
var render = (function(){
ct.renderFns = []
sel
.append('div.chart-title').text(ct.key == 'color1' ? 'Training a model with an isolated red point' : 'Training a model with an isolated blue point')
.st({marginLeft: 10, marginBottom: -18, marginTop: -5})
.parent()
.appendMany('div', ct.randLines[0])
.st({display: 'inline-block'})
.each(function(d, i){ drawBatch(d3.select(this), i, ct)})
return () => ct.renderFns.forEach(d => d())
})()
updateAllFns.push(() => {
calcBatches()
render()
})
}
function drawBatch(sel, batchIndex, ct, ct2){
var size = ct2 ? 300 : 150
var mScale = ct2 ? 0 : 1
var c = d3.conventions({
sel,
width: size,
height: size,
margin: {left: 10*mScale, right: 10*mScale, top: 20*mScale, bottom: ct2 ? 50 : 20},
layers: 'scsd',
})
var divSel = c.layers[3].st({pointerEvents: 'none'})
c.layers[0].append('rect')
.at({width: c.width, height: c.height, fill: '#efefef'})
c.svg = c.layers[2]
c.svg.append('rect')
.at({width: c.width, height: c.height, fill: 'rgba(0,0,0,0)'})
c.svg.append('text')
.text('Step ' + (batchIndex + 1))
.translate([c.width/2, c.height + 13])
.at({textAnchor: 'middle', fontSize: 10, fill: '#999'})
.st({opacity: ct2 ? 0 : 1})
c.x.domain([0, 1]).clamp(1)
c.y.domain([0, 1]).clamp(1)
var drag = d3.drag()
.on('start', () => c.svg.classed('dragging', 1))
.on('end', () => c.svg.classed('dragging', 0))
.on('drag', function(d){
d.x = d3.clamp(.03, c.x.invert(d3.event.x), .97)
d.y = d3.clamp(.03, c.y.invert(d3.event.y), .97)
updateCircle(d)
updateAll()
})
.subject(function(d){ return {x: c.x(d.x), y: c.y(d.y)} })
var circleSel = c.svg.appendMany('circle.point', pointData)
.at({r: 4, fill: d => util.colors[d[ct.key]]})
.call(drag)
.classed('swapped', d => d.color0 != d.color1)
.translate(d => [c.x(d.x), c.y(d.y)])
// .call(d3.attachTooltip)
updateCircleFns.push(d => {
circleSel
.filter(e => e == d) // rendering circles is dropping frames ?
.translate(d => [c.x(d.x), c.y(d.y)])
})
if (ct2){
var defs = c.svg.append('defs');
defs.append('linearGradient#red-blue-def')
.append('stop').at({offset: '0%', 'stop-color': util.colors[0]}).parent()
.append('stop').at({offset: '45%', 'stop-color': util.colors[0]}).parent()
.append('stop').at({offset: '55%', 'stop-color': util.colors[1]}).parent()
.append('stop').at({offset: '100%', 'stop-color': util.colors[1]})
defs.append('linearGradient#blue-red-def')
.append('stop').at({offset: '0%', 'stop-color': util.colors[1]}).parent()
.append('stop').at({offset: '45%', 'stop-color': util.colors[1]}).parent()
.append('stop').at({offset: '55%', 'stop-color': util.colors[0]}).parent()
.append('stop').at({offset: '100%', 'stop-color': util.colors[0]})
circleSel
// .at({r: 1.2})
.filter(d => d.color0 != d.color1)
.st({r: 7, fillOpacity: 1})
.st({fill: 'url(#red-blue-def)'})//, stroke: 'url(#blue-red-def)'})
var gradientClipAnnoSel = c.svg.append('text.annotation')
.translate([c.width + 20, -40])
.tspans(d3.wordwrap('Completely clipping the gradient stops the model from learning anything from the training data.', 25), 14)
divSel.append('div.annotation')
.translate([30, c.height + 5])
.html(`
<span style='color:${util.colors[0]}'>〰</span> Models trained with the isolated blue point
<div>
<span style='color:${util.colors[1]}'>〰</span> Models trained with the isolated red point
`)
.st({lineHeight: '1.3em'})
.selectAll('span').st({fontSize: 20, height: 0, display: 'inline-block', top: 3, position: 'relative', fontWeight: 700})
}
function getRandLines(){
return ct2 ? ct.randLines.concat(ct2.randLines) : ct.randLines
}
var ctx = c.layers[1]
var lineGen = d3.line()
.x(d => c.x(d.x))
.y(d => c.y(d.y))
.curve(d3.curveNatural)
.context(ctx)
ct.renderFns.push(() => {
var scores = ct.buckets[0].scores
var paddedLineData = getRandLines().map(line => {
var xyData = line[batchIndex].map((scoreIndex, i) => {
return {x: ct.buckets[i].x, y: scores[scoreIndex | 0].y}
})
return [
{x: 0, y: batchIndex*state.learningRate ? xyData[0].y : 0},
...xyData,
{x: 1, y: batchIndex*state.learningRate ? _.last(xyData).y : 1}
]
})
ctx.clearRect(-c.margin.left, -c.margin.top, c.width + c.margin.left + c.margin.right, c.height + c.margin.top + c.margin.bottom)
paddedLineData.forEach((d, i) => {
ctx.beginPath()
ctx.lineWidth = .1
ctx.strokeStyle = !ct2 ? '#000' : i < ct.randLines.length ? util.colors[1] : util.colors[0]
lineGen(d)
ctx.stroke()
})
if (ct2){
gradientClipAnnoSel.st({opacity: state.learningRate == 0 ? 1 : 0})
}
})
}
function addSliders(){
var width = 180
var height = 30
var color = '#000'
var sliders = [
{key: 'nMaxRand', label: 'Random Noise', r: [0, 30]},
{key: 'learningRate', label: 'Gradient Clip', r: [30, 0]},
]
sliders.forEach(d => {
d.value = state[d.key]
d.xScale = d3.scaleLinear().range([0, width]).domain(d.r).clamp(1)
})
var svgSel = dbSel.append('div.sliders').lower()
.st({marginTop: 5, marginBottom: 5})
.appendMany('div.slider-container', sliders)
.append('svg').at({width, height})
.append('g').translate(120, 0)
svgSel.append('text.chart-title')
.text(d => d.label)
.at({textAnchor: 'end', dy: '.33em', x: -15})
var sliderSel = svgSel
.on('click', function(d){
d.value = d.xScale.invert(d3.mouse(this)[0])
renderSliders(d)
})
.classed('slider', true)
.st({cursor: 'pointer'})
var textSel = sliderSel.append('text.slider-label-container')
.at({y: -20, fontWeight: 500, textAnchor: 'middle', x: 180/2})
sliderSel.append('rect')
.at({width, height, y: -height/2, fill: 'rgba(0,0,0,0)'})
sliderSel.append('path').at({
d: `M 0 -.5 H ${width}`,
stroke: color,
strokeWidth: 1
})
var leftPathSel = sliderSel.append('path').at({
d: `M 0 -.5 H ${width}`,
stroke: color,
strokeWidth: 3
})
var drag = d3.drag()
.on('drag', function(d){
var x = d3.mouse(this)[0]
d.value = d.xScale.invert(x)
renderSliders(d)
})
var circleSel = sliderSel.append('circle').call(drag)
.at({r: 7, stroke: '#000'})
function renderSliders(d){
if (d) state[d.key] = d.value
circleSel.at({cx: d => d.xScale(d.value)})
leftPathSel.at({d: d => `M 0 -.5 H ${d.xScale(d.value)}`})
updateAll()
}
renderSliders()
}
addSliders()
updateAll()
|