31:T61d,draw = { let inp = tf.randomNormal([1, 256, 256, 3], 0.0, 1.0, 'float32') let finalSize = 256 * Math.pow(octaveScale, numOctaves - 1) let inpV = tf.variable(inp, true) invalidation.then(() => { inp.dispose(), inpV.dispose(), target.dispose() }) const obj = x => tf.tidy(() => { const ret = graphExec(net, layerSelector, {'input_1': x}) return ret.shape.length == 2 ? ret : ret.sum([1, 2]) }) const grads = tf.grad(obj) const target = tf.tidy(() => tf.oneHot(tf.scalar(neuron, 'int32').expandDims(0), numFilters).toFloat()) for(let octave = 0; octave < numOctaves; octave++) { if(octave > 0) { const oldInpV = inpV const [newH, newW] = [Math.floor(octaveScale * inpV.shape[1]), Math.floor(octaveScale * inpV.shape[2])] inpV = tf.tidy(() => tf.variable(tf.image.resizeBilinear(inpV, [newH, newW]))) oldInpV.dispose() } for(let i = 0; i < nSteps; i++) { const normed = tf.tidy(() => { let g; if(useTiledGrad) g = gradTiled(inpV, x => grads(x, target), 256) else g = grads(inpV, target) const gNormed = laplacianNormalization(g, 4) inpV.assign(tf.tidy(() => inpV.add(gNormed.mul(tf.scalar(stepSize))))) return tf.image.resizeBilinear(visNormalise(inpV.squeeze()), [finalSize, finalSize]) }) invalidation.then(() => normed.dispose()) await tf.toPixels(normed, canvas) normed.dispose() await tf.nextFrame() yield md`octave: ${octave} step: ${i}` } } target.dispose() inpV.dispose() inp.dispose() }32:T4e5,function laplacianNormalization(img, scales) { const lapKernel4 = lapKernel.mul(tf.scalar(4.0)) const downConv = x => tf.conv2d(x, lapKernel, [2, 2,], 'same') const upConv = (x, s) => tf.conv2dTranspose(x, lapKernel4, s, [2, 2], 'same') function lapSplit(img) { return tf.tidy(() => { const lo = downConv(img) const lo2 = upConv(lo, img.shape) const hi = img.sub(lo2) return [lo, hi] }) } function lapSplitN(img, n) { return tf.tidy(() => { const levels = [] for(let i = 0; i < n; i++) { const [lo, hi] = lapSplit(img) img = lo levels.push(hi) } levels.push(img) return reverse(levels) }) } function lapMerge(levels) { let img = levels[0] for(const hi of levels.slice(1)) { img = tf.tidy(() => upConv(img, hi.shape, [1, 2, 2, 1]).add(hi)) } return img } function normalizeStd(img) { const EPS = 1e-10 const std = tf.tidy(() => img.square().mean().sqrt()) return tf.tidy(() => img.div(tf.maximum(std, tf.scalar(EPS)))) } const tlevels = tf.tidy(() => { const levels = lapSplitN(img, scales) return levels.map(normalizeStd) }) return tf.tidy(() => lapMerge(tlevels)) }33:[] 2f:{"title":"Deepdream Experiments","cells":[{"id":0,"value":"md`# Deepdream Experiments`","pinned":false,"hidden":false,"mode":"ojs"},{"id":213,"value":"md`\nTry pasting these model urls into the textbox below to visualise them with DeepDream:\n* Squeezenet \\`(inputSide = 227)\\`: https://storage.googleapis.com/tfjs-models/tfjs/squeezenet_v1.1/model.json\n* Mobilenet: \\`(inputSide = 224)\\`: https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json`","pinned":false,"hidden":false,"mode":"ojs"},{"id":202,"value":"viewof modelUrl = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":25,"value":"canvas = html`

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":359,"value":"numFilters = {\n const layer = net.getLayer(layerSelector)\n const outputShape = layer.output.shape\n return outputShape[outputShape.length - 1]\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":191,"value":"viewof neuron = {\n return html``\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":197,"value":"viewof stepSize = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":199,"value":"viewof nSteps = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":495,"value":"viewof octaveScale = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":496,"value":"viewof numOctaves = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":231,"value":"viewof layerSelector = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":628,"value":"md`# Implementation\n---`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4,"value":"tf = require('@tensorflow/tfjs')","pinned":false,"hidden":false,"mode":"ojs"},{"id":13,"value":"net = await tf.loadModel(modelUrl)","pinned":false,"hidden":false,"mode":"ojs"},{"id":477,"value":"function roll(arr, offset, axis) {\n const s = arr.shape[axis]\n return tf.tidy(() => {\n const split = offset >= 0 ? [s - offset, offset] : [-offset, s + offset]\n const [lower, upper] = tf.split(arr, split, axis)\n return tf.concat([upper, lower], axis)\n })\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":471,"value":"function gradTiled(img, grad, tileSize = 512) {\n return tf.tidy(() => {\n const sz = tileSize\n\n const [h, w] = img.shape.slice(1, 3)\n const [offY, offX] = [Math.max(1, Math.floor(Math.random() * sz)), \n Math.max(1, Math.floor(Math.random() * sz))]\n const imgShift = roll(roll(img, offY, 1), offX, 2) \n const gradDeep = tf.tidy(() => {\n const gradTiles = []\n for(let y = 0; y < Math.max(h - Math.floor(sz/2), sz); y += sz) {\n for(let x = 0; x < Math.max(w - Math.floor(sz/2), sz); x += sz) {\n const g = tf.tidy(() => {\n const sub = img.slice([0, y, x], [-1, sz, sz])\n return grad(sub)\n })\n gradTiles.push(g)\n }\n }\n return tf.stack(gradTiles)\n })\n const tiledGrad = gradDeep.reshape(img.shape)\n return roll(roll(tiledGrad, -offY, 1), -offX, 2); //roll(tiledGrad, h - offY, w - offX)\n })\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":531,"value":"viewof useTiledGrad = html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":12,"value":"$31","pinned":true,"hidden":false,"mode":"ojs"},{"id":606,"value":"lapKernel = {\n const k1d = tf.tensor1d([1.0, 4.0, 6.0, 4.0, 1.0], 'float32') // [5,]\n const k = tf.outerProduct(k1d, k1d) // [5, 5]\n const kexp = k.expandDims(2).expandDims(3) // [5, 5, 1, 1]\n const eye = tf.eye(3) // [3, 3]\n return kexp.div(k.sum()).mul(eye) // [5, 5, 3, 3]\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":388,"value":"$32","pinned":false,"hidden":false,"mode":"ojs"},{"id":406,"value":"function reverse(arr) {\n return arr.slice().map((_, i) => arr[arr.length - i - 1]) \n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":256,"value":"dot = require(\"@observablehq/graphviz@0.0.2/dist/graphviz.min.js\")","pinned":false,"hidden":false,"mode":"ojs"},{"id":304,"value":"import { toDot, graphExec, inputMap, Tidier, TFC } from '@aman-tiwari/tensorflow-js-helpers'","pinned":false,"hidden":false,"mode":"ojs"},{"id":463,"value":"tidier = Tidier((tf))","pinned":false,"hidden":false,"mode":"ojs"},{"id":323,"value":"vis = dot`${(toDot(net))}`","pinned":false,"hidden":false,"mode":"ojs"},{"id":52,"value":"variance = (t) => tf.tidy(() => {\n const norm = t.squaredDifference(t.mean()).sum()\n const n = t.shape.reduce((x, y) => x * y, 1)\n return norm.div(tf.scalar(n))\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":47,"value":"visNormalise = t => tf.tidy(() => {\n const tprime = t.sub(t.mean())\n const std = variance(t, -1, true).sqrt()\n return tf.tidy(() => tprime.div(std).div(tf.scalar(10.0)).add(tf.scalar(0.5)).clipByValue(0.0, 1.0))\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":8,"value":"imageMeta = await fetch('https://cdn.rawgit.com/PAIR-code/saliency/gh-pages/metadata.json').then(x => x.json())","pinned":false,"hidden":false,"mode":"ojs"},{"id":6,"value":"viewof imageRoll = {\n const dep = [tf]\n const id2Url = id => `https://storage.googleapis.com/learnjs-data/images/saliency-data/${id}.png`\n const handler = id => (evt) => { \n mutable inputUrl = evt.target.src;\n }\n const imgs = imageMeta.map(({ id, label }) => {\n const elt = html`

`\n elt.onclick = handler(id)\n return elt\n })\n return html`

${imgs}

`\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":10,"value":"mutable inputUrl = 'https://storage.googleapis.com/learnjs-data/images/saliency-data/4.png'","pinned":false,"hidden":false,"mode":"ojs"},{"id":604,"value":"{ while(true) { yield tf.memory() } }","pinned":false,"hidden":false,"mode":"ojs"}],"files":"$W33","stdlib":"1","titling":"auto"}