Published
Edited
Sep 24, 2020
17 forks
115 stars
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
async function createModel() {
const model = tf.sequential()
model.add(tf.layers.conv2d({
inputShape: [28, 28, 1],
kernelSize: 5,
filters: 8,
strides: 1,
activation: 'relu',
kernelInitializer: 'VarianceScaling'
}))
model.add(tf.layers.maxPooling2d({
poolSize: [2, 2],
strides: [2, 2]
}))
model.add(tf.layers.conv2d({
kernelSize: 5,
filters: 16,
strides: 1,
activation: 'relu',
kernelInitializer: 'VarianceScaling'
}))
model.add(tf.layers.maxPooling2d({
poolSize: [2, 2],
strides: [2, 2]
}))
model.add(tf.layers.flatten())
model.add(tf.layers.dense({
units: 10,
kernelInitializer: 'VarianceScaling',
activation: 'softmax',
// useBias: false
}))
const optimizer = tf.train.sgd(LEARNING_RATE)
await model.compile({
optimizer: optimizer,
loss: 'categoricalCrossentropy',
metrics: ['accuracy']
})
return model
}
Insert cell
function readTopology({ model }) {
const input = model.inputs[0]
let inputNodes = 1
let size = IMG_SIZE * IMG_SIZE
const layers = [{
nodes: inputNodes,
outSize: size,
outShape: input.shape.slice(1),
name: 'input',
}]
for (let tfLayer of model.layers) {
const name = tfLayer.name
const nodes = tfLayer.filters || tfLayer.units || (name.match(/flatten/) ? 1 : inputNodes)
const layer = {
nodes,
inSize: size
}
// don't know why but the output of conv2d layers seem to always loose 4px size
if (name.match(/^conv2d/)) {
size = Math.pow(Math.sqrt(size) - 4, 2)
}
// max sample half (not supporting other configurations)
if (name.match(/^max_pooling2d/)) {
size = Math.pow(Math.sqrt(size) / 2, 2)
}
if (name.match(/^flatten/)) {
size = inputNodes * size
}
if (name.match(/^dense/)) {
size = 1
}
layer.outSize = size
layer.outShape = tfLayer.output.shape.slice(1)
layer.name = name
if (tfLayer.useBias) {
layer.bias = tfLayer.bias.val.dataSync()
}
if (tfLayer.kernel) {
const kernelVal = tfLayer.kernel.val
// expect weights per node
if (nodes > 1 && kernelVal.shape[kernelVal.shape.length - 1] !== nodes) {
console.warn('unexpected kernel val shape, ignoring weights')
return
}
const weightsPerNode = nodes > 1
? tf.split(kernelVal, nodes, kernelVal.shape.length - 1).map(val => val.squeeze())
: [tfLayer.kernel.val]

layer.weights = weightsPerNode
.map(perNode => {
if (inputNodes > 1) {
// expect weights per input node
if (inputNodes > 1 && perNode.shape[perNode.shape.length - 1] !== inputNodes) {
console.warn('unexpected kernel val shape, ignoring weights')
return
}
return tf.split(perNode, inputNodes, perNode.shape.length - 1).map(val => val.squeeze())
}
return [perNode]
})
.map(vals => vals.map(val => val.dataSync()))

// layer.weightExtent = layer.weights.map(n => n.map(i => d3.extent(i)))
// layer.weightSums = layer.weights.map(n => n.map(i => d3.sum(i)))
}
layers.push(layer)
inputNodes = nodes
}
return layers
}
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell
Insert cell

One platform to build and deploy the best data apps

Experiment and prototype by building visualizations in live JavaScript notebooks. Collaborate with your team and decide which concepts to build out.
Use Observable Framework to build data apps locally. Use data loaders to build in any language or library, including Python, SQL, and R.
Seamlessly deploy to Observable. Test before you ship, use automatic deploy-on-commit, and ensure your projects are always up-to-date.
Learn more