joeytwiddle · November 17, 2019 06:51 · joeytwiddle · Nov 17, 2019
diff --git a/nn-images.html b/nn-images.html
 <html>
 <body>
 <canvas id="canvas" width="512" height="512"></canvas>
 <!-- <script src="./nn-image.js"></script> -->
 <script>

 class Network {
  layers = []
  weightsLayers = []

  constructor(opts) {
    this.layers.push(new Layer(opts.numInputs))
    for (let i = 0; i < opts.numLayers; i++) {
      this.layers.push(new Layer(opts.nodesPerLayer))
    }
    this.layers.push(new Layer(opts.numOutputs))

    for (let i = 0; i < this.layers.length - 1; i++) {
      this.weightsLayers.push(new WeightsLayer(this.layers[i].numNodes, this.layers[i + 1].numNodes))
    }
  }

  setInputValues(values) {
    this.layers[0].setValues(values)
  }

  getOutputValues() {
    return this.layers[this.layers.length - 1].getValues()
  }

  propagateData() {
    for (let l = 1; l < this.layers.length; l++) {
      const previousLayer = this.layers[l - 1]
      const nextLayer = this.layers[l]
      const inputValues = previousLayer.getValues()
      const outputValues = nextLayer.getValues() // must be writable
      const weightsLayer = this.weightsLayers[l - 1]
      for (let j = 0; j < nextLayer.numNodes; j++) {
        let sum = 0
        for (let i = 0; i < previousLayer.numNodes; i++) {
          sum += inputValues[i] * weightsLayer.getWeight(i, j)
        }
        // Apply sigma function, and store new value
        //outputValues[j] = Math.min(Math.max(sum, 0), 1) // always 0 to 1
        //outputValues[j] = Math.min(Math.max(sum, -1), 1) // always -1 to 1
        //outputValues[j] = Math.atan(sum) / Math.PI // always -1 to 1
        //outputValues[j] = (1 + Math.atan(sum) / Math.PI) / 2 // always 0 to 1
        outputValues[j] = 2 * Math.atan(5 * sum) / Math.PI
      }
    }
  }
 }

 class Layer {
  numNodes = 5
  nodeValues = []

  constructor(numNodes) {
    this.numNodes = numNodes
  }

  setValues(newValues) {
    if (newValues.length !== this.numNodes) {
      throw new Error(`newValues.length (${newValues.length}) does not match this.numNodes (${this.numNodes})`)
    }
    this.nodeValues = newValues
  }

  getValues() {
    return this.nodeValues
  }
 }

 class WeightsLayer {
  numInputs = 3
  numOutputs = 3
  weights = []

  constructor(numInputs, numOutputs) {
    this.numInputs = numInputs
    this.numOutputs = numOutputs

    // We could pass this in as an option
    const generateRandomWeight = () => (2 * Math.random() - 1) * 2
    //const generateRandomWeight = () => (2 * Math.random() - 1)

    for (let i = 0; i < numInputs; i++) {
      for (let j = 0; j < numOutputs; j++) {
        this.weights[j * numInputs + i] = generateRandomWeight()
      }
    }
  }

  getWeight(i, j) {
    return this.weights[j * this.numInputs + i]
  }
 }

 function getOutputForInput(network, input) {
  network.setInputValues(input)
  network.propagateData()
  return network.getOutputValues()
 }

 // TODO I guess this should really be stored in the network.
 //      We could put this into the input layer, we will just need to not clobber it with setInputValues()
 //const rnd1 = Math.random()

 function drawImage(canvas, network) {
  // For faster rendering, sample fewer points
  const downsample = 4

  //const rnd2 = Math.random()
  const rnd1 = Math.cos(Date.now() / 1000 / 3)
  const rnd2 = Math.sin(Date.now() / 1000 / 5)

  const ctx = canvas.getContext('2d')

  for (let x = 0; x < canvas.width; x += downsample) {
    for (let y = 0; y < canvas.height; y += downsample) {
      const input = [x / canvas.width - 0.5, y / canvas.width - 0.5, rnd1, rnd2]
      const result = getOutputForInput(network, input)
      // CONSIDER Might be faster to use pixel buffer, but this is likely not the bottleneck anyway!
      ctx.fillStyle = `rgb(${result[0] * 255}, ${result[1] * 255}, ${result[2] * 255})`
      ctx.fillRect(x, y, downsample, downsample)
    }
  }
 }

 const network = new Network({
  numInputs: 4,
  numOutputs: 3,
  numLayers: 4,
  nodesPerLayer: 9,
 })

 const canvas = document.getElementById('canvas')

 //drawImage(canvas, network)
 setInterval(() => drawImage(canvas, network), 100)

 </script>
 </body>
 <!-- By joeytwiddle.  Free to use as you wish. -->
 <!-- Thanks to Purnima Kamath for the inspiration.  My apologies for the numerous bugs. -->
 </html>
	<html>
	<body>
	<canvas id="canvas" width="512" height="512"></canvas>
	<!-- <script src="./nn-image.js"></script> -->
	<script>

	class Network {
	layers = []
	weightsLayers = []

	constructor(opts) {
	this.layers.push(new Layer(opts.numInputs))
	for (let i = 0; i < opts.numLayers; i++) {
	this.layers.push(new Layer(opts.nodesPerLayer))
	}
	this.layers.push(new Layer(opts.numOutputs))

	for (let i = 0; i < this.layers.length - 1; i++) {
	this.weightsLayers.push(new WeightsLayer(this.layers[i].numNodes, this.layers[i + 1].numNodes))
	}
	}

	setInputValues(values) {
	this.layers[0].setValues(values)
	}

	getOutputValues() {
	return this.layers[this.layers.length - 1].getValues()
	}

	propagateData() {
	for (let l = 1; l < this.layers.length; l++) {
	const previousLayer = this.layers[l - 1]
	const nextLayer = this.layers[l]
	const inputValues = previousLayer.getValues()
	const outputValues = nextLayer.getValues() // must be writable
	const weightsLayer = this.weightsLayers[l - 1]
	for (let j = 0; j < nextLayer.numNodes; j++) {
	let sum = 0
	for (let i = 0; i < previousLayer.numNodes; i++) {
	sum += inputValues[i] * weightsLayer.getWeight(i, j)
	}
	// Apply sigma function, and store new value
	//outputValues[j] = Math.min(Math.max(sum, 0), 1) // always 0 to 1
	//outputValues[j] = Math.min(Math.max(sum, -1), 1) // always -1 to 1
	//outputValues[j] = Math.atan(sum) / Math.PI // always -1 to 1
	//outputValues[j] = (1 + Math.atan(sum) / Math.PI) / 2 // always 0 to 1
	outputValues[j] = 2 * Math.atan(5 * sum) / Math.PI
	}
	}
	}
	}

	class Layer {
	numNodes = 5
	nodeValues = []

	constructor(numNodes) {
	this.numNodes = numNodes
	}

	setValues(newValues) {
	if (newValues.length !== this.numNodes) {
	throw new Error(`newValues.length (${newValues.length}) does not match this.numNodes (${this.numNodes})`)
	}
	this.nodeValues = newValues
	}

	getValues() {
	return this.nodeValues
	}
	}

	class WeightsLayer {
	numInputs = 3
	numOutputs = 3
	weights = []

	constructor(numInputs, numOutputs) {
	this.numInputs = numInputs
	this.numOutputs = numOutputs

	// We could pass this in as an option
	const generateRandomWeight = () => (2 * Math.random() - 1) * 2
	//const generateRandomWeight = () => (2 * Math.random() - 1)

	for (let i = 0; i < numInputs; i++) {
	for (let j = 0; j < numOutputs; j++) {
	this.weights[j * numInputs + i] = generateRandomWeight()
	}
	}
	}

	getWeight(i, j) {
	return this.weights[j * this.numInputs + i]
	}
	}

	function getOutputForInput(network, input) {
	network.setInputValues(input)
	network.propagateData()
	return network.getOutputValues()
	}

	// TODO I guess this should really be stored in the network.
	// We could put this into the input layer, we will just need to not clobber it with setInputValues()
	//const rnd1 = Math.random()

	function drawImage(canvas, network) {
	// For faster rendering, sample fewer points
	const downsample = 4

	//const rnd2 = Math.random()
	const rnd1 = Math.cos(Date.now() / 1000 / 3)
	const rnd2 = Math.sin(Date.now() / 1000 / 5)

	const ctx = canvas.getContext('2d')

	for (let x = 0; x < canvas.width; x += downsample) {
	for (let y = 0; y < canvas.height; y += downsample) {
	const input = [x / canvas.width - 0.5, y / canvas.width - 0.5, rnd1, rnd2]
	const result = getOutputForInput(network, input)
	// CONSIDER Might be faster to use pixel buffer, but this is likely not the bottleneck anyway!
	ctx.fillStyle = `rgb(${result[0] * 255}, ${result[1] * 255}, ${result[2] * 255})`
	ctx.fillRect(x, y, downsample, downsample)
	}
	}
	}

	const network = new Network({
	numInputs: 4,
	numOutputs: 3,
	numLayers: 4,
	nodesPerLayer: 9,
	})

	const canvas = document.getElementById('canvas')

	//drawImage(canvas, network)
	setInterval(() => drawImage(canvas, network), 100)

	</script>
	</body>
	<!-- By joeytwiddle. Free to use as you wish. -->
	<!-- Thanks to Purnima Kamath for the inspiration. My apologies for the numerous bugs. -->
	</html>