{"cells":[{"cell_type":"code","metadata":{"id":"1215b33d681","hideRun":false},"source":["import {NeuroModule, nm, BinLayer, Linear} from '/web/oda/apps/neuronet/genius/neuro-module/neuro-module.js';\nimport {tensor} from '/web/oda/apps/neuronet/genius/torus/torus.js';"],"outputs":[]},{"cell_type":"code","metadata":{"id":"110ebbffa46","hideRun":false},"source":["MNIST = class MNIST extends NeuroModule{\n constructor(dim = 28) {\n super(arguments);\n }\n \n __init__(){\n this.layer1 = new Linear(this.dim*this.dim,this.dim*this.dim/2,true)\n this.layer2 = new Linear(this.dim*this.dim/2,10,true)\n }\n \n forward(x){\n x=tensor.from(x)\n x = this.layer1(x)\n x = x.sigm()\n x = this.layer2(x)\n x = x.sigm()\n return x\n }\n}\n"],"outputs":[]},{"cell_type":"code","metadata":{"id":"10bdb2bf168","hideRun":false},"source":["// model = await this.contextItem.owner.getFile('mnist/model.json');\n// out_size = 10\n// if(model){\n// // model = await model.load();\n// mnist = new MNIST(model);\n// }\n// else\nmnist = new MNIST();\n//loss.back();\n//this.contextItem.$owner.saveFile(mnist.toJSON(), 'mnist/model.json')\nprint(mnist)"],"outputs":[{"data":{"text/plain":"{\n \"$\": \"MNIST\",\n \"dim\": 28,\n \"prototype\": {}\n}"}}]},{"cell_type":"code","metadata":{"id":"5df214bef8","hideRun":false},"source":["datasetRaw = await ODA.loadJSON('/web/oda/apps/neuronet/mnist/mnist.json');\n\ngetTrain = (dataset = datasetRaw, batch=5, test_size=0.2) => {\n let [result, res] = [{inputs:[],targets:[]},[]] \n\n let suBatch = Array(10).fill(0).map((_,i)=>i) //.sort(()=>Math.random() - 0.5)\n suBatch.forEach(n=>{\n let target = Array(10).fill(0);\n target[n] = 1\n for (let j=0; jMath.random() - 0.5)\n res.forEach ( ([inp,tar]) => {result.inputs.push(inp); result.targets.push(tar)} )\n return result\n}\n\n// print(datasetRaw[0][0].flat())"],"outputs":[{"data":{"text/plain":"{}"}}]},{"cell_type":"code","metadata":{"id":"1123b3d3bb","hideRun":false},"source":["// dat = mnist.getTrain(datasetRaw , 1)\n// print(`getTrain возвращает { inputs (${dat.inputs.length} по ${dat.inputs[0].length}) ,targets (${dat.targets.length} по ${dat.targets[0].length}) }.`)\n\n// let loss='',rez=''\n\n// // print(dat.inputs[0])\n\n// for (let i=0; i\n :host {max-width:1000px; margin:auto; display:grid; grid-template-columns:1fr 1fr; grid-template-areas: 'input output' 'loss loss'; }\n .block {border: 1px solid #4d85cf; padding:10px; margin:10px;}\n .box { margin:10px}\n .block h3 {margin:0 10px;}\n oda-button {border: 1px solid #4d85cf; margin:10px;}}\n \n
\n

input:

\n \n \n \n \n
\n
\n

output:

\n \n \n \n \n {{$for.index}}\n {{($for.item*100).toFixed(2)}} %\n \n {{maxIndex}} \n \n
\n
\n

loss:

\n \n \n \n \n
\n \n `,\n get rects() {\n return this.input.map((v,i)=> {\n let [x,y] = [i%28, ~~(i/28)].map(a=>a*10+1)\n return {x,y,v}\n } )\n },\n get maxIndex() { return this.result.reduce((a,b,i,c)=>c[a]>b?a:i,0) },\n\n get pathD(){\n if (this.pointsL.length<2) return ''\n let padding = Math.min(this.width,this.height)/10\n let [w,h] = [this.width,this.height].map(x=>x-padding*2)\n let dx = w/(this.pointsL.length-1)\n let [min,max] = [Math.min(...this.pointsL),Math.max(...this.pointsL)]\n let points = this.pointsL.map((v,i)=>[i*dx+padding,h-(v-min)/(max-min)*h+padding])\n let rez = 'M ' + points[0].join(' ') + ' '\n if (this.bezier) rez+= 'Q ' + points[0].join(' ') + ', ' + points[1].join(' ') + ' ' \n else rez += 'L ' + points[1].join(' ') + ' '\n points.slice(2).forEach(e => rez += (this.bezier?'T ':'L ') + e.join(' ') + ' ' )\n return rez\n },\n\n get pointsL() {\n let [lL,lP] = [this.loss.length,this.maxPoint]\n if (lL{\n let idx = Math.floor(i*(lL/(lP-1)))\n let dL = Math.floor(lL/(lP-1))+1\n return this.loss.slice(idx,idx+dL).reduce((a,c)=>a+c)/dL\n })\n newP.push(this.loss[lL-1])\n return newP\n },\n\n go:false,\n _go(){\n this.go = !this.go\n this.steps()\n },\n\n width:281*2+10*6+2,\n height:281,\n maxPoint:100,\n\n input: {\n $def: new Array(28*28).fill(0)\n },\n result:{\n $def: new Array(10).fill(0.1)\n },\n loss:{\n $def:[]//new Array(20000).fill(0).map(_=> Math.random()*5-2)\n },\n\n mnist: new MNIST(),\n getTrain:getTrain,\n\n step() {\n let dat = this.getTrain(undefined, 1)\n let loss, rez;\n for (let i=0; i {\n this.step()\n this.steps()\n }, 0) \n }\n\n})\n"],"outputs":[]},{"cell_type":"html","metadata":{"id":"189f97824ef"},"source":[""]},{"cell_type":"code","metadata":{"id":"f7b1365c51","hideRun":false},"source":["layer1 = nm.linear(10, 10);\nlayer2 = nm.linear(10, 10);\nprint(`layer1 (${layer1.dim_in} x ${layer1.dim_out}), layer2 (${layer2.dim_in} x ${layer2.dim_out})`)\n\n// layer3 = new BinLayer(input.length, target.length);"],"outputs":[{"data":{"text/plain":"layer1 (undefined x undefined), layer2 (undefined x undefined)"}}]},{"cell_type":"code","metadata":{"id":"12d005b792f","hideRun":false},"source":["// print(tensor.LEARNING_RATE)\ntensor.LEARNING_RATE = 0.3;\n// print(tensor.LEARNING_RATE)\nloss = 0\nfor (let i=0; i<10; i++) {\n input = Array(10).fill(0)\n input[Math.floor(Math.random()*10)]=1\n // print(input.join(''))\n \n result = layer1(input);\n // print('layer1', result);\n result = result.sigm();\n // print('sigm1', result);\n result = layer2(result);\n // print('layer2', result.data);\n result = result.sigm();\n \n result = result.softmax();\n \n \n \n let res = ''\n result.data.forEach(x=> res += (' ' + x.toFixed(1)) )\n // print(res)\n\n print(` input = ${input.join(' ')} \\n result=${res} \\n max= ${softmax(result.data).join(' ')}` )\n // print('sigm2',, , ;\n \n loss = result.crossEntropy (input);\n // print('loss', loss.data[0]);\n loss.back();\n}\nprint('loss', loss.data[0]);\n\n\nprint([...result.data])\n\n\n// i = 0 "],"outputs":[{"data":{"text/plain":"ReferenceError:\nsoftmax is not defined\n(27:54)"}}]},{"cell_type":"code","metadata":{"id":"120f8586aca","hideRun":false},"source":["\n// if (iMath.random() // +1 with probability p = σ(x), -1 otherwise.\r\n}\r\n\r\nlet grad = -0.7\r\nprint (`σ(x) = ${Math.max(0,Math.min(1,(grad+1)/2))} = ${Math.max(0,Math.min(1,(grad+1)/2))*100} % ` )\r\n\r\nlet n = 1000\r\nlet rezz = Array(n).fill(grad).map(g => binarize(g)).reduce((akk,b) => { b? akk.true++:akk.false++; return akk },{true:0,false:0} )\r\n\r\nprint(`true: ${rezz.true}, (${rezz.true/n*100} %) `)"],"outputs":[{"data":{"text/plain":"σ(x) = 0.15000000000000002 = 15.000000000000002 % "}},{"data":{"text/plain":"true: 159, (15.9 %) "}}]},{"cell_type":"code","metadata":{"id":"14213e4b005","hideRun":false},"source":["v = Array(15).fill(0).map(a => (Math.random()>0.5)?0:1 ) // [0,1,1,1,0]\r\nmul = v.filter(b=>b).reduce((akk,b) => akk*(-1), 1 )>0\r\nprint (mul)\r\nresult = 'return (' + v.map((_,i)=>{\r\n return '(v['+i+'] - 1)';\r\n }).filter(l=>l).join(` + `) + ')&1;\\n';\r\nprint(result)\r\nconst fn = new Function('v',result);\r\nprint(fn(v))\r\n// print(`in= ${v.join('')} mul =`fn(v))"],"outputs":[{"data":{"text/plain":"true"}},{"data":{"text/plain":"return ((v[0] - 1) + (v[1] - 1) + (v[2] - 1) + (v[3] - 1) + (v[4] - 1) + (v[5] - 1) + (v[6] - 1) + (v[7] - 1) + (v[8] - 1) + (v[9] - 1) + (v[10] - 1) + (v[11] - 1) + (v[12] - 1) + (v[13] - 1) + (v[14] - 1))&1;\n"}},{"data":{"text/plain":"1"}}]},{"cell_type":"code","metadata":{"id":"91775068b6","hideRun":false},"source":[" // print(`input= ${input.join('')} result = ${result.data.map(x=>(Math.round(x*10)/10))} ` )\n \nprint(`-1- ${Array(5).fill(0).map(y=>Math.random()).map(x=>(Math.round(x*10)/10))}) -1-`)\n\nh = Array(5).fill(0).map(y=>Math.random())\nprint(`-2- ${h.map(x=>(Math.round(x*10)/10))}) -2-`)\n"],"outputs":[{"data":{"text/plain":"-1- 0.8,0.9,0.6,0.7,1) -1-"}},{"data":{"text/plain":"-2- 0.8,0.7,0.2,0.5,0.3) -2-"}}]},{"cell_type":"code","metadata":{"id":"13884c43847","hideRun":false},"source":["// let softmax = (output) => {\n// let maximum = output.reduce(function(p,c) { return p>c ? p : c; });\n// let nominators = output.map(function(e) { return Math.exp(e - maximum); });\n// let denominator = nominators.reduce(function (p, c) { return p + c; });\n// let softmax = nominators.map(function(e) { return e / denominator; });\n// let maxIndex = 0;\n// softmax.reduce(function(p,c,i){if(p {\n let exps = output.map(x=>Math.exp(x))\n let sum = exps.reduce((akk,v)=>akk+v)\n return exps.map(e=>e/sum)\n}\nprint(softmax([1,3,8]))\n\nx = tensor.from([1,3,8])\n\nprint(x.softmax())"],"outputs":[{"data":{"text/plain":"[\n 0.0009049591825877395,\n 0.006686794167383233,\n 0.992408246650029\n]"}},{"data":{"text/plain":"{\n \"$\": \"tensor\",\n \"shape\": \"3\",\n \"dType\": \"Float32Array\",\n \"data\": \"0.0009049591608345509 0.006686794105917215 0.992408275604248\"\n}"}}]},{"cell_type":"code","metadata":{"id":"15a0ed373ae","hideRun":false},"source":["xor = [\n { input: [0, 0], output: [0]},\n { input: [0, 1], output: [1]},\n { input: [1, 0], output: [1]},\n { input: [1, 1], output: [1]}];\n \nlayer_xor = nm.linear(2, 1, true);\ntensor.LEARNING_RATE = 0.3;\nloss = 0"],"outputs":[]},{"cell_type":"code","metadata":{"id":"6fd7cadaff","hideRun":false},"source":["\nloss = 0\nfor (let i=0; i<1000; i++) {\n let x = xor[Math.floor(Math.random()*4)]\n \n // print (tensor.from(x.input) ) \n\n result = layer_xor(tensor.from(x.input));\n\n // print(` input = ${x.input.join(',')} result= ${result.data[0].toFixed(2)} ` )\n // print('sigm2',, , ;\n \n loss = result.MSE(x.output);\n // print('loss', loss.data[0]);\n loss.back();\n}\nprint('loss', loss.data[0]);"],"outputs":[{"data":{"text/plain":"loss\n0.1590086966753006"}}]},{"cell_type":"code","metadata":{"id":"82ab1f8db2","hideRun":false},"source":["\nloss = 0\nfor (let i=0; i<1000; i++) {\n let x = xor[Math.floor(Math.random()*4)]\n \n // print (tensor.from(x.input) ) \n\n result = layer_xor(tensor.from(x.input));\n\n // print(` input = ${x.input.join(',')} result= ${result.data[0].toFixed(2)} ` )\n // print('sigm2',, , ;\n \n loss = result.MSE(x.output);\n // print('loss', loss.data[0]);\n loss.back();\n}\nprint('loss', loss.data[0]);"],"outputs":[{"data":{"text/plain":" input = 0,0 result= 0.29 "}},{"data":{"text/plain":" input = 0,1 result= 0.96 "}},{"data":{"text/plain":" input = 1,0 result= 0.84 "}},{"data":{"text/plain":" input = 1,1 result= 1.51 "}}]}],"metadata":{},"nbformat":4,"nbformat_minor":2}