jsdoでdeeplearn.js その3

概要

jsdoでdeeplearn.jsやってみた。
学習させてみた。

参考にしたページ

写真

サンプルコード

function mylog(txt) {
    document.getElementById('helloWorld').innerHTML += txt + "<BR>";
}
var dl = deeplearn;
var g = new dl.Graph();
var math = new dl.NDArrayMathCPU();
function leaky_relu(x) {
    var mx = g.multiply(g.constant(-1), g.relu(x));
    var neg_part = g.multiply(g.constant(-0.2), g.relu(mx));
    return g.add(g.relu(x) , neg_part)
}
var x = g.placeholder('x', []);
var W1data = dl.Array2D.randNormal([16, 1]);
var b1data = dl.Array2D.zeros([16, 1]);
var W2data = dl.Array2D.randNormal([32, 16], 0, Math.sqrt(1.0 / 16.0));
var b2data = dl.Array2D.zeros([32, 1]);
var W3data = dl.Array2D.randNormal([1, 32], 0, Math.sqrt(1.0 / 32.0));
var b3data = dl.Array2D.zeros([1, 1]);
var W1 = g.variable('W1', W1data);
var b1 = g.variable('b1', b1data);
var W2 = g.variable('W2', W2data);
var b2 = g.variable('b2', b2data);
var W3 = g.variable('W3', W3data);
var b3 = g.variable('b3', b3data);
var h1 = leaky_relu(g.add(g.multiply(W1, x), b1));
var h2 = leaky_relu(g.add(g.matmul(W2, h1), b2));
var y_ = leaky_relu(g.add(g.matmul(W3, h2), b3));
var y = g.reshape(y_,[]);
var yLabel = g.placeholder('y label', []);
var cost = g.meanSquaredCost(y, yLabel);
var sess = new dl.Session(g, math);
math.scope((keep, track) => {
    var xs = [];
    var ys = [];
    for (var i = 0; i < 100; i++)
    {
        var xr = Math.random();
        xs.push(track(dl.Scalar.new(xr)));
        ys.push(track(dl.Scalar.new(Math.exp(xr))));
    }
    var shuffledInputProviderBuilder = new dl.InCPUMemoryShuffledInputProviderBuilder([xs, ys]);
    var [xProvider, yProvider] = shuffledInputProviderBuilder.getInputProviders();
    var NUM_BATCHES = 100;
    var BATCH_SIZE = xs.length;
    var LEARNING_RATE = 0.01;
    var optimizer = new dl.SGDOptimizer(LEARNING_RATE);
    var startTime = new Date();
    for (var i = 0; i < NUM_BATCHES; i++) 
    {
        var costValue = sess.train(cost, [{ 
            tensor: x, 
            data: xProvider
        }, { 
            tensor: yLabel,
            data: yProvider
        }], BATCH_SIZE, optimizer, dl.CostReduction.MEAN);
        var cost_val = costValue.get();
    }
    var endTime = new Date();
    var timeDiff = endTime.getTime() - startTime.getTime();
    var predicted = sess.eval(y, [{ 
        tensor: x,
        data: track(dl.Scalar.new(0.2))
    }]).getValues();
    mylog('--- prediction check ---');
    mylog('predicted : ' + predicted);
    mylog('truth     :' + Math.exp(0.2));
    mylog('');
    mylog('--- benchmark result ---');
    mylog('elasped time for training: ' + timeDiff / 1000 +'[sec]');
    var startTime = new Date();
    for (let i = 0; i < 1000; i++) 
    {
        var xdata = dl.Scalar.new(Math.random());
        var predicted = sess.eval(y, [{ 
            tensor: x, 
            data: track(xdata)
        }]).getValues();
    }
    var endTime = new Date();
    var timeDiff = endTime.getTime() - startTime.getTime();
    mylog('elasped time for prediction: ' + timeDiff / 1000.0 + '[msec/cycle]');
});

成果物