摘要：原博地址源碼演示開(kāi)始時(shí)需要加載大概的訓(xùn)練數(shù)據(jù)稍等片刻調(diào)整訓(xùn)練集的大小觀察測(cè)試結(jié)果的準(zhǔn)確性數(shù)據(jù)來(lái)源數(shù)據(jù)來(lái)源與中的一道題目題目給出條訓(xùn)練數(shù)據(jù)包含圖片和標(biāo)簽以及條測(cè)試數(shù)據(jù)只包含圖片要求給這些測(cè)試數(shù)據(jù)打上標(biāo)簽要盡可能的準(zhǔn)確網(wǎng)站中還有許多其

原博地址https://laboo.top/2018/11/21/tfjs-dr/

digit-recognizer

https://github-laziji.github.io/digit-recognizer/
演示開(kāi)始時(shí)需要加載大概100M的訓(xùn)練數(shù)據(jù), 稍等片刻

調(diào)整訓(xùn)練集的大小, 觀察測(cè)試結(jié)果的準(zhǔn)確性

數(shù)據(jù)來(lái)源與 https://www.kaggle.com 中的一道題目 digit-recognizer
題目給出42000條訓(xùn)練數(shù)據(jù)(包含圖片和標(biāo)簽)以及28000條測(cè)試數(shù)據(jù)(只包含圖片)
要求給這些測(cè)試數(shù)據(jù)打上標(biāo)簽[0,1,2,3....,9] 要盡可能的準(zhǔn)確

網(wǎng)站中還有許多其他的機(jī)器學(xué)習(xí)的題目以及數(shù)據(jù), 是個(gè)很好的練手的地方

這里我們使用TensorFlow.js來(lái)實(shí)現(xiàn)這個(gè)項(xiàng)目

卷積神經(jīng)網(wǎng)絡(luò)的第一層有兩種作用, 它既是輸入層也是執(zhí)行層, 接收IMAGE_H * IMAGE_W大小的黑白像素
最后一層是輸出層, 有10個(gè)輸出單元, 代表著0-9這十個(gè)值的概率分布, 例如 Label=2 , 輸出為[0.02,0.01,0.9,...,0.01]

function createConvModel() {
  const model = tf.sequential();

  model.add(tf.layers.conv2d({
    inputShape: [IMAGE_H, IMAGE_W, 1],
    kernelSize: 3,
    filters: 16,
    activation: "relu"
  }));

  model.add(tf.layers.maxPooling2d({ poolSize: 2, strides: 2 }));
  model.add(tf.layers.conv2d({ kernelSize: 3, filters: 32, activation: "relu" }));
  model.add(tf.layers.maxPooling2d({ poolSize: 2, strides: 2 }));
  model.add(tf.layers.conv2d({ kernelSize: 3, filters: 32, activation: "relu" }));
  model.add(tf.layers.flatten({}));

  model.add(tf.layers.dense({ units: 64, activation: "relu" }));
  model.add(tf.layers.dense({ units: 10, activation: "softmax" }));

  return model;
}

我們選擇適當(dāng)?shù)膬?yōu)化器和損失函數(shù), 來(lái)編譯模型

async function train() {

  ui.trainLog("Create model...");
  model = createConvModel();
  
  ui.trainLog("Compile model...");
  const optimizer = "rmsprop";
  model.compile({
    optimizer,
    loss: "categoricalCrossentropy",
    metrics: ["accuracy"],
  });
  const trainData = Data.getTrainData(ui.getTrainNum());
  
  ui.trainLog("Training model...");
  await model.fit(trainData.xs, trainData.labels, {});

  ui.trainLog("Completed!");
  ui.trainCompleted();
}

這里測(cè)試一組測(cè)試數(shù)據(jù), 返回對(duì)應(yīng)的標(biāo)簽, 即十個(gè)輸出單元中概率最高的下標(biāo)

function testOne(xs){
  if(!model){
    ui.viewLog("Need to train the model first");
    return;
  }
  ui.viewLog("Testing...");
  let output = model.predict(xs);
  ui.viewLog("Completed!");
  output.print();
  const axis = 1;
  const predictions = output.argMax(axis).dataSync();
  return predictions[0];
}

歡迎關(guān)注我的博客公眾號(hào)