Tensorflow 模型未在 javascript 中训练

Question

TensorFlow.js 版本

1.4.0。

描述

我正在尝试用 javascript 训练模型，但该模型无法学习或收敛。 我在程序的python版本中使用了相同的模型和相同的数据，因此我希望模型在同一阶段学习。 相反，模型没有改进，第一次运行后验证精度保持不变。 python 模型能够达到约 70% 的准确率，而 javascript 模型在 50 个 epoch 后几乎没有达到超过 5%。 如果您想使用相同的数据，则 URL 有效。

重现错误的代码

蟒蛇代码：

 checkpoint = ModelCheckpoint('best_models/model--{val_accuracy:03f}--{epoch:03d}-{accuracy:03f}.h5', verbose=1, monitor='val_accuracy',save_best_only=True, mode='auto') X_train_raw = requests.get("http://tb-test.chatbotech.com/info/get-training-arrays").json().get("xTrain") X_test_raw = requests.get("http://tb-test.chatbotech.com/info/get-training-arrays").json().get("xTest") y_train_raw = requests.get("http://tb-test.chatbotech.com/info/get-training-arrays").json().get("yTrain") y_test_raw = requests.get("http://tb-test.chatbotech.com/info/get-training-arrays").json().get("yTest") X_train = np.array(ast.literal_eval(X_train_raw)) X_test = np.array(ast.literal_eval(X_test_raw)) y_train_hot = np.array(ast.literal_eval(y_train_raw)) y_test_hot = np.array(ast.literal_eval(y_test_raw)) max_pad_length = 220 model = Sequential() model.add(Conv2D(128, kernel_size=(8, 48), activation='relu', input_shape=(20, max_pad_length, 1))) model.add(MaxPooling2D(pool_size=(3, 120))) model.add(Dropout(0.2)) model.add(Dense(128, activation='relu')) model.add(Dropout(0.3)) model.add(Flatten()) model.add(Dense(30, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adadelta(), metrics=['accuracy']) history = model.fit(X_train, y_train_hot, batch_size=20, epochs=2000, verbose=1, validation_data=(X_test, y_test_hot),callbacks=[checkpoint])

Javascript代码：

 async function getData() { const dataReq = await fetch('http://tb-test.chatbotech.com/info/get-training-arrays'); const trainData = await dataReq.json(); return trainData; } async function run() { // Load and plot the original input data that we are going to train on. const data = await getData(); console.log(data); const model = createModel(); // More code will be added below model.fit(tf.tensor(JSON.parse(data.xTrain), [230, 20, 220, 1], 'float32'), tf.tensor(JSON.parse(data.yTrain), [230, 30]), { shuffle: false, epochs: 2000, validationData: [tf.tensor(JSON.parse(data.xTest), [154, 20, 220, 1], 'float32'), tf.tensor(JSON.parse(data.yTest), [154, 30])], callbacks: { async onEpochEnd(epoch, logs) { console.log(logs); }, onBatchEnd(batch, logs) { console.log(logs); console.log(batch); }}}); } function createModel() { const model = tf.sequential(); model.add(tf.layers.conv2d({filters: 128, kernelSize: [8, 48], activation: 'relu', inputShape: [20, 220, 1], strides: [1, 1], padding: 'valid'})); model.add(tf.layers.maxPooling2d({poolSize: [3, 120], strides: [3, 120]})); model.add(tf.layers.dropout({rate: 0.2})); model.add(tf.layers.dense({units: 128, activation: 'relu'})); model.add(tf.layers.dropout({rate: 0.3})); model.add(tf.layers.flatten()); model.add(tf.layers.dense({units: 30, activation: 'softmax'})); model.compile({loss: tf.metrics.categoricalCrossentropy, optimizer: tf.train.adadelta(1, 0.95, 1e-07 ), metrics: ['accuracy']}); return model; } document.addEventListener('DOMContentLoaded', run);

Answer 1

您也可以尝试这种架构。 我认为这将有助于提高您的模型准确性：

async function* data() {
  while (true) {
    for (i in train) {
      // this function return tensor data
    }
  }
}
async function* labels() {
  while (true) {
    for (i in train) {
      // this function return tensor label
    }
  }
}

async function initModel() {
   // model write here
}


(async function () {

const xs = tf.data.generator(data);
const ys = tf.data.generator(labels);

const model = await initModel();
model.summary();

 await model.fit(xs, ys,{    
    epochs: 5
    batchesPerEpoch: 5 
  });
})()

提供您的反馈。