Question

我正在尝试更改官方lstm-text-generation的tfjs-examples示例中使用的模型以输出其隐藏状态。为此，我在createModel中更改了model.js函数，如下所示：

export function createModel(sampleLen, charSetSize, lstmLayerSizes) {
  const inputs = tf.input({ shape: [sampleLen, charSetSize] });

  let outputs = inputs;
  for (let i = 0; i < lstmLayerSizes.length; ++i) {
    const lstmLayerSize = lstmLayerSizes[i];
    const layer = tf.layers.lstm({
      units: lstmLayerSize,
      // Not sure if this is necessary
      returnSequences: i < lstmLayerSizes.length - 1,
      // As far as I understood, this returns the state
      returnState: i === lstmLayerSizes.length - 1,
    });

    outputs = layer.apply(outputs);
  }

  // Destructure the hidden state (optionally also the cell state in the next step)
  const [lstmOutput, hiddenState] = outputs;

  outputs = tf.layers
    .dense({ units: charSetSize, activation: 'softmax' })
    .apply(lstmOutput);

  // Return the hidden state as part of the model outputs
  return tf.model({ inputs, outputs: [outputs, hiddenState] });
}

我不了解我需要如何更改.fit中fitModel的{{1}}参数：

model.js

训练后的模型似乎还不错。但是，当我将其可视化时，就像this Standford演讲中所做的那样，隐藏状态没有任何意义。

我在做什么有意义吗？

在LSTM Tensorflow.js模型的`.fit`中使用什么参数来输出隐藏状态和单元状态？

0 个答案: