Question

查看我的代码：

const cells = [
  tf.layers.simpleRNNCell({units: 4, activation: 'relu'}),
  tf.layers.simpleRNNCell({units: 3, activation: 'relu'}),
];


const startX = tf.tensor([
 [
   [10, 8, 5],
 ],
 [
   [0, 2, 0],
 ],
]);

const y = tf.tensor([
  [
    [0.3, 0.5, 0.4],
  ],
  [
    [0.1, 0, 0]
  ]
])


const initRnn = ({ inputShape, ...rnnProps }) => {
  const rnn = tf.layers.rnn({
    ...rnnProps,
    returnState: true
  });

  const train = (x, state) => {
    if (!state) {
      state = rnnProps.cell
        .map(({ units }) => tf.ones([x.shape[0], units]))
        .reverse();
    }
    const [yPred, ...newState] = rnn.apply(x, { initialState: state });
  
    return [yPred, newState];
  };

  return { train };
}
const optimizer = tf.train.rmsprop(0.1);

const getLoss = (y, yPred) =>
  tf.losses.softmaxCrossEntropy(y, yPred).mean();

const { train } = initRnn({
  cell: cells,
  returnSequences: true,
  inputShape: startX.shape
});

train(tf.zeros(startX.shape)); //to initialize weights

for (let i = 0; i < 10; i++) {
  let state = undefined;
  let x = startX.clone();

  optimizer.minimize(() => {
    for (let ts = 0; ts < 10; ts++) {
      const [yPred, newState] = train(x, state);

      state = newState;
      x = yPred.clone();

      if (ts === 9) {
        const loss = getLoss(y, yPred);
        loss.print()
        return loss; //the error is thrown here
      }
    };
  })
};

<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@latest"></script>

它引发错误Uncaught Error: Argument tensors passed to stack must be a Tensor[] or TensorLike[]

但是看看如果我一直使用startX作为输入会发生什么情况。

const cells = [
  tf.layers.simpleRNNCell({units: 4, activation: 'relu'}),
  tf.layers.simpleRNNCell({units: 3, activation: 'relu'}),
];


const startX = tf.tensor([
 [
   [10, 8, 5],
 ],
 [
   [0, 2, 0],
 ],
]);

const y = tf.tensor([
  [
    [0.3, 0.5, 0.4],
  ],
  [
    [0.1, 0, 0]
  ]
])


const initRnn = ({ inputShape, ...rnnProps }) => {
  const rnn = tf.layers.rnn({
    ...rnnProps,
    returnState: true
  });

  const train = (x, state) => {
    if (!state) {
      state = rnnProps.cell
        .map(({ units }) => tf.ones([x.shape[0], units]))
        .reverse();
    }
    const [yPred, ...newState] = rnn.apply(x, { initialState: state });
  
    return [yPred, newState];
  };

  return { train };
}
const optimizer = tf.train.rmsprop(0.1);

const getLoss = (y, yPred) =>
  tf.losses.softmaxCrossEntropy(y, yPred).mean();

const { train } = initRnn({
  cell: cells,
  returnSequences: true,
  inputShape: startX.shape
});

train(tf.zeros(startX.shape)); //to initialize weights

for (let i = 0; i < 10; i++) {
  let state = undefined;
  let x = startX.clone();

  optimizer.minimize(() => {
    for (let ts = 0; ts < 10; ts++) {
      const [yPred, newState] = train(startX, state);

      state = newState;
      x = yPred.clone();

      if (ts === 9) {
        const loss = getLoss(y, yPred);
        loss.print()
        return loss;
      }
    };
  })
};

<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@latest"></script>

代码贯穿始终。这两个代码的唯一区别是失败的地方，我将以前的时间步结果用作输入，传递的地方，startX一直被用作输入。

令人惊讶的是，它在优化器回调内部返回时引发错误。因此，重用先前的输出10次（在返回损失的位置）没有问题，并且权重应该由优化器调整。但是，如果使用先前的输出作为输入在10个时间步内没有引起任何问题，为什么在计算梯度时会这样做呢？

Tensorflow.js优化器-传递到堆栈的参数张量必须是`Tensor []`或`TensorLike []`

0 个答案: