检查输入时出错：预期 dense_Dense5_input 有 4 个维度。但得到了形状为 5,2,5 的数组答案

【问题标题】：Error when checking input: expected dense_Dense5_input to have 4 dimension(s). but got array with shape 5,2,5检查输入时出错：预期 dense_Dense5_input 有 4 个维度。但得到了形状为 5,2,5 的数组
【发布时间】：2018-10-13 19:52:00
【问题描述】：

我正在学习 tensorflow.js，我正在尝试创建一个模型，该模型可以根据他们的“玩家”预测 2 个“团队”之间随机比赛/游戏的获胜者。

const rawMatches = [
  {
    t1: [2, 99, 3, 5, 7],
    t2: [4, 75, 48, 23, 6],
    winner: 0
  },
  {
    t1: [2, 99, 48, 5, 7],
    t2: [4, 75, 3, 23, 6],
    winner: 1
  },
  {
    t1: [2, 83, 3, 4, 23],
    t2: [4, 75, 58, 25, 78],
    winner: 0
  },
  {
    t1: [26, 77, 11, 5, 7],
    t2: [3, 43, 48, 23, 9],
    winner: 1
  },
  {
    t1: [2, 99, 3, 5, 7],
    t2: [6, 65, 28, 23, 6],
    winner: 0
  }
];

const train = async () => {
  //   [
  //     [[2, 99, 3, 5, 7], [4, 75, 48, 23, 6]],
  //     [[2, 99, 48, 5, 7], [4, 75, 3, 23, 6]],
  //     [[2, 99, 3, 5, 7], [4, 75, 48, 23, 6]]
  //   ];
  const xs = tf.tensor3d(
    rawMatches.map((match, index) => [match.t1, match.t2])
  );

  //   [[1, 0], [0, 1], [1, 0]];
  const labelsTensor = tf.tensor1d(
    rawMatches.map(match => (match.winner === 1 ? 1 : 0)),
    "int32"
  );

  const ys = tf.oneHot(labelsTensor, 2);

  xs.print();
  ys.print();

  let model = tf.sequential();
  const hiddenLayer = tf.layers.dense({
    units: 15,
    activation: "sigmoid",
    inputShape: [5, 2, 5]
  });
  const outputLayer = tf.layers.dense({
    units: 2,
    activation: "softmax"
  });
  model.add(hiddenLayer);
  model.add(outputLayer);

  const optimizer = tf.train.sgd(0.2);

  model.compile({
    optimizer,
    loss: "categoricalCrossentropy"
  });

  model.fit(xs, ys, { epochs: 1 });
};

train();

<html>
  <head>
    <!-- Load TensorFlow.js -->
    <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@0.13.0"> </script>
  </head>

  <body>
  </body>
</html>

尝试拟合模型后会出现此错误：

Error when checking input: expected dense_Dense11_input to have 4 dimension(s). but got array with shape 5,2,5

包含完整代码的代码沙箱：https://codesandbox.io/s/kr37m63w7

【问题讨论】：

标签： javascript tensorflow machine-learning tensorflow.js

【解决方案1】：

这个模型有两个问题：

首先将输入 x 的维度传递给方法fit。 xs 应该比第一个 inputShape 高一个维度。因为xs 是一个包含inputShape 形状数据的数组，所以inputShape 应该是[2, 5]。

第二，由于输入和输出的维度不匹配，需要使用tf.flatten改变数据的维度。两个维度不匹配，因为输入数据形状为[2, 5] (size = 2)，而输出数据形状为[2] (size = 1)

const rawMatches = [
  {
    t1: [2, 99, 3, 5, 7],
    t2: [4, 75, 48, 23, 6],
    winner: 0
  },
  {
    t1: [2, 99, 48, 5, 7],
    t2: [4, 75, 3, 23, 6],
    winner: 1
  },
  {
    t1: [2, 83, 3, 4, 23],
    t2: [4, 75, 58, 25, 78],
    winner: 0
  },
  {
    t1: [26, 77, 11, 5, 7],
    t2: [3, 43, 48, 23, 9],
    winner: 1
  },
  {
    t1: [2, 99, 3, 5, 7],
    t2: [6, 65, 28, 23, 6],
    winner: 0
  }
];

const train = () => {
  const xs = tf.tensor3d(
    rawMatches.map((match, index) => [match.t1, match.t2])
  );
  const labelsTensor = tf.tensor1d(
    rawMatches.map(match => (match.winner === 1 ? 1 : 0)),
    "int32"
  );

  const ys = tf.oneHot(labelsTensor, 2);

  xs.print();
  ys.print();

  let model = tf.sequential();
  const hiddenLayer = tf.layers.dense({
    units: 15,
    activation: "sigmoid",
    inputShape: [2, 5]
  });
  const outputLayer = tf.layers.dense({
    units: 2,
    activation: "softmax"
  });
  model.add(hiddenLayer);
  model.add(tf.layers.flatten())
  model.add(outputLayer);

  const optimizer = tf.train.sgd(0.2);

  model.compile({
    optimizer,
    loss: "categoricalCrossentropy"
  });

  model.fit(xs, ys, { epochs: 1 });
};

train();

<html>
  <head>
    <!-- Load TensorFlow.js -->
    <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@0.13.0"> </script>
  </head>

  <body>
  </body>
</html>

【讨论】：