Tensorflow.js：检查目标时出错...期望层有 n 个维度

Question

我刚开始使用 Tensorflow.js，我正在尝试构建一个简单的模型，将 28 x 28 数组（每个数组代表一张图片）作为输入。但是有些东西连接不正确。运行下面的片段，我得到：

errors.ts:48 Uncaught (in promise) Error: Error when checking target: expected dense_Dense1 to have 2 dimension(s). but got array with shape 100,28,28
    at new e (errors.ts:48)
    at Od (training.ts:147)
    at e.standardizeUserData (training.ts:1133)
    at training_tensors.ts:427
    at common.ts:14
    at Object.next (common.ts:14)
    at common.ts:14
    at new Promise (<anonymous>)
    at op (common.ts:14)
    at kd (training_tensors.ts:408)

这是代码本身：

// build the model
var input = tf.input({shape: [28,28]})
var h1 = tf.layers.reshape({targetShape: [28*28]}).apply(input)
var h2 = tf.layers.dense({units: 100}).apply(h1)
var model = tf.model({inputs: input, outputs: h2})
model.compile({optimizer: 'sgd', loss: 'meanSquaredError', lr: 0.0001})
model.summary();

// get training data and train
var trainX = tf.ones([100,28,28]);

model.fit(trainX, trainX, {
  batchSize: 10,
  epochs: 1,
})

<script src='https://cdnjs.cloudflare.com/ajax/libs/tensorflow/1.1.2/tf.min.js'></script>

让我感到困惑的是 model.summary() 调用 returns:

_________________________________________________________________
layer_utils.ts:152 Layer (type)                 Output shape              Param #   
layer_utils.ts:64 =================================================================
layer_utils.ts:152 input1 (InputLayer)          [null,28,28]              0         
layer_utils.ts:74 _________________________________________________________________
layer_utils.ts:152 reshape_Reshape1 (Reshape)   [null,784]                0         
layer_utils.ts:74 _________________________________________________________________
layer_utils.ts:152 dense_Dense1 (Dense)         [null,100]                78500     
layer_utils.ts:74 =================================================================
layer_utils.ts:83 Total params: 78500
layer_utils.ts:84 Trainable params: 78500
layer_utils.ts:85 Non-trainable params: 0
layer_utils.ts:86 _________________________________________________________________

这表明重塑层应该将形状为 (batch, 784) 的数组传递给密集层，但错误提示并非如此。

有谁知道我在这里做错了什么？欢迎任何建议！

Answer 1

我的输入有形状（批次，28、28），而模型输出有形状（批次，100）。但是，我要求我的模型预测 trainX 给定输入 trainX（分别为 model.fit 的第二个和第一个参数）。

为了解决这个问题，我只需要将要预测的值的形状更新为（批次，100）：

// build the model
var input = tf.input({shape: [28,28]})
var h1 = tf.layers.reshape({targetShape: [28*28]}).apply(input)
var h2 = tf.layers.dense({units: 17}).apply(h1)
var model = tf.model({inputs: input, outputs: h2})
model.compile({optimizer: 'sgd', loss: 'meanSquaredError', lr: 0.0001})
model.summary();

// get training data and train
var trainX = tf.ones([100,28,28]),
    trainY = tf.ones([100, 17])

model.fit(trainX, trainY, {
  batchSize: 10,
  epochs: 1,
}).then(function() {
  console.log( model.predict(trainX).dataSync() )
})

Tensorflow.js：检查目标时出错...期望层有 n 个维度

Tensorflow.js: Error when checking target ... expected layer to have n dimensions

javascript

tensorflow

keras

tensorflow.js