预期 min_ndim=2，发现 ndim=1。收到完整形状：（无，）

Question

在我的模型中，我有一个用于 1 列特征数组的归一化层。我假设这给出了 1 ndim 输出：

single_feature_model = keras.models.Sequential([
    single_feature_normalizer,
    layers.Dense(1)
])

正常化步骤：

single_feature_normalizer = preprocessing.Normalization(axis=None)
single_feature_normalizer.adapt(single_feature)

我收到的错误是：

ValueError                                Traceback (most recent call last)
<ipython-input-98-22191285d676> in <module>()
      2 single_feature_model = keras.models.Sequential([
      3     single_feature_normalizer,
----> 4     layers.Dense(1) # Linear Model
      5 ])

/usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    225       ndim = x.shape.rank
    226       if ndim is not None and ndim < spec.min_ndim:
--> 227         raise ValueError(f'Input {input_index} of layer "{layer_name}" '
    228                          'is incompatible with the layer: '
    229                          f'expected min_ndim={spec.min_ndim}, '

ValueError: Input 0 of layer "dense_27" is incompatible with the layer: expected min_ndim=2, found ndim=1. Full shape received: (None,)

我似乎密集层正在寻找 2 ndim 数组，而归一化层输出 1 ndim 数组。有没有办法解决这个问题并使模型正常工作？

Answer 1

我认为您需要使用输入形状显式定义输入层，因为您的输出层无法推断来自归一化层的张量的形状：

import tensorflow as tf

single_feature_normalizer = tf.keras.layers.Normalization(axis=None)
feature = tf.random.normal((314, 1))
single_feature_normalizer.adapt(feature)

single_feature_model = tf.keras.models.Sequential([
    tf.keras.layers.Input(shape=(1,)),
    single_feature_normalizer,
    tf.keras.layers.Dense(1)
])

或者直接在归一化层中定义输入形状，而不使用输入层：

single_feature_normalizer = tf.keras.layers.Normalization(input_shape=[1,], axis=None)