python - 一维卷积层尺寸不匹配问题

Question

我正在尝试了解卷积层在神经网络中的工作原理。我针对类似问题找到了两个不同的相关帖子并尝试了这些建议，但无法解决。

1. Keras dimensionality in convolutional layer mismatch
2. ValueError: Input 0 is incompatible with layer conv_1: expected ndim=3, found ndim=4

import tensorflow as tf
model_valid = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(10,)),
    tf.keras.layers.Dense(16,  activation='relu'),
    tf.keras.layers.Conv1D(16, kernel_size=(2), activation='relu', padding='same'),     
    tf.keras.layers.MaxPooling1D(pool_size=(4), strides=3, padding='valid'),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(1, activation='softmax')
])
model_valid.summary()

我收到一个不兼容问题，因为层 conv1d_37 的输入 0 与该层不兼容:预期 ndim=3，发现 ndim=2。收到的完整形状:[无，16]。构建卷积层时会出现问题。

Answer 1

第一个 Conv1D 之前的输入形状不正确。正如用户 spb 所建议的那样，Conv1D 的输入必须是 3D 张量，尽管移除展平层(如建议的那样)不会解决问题。

相反，只需在第一个 Dense 层之后添加一个额外的维度。

import tensorflow as tf
model_valid = tf.keras.Sequential([
    tf.keras.layers.Flatten(input_shape=(10,)),
    tf.keras.layers.Dense(16,  activation='relu'),
    tf.keras.layers.Reshape((16, 1)), # ADD THIS LINE OF CODE
    tf.keras.layers.Conv1D(16, kernel_size=(2), activation='relu', padding='same'),     
    tf.keras.layers.MaxPooling1D(pool_size=(4), strides=3, padding='valid'),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(1, activation='softmax')
])
model_valid.summary()

现在模型总结是:

Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
flatten_4 (Flatten)          (None, 10)                0
_________________________________________________________________
dense_4 (Dense)              (None, 16)                176
_________________________________________________________________
reshape (Reshape)            (None, 16, 1)             0
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 16, 16)            48
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 5, 16)             0
_________________________________________________________________
flatten_5 (Flatten)          (None, 80)                0
_________________________________________________________________
dense_5 (Dense)              (None, 1)                 81
=================================================================
Total params: 305
Trainable params: 305
Non-trainable params: 0
_________________________________________________________________