Tensorflow中CNN的MaxPooling1D与输入维度不兼容问题排查

我有形状为(10000 x 2 x 51 x 1)的2D信号数据，将其视为高为2、宽为51的图像构建CNN架构，代码如下：

input_layer = Input(shape=(x_train.shape[1], x_train.shape[2], x_train.shape[3]))
x = Conv2D(filters=32, kernel_size=(2, 3))(input_layer)
x = PReLU()(x)
x = MaxPooling1D(pool_size=2)(x)
x = Dropout(0.2)(x)
x = Flatten()(x)
x = Dense(units=256)(x)
x = PReLU()(x)
output_layer = Dense(units=1, activation='linear')(x)

执行时第一个MaxPooling1D层报错：

ValueError: Input 0 of layer "max_pooling1d" is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: (None, 1, 49, 32)

我认为第一次Conv2D操作后数据形状变为(1, 49, 32)属于1D数据，但不清楚形状表示哪里出错。改用MaxPooling2D时又出现错误：

ValueError: Exception encountered when calling layer "max_pooling2d" (type MaxPooling2D).
Negative dimension size caused by subtracting 2 from 1 for '{{node max_pooling2d/MaxPool}} = MaxPoolT=DT_FLOAT, data_format="NHWC", explicit_paddings=[], ksize=[1, 2, 2, 1], padding="VALID", strides=[1, 2, 2, 1]' with input shapes: [?,1,49,32].

问题根源 #

1. MaxPooling1D维度不匹配：Conv2D输出的是4维张量(None, height, width, channels)（即(None,1,49,32)），而MaxPooling1D要求输入是3维张量(None, steps, features)，两者维度结构完全不同，直接使用必然报错。你误以为(1,49,32)是1D数据，但它本质还是包含高度、宽度、通道的2D空间特征图结构。

2. MaxPooling2D负维度错误：默认MaxPooling2D的池化窗口是(2,2)，但此时特征图的高度已经被Conv2D压缩到1，用(2,2)窗口做VALID padding时，计算后高度维度变为1-2=-1，出现非法负维度导致报错。

修正方案 #

方案一：适配特征图尺寸的MaxPooling2D #

既然Conv2D后高度已为1，只需对宽度维度做池化，指定池化窗口为(1,2)即可：

input_layer = Input(shape=(x_train.shape[1], x_train.shape[2], x_train.shape[3]))
x = Conv2D(filters=32, kernel_size=(2, 3))(input_layer)
x = PReLU()(x)
# 仅对宽度维度池化，避免高度维度报错
x = MaxPooling2D(pool_size=(1, 2))(x)
x = Dropout(0.2)(x)
x = Flatten()(x)
x = Dense(units=256)(x)
x = PReLU()(x)
output_layer = Dense(units=1, activation='linear')(x)

方案二：转3维结构后用MaxPooling1D #

先通过Reshape去掉高度维度，将4维张量转为MaxPooling1D要求的3维结构：

input_layer = Input(shape=(x_train.shape[1], x_train.shape[2], x_train.shape[3]))
x = Conv2D(filters=32, kernel_size=(2, 3))(input_layer)
x = PReLU()(x)
# 把(None,1,49,32)转为(None,49,32)，适配MaxPooling1D输入要求
x = Reshape((x.shape[2], x.shape[3]))(x)
x = MaxPooling1D(pool_size=2)(x)
x = Dropout(0.2)(x)
x = Flatten()(x)
x = Dense(units=256)(x)
x = PReLU()(x)
output_layer = Dense(units=1, activation='linear')(x)

方案三：调整初始Conv2D参数，保留高度维度 #

修改Conv2D的卷积核尺寸为(1,3)，只在宽度方向卷积，保留高度维度为2，后续可正常使用MaxPooling2D：

input_layer = Input(shape=(x_train.shape[1], x_train.shape[2], x_train.shape[3]))
# 用(1,3)卷积核，避免高度维度被压缩到1
x = Conv2D(filters=32, kernel_size=(1, 3))(input_layer)
x = PReLU()(x)
x = MaxPooling2D(pool_size=(2, 2))(x)
x = Dropout(0.2)(x)
x = Flatten()(x)
x = Dense(units=256)(x)
x = PReLU()(x)
output_layer = Dense(units=1, activation='linear')(x)

内容的提问来源于stack exchange，提问作者justRandomLearner