问题描述
我正在尝试构建CNN，并陷入MaxPooling3D图层无法正常工作的情况。这两层的输入形状均为（1、5、32），我想使用poolsize（1、1、32）在深度上进行最大池化，以便输出变为形状（1、5、1）。但这会引发错误：
ValueError: Input 0 of layer max_pooling3d is incompatible with the
layer: expected ndim=5,found ndim=4. Full shape received: [None,1,5,32]
我不明白为什么期望/要求尺寸为5。如果我改为使用具有poolsize（1,1）的MaxPooling2D图层，则所有内容都可以正确编译，并且可以得到下面的模型。
> Model: "functional_1"
> __________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
> ==================================================================================================
input_1 (InputLayer) [(None,1)] 0
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None,32) 192 input_1[0][0]
__________________________________________________________________________________________________
conv2d (Conv2D) (None,32) 192 input_1[0][0]
__________________________________________________________________________________________________
reshape (Reshape) (None,32) 0 conv2d_1[0][0]
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D) (None,32) 0 conv2d[0][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None,32) 0 reshape[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None,64) 0 max_pooling2d[0][0]
max_pooling2d_1[0][0]
==================================================================================================
Total params: 384 Trainable params: 384 Non-trainable params: 0
__________________________________________________________________________________________________
Process finished with exit code 0
n=5
inp_similarity = Input(shape=(n,n,1))
conv11 = Conv2D(32,(n,1))(inp_similarity)
conv12 = Conv2D(32,(1,n))(inp_similarity)
reshape1 = Reshape((1,32))(conv12)
maxpl11 = MaxPooling2D((1,1))(conv11)
maxpl12 = MaxPooling2D((1,1))(reshape1)
merge1 = Concatenate()([maxpl11,maxpl12])
model = Model(inp_similarity,merge1)
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
model.summary()
解决方法
您的目标是在特征维上进行“池化” ...这不是池化层的范围...它们仅在空间维上进行池化。您需要更简单的东西
n=5
inp_similarity = Input(shape=(n,n,1))
conv11 = Conv2D(32,(n,1))(inp_similarity)
conv12 = Conv2D(32,(1,n))(inp_similarity)
reshape1 = Reshape((1,5,32))(conv12)
maxpl11 = Lambda(lambda x: tf.reduce_max(x,axis=-1,keepdims=True))(conv11)
maxpl12 = Lambda(lambda x: tf.reduce_max(x,keepdims=True))(reshape1)
merge1 = Concatenate()([maxpl11,maxpl12])
model = Model(inp_similarity,merge1)
model.compile(optimizer='adam',loss='binary_crossentropy',metrics=['accuracy'])
model.summary()
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_4 (InputLayer) [(None,1)] 0
__________________________________________________________________________________________________
conv2d_35 (Conv2D) (None,1,32) 192 input_4[0][0]
__________________________________________________________________________________________________
conv2d_34 (Conv2D) (None,32) 192 input_4[0][0]
__________________________________________________________________________________________________
reshape_2 (Reshape) (None,32) 0 conv2d_35[0][0]
__________________________________________________________________________________________________
lambda_6 (Lambda) (None,1) 0 conv2d_34[0][0]
__________________________________________________________________________________________________
lambda_7 (Lambda) (None,1) 0 reshape_2[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None,2) 0 lambda_6[0][0]
lambda_7[0][0]
==================================================================================================