ValueError：dense_24 层的输入 0 不兼容：输入形状的预期轴 -1 具有值 1024，但接收到形状为 [16, 512] 的输入

Question

我对使用 tensorflow 还是很陌生，非常感谢您对此提供一些帮助。我正在训练自动编码器，并尝试使用 tensorflow.data 管道加载数据输入。但是，这样做之后，我一直遇到输入形状等问题。有谁知道如何解决这个问题？非常感谢！！！

datatrain1.shape 为 (18820, 16, 256, 1)
这就是我定义数据集的方式 dataset = tf.data.Dataset.from_tensor_slices((datatrain1, datatrain1))
数据集的形状如下：
<TensorSliceDataset shapes: ((16, 256, 1), (16, 256, 1)), types: (tf.float32, tf.float32)>

自动编码器代码 autoencoder.compile("adam", loss="mse")
autoencoder.fit(dataset,epochs=1, shuffle=True)
上面的 fit 调用给了我错误：

ValueError: Input 0 of layer dense_24 is incompatible with the layer: expected axis -1 of input shape to have value 1024 but received input with shape [16, 512]

这是模型定义代码：

n_clusters = 32

input_img = Input(shape=(16, 256, 1))
x = res_conv_block(input_img, 64, 2)

pool_1 = MaxPooling2D((1, 2), padding="same")(x)
x = res_conv_block(pool_1, 64, 2)

pool_2 = MaxPooling2D((2, 2), padding="same")(x)
x = res_conv_block(pool_2, 64, 2)

x = Conv2D(8, (3, 3), activation="relu", padding="same")(x)
x = MaxPooling2D((2, 2), padding="same")(x)


flat = Flatten()(x)
x = Dense(256, activation='relu')(flat)
encoded = Dense(64, activation='relu', name="encoded")(x)
x = Dense(256, activation='relu')(encoded)
x = Dense(1024, activation='relu')(x)
x = Reshape((4, 32, 8))(x)

x = UpSampling2D((2, 2))(x)
up_1 = Conv2DTranspose(64, (3, 3), activation="relu", padding="same")(x)

x = res_deconv_block(up_1, 64, 2)
up_2 = UpSampling2D((2, 2))(x)

x = res_deconv_block(up_2, 64, 2)
up_3 = UpSampling2D((1, 2))(x)

x = res_deconv_block(up_3, 64, 2)
decoded = Conv2DTranspose(1, (3, 3), padding="same", name="decoded")(x)

autoencoder= Model(inputs=input_img, outputs=decoded, name="autoencoder")
encoder = Model(inputs=input_img, outputs=encoded, name="encoder")
clustering_layer = ClusteringLayer(n_clusters, name='clustering_layer')(encoder.output)
# SOM_layer = 
idec = Model(inputs=autoencoder.input, outputs=[clustering_layer, decoded], name="res-idec")
#SOM_layer = model.add(SOM_Layer())

模型总结：

Model: "res-idec"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_11 (InputLayer)           [(None, 16, 256, 1)] 0                                            
__________________________________________________________________________________________________
conv2d_58 (Conv2D)              (None, 16, 256, 64)  640         input_11[0][0]                   
__________________________________________________________________________________________________
conv2d_59 (Conv2D)              (None, 16, 256, 64)  36928       conv2d_58[0][0]                  
__________________________________________________________________________________________________
add_48 (Add)                    (None, 16, 256, 64)  0           conv2d_59[0][0]                  
                                                                 input_11[0][0]                   
__________________________________________________________________________________________________
max_pooling2d_24 (MaxPooling2D) (None, 16, 128, 64)  0           add_48[0][0]                     
__________________________________________________________________________________________________
conv2d_60 (Conv2D)              (None, 16, 128, 64)  36928       max_pooling2d_24[0][0]           
__________________________________________________________________________________________________
conv2d_61 (Conv2D)              (None, 16, 128, 64)  36928       conv2d_60[0][0]                  
__________________________________________________________________________________________________
add_49 (Add)                    (None, 16, 128, 64)  0           conv2d_61[0][0]                  
                                                                 max_pooling2d_24[0][0]           
__________________________________________________________________________________________________
max_pooling2d_25 (MaxPooling2D) (None, 8, 64, 64)    0           add_49[0][0]                     
__________________________________________________________________________________________________
conv2d_62 (Conv2D)              (None, 8, 64, 64)    36928       max_pooling2d_25[0][0]           
__________________________________________________________________________________________________
conv2d_63 (Conv2D)              (None, 8, 64, 64)    36928       conv2d_62[0][0]                  
__________________________________________________________________________________________________
add_50 (Add)                    (None, 8, 64, 64)    0           conv2d_63[0][0]                  
                                                                 max_pooling2d_25[0][0]           
__________________________________________________________________________________________________
conv2d_64 (Conv2D)              (None, 8, 64, 8)     4616        add_50[0][0]                     
__________________________________________________________________________________________________
max_pooling2d_26 (MaxPooling2D) (None, 4, 32, 8)     0           conv2d_64[0][0]                  
__________________________________________________________________________________________________
flatten_8 (Flatten)             (None, 1024)         0           max_pooling2d_26[0][0]           
__________________________________________________________________________________________________
dense_24 (Dense)                (None, 256)          262400      flatten_8[0][0]                  
__________________________________________________________________________________________________
encoded (Dense)                 (None, 64)           16448       dense_24[0][0]                   
__________________________________________________________________________________________________
dense_25 (Dense)                (None, 256)          16640       encoded[0][0]                    
__________________________________________________________________________________________________
dense_26 (Dense)                (None, 1024)         263168      dense_25[0][0]                   
__________________________________________________________________________________________________
reshape_8 (Reshape)             (None, 4, 32, 8)     0           dense_26[0][0]                   
__________________________________________________________________________________________________
up_sampling2d_24 (UpSampling2D) (None, 8, 64, 8)     0           reshape_8[0][0]                  
__________________________________________________________________________________________________
conv2d_transpose_56 (Conv2DTran (None, 8, 64, 64)    4672        up_sampling2d_24[0][0]           
__________________________________________________________________________________________________
conv2d_transpose_57 (Conv2DTran (None, 8, 64, 64)    36928       conv2d_transpose_56[0][0]        
__________________________________________________________________________________________________
conv2d_transpose_58 (Conv2DTran (None, 8, 64, 64)    36928       conv2d_transpose_57[0][0]        
__________________________________________________________________________________________________
add_51 (Add)                    (None, 8, 64, 64)    0           conv2d_transpose_58[0][0]        
                                                                 conv2d_transpose_56[0][0]        
__________________________________________________________________________________________________
up_sampling2d_25 (UpSampling2D) (None, 16, 128, 64)  0           add_51[0][0]                     
__________________________________________________________________________________________________
conv2d_transpose_59 (Conv2DTran (None, 16, 128, 64)  36928       up_sampling2d_25[0][0]           
__________________________________________________________________________________________________
conv2d_transpose_60 (Conv2DTran (None, 16, 128, 64)  36928       conv2d_transpose_59[0][0]        
__________________________________________________________________________________________________
add_52 (Add)                    (None, 16, 128, 64)  0           conv2d_transpose_60[0][0]        
                                                                 up_sampling2d_25[0][0]           
__________________________________________________________________________________________________
up_sampling2d_26 (UpSampling2D) (None, 16, 256, 64)  0           add_52[0][0]                     
__________________________________________________________________________________________________
conv2d_transpose_61 (Conv2DTran (None, 16, 256, 64)  36928       up_sampling2d_26[0][0]           
__________________________________________________________________________________________________
conv2d_transpose_62 (Conv2DTran (None, 16, 256, 64)  36928       conv2d_transpose_61[0][0]        
__________________________________________________________________________________________________
add_53 (Add)                    (None, 16, 256, 64)  0           conv2d_transpose_62[0][0]        
                                                                 up_sampling2d_26[0][0]           
__________________________________________________________________________________________________
clustering_layer (ClusteringLay (None, 32)           2048        encoded[0][0]                    
__________________________________________________________________________________________________
decoded (Conv2DTranspose)       (None, 16, 256, 1)   577         add_53[0][0]                     
==================================================================================================
Total params: 977,417
Trainable params: 977,417
Non-trainable params: 0

Answer 1

您不需要 Flatten 层，因为输入是 2D。

删除flat = Flatten()(x)

x = Dense(256, activation='relu')(x)
encoded = Dense(64, activation='relu', name="encoded")(x)
x = Dense(256, activation='relu')(encoded)
x = Dense(1024, activation='relu')(x)
x = Reshape((4, 32, 8))(x)