不兼容的形状：[128,1] 与 [128,3,3]

Question

我正在尝试创建一个 CNN 来对 SVHN 数据集进行分类，但在创建我的模型时遇到了不兼容的形状错误：不兼容的形状：[128,3,3,10] 与 [128,1]。我该如何解决？

         model = Sequential([
                          Conv2D(filters=8, kernel_size=(3, 3), 
                           activation='relu', input_shape=(32, 32,3 
                           name='conv_1'),
                            
                           Conv2D(filters=8, kernel_size=(3, 3), 
                          activation='relu', padding= 'SAME',  
                                   `name='conv_2'),
                           MaxPooling2D(pool_size=(8, 8), name='pool_1'),
                            Dense(64,  kernel_regularizer = 
                            regularizers.l2(0.5),bias_initializer='ones',
                            activation='relu' , name='dense_1'),
                            Dropout(0.3),  
                            Dense(64,kernel_regularizer = 
                            regularizers.l2(0.5) , activation='relu' 
                           ,name='dense_2'),
                            BatchNormalization(),  
                            Dense(64,  kernel_regularizer = 
                            regularizers.l2(0.5) ,  activation='relu' 
                             ,name='dense_3'),
                             Dense(10,  activation='softmax' 
                             ,name='dense_4')
             ])


           model.compile(
           optimizer = 'adam',
           loss = 'sparse_categorical_crossentropy',
           metrics= ['accuracy' ])
   

           history = model.fit(train_images,train_labels , epochs = 30 
           ,validation_split = 0.15,
                batch_size= 128, verbose = False )

Answer 1

在最后一个Dense 层之前放置一个Flatten 层。因为您没有这样做，所以在为您提供类的层之前，张量不会减少为单维张量。

TensorFlow 中的一般模式是在输出该类的层之前使用Flatten。

另外，我已经删除了您放置的随机密集层中的BatchNormalization，BatchNormalization 层通常放在 Conv 层之后，但是您可以将它们放在 Dense 层之后。如果你是 BatchNormalization，确保整个网络或相关的都有它。不要只放一个随机的BatchNormalization 层。

以下是您更改代码的方法。

 model = Sequential([Conv2D(filters=8, kernel_size=(3, 3), 
                           activation='relu', input_shape=(32, 32,3), 
                           name='conv_1'),
                     BatchNormalization(),
                     Conv2D(filters=8, kernel_size=(3, 3), 
                          activation='relu', padding= 'SAME',  
                          name='conv_2'),
                     BatchNormalization(),
                     MaxPooling2D(pool_size=(8, 8), name='pool_1'),
                     Flatten(),
                     Dense(64,  kernel_regularizer = 
                           regularizers.l2(0.5), bias_initializer='ones',
                           activation='relu' , name='dense_1'),
                     Dropout(0.3),  
                     Dense(64,kernel_regularizer = 
                           regularizers.l2(0.5) , activation='relu', 
                           name='dense_2'),
                     Dense(64,  kernel_regularizer = 
                           regularizers.l2(0.5) ,  activation='relu', 
                           name='dense_3'),
                     Dense(10,  activation='softmax', 
                           name='dense_4')
             ])


           model.compile(
           optimizer = 'adam',
           loss = 'sparse_categorical_crossentropy',
           metrics= ['accuracy' ])
   

           history = model.fit(train_images,train_labels , epochs = 30)

Answer 2

我认为你的代码有两个问题。

首先请确认train_labels的形状。 Incompatible shapes 如果张量的形状错误，就会出现错误。我认为[128, 1] 的形状意味着train_label 不是一种单热向量。如果 train_label 形状如 [1, 3, 8, ..]，则应将形状更改为 [[0, 1, 0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 1, 0, 0, 0, 0, 0, 0], ...]

其次，您应该如上所述在Dense层之前添加Flatten层。

 model = Sequential([
       Conv2D(filters=8, kernel_size=(3, 3), activation='relu', input_shape=(32, 32, 3), name='conv_1'),
       ...
       MaxPooling2D(pool_size=(8, 8), name='pool_1'),
       
       Flatten(),

       Dense(64,  kernel_regularizer ...
             ])

---

(32, 32, 1) 的形状表示输入形状的最后一个暗淡应该是 1。所以你应该把Conv2D的input_shape改成(32, 32, 1)

Conv2D(filters=8, kernel_size=(3, 3), activation='relu', input_shape=(32, 32, 1) ...

另外，train_images 也应该改为(32, 32, 1)，因为图像的通道是一个。

train_images = tf.expand_dims(train_images, -1)

另外，你可以像这样得到train_labels的one-hot向量：

train_labels = tf.squeeze(train_labels)
train_labels = tf.one_hot(train_labels, depth=10)