Question

我尝试遵循Keras tutorial来为MNIST构建自动编码器。自动编码器正常工作，然后我尝试将图像更改，因此将形状从28, 28, 1更改为150, 150, 3，然后收到以下错误：

ValueError：检查目标时出错：预期conv2d_6具有形状（148，148，1），但数组的形状为（150，150，3）

自动编码器体系结构：

input_img = Input(shape=(150, 150, 3))

x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)

x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation='relu')(x)
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)

autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer=Adam(0.01), loss='binary_crossentropy')

火车设置：

autoencoder.fit(x_train, y_train,
                epochs=50,
                batch_size=512,
                shuffle=True,
                validation_data=(x_test, y_test))

我的数据形状如下：

x_train shape: (4022, 150, 150, 3)
y_train shape: (4022, 150, 150, 3)
x_test shape: (447, 150, 150, 3)
y_test shape: (447, 150, 150, 3)

到我的工作空间的协作链接：

https://colab.research.google.com/drive/1C8RX7OYS2BXaHJh6VOMscxEbrFTuQY5H

Answer 1

使用此代码，它将起作用

input_img = Input(shape=(150, 150, 3))
x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
encoded = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = ZeroPadding2D(padding=(1, 1), input_shape=(148, 148, 16))(x)
x = Conv2D(16, (3, 3), activation='relu')(x)
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(3, (3, 3), activation='sigmoid', padding='valid')(x)
autoencoder = Model(input_img, decoded)
autoencoder.compile(optimizer=Adam(0.01), loss='binary_crossentropy')

我添加了零填充并更改了最后一层转换，以输出3个通道

这将打印以下摘要

Layer (type)                 Output Shape              Param #   
=================================================================
input_10 (InputLayer)        (None, 150, 150, 3)       0         
_________________________________________________________________
conv2d_58 (Conv2D)           (None, 150, 150, 16)      448       
_________________________________________________________________
max_pooling2d_25 (MaxPooling (None, 75, 75, 16)        0         
_________________________________________________________________
conv2d_59 (Conv2D)           (None, 75, 75, 8)         1160      
_________________________________________________________________
max_pooling2d_26 (MaxPooling (None, 38, 38, 8)         0         
_________________________________________________________________
conv2d_60 (Conv2D)           (None, 38, 38, 8)         584       
_________________________________________________________________
max_pooling2d_27 (MaxPooling (None, 19, 19, 8)         0         
_________________________________________________________________
conv2d_61 (Conv2D)           (None, 19, 19, 8)         584       
_________________________________________________________________
up_sampling2d_25 (UpSampling (None, 38, 38, 8)         0         
_________________________________________________________________
conv2d_62 (Conv2D)           (None, 38, 38, 8)         584       
_________________________________________________________________
up_sampling2d_26 (UpSampling (None, 76, 76, 8)         0         
_________________________________________________________________
zero_padding2d_4 (ZeroPaddin (None, 78, 78, 8)         0         
_________________________________________________________________
conv2d_63 (Conv2D)           (None, 76, 76, 16)        1168      
_________________________________________________________________
up_sampling2d_27 (UpSampling (None, 152, 152, 16)      0         
_________________________________________________________________
conv2d_64 (Conv2D)           (None, 150, 150, 3)       435       
=================================================================
Total params: 4,963
Trainable params: 4,963
Non-trainable params: 0
_________________________________________________________________
None

Answer 2

我认为您的最终解码层应解码为三个通道，而不是一个。

decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)

应该是

decoded = Conv2D(3, (3, 3), activation='sigmoid', padding='same')(x)

不？

自动编码器适用于MNIST，但对于较大尺寸的图像会崩溃

2 个答案: