ValueError：检查目标时出错：预期 conv2d_120 的形状为 (256, 256, 16) 但得到的数组形状为 (256, 256, 3)

Question

我的输入是 256x256 rgb 图像，我的自动编码器 output 想要是 256x256 rgb（但只有黑白颜色），输入和输出如下例如输入图像outputimage这是我的代码

train_data = np.empty((train_N,256,256,3))
train_labels = np.empty((trainL_N,256,256,3))
test_data = np.empty((test_N,256,256,3))
test_labels = np.empty((testL_N,256,256,3))
def loadIMG(imagePath , number, Array):
while number >0:
    img = cv2.imread(imagePath[number-1])
    img = cv2.resize(img,(256,256),interpolation=cv2.INTER_AREA)
    img_ndarray=np.asarray(img,dtype='float64')
    Array[number-1] = img_ndarray
    number = number - 1

loadIMG(imagePath1,train_N,train_data)
loadIMG(imagePath2,trainL_N,train_labels)
loadIMG(imagePath3,test_N,test_data)
loadIMG(imagePath4,testL_N,test_labels)
def train_model():

global history

input_img= Input(shape=(256, 256, 3))
#大小 = 256*256
x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
#大小 = 127*127
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
#大小 = 62*62
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
#大小 = 30*30
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)    
#大小 = 14*14
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)   
#大小 = 6*6
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
#大小 = 2*2
#這邊直接再次maxpooling 來達到1*1
encoded = MaxPooling2D((2, 2), padding='same', name='encoder')(x)   #大小 = 1*1 

x = UpSampling2D((2, 2))(encoded)
#大小2*2
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
#大小6*6
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)    
#大小14*14
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)    
#大小30*30
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
#大小62*62
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
#大小127*127
x = UpSampling2D((2, 2))(x)
decoded = Conv2D(16, (3, 3), activation='softmax', padding='same')(x)
    
autoencoder = Model(input_img, decoded)  

print(autoencoder.summary())
autoencoder.compile(optimizer='adam', loss='categorical_crossentropy',metrics=[tf.keras.metrics.CategoricalAccuracy()])

history = autoencoder.fit(train_data, train_labels,
                epochs=20,
                batch_size=24,
                shuffle=True,
                validation_data=(test_data, test_labels),
                callbacks=[TensorBoard(log_dir='/tmp/autoencoder', histogram_freq=0, write_graph=False)])

autoencoder.save('autoencoder.h5')

这是 model 摘要model 摘要

如何使 (256,256,16) 变回 (256,256,3)？ 我读过其他类似的问题，但没有找到关于我的情况的解决方案

Answer 1

如果您希望 output 成为灰度图像，您需要像这样更改 model 中的最后一层：

解码 = Conv2D(1, (3, 3), activation='softmax', padding='same')(x)

如果你想让你的 output 形状为 (256,256,3) 那么只需使用 3 作为过滤器的数量。

虽然不确定您要在这里解决什么问题。 缺少这里的上下文..