CNN图像到图像的回归输出损失非常高99.3％

问题描述

我有一个深度学习模型，该模型将4d图像作为输入并预测1D图像。但是我的损失很高。谁能帮我找出原因。

样本输入图像： [第1维[] [1]] [1] + [第2维] [2] + [第3维] [3] + [第4维] [4] ====输出[所需的输出图像] [5]

包含输出图像的信息很少。

我使用RMSE来计算tf.keras的损失。它似乎没有收敛。

这是我的损失的样子：

第1/5集 25/27 [========================= ......]-ETA：1：16-损失：99.7717-acc：0.0000e +00

模型架构和模型拟合代码如下：

def unet(pretrained_weights = None,input_size = (512,512,4)):
    inputs = Input(input_size)
    conv1 = Conv2D(64,3,activation = 'relu',padding = 'same',kernel_initializer = 'he_normal')(inputs)
    #conv1 = Conv2D(64,kernel_initializer = 'he_normal')(conv1)
    pool1 = MaxPooling2D(pool_size=(2,2))(conv1)
    conv2 = Conv2D(128,kernel_initializer = 'he_normal')(pool1)
    #conv2 = Conv2D(128,kernel_initializer = 'he_normal')(conv2)
    pool2 = MaxPooling2D(pool_size=(2,2))(conv2)
    conv3 = Conv2D(256,kernel_initializer = 'he_normal')(pool2)
    #conv3 = Conv2D(256,kernel_initializer = 'he_normal')(conv3)
    pool3 = MaxPooling2D(pool_size=(2,2))(conv3)
    conv4 = Conv2D(512,kernel_initializer = 'he_normal')(pool3)
    #conv4 = Conv2D(512,kernel_initializer = 'he_normal')(conv4)
    drop4 = Dropout(0.5)(conv4)
    pool4 = MaxPooling2D(pool_size=(2,2))(drop4)

    conv5 = Conv2D(1024,kernel_initializer = 'he_normal')(pool4)
    #conv5 = Conv2D(1024,kernel_initializer = 'he_normal')(conv5)
    drop5 = Dropout(0.5)(conv5)

    up6 = Conv2DTranspose(512,(2,2),strides=(2,padding='same')(drop5)
    #merge6 = concatenate([drop4,up6],axis = 3)
    merge6=up6
    conv6 = Conv2D(512,kernel_initializer = 'he_normal')(merge6)
    #conv6 = Conv2D(512,kernel_initializer = 'he_normal')(conv6)

    up7 =Conv2DTranspose(256,padding='same')(conv6)
    #merge7 = concatenate([conv3,up7],axis = 3)
    merge7=up7
    conv7 = Conv2D(256,kernel_initializer = 'he_normal')(merge7)
    #conv7 = Conv2D(256,kernel_initializer = 'he_normal')(conv7)

    #up8 = Conv2D(128,2,kernel_initializer = 'he_normal')(Conv2DTranspose(1,(3,3),padding='same')(conv7))
    up8 = Conv2DTranspose(128,padding='same')(conv7)
    #merge8 = concatenate([conv2,up8],axis = 3)
    merge8=up8
    Conv2DTranspose(1,padding='same')
    conv8 = Conv2D(128,kernel_initializer = 'he_normal')(merge8)
    #conv8 = Conv2D(128,kernel_initializer = 'he_normal')(conv8)

    #up9 = Conv2D(64,padding='same')(conv8))
    up9 = Conv2DTranspose(64,padding='same')(conv8)
    #merge9 = concatenate([conv1,up9],axis = 3)
    merge9=up9
    conv9 = Conv2D(64,kernel_initializer = 'he_normal')(merge9)
    #conv9 = Conv2D(64,kernel_initializer = 'he_normal')(conv9)
    #conv9 = Conv2D(2,kernel_initializer = 'he_normal')(conv9)
    conv10 = Conv2D(1,1,activation = 'sigmoid')(conv9)
    conv11 = tf.keras.layers.Reshape((512,512))(conv10)
    model = tf.keras.Model(inputs,conv11)
    
    model.compile(optimizer = Adam(lr = 1e-4),loss = 'binary_crossentropy',metrics = ['accuracy'])
    
    #model.summary()

    if(pretrained_weights):
        model.load_weights(pretrained_weights)

    return model

model=unet()
model.compile(loss=tf.keras.losses.MeanAbsolutePercentageError(),optimizer=tf.keras.optimizers.Adadelta(),metrics = ['accuracy'])

model.fit(train_ds,epochs=5,verbose=1,validation_data=validation_ds)


  [1]: https://i.stack.imgur.com/h9x3C.png
  [2]: https://i.stack.imgur.com/0LjZ4.png
  [3]: https://i.stack.imgur.com/qu0cm.png
  [4]: https://i.stack.imgur.com/ZiKlg.png
  [5]: https://i.stack.imgur.com/f0izQ.png

解决方法

我认为您可以执行以下一些检查来找出问题所在：

• 检查目标图像是否正确加载
• 更改优化程序并尝试学习率
• 再培训几个纪元，看看损失是否正在收敛
• 有时损失函数的收敛性较差。我看到您正在使用mean_absolute_percentage_error。它不是一个平稳的损失函数，可能会出现收敛问题。可能是使用rmse训练了几个时期的火车模型，然后切换了